Merge pull request #28 from NVIDIA/cutlass_1.1

Fixed typeo

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "tools/external/googletest"]
+	path = tools/external/googletest
+	url = https://github.com/google/googletest.git

CHANGELOG.md

@@ -0,0 +1,63 @@
+# NVIDIA CUTLASS Changelog
+
+
+## 1.1.0 (2018-09-19)
+  * Turing Features
+    * WMMA GEMM targeting TensorCores - INT8, INT4, 1-bit
+  * Batched Strided GEMM
+  * Threadblock rasterization strategies
+    * Improved performance for adverse problem sizes and data layouts
+  * Extended CUTLASS Core comonents
+    * Tensor views support arbitrary matrix and tensor layouts
+    * Zip iterators for structuring multiple data streams
+  * Enhanced CUTLASS utilities
+    * Reference code for tensor operations in host and device code
+    * Added HostMatrix<> for simplified matrix creation
+  * Examples
+    * Basic GEMM, tensor views, CUTLASS utilities, batched GEMM, WMMA GEMM
+
+## 1.0.1 (2018-06-11)
+
+  * Intra-threadblock reduction added for small threadblock tile sizes
+    * sgemm_64x128x16, sgemm_128x128x16, sgemm_128x64x16, sgemm_128x32x16, sgemm_64x64x16, sgemm_64x32x16
+    * igemm_32x32x128
+  * GEMM _K_ residue handled during prologue prior to mainloop
+  * Replaced Google Test copy with submodule. Use `git submodule init`
+
+## [1.0.0](https://github.com/NVIDIA/cutlass/commit/2028ebe120aab22bfd0b2baf8902d4c9627eb33f) (2018-05-16)
+
+  * Substantial rewrite to accommodate new architecture
+  * Kernels: SGEMM, DGEMM, IGEMM, HGEMM, WMMA GEMM
+  * Unit and performance tests
+
+## [0.0.1](https://github.com/NVIDIA/cutlass/commit/d08ba8ac46e2fa3f745e070c390182edb56b2e91) (2017-12-04)
+
+  * Initial release
+
+
+## Copyright
+
+Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+
+```
+  Redistribution and use in source and binary forms, with or without modification, are permitted
+  provided that the following conditions are met:
+      * Redistributions of source code must retain the above copyright notice, this list of
+        conditions and the following disclaimer.
+      * Redistributions in binary form must reproduce the above copyright notice, this list of
+        conditions and the following disclaimer in the documentation and/or other materials
+        provided with the distribution.
+      * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+        to endorse or promote products derived from this software without specific prior written
+        permission.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+  IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+  FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+  FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+  BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+  OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+  STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+```
+

CMake/bin2hex.cmake

@@ -0,0 +1,26 @@
+# A small utility function which generates a C-header from an input file
+function(FILE_TO_C_STRING FILENAME VARIABLE_NAME OUTPUT_STRING ZERO_TERMINATED)
+  FILE(READ "${FILENAME}" HEX_INPUT HEX)
+  if (${ZERO_TERMINATED})
+    string(APPEND HEX_INPUT "00")
+  endif()
+
+  string(REGEX REPLACE "(....)" "\\1\n" HEX_OUTPUT ${HEX_INPUT})
+  string(REGEX REPLACE "([0-9a-f][0-9a-f])" "0x\\1," HEX_OUTPUT ${HEX_OUTPUT})
+
+  set(HEX_OUTPUT "static char const ${VARIABLE_NAME}[] = {\n  ${HEX_OUTPUT}\n};\n")
+
+  set(${OUTPUT_STRING} "${HEX_OUTPUT}" PARENT_SCOPE)
+endfunction()
+
+message("Create header file for ${FILE_IN}")
+message("Create header file for ${FILE_OUT}")
+file_to_c_string(${FILE_IN} ${VARIABLE_NAME} OUTPUT_STRING ZERO_TERMINATED)
+
+set(RESULT "#pragma once\n")
+string(APPEND RESULT "namespace cutlass {\n")
+string(APPEND RESULT "namespace nvrtc {\n")
+string(APPEND RESULT "${OUTPUT_STRING}")
+string(APPEND RESULT "} // namespace nvrtc\n")
+string(APPEND RESULT "} // namespace cutlass\n")
+file(WRITE "${FILE_OUT}" "${RESULT}")

CMakeLists.txt

@@ -0,0 +1,227 @@
+# Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without modification, are permitted
+# provided that the following conditions are met:
+#     * Redistributions of source code must retain the above copyright notice, this list of
+#       conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above copyright notice, this list of
+#       conditions and the following disclaimer in the documentation and/or other materials
+#       provided with the distribution.
+#     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+#       to endorse or promote products derived from this software without specific prior written
+#       permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+# IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+# FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+# BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+# OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+# STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+cmake_minimum_required(VERSION 3.3.0)
+
+set(CUTLASS_LANGUAGES CXX)
+
+# CMake 3.9.0 has native support for CUDA without the need of the CUDA package. Use it!
+if(WIN32 AND NOT ${CMAKE_VERSION} VERSION_LESS "3.9.0")
+  list(APPEND CUTLASS_LANGUAGES CUDA)
+  set(CUTLASS_NATIVE_CUDA TRUE)
+
+  macro(cutlass_add_executable)
+    add_executable(${ARGN})
+  endmacro()
+else()
+  # FindCUDA fails to detect VS 2017 due to a changed directory format of the toolkits.
+  # For this configuration we need CMake >= 3.9.0 to use the native CUDA support.
+  if (WIN32 AND MSVC_VERSION GREATER 1800)
+    message(FATAL_ERROR "Please upgrade CMake to version >= 3.9.0 to support Visual Studio 2017 or higher")
+  endif()
+
+  # Fall back to the FindCUDA version to create an executable with CUDA files
+  macro(cutlass_add_executable)
+    cuda_add_executable(${ARGN})
+  endmacro()
+endif()
+
+project(CUTLASS ${CUTLASS_LANGUAGES})
+
+# check if the configuration is supported
+if( NOT CMAKE_SIZEOF_VOID_P EQUAL 8 )
+    message(FATAL_ERROR "CUTLASS requires a 64-bit compiler!")
+endif()
+
+find_package(CUDA)
+find_package(Doxygen QUIET)
+
+###################################################################################################
+#
+# Configure CMake variables
+#
+###################################################################################################
+
+find_library(CUBLAS_LIBRARY cublas HINTS
+                                   ${CUDA_TOOLKIT_ROOT_DIR}/lib64
+                                   ${CUDA_TOOLKIT_ROOT_DIR}/lib/x64)
+
+# By default we want to build in Release mode to ensure that we're getting best performance
+if (NOT (CMAKE_BUILD_TYPE OR CONFIGURATION_TYPES))
+  set(CMAKE_BUILD_TYPE Release CACHE STRING "Choose build level" FORCE)
+  # We do support Debug or Release builds
+  set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "RelWithDebInfo" "Release")
+endif()
+
+if(WIN32)
+  # On Windows we link against the shared (DLL) runtime. Change gtest settings to match this.
+  set(gtest_force_shared_crt ON CACHE BOOL "Use shared (DLL) run-time lib even when Google Test is built as static lib" FORCE)
+endif()
+
+if (WIN32)
+    # Enable more warnings and treat as errors
+    string(APPEND NVCC_FLAGS " -Xcompiler /W3 -Xcompiler /WX")
+
+    # Disable warning on Unicode characters
+    string(APPEND NVCC_FLAGS " -Xcompiler /wd4819")
+
+    # Disable excess x86 floating point precision that can lead to results being labeled incorrectly
+    string(APPEND NVCC_FLAGS " -Xcompiler /fp:strict")
+
+    # Verbose option
+    if (${CUTLASS_NVCC_VERBOSE})
+        string(APPEND NVCC_FLAGS " -v")
+    endif()
+endif(WIN32)
+
+set(CUTLASS_NVCC_ARCHS "50;60;61;70;75" CACHE STRING "The SM architectures to build code for.")
+set(CUTLASS_NVCC_EMBED_CUBIN ON CACHE BOOL "Embed compiled CUDA kernel binaries into executables.")
+set(CUTLASS_NVCC_EMBED_PTX ON CACHE BOOL "Embed compiled PTX into executables.")
+set(CUTLASS_NVCC_KEEP OFF CACHE BOOL "Keep intermediate files generated by NVCC.")
+
+#
+# NOTE: running with asan and CUDA requires the following environment variable:
+#
+#  ASAN_OPTIONS=protect_shadow_gap=0:replace_intrin=0:detect_leaks=0
+#
+# without the above environment setting, an error like the following may be generated:
+#
+#  *** Error: Could not detect active GPU device ID [out of memory]
+#  ...
+#  ==9149==ERROR: LeakSanitizer: detected memory leaks
+#  ...
+#
+if(ENABLE_ASAN)  # https://github.com/google/sanitizers/wiki/AddressSanitizer
+  string(APPEND NVCC_FLAGS " --compiler-options -fsanitize=address --compiler-options -fno-omit-frame-pointer")
+  string(APPEND CMAKE_EXE_LINKER_FLAGS " -fsanitize=address")
+endif()
+
+###################################################################################################
+#
+# Configure CUDA build options
+#
+###################################################################################################
+
+# Set NVCC arguments
+foreach(ARCH ${CUTLASS_NVCC_ARCHS})
+  if(CUTLASS_NVCC_EMBED_CUBIN)
+    string(APPEND NVCC_FLAGS " -gencode arch=compute_${ARCH},code=sm_${ARCH}")
+  endif()
+  if(CUTLASS_NVCC_EMBED_PTX)
+    string(APPEND NVCC_FLAGS " -gencode arch=compute_${ARCH},code=compute_${ARCH}")
+  endif()
+endforeach()
+
+if (CUTLASS_NVCC_KEEP)
+    string(APPEND NVCC_FLAGS " -keep")
+endif()
+
+if (WIN32 AND CUTLASS_NATIVE_CUDA)
+  string(APPEND NVCC_FLAGS_RELEASE " -lineinfo")
+else()
+  string(APPEND NVCC_FLAGS " -lineinfo")
+endif()
+
+string(APPEND NVCC_FLAGS_DEBUG " -g")
+string(APPEND NVCC_FLAGS_RELWITHDEBINFO " -O3")
+string(APPEND NVCC_FLAGS_RELEASE " -O3")
+
+# define NDEBUG for release mode to disable assertions
+string(APPEND NVCC_FLAGS_RELEASE " -DNDEBUG")
+
+if (CUTLASS_NATIVE_CUDA)
+  set(CMAKE_CUDA_FLAGS "${NVCC_FLAGS}")
+  set(CMAKE_CUDA_FLAGS_RELEASE "${NVCC_FLAGS_RELEASE}")
+  set(CMAKE_CUDA_FLAGS_RELWITHDEBINFO "${NVCC_FLAGS_RELWITHDEBINFO}")
+  set(CMAKE_CUDA_FLAGS_DEBUG "${NVCC_FLAGS_DEBUG}")
+else()
+  set(CUDA_NVCC_FLAGS ${NVCC_FLAGS})
+  set(CUDA_NVCC_FLAGS_DEBUG ${NVCC_FLAGS_DEBUG})
+  set(CUDA_NVCC_FLAGS_RELWITHDEBINFO ${NVCC_FLAGS_RELWITHDEBINFO})
+  set(CUDA_NVCC_FLAGS_RELEASE ${NVCC_FLAGS_RELEASE})
+endif()
+
+#
+# The following items should eventually be pushed into cutlass/CMakeLists.txt
+#
+
+# GLOB for CUTLASS header files. Should we use a static list instead?
+file(GLOB CUTLASS_GEMM RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} cutlass/gemm/*.h)
+file(GLOB CUTLASS_UTIL RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} cutlass/util/*.h)
+file(GLOB CUTLASS_DEVICE RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} cutlass/device/*.h)
+file(GLOB CUTLASS_CORE RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} cutlass/*.h)
+###################################################################################################
+#
+# Define build targets
+#
+###################################################################################################
+
+source_group("cutlass\\gemm" FILES ${CUTLASS_GEMM})
+source_group("cutlass\\util" FILES ${CUTLASS_UTIL})
+source_group("cutlass\\device" FILES ${CUTLASS_DEVICE})
+source_group("cutlass" FILES ${CUTLASS_CORE})
+
+add_library(CUTLASS INTERFACE)
+include_directories("${CMAKE_CURRENT_SOURCE_DIR}")
+target_sources(CUTLASS INTERFACE
+  ${CUTLASS_GEMM}
+  ${CUTLASS_UTIL}
+  ${CUTLASS_DEVICE}
+  ${CUTLASS_CORE}
+)
+
+target_include_directories(CUTLASS INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
+
+# Create a custom target to ensure that the CUTLASS sources are visible in an IDE
+add_custom_target(cutlass_ide SOURCES
+  ${CUTLASS_GEMM}
+  ${CUTLASS_UTIL}
+  ${CUTLASS_DEVICE}
+  ${CUTLASS_CORE}
+)
+# Doxygen is available. Generate documentation
+if (DOXYGEN_FOUND)
+    # DOT is available. Enable graph generation in the documentation
+    if (DOXYGEN_DOT_EXECUTABLE)
+        set(CUTLASS_ENABLE_DOXYGEN_DOT ON CACHE BOOL "Use dot to generate graphs in the doxygen documentation.")
+    else()
+        set(CUTLASS_ENABLE_DOXYGEN_DOT OFF CACHE BOOL "Use dot to generate graphs in the doxygen documentation." FORCE)
+    endif()
+
+    if (CUTLASS_ENABLE_DOXYGEN_DOT)
+        set(HAVE_DOT "YES")
+    else()
+        set(HAVE_DOT "NO")
+    endif()
+
+    # Add custom target for Doxygen.
+    add_custom_target(cutlass_docs ${CMAKE_COMMAND} -E env
+        "DOT_PATH=${DOXYGEN_DOT_EXECUTABLE}"
+        "HAVE_DOT=${HAVE_DOT}"
+        ${DOXYGEN_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/Doxyfile
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        VERBATIM
+    )
+endif()
+
+add_subdirectory(tools)
+add_subdirectory(examples)

CUTLASS.md

@@ -0,0 +1,338 @@
+![ALT](/media/images/gemm-hierarchy-with-epilogue-no-labels.png "Complete CUDA GEMM decomposition")
+
+# CUTLASS
+
+This document is intended to accompany the CUTLASS source code, to describe the interaction between
+CUTLASS core components, and to identify their role in implementing GEMM computations efficiently in CUDA.
+
+1. [Design Patterns](#S-design-patterns)
+2. [General Matrix Multiply](#S-general-matrix-multiply)
+3. [Core Components](#S-core-components)
+4. [Utilities](#S-utilities)
+
+# <a name="S-design-patterns"></a> 1. Design Patterns
+
+CUTLASS strives to achieve the highest performance possible on NVIDIA GPUs while also offering a
+flexible composition that an be easily applied to solve new problems related to Deep Learning and
+linear algebra. Though we intend to make CUTLASS as simple and straightforward as possible, given
+a tradeoff between simplicity and performance, CUTLASS chooses performance. Consequently, several
+design patterns are necessary to yield a composable structure while also satisfying these performance
+objectives. This section is intended to provide more detail.
+
+* [Sequencing and Nesting](#S-patterns-sequencing-nesting)
+* [Tiles and Iterators](#S-patterns-tiles-iterators)
+* [Host-side Params](#S-patterns-host-side-params)
+* [Composable Shared Memory](#S-patterns-composable-shared-memory)
+
+## <a name="S-patterns-sequencing-nesting"></a> Sequencing and Nesting of Collective Primitives
+
+CUTLASS embodies a design paradigm exemplified by the [CUB library](https://nvlabs.github.io/cub/) for expressing collective operations. Objects expose an interface for a problem that is then decomposed into concurrent subtasks executed by cooperating threadblocks, warps, and threads. For example, a grid-level object may be constructed with base pointers to the start of a GEMM operation, add a threadblock-dependent offset to partition the problem, and then compute a per-threadblock GEMM. This in turn performs some operations as a collection of cooperating threads, while it may partition other parts of the task into warp-level subtasks. 
+
+## <a name="S-patterns-tiles-iterators"></a> Tiles and Iterators
+
+Efficient dense linear algebra computations emphasize data movement to match the execution of mathemtical operators to the flow of data. Consequently, CUTLASS defines a rich set of primitives for partitioning a tile of data among participating threads, warps, and threadblocks. CUTLASS applies the familiar iterator design pattern to provide an abstraction layer to (1.) access these tile objects and (2.) traverse a sequence of objects embedded in a higher level data structure. These subpartitions are typically defined by compile-time constants
+specifying element type, size, and data layout. CUTLASS refers to subpartitions as _tiles_.
+
+_Iterators_ are familiar design patterns in C++ that provide an abstraction for accessing individual
+elements in memory as well as traversing over a collection. GEMM kernels in CUTLASS depend on accessing
+a sequence of tiles from global memory, from shared memory, and in registers. Consequently, _tile iterators_
+are prevalent throughout the CUTLASS implementation.
+
+The canonical CUTLASS tile iterator template is defined in [cutlass/tile_iterator.h](cutlass/tile_iterator.h).
+
+## <a name="S-patterns-host-side-params"></a> Host-side Params structure
+
+Several CUTLASS template classes exhibit a pattern in which problem-specific internal state is known at kernel launch time and remains invariant throughout the execution of a kernel. For example, tile iterators compute several offsets based on the strides of the input tensor that is added to an internal pointer when loading the elements of a tile. These are computed from the tensor stride and never updated; the per-thread internal state consists only of the internal global memory pointer.
+
+CUTLASS can take advantage of this CUDA grid-invariant property by constructing the object in host code and passing a composed parameters structure to the kernel. This confers two benefits: (1.) invariant state is held in constant memory, and (2.) there is no overhead to compute the initial state by each thread.
+
+The design pattern in CUTLASS is for classes with nontrivial constructors to define `struct Params` as an inner class which contains grid-invariant state. These should define a constructor and an `initialize()` method. The `Params` structure should also include a data member corresponding to each data member in the parent class, so these too can be properly constructed in host code. The parent class should define a constructor which accepts `Params const &` as its first argument.
+
+For example, `cutlass::gemm::Gemm<>` should define `struct cutlass::gemm::Gemm::Params`. The latter should define data members for each data member in `cutlass::gemm::Gemm<>`. 
+
+
+## <a name="S-patterns-composable-shared-memory"></a> Composable shared memory allocation
+
+Shared memory requires explicit effort by the programmer to allocate and de-allocate. CUTLASS follows the paradigm introduced by [CUB](https://nvlabs.github.io/cub/) to define composed structures for storing data intended to be held in shared memory. Any object requiring shared memory storage for itself or its data members should define a child structure called SharedStorage. This holds data needed by the class and also instantiates SharedStorage objects for each data member.
+
+To be consistent, this pattern defines a convention in which classes define internal shared memory storage requirements. Classes should consider all SharedStorage structures to be opaque other than their own child class. When the lifetimes of child objects are known to be non-overlapping, unions may be used to alias multiple SharedStorage objects to the same shared memory region and reduce overall SMEM capacity.
+
+## <a name="S-patterns-loop-unrolling"></a> Loop Unrolling
+
+CUTLASS requires tiles of data to be stored in registers for high-bandwidth access. Simultaneously, high-throughput math instructions
+must be issued concurrently with memory instructions to hide latency with relatively few concurrent threads. These objectives are
+achieved by unrolling loops whose iteration counts are known at compile time.
+
+Consequently, most loops within the CUTLASS GEMM implementation are specified by constant values and template arguments. The CUDA compiler
+is able to unroll the loop bodies, map array elements to registers, and construct an efficient instruction schedule.
+
+## <a name="S-patterns-loop-unrolling"></a> Templates
+
+CUDA C++ templates and modern generic programming techniques enable CUTLASS device code to span a large design space.
+
+This design space includes:
+* Mixed precision arithmetic and data storage
+* Kernels specialized for layout and problem size
+* Support for kernel fusion
+
+Moreover, templates provided a structured approach to collecting compile-time constants such as tile dimensions. These
+must be template arguments to target static array allocation and take advantage of loop unrolling, constant folding,
+and function inlining.
+
+# <a name="S-general-matrix-multiply"></a> 2. General Matrix Multiply
+
+The following figure illustrates the hierarchical GEMM computation embodied by CUTLASS. Each stage depicts a nested level of tiling which corresponds to a layer of concurrency within the CUDA execution model and to a level within the memory hierarchy, becoming increasingly finer moving left to right.
+
+![ALT](/media/images/gemm-structural-components.png "CUTLASS GEMM Structural Components")
+
+## Threadblock-level GEMM
+
+The CUTLASS GEMM kernel partitions the _C_ matrix into a 2D tiling of threadblocks.
+Each threadblock computes a matrix product whose outer dimensions _M_ and _N_ are compile-time constants. The
+GEMM's _K_ dimension is partitioned into tiles and iterated over by the GEMM _mainloop_. The shape of the matrix
+multiply operation performed by each iteration of the mainloop is referred to as _OutputTile_.
+
+The threadblock loads a sequence of tiles from global memory and stores this data to shared memory. The iterative
+access and traversal of tiles in global memory are performed by a _TileLoadIterator_, and storing to a circular
+buffer in shared memory is performed by a _GlobalLoadIterator_. 
+
+**[Global Load Stream](cutlass/gemm/gemm_global_stream.h)** manages loading of the threadblock-scope multiplicands to the GEMM kernel. It owns an iterator into global memory for loading tiles of data, a TensorAllocation in shared memory to hold the resulting tile, and an iterator for writing the tile into this allocation. A transformer exists to optionally transform the data as it is loaded which may of use to perform type conversion or, in the case of int8 GEMM, transpose 4x4 tiles held in registers.
+
+The Global Load Stream template contains members defined by the following templates:
+
+* [GemmGlobalIteratorAb](cutlass/gemm/gemm_global_tile.h)
+* [Transformer](cutlass/convert.h)
+* [GemmSharedStoreTileAb](cutlass/gemm/gemm_shared_tile.h)
+
+## Warp-level GEMM
+
+The threadblock's _OutputTile_ is partitioned among the warps, and each computes a warp-level matrix product.
+Data is loaded from shared memory into registers, and math instructions are dispatched to CUDA Cores or Tensor Cores.
+
+[**Shared Load Stream**](cutlass/gemm/gemm_shared_stream.h) manages loading of warp-level multiplicands from shared memory into registers. This owns an iterator for fetching data and the destination fragments for holding the results. 
+
+* [GemmSharedLoadTile{A,B}](cutlass/gemm/gemm_shared_tile.h)
+
+**Matrix Multiply** computes a matrix product operation on data held in registers. Specializations exist for thread-level instructions such as single-precision fused multiply-add as well as warp-level matrix operations targeting TensorCores. 
+
+* [WMMA Multiply Add](cutlass/gemm/wmma_gemm_multiply_add.h)
+
+## Thread-level GEMM
+
+SGEMM, IGEMM, HGEMM, and DGEMM are computed by SIMT math instructions issued by thread-level matrix multiply
+procedures. 
+
+* [ThreadMultiplyAdd](cutlass/gemm/thread_multiply_add.h)
+* [IGEMM specialization](cutlass/gemm/igemm_multiply_add.h)
+* [HGEMM specialization](cutlass/gemm/hgemm_multiply_add.h)
+
+## Epilogue 
+
+The [**epilogue**](cutlass/gemm/gemm_epilogue.h) iteratively selects a subset of accumulator elements held by a warp, writes them to shared memory, and loads them by different threads such that a threadblock-scoped tile store operation will make contiguous, striped accesses to global memory. Thus, the flow of data utilizes the following components:
+
+1. [Transformer](cutlass/convert.h) for converting the data types of accumulator elements
+2. [GemmSharedStoreTileD](cutlass/gemm/gemm_shared_tile.h) to store to shared memory specialized to the accumulator layout.
+3. [GemmSharedLoadTileD](cutlass/gemm/gemm_shared_tile.h) to load the data from shared memory.
+4. [GemmGlobalIteratorC](cutlass/gemm/gemm_global_tile.h) to load a tile from global memory.
+5. A [functor](cutlass/gemm/linear_scaling.h) to compute an element-wise operation on the matrix product and source data (such as alpha*AB+beta*C).
+6. [GemmGlobalIteratorD](cutlass/gemm/gemm_global_tile.h) to write the output to global memory.
+
+## GEMM Traits
+
+[**cutlass::gemm::GemmTraits**](cutlass/gemm/gemm_traits.h) collects the structural properties of a complete GEMM computation into a single template class. As a result, the Traits classes encapsulate the the iterators and transformers for all supported GEMM operands and layouts. Low-level details needed by Traits (such as scalar types for operands, thread-block tile size, number of scalar elements per memory access within each phase, number of stages in shared memory, as well as other implementation-specific properties of the GEMM computation) are specified in class [**cutlass::gemm::GemmConfig**](cutlass/gemm/gemm_config.h).
+
+
+# <a name="S-core-components"></a> 3. Core Components
+
+CUTLASS GEMM kernels are implemented by a set of Core components for interacting with mathematical tensor and matrix
+objects as well as constructing efficient CUDA kernels.
+
+* [Tensor views](#S-core-tensor-views)
+* [Shape](#S-core-shape)
+* [Tile structure](#S-core-tile-structure)
+* [Fragment](#S-core-fragment)
+* [Predicate vector](#S-core-predicate-vector)
+
+## <a name="S-core-tensor-views"></a> Tensor View
+
+Matrices and tensors are typically represented as n-D arrays held in linear memory with a single base pointer and a stride vector. Element _i_ of the stride vector indicates the offset in linear memory between consecutive elements in dimension i. Consequently, the linear offset for an arbitrary element specified as an n-tuple may be computed as the dot product of the coordinate and the stride vector.
+
+CUTLASS provides abstractions for interacting with multidimension tensors in device memory.
+Consequently, we define a hierarchy of pointer-like types for referencing tensors.
+
+`T *` - raw pointer to elements of type T
+
+`cutlass::TensorRef<T, Rank>`  - reference to a tensor of elements of type T and given rank. Includes a mapping function and associated stride vector for accessing elements in linear memory.
+
+`cutlass::TensorView<T, Rank>` - extends `TensorRef<>` by adding bounds information. This is a complete mathematical object which may be used as the argument to CUTLASS functions.
+
+The above provide an identity maping of a logical index space to linear memory. An element
+at logical coordinate X has an offset computed as follows:
+```
+offset = dot(X, stride)
+```
+where `dot()` computes the inner product of X and a vector of "strides."
+
+CUTLASS 1.1 introduces a mapping function and an additional "storage rank" to offer a flexible way to
+map the logical index space of the tensor to memory. The mapping function maps a coordinate
+of rank _R_ to an index space of rank _S_. The linear offset is computed as:
+```
+offset = dot( MapFunc(X), stride )
+```
+where stride is a vector of rank _S_.
+
+CUTLASS kernels make extensive use of vectorization of memory accesses for efficiency and
+correctness. Consequently, we enforce a constraint on the strides used by mapping functions
+such that:
+
+1. The "fastest-changing" stride is always 1 thereby mandating that consecutive elements in
+   that rank are consecutive in linear memory.
+
+2. The fastest changing rank is always last in the stride vector and not explicitly stored.
+
+Thus, the stride vector used by mapping functions has length of one fewer than the rank of the
+storage tensor. These constraints are consistent with the BLAS interface of passing matrices as
+a tuple consisting of a pointer and a "leading dimension." In fact, these are rank=2 tensors
+whose fastest changing dimension is 1, and only the strided dimension is explicitly represented.
+
+A typical mapping function might simply map the rows and columns of a matrix, a rank=2 tensor,
+to linear memory such that (1.) elements in the same column are consecutive in memory
+(column-major), or (2.) elements in the same row are consecutive (row-major). These can be
+accomplished by two different mapping functions whose stride vector is length=2. The first
+element is the "leading dimension."
+
+The requirement that the fastest-changing stride always be of unit size need not be a limitation.
+To implement "sparse" computations or matrix operations in which matrix elements have arbitrary
+stride along both row and column, define a mapping function whose storage rank is 3. This permits
+two elements of the stride vector to have a non-unit value.
+
+`cutlass::TensorView<>` extends this concept by including a size vector to specify the bounds of
+the index space. The value of each coordinate in the size vector defines the half-open range of
+indices whose smallest value is zero.
+
+## <a name="S-core-shape"></a> Shape
+
+To avoid complicated template metaprogramming, CUTLASS targets fixed compile-time tile sizes specified
+by a four-dimensional template `cutlass::Shape<>`. This defines the following dimensions, mirroring
+the NHWC tensor format used for convolution in Deep Learning frameworks.
+
+- `D`: depth of tensor
+- `H`: first strided dimension
+- `W`: contiguous sequence of tensor elements
+- `C`: number of channels, usually used for vectorized access
+
+Template specializations of `Shape` appear as arguments to numerous dependent template classes which
+must specify compile-time constant tile sizes.
+
+## <a name="S-core-tile-structure"></a> Tile Structure
+
+Tiled structures express an arrangement of data in memory as well as a logical mapping of concurrent CUDA
+threads to the problem space. For example, the CUTLASS GEMM 
+
+Tiled structures can be defined using the `cutlass::TileTraits<>` concept which defines the following
+members. Collectively, these members offer a flexible way to define a 4-D subpartition of an integer
+lattice, partition its elements among a collection of threads, and map each unique thread ID to a unique
+offset.
+
+- _Tile_ (concept `Shape<>`) - describes the dimensions of the tile in terms of scalar elements
+- _Delta_ (concept `Shape<>`) - describes the distance along each logical dimension between items
+- _Iterations_ (concept `Shape<>`) - describes the number of items along each logical dimension
+- _ThreadOffset_ (concept _functor_) - implements `Coord<4> operator()() const` to determine a thread's
+  initial offset in the logical 4-D coordinate space
+
+The following figure illustrates the CUTLASS tile structure. The overall shape, 16-by-16, is partitioned into
+vectors of length two among 32 threads. The elements stored by thread 9 are highlighted.
+
+<img src="/media/images/cutlass-tile-structure.png" alt="CUTLASS tile structure" width="30%" />
+
+The `cutlass::TileTraits<>` definition that describes this arrangement may be defined as follows:
+
+```
+struct ExampleTileTraits {
+
+  /// Overall shape of tile
+  typedef Shape<1, 16, 16, 1> Tile;
+
+  /// Distance along each dimension of accesses
+  typedef Shape<1, 4, 1, 1> Delta;
+
+  /// Number of memory accesses performed by each thread
+  typedef Shape<1, 4, 1, 1> Iterations;
+
+  /// Offset function - maps each thread to a unique starting offset within the 4D tile
+  struct ThreadOffset {
+
+    CUTLASS_DEVICE Coord<4> operator()() const {
+
+      typdef Shape<1, 16, 8, 2> Vectorized;
+
+      return make_Coord(
+        0,                              // depth "D" dimension
+        threadIdx.x / Vectorized::kW,   // horisontal "H" dimension - first strided dimension
+        threadIdx.x % Vectorized::kW,   // vertical "W" dimension - contiguous dimension
+        0
+      );
+    }
+  };
+};
+```
+
+## <a name="S-core-tile-iterator"></a> Tile Iterator
+
+The iterator design pattern provides an abstraction for accessing the items in a collection in sequence. Basic
+operators defined by iterators consist of accessing an item - either a load or store - followed by traversal to
+the next item in sequence.
+
+<img src="/media/images/cutlass-tile-iteration.png" alt="CUTLASS tile access and traversal" width="50%" />
+
+To offer a generic solution that spans numerous data types and layouts, CUTLASS defines the _TileIterator_ concept.
+This concept provides access to a sequence of _tiles_ embedded in a tensor in addressable memory. 
+
+The canonical CUTLASS tile iterator template is defined in [cutlass/tile_iterator.h](cutlass/tile_iterator.h).
+
+## <a name="S-core-fragment"></a> Fragment
+
+A fragment is analogous to `std::array<>` in that it is a constant-sized array of elements. Typically backed by storage in the SM's register file, CUTLASS `Fragment<>` objects are used to store tiles. For threadblock- and warp-scope operations, the contents of these tiles are distributed across the partipcipating threads. In such cases, a thread's `Fragment<>` contains the part of the tile held by that thread.
+
+## <a name="S-core-predicate-vector"></a> Predicate Vector
+
+SIMT architectures utilize predicated execution in place of control flow when conditional code sequences are fairly short, on the order of a few machine instructions. While CUDA C++ does not include constructs at the language level for predication, PTX makes this explicit, and compilation to SASS is assumed to aggressively utilize predication. Typical applications are to initialize a sequence of bits used to mask memory operations and use these bits as predicates guarding memory load and store instructions. 
+
+CUTLASS provides `PredicateVector` defined in [cutlass/predicate_vector.h](cutlass/predicate_vector.h) to manage a statically-sized bit vector, store them into general purpose registers, and efficiently access them in sequence. By storing four predicates per byte in hardware registers, the CUDA compiler is able to issue specialized instructions to achieve very efficient unpacking. 
+
+
+# <a name="S-utilities"></a> 4. Utilities
+
+CUTLASS implements efficient matrix multiply computations on GPUs. It is accompanied by an extensive utility
+framework offering features such as:
+
+* [cutlass::half_t](tools/util/half.h) - a host-side half-precision type
+* Components for allocating and initializing [host-side and device-side tensors](tools/util/host_tensor.h) usable by CUTLASS
+* Reference implementations of [GEMM](tools/util/reference/host/gemm.h) and [element-wise operations](tools/util/reference/host/tensor_elementwise.h)
+
+# Copyright
+
+Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+
+```
+  Redistribution and use in source and binary forms, with or without modification, are permitted
+  provided that the following conditions are met:
+      * Redistributions of source code must retain the above copyright notice, this list of
+        conditions and the following disclaimer.
+      * Redistributions in binary form must reproduce the above copyright notice, this list of
+        conditions and the following disclaimer in the documentation and/or other materials
+        provided with the distribution.
+      * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+        to endorse or promote products derived from this software without specific prior written
+        permission.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+  IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+  FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+  FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+  BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+  OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+  STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+```
+

Doxyfile

@@ -218,7 +218,8 @@ TAB_SIZE               = 4
 # "Side Effects:". You can put \n's in the value part of an alias to insert
 # newlines.
 
-ALIASES                =
+#ALIASES += "concept{1}=@ingroup \1\n@par Implemented concepts:\n@ref \1"
+ALIASES += "concept{1}=@ingroup \1"
 
 # This tag can be used to specify a number of word-keyword mappings (TCL only).
 # A mapping has the form "name=value". For example adding "class=itcl::class"
@@ -396,7 +397,7 @@ LOOKUP_CACHE_SIZE      = 0
 # normally produced when WARNINGS is set to YES.
 # The default value is: NO.
 
-EXTRACT_ALL            = NO
+EXTRACT_ALL            = YES
 
 # If the EXTRACT_PRIVATE tag is set to YES all private members of a class will
 # be included in the documentation.
@@ -733,7 +734,7 @@ WARN_LOGFILE           =
 # spaces.
 # Note: If this tag is empty the current directory is searched.
 
-INPUT                  = cutlass cutlass/gemm cutlass/util
+INPUT                  = cutlass
 
 # This tag can be used to specify the character encoding of the source files
 # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
@@ -759,7 +760,7 @@ FILE_PATTERNS          =
 # be searched for input files as well.
 # The default value is: NO.
 
-RECURSIVE              = NO
+RECURSIVE              = YES
 
 # The EXCLUDE tag can be used to specify files and/or directories that should be
 # excluded from the INPUT source files. This way you can easily exclude a
@@ -2032,7 +2033,7 @@ HIDE_UNDOC_RELATIONS   = YES
 # set to NO
 # The default value is: NO.
 
-HAVE_DOT               = NO
+HAVE_DOT               = $(HAVE_DOT)
 
 # The DOT_NUM_THREADS specifies the number of dot invocations doxygen is allowed
 # to run in parallel. When set to 0 doxygen will base this on the number of
@@ -2204,7 +2205,7 @@ INTERACTIVE_SVG        = NO
 # found. If left blank, it is assumed the dot tool can be found in the path.
 # This tag requires that the tag HAVE_DOT is set to YES.
 
-DOT_PATH               =
+DOT_PATH               = $(DOT_PATH)
 
 # The DOTFILE_DIRS tag can be used to specify one or more directories that
 # contain dot files that are included in the documentation (see the \dotfile

README.md

@@ -1,106 +1,249 @@
-![ALT](/media/fig-09-complete-hierarchy.png "Complete CUDA GEMM decomposition")
+![ALT](/media/images/gemm-hierarchy-with-epilogue-no-labels.png "Complete CUDA GEMM decomposition")
 
-# Introduction
+# CUTLASS 1.1
 
-CUTLASS is a collection of CUDA C++ template abstractions for implementing 
-high-performance matrix-multiplication (GEMM) at all levels and scales within CUDA. 
-It incorporates strategies for hierarchical decomposition and data movement similar 
-to those used to implement cuBLAS.  CUTLASS decomposes these "moving parts" into 
+_CUTLASS 1.1.0 - September 2018_
+
+CUTLASS 1.1 is a collection of CUDA C++ template abstractions for implementing
+high-performance matrix-multiplication (GEMM) at all levels and scales within CUDA.
+It incorporates strategies for hierarchical decomposition and data movement similar
+to those used to implement cuBLAS.  CUTLASS decomposes these "moving parts" into
 reusable, modular software components abstracted by C++ template classes.  These
-thread-wide, warp-wide, block-wide, and device-wide primitives can be specialized 
-and tuned via custom tiling sizes, data types, and other algorithmic policy. The 
-resulting flexibility simplifies their use as building blocks within custom kernels 
+thread-wide, warp-wide, block-wide, and device-wide primitives can be specialized
+and tuned via custom tiling sizes, data types, and other algorithmic policy. The
+resulting flexibility simplifies their use as building blocks within custom kernels
 and applications.
 
 To support a wide variety of applications, CUTLASS provides extensive support for
-mixed-precision computations, providing specialized data-movement and 
-multiply-accumulate abstractions for 8-bit integer, half-precision floating 
-point (FP16), single-precision floating point (FP32), and double-precision floating 
+mixed-precision computations, providing specialized data-movement and
+multiply-accumulate abstractions for 8-bit integer, half-precision floating
+point (FP16), single-precision floating point (FP32), and double-precision floating
 point (FP64) types.  Furthermore, CUTLASS demonstrates CUDA's WMMA API for targeting
-the programmable, high-throughput _Tensor Cores_ provided by NVIDIA's Volta architecture 
+the programmable, high-throughput _Tensor Cores_ provided by NVIDIA's Volta architecture
 and beyond.
 
-For more exposition, see our Parallel Forall blog post [CUTLASS: Fast Linear Algebra 
-in CUDA C++](https://devblogs.nvidia.com/parallelforall/cutlass-linear-algebra-cuda). 
+CUTLASS 1.1 is described in the [CUTLASS Documentation](CUTLASS.md) and the accompanying
+[Doxygen documentation](https://nvidia.github.io/cutlass).
+We describe the structure of an efficient GEMM in our talk at the
+[GPU Technology Conference 2018](http://on-demand.gputechconf.com/gtc/2018/presentation/s8854-cutlass-software-primitives-for-dense-linear-algebra-at-all-levels-and-scales-within-cuda.pdf).
+
+# What's New in CUTLASS 1.1
+
+* [CUTLASS Documentation](CUTLASS.md)
+* [Examples](examples/)
+  * Basic GEMM, tensor views, CUTLASS utilities, batched GEMM, WMMA GEMM
+* Turing Features
+  * [WMMA GEMM targeting TensorCores](tools/test/unit/gemm/wmma_integer_gemm.cu) - INT8, INT4, 1-bit
+* [Batched Strided GEMM](tools/test/unit/gemm/batched_strided_sgemm_128x128x8.cu)
+* [Threadblock rasterization strategies](tools/test/unit/gemm/sgemm_threadblock_swizzle_nt.cu)
+  * Improved performance for adverse problem sizes and data layouts
+* Extended CUTLASS Core components
+  * Tensor views support arbitrary matrix and tensor layouts
+  * Zip iterators for structuring multiple data streams
+* Enhanced CUTLASS utilities
+  * [Reference implementations](tools/util/reference) for tensor operations in [host](tools/util/reference/host) and [device](tools/util/reference/device) code
+  * Added `HostMatrix<>` for simplified matrix creation
 
 # Performance
 
-<p align="center"><img src=/media/cutlass-performance-plot.png></p>
+<p align="center"><img src=/media/images/cutlass-performance-plot.png></p>
 
-CUTLASS primitives are very efficient.  When used to construct device-wide GEMM kernels, 
-they exhibit performance comparable to cuBLAS for scalar GEMM 
-computations. The above figure shows CUTLASS performance relative to cuBLAS 
-for large matrix dimensions (M=10240, N=K=4096) running on an NVIDIA Tesla V100 GPU 
-when compiled with CUDA 9.0. 
+CUTLASS primitives are very efficient.  When used to construct device-wide GEMM kernels,
+they exhibit performance comparable to cuBLAS for scalar GEMM
+computations. The above figure shows CUTLASS performance relative to cuBLAS
+for large matrix dimensions (M=10240, N=K=4096) running on an NVIDIA Titan V GPU
+when compiled with CUDA 10.0.
+
+# Compatibility
+
+CUTLASS performs best when compiled with the [CUDA 10.0 Toolkit](ttps://developer.nvidia.com/cuda-toolkit).
+It is compatible with CUDA 9.0, 9.1, and 9.2, but these versions of the CUDA Toolkit do not support new Turing WMMA features.
+
+We have tested the following environments.
+
+|**Operating System** | **Compiler** |
+|-----------------|----------|
+| Windows 10      | Microsoft Visual Studio 2015|
+|                 | Microsoft Visual Studio 2017|
+| Ubuntu 14.04 | GCC 4.8.2 |
+| Ubuntu 16.04 | GCC 5.4.0 |
+| Ubuntu 18.04 | GCC 7.3.0 |
+
+CUTLASS runs successfully on the following NVIDIA GPUs, and it is expected to be efficient on
+any Maxwell-, Pascal-, or Volta-architecture NVIDIA GPU.
+
+|**GPU**|
+|---|
+|NVIDIA GeForce 1080|
+|NVIDIA TitanXP|
+|NVIDIA Tesla P100|
+|NVIDIA Tesla V100|
+|NVIDIA TitanV|
+|NVIDIA GeForce RTX 2080 TI, 2080, 2070|
+
+# Building CUTLASS
+
+CUTLASS is a header-only template library and does not need to be built to be used by other
+projects. However, we distribute extensive unit tests and utility programs to demonstrate
+CUTLASS. These instructions are for building those test programs.
+
+CUTLASS's unit tests depend on Google Test which exists as a git submodule. You can fetch
+submodules as follows.
+
+```
+$ git submodule update --init --recursive
+```
+
+CUTLASS can be build with CMake starting version 3.10. By default CUTLASS will build kernels
+for CUDA architecture versions 5.0, 6.0, 6.1, 7.0 and 7.5. To reduce compile time you can specify
+the architectures to build CUTLASS for by changing the CMake configuration setting
+`CUTLASS_NVCC_ARCHS`.
+
+Create a build directory within the CUTLASS project, then run CMake once.
+
+```
+$ mkdir build && cd build
+$ cmake ..
+```
+
+Compile the CUTLASS project by running Make. Include the -j argument to compile sources in
+parallel and speed up the build process.
+
+```
+$ make -j12
+...
+$
+```
+
+Verify CUTLASS has been built correctly by running the unit tests from the build/ directory.
+
+```
+$ ./tools/test/unit/cutlass_unit_test
+...
+...
+...
+[----------] Global test environment tear-down
+[==========] 946 tests from 57 test cases ran. (10812 ms total)
+[  PASSED  ] 946 tests.
+```
+
+All tests should pass, though the exact number of tests may vary over time.
 
 # Project Structure
 
 CUTLASS is arranged as a header-only library with several example test programs
-that demonstrate instantiating a GEMM task within a CUDA kernel. Comments inline
-with the source explain the individual components.
+that demonstrate instantiating a GEMM task within a CUDA kernel. The Doxygen documentation
+provides a complete list of files, classes, and template concepts defined in the CUTLASS
+project. A brief summary is described below.
 
-The repository is organized in the following arrangement.
-
-    cutlass/                Root of header-only source library for matrix multiply
-      gemm/                 Implementation of GEMM __device__ code and supporting components
-      util/                 Utility components for CUDA device-side CUDA development
-
-A test program is provided to illustrate the use of CUTLASS. This is implemented
-in the following directory.
-
-    cutlass_test            Root of test programs depicting CUTLASS kernels
-      util/                 Utilities
-      gemm.cu               Simple example calling CUTLASS and CUBLAS GEMM kernels
-      Makefile              Build script for test programs
-
-
-# Makefile usage
-
-There are different sample targets for different GEMM data types and
-transposititions.  Be sure to specify your target architecture.
-
-    make <sgemm|dgemm|hgemm|igemm|wgemm> sm=<60|61|70> \
-      [transpose=<nn|nt|tn|tt>] [verbose=<0|1>] [keep=<0|1>]
-
-
-# Program usage
-
- Program usage:
-
-     <s|d|h|i|w>gemm_<nn|nt|tn|tt>
-           [--help]
-           [--schmoo=<#schmoo-samples> || --m=<height> --n=<width> --k=<depth>]
-           [--i=<timing iterations>]
-           [--device=<device-id>]
-           [--alpha=<alpha> --beta=<beta>]
-
-
-# Open Source License
-
-CUTLASS is released by NVIDIA Corporation under the "New BSD" open-source license:
+The CUTLASS library is defined in the cutlass/ directory and consists of CUDA C++ template
+classes and other definitions for implementing efficient GPU GEMM kernels. A set of core
+classes and templates define basic primitives that are then applied to compute GEMM via
+templates in the cutlass/gemm directory.
 
 ```
-Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
-
-Redistribution and use in source and binary forms, with or without
-modification, are permitted provided that the following conditions are met:
-   *  Redistributions of source code must retain the above copyright
-      notice, this list of conditions and the following disclaimer.
-   *  Redistributions in binary form must reproduce the above copyright
-      notice, this list of conditions and the following disclaimer in the
-      documentation and/or other materials provided with the distribution.
-   *  Neither the name of the NVIDIA CORPORATION nor the
-      names of its contributors may be used to endorse or promote products
-      derived from this software without specific prior written permission.
-
-THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
-ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
-WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
-DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
-DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
-(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
-LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
-ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+cutlass/
+  gemm/
+  util/
+  <core API components>
 ```
+
+Several tools and test programs are also distributed with the CUTLASS library. They are
+contained in the following directories.
+
+```
+examples/
+  00_basic_gemm/
+  01_tensor_view/
+  02_cutlass_utilities/
+  03_batched_gemm/
+  04_tile_iterator/
+  05_wmma_gemm/
+tools/
+  test/
+    unit/
+      core/
+      gemm/
+    perf/
+  util/
+    reference/
+      device/
+      host/
+    <utilities>
+```
+
+The `test/unit/` directory consist of unit tests implemented with Google Test that demonstrate
+basic usage of Core API components and complete tests of the CUTLASS GEMM computations.
+
+The `tools/util` directory contains CUTLASS utilities including reference implementations of GEMM and
+several element-wise tensor operations.
+
+# Performance Profiling
+
+The `test/perf/` directory contains a command-line utility for launching each of the GEMM kernels.
+Its usage is shown below.
+
+Program usage:
+
+```
+  cutlass_perf_test [options]
+
+    --help
+    --append=<true|false*>                            If true, appends output to existing CSV file. If false, overwrites.
+    --alpha=<alpha>                                   Value for alpha to be used in GEMM experiments
+    --beta=<beta>                                     Value for beta to be used in GEMM experiments
+    --dist=<distribution>                             Describes the random distribution of each of the input matrix operands.
+    --execution_mode=<mode>                           Specifies execution mode: profile, verify, single
+    --output=<filename.csv>                           Writes summary of profiling to specified .csv file
+    --iterations=<timing iterations>                  maximum number of iterations to execute when profiling
+    --m=<height>[:max height[:step]]                  Height of GEMM problem (number of rows of C). May specify a range with optional step size.
+    --n=<width>[:max width[:step]]                    Width of GEMM problem (number of columns of C). May specify a range with optional step size.
+    --k=<depth>[:max depth[:step]]                    Size of inner dimension of A and B. May specify a range with optional step size.
+    --kernels=<{s|d|h|i|wmma_}gemm_{nn,nt,tn,tt}>     Select GEMM datatype and layout to use for tests
+    --peak=<bool>                                     If true, only reports peak performance per kernel after profiling specified problem space.
+    --save_workspace={*never,incorrect,always}        Specifies when to save the GEMM inputs and results to the filesystem.
+    --seed=<seed>                                     Random seed used by the random number generator in initializing input matrices.
+    --tags=<column:tag,...>                           Inserts leading columns in output table and uniform values for each column.
+
+
+  Example usage:
+
+  # Runs one problem size for all kernels
+  $ ./tools/test/perf/cutlass_perf_test --m=10240 --n=1024 --k=1024
+
+  # Varies GEMM K dimension for SGEMM and IGEMM with column-major multiplicands
+  $ ./tools/test/perf/cutlass_perf_test --m=10240 --n=4096 --k=1024:8192:128 --kernels=sgemm_nn,igemm_nn
+```
+
+# About
+
+CUTLASS is released by NVIDIA Corporation as Open Source software under the
+3-clause "New" BSD license.
+
+
+# Copyright
+
+Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+
+```
+  Redistribution and use in source and binary forms, with or without modification, are permitted
+  provided that the following conditions are met:
+      * Redistributions of source code must retain the above copyright notice, this list of
+        conditions and the following disclaimer.
+      * Redistributions in binary form must reproduce the above copyright notice, this list of
+        conditions and the following disclaimer in the documentation and/or other materials
+        provided with the distribution.
+      * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+        to endorse or promote products derived from this software without specific prior written
+        permission.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+  IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+  FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+  FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+  BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+  OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+  STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+```
+

common.mk

@@ -1,181 +0,0 @@
-#/******************************************************************************
-# * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
-# *
-# * Redistribution and use in source and binary forms, with or without
-# * modification, are permitted provided that the following conditions are met:
-# *	 * Redistributions of source code must retain the above copyright
-# *	   notice, this list of conditions and the following disclaimer.
-# *	 * Redistributions in binary form must reproduce the above copyright
-# *	   notice, this list of conditions and the following disclaimer in the
-# *	   documentation and/or other materials provided with the distribution.
-# *	 * Neither the name of the NVIDIA CORPORATION nor the
-# *	   names of its contributors may be used to endorse or promote products
-# *	   derived from this software without specific prior written permission.
-# *
-# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
-# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
-# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
-# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
-# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
-# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
-# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
-# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-# *
-#******************************************************************************/
-
-
-#-------------------------------------------------------------------------------
-# Commandline Options
-#-------------------------------------------------------------------------------
-
-# sm=<XX,...> Compute-capability to compile for, e.g., "sm=200,300,350" (SM2.0 by default).
-
-COMMA := ,
-ifdef sm
-	SM_ARCH := $(subst $(COMMA),-,$(sm))
-else
-    $(error Please specify SM architecture makefile argument: "sm=XX")
-endif
-
-ifeq (70, $(findstring 70, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_70,code=\"sm_70,compute_70\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_70
-endif
-ifeq (62, $(findstring 62, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_62,code=\"sm_62,compute_62\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_62
-endif
-ifeq (61, $(findstring 61, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_61,code=\"sm_61,compute_61\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_61
-endif
-ifeq (60, $(findstring 60, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_60,code=\"sm_60,compute_60\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_60
-endif
-ifeq (52, $(findstring 52, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_52,code=\"sm_52,compute_52\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_52
-endif
-ifeq (37, $(findstring 37, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_37,code=\"sm_37,compute_37\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_37
-endif
-ifeq (35, $(findstring 35, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_35,code=\"sm_35,compute_35\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_35
-endif
-ifeq (30, $(findstring 30, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_30,code=\"sm_30,compute_30\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_30
-endif
-ifeq (21, $(findstring 21, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_20,code=\"sm_21,compute_20\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_21
-endif
-ifeq (20, $(findstring 20, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_20,code=\"sm_20,compute_20\"
-    CLANG_SM_TARGETS += --cuda-gpu-arch=sm_20
-endif
-
-
-# [verbose=<0|1>] Verbose toolchain output from nvcc option
-ifeq ($(verbose), 1)
-	NVCCFLAGS += -v
-	CLANG_CFLAGS += -v
-endif
-
-
-# [keep=<0|1>] Keep intermediate compilation artifacts option
-ifeq ($(keep), 1)
-	NVCCFLAGS += -keep
-	CLANG_CFLAGS += --save-temps
-endif
-
-
-# [debug=<0|1>] Generate debug mode code
-ifeq ($(debug), 1)
-	NVCCFLAGS += -G
-	CLANG_CFLAGS += --cuda-noopt-device-debug
-endif
-
-
-#-------------------------------------------------------------------------------
-# Compiler and compilation platform
-#-------------------------------------------------------------------------------
-
-BASE_DIR := $(dir $(lastword $(MAKEFILE_LIST)))
-
-NVCC := "$(shell which nvcc)"
-ifdef nvccver
-    NVCC_VERSION := $(nvccver)
-else
-    NVCC_VERSION := $(strip $(shell nvcc --version | grep release | sed 's/.*release //' |  sed 's/,.*//'))
-endif
-
-# Detect OS
-OSUPPER := $(shell uname -s 2>/dev/null | tr [:lower:] [:upper:])
-
-# Default flags: verbose kernel properties (regs, smem, cmem, etc.); runtimes for compilation phases
-NVCCFLAGS += -O3 -Xptxas -v
-CLANG_CFLAGS += -O3 -Xcuda-ptxas -v
-ifeq (WIN_NT, $(findstring WIN_NT, $(OSUPPER)))
-    # For MSVC
-
-    # Enable more warnings and treat as errors
-    NVCCFLAGS += -Xcompiler /W3 -Xcompiler /WX
-
-    # Disable excess x86 floating point precision that can lead to results being labeled incorrectly
-    NVCCFLAGS += -Xcompiler /fp:strict
-
-	# Compiler
-	CC := cl
-
-	# Multithreaded runtime
-	NVCCFLAGS += -Xcompiler /MT
-
-	CUDART_CYG := "$(shell dirname $(NVCC))/../lib/x64/cudart.lib"
-	CUDART := "$(shell cygpath -w $(CUDART_CYG))"
-
-else
-    # For g++
-
-    # Disable excess x86 floating point precision that can lead to results being labeled incorrectly
-    #NVCCFLAGS += -Xcompiler -ffloat-store
-
-	# Compiler
-    CC := g++
-
-    CUDART := "$(shell dirname $(NVCC))/../lib64/libcudart_static.a"
-
-endif
-
-# compiler=clang Enables compilation with clang. 
-
-ifeq ($(compiler), clang)
-	# NVCC_VERSION is used as the proxy for the CUDA version.
-	BIN_SUFFIX := sm$(SM_ARCH)_clang_cuda_$(NVCC_VERSION)
-	# Clangs needs few extra flags to point it to CUDA SDK
-	# and link the binaries with CUDA runtime.
-	CUDA_BASE=$(realpath $(join $(dir $(shell which nvcc)), ..))
-	CLANG_CFLAGS += --cuda-path=$(CUDA_BASE)
-	LIBINC += -L$(CUDA_BASE)/lib64 -Wl,-rpath=$(CUDA_BASE)/lib64
-	LIBS   += -lcudart
-
-	# Replace NVCC and its options with clang++.
-	NVCC = clang++
-	NVCCFLAGS = $(CLANG_CFLAGS)
-	SM_TARGETS = $(CLANG_SM_TARGETS)
-else
-	# Suffix to append to each binary
-	BIN_SUFFIX := sm$(SM_ARCH)_nvcc_$(NVCC_VERSION)
-endif
-
-
-#-------------------------------------------------------------------------------
-# Function for computing dependency Lists
-#-------------------------------------------------------------------------------
-
-rwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))

cutlass/convert.h

@@ -0,0 +1,102 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*!
+  \file
+  \brief Defines conversion operations among Fragments of different base type.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename InputFragment_, typename OutputFragment_>
+struct Convert {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename InputScalar_, typename OutputScalar_, int kScalars_>
+struct Convert<Fragment<InputScalar_, kScalars_>, Fragment<OutputScalar_, kScalars_> > {
+  /// The input fragment.
+  typedef Fragment<InputScalar_, kScalars_> InputFragment;
+  /// The output fragment.
+  typedef Fragment<OutputScalar_, kScalars_> OutputFragment;
+
+  /// Ctor.
+  CUTLASS_DEVICE Convert() {}
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(InputFragment const& src, OutputFragment& dst) {
+    transform(src, 0, dst);
+  }
+
+  /// Transform a fragment.
+  template <typename Fragment_>
+  CUTLASS_DEVICE void transform(Fragment_ const& src, int offset, OutputFragment& dst) {
+    for (int i = 0; i < kScalars_; ++i) {
+      dst[i] = static_cast<OutputScalar_>(src[i + offset]);
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Fragment_>
+struct Copy {
+  /// The input fragment.
+  typedef Fragment_ InputFragment;
+  /// The output fragment.
+  typedef Fragment_ OutputFragment;
+
+  /// Ctor.
+  CUTLASS_DEVICE Copy() {}
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(Fragment_ const& src, Fragment_& dst) { transform(src, 0, dst); }
+
+  /// Transform a fragment.
+  template <typename InputFragment_>
+  CUTLASS_DEVICE void transform(InputFragment_ const& src, int offset, Fragment_& dst) {
+    if (sizeof(typename Fragment_::Element) == 8) {
+      uint64_t const* src_ptr = reinterpret_cast<uint64_t const*>(&src[offset]);
+      uint64_t* dst_ptr = reinterpret_cast<uint64_t*>(&dst[0]);
+      for (int i = 0; i < sizeof(Fragment_) / 8; ++i) {
+        dst_ptr[i] = src_ptr[i];
+      }
+    } else {
+      uint32_t const* src_ptr = reinterpret_cast<uint32_t const*>(&src[offset]);
+      uint32_t* dst_ptr = reinterpret_cast<uint32_t*>(&dst[0]);
+      for (int i = 0; i < sizeof(Fragment_) / 4; ++i) {
+        dst_ptr[i] = src_ptr[i];
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/coord.h

@@ -0,0 +1,353 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief A Coord is a coordinate of arbitrary rank into a tensor or matrix
+*/
+
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cutlass/util/platform.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Describes identity elements
+struct Identity {
+  /// Enumeration describing identity elements. Value assignments are significant.
+  /// Feel free to add or multiply by these, respectively.
+  enum Kind { Additive = 0, Multiplicative = 1 };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Statically-sized array specifying Coords within a tensor
+template <int Rank_, typename Index_ = int>
+struct Coord {
+  //
+  // Type and constant definitions
+  //
+
+  /// Number of elements in Coord
+  static int const kRank = Rank_;
+
+  /// Number of elements in Coord, aliased for compatibility
+  static int const N = Rank_;
+
+  /// Index type used to store elements
+  typedef Index_ Index;
+
+  //
+  // Data members
+  //
+
+  /// Indices
+  Index idx[kRank];
+
+  //
+  // Methods
+  //
+
+  /// Default ctor initializes uniformly
+  CUTLASS_HOST_DEVICE
+  Coord(Index value = 0) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] = value;
+    }
+  }
+
+  /// Constructs from an array of integers
+  CUTLASS_HOST_DEVICE
+  Coord(Index _idx[]) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] = _idx[i];
+    }
+  }
+
+  /// Constructs from an array of integers
+  CUTLASS_HOST_DEVICE
+  Coord(Coord<kRank> const &coord) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] = coord[i];
+    }
+  }
+
+  /// Returns a slice of the Coord which may be larger or smaller in rank
+  /// than this.
+  template <int Slice>
+  CUTLASS_HOST_DEVICE
+  Coord<Slice> slice(int start = 0, Index identity = 0) const {
+    Coord<Slice> result;
+    for (int i = 0; i < Slice; ++i) {
+      if (i + start < kRank) {
+        slice[i] = idx[i + start];
+      }
+      else {
+        slice[i] = identity;
+      }
+    }
+    return result;
+  }
+
+  /// Returns true if Coord is non-zero.
+  CUTLASS_HOST_DEVICE
+  operator bool() const {
+    for (int i = 0; i < kRank; ++i) {
+      if (idx[i]) {
+        return true;
+      }
+    }
+    return false;
+  }
+
+  /// Returns true if Coord is uniformly zero.
+  CUTLASS_HOST_DEVICE
+  bool operator!() const {
+    for (int i = 0; i < kRank; ++i) {
+      if (idx[i]) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  /// Element-wise addition
+  CUTLASS_HOST_DEVICE
+  Coord operator+(Coord const& b) const {
+    Coord c;
+    for (int i = 0; i < kRank; ++i) {
+      c.idx[i] = idx[i] + b.idx[i];
+    }
+    return c;
+  }
+
+  /// Element-wise subtraction
+  CUTLASS_HOST_DEVICE
+  Coord operator-(Coord const& b) const {
+    Coord c;
+    for (int i = 0; i < kRank; ++i) {
+      c.idx[i] = idx[i] - b.idx[i];
+    }
+    return c;
+  }
+
+  /// Element-wise multiplication
+  CUTLASS_HOST_DEVICE
+  Coord operator*(Coord const& b) const {
+    Coord c;
+    for (int i = 0; i < kRank; ++i) {
+      c.idx[i] = idx[i] * b.idx[i];
+    }
+    return c;
+  }
+
+  /// Element-wise division
+  CUTLASS_HOST_DEVICE
+  Coord operator/(Coord const& b) const {
+    Coord c;
+    for (int i = 0; i < kRank; ++i) {
+      c.idx[i] = idx[i] / b.idx[i];
+    }
+    return c;
+  }
+
+  /// In-place addition
+  CUTLASS_HOST_DEVICE
+  Coord& operator+=(Coord const& b) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] += b.idx[i];
+    }
+    return *this;
+  }
+
+  /// In-place subtraction
+  CUTLASS_HOST_DEVICE
+  Coord& operator-=(Coord const& b) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] -= b.idx[i];
+    }
+    return *this;
+  }
+
+  /// In-place multiplication
+  CUTLASS_HOST_DEVICE
+  Coord& operator*=(Coord const& b) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] *= b.idx[i];
+    }
+    return *this;
+  }
+
+  /// In-place division
+  CUTLASS_HOST_DEVICE
+  Coord& operator/=(Coord const& b) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] /= b.idx[i];
+    }
+    return *this;
+  }
+
+  /// Member access operator
+  CUTLASS_HOST_DEVICE Index& operator[](int dim) { return idx[dim]; }
+
+  /// Member access operator
+  CUTLASS_HOST_DEVICE Index const& operator[](int dim) const { return idx[dim]; }
+
+  /// Computes the dot product of two Coord instances
+  template <typename T>
+  CUTLASS_HOST_DEVICE T dot(Coord const& b, T sum) const {
+    for (int i = 0; i < kRank; ++i) {
+      sum += idx[i] * b.idx[i];
+    }
+    return sum;
+  }
+
+  /// Computes the dot product of two Coord instances
+  template <typename T>
+  CUTLASS_HOST_DEVICE T dot(Coord const& b) const {
+    T sum = T(0);
+    for (int i = 0; i < kRank; ++i) {
+      sum += idx[i] * b.idx[i];
+    }
+    return sum;
+  }
+
+  /// Gets the index of a given Coord element
+  template <int Dim>
+  CUTLASS_HOST_DEVICE Index& at() {
+    return idx[Dim];
+  }
+
+  /// Access via index; may limit unrolling potential
+  CUTLASS_HOST_DEVICE
+  Index& at(int dim) { return idx[dim]; }
+
+  /// Gets the index of a given Coord element
+  template <int Dim>
+  CUTLASS_HOST_DEVICE Index const& at() const {
+    return idx[Dim];
+  }
+
+  /// Access via index; may limit unrolling potential
+  CUTLASS_HOST_DEVICE
+  Index const& at(int dim) const { return idx[dim]; }
+
+  /// Determines if two Coord<> objects are equal
+  CUTLASS_HOST_DEVICE
+  bool operator==(Coord<kRank> const& b) const {
+    bool equal = true;
+    for (int i = 0; equal && i < kRank; ++i) {
+      equal = (idx[i] == b.idx[i]);
+    }
+    return equal;
+  }
+
+  /// Not equal
+  CUTLASS_HOST_DEVICE
+  bool operator!=(Coord<kRank> const& b) const { return !(*this == b); }
+
+  /// Clamps a coordinate to a range specified by maximum and minimum values
+  CUTLASS_HOST_DEVICE
+  Coord& clamp(Coord<kRank> const& max, Coord<kRank> const& min = Coord<kRank>()) {
+    for (int i = 0; i < kRank; ++i) {
+      idx[i] = __NV_STD_MAX(__NV_STD_MIN(idx[i], max.idx[i]), min.idx[i]);
+    }
+    return *this;
+  }
+
+  /// Returns the product of all elements
+  CUTLASS_HOST_DEVICE
+  Index count() const {
+    Index product = idx[0];
+    for (int i = 1; i < kRank; ++i) {
+      product *= idx[i];
+    }
+    return product;
+  }
+
+  /// Less than operator
+  CUTLASS_HOST_DEVICE
+  bool operator<(Coord<kRank> const &b) const {
+    for (int i = 0; i < kRank; ++i) {
+      if (!(idx[i] < b[i])) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  /// Less than or equals operator
+  CUTLASS_HOST_DEVICE
+  bool operator<=(Coord<kRank> const &b) const {
+    for (int i = 0; i < kRank; ++i) {
+      if (!(idx[i] <= b[i])) {
+        return false;
+      }
+    }
+    return true;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Helper to make a 2-element coordinate
+CUTLASS_HOST_DEVICE
+Coord<1> make_Coord(int _0) {
+  int values[1] = {_0};
+  return Coord<1>(values);
+}
+
+/// Helper to make a 2-element coordinate
+CUTLASS_HOST_DEVICE
+Coord<2> make_Coord(int _0, int _1) {
+  int values[2] = {_0, _1};
+  return Coord<2>(values);
+}
+
+/// Helper to make a 3-element coordinate
+CUTLASS_HOST_DEVICE
+Coord<3> make_Coord(int _0, int _1, int _2) {
+  int values[3] = {_0, _1, _2};
+  return Coord<3>(values);
+}
+
+/// Helper to make a 4-element coordinate
+CUTLASS_HOST_DEVICE
+Coord<4> make_Coord(int _0, int _1, int _2, int _3) {
+  int values[4] = {_0, _1, _2, _3};
+  return Coord<4>(values);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Shape_>
+CUTLASS_HOST_DEVICE Coord<3> make_Coord_from_shape() {
+  return make_Coord(Shape_::kD, Shape_::kH, Shape_::kW);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/core_io.h

@@ -0,0 +1,126 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Helpers for printing cutlass/core objects
+*/
+
+#pragma once
+
+#include <iosfwd>
+#include <typeinfo>
+
+#include "cutlass/coord.h"
+#include "cutlass/vector.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int Rank>
+std::ostream& operator<<(std::ostream& out, Coord<Rank> const& coord) {
+  for (int i = 0; i < Rank; ++i) {
+    out << (i ? ", " : "") << coord.idx[i];
+  }
+  return out;
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Helper to enable formatted printing of CUTLASS scalar types to an ostream
+template <typename T>
+struct ScalarIO {
+
+  /// Value to print
+  T value;
+
+  /// Default ctor
+  ScalarIO() { }
+
+  /// Constructs from a value
+  ScalarIO(T value): value(value) {}
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Default printing to ostream
+template <typename T>
+inline std::ostream &operator<<(std::ostream &out, ScalarIO<T> const &scalar) {
+  return out << scalar.value;
+}
+
+/// Printing to ostream of int8_t as integer rather than character
+template <>
+inline std::ostream &operator<<(std::ostream &out, ScalarIO<int8_t> const &scalar) {
+  return out << int(scalar.value);
+}
+
+/// Printing to ostream of uint8_t as integer rather than character
+template <>
+inline std::ostream &operator<<(std::ostream &out, ScalarIO<uint8_t> const &scalar) {
+  return out << unsigned(scalar.value);
+}
+
+/// Printing to ostream of vector of 1b elements
+template <>
+inline std::ostream &operator<<(
+  std::ostream &out, 
+  ScalarIO<cutlass::Vector<cutlass::bin1_t, 32> > const &scalar) {
+
+  for (int i = 0; i < 32; i++) {
+    out << int(scalar.value[i]);
+    out << ((i != 31) ? ", " : "");
+  }
+  return out;
+}
+
+/// Printing to ostream of vector of 4b signed integer elements
+template <>
+inline std::ostream &operator<<(
+  std::ostream &out, 
+  ScalarIO<cutlass::Vector<cutlass::int4_t, 8> > const &scalar) {
+
+  for (int i = 0; i < 8; i++) {
+    out << int(scalar.value[i]);
+    out << ((i != 7) ? ", " : "");
+  }
+  return out;
+}
+
+/// Printing to ostream of vector of 4b unsigned integer elements
+template <>
+inline std::ostream &operator<<(
+  std::ostream &out, 
+  ScalarIO<cutlass::Vector<cutlass::uint4_t, 8> > const &scalar) {
+
+  for (int i = 0; i < 8; i++) {
+    out << unsigned(scalar.value[i]);
+    out << ((i != 7) ? ", " : "");
+  }
+  return out;
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/cutlass.h

@@ -0,0 +1,90 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+/*! \file
+    \brief Basic include for CUTLASS macros
+*/
+
+#pragma once
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#define CUTLASS_MAJOR 1
+#define CUTLASS_MINOR 1
+#define CUTLASS_PATCH 0
+#define CUTLASS_VERSION ((CUTLASS_MAJOR)*100 + (CUTLASS_MINOR)*10 + CUTLASS_PATCH)
+
+#ifdef __NVCC__
+#define CUTLASS_HOST_DEVICE __forceinline__ __device__ __host__
+#define CUTLASS_DEVICE __forceinline__ __device__
+#elif defined(__CUDACC_RTC__)
+#define CUTLASS_HOST_DEVICE __forceinline__ __device__
+#define CUTLASS_DEVICE __forceinline__ __device__
+#else
+#define CUTLASS_HOST_DEVICE
+// CUTLASS_DEVICE is an error if not compiling device code
+#endif
+
+#define CUTLASS_ASSERT(x) assert(x)
+
+// CUTLASS_PRAGMA_(UNROLL|NO_UNROLL) optimization directives for the CUDA compiler.
+#if defined(__CUDA_ARCH__)
+#if defined(_MSC_VER)
+#define CUTLASS_PRAGMA_UNROLL __pragma("unroll")
+#define CUTLASS_PRAGMA_NO_UNROLL __pragma("unroll 1")
+#else
+#define CUTLASS_PRAGMA_UNROLL _Pragma("unroll")
+#define CUTLASS_PRAGMA_NO_UNROLL _Pragma("unroll 1")
+#endif
+#else
+#define CUTLASS_PRAGMA_UNROLL
+#define CUTLASS_PRAGMA_NO_UNROLL
+#endif
+
+#define CUTLASS_GEMM_LOOP CUTLASS_PRAGMA_NO_UNROLL
+
+// A small helper class to dump a type at compile time
+// Usage:: DumpType<Class>::Class
+template <typename T>
+struct DebugType {};
+
+template <typename T>
+void DebugTypeFunc(T const& t) {
+  T::t;
+}
+
+// A small helper class to dump a compile time constant at compile time
+// Usage: DumpValue<Class::kConstant>::kConstant
+template <int Value>
+struct DebugValue {};
+
+namespace cutlass {
+
+/// NVIDIA GPU Warp size
+static const int kWarpSize = 32;
+
+}  // namespace cutlass
+
+////////////////////////////////////////////////////////////////////////////////////////////////////

cutlass/fragment.h

@@ -0,0 +1,276 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines Fragment, a statically-sized array for storing parts of matrices within a
+   thread's registers.
+*/
+#pragma once
+
+#include <assert.h>
+#include "cutlass/shape.h"
+#include "cutlass/util/cutlass_math.h"
+#include "cutlass/vector.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup fragment_concept Fragment Concept
+@{
+
+\ref fragment_concept is a statically sized array for storing parts of tiles held by individual CUDA
+threads.
+
+@par \ref fragment_concept
+   Types satisfying \ref fragment_concept define the following members
+  - <b>Element</b> - type of each access held within the fragment
+  - <b>kElements</b> - number of elements stored by the fragment
+  - <b>clear()</b> - overwrites the fragment storage with zeros
+  - <b>Element & operator[](int i)</b> - by-reference access of the ith element
+  - <b>Element const & operator[](int i) const</b> - const by-reference access of the ith element
+@}
+*/
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup fragment_iterator_concept Fragment Iterator Concept
+@{
+
+\ref fragment_iterator_concept provides structured access to the elements within a fragment with an
+optional bitcast to the desired access type
+
+@par \ref fragment_iterator_concept
+   Types satisfying \ref fragment_iterator_concept define the following members
+  - <b>AccessType& operator[](int i)</b> - provides access to the ith element of the fragment
+  - <b>AccessType& at(int d, int h, int w, int c)</b> - applies \ref layout_concept to fragment and
+provides access to element at (d, h, w, c)
+
+@}
+*/
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int alignment>
+struct StorageType {
+  typedef uint64_t Type;
+};
+template <>
+struct StorageType<4> {
+  typedef uint32_t Type;
+};
+template <>
+struct StorageType<2> {
+  typedef uint16_t Type;
+};
+template <>
+struct StorageType<1> {
+  typedef uint8_t Type;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief A template defining \ref fragment_concept
+* @concept{fragment_concept}
+*/
+template <typename Element_, int kElements_, size_t kAlignment_ = 16>
+struct Fragment : public AlignedStruct<kAlignment_> {
+  /// Make sure the alignment makes sense wrt the size of elements.
+  static_assert(kAlignment_ == 16 || kAlignment_ >= sizeof(Element_), "Alignment is too small");
+  /// Alignment must be a power of two
+  static_assert(is_pow2<kAlignment_>::value, "Alignment must be a power of two");
+
+  /// This class.
+  typedef Fragment<Element_, kElements_> This_;
+  /// The element.
+  typedef Element_ Element;
+  /// The number of elements.
+  static int const kElements = kElements_;
+  /// Alignment
+  static int const kAlignment = kAlignment_;
+
+  /// Clear a fragment.
+  CUTLASS_HOST_DEVICE void clear() {
+    // Avoid element-wise access for sub 32b element type
+    if (kAlignment_ >= 8 && (kElements * sizeof(Element)) % 8 == 0) {
+      uint64_t* ptr = reinterpret_cast<uint64_t*>(storage);
+      for (int i = 0; i < (kElements * sizeof(Element)) / 8; ++i) {
+        ptr[i] = uint64_t(0);
+      }
+    } else if (kAlignment_ >= 4 && (kElements * sizeof(Element)) % 4 == 0) {
+      uint32_t* ptr = reinterpret_cast<uint32_t*>(storage);
+      for (int i = 0; i < (kElements * sizeof(Element)) / 4; ++i) {
+        ptr[i] = uint32_t(0);
+      }
+    } else if (kAlignment_ >= 2 && (kElements * sizeof(Element)) % 2 == 0) {
+      uint16_t* ptr = reinterpret_cast<uint16_t*>(storage);
+      for (int i = 0; i < (kElements * sizeof(Element)) / 2; ++i) {
+        ptr[i] = uint16_t(0);
+      }
+    } else {
+      for (int i = 0; i < kElements; ++i) {
+        storage[i] = 0;
+      }
+    }
+  }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE Element& operator[](int i) { return reinterpret_cast<Element*>(storage)[i]; }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE Element const& operator[](int i) const {
+    return reinterpret_cast<Element const*>(storage)[i];
+  }
+
+ private:
+  /// Storage type to use for Elements
+  typedef typename StorageType<kAlignment_>::Type StorageType;
+
+  /// Number of elements in the storage
+  static int const kStorageCount =
+      (sizeof(Element_) * kElements_ + sizeof(StorageType) - 1) / sizeof(StorageType);
+  /// The storage.
+  StorageType storage[kStorageCount];
+
+  /// Ensure that there's enough storage for all elements
+  static_assert(sizeof(StorageType) <= kAlignment_, "StorageType is too big for given alignment");
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief A template defining \ref fragment_iterator_concept
+* @concept{fragment_iterator_concept}
+*/
+template <typename Fragment_, typename Iterations_, typename AccessType_>
+struct FragmentIterator {
+  /// This class.
+  typedef FragmentIterator<Fragment_, Iterations_, AccessType_> This_;
+  /// The fragment.
+  typedef Fragment_ Fragment;
+  /// The number of iterations.
+  typedef Iterations_ Iterations;
+  /// The access type.
+  typedef AccessType_ AccessType;
+
+  /// The element.
+  typedef typename Fragment::Element Element;
+  /// The number of elements per access.
+  static int const kElementsPerAccess = (int)(sizeof(AccessType) / sizeof(Element));
+  /// The shape of the the fragment.
+  typedef typename ShapeMul<Iterations, Shape<1, 1, 1, kElementsPerAccess> >::Shape FragmentShape;
+  /// The linear strides for iterations.
+  typedef typename ShapeStrides<FragmentShape, kElementsPerAccess>::Shape Strides;
+
+  /// Ctor.
+  template <typename OtherFragment_>
+  CUTLASS_HOST_DEVICE FragmentIterator(OtherFragment_& fragment, int offset = 0)
+      : pointer(reinterpret_cast<Element*>(&fragment[offset])) {
+    static_assert(OtherFragment_::kElements >= Fragment::kElements, "");
+  }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE AccessType const& at(int d, int h, int w, int c = 0) const {
+    int const imm = ComputeOffsetFromStrides<Strides>::get(d, h, w, c);
+    return reinterpret_cast<AccessType const&>(pointer[imm]);
+  }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE AccessType& at(int d, int h, int w, int c = 0) {
+    int const imm = ComputeOffsetFromStrides<Strides>::get(d, h, w, c);
+    return reinterpret_cast<AccessType&>(pointer[imm]);
+  }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE AccessType const& operator[](int i) const {
+    return reinterpret_cast<AccessType const&>(pointer[i * kElementsPerAccess]);
+  }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE AccessType& operator[](int i) {
+    return reinterpret_cast<AccessType&>(pointer[i * kElementsPerAccess]);
+  }
+
+  /// Is the iterator valid?
+  CUTLASS_HOST_DEVICE bool valid(int d, int h, int w, int c) const { return true; }
+
+  /// The pointer.
+  Element* pointer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Fragment_, typename Iterations_, typename AccessType_>
+struct FragmentConstIterator {
+  /// This class.
+  typedef FragmentIterator<Fragment_, Iterations_, AccessType_> This_;
+  /// The fragment.
+  typedef Fragment_ Fragment;
+  /// The number of iterations.
+  typedef Iterations_ Iterations;
+  /// The access type.
+  typedef AccessType_ AccessType;
+
+  /// The element.
+  typedef typename Fragment::Element Element;
+  /// The number of elements per access.
+  static int const kElementsPerAccess = (int)(sizeof(AccessType) / sizeof(Element));
+  /// The shape of the the fragment.
+  typedef typename ShapeMul<Iterations, Shape<1, 1, 1, kElementsPerAccess> >::Shape FragmentShape;
+  /// The linear strides for iterations.
+  typedef typename ShapeStrides<FragmentShape, kElementsPerAccess>::Shape IterationsStrides;
+
+  /// Ctor.
+  template <typename OtherFragment_>
+  CUTLASS_HOST_DEVICE FragmentConstIterator(OtherFragment_& fragment, int offset = 0)
+      : pointer(reinterpret_cast<Element const*>(&fragment[offset])) {
+    static_assert(OtherFragment_::kElements >= Fragment::kElements, "");
+  }
+  /// Create from non-constant FragmentIterator
+  CUTLASS_HOST_DEVICE FragmentConstIterator(
+      FragmentIterator<Fragment_, Iterations_, AccessType_> const& rhs_)
+      : pointer(reinterpret_cast<Element const*>(rhs_.offset)) {}
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE AccessType const& at(int d, int h, int w, int c = 0) const {
+    int const imm = ComputeOffsetFromStrides<IterationsStrides>::get(d, h, w, c);
+    return reinterpret_cast<AccessType const&>(pointer[imm]);
+  }
+
+  /// The accessor.
+  CUTLASS_HOST_DEVICE AccessType const& operator[](int i) const {
+    return reinterpret_cast<AccessType const&>(pointer[i * kElementsPerAccess]);
+  }
+
+  /// Is the iterator valid?
+  CUTLASS_HOST_DEVICE bool valid(int d, int h, int w, int c) const { return true; }
+
+  /// The pointer.
+  Element const* pointer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/fragment_multiply_add.h

@@ -0,0 +1,159 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines multiply-add operations on fragments within a thread.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template < typename ScalarAlphaBeta_, 
+  typename ScalarAccum_, 
+  bool fragMul2 = true /*number of element per fragment is multiple of 2*/
+>
+struct FragmentMultiplyAdd {
+  /// The shape of the instruction.
+  typedef Shape<1, 1, 1, 1> InstructionShape;
+  /// The type for alpha and beta
+  typedef ScalarAlphaBeta_ ScalarAlphaBeta;
+  /// The type for accumlator
+  typedef ScalarAccum_ ScalarAccum;
+
+  /// Ctor.
+  CUTLASS_DEVICE FragmentMultiplyAdd() {}
+
+  /// Multiply : d = a*b.
+  template <typename FragmentB_, typename FragmentCd_>
+  CUTLASS_DEVICE void multiply(ScalarAlphaBeta a, FragmentB_ const& b, FragmentCd_& d) {
+#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
+    int const kReduction = FragmentB_::kElements / FragmentCd_::kElements;
+    for (int j = 0; j < FragmentCd_::kElements; ++j) {
+      d[j] = b[j * kReduction + 0];
+      for (int k = 1; k < kReduction; ++k) {
+        d[j] += b[j * kReduction + k];
+      }
+      d[j] = a * ScalarAlphaBeta(d[j]);
+    }
+#endif
+  }
+
+  /// Multiply : d = a*b + c.
+  template <typename FragmentB_, typename FragmentCd_>
+  CUTLASS_DEVICE void multiply_add(ScalarAlphaBeta a,
+                                   FragmentB_ const& b,
+                                   FragmentCd_ const& c,
+                                   FragmentCd_& d) {
+#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
+    int const kReduction = FragmentB_::kElements / FragmentCd_::kElements;
+    for (int j = 0; j < FragmentCd_::kElements; ++j) {
+      d[j] = b[j * kReduction + 0];
+      for (int k = 1; k < kReduction; ++k) {
+        d[j] += b[j * kReduction + k];
+      }
+      d[j] = a * ScalarAlphaBeta(d[j]) + ScalarAlphaBeta(c[j]);
+    }
+#endif
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
+template <>
+struct FragmentMultiplyAdd<half, half, true> {
+  /// The shape of the instruction.
+  typedef Shape<1, 1, 1, 1> InstructionShape;
+  /// The type for alpha and beta
+  typedef half ScalarAlphaBeta;
+  /// The type for accumlator
+  typedef half ScalarAccum;
+
+  /// Ctor.
+  CUTLASS_DEVICE FragmentMultiplyAdd() {}
+
+  /// Multiply : d = a*b.
+  template <typename FragmentB_, typename FragmentCd_>
+  CUTLASS_DEVICE void multiply(half a, FragmentB_ const& b, FragmentCd_& d) {
+#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
+    // The input.
+    __half2 const* b_half2 = reinterpret_cast<__half2 const*>(&b[0]);
+    // The output.
+    __half2* d_half2 = reinterpret_cast<__half2*>(&d[0]);
+
+    // Assemble a half2 from a.
+    __half2 const a_half2 = __half2half2(a);
+
+    int const kReduction = (FragmentB_::kElements / FragmentCd_::kElements);
+
+    for (int j = 0; j < FragmentCd_::kElements / 2; ++j) {
+      d_half2[j] = __hmul2(a_half2, b_half2[j * kReduction + 0]);
+
+      for (int k = 1; k < kReduction; ++k) {
+        d_half2[j] = __hfma2(a_half2, b_half2[j * kReduction + k], d_half2[j]);
+      }
+    }
+#endif
+  }
+
+
+  /// Multiply : d = a*b + c.
+  template <typename FragmentB_, typename FragmentCd_>
+  CUTLASS_DEVICE void multiply_add(half a,
+                                   FragmentB_ const& b,
+                                   FragmentCd_ const& c,
+                                   FragmentCd_& d) {
+#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
+    // The inputs.
+    __half2 const* b_half2 = reinterpret_cast<__half2 const*>(&b[0]);
+    __half2 const* c_half2 = reinterpret_cast<__half2 const*>(&c[0]);
+    // The output.
+    __half2* d_half2 = reinterpret_cast<__half2*>(&d[0]);
+
+    // Assemble a half2 from a.
+    __half2 const a_half2 = __half2half2(a);
+
+    int const kReduction = (FragmentB_::kElements / FragmentCd_::kElements);
+    for (int j = 0; j < FragmentCd_::kElements / 2; ++j) {
+      d_half2[j] = __hfma2(a_half2, b_half2[j * kReduction + 0], c_half2[j]);
+
+      for (int k = 1; k < kReduction; ++k) {
+        d_half2[j] = __hfma2(a_half2, b_half2[j * kReduction + k], d_half2[j]);
+      }
+    }
+#endif
+  }
+};
+
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/block_loader.h

@@ -1,162 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * block-wide tile-loading abstractions
- */
-
-#include "../util/util.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * load_algorithm
- ******************************************************************************/
-
-/**
- * \brief Enumeration of matrix loading algorithms
- */
-struct load_algorithm
-{
-    /// \brief Enumerants.  See corresponding tag types.
-    enum kind_t
-    {
-        CongruousCopy  = 0,
-        CrosswiseCopy  = 1,
-    };
-
-    /**
-     * \brief Generic tag
-     */
-    template <kind_t Kind>
-    struct any_tag : nv_std::integral_constant<kind_t, Kind> {};
-
-    /**
-     * \brief Copy from a global matrix that is row-major in relation
-     * to the local row-major tile
-     */
-    typedef any_tag<CongruousCopy> contiguous_tag_t;
-
-    /**
-     * \brief Copy from a global matrix that is column-major in relation
-     * to the local row-major tile
-     */
-    typedef any_tag<CrosswiseCopy> crosswise_tag_t;
-
-};
-
-
-/******************************************************************************
- * block_loader
- ******************************************************************************/
-
-/**
- * \brief A three-phase data loading abstraction (prefetch, commit, and
- * advance) for iterating over ranges of block-wide matrix tiles.
- *
- * Each iteration sequence produces a KxL (height-by-width) block-wide tile of
- * value_t in shared memory. The layout of the shared
- * block-wide tile is a row-major (L-major) tiling of dp_vector_t items, which are
- * themselves column-major (K-major) vectors of value_t.  Its dimensions are:
- *    K = BlockDpVectorsK * (sizeof(dp_vector_t) / sizeof(value_t)
- *    L = BlockDpVectorsL
- *
- * NB: This generic class is not directly constructible.  Architecture- and
- * algorithm-specific template specializations will provide the API
- * functionality prescribed here.
- *
- */
-template <
-    int BlockThreads,                       ///< Number of threads in each thread block (blockDim.x)
-    int BlockDpVectorsK,                    ///< Extent of block-wide tile in dp_vector_t along the K-axis (height)
-    int BlockDpVectorsL,                    ///< Extent of block-wide tile in dp_vector_t along the L-axis (width)
-    typename value_t,                       ///< Input matrix value type
-    int LeadingDimAlignBytes,               ///< Byte alignment of input matrix leading dimension
-    bool AllowRaggedTiles,                  ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
-    typename dp_vector_t,                   ///< Dot-product vector type along the K-axis
-    load_algorithm::kind_t LoadAlgorithm>   ///< Algorithm for loading a shared tile of KxL matrix data
-struct block_loader
-{
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    block_loader(
-        value_t *d_matrix,              ///< Pointer to input matrix
-        int matrix_values_l,            ///< Extent of the input matrix in value_t along the L-axis
-        int matrix_values_stride_k,     ///< Distance in value_t within pitched-linear memory between successive coordinates along the K-axis
-        int matrix_values_stride_l,     ///< Distance in value_t within pitched-linear memory between successive coordinates along the L-axis
-        int2 block_begin_item_coords,   ///< Thread block's starting value_t coordinates (l, k) within the input matrix
-        int block_end_item_k);          ///< Thread block's ending coordinate (k) within the input matrix (one-past)
-
-    //-------------------------------------------------------------------------
-    // Loader API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Request the current block-wide tile
-     */
-    void request();
-
-
-    /**
-     * Advance the loader to the next block-wide tile in the K-axis
-     */
-    void next();
-
-
-    /**
-     * Commit the previously-requested block-wide tile to shared memory
-     *
-     * NB: To facilitate padding for avoiding shared memory bank conflicts, we
-     * allow the row stride _BlockDpVectorsL to be arbitrarily bigger than the
-     * tile width BlockDpVectorsL.
-     */
-    template <int _BlockDpVectorsL>
-    void commit(
-        dp_vector_t (&scratch_tile)[BlockDpVectorsK][_BlockDpVectorsL]);
-
-};
-
-
-} // namespace gemm
-} // namespace cutlass
-
-
-/******************************************************************************
- * Tail-include specializations that adhere to the block_loader API
- ******************************************************************************/
-
-#include "block_loader_crosswise.h"
-#include "block_loader_congruous_dp1.h"
-#include "block_loader_congruous_idp4.h"

cutlass/gemm/block_loader_congruous_dp1.h

@@ -1,406 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Tile-loading abstraction for thread blocks
- */
-
-#include "../util/util.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * block_loader (CongruousCopy + dp1 specialization)
- ******************************************************************************/
-
-/**
- * \brief A three-phase data loading abstraction (prefetch, commit, and
- * advance) for iterating over ranges of block-wide matrix tiles.
- * (CongruousCopy + dp1 specialization)
- *
- * Each iteration sequence produces a KxL (height-by-width) block-wide tile of
- * value_t in shared memory.  The layout of the shared block-wide tile is
- * a row-major (L-major) tiling of singleton "dp1" dp_vector_t items, where
- * dp_vector_t == value_t.  Its dimensions are:
- *    K = BlockDpVectorsK
- *    L = BlockDpVectorsL
- *
- * The data is copied from a corresponding tile of global matrix data whose
- * layout of value_t is also L-major. This constitutes a CongruousCopy
- * between the L-major global tile and the L-major shared tile.
- *
- * NB: Because they are "dp1" singletons, the K-major orientation of
- * dp_vector_t in shared memory is irrelevant, and the L-major global and
- * shared tile layouts are perfectly congruous.  As a result, we can increase
- * the granularity of data transfer via vectorization of loads and stores
- * without any intermediate {dis|re}assembly.
- *
- * NB: Consecutive threads within a block are mapped in L-major
- * fashion across a first-set of LDG-vectors of dp_vector_t (value_t) within
- * their global tile. Successive sets of LDG-vectors are then strip-mined
- * as necessary down the K-axis.  These discontiguous LDG-vectors comprise the
- * thread's "slice" of the block-wide tile.
- */
-template <
-    int BlockThreads,           ///< Number of threads in each thread block (blockDim.x)
-    int BlockDpVectorsK,        ///< Extent of block-wide tile in dp_vector_t along the K-axis (height)
-    int BlockDpVectorsL,        ///< Extent of block-wide tile in dp_vector_t along the L-axis (width)
-    typename value_t,           ///< Input matrix value type
-    int LeadingDimAlignBytes,   ///< Byte alignment of input matrix leading dimension
-    bool AllowRaggedTiles       ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
->
-struct block_loader<
-    BlockThreads,
-    BlockDpVectorsK,
-    BlockDpVectorsL,
-    value_t,
-    LeadingDimAlignBytes,
-    AllowRaggedTiles,
-    value_t,                        ///< Dot-product vector type along the K-axis (dp1 specialization)
-    load_algorithm::CongruousCopy>  ///< Algorithm for loading a shared tile of KxL matrix data (CongruousCopy specialization)
-{
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    /// Dot-product vector type along the K-axis
-    typedef value_t dp_vector_t;
-
-    enum
-    {
-        /// Number of value_t in a dp_vector_t
-        DpVectorItems = divide_assert<sizeof(dp_vector_t), sizeof(value_t)>::value,
-
-        /// Number of dp_vector_t in a block-wide tile
-        BlockDpVectors = BlockDpVectorsK * BlockDpVectorsL,
-
-        /// Number of dp_vector_t in a thread-tile
-        ThreadDpVectors = divide_assert<BlockDpVectors, BlockThreads>::value,
-    };
-
-    /// Data movement type, coarsened by LeadingDimAlignBytes, capped by the
-    /// smaller of either ThreadDpVectors or BlockDpVectorsL
-    typedef io_vector<
-            dp_vector_t,
-            __NV_STD_MIN(ThreadDpVectors, BlockDpVectorsL),
-            LeadingDimAlignBytes>
-        ldg_vector_t;
-
-    enum
-    {
-        /// Number of dp_vector_t per ldg_vector_t
-        LdgVectorDpVectors = ldg_vector_t::VectorItems,
-
-        /// Number of value_t per ldg_vector_t
-        LdgVectorItems = LdgVectorDpVectors * DpVectorItems,
-
-
-
-        /// Total number of ldg_vector_t within each block-wide tile
-        BlockLdgVectors = divide_assert<BlockDpVectors, LdgVectorDpVectors>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along L-axis
-        BlockLdgVectorsL = divide_assert<BlockDpVectorsL, LdgVectorDpVectors>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along K-axis
-        BlockLdgVectorsK = BlockDpVectorsK,
-
-
-
-        /// Number of ldg_vector_t within each thread-tile
-        ThreadLdgVectors = divide_assert<BlockLdgVectors, BlockThreads>::value,
-
-        /// Extent of the thread tile in ldg_vector_t along L-axis
-        ThreadLdgVectorsL = __NV_STD_MAX(1, (BlockLdgVectorsL / BlockThreads)),
-
-        /// Extent of the thread tile in ldg_vector_t along K-axis
-        ThreadLdgVectorsK = divide_assert<ThreadLdgVectors, ThreadLdgVectorsL>::value,
-
-
-
-        /// Number of ldg_vector_t within each stripmine-tile
-        StripmineLdgVectors = BlockThreads,
-
-        /// Extent of the stripmine tile in ldg_vector_t along L-axis
-        StripmineLdgVectorsL = __NV_STD_MIN(BlockLdgVectorsL, StripmineLdgVectors),
-
-        /// Extent of the stripmine tile in ldg_vector_t along K-axis
-        StripmineLdgVectorsK = divide_assert<StripmineLdgVectors, StripmineLdgVectorsL>::value,
-
-
-
-        /// Alignment in dp_vector_t along L needed for committing prefetch
-        AlignmentDpVectorsL = LdgVectorDpVectors,
-    };
-
-    /// Predicate bit vector
-    typedef uint64_t predicate_mask_t;
-
-
-    //-------------------------------------------------------------------------
-    // Assert assumptions
-    //-------------------------------------------------------------------------
-
-    static_assert(
-        (ThreadLdgVectors <= sizeof(predicate_mask_t) * 8),
-        "Predicate mask type does not contain enough bits for encoding load predicates");
-
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// Input pointer to matrix in ldg_vector_t
-    ldg_vector_t *d_matrix_ldgvecs;
-
-    /// Extent of the input matrix in ldg_vector_t along the L-axis
-    int matrix_ldgvecs_l;
-
-    /// Thread block's ending ldg_vector_t coordinate (k) within the input matrix (one-past)
-    int block_end_ldgvec_k;
-
-    /// Predicate bits for guarding ldg_vector_t loads within "whole-k" block-wide tiles
-    predicate_mask_t guard;
-
-    /// Predicate bits for guarding ldg_vector_t loads within the final block-wide "residue" tile
-    predicate_mask_t residue_guard;
-
-    /// Iteration span in "whole-k" block-wide tiles
-    int wholek_tiles_remaining;
-
-    /// Distance in ldg_vector_t within pitched-linear memory between successive coordinates along the K-axis
-    int matrix_ldgvec_stride_k;
-
-    /// Distance in ldg_vector_t within pitched-linear memory between successive coordinates along the L-axis
-    int matrix_ldgvec_stride_l;
-
-    /// ldg_vector_t coordinates (l, k) of thread-tile within the block-wide tile
-    int2 block_thread_ldgvec_coords;
-
-    /// Thread-wide tile of prefetch data
-    ldg_vector_t thread_tile[ThreadLdgVectorsK][ThreadLdgVectorsL];
-
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    block_loader(
-        value_t *d_matrix_items,        ///< Input pointer to matrix in value_t
-        int matrix_items_l,             ///< Extent of the input matrix in value_t along the L-axis
-        int matrix_items_stride_k,      ///< Distance in value_t within pitched-linear memory between successive coordinates along the K-axis
-        int matrix_items_stride_l,      ///< Distance in value_t within pitched-linear memory between successive coordinates along the L-axis
-        int2 matrix_block_item_coords,  ///< value_t coordinates (l, k) of first block-wide tile within the input matrix
-        int block_end_item_k)           ///< Thread block's ending coordinate (k) within the input matrix (one-past)
-    :
-        block_end_ldgvec_k(block_end_item_k),
-        guard(0),
-        residue_guard(0)
-    {
-        matrix_ldgvecs_l = matrix_items_l / LdgVectorItems;
-        matrix_ldgvec_stride_k = matrix_items_stride_k / LdgVectorItems,
-        matrix_ldgvec_stride_l = matrix_items_stride_l;
-
-        // ldg_vector_t coordinates (l, k) of thread-tile within the block-wide tile
-        block_thread_ldgvec_coords = make_int2(
-            threadIdx.x % BlockLdgVectorsL,                 // l-coordinate
-            threadIdx.x / BlockLdgVectorsL);                // k-coordinate
-
-        // ldg_vector_t coordinates (l, k) of first block-wide tile within the input matrix
-        int2 matrix_block_ldgvec_coords = make_int2(
-            matrix_block_item_coords.x / LdgVectorItems,     // l-coordinate
-            matrix_block_item_coords.y);                    // k-coordinate
-
-        // Iteration span in ldg_vector_t
-        int span_ldgvec_k = (block_end_item_k - matrix_block_item_coords.y);
-
-
-
-        // ldg_vector_t coordinates (l, k) of first thread-tile tile within the input matrix
-        int2 matrix_thread_ldgvec_coords = make_int2(
-            block_thread_ldgvec_coords.x + matrix_block_ldgvec_coords.x,
-            block_thread_ldgvec_coords.y + matrix_block_ldgvec_coords.y);
-
-        // Iteration range in "whole-k" block-wide tiles
-        wholek_tiles_remaining = span_ldgvec_k / BlockLdgVectorsK;
-
-        // Extent of final residue-tile in ldg_vector_t along K-axis
-        int residue_ldgvecs_k = span_ldgvec_k % BlockLdgVectorsK;
-
-        // Initialize I/O predicates
-        if (AllowRaggedTiles)
-        {
-            // Outer thread-tile ldg_vector_t iteration (K-axis)
-            #pragma unroll
-            for (int thread_ldgvec_k = 0; thread_ldgvec_k < ThreadLdgVectorsK; ++thread_ldgvec_k)
-            {
-                int block_ldgvec_k = block_thread_ldgvec_coords.y + (thread_ldgvec_k * StripmineLdgVectorsK);
-
-                // Whether block_ldgvec_coords.y is valid in the final residue tile
-                predicate_mask_t valid_k = (block_ldgvec_k < residue_ldgvecs_k);
-
-                // Inner thread-tile ldg_vector_t iteration (L-axis)
-                #pragma unroll
-                for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-                {
-                    int block_ldgvec_l = block_thread_ldgvec_coords.x + (thread_ldgvec_l * StripmineLdgVectorsL);
-
-                    // Whether block_ldgvec_coords.x is valid any block-wide tile
-                    predicate_mask_t valid_l = (matrix_block_ldgvec_coords.x + block_ldgvec_l < matrix_ldgvecs_l);
-
-                    // Linear index of ldg_vector_t load
-                    int ldgvec_idx = thread_ldgvec_l + (thread_ldgvec_k * ThreadLdgVectorsL);
-
-                    // Set predicate guard bits
-                    guard |= (valid_l << ldgvec_idx);
-                    residue_guard |= ((valid_l & valid_k) << ldgvec_idx);
-                }
-            }
-
-            // Promote residue-guard to primary-guard if no full tiles remain
-            if (!wholek_tiles_remaining)
-            {
-                guard = residue_guard;
-            }
-        }
-
-        // Update the input pointer to be matrix_thread_ldgvec_coords
-        this->d_matrix_ldgvecs =
-            reinterpret_cast<ldg_vector_t*>(d_matrix_items) +
-            (matrix_thread_ldgvec_coords.y * matrix_ldgvec_stride_k) +
-            (matrix_thread_ldgvec_coords.x * matrix_ldgvec_stride_l);
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Loader API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Request the current block-wide tile
-     */
-    inline __device__
-    void request()
-    {
-        // Outer thread-tile ldg_vector_t iteration (K-axis)
-        #pragma unroll
-        for (int thread_ldgvec_k = 0; thread_ldgvec_k < ThreadLdgVectorsK; ++thread_ldgvec_k)
-        {
-            // Inner thread-tile ldg_vector_t iteration (L-axis)
-            #pragma unroll
-            for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-            {
-                // Linear index of ldg_vector_t load
-                int ldgvec_idx = (thread_ldgvec_k * ThreadLdgVectorsL) + thread_ldgvec_l;
-
-                // Unpack predicate guard
-                predicate_mask_t valid = ((guard >> ldgvec_idx) & 1);
-
-                if (!AllowRaggedTiles || valid)
-                {
-                    // Perform load
-                    thread_tile[thread_ldgvec_k][thread_ldgvec_l].load(
-                        d_matrix_ldgvecs +
-                        (thread_ldgvec_k * StripmineLdgVectorsK * matrix_ldgvec_stride_k) +
-                        (thread_ldgvec_l * StripmineLdgVectorsL * matrix_ldgvec_stride_l));
-                }
-                else
-                {
-                    // Zero-initialize
-                    #pragma unroll
-                    for (int dpvec = 0; dpvec < LdgVectorDpVectors; ++dpvec)
-                        thread_tile[thread_ldgvec_k][thread_ldgvec_l].buff[dpvec] = 0;
-                }
-            }
-        }
-    }
-
-
-    /**
-     * Advance the loader to the next block-wide tile in the K-axis
-     */
-    inline __device__
-    void next()
-    {
-        d_matrix_ldgvecs += (matrix_ldgvec_stride_k * BlockLdgVectorsK);
-
-        if (AllowRaggedTiles)
-        {
-            --wholek_tiles_remaining;
-
-            // Promote residue-guard to primary-guard if no full tiles remain
-            if (!wholek_tiles_remaining)
-            {
-                guard = residue_guard;
-            }
-        }
-    }
-
-
-    /**
-     * Commit the previously-requested block-wide tile to shared memory
-     *
-     * NB: To facilitate padding for avoiding shared memory bank conflicts, we
-     * allow the row stride SmemDpVectorsL to be arbitrarily bigger than the
-     * tile width BlockDpVectorsL.
-     */
-    template <int SmemDpVectorsL>
-    inline __device__
-    void commit(
-        dp_vector_t (&scratch_tile)[BlockDpVectorsK][SmemDpVectorsL])
-    {
-        static_assert(SmemDpVectorsL >= BlockDpVectorsL, "Row stride must be >= tile width.");
-
-        // Outer thread-tile ldg_vector_t iteration (K-axis)
-        #pragma unroll
-        for (int thread_ldgvec_k = 0; thread_ldgvec_k < ThreadLdgVectorsK; ++thread_ldgvec_k)
-        {
-            int block_ldgvec_k = block_thread_ldgvec_coords.y + (thread_ldgvec_k * StripmineLdgVectorsK);
-
-            // Inner thread-tile ldg_vector_t iteration (L-axis)
-            #pragma unroll
-            for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-            {
-                int block_ldgvec_l = block_thread_ldgvec_coords.x + (thread_ldgvec_l * StripmineLdgVectorsL);
-
-                thread_tile[thread_ldgvec_k][thread_ldgvec_l].store(
-                    &scratch_tile[block_ldgvec_k][block_ldgvec_l * LdgVectorDpVectors]);
-            }
-        }
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/block_loader_congruous_idp4.h

@@ -1,544 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Tile-loading abstraction for thread blocks
- */
-
-#include "../util/util.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * block_loader (CongruousCopy + idp4 specialization)
- ******************************************************************************/
-
-/**
- * \brief A three-phase data loading abstraction (prefetch, commit, and
- * advance) for iterating over ranges of block-wide matrix tiles.
- * (CongruousCopy + idp4 specialization)
- *
- * Each iteration sequence produces a KxL (height-by-width) block-wide tile of
- * value_t in shared memory. The layout of the shared block-wide tile is
- * a row-major (L-major) tiling of int32_t dp_vector_t, which are themselves
- * column-major (K-major) vectors of int8_t value_t.  Its dimensions are:
- *    K = BlockDpVectorsK * (sizeof(dp_vector_t) / sizeof(value_t)
- *    L = BlockDpVectorsL
- *
- * The data is copied from a corresponding tile of global matrix data whose
- * layout of value_t is also L-major. This constitutes a CongruousCopy between
- * the L-major global tile and the L-major shared tile.
- *
- * NB: The K-major value_t in shared dp_vector_t are imperfectly congruous
- * with the L-major value_t in global memory. As a result, the granularity
- * of data transfer is a "dp-square" of (DpVectorItems * DpVectorItems) values
- * that must be transposed from L-oriented dp_vector_t to K-oriented
- * dp_vector_t prior to commitment.
- *
- * NB: Consecutive threads within a block are mapped in L-major
- * fashion across a first-set of squares within their global tile. Successive
- * sets of squares are then strip-mined as necessary down the K-axis.  These
- * discontiguous squares comprise the thread's "slice" of the block-wide tile.
- */
-template <
-    int BlockThreads,           ///< Number of threads in each thread block (blockDim.x)
-    int _BlockDpVectorsK,        ///< Extent of block-wide tile in dp_vector_t along the K-axis (height)
-    int _BlockDpVectorsL,        ///< Extent of block-wide tile in dp_vector_t along the L-axis (width)
-    int LeadingDimAlignBytes,   ///< Byte alignment of input matrix leading dimension
-    bool AllowRaggedTiles       ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
->
-struct block_loader<
-    BlockThreads,
-    _BlockDpVectorsK,
-    _BlockDpVectorsL,
-    int8_t,                         ///< Input matrix value type (idp4 specialization)
-    LeadingDimAlignBytes,
-    AllowRaggedTiles,
-    int32_t,                        ///< Dot-product vector type along the K-axis (idp4 specialization)
-    load_algorithm::CongruousCopy>  ///< Algorithm for loading a shared tile of KxL matrix data (CrosswiseCopy specialization)
-{
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    /// Input matrix value type
-    typedef int8_t value_t;
-
-    /// Dot-product vector type along the K-axis
-    typedef int32_t dp_vector_t;
-
-    enum
-    {
-        /// Number of value_t in a dp_vector_t
-        DpVectorItems = divide_assert<sizeof(dp_vector_t), sizeof(value_t)>::value,
-
-        /// Number of dp_vector_t in a block-wide tile
-        BlockDpVectors = _BlockDpVectorsK * _BlockDpVectorsL,
-
-        /// Number of dp_vector_t in a thread-tile
-        ThreadDpVectors = divide_assert<BlockDpVectors, BlockThreads>::value,
-
-        /// Number of dp_vector_t in a dp-square
-        SquareDpVectors = DpVectorItems,
-
-        /// Number of dp-square tiles in a thread-tile
-        ThreadSquares = divide_assert<ThreadDpVectors, SquareDpVectors>::value,
-
-        /// Extent of block-wide tile in transposed dp_vector_t along the K-axis (height)
-        BlockTransDpVectorsK = _BlockDpVectorsK * DpVectorItems,
-
-        /// Extent of block-wide tile in transposed dp_vector_t along the L-axis (height)
-        BlockTransDpVectorsL = divide_assert<_BlockDpVectorsL, DpVectorItems>::value,
-
-
-    };
-
-    /// Load-from-global data movement type, coarsened by LeadingDimAlignBytes, capped by the
-    /// smaller of either ThreadSquares or BlockTransDpVectorsL
-    typedef io_vector<
-            dp_vector_t,
-            __NV_STD_MIN(ThreadSquares, BlockTransDpVectorsL),
-            LeadingDimAlignBytes>
-        ldg_vector_t;
-
-    /// Store-to-shared data movement type equivalent to a dp-square
-    typedef io_vector<
-            dp_vector_t,
-            SquareDpVectors>
-        sts_vector_t;
-
-    enum
-    {
-        /// Number of dp_vector_t per ldg_vector_t
-        LdgVectorDpVectors = ldg_vector_t::VectorItems,
-
-        /// Number of value_t per ldg_vector_t
-        LdgVectorItems = LdgVectorDpVectors * DpVectorItems,
-
-
-
-        /// Total number of ldg_vector_t within each block-wide tile
-        BlockLdgVectors = divide_assert<BlockDpVectors, LdgVectorDpVectors>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along L-axis
-        BlockLdgVectorsL = divide_assert<BlockTransDpVectorsL, LdgVectorDpVectors>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along K-axis
-        BlockLdgVectorsK = BlockTransDpVectorsK,
-
-
-
-        /// Number of ldg_vector_t within each thread-tile
-        ThreadLdgVectors = divide_assert<BlockLdgVectors, BlockThreads>::value,
-
-        /// Extent of the thread tile in ldg_vector_t along L-axis
-        ThreadLdgVectorsL = __NV_STD_MAX(1, (BlockLdgVectorsL / BlockThreads)),
-
-        /// Extent of the thread tile in ldg_vector_t along K-axis
-        ThreadLdgVectorsK = divide_assert<ThreadLdgVectors, ThreadLdgVectorsL>::value,
-
-        /// Extent of the thread tile in dp-square tiles along K-axis
-        ThreadSquaresK = divide_assert<ThreadLdgVectorsK, SquareDpVectors>::value,
-
-
-
-        /// Number of ldg_vector_t within each stripmine-tile
-        StripmineLdgVectors = BlockThreads * SquareDpVectors,
-
-        /// Extent of the stripmine tile in ldg_vector_t along L-axis
-        StripmineLdgVectorsL = __NV_STD_MIN(BlockLdgVectorsL, BlockThreads),
-
-        /// Extent of the stripmine tile in ldg_vector_t along K-axis
-        StripmineLdgVectorsK = divide_assert<StripmineLdgVectors, StripmineLdgVectorsL>::value,
-
-        /// Extent of the stripmine tile in dp-square tiles along K-axis
-        StripmineSquaresK = divide_assert<StripmineLdgVectorsK, SquareDpVectors>::value,
-
-
-
-        /// Alignment in dp_vector_t along L needed for committing prefetch
-        AlignmentDpVectorsL = LdgVectorDpVectors,
-    };
-
-    /// Predicate mask type
-    typedef uint32_t predicate_mask_t;
-
-
-    //-------------------------------------------------------------------------
-    // Assert assumptions
-    //-------------------------------------------------------------------------
-
-    static_assert((LeadingDimAlignBytes >= 4) && (LeadingDimAlignBytes % 4 == 0),
-        "Alignment for matrix operands to IGEMM must be a multiple of 4 bytes.");
-
-    static_assert(
-        (ThreadLdgVectors <= sizeof(predicate_mask_t) * 8),
-        "Predicate mask type does not contain enough bits for encoding load predicates");
-
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// Input pointer to matrix in ldg_vector_t
-    ldg_vector_t *d_matrix_ldgvecs;
-
-    /// Extent of the input matrix in ldg_vector_t along the L-axis
-    int matrix_ldgvecs_l;
-
-    /// Thread block's ending ldg_vector_t coordinate (k) within the input matrix (one-past)
-    int block_end_ldgvec_k;
-
-    /// Predicate bits for guarding ldg_vector_t loads within "whole-k" block-wide tiles
-    predicate_mask_t guard;
-
-    /// Predicate bits for guarding ldg_vector_t loads within the final block-wide "residue" tile
-    predicate_mask_t residue_guard;
-
-    /// Iteration span in "whole-k" block-wide tiles
-    int wholek_tiles_remaining;
-
-    /// Distance in ldg_vector_t within pitched-linear memory between successive coordinates along the K-axis
-    int matrix_ldgvec_stride_k;
-
-    /// Distance in ldg_vector_t within pitched-linear memory between successive coordinates along the L-axis
-    int matrix_ldgvec_stride_l;
-
-    /// ldg_vector_t coordinates (l, k) of thread-tile within the block-wide tile
-    int2 block_thread_ldgvec_coords;
-
-    /// Thread-wide tile of prefetch data
-    ldg_vector_t thread_tile[ThreadSquaresK][SquareDpVectors][ThreadLdgVectorsL];
-
-
-
-    //-------------------------------------------------------------------------
-    // Utility methods
-    //-------------------------------------------------------------------------
-
-
-    /**
-     * \brief Byte-permute. Pick four arbitrary bytes from two 32-bit registers, and reassemble them into a 32-bit destination register.  For SM2.0 or later.
-     *
-     * \par
-     * The bytes in the two source registers \p a and \p b are numbered from 0 to 7:
-     * {\p b, \p a} = {{b7, b6, b5, b4}, {b3, b2, b1, b0}}. For each of the four bytes
-     * {b3, b2, b1, b0} selected in the return value, a 4-bit selector is defined within
-     * the four lower "nibbles" of \p index: {\p index } = {n7, n6, n5, n4, n3, n2, n1, n0}
-     *
-     * \par Snippet
-     * The code snippet below illustrates byte-permute.
-     * \par
-     * \code
-     * #include <cub/cub.cuh>
-     *
-     * __global__ void ExampleKernel(...)
-     * {
-     *     int a        = 0x03020100;
-     *     int b        = 0x07060504;
-     *     int index    = 0x00007531;
-     *
-     *     int selected = prmt(a, b, index);    // 0x07050301
-     *
-     * \endcode
-     *
-     */
-    inline __device__
-    int32_t prmt(int32_t a, int32_t b, unsigned int index)
-    {
-        int ret;
-        asm volatile("prmt.b32 %0, %1, %2, %3;" : "=r"(ret) : "r"(a), "r"(b), "r"(index));
-        return ret;
-    }
-
-
-    /**
-     * Convert a "dp-square" from L-major to K-major
-     */
-    inline __device__
-    void transpose_dp_square(dp_vector_t (&dp_square)[SquareDpVectors])
-    {
-        // Transpose dp_vector_t squares
-        int32_t y = prmt(dp_square[0], dp_square[1], 0x00007362);
-        int32_t w = prmt(dp_square[2], dp_square[3], 0x00007362);
-        int32_t x = prmt(dp_square[0], dp_square[1], 0x00005140);
-        int32_t z = prmt(dp_square[2], dp_square[3], 0x00005140);
-
-        dp_square[0] = prmt(x, z, 0x00005410);
-        dp_square[1] = prmt(x, z, 0x00007632);
-        dp_square[2] = prmt(y, w, 0x00005410);
-        dp_square[3] = prmt(y, w, 0x00007632);
-    }
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    block_loader(
-        value_t *d_matrix_items,        ///< Input pointer to matrix in value_t
-        int matrix_items_l,             ///< Extent of the input matrix in value_t along the L-axis
-        int matrix_items_stride_k,      ///< Distance in value_t within pitched-linear memory between successive coordinates along the K-axis
-        int matrix_items_stride_l,      ///< Distance in value_t within pitched-linear memory between successive coordinates along the L-axis
-        int2 matrix_block_item_coords,  ///< value_t coordinates (l, k) of first block-wide tile within the input matrix
-        int block_end_item_k)           ///< Thread block's ending coordinate (k) within the input matrix (one-past)
-    :
-        block_end_ldgvec_k(block_end_item_k),
-        guard(0),
-        residue_guard(0)
-    {
-        matrix_ldgvecs_l = matrix_items_l / LdgVectorItems;
-        matrix_ldgvec_stride_k = matrix_items_stride_k / LdgVectorItems,
-        matrix_ldgvec_stride_l = matrix_items_stride_l;
-
-        // ldg_vector_t coordinates (l, k) of thread-tile within the block-wide tile
-        block_thread_ldgvec_coords = make_int2(
-            threadIdx.x % BlockLdgVectorsL,                      // l-coordinate
-            (threadIdx.x / BlockLdgVectorsL) * SquareDpVectors); // k-coordinate
-
-        // ldg_vector_t coordinates (l, k) of first block-wide tile within the input matrix
-        int2 matrix_block_ldgvec_coords = make_int2(
-            matrix_block_item_coords.x / LdgVectorItems,         // l-coordinate
-            matrix_block_item_coords.y);                        // k-coordinate
-
-        // Iteration span in ldg_vector_t
-        int span_ldgvec_k = (block_end_item_k - matrix_block_item_coords.y);
-
-
-
-        // ldg_vector_t coordinates (l, k) of first thread-tile tile within the input matrix
-        int2 matrix_thread_ldgvec_coords = make_int2(
-            block_thread_ldgvec_coords.x + matrix_block_ldgvec_coords.x,
-            block_thread_ldgvec_coords.y + matrix_block_ldgvec_coords.y);
-
-        // Iteration range in "whole-k" block-wide tiles
-        wholek_tiles_remaining = span_ldgvec_k / BlockLdgVectorsK;
-
-        // Extent of final residue-tile in ldg_vector_t along K-axis
-        int residue_ldgvecs_k = span_ldgvec_k % BlockLdgVectorsK;
-
-        // Initialize I/O predicates
-        if (AllowRaggedTiles)
-        {
-            // Iterate through rows of squares in thread tile
-            #pragma unroll
-            for (int thread_square_k = 0; thread_square_k < ThreadSquaresK; ++thread_square_k)
-            {
-                // Iterate through rows of dp_vector_t in each square
-                #pragma unroll
-                for (int square_dpvec = 0; square_dpvec < SquareDpVectors; ++square_dpvec)
-                {
-                    // ldg_vector_t K-coordinate in block-wide tile (K-axis strip-mining of ldg_vector_t within block-tile)
-                    int block_ldgvec_k =
-                        block_thread_ldgvec_coords.y +
-                        (thread_square_k * StripmineLdgVectorsK) +
-                        square_dpvec;
-
-                    // Whether block_ldgvec_coords.y is valid in the final residue tile
-                    predicate_mask_t valid_k = (block_ldgvec_k < residue_ldgvecs_k);
-
-                    // L-axis strip-mining of block-tile
-                    #pragma unroll
-                    for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-                    {
-                        // ldg_vector_t L-coordinate in block-wide tile (L-axis strip-mining of ldg_vector_t within block-tile)
-                        int block_ldgvec_l = block_thread_ldgvec_coords.x + (thread_ldgvec_l * StripmineLdgVectorsL);
-
-                        // Whether block_ldgvec_coords.x is valid any block-wide tile
-                        predicate_mask_t valid_l = (matrix_block_ldgvec_coords.x + block_ldgvec_l < matrix_ldgvecs_l);
-
-                        // Linear index of ldg_vector_t load
-                        int ldgvec_idx =
-                            (thread_square_k * SquareDpVectors * ThreadLdgVectorsL) +
-                            (square_dpvec * ThreadLdgVectorsL) +
-                            thread_ldgvec_l;
-
-                        // Set predicate guard bits
-                        guard |= (valid_l << ldgvec_idx);
-                        residue_guard |= ((valid_l & valid_k) << ldgvec_idx);
-                    }
-                }
-            }
-
-            // Promote residue-guard to primary-guard if no full tiles remain
-            if (!wholek_tiles_remaining)
-            {
-                guard = residue_guard;
-            }
-        }
-
-        // Update the input pointer to be matrix_thread_ldgvec_coords
-        this->d_matrix_ldgvecs =
-            reinterpret_cast<ldg_vector_t*>(d_matrix_items) +
-            (matrix_thread_ldgvec_coords.y * matrix_ldgvec_stride_k) +
-            (matrix_thread_ldgvec_coords.x * matrix_ldgvec_stride_l);
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Loader API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Request the current block-wide tile
-     */
-    inline __device__
-    void request()
-    {
-        // Each thread iterates through the ldg_vector_t in its thread tile
-
-        // Iterate through rows of squares in thread tile
-        #pragma unroll
-        for (int thread_square_k = 0; thread_square_k < ThreadSquaresK; ++thread_square_k)
-        {
-            // Iterate through rows of dp_vector_t in each square
-            #pragma unroll
-            for (int square_dpvec = 0; square_dpvec < SquareDpVectors; ++square_dpvec)
-            {
-                // Iterate through ldg_vector_t in each row
-                #pragma unroll
-                for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-                {
-                    // Linear index of ldg_vector_t load
-                    int ldgvec_idx =
-                        (thread_square_k * SquareDpVectors * ThreadLdgVectorsL) +
-                        (square_dpvec * ThreadLdgVectorsL) +
-                        thread_ldgvec_l;
-
-                    // Unpack predicate guard
-                    predicate_mask_t valid = ((guard >> ldgvec_idx) & 1);
-
-                    if (!AllowRaggedTiles || valid)
-                    {
-                        // Perform load
-                        thread_tile[thread_square_k][square_dpvec][thread_ldgvec_l].load(
-                            d_matrix_ldgvecs +
-                            (((thread_square_k * StripmineLdgVectorsK) + square_dpvec) * matrix_ldgvec_stride_k) +
-                            (thread_ldgvec_l * StripmineLdgVectorsL * matrix_ldgvec_stride_l));
-                    }
-                    else
-                    {
-                        // Zero-initialize
-                        #pragma unroll
-                        for (int dpvec = 0; dpvec < LdgVectorDpVectors; ++dpvec)
-                            thread_tile[thread_square_k][square_dpvec][thread_ldgvec_l].buff[dpvec] = 0;
-                    }
-                }
-            }
-        }
-    }
-
-
-    /**
-     * Advance the loader to the next block-wide tile in the K-axis
-     */
-    inline __device__
-    void next()
-    {
-        d_matrix_ldgvecs += (matrix_ldgvec_stride_k * BlockLdgVectorsK);
-
-        if (AllowRaggedTiles)
-        {
-            --wholek_tiles_remaining;
-
-            // Promote residue-guard to primary-guard if no full tiles remain
-            if (!wholek_tiles_remaining)
-            {
-                guard = residue_guard;
-            }
-        }
-    }
-
-
-    /**
-     * Commit the previously-requested block-wide tile to shared memory
-     *
-     * NB: To facilitate padding for avoiding shared memory bank conflicts, we
-     * allow the row stride SmemDpVectorsL to be arbitrarily bigger than the
-     * tile width BlockDpVectorsL.
-     */
-    template <int SmemDpVectorsL>
-    inline __device__
-    void commit(
-        dp_vector_t (&scratch_tile)[_BlockDpVectorsK][SmemDpVectorsL])
-    {
-        static_assert(SmemDpVectorsL >= _BlockDpVectorsL, "Row stride must be >= tile width.");
-
-        // Square K-coordinate of thread tile in block-wide tile
-        int block_thread_square_k = block_thread_ldgvec_coords.y / SquareDpVectors;
-
-        // Iterate through rows of squares in thread tile
-        #pragma unroll
-        for (int thread_square_k = 0; thread_square_k < ThreadSquaresK; ++thread_square_k)
-        {
-            // Square K-coordinate in block-wide tile (K-axis strip-mining of squares within block-tile)
-            int block_square_k = block_thread_square_k + (thread_square_k * StripmineSquaresK);
-
-            // Iterate through ldg_vector_t in each row
-            #pragma unroll
-            for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-            {
-                // ldg_vector_t L-coordinate in block-wide tile (L-axis strip-mining of ldg_vector_t within block-tile)
-                int block_ldgvec_l = block_thread_ldgvec_coords.x + (thread_ldgvec_l * StripmineLdgVectorsL);
-
-                // Iterate through squares in each ldg_vector_t
-                #pragma unroll
-                for (int ldgvec_dpvec_l = 0; ldgvec_dpvec_l < LdgVectorDpVectors; ++ldgvec_dpvec_l)
-                {
-                    // Square L-coordinate in block-wide tile (L-axis raking of square-slices within ldg_vector_t)
-                    int block_square_l = (block_ldgvec_l * LdgVectorDpVectors) + ldgvec_dpvec_l;
-
-                    // Assemble square of L-major dp_vector_t from stack of slices
-                    sts_vector_t square;
-
-                    // Iterate through rows of dp_vector_t in each square
-                    #pragma unroll
-                    for (int square_dpvec = 0; square_dpvec < SquareDpVectors; ++square_dpvec)
-                    {
-                        square.buff[square_dpvec] = thread_tile[thread_square_k][square_dpvec][thread_ldgvec_l].buff[ldgvec_dpvec_l];
-                    }
-
-                    // Un-transpose square from L-major to K-major
-                    transpose_dp_square(square.buff);
-
-                    // Store dp-square
-                    square.store(&scratch_tile[block_square_k][block_square_l * SquareDpVectors]);
-                }
-            }
-        }
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/block_loader_crosswise.h

@@ -1,411 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Tile-loading abstraction for thread blocks
- */
-
-#include "../util/util.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * block_loader (CrosswiseCopy specialization)
- ******************************************************************************/
-
-/**
- * \brief A three-phase data loading abstraction (prefetch, commit, and
- * advance) for iterating over ranges of block-wide matrix tiles.
- * (CrosswiseCopy specialization)
- *
- * Each iteration sequence produces a KxL (height-by-width) block-wide tile of
- * value_t in shared memory.  The layout of the shared block-wide tile is
- * a row-major (L-major) tiling of dp_vector_t items, which are themselves
- * column-major (K-major) vectors of value_t.  Its dimensions are:
- *    K = BlockDpVectorsK * (sizeof(dp_vector_t) / sizeof(value_t)
- *    L = BlockDpVectorsL
- *
- * The data is copied from a corresponding tile of global matrix data whose
- * layout of value_t is K-major.  This constitutes a CrosswiseCopy between
- * the K-major global tile and the L-major shared tile.
- *
- * NB: The orientation of dp_vector_t components in shared memory is congruous
- * with the global matrix data, so we can use dp_vector_t as the minimum
- * granularity of data transfer without any intermediate {dis|re}assembly
- * of its value_t components.  However, the global and shared memory layouts
- * of dp_vector_t items are cross-wise with respect to each other, so any
- * further LDG-vectorization of dp_vector_t data requires intermediate
- * disassembly into dp_vector_t components to be stored individually into
- * the shared tile.
- *
- * NB: Consecutive threads within a block are mapped in K-major
- * fashion down a first set of LDG-vectors of dp_vector_t within their global
- * tile. Successive sets of LDG-vectors are then strip-mined as necessary
- * across the L-axis.  These discontiguous LDG-vectors comprise the thread's
- * "slice" of the block-wide tile.
- */
-template <
-    int BlockThreads,           ///< Number of threads in each thread block (blockDim.x)
-    int BlockDpVectorsK,        ///< Extent of block-wide tile in dp_vector_t along the K-axis (height)
-    int BlockDpVectorsL,        ///< Extent of block-wide tile in dp_vector_t along the L-axis (width)
-    typename value_t,           ///< Input matrix value type
-    int LeadingDimAlignBytes,   ///< Byte alignment of input matrix leading dimension
-    bool AllowRaggedTiles,      ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
-    typename dp_vector_t>       ///< Dot-product vector type along the K-axis
-struct block_loader<
-    BlockThreads,
-    BlockDpVectorsK,
-    BlockDpVectorsL,
-    value_t,
-    LeadingDimAlignBytes,
-    AllowRaggedTiles,
-    dp_vector_t,
-    load_algorithm::CrosswiseCopy>  ///< Algorithm for loading a shared tile of KxL matrix data (CrosswiseCopy specialization)
-{
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    enum
-    {
-        /// Number of value_t in a dp_vector_t
-        DpVectorItems = divide_assert<sizeof(dp_vector_t), sizeof(value_t)>::value,
-
-        /// Number of dp_vector_t in a block-wide tile
-        BlockDpVectors = BlockDpVectorsK * BlockDpVectorsL,
-
-        /// Number of dp_vector_t in a thread-tile
-        ThreadDpVectors = divide_assert<BlockDpVectors, BlockThreads>::value,
-    };
-
-    /// Data movement type, coarsened by LeadingDimAlignBytes, capped by the
-    /// smaller of either ThreadDpVectors or BlockDpVectorsK
-    typedef io_vector<
-            dp_vector_t,
-            __NV_STD_MIN(ThreadDpVectors, BlockDpVectorsK),
-            LeadingDimAlignBytes>
-        ldg_vector_t;
-
-    enum
-    {
-        /// Number of dp_vector_t per ldg_vector_t
-        LdgVectorDpVectors = ldg_vector_t::VectorItems,
-
-        /// Number of value_t per ldg_vector_t
-        LdgVectorItems = LdgVectorDpVectors * DpVectorItems,
-
-
-
-        /// Total number of ldg_vector_t within each block-wide tile
-        BlockLdgVectors = divide_assert<BlockDpVectors, LdgVectorDpVectors>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along K-axis
-        BlockLdgVectorsK = divide_assert<BlockDpVectorsK, LdgVectorDpVectors>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along L-axis
-        BlockLdgVectorsL = BlockDpVectorsL,
-
-
-
-        /// Number of ldg_vector_t within each thread-tile
-        ThreadLdgVectors = divide_assert<BlockLdgVectors, BlockThreads>::value,
-
-        /// Extent of the thread tile in ldg_vector_t along K-axis
-        ThreadLdgVectorsK = __NV_STD_MAX(1, (BlockLdgVectorsK / BlockThreads)),
-
-        /// Extent of the thread tile in ldg_vector_t along L-axis
-        ThreadLdgVectorsL = divide_assert<ThreadLdgVectors, ThreadLdgVectorsK>::value,
-
-
-
-        /// Number of ldg_vector_t within each stripmine-tile
-        StripmineLdgVectors = BlockThreads,
-
-        /// Extent of the stripmine tile in ldg_vector_t along K-axis
-        StripmineLdgVectorsK = __NV_STD_MIN(BlockLdgVectorsK, StripmineLdgVectors),
-
-        /// Extent of the stripmine tile in ldg_vector_t along L-axis
-        StripmineLdgVectorsL = divide_assert<StripmineLdgVectors, StripmineLdgVectorsK>::value,
-
-
-
-        /// Alignment in dp_vector_t along L needed for committing prefetch
-        AlignmentDpVectorsL = 1,
-    };
-
-    /// Predicate bit vector
-    typedef uint64_t predicate_mask_t;
-
-
-    //-------------------------------------------------------------------------
-    // Assert assumptions
-    //-------------------------------------------------------------------------
-
-    static_assert(
-        (ThreadLdgVectors <= sizeof(predicate_mask_t) * 8),
-        "Predicate mask type does not contain enough bits for encoding load predicates");
-
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// Input pointer to matrix in ldg_vector_t
-    ldg_vector_t *d_matrix_ldgvecs;
-
-    /// Extent of the input matrix in ldg_vector_t along the L-axis
-    int matrix_ldgvecs_l;
-
-    /// Thread block's ending ldg_vector_t coordinate (k) within the input matrix (one-past)
-    int block_end_ldgvec_k;
-
-    /// Predicate bits for guarding ldg_vector_t loads within "whole-k" block-wide tiles
-    predicate_mask_t guard;
-
-    /// Predicate bits for guarding ldg_vector_t loads within the final block-wide "residue" tile
-    predicate_mask_t residue_guard;
-
-    /// Iteration span in "whole-k" block-wide tiles
-    int wholek_tiles_remaining;
-
-    /// Distance in ldg_vector_t within pitched-linear memory between successive coordinates along the K-axis
-    int matrix_ldgvec_stride_k;
-
-    /// Distance in ldg_vector_t within pitched-linear memory between successive coordinates along the L-axis
-    int matrix_ldgvec_stride_l;
-
-    /// ldg_vector_t coordinates (l, k) of thread-tile within the block-wide tile
-    int2 block_thread_ldgvec_coords;
-
-    /// Thread-wide tile of prefetch data
-    ldg_vector_t thread_tile[ThreadLdgVectorsK][ThreadLdgVectorsL];
-
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    block_loader(
-        value_t *d_matrix_items,        ///< Input pointer to matrix in value_t
-        int matrix_items_l,             ///< Extent of the input matrix in value_t along the L-axis
-        int matrix_items_stride_k,      ///< Distance in value_t within pitched-linear memory between successive coordinates along the K-axis
-        int matrix_items_stride_l,      ///< Distance in value_t within pitched-linear memory between successive coordinates along the L-axis
-        int2 matrix_block_item_coords,  ///< value_t coordinates (l, k) of first block-wide tile within the input matrix
-        int block_end_item_k)           ///< Thread block's ending coordinate (k) within the input matrix (one-past)
-    :
-        block_end_ldgvec_k(block_end_item_k),
-        guard(0),
-        residue_guard(0)
-    {
-        matrix_ldgvecs_l = matrix_items_l;
-        matrix_ldgvec_stride_k = matrix_items_stride_k;
-        matrix_ldgvec_stride_l = (matrix_items_stride_l / LdgVectorItems);
-
-        // ldg_vector_t coordinates (l, k) of thread-tile within the block-wide tile
-        block_thread_ldgvec_coords = make_int2(
-            (threadIdx.x / BlockLdgVectorsK),                // l-coordinate
-            (threadIdx.x % BlockLdgVectorsK));               // k-coordinate
-
-        // ldg_vector_t coordinates (l, k) of first block-wide tile within the input matrix
-        int2 matrix_block_ldgvec_coords = make_int2(
-            matrix_block_item_coords.x,                     // l-coordinate
-            matrix_block_item_coords.y / LdgVectorItems);    // k-coordinate
-
-        // Iteration span in ldg_vector_t
-        int span_ldgvec_k = (block_end_item_k - matrix_block_item_coords.y) / LdgVectorItems;
-
-
-
-        // ldg_vector_t coordinates (l, k) of first thread-tile tile within the input matrix
-        int2 matrix_thread_ldgvec_coords = make_int2(
-            block_thread_ldgvec_coords.x + matrix_block_ldgvec_coords.x,
-            block_thread_ldgvec_coords.y + matrix_block_ldgvec_coords.y);
-
-        // Iteration range in "whole-k" block-wide tiles
-        wholek_tiles_remaining = span_ldgvec_k / BlockLdgVectorsK;
-
-        // Extent of final residue-tile in ldg_vector_t along K-axis
-        int residue_ldgvecs_k = span_ldgvec_k % BlockLdgVectorsK;
-
-        // Initialize I/O predicates
-        if (AllowRaggedTiles)
-        {
-            // Outer thread-tile ldg_vector_t iteration (K-axis)
-            #pragma unroll
-            for (int thread_ldgvec_k = 0; thread_ldgvec_k < ThreadLdgVectorsK; ++thread_ldgvec_k)
-            {
-                int block_ldgvec_k = block_thread_ldgvec_coords.y + (thread_ldgvec_k * StripmineLdgVectorsK);
-
-                // Whether block_ldgvec_coords.y is valid in the final residue tile
-                predicate_mask_t valid_k = (block_ldgvec_k < residue_ldgvecs_k);
-
-                // Inner thread-tile ldg_vector_t iteration (L-axis)
-                #pragma unroll
-                for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-                {
-                    int block_ldgvec_l = block_thread_ldgvec_coords.x + (thread_ldgvec_l * StripmineLdgVectorsL);
-
-                    // Whether block_ldgvec_coords.x is valid any block-wide tile
-                    predicate_mask_t valid_l = (matrix_block_ldgvec_coords.x + block_ldgvec_l < matrix_ldgvecs_l);
-
-                    // Linear index of ldg_vector_t load
-                    int ldgvec_idx = thread_ldgvec_l + (thread_ldgvec_k * ThreadLdgVectorsL);
-
-                    // Set predicate guard bits
-                    guard |= (valid_l << ldgvec_idx);
-                    residue_guard |= ((valid_l & valid_k) << ldgvec_idx);
-                }
-            }
-
-            // Promote residue-guard to primary-guard if no full tiles remain
-            if (!wholek_tiles_remaining)
-            {
-                guard = residue_guard;
-            }
-        }
-
-        // Update the input pointer to be matrix_thread_ldgvec_coords
-        this->d_matrix_ldgvecs =
-            reinterpret_cast<ldg_vector_t*>(d_matrix_items) +
-            (matrix_thread_ldgvec_coords.y * matrix_ldgvec_stride_k) +
-            (matrix_thread_ldgvec_coords.x * matrix_ldgvec_stride_l);
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Loader API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Request the current block-wide tile
-     */
-    inline __device__
-    void request()
-    {
-        // Outer thread-tile ldg_vector_t iteration (K-axis)
-        #pragma unroll
-        for (int thread_ldgvec_k = 0; thread_ldgvec_k < ThreadLdgVectorsK; ++thread_ldgvec_k)
-        {
-            // Inner thread-tile ldg_vector_t iteration (L-axis)
-            #pragma unroll
-            for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-            {
-                // Linear index of ldg_vector_t load
-                int ldgvec_idx = (thread_ldgvec_k * ThreadLdgVectorsL) + thread_ldgvec_l;
-
-                // Unpack predicate guard
-                predicate_mask_t valid = ((guard >> ldgvec_idx) & 1);
-
-                if (!AllowRaggedTiles || valid)
-                {
-                    // Perform load
-                    thread_tile[thread_ldgvec_k][thread_ldgvec_l].load(
-                        d_matrix_ldgvecs +
-                        (thread_ldgvec_k * StripmineLdgVectorsK * matrix_ldgvec_stride_k) +
-                        (thread_ldgvec_l * StripmineLdgVectorsL * matrix_ldgvec_stride_l));
-                }
-                else
-                {
-                    // Zero-initialize
-                    #pragma unroll
-                    for (int dpvec = 0; dpvec < LdgVectorDpVectors; ++dpvec)
-                        thread_tile[thread_ldgvec_k][thread_ldgvec_l].buff[dpvec] = 0;
-                }
-            }
-        }
-    }
-
-
-    /**
-     * Advance the loader to the next block-wide tile in the K-axis
-     */
-    inline __device__
-    void next()
-    {
-        d_matrix_ldgvecs += (matrix_ldgvec_stride_k * BlockLdgVectorsK);
-
-        if (AllowRaggedTiles)
-        {
-            --wholek_tiles_remaining;
-
-            // Promote residue-guard to primary-guard if no full tiles remain
-            if (!wholek_tiles_remaining)
-            {
-                guard = residue_guard;
-            }
-        }
-    }
-
-
-    /**
-     * Commit the previously-requested block-wide tile to shared memory
-     *
-     * NB: To facilitate padding for avoiding shared memory bank conflicts, we
-     * allow the row stride SmemDpVectorsL to be arbitrarily bigger than the
-     * tile width BlockDpVectorsL.
-     */
-    template <int SmemDpVectorsL>
-    inline __device__
-    void commit(
-        dp_vector_t (&scratch_tile)[BlockDpVectorsK][SmemDpVectorsL])
-    {
-        static_assert(SmemDpVectorsL >= BlockDpVectorsL, "Row stride must be >= tile width.");
-
-        // Outer thread-tile ldg_vector_t iteration (K-axis)
-        #pragma unroll
-        for (int thread_ldgvec_k = 0; thread_ldgvec_k < ThreadLdgVectorsK; ++thread_ldgvec_k)
-        {
-            int block_ldgvec_k = block_thread_ldgvec_coords.y + (thread_ldgvec_k * StripmineLdgVectorsK);
-
-            // Inner thread-tile ldg_vector_t iteration (L-axis)
-            #pragma unroll
-            for (int thread_ldgvec_l = 0; thread_ldgvec_l < ThreadLdgVectorsL; ++thread_ldgvec_l)
-            {
-                int block_ldgvec_l = block_thread_ldgvec_coords.x + (thread_ldgvec_l * StripmineLdgVectorsL);
-
-                // Write column of dp_vector_t
-                #pragma unroll
-                for (int dpvec = 0; dpvec < LdgVectorDpVectors; ++dpvec)
-                {
-                    scratch_tile[(block_ldgvec_k * LdgVectorDpVectors) + dpvec][block_ldgvec_l] =
-                        thread_tile[thread_ldgvec_k][thread_ldgvec_l].buff[dpvec];
-                }
-            }
-        }
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/block_loader_wmma.h

@@ -1,322 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Tile-loading abstraction for thread blocks
- */
-
-#include "../util/util.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/**
- * block-wide tile loader supporting congruous mapping of data from source and
- * destination addressable storage. Typically, this will be used to load a
- * block-wide tile from global memory into shared memory.
- *
- * This enables the caller to specify MatrixAlignBytes guarantees of the input pointer
- * and performs memory operations on vectors. This increases the efficiency of
- * memory operations and reduces the number of guard predicates needed.
- *
- */
-template <
-    bool congruous,         ///< Indicates whether the "GEMM K" dimension refers to strided matrix dimension
-    int BlockThreads,       ///< Number of threads participating in the streaming operation
-    int BlockItemsL,       ///< Extent of block-wide tile in value_t along the L-axis (width)
-    int BlockItemsK,        ///< Extent of block-wide tile in value_t along the K-axis (height)
-    typename value_t,       ///< Input matrix value type
-    int MatrixAlignBytes,   ///< Byte alignment of input matrix
-    bool AllowRaggedTiles   ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
->
-struct block_loader_wmma
-{
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    /// Predicate bit vector
-    typedef uint64_t predicate_mask_t;
-
-    /// Data movement type, coarsened by MatrixAlignBytes
-    typedef io_vector<
-            value_t,
-            divide_assert<MatrixAlignBytes, sizeof(value_t)>::value,
-            MatrixAlignBytes>
-        ldg_vector_t;
-
-    enum
-    {
-        /// Number of items per ldg_vector_t
-        LdgVectorItems = ldg_vector_t::VectorItems,
-
-        /// Total number of ldg_vector_t within the block-wide tile
-        BlockLdgVectors = divide_assert<(BlockItemsL * BlockItemsK), LdgVectorItems>::value,
-
-        /// Extent of the block-wide tile in ldg_vector_t along K-axis
-        BlockLdgVectorsK = BlockItemsK,
-
-        /// Extent of the block-wide tile in ldg_vector_t along L-axis
-        BlockLdgVectorsL = divide_assert<BlockItemsL, LdgVectorItems>::value,
-
-        /// Number of ldg_vector_t within each thread tile
-        ThreadLdgVectors = divide_assert<BlockLdgVectors, BlockThreads>::value,
-
-        /// Extent of the thread tile in ldg_vector_t along the L-axis
-        ThreadLdgVectorsL = __NV_STD_MAX(1, BlockLdgVectorsL / BlockThreads),
-
-        /// Block-wide strip-mining distance between ldg_vector_t along the K-axis
-        BlockLdgVectorStrideK = __NV_STD_MAX(1, BlockThreads / BlockLdgVectorsL),
-
-        /// Extent of the thread tile in ldg_vector_t along the K-axis
-        ThreadLdgVectorsK = divide_assert<BlockLdgVectorsK, BlockLdgVectorStrideK>::value,
-    };
-
-
-    //-------------------------------------------------------------------------
-    // Assert assumptions
-    //-------------------------------------------------------------------------
-
-    /// Define assertions
-    static_assert(ThreadLdgVectorsL * ThreadLdgVectorsK == ThreadLdgVectors,
-        "Number of vectors must be fully covered by the thread's 2D vector tile.");
-
-    /// Predicate masks must be large enough to guard every vector load
-    static_assert(sizeof(predicate_mask_t) * 8 >= ThreadLdgVectorsL * ThreadLdgVectorsK,
-        "Predicate bit vector must be large enough to guard every vector load.");
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// pointer to tile in global memory
-    const ldg_vector_t *ptr;
-
-    /// stride of the matrix in the K-axis
-    int matrix_values_stride_k;
-
-    /// Guard predicate
-    predicate_mask_t guard;
-
-    /// Guard for the last request iteration
-    predicate_mask_t residue_guard;
-
-    /// Number of 'whole' request iterations before encountering the residue
-    int request_iterations;
-
-    /// fetch registers
-    ldg_vector_t fetch[ThreadLdgVectors];
-
-    /// Thread's base offset from the start of a block-wide tile
-    int thread_offset_l;
-
-    /// Thread's basae offset from the start of a block-wide tile
-    int thread_offset_k;
-
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    block_loader_wmma(
-        const value_t *d_matrix,        ///< Pointer to input matrix
-        int matrix_values_l,            ///< Extent of the input matrix in value_t along the L-axis
-        int start_l,                    ///< Starting location in tile
-        int dim_k,                      ///< Inner dimension of tile, used for computing guard predicates
-        int _matrix_values_stride_k,    ///< Stride of K-axis of atrix
-        int start_k,                    ///< Tile's starting location
-        int2 block_begin_item_coords)  ///< Thread block's starting value_t coordinates (l, k) within the input matrix
-    :
-        ptr(reinterpret_cast<const ldg_vector_t *>(d_matrix)),
-        matrix_values_stride_k(_matrix_values_stride_k / LdgVectorItems),
-        guard(0),
-        residue_guard(0)
-    {
-        // Compute block's starting coordinates in units of vectors
-        int block_base_l = block_begin_item_coords.x / LdgVectorItems;
-        int block_base_k = block_begin_item_coords.y;
-
-        // Compute a thread tiling of the block-wide tile
-        int tid = threadIdx.x;
-        thread_offset_l = tid % BlockLdgVectorsL;
-        thread_offset_k = tid / BlockLdgVectorsL;
-
-        // Add the block and thread offsets to the source pointer
-        ptr += (block_base_l + thread_offset_l) +
-            (block_base_k + thread_offset_k) * matrix_values_stride_k;
-
-        // When AllowRaggedTiles support is enabled, compute a bit vector of guard
-        // predicates
-        if (AllowRaggedTiles)
-        {
-            if (congruous)
-            {
-                request_iterations = (dim_k - start_k) / BlockItemsK;
-            }
-            else
-            {
-                request_iterations = (matrix_values_l - start_l) / BlockItemsL;
-            }
-
-            #pragma unroll
-            for (int k_idx = 0; k_idx < ThreadLdgVectorsK; ++k_idx)
-            {
-                #pragma unroll
-                for (int l_idx = 0; l_idx < ThreadLdgVectorsL; ++l_idx)
-                {
-                    int item = l_idx + k_idx * ThreadLdgVectorsL;
-
-                    // Global vector L and K indices
-                    int vec_l = l_idx * BlockThreads;
-                    int vec_k = k_idx * BlockLdgVectorStrideK;
-
-                    predicate_mask_t pred;
-                    predicate_mask_t residue_pred;
-
-                    if (congruous)
-                    {
-                        pred = (((block_base_l + thread_offset_l + vec_l) * LdgVectorItems < matrix_values_l) ? 1 : 0);
-                        residue_pred = ((block_base_k + thread_offset_k + vec_k < (dim_k % BlockItemsK)) ? 1 : 0);
-                    }
-                    else
-                    {
-                        pred = ((block_base_k + thread_offset_k + vec_k < dim_k) ? 1 : 0);
-                        residue_pred = (((block_base_l + thread_offset_l + vec_l) * LdgVectorItems < (matrix_values_l % BlockItemsL)) ? 1 : 0);
-                    }
-
-                    // Update the guard and residue_guard word with predicate bits
-                    guard |= (pred << item);
-                    residue_guard |= (residue_pred << item);
-                }
-            }
-
-            // If there are zero full request iterations, compute the intersection
-            // with the residue guard.
-            if (!request_iterations)
-            {
-                guard &= residue_guard;
-            }
-        }
-    }
-
-
-
-    /**
-     * Request the current block-wide tile from source memory
-     */
-    inline __device__
-    void request()
-    {
-        #pragma unroll
-        for (int k_idx = 0; k_idx < ThreadLdgVectorsK; ++k_idx)
-        {
-            #pragma unroll
-            for (int l_idx = 0; l_idx < ThreadLdgVectorsL; ++l_idx)
-            {
-                int load_idx = l_idx + (k_idx * ThreadLdgVectorsL);
-                bool pred = !AllowRaggedTiles || (guard & (predicate_mask_t(1) << load_idx));
-                if (pred)
-                {
-                    fetch[load_idx].load(
-                        ptr +
-                        (k_idx * BlockLdgVectorStrideK * matrix_values_stride_k) + (l_idx * BlockThreads));
-                }
-                else
-                {
-                    #pragma unroll
-                    for (int elem_idx = 0; elem_idx < LdgVectorItems; ++elem_idx)
-                    {
-                        fetch[load_idx].buff[elem_idx] = 0;
-                    }
-                }
-            }
-        }
-    }
-
-
-    /// Advance to the next block-wide tile
-    inline __device__
-    void next()
-    {
-        if (congruous)
-        {
-            ptr += BlockItemsK * matrix_values_stride_k;
-        }
-        else
-        {
-            ptr += BlockLdgVectorsL;
-        }
-
-        // Track number of full iterations to intersect with the residue guard predicates.
-        if (AllowRaggedTiles)
-        {
-            --request_iterations;
-            if (!request_iterations)
-            {
-                guard &= residue_guard;
-            }
-        }
-    }
-
-
-    /// Commit the values to the scratch tile to destination memory.
-    template <int SmemStride>
-    inline __device__
-    void commit(value_t *scratch_tile)
-    {
-        static_assert(SmemStride % LdgVectorItems == 0,
-            "SMEM stride must be divisible by the size of vector loads");
-
-        ldg_vector_t *smem_ptr = reinterpret_cast<ldg_vector_t *>(scratch_tile);
-        smem_ptr += thread_offset_l + thread_offset_k * SmemStride / LdgVectorItems;
-
-        #pragma unroll
-        for (int k_idx = 0; k_idx < ThreadLdgVectorsK; ++k_idx)
-        {
-            #pragma unroll
-            for (int l_idx = 0; l_idx < ThreadLdgVectorsL; ++l_idx)
-            {
-                int load_idx = l_idx + (k_idx * ThreadLdgVectorsL);
-
-                fetch[load_idx].store(smem_ptr +
-                    (k_idx * BlockLdgVectorStrideK * SmemStride / LdgVectorItems) +
-                    (l_idx * BlockThreads));
-            }
-        }
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass
-

cutlass/gemm/block_task.h

@@ -1,677 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * A block-wide task abstraction for computing device-wide GEMM
- */
-
-#include <stdint.h>
-
-#include "../util/util.h"
-
-#include "grid_raster.h"
-#include "block_loader.h"
-#include "k_split_control.h"
-#include "thread_accumulator.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * block_task_policy
- ******************************************************************************/
-
-/**
- * \brief Parameterizable tuning policy for \p block_task
- *
- * Once parameterized, \p block_task_policy provides the member constant
- * \p BlockThreads indicating to the required thread block size
- */
-template <
-    int _BlockItemsY,                               ///< Height in rows of a block-wide tile in matrix C
-    int _BlockItemsX,                               ///< Width in columns of a block-wide tile in matrix C
-    int _BlockItemsK,                               ///< Extent of block-wide A|B tiles in value_t along the K-axis
-    int _ThreadItemsY,                              ///< Height in rows of a thread tile in C
-    int _ThreadItemsX,                              ///< Width in columns of a thread tile in C
-    bool _UseDoubleScratchTiles,                    ///< Whether to halve synchronization overhead at the expense of doubled shared memory and addressing overhead
-    grid_raster_strategy::kind_t _RasterStrategy>   ///< Strategy for enumerating \p block_task within an input matrix
-struct block_task_policy
-{
-    enum
-    {
-        /// Height in rows of a block-wide tile in matrix C
-        BlockItemsY = _BlockItemsY,
-
-        /// Width in columns of a block-wide tile in matrix C
-        BlockItemsX = _BlockItemsX,
-
-        /// Height in rows of a thread tile in C
-        ThreadItemsY = _ThreadItemsY,
-
-        /// Width in columns of a thread tile in C
-        ThreadItemsX = _ThreadItemsX,
-
-        /// Extent of block-wide A|B tiles in value_t along the K-axis
-        BlockItemsK = _BlockItemsK,
-
-        /// Whether to halve synchronization overhead at the expense of doubled shared memory and addressing overhead
-        UseDoubleScratchTiles = _UseDoubleScratchTiles,
-
-        /// Number of threads in each thread block (blockDim.x)
-        BlockThreads = divide_assert<
-                            (BlockItemsY * BlockItemsX),
-                            (ThreadItemsY * ThreadItemsX)>::value,
-    };
-
-    /// Strategy for enumerating \p block_task within an input matrix
-    static const grid_raster_strategy::kind_t RasterStrategy = _RasterStrategy;
-};
-
-
-/******************************************************************************
- * block_task
- ******************************************************************************/
-
-/**
- * \brief A block-wide task abstraction for computing device-wide GEMM
- *
- * Each thread_block is assigned a unique tile of output matrix C to compute by
- * consuming the corresponding stripes of the input matrices A and B.
- */
-template <
-    typename                    block_task_policy_t,    ///< Parameterization of block_task_policy
-    typename                    value_t,                ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,                ///< Accumulator value type (matrix C and scalars)
-    matrix_transform_t::kind_t  TransformA,             ///< View transform enumerant for matrix A
-    int                         LdgAlignA,              ///< Alignment (in bytes) for A operand
-    matrix_transform_t::kind_t  TransformB,             ///< View transform enumerant for matrix B
-    int                         LdgAlignB,              ///< Alignment (in bytes) for B operand
-    typename                    epilogue_op_t,          ///< Epilogue operation applied to GEMM
-    int                         LdgAlignC,              ///< Alignment (in bytes) for C operand
-    bool                        AllowRaggedTiles        ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
->
-struct block_task
-{
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    enum
-    {
-        /// Number of threads in each thread block (blockDim.x)
-        BlockThreads = block_task_policy_t::BlockThreads,
-
-        /// Extent of thread tile in value_t along M-axis
-        ThreadItemsY = block_task_policy_t::ThreadItemsY,
-
-        /// Extent of thread tile in value_t along N-axis
-        ThreadItemsX = block_task_policy_t::ThreadItemsX,
-    };
-
-    /// Accumulator type
-    typedef thread_accumulator<
-            ThreadItemsY,
-            ThreadItemsX,
-            value_t,
-            accum_t>
-        thread_accumulator_t;
-
-    /// Dot-product vector type along the K-axis (e.g, uchar4 when using IDP4A)
-    typedef typename thread_accumulator_t::dp_vector_t dp_vector_t;
-
-    enum
-    {
-        /// Whether this is a small, latency-bound tile
-        IsSmallTile = (ThreadItemsY < 4) && (ThreadItemsX < 4),
-
-        /// Number of value_t in dp_vector_t
-        DpVectorItems = divide_assert<sizeof(dp_vector_t), sizeof(value_t)>::value,
-
-        /// Extent of block-wide C-tile in accum_t (and A-tiles in value_t) along M-axis (height)
-        BlockItemsY = block_task_policy_t::BlockItemsY,
-
-        /// Extent of block-wide C-tile in accum_t (and B-tiles in value_t) along N-axis (width)
-        BlockItemsX = block_task_policy_t::BlockItemsX,
-
-        /// Extent of block-wide A|B tiles in value_t along the K-axis
-        BlockItemsK = block_task_policy_t::BlockItemsK,
-
-        /// Whether to halve synchronization overhead at the expense of doubled shared memory and addressing overhead
-        UseDoubleScratchTiles = block_task_policy_t::UseDoubleScratchTiles,
-
-        /// Extent of block-wide A|B tiles in dp_vector_t along the K-axis
-        BlockDpVectorsK = divide_assert<BlockItemsK, DpVectorItems>::value,
-
-        /// Number of dp_vector_t along M-axis that can be read in a single LDS from the shared A-tile (up to 128b if more than one value_t)
-        LdsVectorDpVectorsA = __NV_STD_MIN(
-            ThreadItemsY, 
-            __NV_STD_MAX(1, (128 / (__NV_STD_MAX(sizeof(dp_vector_t), sizeof(accum_t)) * 8)))),
-
-        /// Number of dp_vector_t along N-axis that can be read in a single LDS from the shared B-tile (up to 128b if more than one value_t)
-        LdsVectorDpVectorsB = __NV_STD_MIN(
-            ThreadItemsX, 
-            __NV_STD_MAX(1, (128 / (__NV_STD_MAX(sizeof(dp_vector_t), sizeof(accum_t)) * 8)))),
-
-        /// Number of strip-mined LDS vector reads from shared A-tile
-        ThreadLdsVectorsA = divide_assert<ThreadItemsY, LdsVectorDpVectorsA>::value,
-
-        /// Number of strip-mined LDS vector reads from shared B-tile
-        ThreadLdsVectorsB = divide_assert<ThreadItemsX, LdsVectorDpVectorsB>::value,
-
-        /// Number of elements in one LDG/STG vector of C-tile
-        ThreadLdgVectorSizeC = __NV_STD_MIN(LdgAlignC, 16) / (sizeof(accum_t)),
-
-        /// Number of threads in warp
-        WarpThreads = 32,
-
-        /// Extent of warp in threads along the M-axis
-        WarpThreadsY = (BlockItemsY > BlockItemsX) ? 8 : 4,
-
-        /// Extent of warp in threads along the N-axis
-        WarpThreadsX = divide_assert<WarpThreads, WarpThreadsY>::value,
-
-        /// Extent of warp-wide tile in items along the M-axis
-        WarpItemsY = WarpThreadsY * ThreadItemsY,
-
-        /// Extent of warp-wide tile in items along the N-axis
-        WarpItemsX = WarpThreadsX * ThreadItemsX,
-
-        /// Extent of block in warps along M-axis
-        BlockWarpsY = divide_assert<BlockItemsY, WarpItemsY>::value,
-
-        /// Extent of block in warps along N-axis
-        BlockWarpsX = divide_assert<BlockItemsX, WarpItemsX>::value,
-    };
-
-    /// Load-from-shared data movement type for A-tile, coarsened by LdsVectorDpVectorsA
-    typedef io_vector<dp_vector_t, LdsVectorDpVectorsA> lds_vector_a_t;
-
-    /// Load-from-shared data movement type for B-tile, coarsened by LdsVectorDpVectorsB
-    typedef io_vector<dp_vector_t, LdsVectorDpVectorsB> lds_vector_b_t;
-
-    /// Thread block rasterization helper type
-    typedef grid_raster<
-            BlockItemsY,
-            BlockItemsX,
-            TransformA,
-            TransformB,
-            block_task_policy_t::RasterStrategy>
-        grid_raster_t;
-
-
-    /// Tile loader type for matrix A
-    typedef block_loader<
-            BlockThreads,                                       // BlockThreads
-            BlockDpVectorsK,                                    // BlockDpVectorsK
-            BlockItemsY,                                        // BlockItemsL
-            value_t,                                            // value_t
-            LdgAlignA,                                          // MatrixAlignBytes
-            AllowRaggedTiles,                                   // AllowRaggedTiles
-            dp_vector_t,                                        // dp_vector_t
-            (TransformA == matrix_transform_t::NonTranspose) ?  // LoadAlgorithm
-                load_algorithm::CongruousCopy :
-                load_algorithm::CrosswiseCopy>
-        block_loader_a_t;
-
-
-    /// Tile loader type for matrix B
-    typedef block_loader<
-            BlockThreads,                                       // BlockThreads
-            BlockDpVectorsK,                                    // BlockDpVectorsK
-            BlockItemsX,                                        // BlockItemsL
-            value_t,                                            // value_t
-            LdgAlignB,                                          // MatrixAlignBytes
-            AllowRaggedTiles,                                   // AllowRaggedTiles
-            dp_vector_t,                                        // dp_vector_t
-            (TransformB == matrix_transform_t::NonTranspose) ?  // LoadAlgorithm
-                load_algorithm::CrosswiseCopy :
-                load_algorithm::CongruousCopy>
-        block_loader_b_t;
-
-
-    enum
-    {
-        /// Number of value_t to pad the end of each row of the shared A-tile
-        PadItemsA = (TransformA == matrix_transform_t::NonTranspose) ?
-            __NV_STD_MAX(LdsVectorDpVectorsA, block_loader_a_t::AlignmentDpVectorsL) :
-            LdsVectorDpVectorsA,
-
-        /// Number of value_t to pad the end of each row of the shared B-tile
-        PadItemsB = (TransformB == matrix_transform_t::NonTranspose) ?
-            LdsVectorDpVectorsB :
-            __NV_STD_MAX(LdsVectorDpVectorsB, block_loader_b_t::AlignmentDpVectorsL),
-    };
-
-
-    /// Shared memory layout for a prefetch page
-    struct page_storage_t
-    {
-        /// Tile of A
-        dp_vector_t __align__(16) block_a[BlockDpVectorsK][BlockItemsY + PadItemsA];
-
-        /// Tile of B
-        dp_vector_t __align__(16) block_b[BlockDpVectorsK][BlockItemsX + PadItemsB];
-    };
-
-
-    /// Shared memory layout for scratch storage
-    struct scratch_storage_t
-    {
-        /// Prefetch pages
-        page_storage_t pages[UseDoubleScratchTiles ? 2 : 1];
-
-        /// Accumulator shared scratch
-        typename thread_accumulator_t::scratch_storage_t accum_scratch;
-    };
-
-
-    //-------------------------------------------------------------------------
-    // Assert assumptions
-    //-------------------------------------------------------------------------
-
-    // Ensure we have at least two unrolled innermost loop iterations (one to prefetch
-    // the next global tile and then one to prefetch the first strip of it from shared)
-    static_assert ((BlockDpVectorsK >= 2), "BlockDpVectorsK must be >= 2.");
-
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// Scratch storage reference
-    scratch_storage_t *scratch;
-
-    /// Which page of scratch tiles we're currently reading from
-    int page_idx;
-
-    /// Pointer to matrix C
-    accum_t *d_c;
-
-    /// Epilogue operation applied to update matrix C
-    epilogue_op_t epilogue_op;
-
-    /// Matrix height in rows of trans_op(A) and C
-    int dim_m;
-
-    /// Matrix width in columns of trans_op(B) and C
-    int dim_n;
-
-    /// Control for inter-block k-splitting
-    k_split_control k_split;
-
-    /// Thread block's base value_t coordinates (m, n) in matrix C
-    grid_raster_t grid_raster;
-
-    /// Thread block's current coordinate (k) within A|B matrices
-    int block_item_coords_k;
-
-    /// Thread block's ending coordinate (k) within A|B matrices (one-past)
-    int block_end_item_k;
-
-    /// Warp's coordinates (x, y) in thread block
-    int2 block_warp_coords;
-
-    /// Thread's coordinates (x, y) in warp
-    int2 warp_thread_coords;
-
-    /// Thread's base item offset within strip of A tile
-    int thread_strip_offset_a;
-
-    /// Thread's base item offset within strip of B tile
-    int thread_strip_offset_b;
-
-    /// Thread's active-k/prefetch-k slices from shared A tile
-    lds_vector_a_t local_slices_a[2][ThreadLdsVectorsA];
-
-    /// Thread's active-k/prefetch-k slices from shared B tile
-    lds_vector_b_t local_slices_b[2][ThreadLdsVectorsB];
-
-    /// A tile loader
-    block_loader_a_t loader_a;
-
-    /// B tile loader
-    block_loader_b_t loader_b;
-
-    /// C tile accumulator
-    thread_accumulator_t accumulator;
-
-
-    //-------------------------------------------------------------------------
-    // Coordinate system helpers
-    //-------------------------------------------------------------------------
-
-    /// Compute the warp's coordinates (x, y) in thread block
-    inline __device__
-    int2 warp_coords()
-    {
-        int warp_id = threadIdx.x / WarpThreads;
-        return make_int2(
-            warp_id % BlockWarpsX,
-            warp_id / BlockWarpsX);
-    }
-
-
-    /// Compute the thread's lane-coordinates (x, y) in warp
-    inline __device__
-    int2 thread_coords()
-    {
-        int lane_id = threadIdx.x % WarpThreads;
-
-        // Maxwell+ mapping of threads within a 2D warp for maximal LDS bandwidth
-        return make_int2(
-            lane_id / WarpThreadsY,
-            lane_id % WarpThreadsY);
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    block_task(
-        scratch_storage_t *scratch,
-        value_t *d_a,
-        value_t *d_b,
-        accum_t *d_c,
-        epilogue_op_t epilogue_op,
-        int dim_m,
-        int dim_n,
-        int dim_k,
-        k_split_control k_split)
-    :
-        scratch(scratch),
-        page_idx(0),
-        d_c(d_c),
-        epilogue_op(epilogue_op),
-        dim_m(dim_m),
-        dim_n(dim_n),
-        k_split(k_split),
-        block_item_coords_k(k_split.block_begin_item_k()),
-        block_end_item_k(k_split.block_end_item_k(dim_k)),
-        block_warp_coords(warp_coords()),
-        warp_thread_coords(thread_coords()),
-        thread_strip_offset_a((warp_thread_coords.y * LdsVectorDpVectorsA) + (block_warp_coords.y * WarpItemsY)),
-        thread_strip_offset_b((warp_thread_coords.x * LdsVectorDpVectorsB) + (block_warp_coords.x * WarpItemsX)),
-
-        loader_a(
-            d_a,                                                            // d_matrix
-            dim_m,                                                          // matrix_values_l
-            (TransformA == matrix_transform_t::NonTranspose) ? dim_m : 1,   // matrix_values_stride_k
-            (TransformA == matrix_transform_t::NonTranspose) ? 1 : dim_k,   // matrix_values_stride_l
-            make_int2(                                                      // block_begin_item_coords
-                grid_raster.block_item_coords.y,
-                block_item_coords_k),
-            block_end_item_k),                                              // block_end_item_k
-
-        loader_b(
-            d_b,                                                            // d_matrix
-            dim_n,                                                          // matrix_values_l
-            (TransformB == matrix_transform_t::NonTranspose) ? 1 : dim_n,   // matrix_values_stride_k
-            (TransformB == matrix_transform_t::NonTranspose) ? dim_k : 1,   // matrix_values_stride_l
-            make_int2(                                                      // block_begin_item_coords
-                grid_raster.block_item_coords.x,
-                block_item_coords_k),
-            block_end_item_k),                                              // block_end_item_k
-
-        accumulator(scratch->accum_scratch)
-    {}
-
-
-    //-------------------------------------------------------------------------
-    // Prefetching utility methods
-    //-------------------------------------------------------------------------
-
-    /**
-     * Request the calling thread's slices of the shared tiles at depth \p tile_offset_k
-     */
-    inline __device__ void request_local_prefetch(
-        lds_vector_a_t (&slice_a)[ThreadLdsVectorsA],  ///< Slice from A
-        lds_vector_b_t (&slice_b)[ThreadLdsVectorsB],  ///< Slice from B
-        int tile_offset_k)
-    {
-        // Load B strip
-        for (int i = 0; i < ThreadLdsVectorsB; ++i)
-        {
-            slice_b[i].load(
-                &scratch->pages[page_idx].block_b[tile_offset_k][thread_strip_offset_b + (i * WarpThreadsX * LdsVectorDpVectorsB)]);
-        }
-
-        // Load A strip
-        for (int i = 0; i < ThreadLdsVectorsA; ++i)
-        {
-            slice_a[i].load(
-                &scratch->pages[page_idx].block_a[tile_offset_k][thread_strip_offset_a + (i * WarpThreadsY * LdsVectorDpVectorsA)]);
-        }
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Epilogue
-    //-------------------------------------------------------------------------
-
-    /**
-     * Performs the GEMM epilogue:
-     *   - Applies the scalar multipliers and addends to the accumulators
-     *   - Write the result to the output matrix
-     */
-    __forceinline__ __device__
-    void epilogue()
-    {
-        // Wait for predecessor thread block(s) to produce block-wide tile of
-        // exclsuive partial-sums
-        k_split.wait();
-
-        // Configure epilogue as to whether the thread block is a secondary
-        // accumulator in an inter-block k-splitting scheme
-        if (k_split.is_secondary_accumulator())
-            epilogue_op.set_secondary_accumulator();
-
-        // Whether the addend from C needs loading
-        bool must_init_addend = epilogue_op.must_init_addend();
-
-        #pragma unroll
-        for (int x = 0; x < ThreadItemsX; ++x)
-        {
-            #pragma unroll
-            for (int y = 0; y < ThreadItemsY; y += LdsVectorDpVectorsA)
-            {
-                int thread_strip_b = x / LdsVectorDpVectorsB;
-                int thread_strip_a = y / LdsVectorDpVectorsA;
-
-                int thread_item_coords_tile_x = thread_strip_offset_b + (thread_strip_b * WarpThreadsX * LdsVectorDpVectorsB) + (x % LdsVectorDpVectorsB);
-                int thread_item_coords_tile_y = thread_strip_offset_a + (thread_strip_a * WarpThreadsY * LdsVectorDpVectorsA) + (y % LdsVectorDpVectorsA);
-
-                int c_idx = (grid_raster.block_item_coords.x + thread_item_coords_tile_x) * dim_m +
-                    grid_raster.block_item_coords.y + thread_item_coords_tile_y;
-
-                accum_t *my_c = d_c + c_idx;
-
-                #pragma unroll
-                for (int i = 0; i < LdsVectorDpVectorsA; ++i)
-                {
-                    accum_t c_slice = accum_t(0);
-                    accum_t *c_ptr = my_c + i;
-
-                    if ((grid_raster.block_item_coords.x + thread_item_coords_tile_x) < dim_n &&
-                        (grid_raster.block_item_coords.y + thread_item_coords_tile_y + i) < dim_m)
-                    {
-                        if (must_init_addend)
-                        {
-                            ldg_cg(c_slice, c_ptr);
-                        }
-
-                        c_slice = epilogue_op(accumulator.get(x, y + i), c_slice, c_idx + i);
-
-                        stg_cg(c_ptr, c_slice);
-                    }
-                }
-            }
-        }
-
-        // Signal k-split successor thread_block that we have produced our block-wide
-        // tile of inclusive partial-sums
-        k_split.signal();
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Tile consumption
-    //-------------------------------------------------------------------------
-
-    /**
-     * Consume a tile of A and B each
-     */
-    template <bool DoGlobalPrefetch>
-    __forceinline__ __device__
-    void consume_tile()
-    {
-        // Unroll BlockDpVectorsK iterations of outer-product accumulations
-        #pragma unroll
-        for (int tile_offset_k = 0; tile_offset_k < BlockDpVectorsK; tile_offset_k += 1)
-        {
-            // Last strip commits global prefetch for next tile
-            if ((tile_offset_k == BlockDpVectorsK - 1) && DoGlobalPrefetch)
-            {
-                // If not using two pages of scratch tiles, protect the above prefetch loads from the committing writes below
-                if (!UseDoubleScratchTiles)
-                    __syncthreads();
-
-                // If using two pages of scratch tiles, switch to next page before writing
-                if (UseDoubleScratchTiles)
-                {
-                    page_idx = (page_idx ? 0 : 1);
-                }
-
-                // Commit global prefetch data to scratch page
-                loader_a.commit(scratch->pages[page_idx].block_a);
-                loader_b.commit(scratch->pages[page_idx].block_b);
-
-                __syncthreads();
-            }
-
-            // Request local prefetch for next strip
-            request_local_prefetch(
-                local_slices_a[(tile_offset_k + 1) % 2],
-                local_slices_b[(tile_offset_k + 1) % 2],
-                (tile_offset_k + 1) % BlockDpVectorsK);
-
-            // Request global prefetch for next tile on first strip
-            if ((tile_offset_k == 0) && DoGlobalPrefetch)
-            {
-                loader_b.request();
-                loader_b.next();
-                loader_a.request();
-                loader_a.next();
-            }
-
-            // Cast strip-mined loads to contiguous array of dp_vector_t
-            typedef dp_vector_t thread_tile_a_t[ThreadLdsVectorsA * LdsVectorDpVectorsA];
-            typedef dp_vector_t thread_tile_b_t[ThreadLdsVectorsB * LdsVectorDpVectorsB];
-            thread_tile_a_t &thread_tile_a = reinterpret_cast<thread_tile_a_t&>(local_slices_a[(tile_offset_k) % 2]);
-            thread_tile_b_t &thread_tile_b = reinterpret_cast<thread_tile_b_t&>(local_slices_b[(tile_offset_k) % 2]);
-
-            // Accumulate this dp-stripe product
-            accumulator.multiply_accumulate(thread_tile_a, thread_tile_b);
-        }
-    }
-
-
-    //-------------------------------------------------------------------------
-    // GEMM API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Compute GEMM
-     */
-    __forceinline__ __device__
-    void run()
-    {
-        // Quit if the thread block is fully out-of-bounds
-        if (grid_raster.is_block_oob(dim_m, dim_n))
-        {
-            asm volatile("exit;");
-        }
-
-        // Request global prefetch of first tile
-        loader_a.request();
-        loader_a.next();
-        loader_b.request();
-        loader_b.next();
-
-        // Commit global prefetch of first tile to shared memory
-        loader_a.commit(scratch->pages[page_idx].block_a);
-        loader_b.commit(scratch->pages[page_idx].block_b);
-
-        // Advance to next A,B tiles in K-axis
-        block_item_coords_k += BlockItemsK;
-
-        // Synchronize shared tiles and prepared accumulator
-        __syncthreads();
-
-        // Initialize thread's slice of accumulators
-        accumulator.init();
-
-        // Request first iteration of local prefetch strips
-        request_local_prefetch(
-            local_slices_a[0],
-            local_slices_b[0],
-            0);
-
-        //
-        // Main loop
-        //
-
-        // Consume tiles in A and B along the K-axis (all but last tile)
-        #pragma unroll 1
-        while (block_item_coords_k < block_end_item_k)
-        {
-            consume_tile<true>();
-
-            // Advance to next A,B tiles in K-axis
-            block_item_coords_k += BlockItemsK;
-        }
-
-        // Consume last tile
-        consume_tile<false>();
-
-        //
-        // Eplilogue
-        //
-
-        epilogue();
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/block_task_wmma.h

@@ -1,767 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-/**
- * \file
- * A block-wide task abstraction for computing device-wide GEMM
- */
-
-#pragma once
-
-// Compiler guard conditional to avoid compilation errors on versions of CUDA that
-// do not support the WMMA API.
-#if defined (WMMA)
-
-#include <stdint.h>
-
-#include "../util/util.h"
-
-#include "grid_raster.h"
-#include "block_loader.h"
-#include "block_loader_wmma.h"
-#include "wmma_accumulator.h"
-
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * block_task_wmma_policy
- ******************************************************************************/
-
-/**
- * \brief Parameterizable tuning policy for block-wide WMMA GEMM tasks
- *
- * Once parameterized, \p block_task_policy provides the member constant
- * \p BlockThreads indicating to the required thread block size
- */
-template <
-    int _BlockItemsY,                               ///< Height in rows of a block-wide tile in matrix C
-    int _BlockItemsX,                               ///< Width in columns of a block-wide tile in matrix C
-    int _BlockItemsK,                               ///< Extent of block-wide A|B tiles in value_t along the K-axis
-    int _WarpItemsY,                                ///< Height in rows of a Warp tile's accumulators
-    int _WarpItemsX,                                ///< Width in columns of a Warp tile's accumulators
-    int _WmmaItemsY,                                ///< Height in rows of a discrete WMMA block's accumulators
-    int _WmmaItemsX,                                ///< Width in columns of a discrete WMMA block's accumulators
-    int _WmmaItemsK,                                ///< Depth of each discrete WMMA block
-    bool _UseDoubleScratchTiles,                    ///< Whether to halve synchronization overhead at the expense of doubled shared memory and addressing overhead
-    grid_raster_strategy::kind_t _RasterStrategy>   ///< Strategy for enumerating \p block_task within an input matrix
-struct block_task_wmma_policy
-{
-    /// Strategy for enumerating \p block_task within an input matrix
-    static const grid_raster_strategy::kind_t RasterStrategy = _RasterStrategy;
-
-    enum
-    {
-        /// Height in rows of a block-wide tile in matrix C
-        BlockItemsY = _BlockItemsY,
-
-        /// Width in columns of a block-wide tile in matrix C
-        BlockItemsX = _BlockItemsX,
-
-        /// Extent of block-wide A|B tiles in value_t along the K-axis
-        BlockItemsK = _BlockItemsK,
-
-        /// Height in rows of a Warp tile's accumulators
-        WarpItemsX = _WarpItemsX,
-
-        /// Width in columns of a Warp tile's accumulators
-        WarpItemsY = _WarpItemsY,
-
-        /// Width in columns of a discrete WMMA block's accumulators
-        WmmaItemsX = _WmmaItemsX,
-
-        /// Height in rows of a discrete WMMA block's accumulators
-        WmmaItemsY = _WmmaItemsY,
-
-        /// Depth of each discrete WMMA block
-        WmmaItemsK = _WmmaItemsK,
-
-        /// Whether to halve synchronization overhead at the expense of doubled shared memory and addressing overhead
-        UseDoubleScratchTiles = _UseDoubleScratchTiles,
-
-
-        //
-        // Derived quantities
-        //
-
-        /// Machine warp size
-        WarpThreads = 32,
-
-        /// Number of WMMA operations in the height dimension
-        WmmaBlocksY = divide_assert<WarpItemsY, WmmaItemsY>::value,
-
-        /// Number of WMMA operations in the height dimension
-        WmmaBlocksX = divide_assert<WarpItemsX, WmmaItemsX>::value,
-
-        /// Number of warps in each thread block
-        BlockWarps = divide_assert<BlockItemsY * BlockItemsX, WarpItemsX * WarpItemsY>::value,
-
-        /// Number of threads in each thread block (blockDim.x)
-        BlockThreads = BlockWarps * WarpThreads,
-    };
-};
-
-
-/******************************************************************************
- * block_task_wmma
- ******************************************************************************/
-
-/**
- * \brief A block-wide task abstraction for computing device-wide GEMM
- *
- * Each thread_block is assigned a unique tile of output matrix C to compute by
- * consuming the corresponding stripes of the input matrices A and B.
- */
-template <
-    typename                    block_task_policy_t,    ///< Parameterization of block_task_policy
-    typename                    value_t,                ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,                ///< Accumulator value type (matrix C and scalars)
-    matrix_transform_t::kind_t  TransformA,             ///< View transform enumerant for matrix A
-    int                         LdgAlignA,              ///< Alignment (in bytes) for A operand
-    matrix_transform_t::kind_t  TransformB,             ///< View transform enumerant for matrix B
-    int                         LdgAlignB,              ///< Alignment (in bytes) for B operand
-    typename                    epilogue_op_t,          ///< Epilogue operation to update matrix C
-    int                         LdgAlignC,              ///< Alignment (in bytes) for C operand
-    bool                        AllowRaggedTiles        ///< Whether the input matrix's dimensions need not be an even-multiple of the block-wide tile dimensions
->
-struct block_task_wmma
-{
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    enum
-    {
-        /// Number of threads in each thread block (blockDim.x)
-        BlockThreads = block_task_policy_t::BlockThreads,
-
-        /// Extent of block-wide C-tile in accum_t (and A-tiles in value_t) along M-axis (height)
-        BlockItemsY = block_task_policy_t::BlockItemsY,
-
-        /// Extent of block-wide C-tile in accum_t (and B-tiles in value_t) along N-axis (width)
-        BlockItemsX = block_task_policy_t::BlockItemsX,
-
-        /// Extent of block-wide A|B tiles in value_t along the K-axis
-        BlockItemsK = block_task_policy_t::BlockItemsK,
-
-        /// Extent of warp C-tile in accum_t (and A-tiles in value_t) along M-axis (height)
-        WarpItemsY = block_task_policy_t::WarpItemsY,
-
-        /// Extent of warp C-tile in accum_t (and B-tiles in value_t) along N-axis (width)
-        WarpItemsX = block_task_policy_t::WarpItemsX,
-
-        /// Extent of warp C-tile in accum_t (and A-tiles in value_t) along M-axis (height)
-        WmmaItemsY = block_task_policy_t::WmmaItemsY,
-
-        /// Extent of warp C-tile in accum_t (and B-tiles in value_t) along N-axis (width)
-        WmmaItemsX = block_task_policy_t::WmmaItemsX,
-
-        /// Extent of warp-wide A|B-tiles in value_t along K-axis
-        WmmaItemsK = block_task_policy_t::WmmaItemsK,
-
-        /// Whether to halve synchronization overhead at the expense of doubled shared memory and addressing overhead
-        UseDoubleScratchTiles = block_task_policy_t::UseDoubleScratchTiles,
-
-        /// Number of threads in warp
-        WarpThreads = block_task_policy_t::WarpThreads,
-
-        /// Number of warps participating
-        BlockWarps = block_task_policy_t::BlockWarps,
-
-        /// Extent of block in warps along M-axis
-        BlockWarpsY = divide_assert<BlockItemsY, WarpItemsY>::value,
-
-        /// Extent of block in warps along N-axis
-        BlockWarpsX = divide_assert<BlockItemsX, WarpItemsX>::value,
-
-        /// Number of MMA unrolls
-        WmmaUnrollCount = divide_assert<BlockItemsK,  WmmaItemsK>::value,
-
-        /// True if the A matrix layout is column major (K is the strided dimension)
-        IsLayoutCongruousA = (TransformA == matrix_transform_t::NonTranspose),
-
-        /// True if the B matrix layout is row mayor (K is the strided dimension)
-        IsLayoutCongruousB = (TransformB == matrix_transform_t::Transpose),
-
-    };
-
-    /// WMMA may support unique types for A and B, so plan ahead for this
-    typedef value_t value_a_t;
-
-    /// WMMA may support unique types for A and B, so plan ahead for this
-    typedef value_t value_b_t;
-
-    /// WMMA accumulator type
-    typedef wmma_accumulator<
-            WarpItemsY,
-            WarpItemsX,
-            WmmaItemsY,
-            WmmaItemsX,
-            WmmaItemsK,
-            value_a_t,
-            value_b_t,
-            accum_t,
-            TransformA,
-            TransformB>
-        accumulator_t;
-
-    /// Thread block rasterization helper type
-    typedef grid_raster<
-            BlockItemsY,
-            BlockItemsX,
-            TransformA,
-            TransformB,
-            block_task_policy_t::RasterStrategy>
-        grid_raster_t;
-
-    /// Tile loader type for matrix A
-    typedef block_loader_wmma<
-            IsLayoutCongruousA,
-            BlockThreads,
-            (IsLayoutCongruousA ? BlockItemsY : BlockItemsK),
-            (IsLayoutCongruousA ? BlockItemsK : BlockItemsY),
-            value_a_t,
-            LdgAlignA,
-            AllowRaggedTiles>
-        block_loader_a_t;
-
-    /// Tile loader type for matrix A
-    typedef block_loader_wmma<
-            IsLayoutCongruousB,
-            BlockThreads,
-            (IsLayoutCongruousB ? BlockItemsX : BlockItemsK),
-            (IsLayoutCongruousB ? BlockItemsK : BlockItemsX),
-            value_b_t,
-            LdgAlignB,
-            AllowRaggedTiles>
-        block_loader_b_t;
-
-    /// Type alias for matrix A fragment type
-    typedef typename accumulator_t::fragment_a_t fragment_a_t;
-
-    /// Type alias for matrix B fragment type
-    typedef typename accumulator_t::fragment_b_t fragment_b_t;
-
-    enum
-    {
-        /// Number of fragments from A matrix
-        WmmaBlocksY = accumulator_t::WmmaBlocksY,
-
-        /// Number of fragments from B matrix
-        WmmaBlocksX = accumulator_t::WmmaBlocksX,
-
-        /// Number of value_t to pad the outer dimension of the shared A-tile
-        PadItemsA = 16,
-
-        /// Number of value_t to pad the outer dimension of the shared B-tile
-        PadItemsB = 16,
-
-        /// Leading dimension of A matrix tile
-        LdmSmemA = (IsLayoutCongruousA ? BlockItemsY: BlockItemsK) + PadItemsA,
-
-        /// Leading dimension of A matrix tile
-        StridedSmemA = (IsLayoutCongruousA ? BlockItemsK : BlockItemsY ),
-
-        /// Leading dimension of B matrix tile
-        LdmSmemB = (IsLayoutCongruousB? BlockItemsX : BlockItemsK)  + PadItemsB,
-
-        StridedSmemB = (IsLayoutCongruousB ? BlockItemsK : BlockItemsX),
-    };
-
-    /// Shared memory layout for a prefetch page
-    struct page_storage_t
-    {
-        /// Tile of A
-        value_a_t __align__(16) block_a[StridedSmemA][LdmSmemA];
-
-        /// Tile of B
-        value_b_t __align__(16) block_b[StridedSmemB][LdmSmemB];
-    };
-
-    /// Shared memory layout for scratch storage
-    struct scratch_storage_t
-    {
-        union
-        {
-            /// Prefetch pages
-            uninitialized<page_storage_t> pages[UseDoubleScratchTiles ? 2 : 1];
-
-            /// Scratch storage for warps
-            accum_t epilogue[BlockWarps][WmmaItemsX * WmmaItemsY];
-        };
-    };
-
-    //-------------------------------------------------------------------------
-    // Assert assumptions
-    //-------------------------------------------------------------------------
-
-    // Ensure we have at least two unrolled innermost loop iterations (one to prefetch
-    // the next global tile and then one to prefetch the first strip of it from shared)
-    static_assert ((BlockItemsK >= 2), "BlockItemsK must be >= 2.");
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// Scratch storage reference
-    scratch_storage_t *scratch;
-
-    /// Which page of scratch tiles we're currently reading from
-    int page_idx;
-
-    /// Pointer to matrix C
-    accum_t *d_c;
-
-    /// Epilogue operation applied to update matrix C
-    epilogue_op_t epilogue_op;
-
-    /// Matrix height in rows of trans_op(A) and C
-    int dim_m;
-
-    /// Matrix width in columns of trans_op(B) and C
-    int dim_n;
-
-    /// Control for inter-block k-splitting
-    k_split_control k_split;
-
-    /// Thread block's base value_t coordinates (m, n) in matrix C
-    grid_raster_t grid_raster;
-
-    /// Thread block's current coordinate (k) within A|B matrices
-    int block_item_coords_k;
-
-    /// Thread block's ending coordinate (k) within A|B matrices (one-past)
-    int block_end_item_k;
-
-    /// Warp's coordinates (x, y) in thread block
-    int2 block_warp_item_coords;
-
-    /// A tile loader
-    block_loader_a_t loader_a;
-
-    /// B tile loader
-    block_loader_b_t loader_b;
-
-    /// Thread's active-k/prefetch-k slices from shared A tile
-    fragment_a_t local_slices_a[2][WmmaBlocksY];
-
-    /// Thread's active-k/prefetch-k slices from shared B tile
-    fragment_b_t local_slices_b[2][WmmaBlocksX];
-
-    /// Accumulator tile
-    accumulator_t accumulator;
-
-
-    //-------------------------------------------------------------------------
-    // Coordinate system helpers
-    //-------------------------------------------------------------------------
-
-    /// Compute the warp's item-coordinates (x, y) in thread block
-    inline __device__
-    int2 warp_item_coords()
-    {
-        int warp_id = threadIdx.x / WarpThreads;
-
-        return make_int2(
-            (warp_id / BlockWarpsY) * WarpItemsX,
-            (warp_id % BlockWarpsY) * WarpItemsY);
-    }
-
-    /// Compute the thread block's base item-coordinates in matrix A
-    inline __device__
-    int2 a_block_item_coords()
-    {
-        if (TransformA == matrix_transform_t::NonTranspose)
-        {
-            return make_int2(grid_raster.block_item_coords.y, block_item_coords_k);
-        }
-        else
-        {
-            return make_int2(block_item_coords_k, grid_raster.block_item_coords.y);
-        }
-    }
-
-    /// Compute the thread block's base item-coordinates in matrix B
-    inline __device__
-    int2 b_block_item_coords()
-    {
-        if (TransformB == matrix_transform_t::Transpose)
-        {
-            return make_int2(grid_raster.block_item_coords.x, block_item_coords_k);
-        }
-        else
-        {
-            return make_int2(block_item_coords_k, grid_raster.block_item_coords.x);
-        }
-    }
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    block_task_wmma(
-        scratch_storage_t *scratch,
-        value_t *d_a,
-        value_t *d_b,
-        accum_t *d_c,
-        epilogue_op_t epilogue_op,
-        int dim_m,
-        int dim_n,
-        int dim_k,
-        k_split_control k_split)
-    :
-        scratch(scratch),
-        page_idx(0),
-        d_c(d_c),
-        epilogue_op(epilogue_op),
-        dim_m(dim_m),
-        dim_n(dim_n),
-        k_split(k_split),
-        block_item_coords_k(k_split.block_begin_item_k()),
-        block_end_item_k(k_split.block_end_item_k(dim_k)),
-        block_warp_item_coords(warp_item_coords()),
-
-        loader_a(
-            reinterpret_cast<value_a_t const *>(d_a),
-            (IsLayoutCongruousA ? dim_m : block_end_item_k),
-            (IsLayoutCongruousA ? 0 : block_item_coords_k),
-            (IsLayoutCongruousA ? block_end_item_k : dim_m),
-            (IsLayoutCongruousA ? dim_m : dim_k),
-            (IsLayoutCongruousA ? block_item_coords_k : 0),
-            a_block_item_coords()),
-
-        loader_b(
-            reinterpret_cast<value_b_t const *>(d_b),
-            (IsLayoutCongruousB ? dim_n : block_end_item_k),
-            (IsLayoutCongruousB ? 0 : block_item_coords_k),
-            (IsLayoutCongruousB ? block_end_item_k : dim_n),
-            (IsLayoutCongruousB ? dim_n : dim_k),
-            (IsLayoutCongruousB ? block_item_coords_k : 0),
-            b_block_item_coords())
-    {}
-
-
-    //-------------------------------------------------------------------------
-    // Prefetching utility methods
-    //-------------------------------------------------------------------------
-
-    /**
-     * Request the calling thread's slices of the shared tiles at depth \p tile_offset_k
-     */
-    inline __device__ void request_local_prefetch(
-        fragment_a_t local_slices_a[WmmaBlocksY],  ///< Slice from A
-        fragment_b_t local_slices_b[WmmaBlocksX],  ///< Slice from B
-        int tile_offset_k)
-    {
-        value_b_t const *smem_A_base = &scratch->pages[page_idx].alias().block_a[0][0];
-        value_b_t const *smem_B_base = &scratch->pages[page_idx].alias().block_b[0][0];
-
-        int constexpr kstride_a = (IsLayoutCongruousA ? LdmSmemA : 1);
-        int constexpr lstride_a = (IsLayoutCongruousA ? 1 : LdmSmemA);
-
-        int constexpr kstride_b = (IsLayoutCongruousB ? LdmSmemB : 1);
-        int constexpr lstride_b = (IsLayoutCongruousB ? 1 : LdmSmemB);
-
-        // Load B strip
-        #pragma unroll
-        for (int i = 0; i < WmmaBlocksX; ++i)
-        {
-            value_b_t const *smem_B_ptr =
-                &smem_B_base[tile_offset_k * kstride_b + (block_warp_item_coords.x + WmmaItemsX * i) * lstride_b];
-
-            nvcuda::wmma::load_matrix_sync(local_slices_b[i], smem_B_ptr, LdmSmemB);
-        }
-
-        // Load A strip
-        #pragma unroll
-        for (int i = 0; i < WmmaBlocksY; ++i)
-        {
-            value_a_t const *smem_A_ptr =
-                &smem_A_base[tile_offset_k * kstride_a + (block_warp_item_coords.y + WmmaItemsY * i) * lstride_a];
-
-            nvcuda::wmma::load_matrix_sync(local_slices_a[i], smem_A_ptr, LdmSmemA);
-        }
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Epilogue
-    //-------------------------------------------------------------------------
-
-    /**
-     * Performs the GEMM epilogue:
-     *   - Applies the scalar multipliers and addends to the accumulators
-     *   - Write the result to the output matrix
-     */
-    inline __device__ void epilogue()
-    {
-        // Wait for predecessor thread block(s) to produce partial-sums
-        k_split.wait();
-
-        // Configure epilogue as to whether the thread block is a secondary
-        // accumulator in an inter-block k-splitting scheme
-        if (k_split.is_secondary_accumulator())
-            epilogue_op.set_secondary_accumulator();
-
-        // Whether or not the addend from C needs loading
-        bool must_init_addend = epilogue_op.must_init_addend();
-
-        int warp_base_x = grid_raster.block_item_coords.x + block_warp_item_coords.x;
-        int warp_base_y = grid_raster.block_item_coords.y + block_warp_item_coords.y;
-
-        int constexpr SmemStride = WmmaItemsY;
-
-        int warp_id = threadIdx.x / 32;
-
-        // Compute shape of one accumulator read/modify/write operation
-        int constexpr ItemsY = (WmmaItemsY);
-        int constexpr ItemsX = (32 / ItemsY);
-        int constexpr IterationsX = WmmaItemsX / ItemsX;
-
-        // Compute a rasterization of warp lanes across the WMMA tile.
-        int lane_id = (threadIdx.x % 32);
-        int lane_read_x = (lane_id / ItemsY);
-        int lane_read_y = (lane_id % ItemsY);
-
-        accum_t *smem_scratch = scratch->epilogue[warp_id];
-        accum_t const *smem_read_ptr = smem_scratch + lane_read_y + lane_read_x * SmemStride;
-
-        #pragma unroll
-        for (int xb = 0; xb < WmmaBlocksX; ++xb)
-        {
-            #pragma unroll
-            for (int yb = 0; yb < WmmaBlocksY; ++yb)
-            {
-                // Store accumulator tile to SMEM
-                nvcuda::wmma::store_matrix_sync(
-                    smem_scratch,
-                    accumulator.accumulators[xb][yb],
-                    SmemStride,
-                    matrix_layout<matrix_transform_t::NonTranspose>::kind);
-
-                // Synchronize threads within the warp
-                __syncthreads();
-
-                // Compute lane coordinates so that each thread efficiently accesses SMEM.
-                int c_x = (warp_base_x + (xb) * WmmaItemsX + lane_read_x);
-                int c_y = (warp_base_y + (yb) * WmmaItemsY + lane_read_y);
-
-                // Compute guard predicate by comparing against problem dimensions.
-                bool pred = c_y < dim_m;
-
-                // Compute output pointer from lane coordinates
-                int c_index = c_x * dim_m + c_y;
-                accum_t *c_ptr = reinterpret_cast<accum_t *>(d_c) + c_x * dim_m + c_y;
-
-                // Iterate over columns of output tile. Load from SMEM, compute epilogue operation,
-                // and stream output to global memory
-                #pragma unroll
-                for (int item_x = 0; item_x < IterationsX; ++item_x)
-                {
-                    accum_t accum = smem_read_ptr[item_x * ItemsX * SmemStride];
-                    accum_t c_element = 0;
-
-                    // Filter against problem dimensions as the warp iterates across the columns of
-                    // output.
-                    pred = (pred && ((c_x + item_x * ItemsX) < dim_n));
-
-                    if (must_init_addend && pred)
-                    {
-                        // NB: inline PTX to utilize strong operations for inter-block synchronization.
-                        //     The following is equivalent to:
-                        //
-                        //         c_element = c_ptr[0];
-                        asm volatile ("ld.global.cg.f32 %0, [%1];\n" : "=f"(c_element) : "l"(c_ptr));
-                    }
-
-                    c_element = epilogue_op(accum, c_element, c_index);
-
-                    if (pred)
-                    {
-                        // NB: inline PTX to utilize strong operations for inter-block synchronization.
-                        //     The following is equivalent to:
-                        //
-                        //         c_ptr[0] = c_element;
-
-                        asm volatile ("st.global.cg.f32 [%0], %1;\n" : : "l"(c_ptr), "f"(c_element));
-                    }
-
-                    // Increment output pointer
-                    c_ptr += dim_m * ItemsX;
-                    c_index += dim_m * ItemsX;
-                }
-                __syncthreads();
-            }
-        }
-
-        // Signal k-split successor thread_block
-        k_split.signal();
-    }
-
-    //-------------------------------------------------------------------------
-    // Tile consumption
-    //-------------------------------------------------------------------------
-
-    /**
-     * Consume a tile of A and B each
-     */
-    template <bool DoGlobalPrefetch>
-    inline __device__
-    void consume_tile()
-    {
-        // Request global prefetch for next tile on first strip
-        if (DoGlobalPrefetch)
-        {
-            loader_b.request();
-            loader_b.next();
-            loader_a.request();
-            loader_a.next();
-        }
-
-        // Unroll BlockDpVectorsK iterations of outer-product accumulations
-        #pragma unroll
-        for (int iteration = 0; iteration < WmmaUnrollCount; ++iteration)
-        {
-            int tile_offset_k = iteration * WmmaItemsK;
-
-            // Active load-from-shared index
-            int active_lds_idx = __NV_STD_MIN(WmmaUnrollCount - 1, (iteration) % 2);
-
-            // Next load-from-shared index
-            int next_lds_idx = __NV_STD_MIN(WmmaUnrollCount - 1, (iteration + 1) % 2);
-
-            // The last unrolled iteration commits the global fetches
-            if ((iteration == WmmaUnrollCount - 1) && DoGlobalPrefetch)
-            {
-                // If not using two pages of scratch tiles, protect the above prefetch loads from
-                // the committing writes below
-                if (!UseDoubleScratchTiles)
-                {
-                    __syncthreads();
-                }
-                else
-                {
-                    page_idx = (page_idx ? 0 : 1);
-                }
-
-                // Commit global prefetch data to scratch page
-                loader_a.template commit<LdmSmemA>(&scratch->pages[page_idx].alias().block_a[0][0]);
-                loader_b.template commit<LdmSmemB>(&scratch->pages[page_idx].alias().block_b[0][0]);
-
-                __syncthreads();
-            }
-
-            // Accumulate this dp-stripe product
-            accumulator.multiply_accumulate(
-                local_slices_a[active_lds_idx],
-                local_slices_b[active_lds_idx]);
-
-            // Request local prefetch for next strip
-            request_local_prefetch(
-                local_slices_a[next_lds_idx],
-                local_slices_b[next_lds_idx],
-                (tile_offset_k + WmmaItemsK) % BlockItemsK);
-        }
-    }
-
-    //-------------------------------------------------------------------------
-    // GEMM API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Compute GEMM
-     */
-    inline __device__
-    void run()
-    {
-        // Quit if the thread block is fully out-of-bounds
-        if (grid_raster.is_block_oob(dim_m, dim_n))
-        {
-            asm volatile("exit;");
-        }
-
-        // Request global prefetch of first tile
-        loader_a.request();
-        loader_a.next();
-        loader_b.request();
-        loader_b.next();
-
-        // Commit global prefetch of first tile to shared memory
-        loader_a.template commit<LdmSmemA>(&scratch->pages[page_idx].alias().block_a[0][0]);
-        loader_b.template commit<LdmSmemB>(&scratch->pages[page_idx].alias().block_b[0][0]);
-
-        // Advance to next A,B tiles in K-axis
-        block_item_coords_k += BlockItemsK;
-
-        // Synchronize shared tiles and prepared accumulator
-        __syncthreads();
-
-        // Initialize thread's slice of accumulators
-        accumulator.init();
-
-        // Request first iteration of local prefetch strips
-        request_local_prefetch(
-            local_slices_a[0],
-            local_slices_b[0],
-            0);
-
-        //
-        // Main loop
-        //
-
-        // Consume tiles in A and B along the K-axis (all but last tile)
-        #pragma unroll 1
-        while (block_item_coords_k < block_end_item_k)
-        {
-            consume_tile<true>();
-
-            // Advance to next A,B tiles in K-axis
-            block_item_coords_k += BlockItemsK;
-        }
-
-        consume_tile<false>();
-
-        //
-        // Eplilogue
-        //
-
-        // prevent overwriting SMEM until all warps have finished loading data
-        __syncthreads();
-
-        // store accumulator tile to global memory
-        epilogue();
-    }
-};
-
-} // namespace gemm
-} // namespace cutlass
-
-#endif

cutlass/gemm/clear_accumulators.h

@@ -0,0 +1,58 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines abstractions for efficiently clearing accumulator tiles.
+*/
+#pragma once
+
+#include "cutlass/vector.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kLanes_ = 1>
+struct ClearAccumulators {
+  /// The shared storage.
+  struct SharedStorage {};
+
+  /// Ctor.
+  CUTLASS_DEVICE ClearAccumulators(SharedStorage& shared_storage) {}
+
+  /// Ctor.
+  CUTLASS_DEVICE ClearAccumulators() {}
+
+  /// Clear the fragment.
+  template <typename Fragment_>
+  CUTLASS_DEVICE void clear(Fragment_& fragment) {
+    fragment.clear();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/dgemm_traits.h

@@ -0,0 +1,134 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines structural traits of double-precision GEMM.
+*/
+#pragma once
+
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/gemm/gemm_epilogue.h"
+#include "cutlass/gemm/gemm_epilogue_traits.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/gemm/gemm_shared_tile.h"
+#include "cutlass/gemm/gemm_traits.h"
+#include "cutlass/gemm/thread_multiply_add.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The tile size for threadblock-level GEMM (K-by-N-by-M).
+    typename OutputTile_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_,
+    /// The number of scalars per LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of scalars per LDG for B.
+    int kScalarsPerLdgB_ = 1>
+struct DgemmConfig
+    : public GemmConfig<
+          /// The scalar type for A.
+          double,
+          /// The scalar type for B.
+          double,
+          /// The scalar type for C.
+          double,
+          /// The scalar type for D.
+          double,
+          /// The tile size for the GEMM KxNxM.
+          OutputTile_,
+          /// The functor to do the math in the main loop.
+          ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, double, double, double>,
+          /// The number of scalars per LDG for A.
+          kScalarsPerLdgA_,
+          /// The number of scalars per STS for A.
+          kScalarsPerLdgA_,
+          /// The number of scalars per LDS for A.
+          2,
+          /// The number of scalars per LDG for B.
+          kScalarsPerLdgB_,
+          /// The number of scalars per STS for B.
+          kScalarsPerLdgB_,
+          /// The number of scalars per LDS for B.
+          2,
+          /// The number of scalars per LDG for C and STG for D.
+          1,
+          /// The number of scalars per STS for D.
+          2,
+          /// The number of scalars per LDS for D.
+          1,
+          /// The number of stages in shared memory.
+          2,
+          /// kResidueSeparate
+          false,
+          /// kResidueInPrologue
+          false,
+          /// kLaunchBounds
+          false
+          >{};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The tile size for threadblock-level GEMM (K-by-N-by-M)
+    typename OutputTile_ = Shape<8, 64, 128>,
+    /// The functor to use in the epilogue.
+    typename EpilogueFunctor_ = LinearScaling<double>,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<8, 8, 8>,
+    /// The number of doubles loaded in one LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of doubles loaded in one LDG for B.
+    int kScalarsPerLdgB_ = 1,
+    /// The index.
+    typename Index_ = int,
+    /// The DGEMM config.
+    typename GemmConfig_ =
+        DgemmConfig<OutputTile_, ThreadGemmShape_, kScalarsPerLdgA_, kScalarsPerLdgB_>,
+    /// The traits class for the epilogue.
+    typename GemmEpilogueTraits_ =
+        SimplifiedGemmEpilogueTraits<GemmConfig_, EpilogueFunctor_, Index_> >
+struct DgemmTraits : public SimplifiedGemmTraits<
+                         // The layout for A.
+                         kLayoutA_,
+                         // The layout for B.
+                         kLayoutB_,
+                         // The config.
+                         GemmConfig_,
+                         // The epilogue.
+                         GemmEpilogue<GemmEpilogueTraits_>,
+                         // The index.
+                         Index_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/dispatch.h

@@ -1,542 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * GEMM kernel entrypoint and dispatch stub
- */
-
-#include <stdint.h>
-
-#include "../util/util.h"
-#include "block_task.h"
-#include "block_task_wmma.h"
-#include "grid_raster.h"
-#include "dispatch_policies.h"
-#include "k_split_control.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * param_pack
- ******************************************************************************/
-
-/**
- * Parameter-pack structure
- *
- * Kernel launch latency is reduced when kernel arguments are wrapped into
- * a single parameter
- */
-template <
-    typename value_t,
-    typename accum_t,
-    typename epilogue_op_t>
-struct param_pack
-{
-    int m;                      ///< Height in rows of op(A) and C
-    int n;                      ///< Width in columns of op(B) and C
-    int k;                      ///< Width in columns of op(A) and height in rows of op(B)
-    k_split_control k_split;    ///< Abstraction for controlling inter-block k-splitting
-    value_t *d_a;               ///< Pointer to matrix A array values
-    value_t *d_b;               ///< Pointer to matrix B array values
-    accum_t *d_c;               ///< Pointer to matrix C array values
-    epilogue_op_t epilogue_op;
-
-    param_pack(
-        int m,                      ///< Height in rows of op(A) and C
-        int n,                      ///< Width in columns of op(B) and C
-        int k,                      ///< Width in columns of op(A) and height in rows of op(B)
-        k_split_control k_split,    ///< Abstraction for controlling inter-block k-splitting
-        epilogue_op_t op,           ///< Epilogue operation to update matrix C
-        value_t *d_a,               ///< Pointer to matrix A array values
-        value_t *d_b,               ///< Pointer to matrix B array values
-        accum_t *d_c)               ///< Pointer to matrix C array values
-    :
-        m(m),
-        n(n),
-        k(k),
-        k_split(k_split),
-        epilogue_op(op),
-        d_a(d_a),
-        d_b(d_b),
-        d_c(d_c)
-    {}
-
-};
-
-
-/******************************************************************************
- * Conditionally select the appropriate GEMM threadblock task
- ******************************************************************************/
-
-/// Conditional selection for block task
-template <
-    math_operation_class_t      math_op,            ///<
-    typename                    block_task_policy_t,  ///< Parameterization of block_task_policy
-    typename                    value_t,            ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,            ///< Accumulator value type (matrix C and scalars)
-    matrix_transform_t::kind_t  TransformA,         ///< View transform enumerant for matrix A
-    int                         LdgAlignA,          ///< Alignment (in bytes) for A operand
-    matrix_transform_t::kind_t  TransformB,         ///< View transform enumerant for matrix B
-    int                         LdgAlignB,          ///< Alignment (in bytes) for B operand
-    typename                    epilogue_op_t,      ///< Epilogue operation applied to GEMM
-    int                         LdgAlignC,          ///< Alignment (in bytes) for C operand
-    bool                        AllowRaggedTiles    ///< Whether GEMM supports matrix sizes other than multiple of BlockItems{XY}
->
-struct gemm_block_task;
-
-/// Scalar math operations
-template <
-    typename                    block_task_policy_t,  ///< Parameterization of block_task_policy
-    typename                    value_t,            ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,            ///< Accumulator value type (matrix C and scalars)
-    matrix_transform_t::kind_t  TransformA,         ///< View transform enumerant for matrix A
-    int                         LdgAlignA,          ///< Alignment (in bytes) for A operand
-    matrix_transform_t::kind_t  TransformB,         ///< View transform enumerant for matrix B
-    int                         LdgAlignB,          ///< Alignment (in bytes) for B operand
-    typename                    epilogue_op_t,      ///< Epilogue operation applied to GEMM
-    int                         LdgAlignC,          ///< Alignment (in bytes) for C operand
-    bool                        AllowRaggedTiles    ///< Whether GEMM supports matrix sizes other than multiple of BlockItems{XY}
->
-struct gemm_block_task<
-    math_operation_class_t::scalar,
-    block_task_policy_t,
-    value_t,
-    accum_t,
-    TransformA,
-    LdgAlignA,
-    TransformB,
-    LdgAlignB,
-    epilogue_op_t,
-    LdgAlignC,
-    AllowRaggedTiles
->
-{
-    // Parameterize task type
-    typedef block_task<
-            block_task_policy_t,
-            value_t,
-            accum_t,
-            TransformA,
-            LdgAlignA,
-            TransformB,
-            LdgAlignB,
-            epilogue_op_t,
-            LdgAlignC,
-            AllowRaggedTiles> type;
-};
-
-/// Matrix math operations
-template <
-    typename                    block_task_policy_t,  ///< Parameterization of block_task_policy
-    typename                    value_t,            ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,            ///< Accumulator value type (matrix C and scalars)
-    matrix_transform_t::kind_t  TransformA,         ///< View transform enumerant for matrix A
-    int                         LdgAlignA,          ///< Alignment (in bytes) for A operand
-    matrix_transform_t::kind_t  TransformB,         ///< View transform enumerant for matrix B
-    int                         LdgAlignB,          ///< Alignment (in bytes) for B operand
-    typename                    epilogue_op_t,      ///< Epilogue operation applied to GEMM
-    int                         LdgAlignC,          ///< Alignment (in bytes) for C operand
-    bool                        AllowRaggedTiles    ///< Whether GEMM supports matrix sizes other than multiple of BlockItems{XY}
->
-struct gemm_block_task<
-    math_operation_class_t::matrix,
-    block_task_policy_t,
-    value_t,
-    accum_t,
-    TransformA,
-    LdgAlignA,
-    TransformB,
-    LdgAlignB,
-    epilogue_op_t,
-    LdgAlignC,
-    AllowRaggedTiles>
-{
-
-#if defined(WMMA)   // conditional compilation with WMMA headers
-
-    // Parameterize task type
-    typedef block_task_wmma<
-            block_task_policy_t,
-            value_t,
-            accum_t,
-            TransformA,
-            LdgAlignA,
-            TransformB,
-            LdgAlignB,
-            epilogue_op_t,
-            LdgAlignC,
-            AllowRaggedTiles> type;
-
-#endif
-};
-
-/******************************************************************************
- * GEMM kernel entrypoint
- ******************************************************************************/
-
-/**
- * GEMM kernel
- *
- * NB: Not sure why NVVM is doing stuff with "__launch_bounds__" instead of just
- * passing it along to PTXAS, but it is currently resulting in less optimal codegen
- */
-template <
-    math_operation_class_t      math_op,            ///< Indicates which class of math operation to select
-    typename                    block_task_policy_t,  ///< Parameterization of block_task_policy
-    matrix_transform_t::kind_t  TransformA,         ///< Transformation op for matrix A
-    int                         LdgAlignA,          ///< Alignment of A matrix elements in bytes
-    matrix_transform_t::kind_t  TransformB,         ///< Transformation op for matrix B
-    int                         LdgAlignB,          ///< Alignment of B matrix elements in bytes
-    typename                    value_t,            ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,            ///< Accumulator value type (matrix C and scalars)
-    typename                    epilogue_op_t,      ///< Epilogue operation applied to update matrix C
-    int                         LdgAlignC,          ///< Alignment of C elements in bytes
-    bool                        AllowRaggedTiles>   ///< Boolean to indicate whether AllowRaggedTiles handling is enabled
-__global__ void kernel(param_pack<value_t, accum_t, epilogue_op_t> pack)
-{
-    // Parameterize task type
-    typedef typename gemm_block_task<
-        math_op,
-        block_task_policy_t,
-        value_t,
-        accum_t,
-        TransformA,
-        LdgAlignA,
-        TransformB,
-        LdgAlignB,
-        epilogue_op_t,
-        LdgAlignC,
-        AllowRaggedTiles>::type block_task_t;
-
-    // Declare statically-allocated shared storage
-    __shared__ typename block_task_t::scratch_storage_t smem;
-
-    // Construct and run the task
-    block_task_t(
-        &smem,
-        pack.d_a,
-        pack.d_b,
-        pack.d_c,
-        pack.epilogue_op,
-        pack.m,
-        pack.n,
-        pack.k,
-        pack.k_split).run();
-}
-
-
-/******************************************************************************
- * Launch configuration description returned to the caller
- ******************************************************************************/
-
-/// Return details about the launch configuration to the caller
-struct launch_configuration
-{
-    //
-    // Data members
-    //
-
-    /// cudaError_t resulting from grid launch
-    cudaError_t result;
-
-    /// Extent of a thread block's partition along the GEMM K-axis
-    int split_k;
-
-    /// Kernel grid extents in thread blocks
-    dim3 grid;
-
-    /// Thread block extents in threads
-    dim3 block;
-
-    //
-    // Methods
-    //
-
-    /// Constructor
-    launch_configuration():
-        result(cudaSuccess),
-        split_k(0),
-        grid(0, 0, 0),
-        block(0, 0, 0) {
-
-    }
-
-    /// Conversion from cudaError_t
-    launch_configuration(cudaError_t result):
-        result(result),
-        split_k(1),
-        grid(0, 0, 0),
-        block(0, 0, 0) {
-
-    }
-
-    /// Launch configuration for Cutlass kernels
-    launch_configuration(
-        cudaError_t result,
-        int split_k,
-        dim3 grid,
-        dim3 block
-    ):
-        result(result),
-        split_k(split_k),
-        grid(grid),
-        block(block) {
-
-    }
-};
-
-
-/******************************************************************************
- * Dispatch stub
- ******************************************************************************/
-
-/**
- * GEMM dispatch stub
- *
- * This function also serves as the autotuning entrypoint to evaluate different
- * tuning parameterizations of kernel.
- */
-template <
-    math_operation_class_t      math_op,            ///< Indicates which class of math operation to select
-    typename                    block_task_policy_t,  ///< Parameterization of block_task_policy
-    matrix_transform_t::kind_t  TransformA,         ///< Transformation op for matrix A
-    int                         LdgAlignA,          ///< Alignment of A matrix elements in bytes
-    matrix_transform_t::kind_t  TransformB,         ///< Transformation op for matrix B
-    int                         LdgAlignB,          ///< Alignment of B matrix elements in bytes
-    typename                    value_t,            ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,            ///< Accumulator value type (matrix C and scalars)
-    typename                    epilogue_op_t,      ///< Epilogue operation
-    int                         LdgAlignC,          ///< Alignment of C matrix elements in bytes
-    bool                        AllowRaggedTiles,   ///< Boolean to indicate whether AllowRaggedTiles handling is enabled
-    typename                    kernel_ptr_t>       ///< GEMM kernel function pointer type
-launch_configuration dispatch(
-    kernel_ptr_t    kernel_ptr,                     ///< GEMM kernel function pointer
-    int             m,                              ///< Height in rows of op(A) and C
-    int             n,                              ///< Width in columns of op(B) and C
-    int             k,                              ///< Width in columns of op(A) and height in rows of op(B)
-    epilogue_op_t   epilogue_op,                    ///< Epilogue operation to update matrix C
-    value_t         *d_a,                           ///< Device pointer to matrix A array values
-    value_t         *d_b,                           ///< Device pointer to matrix B array values
-    accum_t         *d_c,                           ///< Device pointer to matrix C array values
-    cudaStream_t    stream = 0,                     ///< CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
-    bool            debug_synchronous = true)       ///< Whether or not to synchronize the stream after every kernel launch
-                                                    ///  to check for errors.  Also causes launch configurations to be printed
-                                                    ///  to the console if DEBUG is defined.  Default is \p false.
-{
-    // Thread block rasterization type
-    typedef grid_raster<
-            block_task_policy_t::BlockItemsY,
-            block_task_policy_t::BlockItemsX,
-            TransformA,
-            TransformB,
-            block_task_policy_t::RasterStrategy>
-        grid_raster_t;
-
-    launch_configuration config;
-
-    // Compute block dims
-    config.block = dim3(block_task_policy_t::BlockThreads);
-
-    // Compute shared memory
-    int dynamic_smem_bytes = 0;
-
-    // Compute occupancy
-    int max_sm_occupancy;
-    if (CUDA_PERROR_DEBUG(config.result = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
-        &max_sm_occupancy,
-        kernel_ptr,
-        config.block.x * config.block.y,
-        dynamic_smem_bytes)))
-    {
-        return config;
-    }
-
-    // Compute grid extents
-    config.grid = grid_raster_t::grid_dims(m, n);
-
-    // Get SM count
-    int sm_count;
-    if (CUDA_PERROR_DEBUG(config.result = get_sm_count(sm_count)))
-        return config;
-
-    // Get k-split flag storage (TODO: make a pool)
-    int *d_flags;
-    if (CUDA_PERROR_DEBUG(config.result = cudaGetSymbolAddress((void**) &d_flags, d_flags_split_k)))
-        return config;
-
-    // Construct k-split coordinator
-    k_split_control k_split(
-        d_flags,
-        sm_count,
-        max_sm_occupancy,
-        k,
-        block_task_policy_t::BlockItemsK,
-        config.block,
-        config.grid);     // in,out
-
-    config.split_k = k_split.split_k;
-
-    // Log kernel configuration
-    if (debug_synchronous)
-    {
-        // Compute tiling efficiency
-        float block_tiling_efficiency = float(block_task_policy_t::BlockItemsY * block_task_policy_t::BlockItemsX) /
-            float(block_task_policy_t::BlockItemsY + block_task_policy_t::BlockItemsX);
-
-        float tiling_efficiency = block_tiling_efficiency;
-
-        float wave_efficiency = k_split.get_wave_efficiency(
-            sm_count, max_sm_occupancy, config.block, config.grid);
-
-        CUDA_LOG_DEBUG("Final wave_efficiency %.4f, tiling_efficiency %.4f\n",
-            wave_efficiency, tiling_efficiency);
-
-        CUDA_LOG_DEBUG("Invoking kernel<<<(%d, %d, %d), (%d.y,%d.x), %d, %lld>>>(), %d SM occupancy, %d split_k\n",
-            config.grid.x, config.grid.y, config.grid.z,
-            config.block.y, config.block.x,
-            dynamic_smem_bytes,
-            (long long) stream,
-            max_sm_occupancy,
-            k_split.split_k);
-    }
-
-    // Construct parameter-pack
-    param_pack<value_t, accum_t, epilogue_op_t> pack(
-        m,
-        n,
-        k,
-        k_split,
-        epilogue_op,
-        d_a,
-        d_b,
-        d_c);
-
-    // Prepare k-split coordinator
-    if (CUDA_PERROR_DEBUG(config.result = k_split.prepare(stream, debug_synchronous)))
-    {
-        return config;
-    }
-
-    // Invoke kernel
-    kernel_ptr<<< config.grid, config.block, dynamic_smem_bytes, stream >>>(pack);
-
-    // Check for failure to launch
-    if (CUDA_PERROR_DEBUG(config.result = cudaPeekAtLastError()))
-        return config;
-
-    // Sync the stream if specified to flush runtime errors
-    if (debug_synchronous && (CUDA_PERROR_DEBUG(config.result = cudaStreamSynchronize(stream))))
-        return config;
-
-    return config;
-}
-
-
-/******************************************************************************
- * GEMM
- ******************************************************************************/
-
-/**
- * Computes gemm on device matrices
- */
-template <
-    tiling_strategy::kind_t      TilingStrategy,    ///< Tile-sizing classification
-    math_operation_class_t      math_op,        ///< Indicates which class of math operation to select
-    matrix_transform_t::kind_t  TransformA,     ///< Transformation op for matrix A
-    int                         LdgAlignA,      ///< Alignment (in bytes) of A operand
-    matrix_transform_t::kind_t  TransformB,     ///< Transformation op for matrix B
-    int                         LdgAlignB,      ///< Alignment (in bytes) of B operand
-    typename                    value_t,        ///< Multiplicand value type (matrices A and B)
-    typename                    accum_t,        ///< Accumulator value type (matrix C and scalars)
-    typename                    epilogue_op_t,  ///< Epilogue operation to update matrix C
-    int                         LdgAlignC>      ///< Alignment (in bytes) of C operand
-launch_configuration device_gemm(
-    int             m,                          ///< Height in rows of op(A) and C
-    int             n,                          ///< Width in columns of op(B) and C
-    int             k,                          ///< Width in columns of op(A) and height in rows of op(B)
-    epilogue_op_t   epilogue_op,                ///< Epilogue operation to update matrix C
-    value_t         *d_a,                       ///< Device pointer to matrix A array values
-    value_t         *d_b,                       ///< Device pointer to matrix B array values
-    accum_t         *d_c,                       ///< Device pointer to matrix C array values
-    cudaStream_t    stream = 0,                 ///< CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
-    bool            debug_synchronous = false)  ///< Whether or not to synchronize the stream after every kernel launch to
-                                                ///  check for errors.  Also causes launch configurations to be printed to
-                                                ///  the console if DEBUG is defined.  Default is \p false.
-{
-    // Parameterize an task policy type
-    // (TODO: use a policy dispatch mechanism based upon SM version)
-    typedef gemm_policy<value_t, accum_t, TransformA, TransformB, TilingStrategy> block_task_policy_t;
-
-    // AllowRaggedTiles-tile check
-    if ((m % block_task_policy_t::BlockItemsY != 0) ||
-        (n % block_task_policy_t::BlockItemsX != 0) ||
-        (k % block_task_policy_t::BlockItemsK != 0))
-    {
-        // Needs ragged tile-handling
-        static const bool AllowRaggedTiles = true;
-
-        return dispatch<math_op, block_task_policy_t, TransformA, LdgAlignA, TransformB, LdgAlignB, value_t, accum_t, epilogue_op_t, LdgAlignC, AllowRaggedTiles>(
-            kernel<math_op,block_task_policy_t, TransformA, LdgAlignA, TransformB, LdgAlignB, value_t, accum_t, epilogue_op_t, LdgAlignC, AllowRaggedTiles>,
-            m,
-            n,
-            k,
-            epilogue_op,
-            d_a,
-            d_b,
-            d_c,
-            stream,
-            debug_synchronous);
-    }
-    else
-    {
-        // Does not need ragged tile-handling
-        static const bool AllowRaggedTiles = false;
-
-        return dispatch<math_op, block_task_policy_t, TransformA, LdgAlignA, TransformB, LdgAlignB, value_t, accum_t, epilogue_op_t, LdgAlignC, AllowRaggedTiles>(
-            kernel<math_op,block_task_policy_t, TransformA, LdgAlignA, TransformB, LdgAlignB, value_t, accum_t, epilogue_op_t, LdgAlignC, AllowRaggedTiles>,
-            m,
-            n,
-            k,
-            epilogue_op,
-            d_a,
-            d_b,
-            d_c,
-            stream,
-            debug_synchronous);
-    }
-
-
-}
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/dispatch_policies.h

@@ -1,661 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Architecture-specific GEMM block_task policies
- */
-
-#include <stdint.h>
-
-#include "../util/util.h"
-#include "block_task.h"
-#include "grid_raster.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * tiling_strategy
- ******************************************************************************/
-
-/**
- * Enumeration of tile-sizing granularities
- */
-struct tiling_strategy : printable_t
-{
-    /// \brief Enumerants
-    enum kind_t
-    {
-        Unknown,
-        Small,
-        Medium,
-        Large,
-        Tall,
-        Wide,
-        Huge,
-    };
-
-    /// Enumerant value
-    kind_t kind;
-
-    /// Default constructor
-    tiling_strategy() : kind(Unknown) {}
-
-    /// Copy constructor
-    tiling_strategy(const kind_t &other_kind) : kind(other_kind) {}
-
-    /// Cast to kind_t
-    operator kind_t() const { return kind; }
-
-    /// Returns the instance as a string
-    __host__ __device__ inline
-    char const* to_string() const
-    {
-        switch (kind)
-        {
-            case Small:     return "small";
-            case Medium:    return "medium";
-            case Large:     return "large";
-            case Tall:      return "tall";
-            case Wide:      return "wide";
-            case Huge:      return "huge";
-            case Unknown:
-            default:        return "unknown";
-        }
-    }
-
-    /// Insert the formatted instance into the output stream
-    void print(std::ostream& out) const { out << to_string(); }
-};
-
-
-/******************************************************************************
- * GEMM
- ******************************************************************************/
-
-/**
- * GEMM task policy specialization for sgemm
- */
-template <
-    typename value_t,
-    typename accum_t,
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB,      ///< Transformation op for matrix B
-    tiling_strategy::kind_t TilingStrategy>     ///< Tile-sizing classification
-struct gemm_policy;
-
-
-/******************************************************************************
- * SGEMM
- ******************************************************************************/
-
-/**
- * GEMM task policy specialization for small sgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<float, float, TransformA, TransformB, tiling_strategy::Small> :
-    block_task_policy<
-        16,     // _BlockItemsY
-        16,     // _BlockItemsX
-        16,     // _BlockItemsK
-        2,      // _ThreadItemsY
-        2,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/**
- * GEMM task policy specialization for medium sgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<float, float, TransformA, TransformB, tiling_strategy::Medium> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        32,     // _BlockItemsX
-        8,      // _BlockItemsK
-        4,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for large sgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<float, float, TransformA, TransformB, tiling_strategy::Large> :
-    block_task_policy<
-        64,     // _BlockItemsY
-        64,     // _BlockItemsX
-        8,      // _BlockItemsK
-        8,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for tall sgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<float, float, TransformA, TransformB, tiling_strategy::Tall> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        32,     // _BlockItemsX
-        8,      // _BlockItemsK
-        8,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for wide sgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<float, float, TransformA, TransformB, tiling_strategy::Wide> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        128,    // _BlockItemsX
-        8,      // _BlockItemsK
-        4,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for huge sgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<float, float, TransformA, TransformB, tiling_strategy::Huge> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        128,    // _BlockItemsX
-        8,      // _BlockItemsK
-        8,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/******************************************************************************
- * DGEMM
- ******************************************************************************/
-
-/**
- * GEMM task policy specialization for small dgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<double, double, TransformA, TransformB, tiling_strategy::Small> :
-    block_task_policy<
-        16,     // _BlockItemsY
-        16,     // _BlockItemsX
-        16,     // _BlockItemsK
-        2,      // _ThreadItemsY
-        2,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/**
- * GEMM task policy specialization for medium dgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<double, double, TransformA, TransformB, tiling_strategy::Medium> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        32,     // _BlockItemsX
-        16,     // _BlockItemsK
-        4,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for large dgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<double, double, TransformA, TransformB, tiling_strategy::Large> :
-    block_task_policy<
-        64,     // _BlockItemsY
-        64,     // _BlockItemsX
-        8,      // _BlockItemsK
-        4,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for tall dgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<double, double, TransformA, TransformB, tiling_strategy::Tall> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        32,     // _BlockItemsX
-        8,      // _BlockItemsK
-        8,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for wide dgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<double, double, TransformA, TransformB, tiling_strategy::Wide> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        128,    // _BlockItemsX
-        8,      // _BlockItemsK
-        4,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for huge dgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<double, double, TransformA, TransformB, tiling_strategy::Huge> :
-    block_task_policy<
-        64,     // _BlockItemsY
-        128,    // _BlockItemsX
-        8,      // _BlockItemsK
-        8,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/******************************************************************************
- * HGEMM
- ******************************************************************************/
-
-/**
- * GEMM task policy specialization for small hgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<__half, __half, TransformA, TransformB, tiling_strategy::Small> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        32,     // _BlockItemsX
-        8,      // _BlockItemsK
-        4,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for medium hgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<__half, __half, TransformA, TransformB, tiling_strategy::Medium> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        32,     // _BlockItemsX
-        16,     // _BlockItemsK
-        8,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for large hgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<__half, __half, TransformA, TransformB, tiling_strategy::Large> :
-    block_task_policy<
-        64,     // _BlockItemsY
-        64,     // _BlockItemsX
-        8,      // _BlockItemsK
-        16,     // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for tall hgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<__half, __half, TransformA, TransformB, tiling_strategy::Tall> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        32,     // _BlockItemsX
-        8,      // _BlockItemsK
-        16,     // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/**
- * GEMM task policy specialization for wide hgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<__half, __half, TransformA, TransformB, tiling_strategy::Wide> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        128,    // _BlockItemsX
-        8,      // _BlockItemsK
-        8,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for huge hgemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<__half, __half, TransformA, TransformB, tiling_strategy::Huge> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        128,    // _BlockItemsX
-        8,      // _BlockItemsK
-        16,     // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/******************************************************************************
- * IGEMM
- ******************************************************************************/
-
-/**
- * GEMM task policy specialization for small igemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<int8_t, int32_t, TransformA, TransformB, tiling_strategy::Small> :
-    block_task_policy<
-        16,     // _BlockItemsY
-        32,     // _BlockItemsX
-        32,     // _BlockItemsK
-        4,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-
-/**
- * GEMM task policy specialization for medium igemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<int8_t, int32_t, TransformA, TransformB, tiling_strategy::Medium> :
-    block_task_policy<
-        32,     // _BlockItemsY
-        32,     // _BlockItemsX
-        32,     // _BlockItemsK
-        4,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for large igemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<int8_t, int32_t, TransformA, TransformB, tiling_strategy::Large> :
-    block_task_policy<
-        64,     // _BlockItemsY
-        64,     // _BlockItemsX
-        32,     // _BlockItemsK
-        8,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>  // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for large igemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<int8_t, int32_t, TransformA, TransformB, tiling_strategy::Tall> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        64,     // _BlockItemsX
-        64,     // _BlockItemsK
-        8,      // _ThreadItemsY
-        4,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>   // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for large igemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<int8_t, int32_t, TransformA, TransformB, tiling_strategy::Wide> :
-    block_task_policy<
-        64,     // _BlockItemsY
-        128,    // _BlockItemsX
-        64,     // _BlockItemsK
-        4,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>  // _RasterStrategy
-{};
-
-/**
- * GEMM task policy specialization for huge igemm
- */
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<int8_t, int32_t, TransformA, TransformB, tiling_strategy::Huge> :
-    block_task_policy<
-        128,    // _BlockItemsY
-        128,    // _BlockItemsX
-        32,     // _BlockItemsK
-        8,      // _ThreadItemsY
-        8,      // _ThreadItemsX
-        false,  // _UseDoubleScratchTiles
-        grid_raster_strategy::Default>  // _RasterStrategy
-{};
-
-
-/******************************************************************************
- * WMMA GEMM
- ******************************************************************************/
-
-// WMMA is a preview feature in CUDA. Conditionally enable wmma_gemm policies.
-#if defined(WMMA)
-
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<half, float, TransformA, TransformB, tiling_strategy::Small> :
-    gemm::block_task_wmma_policy<
-        16,     // _BlockItemsY
-        16,     // _BlockItemsX
-        16,     // _BlockItemsK
-        16,     // _WarpItemsY
-        16,     // _WarpItemsX
-        16,     // _WmmaItemsY
-        16,     // _WmmaItemsX
-        16,     // _WmmaItemsK
-        false,  // _UseDoubleScratchTiles
-        gemm::grid_raster_strategy::Default>    // _RasterStrategy
-{};
-
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy<half, float, TransformA, TransformB, tiling_strategy::Medium> :
-    gemm::block_task_wmma_policy<
-        32,     // _BlockItemsY
-        32,     // _BlockItemsX
-        32,     // _BlockItemsK
-        32,     // _WarpItemsY
-        32,     // _WarpItemsX
-        16,     // _WmmaItemsY
-        16,     // _WmmaItemsX
-        16,     // _WmmaItemsK
-        false,  // _UseDoubleScratchTiles
-        gemm::grid_raster_strategy::Default>    // _RasterStrategy
-{};
-
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy< half, float, TransformA, TransformB, tiling_strategy::Large> :
-    gemm::block_task_wmma_policy<
-        64,     // _BlockItemsY
-        64,     // _BlockItemsX
-        32,     // _BlockItemsK
-        32,     // _WarpItemsY
-        64,     // _WarpItemsX
-        16,     // _WmmaItemsY
-        16,     // _WmmaItemsX
-        16,     // _WmmaItemsK
-        false,  // _UseDoubleScratchTiles
-        gemm::grid_raster_strategy::Default>    // _RasterStrategy
-{};
-
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy< half, float, TransformA, TransformB, tiling_strategy::Tall> :
-    gemm::block_task_wmma_policy<
-        128,    // _BlockItemsY
-        64,     // _BlockItemsX
-        64,     // _BlockItemsK
-        32,     // _WarpItemsY
-        64,     // _WarpItemsX
-        16,     // _WmmaItemsY
-        16,     // _WmmaItemsX
-        16,     // _WmmaItemsK
-        false,  // _UseDoubleScratchTiles
-        gemm::grid_raster_strategy::Default>    // _RasterStrategy
-{};
-
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy< half, float, TransformA, TransformB, tiling_strategy::Wide> :
-    gemm::block_task_wmma_policy<
-        64,     // _BlockItemsY
-        128,    // _BlockItemsX
-        64,     // _BlockItemsK
-        32,     // _WarpItemsY
-        64,     // _WarpItemsX
-        16,     // _WmmaItemsY
-        16,     // _WmmaItemsX
-        16,     // _WmmaItemsK
-        false,  // _UseDoubleScratchTiles
-        gemm::grid_raster_strategy::Default>    // _RasterStrategy
-{};
-
-template <
-    matrix_transform_t::kind_t TransformA,      ///< Transformation op for matrix A
-    matrix_transform_t::kind_t TransformB>      ///< Transformation op for matrix B
-struct gemm_policy< half, float, TransformA, TransformB, tiling_strategy::Huge> :
-    gemm::block_task_wmma_policy<
-        128,    // _BlockItemsY
-        128,    // _BlockItemsX
-        64,     // _BlockItemsK
-        32,     // _WarpItemsY
-        64,     // _WarpItemsX
-        16,     // _WmmaItemsY
-        16,     // _WmmaItemsX
-        16,     // _WmmaItemsK
-        false,  // _UseDoubleScratchTiles
-        gemm::grid_raster_strategy::Default>    // _RasterStrategy
-{};
-
-#endif
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/dp_accummulate.h

@@ -1,223 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Abstraction for exposing architecture-specific "dot-product-accumulate"
- * ISA operations
- */
-
-#include <stdint.h>
-
-#include "../util/util.h"
-
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * dp_accummulate
- ******************************************************************************/
-
-
-/**
- * \brief Abstraction for exposing architecture-specific "dot-product-accumulate"
- * ISA operations
- *
- * Given two K-component vectors a and b having type value_t[K] and an addend c
- * of type accum_t, the "dot-product-accumulate" of type accum_t is computed
- * as d = x[0]*y[0] + x[1]*y[1] + ...  + x[K-1]*y[K-1] + c.
- *
- * We use the notation "dpK" to connote a K-component dot-product-accumulate.
- * For example, "dp1" is a simple multiply-add.
- *
- * For given pairing of value_t and accum_t types, the corresponding
- * dp_accummulate class will:
- *
- * - Define the member-type dp_vector_t as the appropriate K-component vector
- *   type needed to leverage architecture-specific "dot-product accumulate"
- *   ISA operations.
- * - Implement the corresponding dot-product operation between two dp_vector_t
- *   inputs a and b.
- *
- */
-template <
-    typename value_t,       ///< Component value type
-    typename accum_t>       ///< Accumulator value type
-struct dp_accummulate;
-
-
-
-/// Default "dp1" dot-product-accumulate traits specialization for value_t->accum_t
-template <
-    typename value_t,       ///< Component value type
-    typename accum_t>       ///< Accumulator value type
-struct dp_accummulate
-{
-    /// Single-component "dp1" dot-product vector type
-    typedef value_t dp_vector_t;
-
-
-    /// Compute "dp1" float->float
-    inline __device__
-    static void mad(
-        float &d,
-        const float &a,
-        const float &b,
-        const float &c)
-    {
-        asm volatile ( "fma.rn.f32 %0, %1, %2, %3;\n"
-            : "=f"(d) : "f"(a), "f"(b), "f"(c));
-    }
-
-
-    /// Compute "dp1" double->double
-    inline __device__
-    static void mad(
-        double &d,
-        const double &a,
-        const double &b,
-        const double &c)
-    {
-        asm volatile ("fma.rn.f64 %0, %1, %2, %3;\n"
-            : "=d"(d) : "d"(a), "d"(b), "d"(c));
-    }
-
-
-    /// Compute "dp1" int16_t->int32_t
-    inline __device__
-    static void mad(
-        int32_t &d,
-        const int16_t &a,
-        const int16_t &b,
-        const int32_t &c)
-    {
-        asm volatile ("mad.wide.s16 %0, %1, %2, %3;\n"
-            : "=r"(d) : "h"(a), "h"(b), "r"(c));
-    }
-
-
-    /// Compute "dp1" uint16_t->uint32_t
-    inline __device__
-    static void mad(
-        uint32_t &d,
-        const uint16_t &a,
-        const uint16_t &b,
-        const uint32_t &c)
-    {
-        asm volatile ("mad.wide.u16 %0, %1, %2, %3;\n"
-            : "=r"(d) : "h"(a), "h"(b), "r"(c));
-    }
-
-
-    /// Compute "dp1" int32_t->int32_t
-    inline __device__
-    static void mad(
-        int32_t &d,
-        const int32_t &a,
-        const int32_t &b,
-        const int32_t &c)
-    {
-        asm volatile ("mad.lo.s32 %0, %1, %2, %3;\n"
-            : "=r"(d) : "r"(a), "r"(b), "r"(c));
-    }
-
-
-    /// Compute "dp1" uint32_t->uint32_t
-    inline __device__
-    static void mad(
-        uint32_t &d,
-        const uint32_t &a,
-        const uint32_t &b,
-        const uint32_t &c)
-    {
-        asm volatile ("mad.lo.u32 %0, %1, %2, %3;\n"
-            : "=r"(d) : "r"(a), "r"(b), "r"(c));
-    }
-
-};
-
-
-
-#if (CUTLASS_ARCH >= 610)   // Specializations only enabled for Pascal SM610+
-
-
-/// "dp4" dot-product-accumulate traits specialization for int8_t->int32_t
-template <>
-struct dp_accummulate<
-    int8_t,                 ///< Component value type
-    int32_t>                ///< Accumulator value type
-{
-    /// Four-component signed "idp4"
-    typedef int32_t dp_vector_t;
-
-    /// Compute "dp4" int16_t->int32_t
-    inline __device__
-    static void mad(
-        int32_t &d,
-        const int32_t &a,
-        const int32_t &b,
-        const int32_t &c)
-    {
-        asm volatile ( "dp4a.s32.s32 %0, %1, %2, %3;\n"
-            : "=r"(d) : "r"(a), "r"(b), "r"(c));
-    }
-};
-
-
-/// "dp4" dot-product-accumulate traits specialization for uint8_t->uint32_t
-template <>
-struct dp_accummulate<
-    uint8_t,                ///< Component value type
-    uint32_t>               ///< Accumulator value type
-{
-    /// Four-component unsigned "idp4"
-    typedef uint32_t dp_vector_t;
-
-    /// Compute "dp4" uint16_t->uint32_t
-    inline __device__
-    static void mad(
-        uint32_t &d,
-        const uint32_t &a,
-        const uint32_t &b,
-        const uint32_t &c)
-    {
-        asm volatile ( "dp4a.u32.u32 %0, %1, %2, %3;\n"
-            : "=r"(d) : "r"(a), "r"(b), "r"(c));
-    }
-};
-
-
-#endif // Specializations only enabled for Pascal SM610+
-
-
-} // namespace gemm
-} // namespace cutlass
-

cutlass/gemm/epilogue_function.h

@@ -1,104 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Epilogue operation to compute final output
- */
-
-namespace cutlass {
-namespace gemm {
-
-    //// Used by GEMM to compute the final result C <= alpha * accumulator + beta * C
-    template <
-        typename accum_t,
-        typename output_t,
-        typename scalar_t
-    >
-    class blas_scaled_epilogue
-    {
-    public:
-
-        scalar_t alpha;
-        scalar_t beta;
-
-        inline __device__ __host__
-        blas_scaled_epilogue(
-            scalar_t alpha,
-            scalar_t beta)
-        :
-            alpha(alpha),
-            beta(beta)
-        {}
-
-
-        /// Epilogue operator
-        inline __device__ __host__
-        output_t operator()(
-            accum_t accumulator,
-            output_t c,
-            size_t idx) const
-        {
-            return output_t(alpha * scalar_t(accumulator) + beta * scalar_t(c));
-        }
-
-
-        /// Epilogue operator
-        inline __device__ __host__
-        output_t operator()(
-            accum_t accumulator,
-            size_t idx) const
-        {
-            return output_t(alpha * scalar_t(accumulator));
-        }
-
-        /**
-         * Configure epilogue as to whether the thread block is a secondary
-         * accumulator in an inter-block k-splitting scheme
-         */
-        inline __device__
-        void set_secondary_accumulator()
-        {
-            beta = scalar_t(1);
-        }
-
-
-        /// Return whether the beta-scaled addend needs initialization
-        inline __device__
-        bool must_init_addend()
-        {
-            return (beta != scalar_t(0));
-        }
-    };
-
-
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/fp16_sgemm_multiply_add.h

@@ -0,0 +1,83 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Template implementing matrix multiply-add operations on fragments.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+#include "cutlass/gemm/thread_multiply_add.h"
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Template performing matrix multiply-add operation within a thread
+template <typename ThreadGemmShape_,
+          typename ThreadsPerWarp_>
+struct ThreadMultiplyAdd<ThreadGemmShape_, ThreadsPerWarp_, half, half, float> {
+  /// The shape of the instruction.
+  typedef Shape<1, 1, 1, 1> InstructionShape;
+  /// The shape of a thread-leveel matrix multiply accumulate.
+  typedef ThreadGemmShape_ ThreadGemmShape;
+  /// Aliased to "AccumulatorsPerThread" for compatibility. Expect to be renamed in CUTLASS v2.0
+  typedef ThreadGemmShape AccumulatorsPerThread;
+  /// The number of threads per warp.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of accumulators per warp.
+  typedef typename ShapeMul<ThreadGemmShape, ThreadsPerWarp>::Shape AccumulatorsPerWarp;
+  /// The type for A. specialized to half
+  typedef half ScalarA;
+  /// The fragment for A.
+  typedef Fragment<ScalarA, AccumulatorsPerThread::kW> FragmentA;
+  /// The type for B. specialized to half
+  typedef half ScalarB;
+  /// The fragment for B.
+  typedef Fragment<ScalarB, AccumulatorsPerThread::kH> FragmentB;
+  /// The type for C and D. specialized to float
+  typedef float ScalarC;
+  /// The accumulators.
+  typedef Fragment<ScalarC, AccumulatorsPerThread::kH * AccumulatorsPerThread::kW, 16> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE ThreadMultiplyAdd() {}
+
+  /// Multiply : d = a*b + c.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    for (int j = 0; j < AccumulatorsPerThread::kH; ++j) {
+      for (int i = 0; i < AccumulatorsPerThread::kW; ++i) {
+        d[j * AccumulatorsPerThread::kW + i] = static_cast<ScalarC>(a[i]) * static_cast<ScalarC>(b[j]) + c[j * AccumulatorsPerThread::kW + i];
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/fp16_sgemm_traits.h

@@ -0,0 +1,152 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defies structural properties of single-precision GEMM where any number of the input/output
+    could be fp16 or fp32. The accumulator type stays in fp32
+*/
+#pragma once
+
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/gemm/gemm_epilogue.h"
+#include "cutlass/gemm/gemm_epilogue_traits.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/gemm/gemm_shared_tile.h"
+#include "cutlass/gemm/gemm_traits.h"
+#include "cutlass/gemm/fp16_sgemm_multiply_add.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The tile size for the GEMM KxNxM.
+    typename OutputTile_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_,
+    /// The type for A
+    typename ScalarA_,
+    /// The type for B
+    typename ScalarB_,
+    /// The type for C
+    typename ScalarC_,
+    /// The type for D
+    typename ScalarD_,
+    /// The number of scalars per LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of scalars per LDG for B.
+    int kScalarsPerLdgB_ = 1>
+struct Fp16SgemmConfig : public GemmConfig<
+                         /// The scalar type for A.
+                         ScalarA_,
+                         /// The scalar type for B.
+                         ScalarB_,
+                         /// The scalar type for C.
+                         ScalarC_,
+                         /// The scalar type for D.
+                         ScalarD_,
+                         /// The tile size for the GEMM KxNxM.
+                         OutputTile_,
+                         /// The functor to do the math in the main loop.
+                         ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, ScalarA_, ScalarB_, float /*for sgemm accum is float*/>,
+                         /// The number of scalars per LDG for A.
+                         kScalarsPerLdgA_,
+                         /// The number of scalars per STS for A.
+                         kScalarsPerLdgA_,
+                         /// The number of scalars per LDS for A.
+                         4,
+                         /// The number of scalars per LDG for B.
+                         kScalarsPerLdgB_,
+                         /// The number of scalars per STS for B.
+                         kScalarsPerLdgB_,
+                         /// The number of scalars per LDS for B.
+                         4,
+                         /// The number of scalars per LDG for C and STG for D.
+                         1,
+                         /// The number of scalars per STS for D.
+                         4,
+                         /// The number of scalars per LDS for D.
+                         1,
+                         /// The number of stages in shared memory.
+                         2> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_ = Shape<8, 128, 128>,
+    /// The type for A
+    typename ScalarA_ = half,
+    /// The type for B
+    typename ScalarB_ = half,
+    /// The type for C
+    typename ScalarC_ = half,
+    /// The type for D
+    typename ScalarD_ = half,
+    /// the Type for alpha and beta,
+    typename Scalar_ = half,
+    /// The functor to use in the epilogue.
+    typename EpilogueFunctor_ = LinearScaling<Scalar_, FragmentMultiplyAdd<Scalar_, float/*accumulator type*/> >,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<8, 8, 8>,
+    /// The number of floats loaded in one LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of floats loaded in one LDG for B.
+    int kScalarsPerLdgB_ = 1,
+    /// The index.
+    typename Index_ = int,
+    /// The SGEMM config.
+    typename GemmConfig_ =
+        Fp16SgemmConfig<OutputTile_, 
+                        ThreadGemmShape_, 
+                        ScalarA_,
+                        ScalarB_,
+                        ScalarC_,
+                        ScalarD_,
+                        kScalarsPerLdgA_, 
+                        kScalarsPerLdgB_>,
+    /// The traits class for the epilogue.
+    typename GemmEpilogueTraits_ =
+        SimplifiedGemmEpilogueTraits<GemmConfig_, EpilogueFunctor_, Index_> >
+struct Fp16SgemmSgemmTraits : public SimplifiedGemmTraits<
+                         // The layout for A.
+                         kLayoutA_,
+                         // The layout for B.
+                         kLayoutB_,
+                         // The config.
+                         GemmConfig_,
+                         // The epilogue.
+                         GemmEpilogue<GemmEpilogueTraits_>,
+                         // The index.
+                         Index_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm.h

@@ -0,0 +1,355 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements a software-pipelined efficient GEMM.
+*/
+#pragma once
+
+#if !defined(__CUDACC_RTC__)
+#include <cuda.h>
+#endif
+
+#include "cutlass/coord.h"
+#include "cutlass/util/platform.h"
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// GEMM kernel with launch bounds specified
+template <typename Gemm_>
+__global__  __launch_bounds__(Gemm_::kThreads)
+void gemm_kernel(typename Gemm_::Params params) {
+  // Declare shared memory.
+  __shared__ typename Gemm_::SharedStorage shared_storage;
+
+  // Construct the GEMM object.
+  Gemm_ gemm(params, shared_storage);
+  // Run GEMM.
+  gemm.multiply_add();
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// GEMM kernel without launch bounds specified
+template <typename Gemm_>
+__global__ /* __launch_bounds__(Gemm_::kThreads) */
+void gemm_kernel_nolb(typename Gemm_::Params params) {
+  // Declare shared memory.
+  __shared__ typename Gemm_::SharedStorage shared_storage;
+
+  // Construct the GEMM object.
+  Gemm_ gemm(params, shared_storage);
+  // Run GEMM.
+  gemm.multiply_add();
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Partial specialization for launching the GEMM kernel with or without launch bounds
+template <typename Gemm, bool WithLaunchBounds>
+struct Launch {
+  Launch(typename Gemm::Params params, dim3 grid, dim3 block, cudaStream_t stream = 0) {
+    gemm_kernel<Gemm><<< grid, block, 0, stream >>>(params);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Partial specialization for launching the GEMM kernel with or without launch bounds
+template <typename Gemm>
+struct Launch<Gemm, false> {
+  Launch(typename Gemm::Params params, dim3 grid, dim3 block, cudaStream_t stream = 0) {
+    gemm_kernel_nolb<Gemm><<< grid, block, 0, stream >>>(params);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmTraits_>
+struct Gemm {
+  /// This class.
+  typedef Gemm<GemmTraits_> This_;
+  /// The traits.
+  typedef GemmTraits_ Traits;
+  /// The shared storage.
+  typedef typename Traits::SharedStorage SharedStorage;
+
+  /// The scalar for A.
+  typedef typename Traits::ScalarA ScalarA;
+  /// The scalar for B.
+  typedef typename Traits::ScalarB ScalarB;
+  /// The scalar in the epilogue.
+  typedef typename Traits::Epilogue::Scalar ScalarEpilogue;
+  /// The scalar for C.
+  typedef typename Traits::Epilogue::ScalarC ScalarC;
+  /// The scalar for D.
+  typedef typename Traits::Epilogue::ScalarD ScalarD;
+  /// The index.
+  typedef typename Traits::Index Index;
+
+  /// Define the mainloop iteration size
+  typedef typename Traits::MultiplyAdd MultiplyAdd;
+
+  /// The number of threads.
+  static int const kThreads = Traits::GemmConfig::kThreads;
+
+  // Number of warp-level multiply-accumulate steps executed by each warp.
+  static Index const kWarpGemmSteps =
+      Traits::GemmConfig::AccumulatorsPerWarp::kD / MultiplyAdd::InstructionShape::kD;
+
+  // Make sure we have at least 2 unrolling steps or our pipeling is not going to work.
+  static_assert(kWarpGemmSteps >= 2, "The pipelining assumes at least two steps");
+
+  /// Use the params object defined in traits
+  typedef typename Traits::Params Params;
+
+//
+// Static function members
+//
+
+/// Support for NVRTC
+#if !defined(__CUDACC_RTC__)
+  /// Launch the kernel.
+  static __host__ cudaError_t launch(Params const& params,
+                                     cudaStream_t stream = cudaStreamDefault) {
+
+    // Launch the kernel.
+    Launch<This_, GemmTraits_::GemmConfig::kLaunchBounds>(
+      params, params.grid, params.block, stream);
+
+    return cudaGetLastError();
+  }
+
+  /// Launch the kernel.
+  static __host__ cudaError_t launch(CUfunction kernel,
+                                     Params const& params,
+                                     CUstream stream = CU_STREAM_LEGACY) {
+
+    // Launch the kernel.
+    void* params_[] = {const_cast<void*>(reinterpret_cast<void const*>(&params))};
+
+    CUresult result = cuLaunchKernel(
+        kernel,
+        params.grid.x, params.grid.y, params.grid.z,
+        params.block.x, params.block.y, params.block.z,
+        0, stream, params_, 0);
+
+    if (result != CUDA_SUCCESS) {
+      return cudaErrorLaunchFailure;
+    }
+    return cudaSuccess;
+  }
+
+#endif
+
+  //
+  // Methods
+  //
+
+  /// Ctor.
+  CUTLASS_DEVICE Gemm(Params const& params_, SharedStorage& shared_storage_)
+      : params(params_), shared_storage(shared_storage_) {}
+
+  /// Computes a warp-level GEMM on data held in shared memory
+  template <bool Residue, bool LastIteration>
+  CUTLASS_DEVICE void consume_tile(typename Traits::GlobalLoadStream& global_to_shared_stream,
+                                   typename Traits::SharedStream& shared_load_stream,
+                                   typename MultiplyAdd::Accumulators& accumulators,
+                                   Index outer_k) {
+    // If residue portion and not calculating residue in prolog, update residue predicates now.
+    if (Residue && outer_k <= Traits::OutputTile::kD) {
+      global_to_shared_stream.residue(outer_k);
+    }
+
+    // Load data for the next iteration of the main loop (unless it's the last iteration).
+    if (!LastIteration) {
+      global_to_shared_stream.copy();
+    }
+
+    CUTLASS_PRAGMA_UNROLL
+    for (int step = 0; step < kWarpGemmSteps - 1; ++step) {
+      // Trigger the copy from shared memory for the next A/B values.
+      shared_load_stream.copy(step + 1);
+
+      // Make sure the values are available for the current iteration to do the multiply-add.
+      shared_load_stream.commit(step);
+
+      MultiplyAdd multiply_add;
+
+      // Do the math on the fragments of the current iteration.
+      multiply_add.multiply_add(shared_load_stream.fragment_a(step),
+                                shared_load_stream.fragment_b(step),
+                                accumulators,
+                                accumulators);
+    }
+
+    // Make sure the data from shared memory has been entirely consumed.
+    Traits::shared_load_fence(true);
+
+    // Commit the data in shared memory for A/B.
+    if (!LastIteration) {
+      global_to_shared_stream.commit();
+    }
+    // Make sure the data is in shared memory.
+    Traits::shared_store_fence(true);
+
+    if (!LastIteration) {
+      // Move to the next stage for the load (if it makes sense).
+      shared_load_stream.inc_stage();
+      // Trigger the copy from shared memory for the next loop iteration.
+      shared_load_stream.copy(0);
+    }
+    // Make sure the values are available for the current iteration to do the multiply-add.
+    shared_load_stream.commit(kWarpGemmSteps - 1);
+
+    // Do the math on the fragments of the current iteration.
+    MultiplyAdd multiply_add;
+    multiply_add.multiply_add(shared_load_stream.fragment_a(kWarpGemmSteps - 1),
+                              shared_load_stream.fragment_b(kWarpGemmSteps - 1),
+                              accumulators,
+                              accumulators);
+  }
+
+  /// Do the GEMM.
+  CUTLASS_DEVICE void multiply_add() {
+    // Swizzle the IDs of the block (to enable better cache behavior).
+    typename Traits::BlockSwizzle block_swizzle;
+    Coord<3> threadblock_offset =
+        block_swizzle.get_threadblock_offset(make_Coord_from_shape<Traits::OutputTile>());
+
+    // We may want to use shared memory to clear the registers.
+    typedef typename Traits::ClearAccumulators ClearAccumulators;
+
+    // The streams to read A/B from global memory to shared memory.
+    typename Traits::GlobalLoadStream global_to_shared_stream(
+        params.global_to_shared_stream,
+        shared_storage.main_loop.global_to_shared_stream,
+        shared_storage.main_loop.threadblock_tile.reference(),
+        params.problem_size.knm(),
+        threadblock_offset);
+
+    // update A and B pointer offset based on batch_id and batch_stride_offset
+    //global_to_shared_stream.add_pointer_offset(block_swizzle.get_batch_id(), params.batch_stride_A, params.batch_stride_B);
+    global_to_shared_stream += make_Coord(block_swizzle.get_batch_id(), 0, 0);
+
+    // Create the accumulator clear.
+    ClearAccumulators clear;
+
+    // Deal with residue in prolog.
+    global_to_shared_stream.move_to_residue(params.problem_size[0], Traits::OutputTile::kD);
+
+    // Fetch the fragments for A and B from global memory.
+    global_to_shared_stream.copy();
+
+    // Copy the elements to shared memory (after transformation if needed).
+    global_to_shared_stream.commit();
+
+    // Make sure the data is in shared memory.
+    Traits::shared_store_fence(false);
+
+    // Rollback to the beginning of the first tile (if residue exists).
+    global_to_shared_stream.rollback(params.problem_size[0] % Traits::OutputTile::kD);
+
+    // The stream of data from shared memory to fragments.
+    typename Traits::SharedStream shared_load_stream(
+        params.shared_stream,
+        shared_storage.main_loop.threadblock_tile.reference());
+
+    // Trigger the copy from shared memory for the 1st stream.
+    shared_load_stream.copy(0);
+
+    // Allocate the accumulators.
+    typename MultiplyAdd::Accumulators accumulators;
+
+    // Clear the accumulators.
+    clear.clear(accumulators);
+
+    // Initial index
+    Index outer_k = params.problem_size[0] - Traits::OutputTile::kD;
+
+    // Check if we are computing residue in prolog or not.
+    if (Traits::GemmConfig::kResidueInProlog) {
+
+      // Execute all mainloop iterations but the last one.
+
+      CUTLASS_GEMM_LOOP
+      for (; outer_k > 0; outer_k -= Traits::OutputTile::kD) {
+        consume_tile<false, false>(
+            global_to_shared_stream, shared_load_stream, accumulators, outer_k);
+
+      }
+
+      // Don't load data for the last "residue" portion since we've already computed the residue.
+      CUTLASS_GEMM_LOOP
+      for (; outer_k > -Traits::OutputTile::kD; outer_k -= Traits::OutputTile::kD) {
+        consume_tile<false, true>(
+            global_to_shared_stream, shared_load_stream, accumulators, outer_k);
+
+      }
+    } else {
+      // When kResidueSeparate = true, execute all mainloop iterations but the last two without any
+      // consideration for K-residue or predicate updates. This improves the steady state of some
+      // kernels.
+      if (Traits::GemmConfig::kResidueSeparate) {
+
+        CUTLASS_GEMM_LOOP
+        for (; outer_k > Traits::OutputTile::kD; outer_k -= Traits::OutputTile::kD) {
+          consume_tile<false, false>(
+              global_to_shared_stream, shared_load_stream, accumulators, outer_k);
+
+        }
+      }
+
+      // Execute remaining tiles with K-residue predicate updates enabled.
+
+      CUTLASS_GEMM_LOOP
+      for (; outer_k > -Traits::OutputTile::kD; outer_k -= Traits::OutputTile::kD) {
+        consume_tile<true, false>(
+            global_to_shared_stream, shared_load_stream, accumulators, outer_k);
+
+      }
+    }
+
+    // Epilogue.
+    typedef typename Traits::Epilogue Epilogue;
+    Epilogue epilogue(params.epilogue, shared_storage.epilogue, params.problem_size.knm());
+    epilogue.epilogue(accumulators, threadblock_offset, block_swizzle.get_batch_id());
+  }
+
+  //
+  // Data members
+  //
+
+  /// The params.
+  Params const& params;
+  /// The shared storage.
+  SharedStorage& shared_storage;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_config.h

@@ -0,0 +1,145 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *   * Redistributions of source code must retain the above copyright notice, this list of
+ *     conditions and the following disclaimer.
+ *   * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *     conditions and the following disclaimer in the documentation and/or other materials
+ *     provided with the distribution.
+ *   * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *     to endorse or promote products derived from this software without specific prior written
+ *     permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+  \brief Defines properties of GEMM computation that impose some constraints on caller.
+*/
+#pragma once
+
+#include "cutlass/shape.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The scalar type for A.
+    typename ScalarA_,
+    /// The scalar type for B.
+    typename ScalarB_,
+    /// The scalar type for C.
+    typename ScalarC_,
+    /// The scalar type for D.
+    typename ScalarD_,
+    /// The threadblock tile size for the GEMM KxNxM.
+    typename OutputTile_,
+    /// The functor to do the math.
+    typename MultiplyAdd_,
+    /// The number of scalars per LDG for A.
+    int kScalarsPerLdgA_,
+    /// The number of scalars per STS for A.
+    int kScalarsPerStsA_,
+    /// The number of scalars per LDG for A.
+    int kScalarsPerLdsA_,
+    /// The number of scalars per LDG for B.
+    int kScalarsPerLdgB_,
+    /// The number of scalars per STS for B.
+    int kScalarsPerStsB_,
+    /// The number of scalars per LDS for B.
+    int kScalarsPerLdsB_,
+    /// The number of scalars per LDG for C and STG for D.
+    int kScalarsPerLdgCAndStgD_,
+    /// The number of scalars per STS for D.
+    int kScalarsPerStsD_,
+    /// The number of scalars per LDS for D.
+    int kScalarsPerLdsD_,
+    /// The number of stages in shared memory to do single/double/triple-buffering.
+    int kStages_,
+    /// If true, residue is computed in mainloop. If false, separate loops are instantiated.
+    bool kResidueSeparate_ = false,
+    /// Is residue performed in prologue?
+    bool kResidueInProlog_ = false,
+    /// If true, kernel is launched with CUDA launch bounds specified
+    bool kLaunchBounds_ = true>
+struct GemmConfig {
+  //
+  /// The scalar for A.
+  typedef ScalarA_ ScalarA;
+  /// The scalar for B.
+  typedef ScalarB_ ScalarB;
+  /// The scalar for C.
+  typedef ScalarC_ ScalarC;
+  /// The scalar for D.
+  typedef ScalarD_ ScalarD;
+
+  /// The tile.
+  typedef OutputTile_ OutputTile;
+  /// The functor to do D = A*B + C.
+  typedef MultiplyAdd_ MultiplyAdd;
+  /// The shape of the instruction.
+  typedef typename MultiplyAdd::InstructionShape InstructionShape;
+  /// The shape of warp-level GEMM
+  typedef typename MultiplyAdd::AccumulatorsPerWarp AccumulatorsPerWarp;
+  /// The accumulators.
+  typedef typename MultiplyAdd::Accumulators Accumulators;
+
+  /// The number of warps.
+  typedef typename ShapeDiv<OutputTile, AccumulatorsPerWarp>::Shape Warps;
+  /// The default warp size (32 threads per warp).
+  static int const kWarpSize = cutlass::kWarpSize;
+  /// The numnber of threads.
+  static int const kThreads = ShapeCount<Warps>::kCount * kWarpSize;
+
+  /// The number of scalars per LDG/STS/LDS for A.
+  static int const kScalarsPerLdgA = kScalarsPerLdgA_;
+  static int const kScalarsPerStsA = kScalarsPerStsA_;
+  static int const kScalarsPerLdsA = kScalarsPerLdsA_;
+
+  /// The number of scalars per LDG/STS/LDS for B.
+  static int const kScalarsPerLdgB = kScalarsPerLdgB_;
+  static int const kScalarsPerStsB = kScalarsPerStsB_;
+  static int const kScalarsPerLdsB = kScalarsPerLdsB_;
+
+  /// The number of scalars per LDG for C.
+  static int const kScalarsPerLdgC = kScalarsPerLdgCAndStgD_;
+
+  /// The number of scalars per STS/LDS/STG for D.
+  static int const kScalarsPerStgD = kScalarsPerLdgCAndStgD_;
+  static int const kScalarsPerStsD = kScalarsPerStsD_;
+  static int const kScalarsPerLdsD = kScalarsPerLdsD_;
+
+  /// The number of accumulators that are going to be fed from one LDS A/B.
+  static int const kAccumulatorsPerLdsA = kScalarsPerLdsA / InstructionShape::kD;
+  static int const kAccumulatorsPerLdsB = kScalarsPerLdsB / InstructionShape::kD;
+
+  /// The number of stages in shared memory to implement double, triple, more-buffering.
+  static int const kStages = kStages_;
+
+  /// If true, mainloop is instantiated twice. The first instantiation contains no predicate
+  // updates and is more efficient for some kernels. If false, only a single mainloop is
+  // instantaited.
+  static bool const kResidueSeparate = kResidueSeparate_;
+
+  /// If true, residue is computed in the prologue.
+  static bool const kResidueInProlog = kResidueInProlog_;
+
+  /// If true, kernel is launched with launch bounds specified
+  static bool const kLaunchBounds = kLaunchBounds_;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_coord.h

@@ -0,0 +1,203 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief GemmCoord is a structure derived from Coord<4> that specifies a location within the
+      coordinate system of a GEMM problem.
+*/
+
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cutlass/coord.h"
+#include "cutlass/util/platform.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// GemmCoord is a structure derived from Coord<4> that specifies a location within the
+/// coordinate space of a GEMM problem.
+struct GemmCoord : public Coord<4, int> {
+
+  /// Integer-valued index
+  typedef int Index;
+
+  /// Base type is a Coord of rank=4
+  typedef Coord<4, Index> Base;
+
+  /// GEMM K dimension - inner dimension of the GEMM problem
+  static int const kK = 0;
+
+  /// GEMM N dimension - columns of the output C matrix
+  static int const kN = 1;
+
+  /// GEMM M dimension - rows of the output C matrix
+  static int const kM = 2;
+
+  /// Batch dimension - for generalizing to larger problems
+  static int const kBatch = 3;
+
+  //
+  // Methods
+  //
+
+  /// Default ctor
+  CUTLASS_HOST_DEVICE
+  GemmCoord() { }
+
+  /// Constructs from Coord<3> and a batch
+  CUTLASS_HOST_DEVICE
+  GemmCoord(Coord<3, Index> const &coord, Index _batch = 0): Base(make_Coord(coord[0], coord[1], coord[2], _batch)) { }
+
+  /// Constructs from Coord<4>
+  CUTLASS_HOST_DEVICE
+  GemmCoord(Coord<4, Index> const &coord): Base(coord) { }
+
+  /// Constructs from an array of coordinate elements
+  CUTLASS_HOST_DEVICE
+  GemmCoord(Index coord[4]): Base(coord) { }
+
+  /// Helper to construct from a K, N, M, batch variables
+  CUTLASS_HOST_DEVICE
+  GemmCoord(Index k, Index n, Index m, Index batch = 0): Base(make_Coord(k, n, m, batch)) { }
+
+  /// Returns the GEMM M coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & m() const { return this->at(kM); }
+
+  /// Returns reference to the GEMM M coordinate
+  CUTLASS_HOST_DEVICE
+  Index & m() { return this->at(kM); }
+
+  /// Returns the GEMM N coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & n() const { return this->at(kN); }
+
+  /// Returns reference to the GEMM N coordinate
+  CUTLASS_HOST_DEVICE
+  Index & n() { return this->at(kN); }
+
+  /// Returns the GEMM K coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & k() const { return this->at(kK); }
+
+  /// Returns reference to the GEMM K coordinate
+  CUTLASS_HOST_DEVICE
+  Index & k() { return this->at(kK); }
+
+  /// Returns the GEMM batch coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & batch() const { return this->at(kBatch); }
+
+  /// Returns reference to the GEMM batch coordinate
+  CUTLASS_HOST_DEVICE
+  Index & batch() { return this->at(kBatch); }
+
+  /// Obtains a Coord<3> from GemmCoord
+  CUTLASS_HOST_DEVICE
+  Coord<3> knm() const {
+    return make_Coord(k(), n(), m());
+  }
+
+  /// Obtains a Coord<2> from GemmCoord
+  CUTLASS_HOST_DEVICE
+  Coord<2> nm() const {
+    return make_Coord(n(), m());
+  }
+
+  /// Obtains a Coord<2> from GemmCoord
+  CUTLASS_HOST_DEVICE
+  Coord<2> km() const {
+    return make_Coord(k(), m());
+  }
+
+  /// Obtains a Coord<2> from GemmCoord
+  CUTLASS_HOST_DEVICE
+  Coord<2> kn() const {
+    return make_Coord(k(), n());
+  }
+
+  //
+  // Coord operators
+  //
+
+  /// Element-wise addition
+  CUTLASS_HOST_DEVICE
+  GemmCoord operator+(Base const& b) const {
+    return GemmCoord(Base::operator+(b));
+  }
+
+  /// Element-wise subtraction
+  CUTLASS_HOST_DEVICE
+  GemmCoord operator-(Base const& b) const {
+    return GemmCoord(Base::operator-(b));
+  }
+
+  /// Element-wise multiplication
+  CUTLASS_HOST_DEVICE
+  GemmCoord operator*(Base const& b) const {
+    return GemmCoord(Base::operator*(b));
+  }
+
+  /// Element-wise division
+  CUTLASS_HOST_DEVICE
+  GemmCoord operator/(Base const& b) const {
+    return GemmCoord(Base::operator/(b));
+  }
+
+  /// In-place addition
+  CUTLASS_HOST_DEVICE
+  GemmCoord& operator+=(Base const& b) {
+    Base::operator+=(b);
+    return *this;
+  }
+
+  /// In-place subtraction
+  CUTLASS_HOST_DEVICE
+  GemmCoord& operator-=(Base const& b) {
+    Base::operator-=(b);
+    return *this;
+  }
+
+  /// In-place multiplication
+  CUTLASS_HOST_DEVICE
+  GemmCoord& operator*=(Base const& b) {
+    Base::operator*=(b);
+    return *this;
+  }
+
+  /// In-place division
+  CUTLASS_HOST_DEVICE
+  GemmCoord& operator/=(Base const& b) {
+    Base::operator/=(b);
+    return *this;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace gemm
+} // namespace cutlass

cutlass/gemm/gemm_desc.h

@@ -0,0 +1,205 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements a software-pipelined efficient GEMM.
+*/
+#pragma once
+
+#include "cutlass/tensor_ref.h"
+#include "cutlass/gemm/gemm_coord.h"
+
+namespace cutlass {
+namespace gemm {
+
+/// GEMM problem description
+template <
+    /// Source accumulator matrix type
+    typename AType_,
+    /// Destination accumulator type
+    typename BType_,
+    /// Source accumulator matrix type
+    typename CType_,
+    /// Destination accumulator type
+    typename DType_,
+    /// Scalar type for alpha and beta
+    typename SType_,
+    /// Index type for dimensions and strides
+    typename Index_ = int
+> struct GemmDesc {
+  //
+  // Type definitions
+  //
+
+  /// Index type for dimensions and strides
+  typedef Index_ Index;
+
+  /// Source accumulator matrix type
+  typedef AType_ AType;
+
+  /// Tensor reference to A operand
+  typedef TensorRef<AType const, 2> TensorRefA;
+
+  /// Destination accumulator type
+  typedef BType_ BType;
+
+  /// Tensor reference to B operand
+  typedef TensorRef<BType const, 2> TensorRefB;
+
+  /// Source accumulator matrix type
+  typedef CType_ CType;
+
+  /// Tensor reference to C operand
+  typedef TensorRef<CType const, 2> TensorRefC;
+
+  /// Destination accumulator type
+  typedef DType_ DType;
+
+  /// Tensor reference to D operand
+  typedef TensorRef<DType, 2> TensorRefD;
+
+  /// Scalar type for alpha and beta
+  typedef SType_ SType;
+
+  //
+  // Data members
+  //
+
+  /// The dimensions of the GEMM.
+  GemmCoord problem_size;
+
+  /// The alpha scaling values.
+  SType alpha;
+
+  /// The source matrix A.
+  TensorRefA A;
+
+  /// batch stride for A operand
+  long long batch_stride_A;
+
+  /// The source matrix B.
+  TensorRefB B;
+
+  /// batch stride for B operand
+  long long batch_stride_B;
+
+  /// The beta scaling values.
+  SType beta;
+
+  /// The source matrix C.
+  TensorRefC C;
+
+  /// batch stride for C operand
+  long long batch_stride_C;
+
+  /// The destination matrix D.
+  TensorRefD D;
+
+  /// batch stride for D operand
+  long long batch_stride_D;
+
+  //
+  // Methods
+  //
+
+  /// Default ctor
+  CUTLASS_HOST_DEVICE
+  GemmDesc(): problem_size(0, 0, 0, 1), alpha(1), beta(0) {}
+
+  /// Constructor for basic GEMM with batch count = 1
+  CUTLASS_HOST_DEVICE
+  GemmDesc(Coord<3> _problem_size,
+           SType _alpha,
+           TensorRefA const &_A,
+           TensorRefB const &_B,
+           SType _beta,
+           TensorRefC const &_C,
+           TensorRefD const &_D
+  ):
+    problem_size(_problem_size[0], _problem_size[1], _problem_size[2], 1),
+    alpha(_alpha),
+    A(_A),
+    batch_stride_A(0),
+    B(_B),
+    batch_stride_B(0),
+    beta(_beta),
+    C(_C),
+    batch_stride_C(0),
+    D(_D),
+    batch_stride_D(0) {}
+
+  /// Constructor for basic GEMM with batch count = 1
+  CUTLASS_HOST_DEVICE
+  GemmDesc(GemmCoord _problem_size,
+           SType _alpha,
+           TensorRefA const &_A,
+           TensorRefB const &_B,
+           SType _beta,
+           TensorRefC const &_C,
+           TensorRefD const &_D
+  ):
+    problem_size(_problem_size.k(), _problem_size.n(), _problem_size.m(), 1),
+    alpha(_alpha),
+    A(_A),
+    batch_stride_A(0),
+    B(_B),
+    batch_stride_B(0),
+    beta(_beta),
+    C(_C),
+    batch_stride_C(0),
+    D(_D),
+    batch_stride_D(0) {
+
+    assert(_problem_size.batch() == 1);
+  }
+
+  /// Constructor for strided batch GEMM GEMM
+  CUTLASS_HOST_DEVICE
+  GemmDesc(GemmCoord _problem_size,
+           SType _alpha,
+           TensorRefA const &_A,
+           long long _batch_stride_A,
+           TensorRefB const &_B,
+           long long _batch_stride_B,
+           SType _beta,
+           TensorRefC const &_C,
+           long long _batch_stride_C,
+           TensorRefD const &_D,
+           long long _batch_stride_D
+  ):
+    problem_size(_problem_size),
+    alpha(_alpha),
+    A(_A),
+    batch_stride_A(_batch_stride_A),
+    B(_B),
+    batch_stride_B(_batch_stride_B),
+    beta(_beta),
+    C(_C),
+    batch_stride_C(_batch_stride_C),
+    D(_D),
+    batch_stride_D(_batch_stride_D) {}
+};
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_epilogue.h

@@ -0,0 +1,224 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements the epilogue phase of the GEMM kernel that efficiently updates global memory
+   with
+      the computed matrix product.
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/coord.h"
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmEpilogueTraits_>
+struct GemmEpilogue {
+  /// The traits class.
+  typedef GemmEpilogueTraits_ Traits;
+  /// The params.
+  typedef typename Traits::Params Params;
+  /// The shared storage.
+  typedef typename Traits::SharedStorage SharedStorage;
+
+  /// The output tile.
+  typedef typename Traits::OutputTile OutputTile;
+  /// The number of iterations.
+  typedef typename Traits::Iterations Iterations;
+  /// The accumulators.
+  typedef typename Traits::Accumulators Accumulators;
+  /// The scalar.
+  typedef typename Traits::Scalar Scalar;
+  /// The functor in charge of the math.
+  typedef typename Traits::Functor Functor;
+
+  /// We do not support 3D or 4D shapes.
+  static_assert(Iterations::kD == 1 && Iterations::kC == 1, "Unsupported 3D/4D shapes");
+
+  /// The iterator for C in global memory.
+  typedef typename Traits::GlobalLoadIteratorC GlobalLoadIteratorC;
+  /// The transformer for C.
+  typedef typename Traits::GlobalTransformerC GlobalTransformerC;
+  /// The transformer for D.
+  typedef typename Traits::GlobalTransformerD GlobalTransformerD;
+  /// The iterator for D in global memory.
+  typedef typename Traits::GlobalStoreIteratorD GlobalStoreIteratorD;
+  /// The iterator to store D in shared memory.
+  typedef typename Traits::SharedStoreIteratorD SharedStoreIteratorD;
+  /// The shared store transformer for D.
+  typedef typename Traits::SharedStoreTransformerD SharedStoreTransformerD;
+  /// The iterator to load D in shared memory.
+  typedef typename Traits::SharedLoadStreamD SharedLoadStreamD;
+
+  /// The index.
+  typedef typename Traits::Index Index;
+
+  /// The scalar for C.
+  typedef typename GlobalLoadIteratorC::Scalar ScalarC;
+  /// The scalar for D.
+  typedef typename GlobalStoreIteratorD::Scalar ScalarD;
+
+  /// Ctor.
+  CUTLASS_DEVICE GemmEpilogue(Params const& params_,
+                              SharedStorage& shared_storage_,
+                              Coord<3> const& _problem_size)
+      : params(params_), shared_storage(shared_storage_), problem_size(_problem_size), functor(params_.functor) {}
+
+  /// Execute the epilogue.
+  CUTLASS_DEVICE void epilogue(Accumulators& accumulators,
+                               Coord<3> const& block = make_Coord(0, 0, 0),
+                               int batch_id = 0) {
+    if (functor.source_required()) {
+      epilogue_with_or_without_beta<true>(accumulators, block, batch_id);
+    } else {
+      epilogue_with_or_without_beta<false>(accumulators, block, batch_id);
+    }
+  }
+
+  template <bool kSourceRequired>
+  CUTLASS_DEVICE void epilogue_with_or_without_beta(Accumulators& accumulators,
+                                                    Coord<3> const& block,
+                                                    int batch_id) {
+    // The C fragment.
+    typename GlobalLoadIteratorC::Fragment fragment_c;
+    // The transformed C fragment.
+    typename GlobalTransformerC::OutputFragment transformed_c;
+    CUTLASS_PRAGMA_UNROLL
+    for (int h = 0; h < Iterations::kH; ++h) {
+      // Compute pointer and predicate offsets for C and D global iterators.
+      int const pointer_offset =
+          ((params.iterator_d.inc_h * (GlobalStoreIteratorD::Iterations::kH - 1) +
+            params.iterator_d.inc_advance) *
+               Iterations::kW +
+           params.stride_h) *
+          h;
+
+      int const predicate_offset =
+          ((params.iterator_d.predicate_inc_h * (GlobalStoreIteratorD::Iterations::kH - 1) +
+            params.iterator_d.predicate_inc_advance) *
+               Iterations::kW +
+           Traits::Delta::kH) *
+          h;
+
+      // The iterator to load the elements of the C matrix.
+      GlobalLoadIteratorC global_load_iterator(
+          params.iterator_c, problem_size, block, pointer_offset, predicate_offset);
+
+      // update C pointer offset based on batch_id and batch_stride_offset
+      //global_load_iterator.add_pointer_offset(batch_id * params.batch_stride_offset_c);
+      global_load_iterator += make_Coord(batch_id, 0, 0);
+
+      // The transformer for C.
+      GlobalTransformerC transformer_c;
+      // The transformer for D.
+      GlobalTransformerD transformer_d;
+      // The iterator to store into the D matrix.
+      GlobalStoreIteratorD global_store_iterator(
+          params.iterator_d, problem_size, block, pointer_offset, predicate_offset);
+
+      // update D pointer offset based on batch_id and batch_stride_offset
+      //global_store_iterator.add_pointer_offset(batch_id * params.batch_stride_offset_d);
+      global_store_iterator += make_Coord(batch_id, 0, 0);
+
+      SharedStoreTransformerD shared_store_transformer;
+      typename SharedStoreTransformerD::OutputFragment shared_store_transformed_d;
+
+      SharedStoreIteratorD shared_store_iterator(
+          params.shared_store_iterator_d,
+          reinterpret_cast<typename SharedStoreIteratorD::Scalar*>(shared_storage.data()));
+
+      SharedLoadStreamD shared_load_stream(
+          params.shared_load_stream_d,
+          reinterpret_cast<typename SharedLoadStreamD::Scalar*>(shared_storage.data()));
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int w = 0; w < Iterations::kW; ++w) {
+        // Load the C matrix into fragment.
+        if (kSourceRequired) {
+          global_load_iterator.load_post_increment(fragment_c);
+        }
+
+        // Make sure we can write to shared memory.
+        shared_load_fence();
+
+        // Copy the accumulators to shared memory.
+        int const offset = (h * Iterations::kW + w) * SharedStoreIteratorD::Fragment::kElements;
+
+        shared_store_transformer.transform(accumulators, offset, shared_store_transformed_d);
+        shared_store_iterator.store_post_increment(shared_store_transformed_d);
+
+        // Make sure the data is in shared memory.
+        shared_store_fence();
+
+        // Copy the accumulators back to registers from shared memory.
+        shared_load_stream.copy();
+        shared_load_stream.commit();
+
+        // Do the math.
+        typename GlobalTransformerD::InputFragment fragment_d;
+
+        if (kSourceRequired) {
+          // Transform C fragment.
+          transformer_c.transform(fragment_c, transformed_c);
+          // Do the math.
+          functor.evaluate(shared_load_stream.fragment(), transformed_c, fragment_d);
+        } else {
+          functor.evaluate(shared_load_stream.fragment(), fragment_d);
+        }
+
+        // Transform D fragment.
+        typename GlobalTransformerD::OutputFragment global_transformed_d;
+        transformer_d.transform(fragment_d, global_transformed_d);
+
+        // Copy the results to global memory.
+        global_store_iterator.store_post_increment(global_transformed_d);
+      }
+    }
+  }
+
+  /// The memory fence for shared loads.
+  CUTLASS_DEVICE void shared_load_fence() { __syncthreads(); }
+
+  /// The memory fence for shared stores.
+  CUTLASS_DEVICE void shared_store_fence() { __syncthreads(); }
+
+  /// The params.
+  Params const& params;
+  /// The shared storage.
+  SharedStorage& shared_storage;
+  /// The dimensions of the GEMM.
+  Coord<3> problem_size;
+  // The functor.
+  Functor functor;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_epilogue_traits.h

@@ -0,0 +1,354 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines structural properties of the GEMM epilogue.
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/coord.h"
+#include "cutlass/gemm/gemm_global_stream.h"
+#include "cutlass/gemm/gemm_shared_stream.h"
+#include "cutlass/gemm/linear_scaling.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/tile_iterator.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The output tile.
+    typename OutputTile_,
+    /// The accumulators.
+    typename Accumulators_,
+    /// The iterator to load C from global memory.
+    typename GlobalLoadIteratorC_,
+    /// The transformer for C.
+    typename GlobalTransformerC_,
+    /// The transformer for D.
+    typename GlobalTransformerD_,
+    /// The iterator to store D to global memory.
+    typename GlobalStoreIteratorD_,
+    /// The iterator to store D to shared memory.
+    typename SharedStoreIteratorD_,
+    /// The shared store transformer for D.
+    typename SharedStoreTransformerD_,
+    /// The stream to load D from shared memory.
+    typename SharedLoadStreamD_,
+    /// The number of iterations in the epilogue.
+    typename Iterations_,
+    /// The iterations strides.
+    typename Delta_,
+    /// The functor to be used in the epilogue.
+    typename Functor_,
+    /// The index.
+    typename Index_ = int>
+struct GemmEpilogueTraits {
+  //
+  /// The output tile.
+  typedef OutputTile_ OutputTile;
+  /// The number of iterations.
+  /// The accumulators.
+  typedef Accumulators_ Accumulators;
+  /// The iterator for C in global memory.
+  typedef GlobalLoadIteratorC_ GlobalLoadIteratorC;
+  /// The transformer for C.
+  typedef GlobalTransformerC_ GlobalTransformerC;
+  /// The transformer for D.
+  typedef GlobalTransformerD_ GlobalTransformerD;
+  /// The iterator for D in global memory.
+  typedef GlobalStoreIteratorD_ GlobalStoreIteratorD;
+  /// The iterator to store D in shared memory.
+  typedef SharedStoreIteratorD_ SharedStoreIteratorD;
+  /// The shared store transformer for D.
+  typedef SharedStoreTransformerD_ SharedStoreTransformerD;
+  /// The stream to store D in shared memory.
+  typedef SharedLoadStreamD_ SharedLoadStreamD;
+  /// typedef typename GemmConfig::EpilogueIterations Iterations;
+  typedef Iterations_ Iterations;
+  /// The iterations strides.
+  typedef Delta_ Delta;
+
+  /// The functor in charge of the math.
+  typedef Functor_ Functor;
+  /// The index.
+  typedef Index_ Index;
+
+  /// We do not support 3D or 4D shapes.
+  static_assert(Iterations::kD == 1 && Iterations::kC == 1, "Unsupported 3D/4D shapes");
+
+  /// The scalar.
+  typedef typename Functor::Scalar Scalar;
+  /// The scalar for C.
+  typedef typename GlobalLoadIteratorC::Scalar ScalarC;
+  /// The scalar for D.
+  typedef typename GlobalStoreIteratorD::Scalar ScalarD;
+
+  /// The params.
+  struct Params {
+    /// The strides for H and W in the different iterations of the epilogue.
+    Index stride_h, stride_w;
+    /// The params for the C iterator.
+    typename GlobalLoadIteratorC::Params iterator_c;
+    /// The params for the D global iterator.
+    typename GlobalStoreIteratorD::Params iterator_d;
+    /// The params for the D shared store iterator.
+    typename SharedStoreIteratorD::Params shared_store_iterator_d;
+    /// The params for the D shared load stream.
+    typename SharedLoadStreamD::Params shared_load_stream_d;
+    /// The functor params.
+    typename Functor::Params functor;
+
+    /// Setup the params.
+    template <typename GemmDesc_>
+    CUTLASS_HOST_DEVICE int initialize(GemmDesc_ const& desc) {
+
+      // The parameters for the functor.
+      int error_code = functor.initialize(desc);
+      if (error_code) {
+        return error_code;
+      }
+
+      // At the end of the H iteration, we jump over a number of columns.
+      this->stride_h = desc.D.leading_dim() * Delta::kH;
+      // Nothing to do here.
+      this->stride_w = 0;
+      // Setup the params for the global memory iterator for C.
+      error_code = iterator_c.initialize(desc.C.data(),
+                                         desc.batch_stride_C,
+                                         desc.C.leading_dim(),
+                                         desc.problem_size[1],
+                                         stride_w,
+                                         Delta::kW);
+      if (error_code) {
+        return error_code;
+      }
+
+      // Setup the params for the global memory iterator for D.
+      return iterator_d.initialize(desc.D.data(),
+                                   desc.batch_stride_D,
+                                   desc.D.leading_dim(),
+                                   desc.problem_size[1],
+                                   stride_w,
+                                   Delta::kW);
+    }
+  };
+
+  /// The shared memory storage to exchange data.
+  union StreamSharedStorage {
+    // The storage for the store iterator.
+    typename SharedStoreIteratorD::SharedStorage store;
+    // The storage for the store iterator.
+    typename SharedLoadStreamD::SharedStorage load;
+  };
+
+  /// The shared memory to swizzle the data in the epilogue.
+  struct SharedStorage {
+    // The storage for the shared stream D.
+    StreamSharedStorage shared_stream;
+
+    //
+    //
+    //
+
+    CUTLASS_DEVICE
+    ScalarD* data() { return reinterpret_cast<ScalarD*>(&shared_stream.load); }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_, typename EpilogueFunctor_, typename Index_ = int>
+struct GemmEpilogueTraitsHelper {
+  /// The scalar.
+  typedef typename EpilogueFunctor_::Scalar Scalar;
+  /// The output tile.
+  typedef typename GemmConfig_::OutputTile OutputTile;
+
+  /// The number of iterations in the epilogue.
+  typedef Shape<1,
+                GemmConfig_::MultiplyAdd::AccumulatorsPerThread::kH /
+                    GemmConfig_::kAccumulatorsPerLdsB,
+                GemmConfig_::kAccumulatorsPerLdsB>
+      Iterations;
+  // The iteration strides in the H/W dimension.
+  typedef Shape<0,
+                GemmConfig_::kAccumulatorsPerLdsB*(
+                    GemmConfig_::Warps::kH* GemmConfig_::MultiplyAdd::ThreadsPerWarp::kH - 1),
+                0>
+      Delta;
+  /// The functor to do the math in the epilogue.
+  typedef EpilogueFunctor_ Functor;
+
+  /// The traits class to build the iterator to store to shared memory for D.
+  typedef GemmSharedStoreTileDTraits<
+      // The pointer is float.
+      // typename Functor::Scalar,
+      // Functor::Scalar is alpha, beta type, in mixed precision, alpha and beta may not be the same with accumulation.
+      // In this case Functor::ScalarAccum is needed
+      typename Functor::ScalarAccum,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The number of scalars per STS.
+      GemmConfig_::kScalarsPerStsD,
+      // The skew -- 128 / sizeof(ScalarD) / kScalarsPerStsD is the number of threads involved in
+      // a single STS. We divide by 2 as our objective is to add a skew to the odd threads to
+      // avoid bank conflicts between odd and even threads.
+      128 / sizeof(typename GemmConfig_::ScalarD) / GemmConfig_::kScalarsPerStsD / 2 *
+          GemmConfig_::kScalarsPerStsD>
+      SharedStoreTileTraits;
+
+  /// The iterator to store D to shared memory.
+  typedef TileStoreIterator<SharedStoreTileTraits,
+                            typename SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorD;
+
+  /// The shared store transformer for D.
+  typedef Copy<typename SharedStoreIteratorD::Fragment> SharedStoreTransformerD;
+
+  /// The traits class to build the iterator to load from shared memory for D.
+  typedef GemmSharedLoadTileDTraits<
+      // The pointer is float.
+      // typename Functor::Scalar,
+      // Functor::Scalar is alpha, beta type, in mixed precision, alpha and beta may not be the same with accumulation.
+      // In this case Functor::ScalarAccum is needed
+      typename Functor::ScalarAccum,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The number of columns of the output tile written by iteration.
+      GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
+      // The number of scalars per LDS.
+      GemmConfig_::kScalarsPerLdsD,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+
+  /// The iterator to load D from shared memory.
+  typedef TileLoadIterator<SharedLoadTileTraits,
+                           typename SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorD;
+  /// The stream to load D.
+  typedef SharedLoadStream<SharedLoadIteratorD> SharedLoadStreamD;
+
+  /// The traits class to build the iterator to load data from global memory for C^N.
+  typedef GemmGlobalTileCdTraits<
+      // The pointer is float const.
+      typename GemmConfig_::ScalarC const,
+      // The tile has size (N / Iterations)xM in GEMM's terminology.
+      Shape<1,
+            GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
+            GemmConfig_::OutputTile::kW>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // How many elements do we jump over at each iteration?
+      Iterations::kW,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgC>
+      GlobalLoadTileTraits;
+
+  /// The iterator to load C.
+  typedef GemmGlobalIteratorCd<GlobalLoadTileTraits, Index_> GlobalLoadIteratorC;
+  /// The transformer for C.
+  typedef Copy<typename GlobalLoadIteratorC::Fragment> GlobalTransformerC;
+
+  /// The traits class to build the iterator to store data to global memory for D^N.
+  typedef GemmGlobalTileCdTraits<
+      // The pointer is float.
+      typename GemmConfig_::ScalarD,
+      // The tile has size (N / Iterations)xM in GEMM's terminology.
+      Shape<1,
+            GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
+            GemmConfig_::OutputTile::kW>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // How many elements do we jump over at each iteration?
+      Iterations::kW,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerStgD>
+      GlobalStoreTileTraits;
+
+  /// The iterator to store D.
+  typedef GemmGlobalIteratorCd<GlobalStoreTileTraits, Index_> GlobalStoreIteratorD;
+  /// The transformer for D.
+  typedef Copy<typename GlobalStoreIteratorD::Fragment> GlobalTransformerD;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The GEMM config.
+    typename GemmConfig_,
+    /// The epilogue functor to do the math in the epilogue.
+    typename EpilogueFunctor_,
+    /// The index.
+    typename Index_ = int,
+    /// The helper to create the traits class.
+    typename Helper_ = GemmEpilogueTraitsHelper<GemmConfig_, EpilogueFunctor_, Index_> >
+struct SimplifiedGemmEpilogueTraits : public GemmEpilogueTraits<
+                                          // The output tile.
+                                          typename GemmConfig_::OutputTile,
+                                          // The accumulators.
+                                          typename GemmConfig_::Accumulators,
+                                          // The global iterator for C.
+                                          typename Helper_::GlobalLoadIteratorC,
+                                          // The transformer for C.
+                                          typename Helper_::GlobalTransformerC,
+                                          // The transformer for D.
+                                          typename Helper_::GlobalTransformerD,
+                                          // The global iterator for D.
+                                          typename Helper_::GlobalStoreIteratorD,
+                                          // The iterator to store D to shared memory.
+                                          typename Helper_::SharedStoreIteratorD,
+                                          // The shared store transformer for D.
+                                          typename Helper_::SharedStoreTransformerD,
+                                          // The stream to load D from shared memory.
+                                          typename Helper_::SharedLoadStreamD,
+                                          // The number of iterations.
+                                          typename Helper_::Iterations,
+                                          // The strides between iterations.
+                                          typename Helper_::Delta,
+                                          // The functor to be used in the epilogue.
+                                          EpilogueFunctor_,
+                                          // The index.
+                                          Index_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_global_stream.h

@@ -0,0 +1,225 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements efficient loading of the thread block-level tile from global memory and
+   storing
+      to shared memory.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/convert.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/tile_allocation.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// Identifies multiplicand
+    GemmOperand::Kind Operand,
+    /// The load iterator.
+    typename LoadIterator_,
+    /// The store iterator to copy to shared memory.
+    typename StoreIterator_,
+    /// The transformer to be applied after the data has been copied from global memory.
+    typename Transformer_>
+
+struct GlobalLoadStream {
+  /// Indicates the type of GEMM operand
+  static GemmOperand::Kind const kOperand = Operand;
+  /// The load iterator.
+  typedef LoadIterator_ LoadIterator;
+  /// The transformer.
+  typedef Transformer_ Transformer;
+  /// The store iterator to write to shared memory.
+  typedef StoreIterator_ StoreIterator;
+
+  /// The fragment that is copied from shared memory.
+  typedef typename LoadIterator::Fragment FetchedFragment;
+  /// The fragment that is obtained after the transformation by the transformer.
+  typedef typename Transformer::OutputFragment TransformedFragment;
+  /// Make sure the fragments match.
+  static_assert((platform::is_same<FetchedFragment, typename Transformer::InputFragment>::value),
+                "");
+  /// The output fragment.
+  typedef TransformedFragment Fragment;
+  /// Make sure the transformed fragment is the same as the store fragment.
+  static_assert((platform::is_same<TransformedFragment, typename StoreIterator::Fragment>::value),
+                "");
+
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = LoadIterator::kLayout;
+  /// The scalar type of the iterator.
+  typedef typename LoadIterator::Scalar Scalar;
+  /// The pointer.
+  typedef typename LoadIterator::Pointer Pointer;
+  /// The index.
+  typedef typename LoadIterator::Index Index;
+  /// The tile
+  typedef typename LoadIterator::Tile Tile;
+
+  /// Shared memory allocation for the tile
+  typedef TileAllocation<typename StoreIterator::Scalar, typename StoreIterator::Tile>
+      ThreadblockTileStorage;
+
+  /// Tensor reference to threadblock tile
+  typedef typename ThreadblockTileStorage::TensorRef ThreadblockTileRef;
+
+  /// The params.
+  struct Params {
+    // The load iterator.
+    typename LoadIterator::Params load_iterator;
+    // The store iterator.
+    typename StoreIterator::Params store_iterator;
+    // Offset to residue.
+    Index offset_to_residue;
+
+    /// Setup the params.
+    CUTLASS_HOST_DEVICE int initialize(Pointer pointer,
+                                       long long batch_stride,
+                                       Index ldm,
+                                       Index _offset_to_residue) {
+
+      offset_to_residue = _offset_to_residue;
+      int error_code = load_iterator.initialize(pointer, batch_stride, ldm);
+      if (error_code) {
+        return error_code;
+      }
+      return store_iterator.initialize();
+    }
+  };
+
+  /// Contains private storage in shared memory needed by the objects within this class. Note,
+  /// this is *NOT* the shared memory allocation for the GEMM threadblock tile. That necessarily
+  /// exists outside this class, as it is also needed by the warp-level shared=>RF stream.
+  struct SharedStorage {};
+
+  //
+  // Static member functions
+  //
+
+  /// Maps a coordinate in the GEMM's (K, N, M) coordinate system to global memory
+  CUTLASS_DEVICE static Coord<3> project_coordinate(Coord<3> const& coord, Index d_offset = 0) {
+    bool const kKstrided =
+        GemmMultiplicandTraits<typename LoadIterator::Tile, kOperand, kLayout>::kKstrided;
+    Coord<3> tile_coord = ProjectOperand<kOperand, kKstrided>::project(coord);
+    return make_Coord(
+        tile_coord[0] + d_offset, tile_coord[1], tile_coord[2] / LoadIterator::Tile::kC);
+  }
+
+  /// Ctor.
+  CUTLASS_DEVICE GlobalLoadStream(
+      Params const& _params,
+      SharedStorage& shared_storage,
+      ThreadblockTileRef const& threadblock_tile_ref,
+      Coord<3> const bounds,
+      Coord<3> const& _threadblock_offset)
+      : params(_params),
+        multiplicand_bounds(project_coordinate(bounds, 1)),
+        threadblock_offset(project_coordinate(_threadblock_offset)),
+        load_iterator(params.load_iterator,
+          project_coordinate(bounds, 1), /*multiplicant_bounds*/
+          project_coordinate(_threadblock_offset) /*threablock_offset*/),
+        transformer(),
+        store_iterator(params.store_iterator, threadblock_tile_ref.data())
+  {
+    load_iterator.initialize_predicates(multiplicand_bounds, threadblock_offset);
+    fetched_fragment.clear();
+  }
+
+
+  /// Load the data from shared memory to the fetch fragment.
+  CUTLASS_DEVICE void copy() { load_iterator.load_post_increment(fetched_fragment); }
+
+  /// Commit the data.
+  CUTLASS_DEVICE void commit() {
+    transformer.transform(fetched_fragment, transformed_fragment);
+    store_iterator.store_post_increment(transformed_fragment);
+    store_iterator.inc_stage();
+  }
+
+  /// Execute the residue code.
+  CUTLASS_DEVICE void residue(Index k, bool skip_clear = false) {
+    load_iterator.residue(k);
+    if (!skip_clear) {
+      fetched_fragment.clear();
+    }
+  }
+
+  /// Move to the residue portion.
+  CUTLASS_DEVICE void move_to_residue(Index k, Index kTileK) {
+    Index kResidue = k % kTileK;
+    if (kResidue) {
+      residue(kResidue);
+    }
+    load_iterator.add_pointer_offset(params.offset_to_residue * load_iterator.stride_advance());
+  }
+
+  /// Rollback to the beginning of the first tile
+  CUTLASS_DEVICE void rollback(void) {
+    load_iterator.initialize_predicates(multiplicand_bounds, threadblock_offset);
+
+    int const kBlock = kOperand == GemmOperand::kA
+                           ? (kLayout == MatrixLayout::kColumnMajor ? Tile::kH : Tile::kW)
+                           : (kLayout == MatrixLayout::kRowMajor ? Tile::kH : Tile::kW);
+
+    load_iterator.add_pointer_offset(-(params.offset_to_residue + kBlock) *
+                                     load_iterator.stride_advance());
+  }
+
+  /// Adds a Coord<3> to the underlying global load iterator
+  CUTLASS_DEVICE GlobalLoadStream &operator+=(Coord<3> const &offset) {
+    load_iterator += offset;
+    return *this;
+  }
+
+  //
+  // Data members
+  //
+
+  /// Parameters
+  Params params;
+  /// Multiplicand bounds
+  Coord<3> multiplicand_bounds;
+  /// Threadblock offset
+  Coord<3> threadblock_offset;
+  /// The iterator.
+  LoadIterator load_iterator;
+  /// The fragment to fetch from shared memory.
+  FetchedFragment fetched_fragment;
+  /// The transformer.
+  Transformer transformer;
+  /// The fragment to convert the data after it has been fetched from shared memory.
+  TransformedFragment transformed_fragment;
+  /// The store iterator.
+  StoreIterator store_iterator;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_global_tile.h

@@ -0,0 +1,640 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines iterators for efficiently loading and storing to global memory.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/util/platform.h"
+
+#include "cutlass/gemm/gemm_operand.h"
+#include "cutlass/matrix_traits.h"
+#include "cutlass/predicate_vector.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/tile_iterator.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// The following functor reshapes a tile of threads to match a tile of data. The idea is that when
+// the user wants to build the iterator traits, he/she may want to specify the tile independently
+// from the number of scalars loaded/stored per instruction. For example, in the row-major version
+// with a tile of size 128x8 - the user may want to that the iterator works with 32x8 threads if
+// each thread loads 1 scalar per LDG. If the user changes to 4 scalars per LDG, then the tile of
+// threads has to change. The code below detects that and correct the code automatically - it is
+// a helper when the user does not specify the right configuration.
+
+template <typename Tile_, typename Threads_, bool = (Tile_::kW < Threads_::kW)>
+struct ReshapeThreads {
+  typedef Threads_ Threads;
+};
+
+template <typename Tile_, typename Threads_>
+struct ReshapeThreads<Tile_, Threads_, true> {
+  typedef Shape<Threads_::kD, Threads_::kH * Threads_::kW / Tile_::kW, Tile_::kW, 1> Threads;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <GemmOperand::Kind kOperand_,
+          MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename Tile_,
+          typename Threads_,
+          int kAccessSize_>
+struct GemmGlobalTileTraits {
+  /// Identity of the operand
+  static GemmOperand::Kind const kOperand = kOperand_;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = kLayout_;
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The number of scalars per LDG/STG.
+  static int const kAccessSize = kAccessSize_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kGlobal;
+  /// The tile shape
+  typedef Tile_ Tile;
+  /// The vectorized tile shape
+  typedef typename ReshapeTile<Tile_, kAccessSize_>::Tile VectorizedTile;
+  /// The threads shape
+  typedef typename ReshapeThreads<VectorizedTile, Threads_>::Threads Threads;
+  /// The relative offset between two elements in the H/W dimension in adjacent threads.
+  typedef Shape<1, 1, VectorizedTile::kC> ThreadsDelta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, Threads::kH, Threads::kW * kAccessSize> Delta;
+
+  /// Strides for immediate offset computation
+  typedef Shape<0, 0, Threads::kW * ThreadsDelta::kW, kAccessSize> ImmediateOffsetStrides;
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<1,
+                VectorizedTile::kH / Threads::kH,
+                VectorizedTile::kW / Threads::kW,
+                VectorizedTile::kC / kAccessSize>
+      Iterations;
+
+  typedef GemmMultiplicandTraits<Tile, kOperand, kLayout> MultiplicandTraits;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int thread_offset_h = threadIdx.x / Threads::kW * ThreadsDelta::kH;
+      int thread_offset_w = threadIdx.x % Threads::kW * ThreadsDelta::kW;
+
+      return make_Coord(0, thread_offset_h, thread_offset_w, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, typename Tile_, typename Threads_, int kStrideH_, int kAccessSize_>
+struct GemmGlobalTileCdTraits : public GemmGlobalTileTraits<GemmOperand::kC,
+                                                            MatrixLayout::kColumnMajor,
+                                                            Scalar_,
+                                                            Tile_,
+                                                            Threads_,
+                                                            kAccessSize_> {
+  /// The base class.
+  typedef GemmGlobalTileTraits<GemmOperand::kC,
+                               MatrixLayout::kColumnMajor,
+                               Scalar_,
+                               Tile_,
+                               Threads_,
+                               kAccessSize_>
+      Base;
+
+  /// The stride in the H dimension.
+  static int const kStrideH = kStrideH_;
+  /// Override the strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Base::Delta::kW, Base::Delta::kC> Delta;
+
+  typedef typename Base::Iterations Iterations;
+
+  typedef typename Base::Threads Threads;
+
+  typedef typename Base::ThreadsDelta ThreadsDelta;
+
+  typedef typename Base::ImmediateOffsetStrides ImmediateOffsetStrides;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int thread_offset_h = threadIdx.x / Threads::kW * kStrideH * Iterations::kH;
+      int thread_offset_w = threadIdx.x % Threads::kW * ThreadsDelta::kW;
+
+      return make_Coord(0, thread_offset_h, thread_offset_w, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename TileTraits_, typename Index_ = int>
+struct GemmGlobalIteratorAb
+    : public TileLoadIterator<TileTraits_,
+                              typename TileTraits_::Scalar,
+                              TileTraits_::MultiplicandTraits::kKstrided ? IteratorAdvance::kH
+                                                                         : IteratorAdvance::kW,
+                              MemorySpace::kGlobal,
+                              Index_> {
+  /// This class.
+  typedef GemmGlobalIteratorAb<TileTraits_, Index_> This_;  /// The base class.
+  typedef TileLoadIterator<TileTraits_,
+                           typename TileTraits_::Scalar,
+                           TileTraits_::MultiplicandTraits::kKstrided ? IteratorAdvance::kH
+                                                                      : IteratorAdvance::kW,
+                           MemorySpace::kGlobal,
+                           Index_>
+      Base;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = TileTraits_::kLayout;
+  /// The tile
+  typedef typename TileTraits_::Tile Tile;
+  /// Fragment type loaded by the iterator
+  typedef typename Base::Fragment Fragment;
+  /// The scalar.
+  typedef typename TileTraits_::Scalar Scalar;
+  /// The threads.
+  typedef typename TileTraits_::Threads Threads;
+  /// The index.
+  typedef Index_ Index;
+  /// The thread offset
+  typedef typename TileTraits_::ThreadOffset ThreadOffset;
+  /// Specifies in which dimension post-increment accesses advance.
+  static IteratorAdvance::Kind const kAdvance = Base::kAdvance;
+
+  typedef cutlass::PredicateVector<ShapeCount<typename Base::Iterations>::kCount> PredicateVector;
+
+  /// Iterator parameters type
+  typedef typename Base::Params BaseParams;
+
+  struct Params : public BaseParams {
+    /// Initializes params to load a strip-mined tile, given pointer and stride_h.
+    CUTLASS_HOST_DEVICE int initialize(Scalar const* ptr,
+                                       long long stride_d,
+                                       Index stride_h) {
+      Index inc_d = 0;
+      Index inc_advance = 0;
+      // Move by some columns for each iteration in the H dimension.
+      Index inc_h = Base::Delta::kH * stride_h;
+
+      // Move by some more columns in the number of iterations if the D dimension is > 1.
+      if (Base::Delta::kD > 0) {
+        inc_d = Base::Delta::kD * stride_h - (Base::Iterations::kH - 1) * inc_h;
+      }
+
+      // Move to the beginning of the next iteration.
+      if (kAdvance == IteratorAdvance::kH && Base::Delta::kD > 0) {
+        inc_advance = inc_d;
+      } else if (kAdvance == IteratorAdvance::kH) {
+        inc_advance = inc_h;
+      } else if (Base::Delta::kD > 0) {
+        inc_advance = (Base::Iterations::kW + 0) * ShapeCount<typename Base::Delta>::kWc -
+                      (Base::Iterations::kH - 1) * inc_h -
+                      (Base::Iterations::kD - 1) * Base::Delta::kD * stride_h;
+      } else {
+        inc_advance = (Base::Iterations::kW + 0) * ShapeCount<typename Base::Delta>::kWc -
+                      (Base::Iterations::kH - 1) * inc_h;
+      }
+
+      Base::Params::initialize(
+          ptr, stride_d, stride_h, 1, inc_d, inc_h, 0, inc_advance);
+      return 0;
+    }
+  };
+
+  /// Offset of an individual lane from the start of the tile
+  Coord<4> thread_offset;
+  /// The parameters
+  Params params;
+  /// The predicates.
+  PredicateVector predicates;
+
+  CUTLASS_HOST_DEVICE void initialize_predicates(const Coord<3>& bounds, const Coord<3>& block_offset) {
+    // Setup the masks to control loads.
+    predicates.fill(0);
+
+    // Fill in the bits of the predicate vector.
+    for (int d = 0; d < Base::Iterations::kD; ++d) {
+      for (int h = 0; h < Base::Iterations::kH; ++h) {
+        for (int w = 0; w < Base::Iterations::kW; ++w) {
+          for (int c = 0; c < Base::Iterations::kC; ++c) {
+            bool flag = w * Base::Delta::kW + thread_offset[2] + block_offset[2] < bounds[2];
+            if (kAdvance == IteratorAdvance::kH) {
+              flag =
+                  flag &&
+                  (h * Base::Delta::kH + d * Base::Delta::kD) + thread_offset[1] + block_offset[1] <
+                      bounds[1];
+            } else {
+              flag = flag && (h * Base::Delta::kH) + thread_offset[1] + block_offset[1] < bounds[1];
+            }
+            int const bit = ComputeOffsetFromShape<typename Base::Iterations>::get(d, h, w, c);
+            predicates.set(bit, flag);
+          }
+        }
+      }
+    }
+  }
+
+  /// Ctor.
+  CUTLASS_HOST_DEVICE GemmGlobalIteratorAb(Params const& _params,
+                                           const Coord<3>& bounds,
+                                           const Coord<3>& threadblock_offset,
+                                           ThreadOffset thread_offset_func = ThreadOffset())
+      : params(_params) {
+    thread_offset = thread_offset_func();
+    // Setup the pointer.
+    params.pointer += ((threadblock_offset[1] + thread_offset[1]) * params.stride_h +
+                       (threadblock_offset[2] + thread_offset[2]));
+
+  }
+
+  /// Increment the pointer in the W dimension.
+  CUTLASS_HOST_DEVICE void inc_w() { Base::inc_w(); }
+  /// Increment the pointer in the H dimension.
+  CUTLASS_HOST_DEVICE void inc_h() { params.pointer += params.inc_h; }
+  /// Increment the pointer in the D dimension.
+  CUTLASS_HOST_DEVICE void inc_d() { params.pointer += params.inc_d; }
+  /// Increment the pointer to move to the next iteration.
+  CUTLASS_HOST_DEVICE void inc_advance() { params.pointer += params.inc_advance; }
+
+  /// Loads a single fragment element from memory
+  CUTLASS_HOST_DEVICE void load_element(
+      typename Base::AccessType& value, int d, int h, int w, int c) const {
+    int const offset =
+        ComputeOffsetFromStrides<typename Base::ImmediateOffsetStrides>::get(0, 0, w, c);
+    Load<Scalar,
+         Base::kAccessSize,
+         Base::kMemorySpace,
+         Base::kFragmentElementType,
+         typename Base::FragmentElement,
+         Base::Tile::kW,
+         Base::kAccessSize * sizeof(Scalar)>::load(value, params.pointer, offset);
+  }
+
+  /// That's the residue! Update the predicates.
+  CUTLASS_HOST_DEVICE void residue(Index k) {
+    // The coordinates of the thread.
+    Index block_h = thread_offset[1];
+    // The contiguous dimension.
+    Index block_w = thread_offset[2];
+
+    // Update the predicate vector.
+    for (int d = 0; d < Base::Iterations::kD; ++d) {
+      for (int h = 0; h < Base::Iterations::kH; ++h) {
+        for (int w = 0; w < Base::Iterations::kW; ++w) {
+          for (int c = 0; c < Base::Iterations::kC; ++c) {
+            Index offset = 0;
+            if (kAdvance == IteratorAdvance::kH) {
+              offset += block_h + h * Base::Delta::kH + d * Base::Delta::kD;
+            } else {
+              offset += block_w + w * Base::Delta::kW;
+            }
+
+            int const bit = ComputeOffsetFromShape<typename Base::Iterations>::get(d, h, w, c);
+            if (offset >= k) {
+              predicates.set(bit, false);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  /// Is the valid?
+  CUTLASS_HOST_DEVICE bool valid(int d, int h, int w, int c) const {
+    int const bit = ComputeOffsetFromShape<typename Base::Iterations>::get(d, h, w, c);
+    return predicates[bit];
+  }
+
+  /// Adds a vector offset to the iterator
+  CUTLASS_HOST_DEVICE GemmGlobalIteratorAb & operator+=(Coord<3> const &offset) {
+
+    long long _offset = offset.template dot<long long>(
+      make_Coord(params.stride_d, params.stride_h, params.stride_w)
+    );
+
+    params.pointer += _offset;
+    return *this;
+  }
+
+  CUTLASS_HOST_DEVICE void add_pointer_offset(Index offset) { params.pointer += offset; }
+
+  CUTLASS_HOST_DEVICE Index stride_advance(void) {
+    Index stride = params.stride_h;
+    if (kAdvance == IteratorAdvance::kW) {
+      stride = params.stride_w;
+    }
+    return stride;
+  }
+
+  template <typename Fragment>
+  CUTLASS_HOST_DEVICE void load_post_increment(Fragment& fragment) {
+    typename Base::FragmentIterator frag_iterator(fragment);
+    for (int d = 0; d < Base::Iterations::kD; ++d) {
+      for (int h = 0; h < Base::Iterations::kH; ++h) {
+        for (int w = 0; w < Base::Iterations::kW; ++w) {
+          for (int c = 0; c < Base::Iterations::kC; ++c) {
+            if (valid(d, h, w, c)) {
+              load_element(
+                  reinterpret_cast<typename Base::AccessType&>(frag_iterator.at(d, h, w, c)),
+                  d,
+                  h,
+                  w,
+                  c);
+            }
+          }
+          if (w < Base::Iterations::kW - 1) {
+            inc_w();
+          }
+        }
+        if (h < Base::Iterations::kH - 1) {
+          inc_h();
+        }
+      }
+      if (d < Base::Iterations::kD - 1) {
+        inc_d();
+      }
+    }
+    inc_advance();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename TileTraits_, typename Index_ = int>
+struct GemmGlobalIteratorCd : public TileIteratorBase<TileTraits_,
+                                                      typename TileTraits_::Scalar,
+                                                      IteratorAdvance::kH,
+                                                      MemorySpace::kGlobal,
+                                                      Index_> {
+  /// This class.
+  typedef GemmGlobalIteratorCd<TileTraits_, Index_> This_;
+  /// The base class.
+  typedef TileIteratorBase<TileTraits_,
+                           typename TileTraits_::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kGlobal,
+                           Index_>
+      Base;
+
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = TileTraits_::kLayout;
+
+  /// The scalar.
+  typedef typename TileTraits_::Scalar Scalar;
+  /// The pointer.
+  typedef typename TileTraits_::Pointer Pointer;
+  /// The threads.
+  typedef typename TileTraits_::Threads Threads;
+  /// The index.
+  typedef Index_ Index;
+  /// The thread offset
+  typedef typename TileTraits_::ThreadOffset ThreadOffset;
+
+  /// The params.
+  struct Params {
+    /// The pointer.
+    Pointer pointer;
+    /// The stride in the D dimension
+    long long stride_d;
+    /// The stride in the H dimension to setup the thread in the block.
+    Index stride_h;
+    /// The strides to increment the pointer.
+    Index inc_advance, inc_h;
+    /// The strides to increment the predicate offset
+    Index predicate_inc_advance, predicate_inc_h;
+    /// The column offset to compute the predicate for the columns.
+    Index predicate_offset;
+
+    /// Setup the params.
+    CUTLASS_HOST_DEVICE int initialize(Pointer pointer,
+                                       long long batch_stride,
+                                       Index ldm,
+                                       Index bound,
+                                       Index epilogue_stride_w,
+                                       Index epilogue_delta_w) {
+      // The pointer.
+      this->pointer = pointer;
+      // Stride per batch
+      stride_d = batch_stride;
+      // Each column of the matrix.
+      stride_h = TileTraits_::ThreadsDelta::kH * ldm;
+      // Each thread output 1 column per iteration. The stride between columns is given by the
+      // number of scalars that are loaded per LDS for B.
+      inc_h = ldm * TileTraits_::kStrideH;
+      inc_advance =
+          (ldm - ldm * TileTraits_::kStrideH * (Base::Iterations::kH - 1)) + epilogue_stride_w;
+
+      predicate_offset = bound;
+      predicate_inc_h = TileTraits_::kStrideH;
+      predicate_inc_advance =
+          -((TileTraits_::kStrideH * (Base::Iterations::kH - 1) - 1) + epilogue_delta_w);
+
+      return 0;
+    }
+  };
+
+  /// Parameters.
+  Params params;
+  /// Offset of an individual lane from the start of the tile
+  Coord<4> thread_offset;
+  /// The predicates for the row.
+  cutlass::PredicateVector<Base::Iterations::kW> predicates;
+
+  /// Ctor.
+  CUTLASS_HOST_DEVICE GemmGlobalIteratorCd(Params const& _params,
+                                           const Coord<3>& bounds,
+                                           const Coord<3>& block_offset,
+                                           ThreadOffset thread_offset_func = ThreadOffset())
+      : params(_params) {
+    thread_offset = thread_offset_func();
+    // Prepare the vector of predicates.
+    for (int i = 0; i < Base::Iterations::kW; ++i) {
+      predicates.set(i, thread_offset[2] + i * Base::Delta::kW < bounds[2]);
+    }
+  }
+
+  /// Ctor.
+  CUTLASS_HOST_DEVICE GemmGlobalIteratorCd(Params const& _params,
+                                           const Coord<3>& bounds,
+                                           const Coord<3>& block,
+                                           int offset = 0,
+                                           int pred_offset = 0,
+                                           ThreadOffset thread_offset_func = ThreadOffset())
+      : params(_params) {
+    thread_offset = thread_offset_func();
+    // Each warp works on a different column of the tile.
+    int const h = thread_offset[1] + block[1];
+    // Each lane writes a different element.
+    int const w = thread_offset[2] + block[2];
+    // Setup the pointer.
+    params.pointer += ((h * params.stride_h + w) + offset);
+
+    // Prepare the vector of predicates.
+    for (int i = 0; i < Base::Iterations::kW; ++i) {
+      predicates.set(i, w + i * Base::Delta::kW < bounds[2]);
+    }
+    params.predicate_offset -= (h + pred_offset);
+  }
+
+  /// Increment the pointer in the C dimension.
+  CUTLASS_HOST_DEVICE void inc_c() {}
+  /// Increment the pointer in the W dimension.
+  CUTLASS_HOST_DEVICE void inc_w() {}
+  /// Increment the pointer in the H dimension.
+  CUTLASS_HOST_DEVICE void inc_h() {
+    params.pointer += params.inc_h;
+    params.predicate_offset -= params.predicate_inc_h;
+  }
+  /// Increment the pointer in the D dimension.
+  CUTLASS_HOST_DEVICE void inc_d() {}
+  /// Increment the pointer to move to the next iteration.
+  CUTLASS_HOST_DEVICE void inc_advance() {
+    params.pointer += params.inc_advance;
+    params.predicate_offset -= params.predicate_inc_advance;
+  }
+
+  /// Adds a vector offset to the iterator
+  CUTLASS_HOST_DEVICE GemmGlobalIteratorCd & operator+=(Coord<3> const &offset) {
+    long long _offset = offset.template dot<long long>(
+      make_Coord(params.stride_d, params.stride_h, 1)
+    );
+    params.pointer += _offset;
+    return *this;
+  }
+
+  /// Loads a single fragment element from memory.
+  CUTLASS_HOST_DEVICE void load_element(
+      typename Base::AccessType& value, int d, int h, int w, int c) const {
+    int const offset =
+        ComputeOffsetFromStrides<typename Base::ImmediateOffsetStrides>::get(d, h, w, c);
+    Load<Scalar,
+         Base::kAccessSize,
+         Base::kMemorySpace,
+         Base::kFragmentElementType,
+         typename Base::FragmentElement,
+         Base::Tile::kW,
+         Base::kAccessSize * sizeof(Scalar)>::load(value, params.pointer, offset);
+  }
+
+  /// Stores a single fragment element into memory.
+  CUTLASS_HOST_DEVICE void store_element(
+      typename Base::AccessType const& value, int d, int h, int w, int c) {
+    int const offset =
+        ComputeOffsetFromStrides<typename Base::ImmediateOffsetStrides>::get(d, h, w, c);
+    Store<Scalar,
+          Base::kAccessSize,
+          Base::kMemorySpace,
+          Base::kFragmentElementType,
+          typename Base::FragmentElement,
+          Base::Tile::kW,
+          Base::kAccessSize * sizeof(Scalar)>::store(value, params.pointer, offset);
+  }
+
+  /// Test the validity of the
+  CUTLASS_HOST_DEVICE bool valid(int d, int h, int w, int c) const {
+    return predicates.at(w) && params.predicate_offset > 0;
+  }
+
+  /// add pointer offset
+  CUTLASS_HOST_DEVICE void add_pointer_offset(Index offset) { params.pointer += offset; }
+
+  /// Loads and increments iterator
+  template <typename Fragment>
+  CUTLASS_HOST_DEVICE void load_post_increment(Fragment& fragment) {
+    typename Base::FragmentIterator frag_iterator(fragment);
+    for (int d = 0; d < Base::Iterations::kD; ++d) {
+      for (int h = 0; h < Base::Iterations::kH; ++h) {
+        for (int w = 0; w < Base::Iterations::kW; ++w) {
+          for (int c = 0; c < Base::Iterations::kC; ++c) {
+            if (valid(d, h, w, c)) {
+              load_element(
+                  reinterpret_cast<typename Base::AccessType&>(frag_iterator.at(d, h, w, c)),
+                  d,
+                  h,
+                  w,
+                  c);
+            }
+          }
+          if (w < Base::Iterations::kW - 1) {
+            inc_w();
+          }
+        }
+        if (h < Base::Iterations::kH - 1) {
+          inc_h();
+        }
+      }
+      if (d < Base::Iterations::kD - 1) {
+        inc_d();
+      }
+    }
+    inc_advance();
+  }
+
+  template <typename Fragment>
+  CUTLASS_HOST_DEVICE void store_post_increment(Fragment& fragment) {
+    typename Base::FragmentIterator frag_iterator(fragment);
+    for (int d = 0; d < Base::Iterations::kD; ++d) {
+      for (int h = 0; h < Base::Iterations::kH; ++h) {
+        for (int w = 0; w < Base::Iterations::kW; ++w) {
+          for (int c = 0; c < Base::Iterations::kC; ++c) {
+            if (valid(d, h, w, c)) {
+              store_element(
+                  reinterpret_cast<typename Base::AccessType&>(frag_iterator.at(d, h, w, c)),
+                  d,
+                  h,
+                  w,
+                  c);
+            }
+          }
+          if (w < Base::Iterations::kW - 1) {
+            inc_w();
+          }
+        }
+        if (h < Base::Iterations::kH - 1) {
+          inc_h();
+        }
+      }
+      if (d < Base::Iterations::kD - 1) {
+        inc_d();
+      }
+    }
+    inc_advance();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_operand.h

@@ -0,0 +1,141 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines constant expressions for mapping GEMM problem size and strides onto pitch-linear
+   memory.
+*/
+#pragma once
+
+#include "cutlass/matrix_traits.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/util/platform.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Helper to describe attributes of GEMM matrix operands
+template <GemmOperand::Kind kOperand_, MatrixLayout::Kind kLayout_>
+struct GemmOperandTraitsAb {
+  static const bool Congruous =
+      (kOperand_ == GemmOperand::kA ^ kLayout_ == MatrixLayout::kRowMajor);
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmOperand::Kind kOperand_, typename Tile_>
+struct GetExtent;
+
+template <typename Tile_>
+struct GetExtent<GemmOperand::kA, Tile_> {
+  static const int kExtent = Tile_::kW;
+};
+
+template <typename Tile_>
+struct GetExtent<GemmOperand::kB, Tile_> {
+  static const int kExtent = Tile_::kH;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Determines the shape of a multiplicand tile in terms of strided (H) and contiguous (W)
+/// dimensions
+template <typename ThreadBlockTile_, GemmOperand::Kind Usage, MatrixLayout::Kind Layout>
+struct GemmMultiplicandTraits {
+  // Only defined for A or B
+  static_assert(Usage == GemmOperand::kA || Usage == GemmOperand::kB,
+                "MultiplicandTileShape defined only for A or B operands.");
+
+  /// Shape of GEMM thread block tile (K, N, M)
+  typedef ThreadBlockTile_ ThreadBlockTile;
+
+  /// Identifies multiplicand
+  static GemmOperand::Kind const kUsage = Usage;
+
+  /// Layout of tile
+  static MatrixLayout::Kind const kLayout = Layout;
+
+  // True if K is the strided dimension
+  static bool const kKstrided = (kUsage == GemmOperand::kA ^ kLayout == MatrixLayout::kRowMajor);
+
+  /// Map the ThreadBlockShape onto (kH, kW) dimensions for A and B operand
+  typedef typename platform::conditional<
+      kKstrided,
+      Shape<1, ThreadBlockTile::kD, GetExtent<Usage, ThreadBlockTile>::kExtent>,
+      Shape<1, GetExtent<Usage, ThreadBlockTile>::kExtent, ThreadBlockTile::kD> >::type Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Project's a coordinate (K, N, M) onto inner and outer dimensions defined for each
+/// operand.
+template <GemmOperand::Kind operand, bool Kstrided = true>
+struct ProjectOperand;
+
+/// Project A operand - (0, K, M)
+template <bool Kstrided>
+struct ProjectOperand<GemmOperand::kA, Kstrided> {
+  CUTLASS_HOST_DEVICE
+  static Coord<3> project(Coord<3> const &coord) {
+    if (Kstrided) {
+      return make_Coord(0, coord[0], coord[2]);
+    } else {
+      return make_Coord(0, coord[2], coord[0]);
+    }
+  }
+};
+
+/// Project B operand - (0, K, N)
+template <bool Kstrided>
+struct ProjectOperand<GemmOperand::kB, Kstrided> {
+  CUTLASS_HOST_DEVICE
+  static Coord<3> project(Coord<3> const &coord) {
+    if (Kstrided) {
+      return make_Coord(0, coord[0], coord[1]);
+    } else {
+      return make_Coord(0, coord[1], coord[0]);
+    }
+  }
+};
+
+/// Project C operand - (0, N, M)
+template <>
+struct ProjectOperand<GemmOperand::kC, true> {
+  CUTLASS_HOST_DEVICE
+  static Coord<3> project(Coord<3> const &coord) { return make_Coord(0, coord[1], coord[2]); }
+};
+
+/// Project D operand - (0, N, M)
+template <>
+struct ProjectOperand<GemmOperand::kD, true> {
+  CUTLASS_HOST_DEVICE
+  static Coord<3> project(Coord<3> const &coord) { return make_Coord(0, coord[1], coord[2]); }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_shared_stream.h

@@ -0,0 +1,132 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines abstractions for managing loading and storing fragments to shared memory in the
+      efficient GEMM pipeline.
+*/
+#pragma once
+
+#include "cutlass/tensor_ref.h"
+#include "cutlass/gemm/gemm_shared_tile.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The load iterator.
+    typename Iterator_,
+    /// The transformer to be applied after the data has been copied from shared memory.
+    typename Transformer_ = Copy<typename Iterator_::Fragment> >
+
+struct SharedLoadStream {
+  /// The load iterator.
+  typedef Iterator_ Iterator;
+  /// The transformer.
+  typedef Transformer_ Transformer;
+
+  /// The fragment that is copied from shared memory.
+  typedef typename Iterator::Fragment FetchedFragment;
+  /// The fragment that is obtained after the transformation by the transformer.
+  typedef typename Transformer::OutputFragment TransformedFragment;
+  /// Make sure the fragments match.
+  static_assert((platform::is_same<FetchedFragment, typename Transformer::InputFragment>::value),
+                "");
+  /// The output fragment.
+  typedef TransformedFragment Fragment;
+  /// Scalar data type
+  typedef typename Iterator::Scalar Scalar;
+
+  /// Reference type to a tensor
+  typedef TensorRef<Scalar, 4> TensorRef;
+
+  /// The params.
+  struct Params {
+    /// The iterator params.
+    typename Iterator::Params iterator;
+
+    /// Setup the params.
+    CUTLASS_HOST_DEVICE int initialize() { return iterator.initialize(); }
+  };
+
+  /// The storage in shared memory needed by that stream.
+  typedef typename Iterator::Storage SharedStorage;
+
+  /// Ctor.
+  CUTLASS_DEVICE SharedLoadStream() {}
+
+  /// Ctor.
+  CUTLASS_DEVICE SharedLoadStream(Params const &params, TensorRef const &ref) {
+    this->initialize(params, ref);
+  }
+
+  /// Initialize the stream.
+  CUTLASS_DEVICE void initialize(Params const &params, TensorRef const &ref) {
+    // The iterator.
+    iterator = Iterator(params.iterator, ref.data());
+    // The transformer.
+    transformer = Transformer();
+  }
+
+  /// Load the data from shared memory to the fetch fragment.
+  CUTLASS_DEVICE void copy() { iterator.load_post_increment(fetched[0]); }
+
+  /// Load the data from shared memory to the fetch fragment.
+  CUTLASS_DEVICE void copy(int step) { iterator.load(fetched[step % 2], step); }
+
+  /// Commit the data.
+  CUTLASS_DEVICE void commit() { transformer.transform(fetched[0], transformed[0]); }
+
+  /// Commit the data.
+  CUTLASS_DEVICE void commit(int step) {
+    transformer.transform(fetched[step % 2], transformed[step % 2]);
+  }
+
+  /// Returns the fragment for the given step
+  CUTLASS_DEVICE TransformedFragment &fragment(int step = 0) { return transformed[step % 2]; }
+
+  /// Returns the fragment for the given step
+  CUTLASS_DEVICE TransformedFragment const &fragment(int step = 0) const {
+    return transformed[step % 2];
+  }
+
+  /// Increment the stage.
+  CUTLASS_DEVICE void inc_stage() { iterator.inc_stage(); }
+
+  /// The iterator.
+  Iterator iterator;
+  /// Fetched fragment
+  FetchedFragment fetched[2];
+  /// The transformer.
+  Transformer transformer;
+  /// Transformed fragment
+  TransformedFragment transformed[2];
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_shared_tile.h

@@ -0,0 +1,417 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines iterators for efficiently loading and storing tiles to and from shared memory.
+*/
+#pragma once
+
+#include "cutlass/gemm/gemm_operand.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, typename Tile_, typename Threads_, int kScalarsPerSts_>
+struct GemmSharedStoreTileAbTraits {
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The tile.
+  typedef typename ReshapeTile<Tile_, kScalarsPerSts_>::Tile Tile;
+  /// The threads.
+  typedef Threads_ Threads;
+  /// The strides to compute the base position of the thread.
+  typedef Shape<0, ShapeCount<Tile>::kWc, Tile::kC, kScalarsPerSts_> ThreadsStrides;
+  /// The skew.
+  static int const kSkew = 0;
+  /// The number of scalars per LDG/STG.
+  static int const kAccessSize = kScalarsPerSts_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<1,
+                Tile::kH / Threads::kH,
+                Tile::kW / Threads::kW,
+                Tile::kC / Threads::kC / kAccessSize>
+      Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, Threads::kH * ShapeCount<Tile>::kWc, Threads::kW * kAccessSize> Delta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, Threads::kH * ShapeCount<Tile>::kWc, Threads::kW * kAccessSize>
+      ImmediateOffsetStrides;
+
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int offset = ComputeThreadOffsetFromStrides<Threads, ThreadsStrides>::get();
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, typename Tile_, typename Threads_, int kScalarsPerSts_, int kSkew_>
+struct GemmSharedStoreWithSkewTileAbTraits {
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The tile without skews.
+  typedef typename ReshapeTile<Tile_, kScalarsPerSts_>::Tile TileWithoutSkew;
+  /// The tile.
+  typedef typename ReshapeTile<Shape<Tile_::kD, Tile_::kH, Tile_::kW + kSkew_>,
+                               kScalarsPerSts_>::Tile Tile;
+  /// The threads.
+  typedef Threads_ Threads;
+  /// The skew.
+  static int const kSkew = kSkew_;
+  /// The number of scalars per STS.
+  static int const kAccessSize = kScalarsPerSts_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<1, TileWithoutSkew::kH / Threads::kW, TileWithoutSkew::kW / Threads::kH> Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, ShapeCount<Tile>::kWc, Threads::kH * kAccessSize> Delta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, ShapeCount<Tile>::kWc, Threads::kH * kAccessSize> ImmediateOffsetStrides;
+
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE Coord<4> operator()() const {
+      int offset = ComputeThreadOffsetFromStrides<Threads, ThreadsStrides>::get();
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+
+ protected:
+  /// The strides to compute the base position of the thread.
+  typedef Shape<0, kScalarsPerSts_, ShapeCount<Tile>::kHwc / Threads::kW> ThreadsStrides;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          typename OutputTile_,
+          typename Warps_,
+          typename ThreadsPerWarp_,
+          typename InstructionShape_,
+          int kStages_,
+          int kScalarsPerLds_,
+          int kSkew_ = 0>
+struct GemmSharedLoadTileATraits {
+  static GemmOperand::Kind const kOperand = GemmOperand::kA;
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The tile without skew.
+  typedef Shape<kStages_,
+                OutputTile_::kD / InstructionShape_::kD,
+                GetExtent<kOperand, OutputTile_>::kExtent * InstructionShape_::kD>
+      TileWithoutSkew_;
+  /// The tile with skew.
+  typedef Shape<kStages_, TileWithoutSkew_::kH, TileWithoutSkew_::kW + kSkew_> TileWithSkew;
+  /// The tile without skew after reshaping.
+  typedef typename ReshapeTile<TileWithoutSkew_, kScalarsPerLds_>::Tile TileWithoutSkew;
+  /// The tile.
+  typedef typename ReshapeTile<TileWithSkew, kScalarsPerLds_>::Tile Tile;
+  /// The number of warps.
+  typedef Warps_ Warps;
+  /// The threads in a warp.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of scalars per LDG/STG.
+  // static int const kScalarsPerLds = kScalarsPerLds_;
+  static int const kAccessSize = kScalarsPerLds_;
+  /// The skew.
+  static int const kSkew = kSkew_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The number of warps.
+  static int const kWarps = GetExtent<kOperand, Warps>::kExtent;
+  /// The number of threads in one dimension of the warp.
+  static int const kThreadsPerWarp = GetExtent<kOperand, ThreadsPerWarp>::kExtent;
+
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<1, 1, TileWithoutSkew::kW / kWarps / kThreadsPerWarp /* / kScalarsPerLds*/>
+      Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<TileWithSkew::kW * Warps::kD, 0, kWarps * kThreadsPerWarp * kAccessSize, 0>
+      ImmediateOffsetStrides;
+  typedef Shape<TileWithSkew::kW * Warps::kD, 0, kWarps * kThreadsPerWarp * kAccessSize, 0> Delta;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE Coord<4> operator()() const {
+      // Extract the warp.
+      int const warp = threadIdx.x / kWarpSize;
+      // Extract the slice.
+      int const slice = warp / (Warps::kH * Warps::kW);
+      // Compute the row offset for each warp.
+      int const warp_row = warp % Warps::kW;
+      // Compute the row offset for each thread.
+      int const lane_row = (threadIdx.x & 0x0e) / 2;
+      // The offset.
+      int const offset =
+          slice * Tile::kW * Tile::kC + (warp_row * ThreadsPerWarp::kW + lane_row) * kAccessSize;
+      // Embed the offset in a 4D coordinate vector.
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          typename OutputTile_,
+          typename Warps_,
+          typename ThreadsPerWarp_,
+          typename InstructionShape_,
+          int kStages_,
+          int kScalarsPerLds_,
+          int kSkew_ = 0>
+struct GemmSharedLoadTileBTraits {
+  static GemmOperand::Kind const kOperand = GemmOperand::kB;
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The tile without skew.
+  typedef Shape<kStages_,
+                OutputTile_::kD / InstructionShape_::kD,
+                GetExtent<kOperand, OutputTile_>::kExtent * InstructionShape_::kD>
+      TileWithoutSkew_;
+  /// The tile with skew.
+  typedef Shape<kStages_, TileWithoutSkew_::kH, TileWithoutSkew_::kW + kSkew_> TileWithSkew;
+  /// The tile without skew after reshaping.
+  typedef typename ReshapeTile<TileWithoutSkew_, kScalarsPerLds_>::Tile TileWithoutSkew;
+  /// The tile.
+  typedef typename ReshapeTile<TileWithSkew, kScalarsPerLds_>::Tile Tile;
+  /// The number of warps.
+  typedef Warps_ Warps;
+  /// The threads in a warp.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of scalars per LDG/STG.
+  static int const kAccessSize = kScalarsPerLds_;
+  /// The skew.
+  static int const kSkew = kSkew_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The number of warps.
+  static int const kWarps = GetExtent<kOperand, Warps>::kExtent;
+  /// The number of threads in one dimension of the warp.
+  static int const kThreadsPerWarp = GetExtent<kOperand, ThreadsPerWarp>::kExtent;
+
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<1, 1, TileWithoutSkew::kW / kWarps / kThreadsPerWarp /* / kAccessSize*/> Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<TileWithSkew::kW * Warps::kD, 0, kWarps * kThreadsPerWarp * kAccessSize, 0>
+      ImmediateOffsetStrides;
+  typedef Shape<TileWithSkew::kW * Warps::kD, 0, kWarps * kThreadsPerWarp * kAccessSize, 0> Delta;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE Coord<4> operator()() const {
+      // Extract the warp.
+      int const warp = threadIdx.x / kWarpSize;
+      // Extract the slice.
+      int const slice = warp / (Warps::kH * Warps::kW);
+      // The warp in the slice.
+      int const warp_in_slice = warp % (Warps::kH * Warps::kW);
+      // Compute the row offset for each warp.
+      int const warp_col = warp_in_slice / Warps::kW;
+      // Compute the row offset for each thread.
+      int const lane_col = (threadIdx.x & 0x10) / 8 + (threadIdx.x & 0x01);
+      // The offset.
+      int const offset =
+          slice * Tile::kW * Tile::kC + (warp_col * ThreadsPerWarp::kH + lane_col) * kAccessSize;
+      // Embed the offset in a 4D coordinate.
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          typename OutputTile_,
+          typename Warps_,
+          typename ThreadsPerWarp_,
+          int kScalarsPerSts_,
+          int kSkew_ = 0>
+struct GemmSharedStoreTileDTraits {
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The dimension of the output tile.
+  typedef OutputTile_ OutputTile;
+  /// The warps in the tile.
+  typedef Warps_ Warps;
+  /// The threads in the warps.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of scalars per LDG/STG.
+  static int const kAccessSize = kScalarsPerSts_;
+  /// The skew.
+  static int const kSkew = kSkew_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The number of scalars per thread.
+  static int const kScalarsPerThread = OutputTile_::kW / Warps::kW / ThreadsPerWarp::kW;
+  /// The number of threads.
+  static int const kThreads = ShapeCount<Warps>::kCount * kWarpSize;
+  /// The number of scalars per row. We build a tile with 2 rows (to avoid bank conflicts).
+  static int const kScalarsPerRow = kThreads / 2 * kScalarsPerThread + kSkew;
+
+  /// The tile.
+  typedef Shape<1, 2, kScalarsPerRow / kAccessSize, kAccessSize> Tile;
+  /// The number of iterations needed to store the tile.
+  typedef Shape<1, 1, kScalarsPerThread / kAccessSize> Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Warps::kW * ThreadsPerWarp::kW * kAccessSize> Delta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Warps::kW * ThreadsPerWarp::kW * kAccessSize> ImmediateOffsetStrides;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE Coord<4> operator()() const {
+      // The warp.
+      int const warp = threadIdx.x / kWarpSize;
+
+      // The position of the warp in the 2D tile.
+      int const warp_row = warp % Warps::kW;
+      int const warp_col = warp / Warps::kW;
+
+      // We assume that the elements are distributed in a warps as 4 columns of 8 elements. The
+      // columns are stored in threads col0=[0, 2, 4, 6, 8, 10, 12, 14], col1=[1, 3, 5, 7, .., 15],
+      // col2=[16, 18, 20, ..., 30] and col3=[17, 19, ..., 31].
+      int hi_halfwarp_offset = ((threadIdx.x >> 4) & 0x1) * OutputTile::kW;
+      int lo_halfwarp_offset = ((threadIdx.x >> 1) & 0x7) + ThreadsPerWarp::kW * warp_row;
+
+      // Odd threads go to the second half of shared memory.
+      int const row = threadIdx.x & 0x01;
+      int col = warp_col * (ThreadsPerWarp::kH / 2) * OutputTile::kW +
+                lo_halfwarp_offset * kAccessSize + hi_halfwarp_offset;
+      // Embed the offset in a 4D coords.
+      return make_Coord(0, 0, row * kScalarsPerRow + col, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          typename OutputTile_,
+          typename Warps_,
+          typename ThreadsPerWarp_,
+          int kTileH_,
+          int kScalarsPerLds_,
+          int kSkew_ = 0>
+struct GemmSharedLoadTileDTraits {
+  /// The scalar.
+  typedef typename platform::remove_const<Scalar_>::type Scalar;
+  /// The pointer.
+  typedef Scalar_* Pointer;
+  /// The dimension of the output tile.
+  typedef OutputTile_ OutputTile;
+  /// The warps in the tile.
+  typedef Warps_ Warps;
+  /// The threads in the warps.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of scalars per LDG/STG.
+  static int const kAccessSize = kScalarsPerLds_;
+  /// The skew.
+  static int const kSkew = kSkew_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The number of scalars per thread.
+  static int const kScalarsPerThread = OutputTile_::kW / Warps::kW / ThreadsPerWarp::kW;
+  /// The number of threads.
+  static int const kThreads = ShapeCount<Warps>::kCount * kWarpSize;
+  /// The number of scalars per row. We build a tile with 2 rows (to avoid bank conflicts).
+  static int const kScalarsPerRow = kThreads / 2 * kScalarsPerThread + kSkew;
+
+  /// The tile. We have 2 rows of scalars. We use those two rows to make sure we do not have bank
+  /// conflicts in the epilogue.
+  typedef Shape<1, 2, kScalarsPerRow / kAccessSize, kAccessSize> Tile;
+
+  // Compute the number of iterations per warp in the Tile::kH dimension.
+  static int const kIterationsInHPerWarp = kTileH_ / ShapeCount<Warps>::kCount;
+
+  // As explained above, the shared memory tile is composed of 2 rows and each rows is made of
+  // kScalarsPerRow. A warp is expected to read from the 1st row, then move to the 2nd row and go
+  // back to the 1st row. To model that scheme we define the Iterations shape as Shape<X, 2, ...>.
+  // However, in some cases, we have only 1 iteration per warp. In that case, we must define the
+  // shape as Shape<1, 1, ...>. The following code does that except that we hijack the kH dimension
+  // to keep the number of elements to reduce for split-K.
+  static int const kIterationsH = kIterationsInHPerWarp == 1 ? 1 : 2;
+  // As soon as we know kIterationsH, it is trivial to compute kIterationsD:
+  static int const kIterationsD = kIterationsInHPerWarp / kIterationsH;
+
+  // If we have split-K enabled, we have to jump over the elements from the "odd/even" column of
+  // threads to grab the other elements.
+  static int const kSplitK = OutputTile::kW * ThreadsPerWarp::kH / 2 * Warps::kH;
+
+  /// The number of iterations needed to store the tile.
+  typedef Shape<kIterationsD, kIterationsH, OutputTile::kW / kWarpSize / kAccessSize, Warps::kD>
+      Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<OutputTile::kW, kScalarsPerRow, kWarpSize * kAccessSize, kSplitK>
+      ImmediateOffsetStrides;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<OutputTile::kW, kScalarsPerRow, kWarpSize * kAccessSize, kSplitK> Delta;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE Coord<4> operator()() const {
+      // Each warp works on a different column.
+      int const h = threadIdx.x / kWarpSize;
+      // Compute the row.
+      int const w = (threadIdx.x & (kWarpSize - 1)) * kAccessSize;
+      int offset = 0;
+      if (Iterations::kH == 1) {
+        int const row = h & 0x1;
+        int const col = h / 2;
+        offset = row * ShapeCount<Tile>::kWc + col * OutputTile::kW * Iterations::kD + w;
+      } else {
+        offset = h * OutputTile::kW * Iterations::kD + w;
+      }
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_stream_pair.h

@@ -0,0 +1,251 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a pair of GEMM tile streams
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/matrix_traits.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/tile_allocation.h"
+#include "cutlass/tile_iterator.h"
+
+#include "cutlass/gemm/clear_accumulators.h"
+#include "cutlass/gemm/gemm_config.h"
+#include "cutlass/gemm/gemm_global_stream.h"
+#include "cutlass/gemm/gemm_operand.h"
+#include "cutlass/gemm/gemm_shared_stream.h"
+#include "cutlass/gemm/threadblock_swizzle.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Collect the global load streams for multiplicands.
+template <typename StreamA_, typename StreamB_, bool kResidueInProlog_>
+struct GlobalLoadStreamPair {
+  //
+  // Type definitions
+  //
+
+  /// Stream for A multiplicand
+  typedef StreamA_ StreamA;
+
+  /// Stream for B multiplicand
+  typedef StreamB_ StreamB;
+
+  /// Parameters object
+  struct Params {
+    /// Parameters object for StreamA
+    typename StreamA::Params stream_a;
+
+    /// Parameters object for StreamB
+    typename StreamB::Params stream_b;
+
+    /// Default constructor
+    CUTLASS_HOST_DEVICE
+    Params() {}
+
+    /// Constructs a global load stream pair Params object
+    CUTLASS_HOST_DEVICE
+    Params(typename StreamA::Params const &_params_A, typename StreamB::Params const &_params_B)
+        : stream_a(_params_A), stream_b(_params_B) {}
+  };
+
+  /// Assumes the A stream defines the index type
+  typedef typename StreamA::Index Index;
+
+  /// Shared memory allocation for threadblock-scoped GEMM tile
+  typedef ZipTileAllocation<typename StreamA::ThreadblockTileStorage,
+                              typename StreamB::ThreadblockTileStorage>
+      ThreadblockTileStorage;
+
+  /// ZipTensorRef to threadblock tiles
+  typedef typename ThreadblockTileStorage::TensorRef ThreadblockTileRef;
+
+  /// Defines a structure containing shared storage for each pair
+  struct SharedStorage {
+    typename StreamA::SharedStorage stream_a;
+    typename StreamB::SharedStorage stream_b;
+  };
+
+  //
+  // Data members
+  //
+
+  /// Stream for A multiplicand
+  StreamA stream_a;
+
+  /// Stream for B multiplicand
+  StreamB stream_b;
+
+  //
+  // Methods
+  //
+
+  /// Ctor.
+  CUTLASS_DEVICE GlobalLoadStreamPair(Params const &params,
+                                      SharedStorage &shared_storage,
+                                      ThreadblockTileRef const &threadblock_tile_ref,
+                                      Coord<3> const &bounds,
+                                      Coord<3> const &block_offset = make_Coord(0, 0, 0))
+      : stream_a(params.stream_a,
+                 shared_storage.stream_a,
+                 threadblock_tile_ref.first,
+                 bounds,
+                 block_offset),
+        stream_b(params.stream_b,
+                 shared_storage.stream_b,
+                 threadblock_tile_ref.second,
+                 bounds,
+                 block_offset) {}
+
+  CUTLASS_DEVICE
+  GlobalLoadStreamPair & operator+=(Coord<3> const offset) {
+    stream_a += offset;
+    stream_b += offset;
+    return *this;
+  }
+
+  /// Trigger the copies from shared memory to registers.
+  CUTLASS_DEVICE void copy() {
+    stream_a.copy();
+    stream_b.copy();
+  }
+
+  /// Commit the data.
+  CUTLASS_DEVICE void commit() {
+    stream_a.commit();
+    stream_b.commit();
+  }
+
+  /// Execute the residue code.
+  CUTLASS_DEVICE void residue(Index k, bool skip_clear = false) {
+    stream_a.residue(k, skip_clear);
+    stream_b.residue(k, skip_clear);
+  }
+
+  /// Move to residue.
+  CUTLASS_DEVICE void move_to_residue(Index k, Index kTileK) {
+    if (kResidueInProlog_) {
+      stream_a.move_to_residue(k, kTileK);
+      stream_b.move_to_residue(k, kTileK);
+    } else if (k < kTileK) {
+      residue(k, true);
+    }
+  }
+
+  /// Rollback to beginning of first tile.
+  CUTLASS_DEVICE void rollback(bool kRollback) {
+    if (kResidueInProlog_ && kRollback) {
+      stream_a.rollback();
+      stream_b.rollback();
+    }
+  }
+};
+
+/// Collect the global load streams for multiplicands.
+template <typename StreamA_, typename StreamB_>
+struct SharedStreamPair {
+  //
+  // Type definitions
+  //
+
+  /// Stream for A multiplicand
+  typedef StreamA_ StreamA;
+
+  /// Stream for B multiplicand
+  typedef StreamB_ StreamB;
+
+  /// Parameters object passed to load iterators
+  struct Params {
+    ///
+    typename StreamA::Params stream_a;
+
+    ///
+    typename StreamB::Params stream_b;
+  };
+
+  /// Shared memory allocation for threadblock-scoped GEMM tile
+  typedef ZipTensorRef<typename StreamA::TensorRef,
+                       typename StreamB::TensorRef >
+      ThreadblockTileRef;
+
+  //
+  // Data members
+  //
+
+  /// The stream for A.
+  StreamA stream_a;
+
+  /// The stream for B.
+  StreamB stream_b;
+
+  //
+  // Methods
+  //
+
+  /// Construct with the composable structure
+  CUTLASS_DEVICE SharedStreamPair(Params const &params, ThreadblockTileRef const &threadblock_tile_ref)
+      : stream_a(params.stream_a, threadblock_tile_ref.first),
+        stream_b(params.stream_b, threadblock_tile_ref.second) {}
+
+  /// Trigger the copies from shared memory to registers.
+  CUTLASS_DEVICE void copy(int step) {
+    stream_a.copy(step);
+    stream_b.copy(step);
+  }
+
+  /// Commit the data.
+  CUTLASS_DEVICE void commit(int step) {
+    stream_a.commit(step);
+    stream_b.commit(step);
+  }
+
+  /// The fragment A.
+  CUTLASS_DEVICE
+  typename StreamA::TransformedFragment const &fragment_a(int step) const {
+    return stream_a.fragment(step);
+  }
+
+  /// The fragment B.
+  CUTLASS_DEVICE
+  typename StreamB::TransformedFragment const &fragment_b(int step) const {
+    return stream_b.fragment(step);
+  }
+
+  /// Increment the stage.
+  CUTLASS_DEVICE void inc_stage() {
+    stream_a.inc_stage();
+    stream_b.inc_stage();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/gemm_traits.h

@@ -0,0 +1,674 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines structural properties of complete GEMM computation.
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/matrix_traits.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/tile_allocation.h"
+#include "cutlass/tile_iterator.h"
+#include "cutlass/kernel_launch.h"
+
+#include "cutlass/gemm/clear_accumulators.h"
+#include "cutlass/gemm/gemm_config.h"
+#include "cutlass/gemm/gemm_desc.h"
+#include "cutlass/gemm/gemm_stream_pair.h"
+#include "cutlass/gemm/gemm_global_stream.h"
+#include "cutlass/gemm/gemm_operand.h"
+#include "cutlass/gemm/gemm_shared_stream.h"
+#include "cutlass/gemm/threadblock_swizzle.h"
+#include "cutlass/gemm/gemm.h"
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind, typename GemmConfig_>
+struct GemmTileTraitsHelperA {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_>
+struct GemmTileTraitsHelperA<MatrixLayout::kColumnMajor, GemmConfig_> {
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = MatrixLayout::kColumnMajor;
+
+  /// The input scalar.
+  typedef typename GemmConfig_::ScalarA Scalar;
+  /// The scalar stored in shared memory.
+  typedef typename GemmConfig_::MultiplyAdd::ScalarA MultiplyAddScalar;
+
+  /// The traits class to build the iterator to load data from global memory for A^N.
+  typedef GemmGlobalTileTraits<
+      // That's A.
+      GemmOperand::kA,
+      // A is column-major.
+      MatrixLayout::kColumnMajor,
+      // The pointer is float const.
+      Scalar const,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kW>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgA>
+      GlobalTileTraits;
+
+  /// The traits class to build the iterator to store data to shared memory for A^N.
+  typedef GemmSharedStoreTileAbTraits<
+      // The pointer is float.
+      MultiplyAddScalar,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<GemmConfig_::kStages,
+            GemmConfig_::OutputTile::kD / GemmConfig_::InstructionShape::kD,
+            GemmConfig_::OutputTile::kW * GemmConfig_::InstructionShape::kD>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS (STS.32 or STS.128, etc).
+      GemmConfig_::kScalarsPerStsA>
+      SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for A^N.
+  typedef GemmSharedLoadTileATraits<
+      // The pointer is float const.
+      MultiplyAddScalar const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      GemmConfig_::kScalarsPerLdsA,
+      // The skew.
+      0>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_>
+struct GemmTileTraitsHelperA<MatrixLayout::kRowMajor, GemmConfig_> {
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = MatrixLayout::kRowMajor;
+
+  /// The input scalar.
+  typedef typename GemmConfig_::ScalarA Scalar;
+  /// The scalar stored in shared memory.
+  typedef typename GemmConfig_::MultiplyAdd::ScalarA MultiplyAddScalar;
+
+  /// The traits class to build the iterator to load data from global memory for A^T.
+  typedef GemmGlobalTileTraits<
+      // That's A.
+      GemmOperand::kA,
+      // A is row-major.
+      MatrixLayout::kRowMajor,
+      // The pointer is float const.
+      Scalar const,
+      // The tile has size MxK in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kW, GemmConfig_::OutputTile::kD>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      Shape<1, GemmConfig_::kThreads / GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kD>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgA>
+      GlobalTileTraits;
+
+  /// The number of scalars in 4B.
+  static int const kScalarsIn4B = sizeof(MultiplyAddScalar) > 4 ? 1 : 4 / sizeof(MultiplyAddScalar);
+  /// The skew for A.
+  static int const kSkewA = 128 / sizeof(MultiplyAddScalar) / GemmConfig_::kScalarsPerStsA /
+                            GlobalTileTraits::Threads::kW * kScalarsIn4B;
+
+  /// The traits class to build the iterator to store data to shared memory for A^T.
+  typedef GemmSharedStoreWithSkewTileAbTraits <
+      // The pointer is float.
+      MultiplyAddScalar,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<GemmConfig_::kStages,
+            GemmConfig_::OutputTile::kD / GemmConfig_::InstructionShape::kD,
+            GemmConfig_::OutputTile::kW * GemmConfig_::InstructionShape::kD>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS.
+      GemmConfig_::kScalarsPerStsA,
+      // The skew to avoid bank conflicts added in the tile W dimension.
+      kSkewA<GemmConfig_::kScalarsPerLdsA ? GemmConfig_::kScalarsPerLdsA : kSkewA>
+          SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for A^T.
+  typedef GemmSharedLoadTileATraits<
+      // The pointer is float const.
+      MultiplyAddScalar const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      GemmConfig_::kScalarsPerLdsA,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind, typename GemmConfig_>
+struct GemmTileTraitsHelperB {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_>
+struct GemmTileTraitsHelperB<MatrixLayout::kColumnMajor, GemmConfig_> {
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = MatrixLayout::kColumnMajor;
+
+  /// The input scalar.
+  typedef typename GemmConfig_::ScalarB Scalar;
+  /// The scalar stored in shared memory.
+  typedef typename GemmConfig_::MultiplyAdd::ScalarB MultiplyAddScalar;
+
+  /// The traits class to build the iterator to load data from global memory for B^N.
+  typedef GemmGlobalTileTraits<
+      // That's B.
+      GemmOperand::kB,
+      // B is column-major.
+      MatrixLayout::kColumnMajor,
+      // The pointer is float const.
+      Scalar const,
+      // The tile has size MxK in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kH, GemmConfig_::OutputTile::kD>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      Shape<1, GemmConfig_::kThreads / GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kD>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgB>
+      GlobalTileTraits;
+
+  /// The number of scalars in 4B.
+  static int const kScalarsIn4B = sizeof(MultiplyAddScalar) > 4 ? 1 : 4 / sizeof(MultiplyAddScalar);
+  /// The skew for B.
+  static int const kSkewB = 128 / sizeof(MultiplyAddScalar) / GemmConfig_::kScalarsPerStsB /
+                            GlobalTileTraits::Threads::kW * kScalarsIn4B;
+
+  /// The traits class to build the iterator to store data to shared memory for B^N.
+  typedef GemmSharedStoreWithSkewTileAbTraits <
+      // The pointer is float.
+      MultiplyAddScalar,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<GemmConfig_::kStages,
+            GemmConfig_::OutputTile::kD / GemmConfig_::InstructionShape::kD,
+            GemmConfig_::OutputTile::kH * GemmConfig_::InstructionShape::kD>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS.
+      GemmConfig_::kScalarsPerStsB,
+      // The skew to avoid bank conflicts added in the tile W dimension.
+      kSkewB<GemmConfig_::kScalarsPerLdsB ? GemmConfig_::kScalarsPerLdsB : kSkewB>
+          SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for B^N.
+  typedef GemmSharedLoadTileBTraits<
+      // The pointer is float const.
+      MultiplyAddScalar const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      GemmConfig_::kScalarsPerLdsB,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_>
+struct GemmTileTraitsHelperB<MatrixLayout::kRowMajor, GemmConfig_> {
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = MatrixLayout::kRowMajor;
+
+  /// The input scalar.
+  typedef typename GemmConfig_::ScalarB Scalar;
+  /// The scalar stored in shared memory.
+  typedef typename GemmConfig_::MultiplyAdd::ScalarB MultiplyAddScalar;
+
+  /// The traits class to build the iterator to load data from global memory for B^T.
+  typedef GemmGlobalTileTraits<
+      // That's B.
+      GemmOperand::kB,
+      // B is row-major.
+      MatrixLayout::kRowMajor,
+      // The pointer is float const.
+      Scalar const,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kH>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgB>
+      GlobalTileTraits;
+
+  /// The traits class to build the iterator to store data to shared memory for B^T.
+  typedef GemmSharedStoreTileAbTraits<
+      // The pointer is float.
+      MultiplyAddScalar,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<GemmConfig_::kStages,
+            GemmConfig_::OutputTile::kD / GemmConfig_::InstructionShape::kD,
+            GemmConfig_::OutputTile::kH * GemmConfig_::InstructionShape::kD>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS (STS.32 or STS.128, etc).
+      GemmConfig_::kScalarsPerStsB>
+      SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for B^T.
+  typedef GemmSharedLoadTileBTraits<
+      // The pointer is float const.
+      MultiplyAddScalar const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      GemmConfig_::kScalarsPerLdsB,
+      // The skew.
+      0>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The GEMM configuration.
+    typename GemmConfig_,
+    /// The stream to load A from global memory to shared memory.
+    typename GlobalLoadStreamA_,
+    /// The stream to load B from global memory to shared memory.
+    typename GlobalLoadStreamB_,
+    /// The stream to load A from shared memory.
+    typename SharedLoadStreamA_,
+    /// The stream to load B from shared memory.
+    typename SharedLoadStreamB_,
+    /// The epilogue.
+    typename Epilogue_,
+    /// The block swizzle to reorganize the grid.
+    typename BlockSwizzle_ = IdentityBlockSwizzle,
+    /// The index.
+    typename Index_ = int,
+    /// The tool used to clear accumulators.
+    typename ClearAccumulators_ = ClearAccumulators<typename GemmConfig_::Accumulators::Element> >
+
+struct GemmTraits {
+  /// This traits
+  typedef GemmTraits<GemmConfig_,
+    GlobalLoadStreamA_,
+    GlobalLoadStreamB_,
+    SharedLoadStreamA_,
+    SharedLoadStreamB_,
+    Epilogue_,
+    BlockSwizzle_,
+    Index_,
+    ClearAccumulators_> This_;
+
+  /// The struct that consumes this Traits
+  typedef typename cutlass::gemm::Gemm<This_> KernelClass;
+
+  /// The configuration.
+  typedef GemmConfig_ GemmConfig;
+  /// The output tile.
+  typedef typename GemmConfig::OutputTile OutputTile;
+
+  /// The stream to load A from global memory to shared memory.
+  typedef GlobalLoadStreamA_ GlobalLoadStreamA;
+  /// The layout of A.
+  static MatrixLayout::Kind const kLayoutA = GlobalLoadStreamA::kLayout;
+  /// The scalar for A.
+  typedef typename GlobalLoadStreamA_::Scalar ScalarA;
+
+  /// The stream to load B from global memory to shared memory.
+  typedef GlobalLoadStreamB_ GlobalLoadStreamB;
+  /// The layout of B.
+  static MatrixLayout::Kind const kLayoutB = GlobalLoadStreamB::kLayout;
+  /// The scalar for B.
+  typedef typename GlobalLoadStreamB_::Scalar ScalarB;
+
+  /// The iterator for A to load from shared memory.
+  typedef SharedLoadStreamA_ SharedLoadStreamA;
+  /// The iterator for B to load from shared memory.
+  typedef SharedLoadStreamB_ SharedLoadStreamB;
+
+  /// The multiply-add functor.
+  typedef typename GemmConfig::MultiplyAdd MultiplyAdd;
+  /// The epilogue.
+  typedef Epilogue_ Epilogue;
+  /// The scalars in the epilogue.
+  typedef typename Epilogue::ScalarC ScalarC;
+  typedef typename Epilogue::ScalarD ScalarD;
+
+  /// The block swizzle to reorganize the grid.
+  typedef BlockSwizzle_ BlockSwizzle;
+  /// The index.
+  typedef Index_ Index;
+  /// Clear the accumulators.
+  typedef ClearAccumulators_ ClearAccumulators;
+
+  /// Assemble the global load streams for A/B.
+  typedef GlobalLoadStreamPair<GlobalLoadStreamA,
+                               GlobalLoadStreamB,
+                               GemmConfig::kResidueInProlog>
+      GlobalLoadStream;
+
+  /// Memory needed to store the threadblock-scoped GEMM tile
+  typedef typename GlobalLoadStream::ThreadblockTileStorage ThreadblockTileStorage;
+
+  /// Assemble the shared load streams for A/B.
+  typedef SharedStreamPair<SharedLoadStreamA, SharedLoadStreamB> SharedStream;
+
+  /// Parameters object constructable on the host.
+  struct Params : public KernelLaunchConfiguration {
+
+    /// GEMM problem size
+    GemmCoord problem_size;
+
+    /// Parameters object for the global load stream
+    typename GlobalLoadStream::Params global_to_shared_stream;
+
+    /// Parameters object for the shared load stream
+    typename SharedStream::Params shared_stream;
+
+    /// The params for the epilogue.
+    typename Epilogue::Params epilogue;
+
+    /// Initialize the parameters.
+    template <typename GemmDesc_>
+    CUTLASS_HOST_DEVICE int initialize(GemmDesc_ const& desc) {
+      // Set the problem size.
+      problem_size = desc.problem_size;
+
+      // Compute grid dimensions
+      BlockSwizzle block_swizzle;
+      this->block = dim3(GemmConfig::kThreads);
+      this->grid = block_swizzle.get_grid_layout(
+        problem_size,
+        make_Coord_from_shape<OutputTile>());
+
+      // Compute offset to residue.
+      Index gemm_k = problem_size[0];
+      Index offset_to_residue = (gemm_k % OutputTile::kD) ? gemm_k - (gemm_k % OutputTile::kD) : 0;
+
+      // Initialize parameters objects for
+      int error_code = global_to_shared_stream.stream_a.initialize(
+        desc.A.data(),
+        desc.batch_stride_A,
+        desc.A.leading_dim(),
+        offset_to_residue
+      );
+      if (error_code) {
+        return error_code;
+      }
+
+      error_code = global_to_shared_stream.stream_b.initialize(
+        desc.B.data(),
+        desc.batch_stride_B,
+        desc.B.leading_dim(),
+        offset_to_residue
+      );
+
+      if (error_code) {
+        return error_code;
+      }
+
+      // The epilogue.
+      return epilogue.initialize(desc);
+    }
+
+    /// Helper to construct a GEMM params using a BLAS-like API
+    CUTLASS_HOST_DEVICE int initialize(Index m,
+                                       Index n,
+                                       Index k,
+                                       typename Epilogue::Scalar alpha,
+                                       ScalarA const* d_a,
+                                       Index lda,
+                                       ScalarB const* d_b,
+                                       Index ldb,
+                                       typename Epilogue::Scalar beta,
+                                       ScalarC const* d_c,
+                                       Index ldc,
+                                       ScalarD* d_d,
+                                       Index ldd) {
+      GemmDesc<ScalarA, ScalarB, ScalarC, ScalarD, typename Epilogue::Scalar> desc(
+        GemmCoord(k, n, m, 1),
+        alpha,
+        TensorRef<ScalarA const, 2>(d_a, lda),
+        TensorRef<ScalarB const, 2>(d_b, ldb),
+        beta,
+        TensorRef<ScalarC const, 2>(d_c, ldc),
+        TensorRef<ScalarD, 2>(d_d, ldd)
+      );
+
+      return this->initialize(desc);
+    }
+
+    /// Helper to construct a batched GEMM params
+    CUTLASS_HOST_DEVICE int initialize(Index m,
+                                       Index n,
+                                       Index k,
+                                       typename Epilogue::Scalar alpha,
+                                       ScalarA const* d_a,
+                                       Index lda,
+                                       long long int batch_stride_A,
+                                       ScalarB const* d_b,
+                                       Index ldb,
+                                       long long int batch_stride_B,
+                                       typename Epilogue::Scalar beta,
+                                       ScalarC const* d_c,
+                                       Index ldc,
+                                       long long int batch_stride_C,
+                                       ScalarD* d_d,
+                                       Index ldd,
+                                       long long int batch_stride_D,
+                                       Index batch_count) {
+
+      GemmDesc<ScalarA, ScalarB, ScalarC, ScalarD, typename Epilogue::Scalar> desc(
+        GemmCoord(k, n, m, batch_count),
+        alpha,
+        TensorRef<ScalarA const, 2>(d_a, lda),
+        batch_stride_A,
+        TensorRef<ScalarB const, 2>(d_b, ldb),
+        batch_stride_B,
+        beta,
+        TensorRef<ScalarC const, 2>(d_c, ldc),
+        batch_stride_C,
+        TensorRef<ScalarD, 2>(d_d, ldd),
+        batch_stride_D
+      );
+
+      return this->initialize(desc);
+    }
+  };
+
+  // The storage for the main loop + prologue.
+  struct MainLoopSharedStorage {
+    /// Stores the threadblock tile
+    ThreadblockTileStorage threadblock_tile;
+
+    /// Storage for GEMM global stream
+    typename GlobalLoadStream::SharedStorage global_to_shared_stream;
+
+    /// Storage for clearing accumulators
+    typename ClearAccumulators::SharedStorage clear;
+  };
+
+  /// The storage in shared memory.
+  union SharedStorage {
+    // The storage for the main loop.
+    MainLoopSharedStorage main_loop;
+    // The storage for the epilogue.
+    typename Epilogue::SharedStorage epilogue;
+  };
+
+  /// The memory fence for shared loads.
+  static CUTLASS_DEVICE void shared_load_fence(bool in_loop) {
+    if (SharedLoadStreamA::Iterator::kRequiresLoadFence ||
+        SharedLoadStreamB::Iterator::kRequiresLoadFence) {
+        __syncthreads();
+    }
+  }
+
+  /// The memory fence for shared stores.
+  static CUTLASS_DEVICE void shared_store_fence(bool in_loop) {
+      __syncthreads();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmTileTraitsHelperA_, typename GemmTileTraitsHelperB_, typename Index_>
+struct SimplifiedGemmTraitsHelper {
+  /// The global iterator to load A from global memory.
+  typedef GemmGlobalIteratorAb<typename GemmTileTraitsHelperA_::GlobalTileTraits, Index_>
+      GlobalLoadIteratorA;
+  /// The data converter for A before storing to shared memory.
+  typedef Copy<typename GlobalLoadIteratorA::Fragment> GlobalTransformerA;
+  /// The iterator to store A to shared memory.
+  typedef TileStoreIterator<typename GemmTileTraitsHelperA_::SharedStoreTileTraits,
+                            typename GemmTileTraitsHelperA_::SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorA;
+  /// The stream to load A from global memory to shared memory.
+  typedef GlobalLoadStream<GemmOperand::kA,
+                              GlobalLoadIteratorA,
+                              SharedStoreIteratorA,
+                              GlobalTransformerA>
+      GlobalLoadStreamA;
+
+  /// The global iterator to load B from global memory.
+  typedef GemmGlobalIteratorAb<typename GemmTileTraitsHelperB_::GlobalTileTraits, Index_>
+      GlobalLoadIteratorB;
+  /// The data converter for B before storing to shared memory.
+  typedef Copy<typename GlobalLoadIteratorB::Fragment> GlobalTransformerB;
+  /// The iterator to store B to shared memory.
+  typedef TileStoreIterator<typename GemmTileTraitsHelperB_::SharedStoreTileTraits,
+                            typename GemmTileTraitsHelperB_::SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorB;
+  /// The stream to load B from global memory to shared memory.
+  typedef GlobalLoadStream<GemmOperand::kB,
+                              GlobalLoadIteratorB,
+                              SharedStoreIteratorB,
+                              GlobalTransformerB>
+      GlobalLoadStreamB;
+
+  /// The iterator to load A from shared memory.
+  typedef TileLoadIterator<typename GemmTileTraitsHelperA_::SharedLoadTileTraits,
+                           typename GemmTileTraitsHelperA_::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorA;
+  /// The stream to load A from shared memory.
+  typedef SharedLoadStream<SharedLoadIteratorA> SharedLoadStreamA;
+  /// The iterator to load B from shared memory.
+  typedef TileLoadIterator<typename GemmTileTraitsHelperB_::SharedLoadTileTraits,
+                           typename GemmTileTraitsHelperB_::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorB;
+  /// The stream to load B from shared memory.
+  typedef SharedLoadStream<SharedLoadIteratorB> SharedLoadStreamB;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The config for the GEMM.
+    typename GemmConfig_,
+    /// The epilogue.
+    typename Epilogue_,
+    /// The index.
+    typename Index_ = int,
+    // The configuration for the A matrix.
+    typename GemmTileTraitsHelperA_ = GemmTileTraitsHelperA<kLayoutA_, GemmConfig_>,
+    // The configuration for the B matrix.
+    typename GemmTileTraitsHelperB_ = GemmTileTraitsHelperB<kLayoutB_, GemmConfig_>,
+    // The helper class to create the streams and iterators.
+    typename Helper_ =
+        SimplifiedGemmTraitsHelper<GemmTileTraitsHelperA_, GemmTileTraitsHelperB_, Index_> >
+struct SimplifiedGemmTraits : public GemmTraits<
+                                  // The config.
+                                  GemmConfig_,
+                                  // The stream to load A from global memory to shared memory.
+                                  typename Helper_::GlobalLoadStreamA,
+                                  // The stream to load B from global memory to shared memory.
+                                  typename Helper_::GlobalLoadStreamB,
+                                  // The stream to load A from shared memory.
+                                  typename Helper_::SharedLoadStreamA,
+                                  // The stream to load B from shared memory.
+                                  typename Helper_::SharedLoadStreamB,
+                                  // The epilogue.
+                                  Epilogue_,
+                                  // The block swizzle to reorganize the grid.
+                                  IdentityBlockSwizzle,
+                                  // The index.
+                                  Index_,
+                                  // The tool used to clear accumulators.
+                                  ClearAccumulators<typename GemmConfig_::Accumulators::Element> > {
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/grid_raster.h

@@ -1,436 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Abstraction for enumerating \p block_task within an input matrix
- */
-
-#include <stdint.h>
-
-#include "../util/util.h"
-
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * grid_raster_strategy
- ******************************************************************************/
-
-/**
- * \brief Strategies for enumerating \p block_task within an input matrix
- */
-struct grid_raster_strategy
-{
-    /// \brief Enumerants
-    enum kind_t
-    {
-        /**
-         * Default \p block_task assignment (currently ColumnMajor for N*,
-         * RowMajor for TT, and TiledCohort for TN)
-         */
-        Default,
-
-        /**
-         * Column-major \p block_task assignment
-         */
-        ColumnMajor,
-
-        /**
-         * Row-major \p block_task assignment
-         */
-        RowMajor,
-
-        /**
-         * Two-level \p block_task assignment (both column-major)
-         */
-        TiledCohort,
-    };
-};
-
-
-
-/******************************************************************************
- * grid_raster
- ******************************************************************************/
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- *
- * NB: This generic class is not directly constructible.  Algorithm-specific
- * template specializations will provide the API functionality prescribed here.
- */
-template <
-    int                             BlockItemsY,    ///< Height in rows of a block-wide tile in matrix C
-    int                             BlockItemsX,    ///< Width in columns of a block-wide tile in matrix C
-    matrix_transform_t::kind_t      TransformA,     ///< View transform enumerant for matrix A
-    matrix_transform_t::kind_t      TransformB,     ///< View transform enumerant for matrix B
-    grid_raster_strategy::kind_t    RasterStrategy> ///< Strategy for enumerating \p block_task within an input matrix
-struct grid_raster
-{
-    //-------------------------------------------------------------------------
-    // Device API
-    //-------------------------------------------------------------------------
-
-    /// Thread block's base item coordinates (x, y) in matrix C
-    int2 block_item_coords;
-
-    /// Constructor
-    grid_raster();
-
-    /// Whether the thread block base coordinates are out-of-bounds for an m*n matrix C
-    bool is_block_oob(int m, int n);
-
-
-    //-------------------------------------------------------------------------
-    // Grid launch API
-    //-------------------------------------------------------------------------
-
-    /// Compute the kernel grid extents (in thread blocks) for consuming an m*n matrix C
-    static dim3 grid_dims(int m, int n);
-};
-
-
-
-/******************************************************************************
- * grid_raster (ColumnMajor specialization)
- ******************************************************************************/
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- * (ColumnMajor specialization)
- *
- * Maps thread blocksin column-major fashion
- */
-template <
-    int                         BlockItemsY,          ///< Height in rows of a block-wide tile in matrix C
-    int                         BlockItemsX,          ///< Width in columns of a block-wide tile in matrix C
-    matrix_transform_t::kind_t  TransformA,         ///< View transform enumerant for matrix A
-    matrix_transform_t::kind_t  TransformB>         ///< View transform enumerant for matrix B
-struct grid_raster<
-    BlockItemsY,
-    BlockItemsX,
-    TransformA,
-    TransformB,
-    grid_raster_strategy::ColumnMajor>                   ///< Strategy for enumerating \p block_task within an input matrix
-{
-    //-------------------------------------------------------------------------
-    // Device API
-    //-------------------------------------------------------------------------
-
-    /// Thread block's base item coordinates (x, y) in matrix C
-    int2 block_item_coords;
-
-    /// Constructor
-    inline __device__
-    grid_raster()
-    {
-        // blockDim.x is the fastest changing grid dim on current architectures
-        block_item_coords = make_int2(
-            BlockItemsX * blockIdx.y,
-            BlockItemsY * blockIdx.x);
-    }
-
-    /// Whether the base \p block_item_coords are out-of-bounds for an m*n matrix C
-    inline __device__
-    bool is_block_oob(int m, int n)
-    {
-        // ColumnMajor never rasterizes fully out-of-bounds thread blocks
-        return false;
-    }
-
-    //-------------------------------------------------------------------------
-    // Grid launch API
-    //-------------------------------------------------------------------------
-
-    /// Compute the kernel grid extents (in thread blocks) for consuming an m*n matrix C
-    inline __host__ __device__
-    static dim3 grid_dims(int m, int n)
-    {
-        // blockDim.x is the fastest changing grid dim on current architectures
-        return dim3(
-            (m + BlockItemsY - 1) / BlockItemsY,
-            (n + BlockItemsX - 1) / BlockItemsX);
-    }
-};
-
-
-
-/******************************************************************************
- * grid_raster (RowMajor specialization)
- ******************************************************************************/
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- * (RowMajor specialization)
- *
- * Enumerates \p block_task in row-major fashion
- */
-template <
-    int                         BlockItemsY,    ///< Height in rows of a block-wide tile in matrix C
-    int                         BlockItemsX,    ///< Width in columns of a block-wide tile in matrix C
-    matrix_transform_t::kind_t  TransformA,     ///< View transform enumerant for matrix A
-    matrix_transform_t::kind_t  TransformB>     ///< View transform enumerant for matrix B
-struct grid_raster<
-    BlockItemsY,
-    BlockItemsX,
-    TransformA,
-    TransformB,
-    grid_raster_strategy::RowMajor>             ///< Strategy for enumerating \p block_task within an input matrix
-{
-    //-------------------------------------------------------------------------
-    // Device API
-    //-------------------------------------------------------------------------
-
-    /// Thread block's base item coordinates (x, y) in matrix C
-    int2 block_item_coords;
-
-    /// Constructor
-    inline __device__
-    grid_raster()
-    {
-        // blockDim.x is the fastest changing grid dim on current architectures
-        block_item_coords = make_int2(
-            BlockItemsX * blockIdx.x,
-            BlockItemsY * blockIdx.y);
-    }
-
-    /// Whether the base \p block_item_coords are out-of-bounds for an m*n matrix C
-    inline __device__
-    bool is_block_oob(int m, int n)
-    {
-        // RowMajor never rasterizes fully out-of-bounds thread blocks
-        return false;
-    }
-
-    //-------------------------------------------------------------------------
-    // Grid launch API
-    //-------------------------------------------------------------------------
-
-    /// Compute the kernel grid extents (in thread blocks) for consuming an m*n matrix C
-    inline __host__ __device__
-    static dim3 grid_dims(int m, int n)
-    {
-        // blockDim.x is the fastest changing grid dim on current architectures
-        return dim3(
-            (n + BlockItemsX - 1) / BlockItemsX,
-            (m + BlockItemsY - 1) / BlockItemsY);
-    }
-
-};
-
-
-
-/******************************************************************************
- * grid_raster (TiledCohort specialization)
- ******************************************************************************/
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- * (TiledCohort specialization)
- *
- * Enumerates \p block_task in column-major fashion across "cohort" tiles (where
- * cohorts are CohortBlocksY high and CohortBlocksX wide), and enumerates cohorts
- * across the matrix in column-major fashion.
- *
- * Grid layout:
- *  - gridDim.y is the height of the grid in cohorts
- *  - gridDim.x is the width of the grid in cohorts multiplied by the number of
- *    thread blocks per cohort
- */
-template <
-    int                         BlockItemsY,    ///< Height in rows of a block-wide tile in matrix C
-    int                         BlockItemsX,    ///< Width in columns of a block-wide tile in matrix C
-    matrix_transform_t::kind_t  TransformA,     ///< View transform enumerant for matrix A
-    matrix_transform_t::kind_t  TransformB>     ///< View transform enumerant for matrix B
-struct grid_raster<
-    BlockItemsY,
-    BlockItemsX,
-    TransformA,
-    TransformB,
-    grid_raster_strategy::TiledCohort>       ///< Strategy for enumerating \p block_task within an input matrix
-{
-    enum
-    {
-        /// Height in thread blocks of a grid rasterization cohort
-        CohortBlocksY = 2,
-
-        /// Width in thread blocks of a grid rasterization cohort
-        CohortBlocksX = 2,
-
-        /// Number of thread blocks per cohort
-        BlocksPerCohort = CohortBlocksY * CohortBlocksX,
-
-        /// Height in items of a grid rasterization cohort
-        CohortItemsY = CohortBlocksY * BlockItemsY,
-
-        /// Width in items of a grid rasterization cohort
-        CohortItemsX = CohortBlocksX * BlockItemsX,
-
-    };
-
-    //-------------------------------------------------------------------------
-    // Device API
-    //-------------------------------------------------------------------------
-
-    /// Thread block's base item coordinates (x, y) in matrix C
-    int2 block_item_coords;
-
-    /// Constructor
-    inline __device__
-    grid_raster()
-    {
-        int block_idx_cohort = blockIdx.x % BlocksPerCohort;
-        int2 cohort_coords_grid = make_int2(
-            blockIdx.x / BlocksPerCohort,
-            blockIdx.y);
-
-        // Cohort is rastered in column-major order
-        int2 block_coords_cohort = make_int2(
-            block_idx_cohort / CohortBlocksY,
-            block_idx_cohort % CohortBlocksY);
-
-        block_item_coords = make_int2(
-            ((cohort_coords_grid.x * CohortBlocksX) + block_coords_cohort.x) * BlockItemsX,
-            ((cohort_coords_grid.y * CohortBlocksY) + block_coords_cohort.y) * BlockItemsY);
-    }
-
-    /// Whether the base \p block_item_coords are out-of-bounds for an m*n matrix C
-    inline __device__
-    bool is_block_oob(int m, int n)
-    {
-        /// thread blocks within the cohort may be fully out-of-bounds
-        return (block_item_coords.x >= n) || (block_item_coords.y >= m);
-    }
-
-    //-------------------------------------------------------------------------
-    // Grid launch API
-    //-------------------------------------------------------------------------
-
-    /// Compute the kernel grid extents (in thread blocks) for consuming an m*n matrix C
-    inline __host__ __device__
-    static dim3 grid_dims(int m, int n)
-    {
-        // Extents of C matrix in cohorts
-        int2 grid_cohort_dims = make_int2(
-            (n + CohortItemsX - 1) / CohortItemsX,
-            (m + CohortItemsY - 1) / CohortItemsY);
-
-        return dim3(
-            grid_cohort_dims.x * BlocksPerCohort,   // gridDim.x is width of grid in cohorts * size of cohort in blocks
-            grid_cohort_dims.y,                     // gridDim.y is height of grid in cohorts
-            1);                                     // gridDim.z is reserved for optional k-splitting
-    }
-};
-
-
-/******************************************************************************
- * grid_raster (Default specializations)
- ******************************************************************************/
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- * (Default N* specialization)
- *
- * Maps thread blocksin column-major fashion
- */
-template <
-    int BlockItemsY,                        ///< Height in rows of a block-wide tile in matrix C
-    int BlockItemsX,                        ///< Width in columns of a block-wide tile in matrix C
-    matrix_transform_t::kind_t TransformB>  ///< View transform enumerant for matrix B
-struct grid_raster<
-    BlockItemsY,
-    BlockItemsX,
-    matrix_transform_t::NonTranspose,       ///< View transform enumerant for matrix A
-    TransformB,
-    grid_raster_strategy::Default>         ///< Strategy for enumerating \p block_task within an input matrix
-:
-    grid_raster<
-        BlockItemsY,
-        BlockItemsX,
-        matrix_transform_t::NonTranspose,
-        TransformB,
-        grid_raster_strategy::ColumnMajor>
-{};
-
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- * (Default TT specialization)
- *
- * Maps thread blocksin row-major fashion
- */
-template <
-    int BlockItemsY,                    ///< Height in rows of a block-wide tile in matrix C
-    int BlockItemsX>                    ///< Width in columns of a block-wide tile in matrix C
-struct grid_raster<
-    BlockItemsY,
-    BlockItemsX,
-    matrix_transform_t::Transpose,      ///< View transform enumerant for matrix A
-    matrix_transform_t::Transpose,      ///< View transform enumerant for matrix B
-    grid_raster_strategy::Default>     ///< Strategy for enumerating \p block_task within an input matrix
-:
-    grid_raster<
-        BlockItemsY,
-        BlockItemsX,
-        matrix_transform_t::Transpose,
-        matrix_transform_t::Transpose,
-        grid_raster_strategy::RowMajor>
-{};
-
-
-/**
- * \brief Abstraction for enumerating \p block_task within an input matrix
- * (Default TN specialization)
- *
- * Maps thread blocksin blocked cohorts
- */
-template <
-    int BlockItemsY,                    ///< Height in rows of a block-wide tile in matrix C
-    int BlockItemsX>                    ///< Width in columns of a block-wide tile in matrix C
-struct grid_raster<
-    BlockItemsY,
-    BlockItemsX,
-    matrix_transform_t::Transpose,      ///< View transform enumerant for matrix A
-    matrix_transform_t::NonTranspose,   ///< View transform enumerant for matrix B
-    grid_raster_strategy::Default>     ///< Strategy for enumerating \p block_task within an input matrix
-:
-    grid_raster<
-        BlockItemsY,
-        BlockItemsX,
-        matrix_transform_t::Transpose,
-        matrix_transform_t::NonTranspose,
-        grid_raster_strategy::TiledCohort>
-{};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/hgemm_global_tile.h

@@ -0,0 +1,90 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Tile traits used to construct global tile iterator for HGEMM. This is intended to
+      partition the thread block-level tile into 2D subtiles loaded by the threads and facilitate
+      memory accesses larger than 16 bits.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/matrix_traits.h"
+#include "cutlass/reshape_tile.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <GemmOperand::Kind kOperand_,
+          MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename Tile_,
+          typename Threads_,
+          int kAccessSize_>
+struct HgemmCrosswiseGlobalTileTraits : public GemmGlobalTileTraits<
+                                            // Which GEMM operand?
+                                            kOperand_,
+                                            // The layout.
+                                            kLayout_,
+                                            // The scalar.
+                                            Scalar_,
+                                            // The tile.
+                                            Tile_,
+                                            // The threads.
+                                            Threads_,
+                                            // The number of scalars per LDG/STG.
+                                            kAccessSize_> {
+  /// The base class.
+  typedef GemmGlobalTileTraits<kOperand_, kLayout_, Scalar_, Tile_, Threads_, kAccessSize_> Base;
+  /// The threads.
+  typedef typename Base::Threads Threads;
+  /// The threads strides.
+  typedef Shape<1, 2, Base::VectorizedTile::kC> ThreadsDelta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<Base::Threads::kH * 2, 1, Base::Threads::kW, Base::kAccessSize> Delta;
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<Base::VectorizedTile::kH / Base::Threads::kH / 2,
+                2,
+                Base::VectorizedTile::kW / Base::Threads::kW,
+                Base::VectorizedTile::kC / Base::kAccessSize>
+      Iterations;
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int thread_offset_h = threadIdx.x / Threads::kW * ThreadsDelta::kH;
+      int thread_offset_w = threadIdx.x % Threads::kW * ThreadsDelta::kW;
+
+      return make_Coord(0, thread_offset_h, thread_offset_w, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/hgemm_multiply_add.h

@@ -0,0 +1,106 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Specialization implementing multiply-add operation on half-precision floating point
+   fragments.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+
+#include "cutlass/gemm/thread_multiply_add.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Template performing matrix multiply-add operation within a thread
+template <typename ThreadGemmShape_, typename ThreadsPerWarp_>
+struct ThreadMultiplyAdd<ThreadGemmShape_, ThreadsPerWarp_, half, half, half> {
+  /// The shape of the instruction.
+  typedef Shape<1, 1, 2, 1> InstructionShape;
+  /// The number of accumulators per thread.
+  typedef ThreadGemmShape_ ThreadGemmShape;
+  /// Aliased for compatibility. Will be removed for CUTLASS v2.0.
+  typedef ThreadGemmShape AccumulatorsPerThread;
+  /// The number of threads per warp.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of accumulators per warp.
+  typedef typename ShapeMul<ThreadGemmShape, ThreadsPerWarp>::Shape AccumulatorsPerWarp;
+  /// The type for A.
+  typedef half ScalarA;
+  /// The fragment for A.
+  typedef Fragment<ScalarA, AccumulatorsPerThread::kW> FragmentA;
+  /// The type for B.
+  typedef half ScalarB;
+  /// The fragment for B.
+  typedef Fragment<ScalarB, AccumulatorsPerThread::kH> FragmentB;
+  /// The type for C and D.
+  typedef half ScalarC;
+  /// The accumulators.
+  typedef Fragment<half, AccumulatorsPerThread::kH * AccumulatorsPerThread::kW> Accumulators;
+
+  /// Make sure there's an even number of elements in both dimensions.
+  static_assert(AccumulatorsPerThread::kH % 2 == 0, "Invalid size");
+  static_assert(AccumulatorsPerThread::kW % 2 == 0, "Invalid size");
+
+  /// Ctor.
+  CUTLASS_DEVICE ThreadMultiplyAdd() {}
+
+  /// Multiply : d = a*b + c.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
+    // The inputs.
+    __half2 const* a_half2 = reinterpret_cast<__half2 const*>(&a[0]);
+    __half2 const* b_half2 = reinterpret_cast<__half2 const*>(&b[0]);
+    __half2 const* c_half2 = reinterpret_cast<__half2 const*>(&c[0]);
+
+    // The output.
+    __half2* d_half2 = reinterpret_cast<__half2*>(&d[0]);
+
+    for (int j = 0; j < AccumulatorsPerThread::kH / 2; ++j) {
+      for (int i = 0; i < AccumulatorsPerThread::kW / 2; ++i) {
+        // The offsets in the output fragment.
+        int const k0 = (2 * j + 0) * (AccumulatorsPerThread::kW / 2) + i;
+        int const k1 = (2 * j + 1) * (AccumulatorsPerThread::kW / 2) + i;
+
+        // Compute the product a[i] * b[j].low.
+        d_half2[k0] = __hfma2(a_half2[i], __low2half2(b_half2[j]), c_half2[k0]);
+        // Compute the product a[i] * b[j].high.
+        d_half2[k1] = __hfma2(a_half2[i], __high2half2(b_half2[j]), c_half2[k1]);
+      }
+    }
+#endif
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/hgemm_swizzle.h

@@ -0,0 +1,94 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Transposes a tile of 16b elements. Used by HGEMM to construct a K-strided layout in
+      shared memory for multiplicands.
+*/
+#pragma once
+
+#include <cuda_fp16.h>
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GlobalIterator_>
+struct HgemmSwizzle {
+  /// The global iterator.
+  typedef GlobalIterator_ GlobalIterator;
+  /// The source fragment.
+  typedef typename GlobalIterator::Fragment Fragment;
+  /// The shape of the source fragment.
+  typedef typename GlobalIterator::FragmentShape FragmentShape;
+
+  /// The input fragment.
+  typedef Fragment InputFragment;
+  /// The output fragment.
+  typedef Fragment OutputFragment;
+
+  /// The src/dst must be half fragments.
+  static_assert((platform::is_same<typename Fragment::Element, half>::value), "Works on half");
+
+  /// The number of elements must be a multiple of 2.
+  static_assert(FragmentShape::kH == 2 && ShapeCount<FragmentShape>::kWc == 2, "Not multiple of 2");
+
+  /// Ctor.
+  CUTLASS_DEVICE HgemmSwizzle() {}
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(Fragment const& src, Fragment& dst) {
+    // Expose src/dst as int arrays.
+    int const* src_int = reinterpret_cast<int const*>(&src[0]);
+    int* dst_int = reinterpret_cast<int*>(&dst[0]);
+
+    // Transpose the data.
+    for (int d = 0; d < FragmentShape::kD; ++d) {
+      // The indices to read two consecutive "rows".
+      int const i0 = 2 * d + 0;
+      int const i1 = 2 * d + 1;
+
+      int a0 = src_int[i0];
+      int a1 = src_int[i1];
+
+      int b0, b1;
+      asm volatile("prmt.b32 %0, %1, %2, 0x5410;" : "=r"(b0) : "r"(a0), "r"(a1));
+      asm volatile("prmt.b32 %0, %1, %2, 0x7632;" : "=r"(b1) : "r"(a0), "r"(a1));
+
+      // The indices to store with "strides".
+      int const j0 = 0 * (ShapeCount<FragmentShape>::kDhw / 2) + d;
+      int const j1 = 1 * (ShapeCount<FragmentShape>::kDhw / 2) + d;
+
+      dst_int[j0] = b0;
+      dst_int[j1] = b1;
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/hgemm_traits.h

@@ -0,0 +1,406 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defies structural properties of half-precision GEMM computation.
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/reshape_tile.h"
+
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/gemm/gemm_epilogue.h"
+#include "cutlass/gemm/gemm_epilogue_traits.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/gemm/gemm_shared_tile.h"
+#include "cutlass/gemm/gemm_traits.h"
+#include "cutlass/gemm/hgemm_global_tile.h"
+#include "cutlass/gemm/hgemm_multiply_add.h"
+#include "cutlass/gemm/hgemm_swizzle.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The tile size for the GEMM KxNxM.
+    typename OutputTile_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_,
+    /// The number of scalars per LDG for A.
+    int kScalarsPerLdgA_ = 2,
+    /// The number of scalars per LDG for B.
+    int kScalarsPerLdgB_ = 2>
+struct HgemmConfig : public GemmConfig<
+                         /// The scalar type for A.
+                         half,
+                         /// The scalar type for B.
+                         half,
+                         /// The scalar type for C.
+                         half,
+                         /// The scalar type for D.
+                         half,
+                         /// The tile size for the GEMM KxNxM.
+                         OutputTile_,
+                         /// The functor to do the math in the main loop.
+                         ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, half, half, half>,
+                         /// The number of scalars per LDG for A.
+                         kScalarsPerLdgA_,
+                         /// The number of scalars per STS for A.
+                         kScalarsPerLdgA_,
+                         /// The number of scalars per LDS for A.
+                         8,
+                         /// The number of scalars per LDG for B.
+                         kScalarsPerLdgB_,
+                         /// The number of scalars per STS for B.
+                         kScalarsPerLdgB_,
+                         /// The number of scalars per LDS for B.
+                         8,
+                         /// The number of scalars per LDG for C and STG for D.
+                         2,
+                         /// The number of scalars per STS for D.
+                         8,
+                         /// The number of scalars per LDS for D.
+                         2,
+                         /// The number of stages in shared memory.
+                         2,
+                         /// kResidueSeparate
+                         false,
+                         /// kResidueInPrologue
+                         true,
+                         /// kLaunchBounds
+                         false
+                         > {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename Iterator_>
+struct HgemmTransformerA {};
+
+template <typename Iterator_>
+struct HgemmTransformerA<MatrixLayout::kColumnMajor, Iterator_> {
+  typedef Convert<typename Iterator_::Fragment, typename Iterator_::Fragment> Transformer;
+};
+
+template <typename Iterator_>
+struct HgemmTransformerA<MatrixLayout::kRowMajor, Iterator_> {
+  typedef HgemmSwizzle<Iterator_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename Iterator_>
+struct HgemmTransformerB {};
+
+template <typename Iterator_>
+struct HgemmTransformerB<MatrixLayout::kRowMajor, Iterator_> {
+  typedef Convert<typename Iterator_::Fragment, typename Iterator_::Fragment> Transformer;
+};
+
+template <typename Iterator_>
+struct HgemmTransformerB<MatrixLayout::kColumnMajor, Iterator_> {
+  typedef HgemmSwizzle<Iterator_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename GemmConfig_>
+struct HgemmTileTraitsHelperA : public GemmTileTraitsHelperA<kLayout_, GemmConfig_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_>
+struct HgemmTileTraitsHelperA<MatrixLayout::kRowMajor, GemmConfig_>
+    : public GemmTileTraitsHelperA<MatrixLayout::kRowMajor, GemmConfig_> {
+  /// The base config.
+  typedef GemmTileTraitsHelperA<MatrixLayout::kRowMajor, GemmConfig_> Base;
+
+  /// The traits class to build the iterator to load data from global memory for A^T.
+  typedef HgemmCrosswiseGlobalTileTraits<
+      GemmOperand::kA,
+      // The layout.
+      MatrixLayout::kRowMajor,
+      // The pointer.
+      half const,
+      // The tile has size MxK in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kW, GemmConfig_::OutputTile::kD>,
+      // The threads are distributed as (threads / K ) x K (the traits may reorganize).
+      Shape<1, GemmConfig_::kThreads / GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kD>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc)
+      GemmConfig_::kScalarsPerLdgA>
+      GlobalTileTraits;
+
+  static int const kSkewA = 128 / sizeof(half) / GlobalTileTraits::Threads::kW / 2;
+
+  /// The traits class to build the iterator to store data to shared memory for A^T.
+  typedef GemmSharedStoreWithSkewTileAbTraits <
+      // The pointer.
+      half,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<GemmConfig_::kStages,
+            GemmConfig_::OutputTile::kD / GemmConfig_::InstructionShape::kD,
+            GemmConfig_::OutputTile::kW * GemmConfig_::InstructionShape::kD>,
+      // The threads are distributed as warps x 32(the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS (STS.32 or STS.128, etc).
+      2,
+      // The skew to avoid bank conflicts added in the tile W dimension.
+      kSkewA<GemmConfig_::kScalarsPerLdsA ? GemmConfig_::kScalarsPerLdsA : kSkewA>
+          SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for A^T.
+  typedef GemmSharedLoadTileATraits<
+      // The pointer.
+      half const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      8,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename GemmConfig_>
+struct HgemmTileTraitsHelperB : public GemmTileTraitsHelperB<kLayout_, GemmConfig_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_>
+struct HgemmTileTraitsHelperB<MatrixLayout::kColumnMajor, GemmConfig_>
+    : public GemmTileTraitsHelperB<MatrixLayout::kColumnMajor, GemmConfig_> {
+  /// The base config.
+  typedef GemmTileTraitsHelperB<MatrixLayout::kColumnMajor, GemmConfig_> Base;
+
+  /// The traits class to build the iterator to load data from global memory for B^N.
+  typedef HgemmCrosswiseGlobalTileTraits<
+      GemmOperand::kB,
+      // The layout.
+      MatrixLayout::kColumnMajor,
+      // The pointer.
+      half const,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kH, GemmConfig_::OutputTile::kD>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      Shape<1, GemmConfig_::kThreads / GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kD>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc)
+      GemmConfig_::kScalarsPerLdgB>
+      GlobalTileTraits;
+
+  static int const kSkewB = 128 / sizeof(half) / GlobalTileTraits::Threads::kW / 2;
+
+  /// The traits class to build the iterator to store data to shared memory for B^N.
+  typedef GemmSharedStoreWithSkewTileAbTraits <
+      // The pointer.
+      half,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<GemmConfig_::kStages,
+            GemmConfig_::OutputTile::kD / GemmConfig_::InstructionShape::kD,
+            GemmConfig_::OutputTile::kH * GemmConfig_::InstructionShape::kD>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS (STS.32 or STS.128, etc).
+      2,
+      // The skew to avoid bank conflicts added in the tile W dimension.
+      kSkewB<GemmConfig_::kScalarsPerLdsB ? GemmConfig_::kScalarsPerLdsB : kSkewB>
+          SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for B^N.
+  typedef GemmSharedLoadTileBTraits<
+      // The pointer.
+      half const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      8,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_,
+    /// The functor to do the math in the epilogue.
+    typename EpilogueFunctor_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_,
+    /// The number of halfs loaded in one LDG for A.
+    int kScalarsPerLdgA_ = 2,
+    /// The number of halfs loaded in one LDG for B.
+    int kScalarsPerLdgB_ = 2,
+    /// The index.
+    typename Index_ = int>
+struct HgemmTraitsHelper {
+  /// The HGEMM config.
+  typedef HgemmConfig<OutputTile_, ThreadGemmShape_, kScalarsPerLdgA_, kScalarsPerLdgB_> GemmConfig;
+  /// The GEMM config for A.
+  typedef HgemmTileTraitsHelperA<kLayoutA_, GemmConfig> GemmTileTraitsHelperA;
+  /// The GEMM config for B.
+  typedef HgemmTileTraitsHelperB<kLayoutB_, GemmConfig> GemmTileTraitsHelperB;
+
+  /// The iterator to load A from global memory.
+  typedef GemmGlobalIteratorAb<typename GemmTileTraitsHelperA::GlobalTileTraits, Index_>
+      GlobalLoadIteratorA;
+  /// The default transformer for A.
+  typedef typename HgemmTransformerA<GemmTileTraitsHelperA::kLayout,
+                                     GlobalLoadIteratorA>::Transformer GlobalTransformerA;
+  /// The iterator to store A to shared memory.
+  typedef TileStoreIterator<typename GemmTileTraitsHelperA::SharedStoreTileTraits,
+                            typename GemmTileTraitsHelperA::SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorA;
+  /// The stream to load A from global memory to shared memory.
+  typedef GlobalLoadStream<GemmOperand::kA,
+                              GlobalLoadIteratorA,
+                              SharedStoreIteratorA,
+                              GlobalTransformerA>
+      GlobalLoadStreamA;
+
+  /// The iterator to load B from global memory.
+  typedef GemmGlobalIteratorAb<typename GemmTileTraitsHelperB::GlobalTileTraits, Index_>
+      GlobalLoadIteratorB;
+  // The default transformer for B.
+  typedef typename HgemmTransformerB<GemmTileTraitsHelperB::kLayout,
+                                     GlobalLoadIteratorB>::Transformer GlobalTransformerB;
+  /// The iterator to store B to shared memory.
+  typedef TileStoreIterator<typename GemmTileTraitsHelperB::SharedStoreTileTraits,
+                            typename GemmTileTraitsHelperB::SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorB;
+  /// The stream to load B from global memory to shared memory.
+  typedef GlobalLoadStream<GemmOperand::kB,
+                              GlobalLoadIteratorB,
+                              SharedStoreIteratorB,
+                              GlobalTransformerB>
+      GlobalLoadStreamB;
+
+  /// The iterator to load A from shared memory
+  typedef TileLoadIterator<typename GemmTileTraitsHelperA::SharedLoadTileTraits,
+                           typename GemmTileTraitsHelperA::SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorA;
+  /// The stream to load A from shared memory.
+  typedef SharedLoadStream<SharedLoadIteratorA> SharedLoadStreamA;
+  /// The iterator to load B from shared memory.
+  typedef TileLoadIterator<typename GemmTileTraitsHelperB::SharedLoadTileTraits,
+                           typename GemmTileTraitsHelperB::SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorB;
+  /// The stream to load B from shared memory.
+  typedef SharedLoadStream<SharedLoadIteratorB> SharedLoadStreamB;
+
+  /// The functor to do the multiply-add in the main loop.
+  typedef typename GemmConfig::MultiplyAdd MultiplyAdd;
+  /// The object to clear accumulators.
+  typedef ClearAccumulators<typename MultiplyAdd::ScalarC> ClearAccumulators;
+
+  /// The traits class for the epilogue.
+  typedef SimplifiedGemmEpilogueTraits<GemmConfig, EpilogueFunctor_, Index_> GemmEpilogueTraits;
+  /// The epilogue.
+  typedef GemmEpilogue<GemmEpilogueTraits> Epilogue;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_ = Shape<8, 128, 128>,
+    /// The functor to do the math in the epilogue.
+    typename EpilogueFunctor_ = LinearScaling<half>,
+    /// Tile size for warp-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<8, 8, 16>,
+    /// The number of halfs loaded in one LDG for A.
+    int kScalarsPerLdgA_ = 2,
+    /// The number of halfs loaded in one LDG for B.
+    int kScalarsPerLdgB_ = 2,
+    /// The index.
+    typename Index_ = int,
+    /// The helper class.
+    typename Helper_ = HgemmTraitsHelper<kLayoutA_,
+                                         kLayoutB_,
+                                         OutputTile_,
+                                         EpilogueFunctor_,
+                                         ThreadGemmShape_,
+                                         kScalarsPerLdgA_,
+                                         kScalarsPerLdgB_,
+                                         Index_> >
+struct HgemmTraits : public GemmTraits<
+                         // The config.
+                         typename Helper_::GemmConfig,
+                         // The stream to load A from global memory to shared memory.
+                         typename Helper_::GlobalLoadStreamA,
+                         // The stream to load B from global memory to shared memory.
+                         typename Helper_::GlobalLoadStreamB,
+                         // The stream to load A from shared memory.
+                         typename Helper_::SharedLoadStreamA,
+                         // The stream to load B from shared memory.
+                         typename Helper_::SharedLoadStreamB,
+                         // The epilogue.
+                         typename Helper_::Epilogue,
+                         // The block swizzle to reorganize the grid.
+                         IdentityBlockSwizzle,
+                         // The index.
+                         Index_,
+                         // The tool used to clear accumulators.
+                         typename Helper_::ClearAccumulators> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/igemm_epilogue.h

@@ -0,0 +1,318 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines the epilogue phase of the GEMM computation for IGEMM, supporting integer and
+      floating-point output matrix formats.
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/fragment.h"
+#include "cutlass/gemm/gemm_global_stream.h"
+#include "cutlass/gemm/gemm_shared_stream.h"
+#include "cutlass/gemm/igemm_global_tile.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/tile_iterator.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kElements_>
+struct IgemmFloatToInt8Converter {
+  /// The input fragment.
+  typedef Fragment<float, kElements_> InputFragment;
+  /// The output fragment.
+  typedef Fragment<int8_t, kElements_> OutputFragment;
+
+  // We are packing 4 floats into int32 registers so we need kElements to be multiple of 4.
+  static_assert(kElements_ % 4 == 0, "kElements must be multiple of 4");
+
+  /// Ctor.
+  CUTLASS_DEVICE IgemmFloatToInt8Converter() {}
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(InputFragment const& src, OutputFragment& dst) {
+    transform(src, 0, dst);
+  }
+
+  /// Transform a fragment.
+  template <typename Fragment_>
+  CUTLASS_DEVICE void transform(Fragment_ const& src, int offset, OutputFragment& dst) {
+    // The inputs.
+    float4 const* src_f4 = reinterpret_cast<float4 const*>(&src[0]);
+    // The outputs.
+    int* dst_int = reinterpret_cast<int*>(&dst[0]);
+
+    // Iterate over the floats and pack them together to produce ints.
+    for (int i = 0; i < kElements_ / 4; ++i) {
+      // Read the float4.
+      float4 f4 = src_f4[i];
+
+      // Clamp the 4 elements of the floats to the [-128, +127] range.
+      float x = fmaxf(-128.f, fminf(127.f, f4.x));
+      float y = fmaxf(-128.f, fminf(127.f, f4.y));
+      float z = fmaxf(-128.f, fminf(127.f, f4.z));
+      float w = fmaxf(-128.f, fminf(127.f, f4.w));
+
+      // Convert to integers.
+      int ix = (int)x;
+      int iy = (int)y;
+      int iz = (int)z;
+      int iw = (int)w;
+
+      // Extract the lower bytes to build an int32 with 4 int8.
+      asm volatile("prmt.b32 %0, %0, %1, 0x1140;" : "+r"(ix) : "r"(iy));
+      asm volatile("prmt.b32 %0, %0, %1, 0x1140;" : "+r"(iz) : "r"(iw));
+      asm volatile("prmt.b32 %0, %0, %1, 0x5410;" : "+r"(ix) : "r"(iz));
+
+      // Store the int.
+      dst_int[i] = ix;
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename InputScalar_, typename OutputFragment_>
+struct IgemmGlobalStoreTransformer {
+  typedef Convert<Fragment<InputScalar_, OutputFragment_::kElements>, OutputFragment_> Transformer;
+};
+
+template <int kElements_>
+struct IgemmGlobalStoreTransformer<float, Fragment<int8_t, kElements_> > {
+  typedef IgemmFloatToInt8Converter<kElements_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kElements_>
+struct IgemmInt8ToFloatConverter {
+  /// The input fragment.
+  typedef Fragment<int8_t, kElements_> InputFragment;
+  /// The output fragment.
+  typedef Fragment<float, kElements_> OutputFragment;
+
+  // We are unpacking 4 int8s from int32.
+  static_assert(kElements_ % 4 == 0, "kElements must be multiple of 4");
+
+  /// Ctor.
+  CUTLASS_DEVICE IgemmInt8ToFloatConverter() {}
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(InputFragment const& src, OutputFragment& dst) {
+    transform(src, 0, dst);
+  }
+
+  /// Transform a fragment.
+  template <typename Fragment_>
+  CUTLASS_DEVICE void transform(Fragment_ const& src, int offset, OutputFragment& dst) {
+    // The inputs.
+    int const* src_int = reinterpret_cast<int const*>(&src[0]);
+    // The outputs.
+    float4* dst_f4 = reinterpret_cast<float4*>(&dst[0]);
+
+    // Iterate over the int8 and unpack them together to produce floats.
+    for (int i = 0; i < kElements_ / 4; ++i) {
+      // Read the int.
+      int ix, iy, iz, iw = src_int[i];
+
+      // Extract the 4 bytes.
+      asm volatile("prmt.b32 %0, 0x0, %1, 0x4440;" : "=r"(ix) : "r"(iw));
+      asm volatile("prmt.b32 %0, 0x0, %1, 0x4441;" : "=r"(iy) : "r"(iw));
+      asm volatile("prmt.b32 %0, 0x0, %1, 0x4442;" : "=r"(iz) : "r"(iw));
+      asm volatile("prmt.b32 %0, 0x0, %1, 0x4443;" : "=r"(iw) : "r"(iw));
+
+      // The floats.
+      float fx, fy, fz, fw;
+
+      // Convert to floats (make sure we generate I2F.F32.S8).
+      asm volatile("cvt.rn.f32.s8 %0, %1;" : "=f"(fx) : "r"(ix));
+      asm volatile("cvt.rn.f32.s8 %0, %1;" : "=f"(fy) : "r"(iy));
+      asm volatile("cvt.rn.f32.s8 %0, %1;" : "=f"(fz) : "r"(iz));
+      asm volatile("cvt.rn.f32.s8 %0, %1;" : "=f"(fw) : "r"(iw));
+
+      // Store the float4.
+      dst_f4[i] = make_float4(fx, fy, fz, fw);
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename InputFragment_, typename OutputScalar_>
+struct IgemmGlobalLoadTransformer {
+  typedef Convert<InputFragment_, Fragment<OutputScalar_, InputFragment_::kElements> > Transformer;
+};
+
+template <int kElements_>
+struct IgemmGlobalLoadTransformer<Fragment<int8_t, kElements_>, float> {
+  typedef IgemmInt8ToFloatConverter<kElements_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename InputScalar_, typename OutputFragment_>
+struct IgemmSharedStoreTransformer {
+  typedef Convert<Fragment<InputScalar_, OutputFragment_::kElements>, OutputFragment_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename IgemmConfig_, typename EpilogueFunctor_, typename Index_>
+struct IgemmEpilogueTraitsHelper
+    : public GemmEpilogueTraitsHelper<IgemmConfig_, EpilogueFunctor_, Index_> {
+  /// The base class.
+  typedef GemmEpilogueTraitsHelper<IgemmConfig_, EpilogueFunctor_, Index_> Base;
+  /// The config.
+  typedef IgemmConfig_ IgemmConfig;
+
+  /// The scalar type of the epilogue.
+  typedef typename Base::Scalar Scalar;
+  /// The iterations.
+  typedef typename Base::Iterations Iterations;
+  /// The iterations strides.
+  typedef typename Base::Delta Delta;
+
+  /// The traits class for the iterator.
+  typedef typename Base::GlobalLoadTileTraits GlobalLoadTileTraits;
+  /// The iterator to store to shared memory.
+  typedef GemmGlobalIteratorCd<GlobalLoadTileTraits> GlobalLoadIteratorC;
+  /// The fragment that needs to be produced by the load iterator.
+  typedef typename GlobalLoadIteratorC::Fragment GlobalFragmentC;
+  /// The transformer from loaded data to math fragment.
+  typedef
+      typename IgemmGlobalLoadTransformer<GlobalFragmentC, Scalar>::Transformer GlobalTransformerC;
+
+  /// The traits class for the iterator.
+  typedef typename Base::GlobalStoreTileTraits GlobalStoreTileTraits;
+  /// The iterator to store to shared memory.
+  typedef GemmGlobalIteratorCd<GlobalStoreTileTraits> GlobalStoreIteratorD;
+  /// The fragment that needs to be passed to that store iterator.
+  typedef typename GlobalStoreIteratorD::Fragment GlobalFragmentD;
+  /// The transformer from accumulators to shared memory fragments.
+  typedef
+      typename IgemmGlobalStoreTransformer<Scalar, GlobalFragmentD>::Transformer GlobalTransformerD;
+
+  /// The traits class for the shared iterator to store D to shared memory.
+  typedef typename Base::SharedStoreTileTraits SharedStoreTileTraits;
+  /// The shared iterator to store D to shared memory.
+  typedef TileStoreIterator<SharedStoreTileTraits,
+                            typename SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kGlobal>
+      SharedStoreIteratorD;
+  /// The fragment that needs to be passed to that store iterator.
+  typedef typename SharedStoreIteratorD::Fragment SharedStoreFragmentD;
+  /// The transformer from accumulators to shared memory fragments.
+  typedef typename IgemmSharedStoreTransformer<typename IgemmConfig::Accumulators::Element,
+                                               SharedStoreFragmentD>::Transformer
+      SharedStoreTransformerD;
+  /// The traits class for the shared iterator to load D from shared memory.
+  typedef typename Base::SharedLoadTileTraits SharedLoadTileTraits;
+  /// The shared iterator to load D from shared memory.
+  typedef TileLoadIterator<SharedLoadTileTraits,
+                           typename SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorD;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The config.
+    typename IgemmConfig_,
+    /// The functor to do the math in the epilogue.
+    typename EpilogueFunctor_,
+    /// The index.
+    typename Index_ = int,
+    /// The helper class to assemble the traits.
+    typename Helper_ = IgemmEpilogueTraitsHelper<IgemmConfig_, EpilogueFunctor_, Index_> >
+struct IgemmEpilogueTraits : public GemmEpilogueTraits<
+                                 // The output tile.
+                                 typename IgemmConfig_::OutputTile,
+                                 // The accumulators.
+                                 typename IgemmConfig_::Accumulators,
+                                 // The global iterator for C.
+                                 typename Helper_::GlobalLoadIteratorC,
+                                 // The transformer for C.
+                                 typename Helper_::GlobalTransformerC,
+                                 // The transformer for D.
+                                 typename Helper_::GlobalTransformerD,
+                                 // The global iterator for D.
+                                 typename Helper_::GlobalStoreIteratorD,
+                                 // The iterator to store D to shared memory.
+                                 typename Helper_::SharedStoreIteratorD,
+                                 // The shared store transformer for D.
+                                 typename Helper_::SharedStoreTransformerD,
+                                 // The stream to load D from shared memory.
+                                 typename Helper_::SharedLoadStreamD,
+                                 // The iterations.
+                                 typename Helper_::Iterations,
+                                 // The strides between iterations.
+                                 typename Helper_::Delta,
+                                 // The functor to be used in the epilogue.
+                                 EpilogueFunctor_,
+                                 // The index.
+                                 Index_> {
+  /// Do we output in int8?
+  static bool const kInt8Output =
+      platform::is_same<typename IgemmConfig_::ScalarC, int8_t>::value != 0;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmEpilogueTraits_, bool = GemmEpilogueTraits_::kInt8Output>
+struct IgemmEpilogue : public GemmEpilogue<GemmEpilogueTraits_> {
+  /// The base class.
+  typedef GemmEpilogue<GemmEpilogueTraits_> Base;
+
+  /// Ctor.
+  CUTLASS_DEVICE IgemmEpilogue(typename Base::Params const& params_,
+                               typename Base::SharedStorage& shared_storage_,
+                               Coord<3> const& _problem_size)
+      : Base(params_, shared_storage_, _problem_size) {}
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmEpilogueTraits_>
+struct IgemmEpilogue<GemmEpilogueTraits_, true> : public GemmEpilogue<GemmEpilogueTraits_> {
+  /// The base class.
+  typedef GemmEpilogue<GemmEpilogueTraits_> Base;
+
+  /// Ctor.
+  CUTLASS_DEVICE IgemmEpilogue(typename Base::Params const& params_,
+                               typename Base::SharedStorage& shared_storage_,
+                               Coord<3> const& _problem_size)
+      : Base(params_, shared_storage_, _problem_size) {}
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/igemm_global_tile.h

@@ -0,0 +1,132 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements tile iterators to partition the thread block tile into 2D subtiles and
+      efficiently load each. Applies permute transformation to construct 'interleaved K-strided'
+      data layout in which 4-element dot products from the same K index are arranged in consecutive
+      locations within shared memory.
+
+      Supports efficient loads from shared memory to target the DP4A instruction.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/matrix_traits.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <GemmOperand::Kind kOperand_,
+          MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename Tile_,
+          typename Threads_,
+          int kAccessSize_>
+struct IgemmGlobalTileTraits : public GemmGlobalTileTraits<
+                                   // Which GEMM operand?
+                                   kOperand_,
+                                   // The layout.
+                                   kLayout_,
+                                   // The scalar.
+                                   Scalar_,
+                                   // The tile.
+                                   Tile_,
+                                   // The threads.
+                                   Threads_,
+                                   // The number of scalars per LDG/STG.
+                                   kAccessSize_> {
+  /// The base class.
+  typedef GemmGlobalTileTraits<kOperand_, kLayout_, Scalar_, Tile_, Threads_, kAccessSize_> Base;
+  /// The threads.
+  typedef typename Base::Threads Threads;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<Base::Threads::kH * 4, 1, Base::Threads::kW, Base::kAccessSize> Delta;
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<Base::VectorizedTile::kH / Base::Threads::kH / 4,
+                4,
+                Base::VectorizedTile::kW / Base::Threads::kW,
+                Base::VectorizedTile::kC / Base::kAccessSize>
+      Iterations;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int thread_offset_h = threadIdx.x / Threads::kW * ThreadsDelta::kH;
+      int thread_offset_w = threadIdx.x % Threads::kW * ThreadsDelta::kW;
+
+      return make_Coord(0, thread_offset_h, thread_offset_w, 0);
+    }
+  };
+
+ public:
+  /// The threads strides.
+  typedef Shape<1, 4, Base::VectorizedTile::kC> ThreadsDelta;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename TileTraits_, typename Index_ = int>
+struct IgemmGlobalIteratorAb : public GemmGlobalIteratorAb<TileTraits_, Index_> {
+  /// The base class.
+  typedef GemmGlobalIteratorAb<TileTraits_, Index_> Base;
+  /// The functor to compute the thread offset.
+  typedef typename TileTraits_::ThreadOffset ThreadOffset;
+
+  /// Constructor.
+  CUTLASS_DEVICE IgemmGlobalIteratorAb(typename Base::Params const& _params,
+                                       const Coord<3>& bounds,
+                                       const Coord<3>& threadblock_offset,
+                                       ThreadOffset thread_offset_func = ThreadOffset())
+      : Base(_params, bounds, threadblock_offset, thread_offset_func), mask_(0xffffffff) {
+    // The number of elements read in a single iteration.
+    int const kBlock = TileTraits_::Tile::kW;
+    // The residue.
+    int const kResidue = (int)(bounds[1] % kBlock);
+
+    // Compute the number of elements that are valid.
+    int const left = kResidue - Base::thread_offset[2];
+    if (left > 0 && left < 4) {
+      mask_ = (1u << (8 * left)) - 1u;
+    }
+  }
+
+  CUTLASS_DEVICE void load_element(
+      typename Base::AccessType& value, int d, int h, int w, int c) const {
+    Base::load_element(value, d, h, w, c);
+    reinterpret_cast<uint32_t&>(value) &= mask_;
+  }
+
+  /// The mask to clean up the values.
+  uint32_t mask_;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/igemm_multiply_add.h

@@ -0,0 +1,91 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements matrix multiply accumulate operation of 8-bit integer data using DP4A
+   instruction.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+
+#include "cutlass/gemm/thread_multiply_add.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Template performing matrix multiply-add operation within a thread
+template <typename ThreadGemmShape_, typename ThreadsPerWarp_>
+struct ThreadMultiplyAdd<ThreadGemmShape_, ThreadsPerWarp_, int8_t, int8_t, int> {
+  /// The shape of the instruction.
+  typedef Shape<4, 1, 1> InstructionShape;
+  /// Shape of the thread-level GEMM (K-by-N-by-M)
+  typedef ThreadGemmShape_ ThreadGemmShape;
+  /// Aliased for compatibility. Will be removed in CUTLASS v2.0
+  typedef ThreadGemmShape AccumulatorsPerThread;
+  /// The number of threads per warp.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of accumulators per warp.
+  typedef typename ShapeMul<ThreadGemmShape, ThreadsPerWarp>::Shape AccumulatorsPerWarp;
+  /// The type for A.
+  typedef int8_t ScalarA;
+  /// The fragment for A.
+  typedef Fragment<ScalarA, AccumulatorsPerThread::kW * 4> FragmentA;
+  /// The type for B.
+  typedef int8_t ScalarB;
+  /// The fragment for B.
+  typedef Fragment<ScalarB, AccumulatorsPerThread::kH * 4> FragmentB;
+  /// The type for C and D.
+  typedef int ScalarC;
+  /// The accumulators.
+  typedef Fragment<ScalarC, AccumulatorsPerThread::kH * AccumulatorsPerThread::kW> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE ThreadMultiplyAdd() {}
+
+  /// Multiply : d = a*b + c.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    // The inputs.
+    int const* a_int = reinterpret_cast<int const*>(&a[0]);
+    int const* b_int = reinterpret_cast<int const*>(&b[0]);
+
+    for (int j = 0; j < AccumulatorsPerThread::kH; ++j) {
+      for (int i = 0; i < AccumulatorsPerThread::kW; ++i) {
+        asm volatile("dp4a.s32.s32 %0, %1, %2, %3;"
+                     : "=r"(d[j * AccumulatorsPerThread::kW + i])
+                     : "r"(a_int[i]), "r"(b_int[j]), "r"(c[j * AccumulatorsPerThread::kW + i]));
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/igemm_swizzle.h

@@ -0,0 +1,125 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Transposes a fragment of data containing packed 8-bit integer elements.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GlobalIterator_>
+struct IgemmSwizzle {
+  /// The global iterator.
+  typedef GlobalIterator_ GlobalIterator;
+  /// The source fragment.
+  typedef typename GlobalIterator::Fragment Fragment;
+  /// The shape of the source fragment.
+  typedef typename GlobalIterator::FragmentShape FragmentShape;
+
+  /// The source fragment.
+  typedef Fragment InputFragment;
+  /// The destination fragment.
+  typedef Fragment OutputFragment;
+
+  /// The src/dst must be int8 fragments.
+  static_assert((platform::is_same<typename Fragment::Element, int8_t>::value), "Works on int8");
+
+  /// The number of elements must be a multiple of 4.
+  static_assert(FragmentShape::kH % 4 == 0 && ShapeCount<FragmentShape>::kWc % 4 == 0,
+                "Not multiple of 4");
+
+  /// Ctor.
+  CUTLASS_DEVICE IgemmSwizzle() {}
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(Fragment const& src, Fragment& dst) {
+    // Expose src/dst as int arrays.
+    int const* src_int = reinterpret_cast<int const*>(&src[0]);
+    int* dst_int = reinterpret_cast<int*>(&dst[0]);
+
+    // Transpose the data.
+    for (int d = 0; d < FragmentShape::kD; ++d) {
+      for (int h = 0; h < FragmentShape::kH / 4; ++h) {
+        for (int w = 0; w < ShapeCount<FragmentShape>::kWc / 4; ++w) {
+          int const i0 = d * (ShapeCount<FragmentShape>::kHwc / 4) +
+                         (4 * h + 0) * (ShapeCount<FragmentShape>::kWc / 4) + w;
+          int const i1 = d * (ShapeCount<FragmentShape>::kHwc / 4) +
+                         (4 * h + 1) * (ShapeCount<FragmentShape>::kWc / 4) + w;
+          int const i2 = d * (ShapeCount<FragmentShape>::kHwc / 4) +
+                         (4 * h + 2) * (ShapeCount<FragmentShape>::kWc / 4) + w;
+          int const i3 = d * (ShapeCount<FragmentShape>::kHwc / 4) +
+                         (4 * h + 3) * (ShapeCount<FragmentShape>::kWc / 4) + w;
+
+          int a0 = src_int[i0];
+          int a1 = src_int[i1];
+          int a2 = src_int[i2];
+          int a3 = src_int[i3];
+
+          // // DEBUG.
+          // if (threadIdx.x == 0) {
+          //     printf("a=0x%08x 0x%08x 0x%08x 0x%08x\n", a0, a1, a2, a3);
+          // }
+
+          int b0, b1, b2, b3, c0;
+          asm volatile("prmt.b32 %0, %1, %2, 0x0040;" : "=r"(b0) : "r"(a0), "r"(a1));
+          asm volatile("prmt.b32 %0, %1, %2, 0x0040;" : "=r"(c0) : "r"(a2), "r"(a3));
+          asm volatile("prmt.b32 %0, %1, %2, 0x5410;" : "=r"(b0) : "r"(b0), "r"(c0));
+
+          asm volatile("prmt.b32 %0, %1, %2, 0x0051;" : "=r"(b1) : "r"(a0), "r"(a1));
+          asm volatile("prmt.b32 %0, %1, %2, 0x0051;" : "=r"(c0) : "r"(a2), "r"(a3));
+          asm volatile("prmt.b32 %0, %1, %2, 0x5410;" : "=r"(b1) : "r"(b1), "r"(c0));
+
+          asm volatile("prmt.b32 %0, %1, %2, 0x0062;" : "=r"(b2) : "r"(a0), "r"(a1));
+          asm volatile("prmt.b32 %0, %1, %2, 0x0062;" : "=r"(c0) : "r"(a2), "r"(a3));
+          asm volatile("prmt.b32 %0, %1, %2, 0x5410;" : "=r"(b2) : "r"(b2), "r"(c0));
+
+          asm volatile("prmt.b32 %0, %1, %2, 0x0073;" : "=r"(b3) : "r"(a0), "r"(a1));
+          asm volatile("prmt.b32 %0, %1, %2, 0x0073;" : "=r"(c0) : "r"(a2), "r"(a3));
+          asm volatile("prmt.b32 %0, %1, %2, 0x5410;" : "=r"(b3) : "r"(b3), "r"(c0));
+
+          // // DEBUG.
+          // if (threadIdx.x == 0) {
+          //     printf("b=0x%08x 0x%08x 0x%08x 0x%08x\n", b0, b1, b2, b3);
+          // }
+
+          dst_int[i0] = b0;
+          dst_int[i1] = b1;
+          dst_int[i2] = b2;
+          dst_int[i3] = b3;
+        }
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/igemm_traits.h

@@ -0,0 +1,550 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defies structural properties of mixed-precision integer GEMM. Multiplicands are assumed
+      to be packed 8bit integers, accumulators are assumed to be 32b signed integers, and output
+      formats vary.
+*/
+#pragma once
+
+#include "cutlass/convert.h"
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/gemm/gemm_epilogue.h"
+#include "cutlass/gemm/gemm_epilogue_traits.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/gemm/gemm_shared_tile.h"
+#include "cutlass/gemm/gemm_traits.h"
+#include "cutlass/gemm/igemm_epilogue.h"
+#include "cutlass/gemm/igemm_global_tile.h"
+#include "cutlass/gemm/igemm_multiply_add.h"
+#include "cutlass/gemm/igemm_swizzle.h"
+#include "cutlass/reshape_tile.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The tile size for the GEMM KxNxM.
+    typename OutputTile_,
+    /// The output type.
+    typename ScalarD_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_>
+struct IgemmConfig : public GemmConfig<
+                         /// The scalar type for A.
+                         int8_t,
+                         /// The scalar type for B.
+                         int8_t,
+                         /// The scalar type for C.
+                         ScalarD_,
+                         /// The scalar type for D.
+                         ScalarD_,
+                         /// The tile size for the GEMM KxNxM.
+                         OutputTile_,
+                         /// The functor to do the math in the main loop.
+                         ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, int8_t, int8_t, int>,
+                         /// The number of scalars per LDG for A.
+                         4,
+                         /// The number of scalars per STS for A.
+                         4,
+                         /// The number of scalars per LDS for A.
+                         16,
+                         /// The number of scalars per LDG for B.
+                         4,
+                         /// The number of scalars per STS for B.
+                         4,
+                         /// The number of scalars per LDS for B.
+                         16,
+                         /// The number of scalars per LDG for C and STG for D.
+                         1,
+                         /// The number of scalars per STS for D.
+                         4,
+                         /// The number of scalars per LDS for D.
+                         1,
+                         /// The number of stages in shared memory.
+                         2,
+                         /// kResidueSeparate
+                         false,
+                         /// kResidueInPrologue
+                         false,
+                         /// kLaunchBounds
+                         false> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename OutputTile_, typename ThreadGemmShape_>
+struct IgemmConfig<OutputTile_, int8_t, ThreadGemmShape_>
+    : public GemmConfig<
+          /// The scalar type for A.
+          int8_t,
+          /// The scalar type for B.
+          int8_t,
+          /// The scalar type for C.
+          int8_t,
+          /// The scalar type for D.
+          int8_t,
+          /// The tile size for the GEMM KxNxM.
+          OutputTile_,
+          /// The functor to do the math in the main loop.
+          ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, int8_t, int8_t, int>,
+          /// The number of scalars per LDG for A.
+          4,
+          /// The number of scalars per STS for A.
+          4,
+          /// The number of scalars per LDS for A.
+          16,
+          /// The number of scalars per LDG for B.
+          4,
+          /// The number of scalars per STS for B.
+          4,
+          /// The number of scalars per LDS for B.
+          16,
+          /// The number of scalars per LDG for C and STG for D.
+          4,
+          /// The number of scalars per STS for D.
+          4,
+          /// The number of scalars per LDS for D.
+          4,
+          /// The number of stages in shared memory.
+          2,
+          /// If true, separate mainloop is instantiated from residue
+          false,
+          /// Compute residue in prolog?
+          true,
+          /// Launch bounds?
+          false> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename GemmConfig_, typename Index_>
+struct IgemmTileTraitsHelperA : public GemmTileTraitsHelperA<kLayout_, GemmConfig_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_, typename Index_>
+struct IgemmTileTraitsHelperA<MatrixLayout::kColumnMajor, GemmConfig_, Index_>
+    : public GemmTileTraitsHelperA<MatrixLayout::kColumnMajor, GemmConfig_> {
+  /// The base config.
+  typedef GemmTileTraitsHelperA<MatrixLayout::kColumnMajor, GemmConfig_> Base;
+
+  /// The number of scalars per LDG/STS/LDS for A.
+  static int const kScalarsPerStsA = 16;
+
+  /// The traits class to build the iterator to load data from global memory for A^N.
+  typedef IgemmGlobalTileTraits<
+      GemmOperand::kA,
+      // The layout.
+      MatrixLayout::kColumnMajor,
+      // The pointer is float const.
+      int8_t const,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kW>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgA>
+      GlobalTileTraits;
+
+  /// The global load iterator.
+  typedef GemmGlobalIteratorAb<GlobalTileTraits, Index_> GlobalLoadIterator;
+
+  /// The traits class to build the iterator to store data to shared memory for A^N.
+  typedef GemmSharedStoreTileAbTraits<
+      // The pointer is float.
+      int8_t,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<GemmConfig_::kStages, GemmConfig_::OutputTile::kD / 4, GemmConfig_::OutputTile::kW * 4>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS (STS.32 or STS.128, etc).
+      kScalarsPerStsA>
+      SharedStoreTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_, typename Index_>
+struct IgemmTileTraitsHelperA<MatrixLayout::kRowMajor, GemmConfig_, Index_> {
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = MatrixLayout::kRowMajor;
+
+  /// The input scalar.
+  typedef int8_t Scalar;
+  /// The scalar stored in shared memory.
+  typedef int8_t MultiplyAddScalar;
+
+  /// The number of scalars per LDG/STS/LDS for A.
+  static int const kScalarsPerStsA = 16;
+
+  /// The traits class to build the iterator to load data from global memory for A^T.
+  typedef IgemmGlobalTileTraits<
+      GemmOperand::kA,
+      // The layout.
+      MatrixLayout::kRowMajor,
+      // The pointer is float const.
+      int8_t const,
+      // The tile has size NxK in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kW, GemmConfig_::OutputTile::kD>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgA>
+      GlobalTileTraits;
+
+  /// The global load iterator.
+  typedef IgemmGlobalIteratorAb<GlobalTileTraits, Index_> GlobalLoadIterator;
+
+  /// The traits class to build the iterator to store data to shared memory for A^N.
+  typedef GemmSharedStoreWithSkewTileAbTraits<
+      // The pointer is int8.
+      int8_t,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<GemmConfig_::kStages, GemmConfig_::OutputTile::kD / 4, GemmConfig_::OutputTile::kW * 4>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS.
+      kScalarsPerStsA,
+      // The skew to avoid bank conflicts added in the tile W dimension.
+      16>
+      SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for A^N.
+  typedef GemmSharedLoadTileATraits<
+      // The pointer is float const.
+      int8_t const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      16,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename GemmConfig_, typename Index_>
+struct IgemmTileTraitsHelperB : public GemmTileTraitsHelperB<kLayout_, GemmConfig_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_, typename Index_>
+struct IgemmTileTraitsHelperB<MatrixLayout::kColumnMajor, GemmConfig_, Index_> {
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = MatrixLayout::kColumnMajor;
+
+  /// The input scalar.
+  typedef int8_t Scalar;
+  /// The scalar stored in shared memory.
+  typedef int8_t MultiplyAddScalar;
+
+  /// The number of scalars per LDG/STS/LDS for B.
+  static int const kScalarsPerStsB = 16;
+
+  /// The traits class to build the iterator to load data from global memory for B^T.
+  typedef IgemmGlobalTileTraits<
+      GemmOperand::kB,
+      // The layout.
+      MatrixLayout::kColumnMajor,
+      // The pointer is float const.
+      int8_t const,
+      // The tile has size NxK in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kH, GemmConfig_::OutputTile::kD>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgB>
+      GlobalTileTraits;
+
+  /// The global load iterator.
+  typedef IgemmGlobalIteratorAb<GlobalTileTraits, Index_> GlobalLoadIterator;
+
+  /// The traits class to build the iterator to store data to shared memory for B^N.
+  typedef GemmSharedStoreWithSkewTileAbTraits<
+      // The pointer is int8.
+      int8_t,
+      // The tile has size KxN in GEMM's terminology.
+      Shape<GemmConfig_::kStages, GemmConfig_::OutputTile::kD / 4, GemmConfig_::OutputTile::kH * 4>,
+      // The threads are distributed as (threads / K) x K (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS.
+      kScalarsPerStsB,
+      // The skew to avoid bank conflicts added in the tile W dimension.
+      16>
+      SharedStoreTileTraits;
+
+  /// The traits class to build the iterator to load from shared memory for B^N.
+  typedef GemmSharedLoadTileBTraits<
+      // The pointer is float const.
+      int8_t const,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The number of threads per warp.
+      typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
+      // The shape of the FMA instruction.
+      typename GemmConfig_::InstructionShape,
+      // The number of stages.
+      GemmConfig_::kStages,
+      // The number of scalars per LDS.
+      16,
+      // The skew.
+      SharedStoreTileTraits::kSkew>
+      SharedLoadTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_, typename Index_>
+struct IgemmTileTraitsHelperB<MatrixLayout::kRowMajor, GemmConfig_, Index_>
+    : public GemmTileTraitsHelperB<MatrixLayout::kRowMajor, GemmConfig_> {
+  /// The base config.
+  typedef GemmTileTraitsHelperB<MatrixLayout::kRowMajor, GemmConfig_> Base;
+
+  /// The number of scalars per LDG/STS/LDS for B.
+  static int const kScalarsPerStsB = 16;
+
+  /// The traits class to build the iterator to load data from global memory for B^T.
+  typedef IgemmGlobalTileTraits<
+      GemmOperand::kB,
+      // The layout.
+      MatrixLayout::kRowMajor,
+      // The pointer is float const.
+      int8_t const,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<1, GemmConfig_::OutputTile::kD, GemmConfig_::OutputTile::kH>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgB>
+      GlobalTileTraits;
+
+  /// The global load iterator.
+  typedef GemmGlobalIteratorAb<GlobalTileTraits, Index_> GlobalLoadIterator;
+
+  /// The traits class to build the iterator to store data to shared memory for B^N.
+  typedef GemmSharedStoreTileAbTraits<
+      // The pointer is float.
+      int8_t,
+      // The tile has size KxM in GEMM's terminology.
+      Shape<GemmConfig_::kStages, GemmConfig_::OutputTile::kD / 4, GemmConfig_::OutputTile::kH * 4>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      typename GlobalTileTraits::Threads,
+      // The number of scalars per STS (STS.32 or STS.128, etc).
+      kScalarsPerStsB>
+      SharedStoreTileTraits;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename Iterator_>
+struct IgemmTransformerA {};
+
+template <typename Iterator_>
+struct IgemmTransformerA<MatrixLayout::kRowMajor, Iterator_> {
+  typedef Copy<typename Iterator_::Fragment> Transformer;
+};
+
+template <typename Iterator_>
+struct IgemmTransformerA<MatrixLayout::kColumnMajor, Iterator_> {
+  typedef IgemmSwizzle<Iterator_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <enum MatrixLayout::Kind kLayout_, typename Iterator_>
+struct IgemmTransformerB {};
+
+template <typename Iterator_>
+struct IgemmTransformerB<MatrixLayout::kColumnMajor, Iterator_> {
+  typedef Copy<typename Iterator_::Fragment> Transformer;
+};
+
+template <typename Iterator_>
+struct IgemmTransformerB<MatrixLayout::kRowMajor, Iterator_> {
+  typedef IgemmSwizzle<Iterator_> Transformer;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_,
+    /// The output type.
+    typename ScalarD_,
+    /// The functor to do the math in the epilogue.
+    typename EpilogueFunctor_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<32, 8, 8>,
+    /// The index.
+    typename Index_ = int>
+struct IgemmTraitsHelper {
+  /// The IGEMM config.
+  typedef IgemmConfig<OutputTile_, ScalarD_, ThreadGemmShape_> GemmConfig;
+  /// The GEMM config for A.
+  typedef IgemmTileTraitsHelperA<kLayoutA_, GemmConfig, Index_> GemmTileTraitsHelperA;
+  /// The GEMM config for B.
+  typedef IgemmTileTraitsHelperB<kLayoutB_, GemmConfig, Index_> GemmTileTraitsHelperB;
+
+  /// The iterator to load A from global memory.
+  typedef typename GemmTileTraitsHelperA::GlobalLoadIterator GlobalLoadIteratorA;
+  /// The default transformer for A.
+  typedef typename IgemmTransformerA<GemmTileTraitsHelperA::kLayout,
+                                     GlobalLoadIteratorA>::Transformer GlobalTransformerA;
+  /// The iterator to store A to shared memory.
+  typedef TileStoreIterator<typename GemmTileTraitsHelperA::SharedStoreTileTraits,
+                            typename GemmTileTraitsHelperA::SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorA;
+  /// The stream to load A from global memory to shared memory.
+  typedef GlobalLoadStream<GemmOperand::kA,
+                              GlobalLoadIteratorA,
+                              SharedStoreIteratorA,
+                              GlobalTransformerA>
+      GlobalLoadStreamA;
+
+  /// The iterator to load B from global memory.
+  typedef typename GemmTileTraitsHelperB::GlobalLoadIterator GlobalLoadIteratorB;
+  // The default transformer for B.
+  typedef typename IgemmTransformerB<GemmTileTraitsHelperB::kLayout,
+                                     GlobalLoadIteratorB>::Transformer GlobalTransformerB;
+  /// The iterator to store B to shared memory.
+  typedef TileStoreIterator<typename GemmTileTraitsHelperB::SharedStoreTileTraits,
+                            typename GemmTileTraitsHelperB::SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared>
+      SharedStoreIteratorB;
+  /// The stream to load B from global memory to shared memory.
+  typedef GlobalLoadStream<GemmOperand::kB,
+                              GlobalLoadIteratorB,
+                              SharedStoreIteratorB,
+                              GlobalTransformerB>
+      GlobalLoadStreamB;
+
+  /// The iterator to load A from shared memory.
+  typedef TileLoadIterator<typename GemmTileTraitsHelperA::SharedLoadTileTraits,
+                           typename GemmTileTraitsHelperA::SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorA;
+  /// The stream to load A from shared memory.
+  typedef SharedLoadStream<SharedLoadIteratorA, Copy<typename SharedLoadIteratorA::Fragment> >
+      SharedLoadStreamA;
+  /// The iterator to load B from shared memory.
+  typedef TileLoadIterator<typename GemmTileTraitsHelperB::SharedLoadTileTraits,
+                           typename GemmTileTraitsHelperB::SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorB;
+  /// The stream to load B from shared memory.
+  typedef SharedLoadStream<SharedLoadIteratorB, Copy<typename SharedLoadIteratorB::Fragment> >
+      SharedLoadStreamB;
+
+  /// The multiply-add functor.
+  typedef typename GemmConfig::MultiplyAdd MultiplyAdd;
+  /// The object to clear accumulators.
+  typedef ClearAccumulators<typename MultiplyAdd::ScalarC> ClearAccumulators;
+
+  /// The epilogue.
+  typedef IgemmEpilogue<IgemmEpilogueTraits<GemmConfig, EpilogueFunctor_> > Epilogue;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename ScalarD_>
+struct IgemmEpilogueScalar {
+  typedef float Scalar;
+};
+
+template <>
+struct IgemmEpilogueScalar<int> {
+  typedef int Scalar;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_ = Shape<32, 128, 128>,
+    /// The output type.
+    typename ScalarD_ = int,
+    /// The functor to do the math in the epilogue.
+    typename EpilogueFunctor_ = LinearScaling<typename IgemmEpilogueScalar<ScalarD_>::Scalar>,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<32, 8, 8>,
+    /// The index.
+    typename Index_ = int,
+    /// The helper class.
+    typename Helper_ = IgemmTraitsHelper<kLayoutA_,
+                                         kLayoutB_,
+                                         OutputTile_,
+                                         ScalarD_,
+                                         EpilogueFunctor_,
+                                         ThreadGemmShape_,
+                                         Index_> >
+struct IgemmTraits : public GemmTraits<
+                         // The config.
+                         typename Helper_::GemmConfig,
+                         // The stream to load A from global memory to shared memory.
+                         typename Helper_::GlobalLoadStreamA,
+                         // The stream to load B from global memory to shared memory.
+                         typename Helper_::GlobalLoadStreamB,
+                         // The stream to load A from shared memory.
+                         typename Helper_::SharedLoadStreamA,
+                         // The stream to load B from shared memory.
+                         typename Helper_::SharedLoadStreamB,
+                         // The epilogue.
+                         typename Helper_::Epilogue,
+                         // The block swizzle to reorganize the grid.
+                         IdentityBlockSwizzle,
+                         // The index.
+                         Index_,
+                         // The tool used to clear accumulators.
+                         typename Helper_::ClearAccumulators> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/k_split_control.h

@@ -1,310 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Abstraction for coordinating inter-block k-splitting
- */
-
-#include <stdint.h>
-
-#include "../util/util.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * Storage and initialization
- ******************************************************************************/
-
-enum
-{
-    NumFlagsSplitK = 4096
-};
-
-
-/**
- * Global K-split semaphore flags
- *
- * TODO: use demand-allocated storage to provide copies for concurrent streams
- */
-__device__ int d_flags_split_k[NumFlagsSplitK];
-
-
-/**
- * Preparation kernel for zero-initializing semaphore flags
- */
-__global__ void prepare_kernel(int *d_flags_split_k)
-{
-    int tid = (blockIdx.x * blockDim.x) + threadIdx.x;
-    if (tid < NumFlagsSplitK)
-        d_flags_split_k[tid] = 0;
-}
-
-
-/******************************************************************************
- * k_split_control
- ******************************************************************************/
-
-/**
- * \brief Abstraction for coordinating inter-block k-splitting
- */
-struct k_split_control
-{
-    /// Extent of a thread block's partition along the GEMM K-axis
-    int split_k;
-
-    /// Whether or not to use a semaphore for inter-block k-splitting.
-    bool use_semaphore;
-
-    /// Pointer to semaphore
-    int *d_flags;
-
-
-
-    //-------------------------------------------------------------------------
-    // Device API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Return the thread block's starting coordinate (k) within the
-     * multiplicand matrices
-     */
-    inline __device__
-    int block_begin_item_k()
-    {
-        return blockIdx.z * split_k;
-    }
-
-
-    /**
-     * Return the thread block's ending coordinate (k) within the multiplicand
-     * matrices (one-past)
-     */
-    inline __device__
-    int block_end_item_k(int dim_k)
-    {
-        int next_start_k = block_begin_item_k() + split_k;
-        return __NV_STD_MIN(next_start_k, dim_k);
-    }
-
-
-    /**
-     * Whether the thread block is a secondary accumulator in an inter-block
-     * k-splitting scheme
-     */
-    inline __device__
-    bool is_secondary_accumulator()
-    {
-        return (blockIdx.z > 0);
-    }
-
-
-    /**
-     * Wait for predecessor thread block(s) to produce the exclusive
-     * partial-sums for this block-wide tile
-     */
-    inline __device__
-    void wait()
-    {
-        // Wait on semaphore
-        if ((use_semaphore) && (blockIdx.z > 0))
-        {
-            if (threadIdx.x == 0)
-            {
-                int bid = (blockIdx.y * gridDim.x) + blockIdx.x;
-                int hash = bid % NumFlagsSplitK;
-                int found;
-                int looking = blockIdx.z;
-                while (true)
-                {
-                    asm volatile ("ld.global.cg.u32 %0, [%1];\n" : "=r"(found) : "l"(d_flags + hash));
-
-                    if (found == looking)
-                        break;
-
-                    /// Fence to keep load from being hoisted from the loop
-                    __syncwarp(0x00000001);
-                }
-            }
-
-            __syncthreads();
-        }
-    }
-
-
-    /**
-     * Signal the successor thread_block(s) that the inclusive partial-sums
-     * from this block-wide tile are available
-     */
-    inline __device__
-    void signal()
-    {
-        if (use_semaphore)
-        {
-            __syncthreads();
-
-            if (threadIdx.x == 0)
-            {
-                int bid = (blockIdx.y * gridDim.x) + blockIdx.x;
-                int hash = bid % NumFlagsSplitK;
-                int val = blockIdx.z + 1;
-
-                asm volatile ("st.global.cg.u32 [%0], %1;\n" : : "l"(d_flags + hash), "r"(val));
-            }
-        }
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Grid launch API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Constructor
-     */
-    inline
-    k_split_control(
-        int     *d_flags,
-        int     sm_count,
-        int     max_sm_occupancy,
-        int     dim_k,
-        int     block_tile_items_k,
-        dim3    block_dims,
-        dim3    &grid_dims)         ///< [in,out]
-    :
-        d_flags(d_flags),
-        split_k(dim_k)
-    {
-        // Compute wave efficiency
-        float wave_efficiency = get_wave_efficiency(
-            sm_count,
-            max_sm_occupancy,
-            block_dims,
-            grid_dims);
-
-        // Update split-k if wave efficiency is less than some threshold
-        if (wave_efficiency < 0.9)
-        {
-            int num_threadblocks = grid_dims.x * grid_dims.y * grid_dims.z;
-
-            // Ideal number of thread blocks in grid
-            int ideal_threadblocks = lcm(sm_count, num_threadblocks);
-
-            // Desired number of partitions to split K-axis into
-            int num_partitions = ideal_threadblocks / num_threadblocks;
-
-            // Compute new k-split share
-            int new_split_k = (dim_k + num_partitions - 1) / num_partitions;
-
-            // Round split_k share to the nearest block_task_policy_t::BlockItemsK
-            new_split_k = round_nearest(new_split_k, block_tile_items_k);
-
-            // Recompute k-splitting factor with new_split_k
-            num_partitions = (dim_k + new_split_k - 1) / new_split_k;
-
-            // Update grid dims and k if we meet the minimum number of iterations worth the overhead of splitting
-            int min_iterations_k = 8;
-
-            if (((new_split_k / block_tile_items_k) > min_iterations_k) &&    // We're going to go through at least this many k iterations
-                (sm_count * max_sm_occupancy < NumFlagsSplitK))             // We have enough semaphore flags allocated
-            {
-                grid_dims.z = num_partitions;
-                split_k = new_split_k;
-            }
-        }
-
-        use_semaphore = (grid_dims.z > 1);
-    }
-
-
-    /**
-     * Initializer
-     */
-    cudaError_t prepare(
-        cudaStream_t    stream,             ///< CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
-        bool            debug_synchronous)  ///< Whether or not to synchronize the stream after every kernel launch to check for errors.  Also causes launch configurations to be printed to the console if DEBUG is defined.  Default is \p false.
-
-    {
-        cudaError error = cudaSuccess;
-
-        if (use_semaphore)
-        {
-            int block_threads = 128;
-            int grid_dims = (NumFlagsSplitK + block_threads - 1) / block_threads;
-
-            prepare_kernel<<<grid_dims, block_threads, 0, stream>>>(d_flags);
-
-            // Check for failure to launch
-            if (CUDA_PERROR_DEBUG(error = cudaPeekAtLastError()))
-                return error;
-
-            // Sync the stream if specified to flush runtime errors
-            if (debug_synchronous && (CUDA_PERROR_DEBUG(error = cudaStreamSynchronize(stream))))
-                return error;
-        }
-
-        return error;
-    }
-
-
-    /**
-     * Compute the efficiency of dispatch wave quantization
-     */
-    float get_wave_efficiency(
-        int     sm_count,
-        int     max_sm_occupancy,
-        dim3    block_dims,
-        dim3    grid_dims)
-    {
-        // Heuristic for how many warps are needed to saturate an SM for a given
-        // multiply-accumulate genre.  (NB: We could make this more rigorous by
-        // specializing on data types and SM width)
-        int saturating_warps_per_sm = 16;
-
-        int num_threadblocks                = grid_dims.x * grid_dims.y * grid_dims.z;
-        int threads_per_threadblock         = block_dims.x * block_dims.y;
-        int warps_per_threadblock           = threads_per_threadblock / 32;
-        int saturating_threadblocks_per_sm  = (saturating_warps_per_sm + warps_per_threadblock - 1) / warps_per_threadblock;
-
-        int saturating_residency    = sm_count * saturating_threadblocks_per_sm;
-        int full_waves              = num_threadblocks / saturating_residency;
-        int remainder_threadblocks          = num_threadblocks % saturating_residency;
-        int total_waves             = (remainder_threadblocks == 0) ? full_waves : full_waves + 1;
-
-        float last_wave_saturating_efficiency = float(remainder_threadblocks) / saturating_residency;
-
-        return (float(full_waves) + last_wave_saturating_efficiency) / total_waves;
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/linear_scaling.h

@@ -0,0 +1,169 @@
+
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements the BLAS linear scaling function alpha*AB + beta*C
+*/
+#pragma once
+
+#include "cutlass/fragment_multiply_add.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename T>
+CUTLASS_DEVICE bool is_zero(T x) {
+  return x == T(0);
+}
+
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
+CUTLASS_DEVICE bool is_zero(half x) { return reinterpret_cast<int16_t&>(x) == int16_t(0); }
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Functor to compute linear combination of fragments
+template <typename Scalar_, typename FragmentMultiplyAdd_ = FragmentMultiplyAdd<Scalar_, Scalar_> >
+struct LinearScaling {
+  // The scalar.
+  typedef Scalar_ Scalar;
+  // The accumulator Type
+  typedef typename FragmentMultiplyAdd_::ScalarAccum ScalarAccum;
+  // The adapater.
+  typedef FragmentMultiplyAdd_ FragmentMultiplyAdd;
+
+  /// The parameters.
+  struct Params {
+    /// The alpha/beta scaling params.
+    Scalar alpha, beta;
+
+    //
+    // Methods
+    //
+
+    // Constructor
+    CUTLASS_HOST_DEVICE
+    Params(Scalar _alpha = 0, Scalar _beta = 0) : alpha(_alpha), beta(_beta) {}
+
+    /// Initialize the parameters
+    CUTLASS_HOST_DEVICE int initialize(Scalar _alpha, Scalar _beta) {
+      alpha = _alpha;
+      beta = _beta;
+      return 0;
+    }
+
+    /// Initialize the parameters.
+    template <typename GemmDesc_>
+    CUTLASS_HOST_DEVICE int initialize(GemmDesc_ const& desc) {
+      alpha = desc.alpha;
+      beta = desc.beta;
+      return 0;
+    }
+  };
+
+  //
+  // Data members
+  //
+
+  Params params;
+
+  //
+  // Methods
+  //
+
+  /// Ctor.
+  CUTLASS_DEVICE LinearScaling() { }
+
+  /// Ctor.
+  CUTLASS_DEVICE LinearScaling(Params const& _params) : params(_params) {}
+
+  /// Method to determine whether the source accumulator matrix C is ever needed. This method
+  /// may always safely return true, though better performance is possible if the source accumulator
+  /// matrix is never loaded unnecessarily.
+  CUTLASS_DEVICE
+  bool source_required() const {
+    return !is_zero(params.beta);
+  }
+
+  /// Evaluate the functor.
+  template <typename FragmentA_, typename FragmentB_>
+  CUTLASS_DEVICE void evaluate(FragmentA_ const& accum, FragmentB_& output) {
+    FragmentMultiplyAdd mad;
+    mad.multiply(params.alpha, accum, output);
+
+  }
+
+  /// Evaluate the functor, without using fragment in the API
+  template <typename ScalarAccum, typename ScalarOutput, int size>
+  CUTLASS_DEVICE void evaluate(ScalarAccum const *accum, ScalarOutput *output) {
+    Fragment<ScalarAccum, size> FragAccum;
+    Fragment<ScalarOutput, size> FragOutput;
+#pragma unroll
+    for (int i = 0; i < size; i++) {
+      FragAccum[i] = accum[i];
+      FragOutput[i] = output[i];
+    }
+    evaluate(FragAccum, FragOutput);
+#pragma unroll
+    for (int i = 0; i < size; i++) {
+      output[i] = FragOutput[i];
+    }
+  }
+
+  /// Evaluate the functor.
+  template <typename FragmentA_, typename FragmentB_>
+  CUTLASS_DEVICE void evaluate(FragmentA_ const& accum, FragmentB_ const& old, FragmentB_& output) {
+    FragmentMultiplyAdd mad;
+    FragmentB_ tmp;
+    mad.multiply(params.beta, old, tmp);
+    mad.multiply_add(params.alpha, accum, tmp, output);
+  }
+
+  /// Evaluate the functor, without using fragment in the API
+  template <typename ScalarAccum, typename ScalarOutput, int size>
+  CUTLASS_DEVICE void evaluate(ScalarAccum const *accum, ScalarOutput const *old, ScalarOutput *output) {
+    Fragment<ScalarAccum, size> FragAccum;
+    Fragment<ScalarOutput, size> FragOutput;
+    Fragment<ScalarOutput, size> FragOld;
+#pragma unroll
+    for (int i = 0; i < size; i++) {
+      FragAccum[i] = accum[i];
+      FragOutput[i] = output[i];
+      FragOld[i] = old[i];
+    }
+    evaluate(FragAccum, FragOld, FragOutput);
+#pragma unroll
+    for (int i = 0; i < size; i++) {
+      output[i] = FragOutput[i];
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/linear_scaling_device_ptr.h

@@ -0,0 +1,149 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements the BLAS linear scaling function alpha*AB + beta*C
+*/
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cutlass/gemm/scalar_or_pointer.h"
+#include "cutlass/gemm/linear_scaling.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Functor to compute linear combination of fragments. This is intended to support passing scalars
+/// either by value from the host or by reference to device-side scalar elements. This is inspired
+/// by cuBLAS's device pointer mode.
+template <typename Scalar_, typename FragmentMultiplyAdd_ = FragmentMultiplyAdd<Scalar_, Scalar_> >
+struct LinearScalingDevicePtr : public LinearScaling<Scalar_, FragmentMultiplyAdd_> {
+
+  /// Linear Scaling class used
+  typedef LinearScaling<Scalar_, FragmentMultiplyAdd_> Base;
+
+  // The scalar.
+  typedef typename Base::Scalar Scalar;
+
+  /// The parameters.
+  class Params  {
+  private:
+    /// Alpha scalar
+    detail::ScalarOrPointer<Scalar> alpha_;
+
+    /// Beta sclaar
+    detail::ScalarOrPointer<Scalar> beta_;
+
+  public:
+    //
+    // Methods
+    //
+
+    // Constructor
+    CUTLASS_HOST_DEVICE
+    Params() {}
+
+    // Constructor
+    CUTLASS_HOST_DEVICE
+    Params(
+      Scalar alpha,
+      Scalar beta
+    ):
+      alpha_(alpha),
+      beta_(beta) {}
+
+    // Constructor
+    CUTLASS_HOST_DEVICE
+    Params(
+      Scalar const *alpha_ptr,
+      Scalar const *beta_ptr
+    ):
+      alpha_(alpha_ptr),
+      beta_(alpha_ptr) {}
+
+    /// Initialize the parameters
+    CUTLASS_HOST_DEVICE int initialize(
+      Scalar alpha,
+      Scalar beta) {
+
+      alpha_ = alpha;
+      beta_ = beta;
+
+      return 0;
+    }
+
+    /// Initialize the parameters
+    CUTLASS_HOST_DEVICE int initialize(
+      Scalar const *alpha,
+      Scalar const *beta) {
+
+      alpha_ = alpha;
+      beta_= beta;
+
+      return 0;
+    }
+
+    /// Initialize the parameters.
+    template <typename GemmDesc_>
+    CUTLASS_HOST_DEVICE int initialize(GemmDesc_ const& desc) {
+
+      alpha_ = desc.alpha;
+      beta_ = desc.beta;
+
+      return 0;
+    }
+
+    /// Gets the alpha scalar
+    CUTLASS_HOST_DEVICE
+    Scalar alpha() const {
+      return alpha_;
+    }
+
+    /// Gets the beta scalar
+    CUTLASS_HOST_DEVICE
+    Scalar beta() const {
+      return beta_;
+    }
+  };
+
+  //
+  // Methods
+  //
+
+  /// Ctor.
+  CUTLASS_HOST_DEVICE LinearScalingDevicePtr(Params const& _params) {
+    this->params.alpha = _params.alpha();
+    this->params.beta = _params.beta();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace gemm
+} // namespace cutlass

cutlass/gemm/scalar_or_pointer.h

@@ -0,0 +1,129 @@
+
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements the BLAS linear scaling function alpha*AB + beta*C
+*/
+#pragma once
+
+#include "cutlass/cutlass.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+namespace detail {
+
+/// Helper class defines an object  which operates as either a scalar or a pointer. If the pointer
+/// is non-null, it is dereferenced when the object is accessed.
+template <typename Scalar_>
+class ScalarOrPointer {
+public:
+  /// Underlying scalar type
+  typedef Scalar_ Scalar;
+
+private:
+  //
+  // Data members
+  //
+
+  /// Scalar value
+  Scalar scalar;
+
+  /// Pointer to use if non null
+  Scalar const *ptr;
+
+public:
+
+  //
+  // Methods
+  //
+
+  /// Default ctor
+  CUTLASS_HOST_DEVICE
+  ScalarOrPointer(): scalar(0), ptr(nullptr) {}
+
+  /// Object behaves as a scalar
+  CUTLASS_HOST_DEVICE
+  ScalarOrPointer(Scalar const &val): scalar(val), ptr(nullptr) {}
+
+  /// Object behaves as a scalar
+  CUTLASS_HOST_DEVICE
+  ScalarOrPointer(Scalar const *ptr_): scalar(0), ptr(ptr_) {}
+
+  /// Returns true if is pointer
+  CUTLASS_HOST_DEVICE
+  bool is_pointer() const {
+    return bool(ptr);
+  }
+
+  /// Gets the pointer value
+  CUTLASS_HOST_DEVICE
+  Scalar const *get_ptr() const {
+    return ptr;
+  }
+
+  /// Gets the pointer value
+  CUTLASS_HOST_DEVICE
+  Scalar get_scalar() const {
+    return scalar;
+  }
+
+  /// Assigns to a scalar and sets pointer to nullptr
+  CUTLASS_HOST_DEVICE
+  ScalarOrPointer &operator=(Scalar const &scalar_) {
+    scalar = scalar_;
+    ptr = nullptr;
+    return *this;
+  }
+
+  /// Assigns to a pointer value
+  CUTLASS_HOST_DEVICE
+  ScalarOrPointer &operator=(Scalar const *ptr_) {
+    ptr = ptr_;
+    return *this;
+  }
+
+  /// Access the element
+  CUTLASS_HOST_DEVICE
+  Scalar get() const {
+    if (ptr) {
+      return *ptr;
+    }
+    return scalar;
+  }
+
+  /// Accesses the element
+  CUTLASS_HOST_DEVICE
+  operator Scalar() const {
+    return get();
+  }
+};
+
+} // namespace detail
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/gemm/sgemm_traits.h

@@ -0,0 +1,172 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defies structural properties of single-precision GEMM.
+*/
+#pragma once
+
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/gemm/gemm_epilogue.h"
+#include "cutlass/gemm/gemm_epilogue_traits.h"
+#include "cutlass/gemm/gemm_global_tile.h"
+#include "cutlass/gemm/gemm_shared_tile.h"
+#include "cutlass/gemm/gemm_traits.h"
+#include "cutlass/gemm/thread_multiply_add.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The tile size for the GEMM KxNxM.
+    typename OutputTile_,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_,
+    /// The number of scalars per LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of scalars per LDG for B.
+    int kScalarsPerLdgB_ = 1,
+    /// Whether to specify launch bounds
+    bool kLaunchBounds = true>
+struct SgemmConfig : public GemmConfig<
+                         /// The scalar type for A.
+                         float,
+                         /// The scalar type for B.
+                         float,
+                         /// The scalar type for C.
+                         float,
+                         /// The scalar type for D.
+                         float,
+                         /// The tile size for the GEMM KxNxM.
+                         OutputTile_,
+                         /// The functor to do the math in the main loop.
+                         ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, float, float, float>,
+                         /// The number of scalars per LDG for A.
+                         kScalarsPerLdgA_,
+                         /// The number of scalars per STS for A.
+                         kScalarsPerLdgA_,
+                         /// The number of scalars per LDS for A.
+                         4,
+                         /// The number of scalars per LDG for B.
+                         kScalarsPerLdgB_,
+                         /// The number of scalars per STS for B.
+                         kScalarsPerLdgB_,
+                         /// The number of scalars per LDS for B.
+                         4,
+                         /// The number of scalars per LDG for C and STG for D.
+                         1,
+                         /// The number of scalars per STS for D.
+                         4,
+                         /// The number of scalars per LDS for D.
+                         1,
+                         /// The number of stages in shared memory.
+                         2,
+                         /// kResidueSeparate
+                         false,
+                         /// kResidueInPrologue
+                         true,
+                         /// kLaunchBounds
+                         kLaunchBounds> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_ = Shape<8, 128, 128>,
+    /// The functor to use in the epilogue.
+    typename EpilogueFunctor_ = LinearScaling<float>,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<8, 8, 8>,
+    /// The number of floats loaded in one LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of floats loaded in one LDG for B.
+    int kScalarsPerLdgB_ = 1,
+    /// The index.
+    typename Index_ = int,
+    /// The SGEMM config.
+    typename GemmConfig_ =
+        SgemmConfig<OutputTile_, ThreadGemmShape_, kScalarsPerLdgA_, kScalarsPerLdgB_, false>,
+    /// The traits class for the epilogue.
+    typename GemmEpilogueTraits_ =
+        SimplifiedGemmEpilogueTraits<GemmConfig_, EpilogueFunctor_, Index_> >
+struct SgemmTraits : public SimplifiedGemmTraits<
+                         // The layout for A.
+                         kLayoutA_,
+                         // The layout for B.
+                         kLayoutB_,
+                         // The config.
+                         GemmConfig_,
+                         // The epilogue.
+                         GemmEpilogue<GemmEpilogueTraits_>,
+                         // The index.
+                         Index_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Helper to define SGEMM traits using Launch Bounds
+template <
+    /// The layout for A.
+    MatrixLayout::Kind kLayoutA_,
+    /// The layout for B.
+    MatrixLayout::Kind kLayoutB_,
+    /// The output tile.
+    typename OutputTile_ = Shape<8, 128, 128>,
+    /// The functor to use in the epilogue.
+    typename EpilogueFunctor_ = LinearScaling<float>,
+    /// Tile size for thread-level GEMM (K-by-N-by-M)
+    typename ThreadGemmShape_ = Shape<8, 8, 8>,
+    /// The number of floats loaded in one LDG for A.
+    int kScalarsPerLdgA_ = 1,
+    /// The number of floats loaded in one LDG for B.
+    int kScalarsPerLdgB_ = 1,
+    /// The index.
+    typename Index_ = int,
+    /// The SGEMM config.
+    typename GemmConfig_ =
+        SgemmConfig<OutputTile_, ThreadGemmShape_, kScalarsPerLdgA_, kScalarsPerLdgB_, true>,
+    /// The traits class for the epilogue.
+    typename GemmEpilogueTraits_ =
+        SimplifiedGemmEpilogueTraits<GemmConfig_, EpilogueFunctor_, Index_> >
+struct SgemmLBTraits : public SimplifiedGemmTraits<
+                         // The layout for A.
+                         kLayoutA_,
+                         // The layout for B.
+                         kLayoutB_,
+                         // The config.
+                         GemmConfig_,
+                         // The epilogue.
+                         GemmEpilogue<GemmEpilogueTraits_>,
+                         // The index.
+                         Index_> {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/thread_accumulator.h

@@ -1,469 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Thread-level multiply-accumulate abstraction
- */
-
-#include "../util/util.h"
-#include "dp_accummulate.h"
-
-
-namespace cutlass {
-namespace gemm {
-
-
-/******************************************************************************
- * thread_accumulator (generic specialization)
- ******************************************************************************/
-
-/**
- * \brief Thread-level multiply-accumulate abstraction (generic specialization)
- *
- * The thread_accumulator class maintains a MxN tile of accumulators in
- * registers to which MxNxK matrix products of two thread tiles A (MxK)
- * and B (KxN) can be added, where:
- *   M = ThreadItemsY
- *   N = ThreadItemsX
- *   K = sizeof(dp_vector_t) / sizeof(value_t).
- *
- * In order to leverage architecture-specific "dot-product accumulate" ISA
- * operations, K is dictated by the thread_accumulator class in the form of
- * the member-type dp_vector_t, which defines a K-component vector of value_t.
- * The multiplicand inputs A and B are provided as arrays of dp_vector_t having
- * extents ThreadItemsY and ThreadItemsX, respectively. (In the single
- * component "dp1" scenario where dp_vector_t == value_t and thus K == 1, the
- * multiplication is simply the outer product of two vectors.)
- *
- * The accumulators are zero-initialized in a two-phase process (construction +
- * initialization) that requires shared storage in the form of the member-type
- * scratch_storage_t during construction. (A single scratch_storage_t instance
- * can be uniformly referenced across all threads in the block during
- * construction *if* the block is synchronized between construction and
- * initialization.)
- *
- * NB: This generic class is not directly constructible.  Architecture- and
- * algorithm-specific template specializations will provide the API
- * functionality prescribed here.
- */
-template <
-    int ThreadItemsY,                   ///< Height of thread tile in accum_t
-    int ThreadItemsX,                   ///< Width of thread tile in accum_t
-    typename value_t,                   ///< Multiplicand value type
-    typename accum_t,                   ///< Accumulator value type
-    int ACCUM_BYTES =                   ///< Size in bytes of accum_t
-        sizeof(accum_t),
-    arch_family_t::kind_t ArchFamily =  ///< Architectural family enumerant
-        CUTLASS_ARCH_FAMILY>
-struct thread_accumulator
-{
-protected:
-
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    /// Specialized dot-product traits type
-    typedef dp_accummulate<value_t, accum_t> dp_accum_traits_t;
-
-
-public:
-
-    //-------------------------------------------------------------------------
-    // Member types
-    //-------------------------------------------------------------------------
-
-    /// Dot-product vector type
-    typedef typename dp_accum_traits_t::dp_vector_t dp_vector_t;
-
-    /// Scratch storage layout
-    struct scratch_storage_t {};
-
-
-protected:
-
-    //-------------------------------------------------------------------------
-    // Data members
-    //-------------------------------------------------------------------------
-
-    /// Thread's tile of accumulators
-    accum_t accumulators[ThreadItemsY][ThreadItemsX];
-
-
-    //-------------------------------------------------------------------------
-    // Utility methods
-    //-------------------------------------------------------------------------
-
-    /**
-     * Compute a multiply-add at accumulator coordinates (x, y)
-     */
-    inline __device__
-    void mad_xy(
-        dp_vector_t (&tile_a)[ThreadItemsY],
-        dp_vector_t (&tile_b)[ThreadItemsX],
-        int x,
-        int y)
-    {
-        dp_accum_traits_t::mad(
-            accumulators[y][x],
-            tile_a[y],
-            tile_b[x],
-            accumulators[y][x]);
-    }
-
-public:
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    thread_accumulator(
-        scratch_storage_t &scratch)
-    {}
-
-
-    //-------------------------------------------------------------------------
-    // Accumulator API
-    //-------------------------------------------------------------------------
-
-    /**
-     * \brief Zero-initialize thread accumulators.
-     *
-     * If a common reference to a single block-wide shared instance of scratch_storage_t
-     * is used during construction, the block must be synchronized after construction
-     * but prior to the invocation of init().
-     */
-    inline __device__
-    void init()
-    {
-        #pragma unroll
-        for (int y = 0; y < ThreadItemsY; ++y) {
-            #pragma unroll
-            for (int x = 0; x < ThreadItemsX; ++x)
-            {
-                accumulators[y][x] = accum_t(0);
-            }
-        }
-    }
-
-
-    /**
-     * Retrieve the accumulator at thread tile coordinates (x, y)
-     */
-    inline __device__
-    accum_t get(int x, int y)
-    {
-        // Accumulators are row-major
-        return accumulators[y][x];
-    }
-
-
-    /**
-     * \brief Compute the product of tile_a and tile_b and add the result to
-     * the tile of accumulators.
-     */
-    inline __device__
-    void multiply_accumulate(
-        dp_vector_t (&tile_a)[ThreadItemsY],
-        dp_vector_t (&tile_b)[ThreadItemsX])
-    {
-        // Simply traverse the accumulator tile in row-major order
-        #pragma unroll
-        for (int y = 0; y < ThreadItemsY; ++y)
-        {
-            #pragma unroll
-            for (int x = 0; x < ThreadItemsX; ++x)
-            {
-                mad_xy(tile_a, tile_b, x, y);
-            }
-        }
-    }
-};
-
-
-
-
-/******************************************************************************
- * thread_accumulator (__half->__half specialization)
- ******************************************************************************/
-
-/**
- * \brief Thread-level multiply-accumulate abstraction (__half->__half specialization)
- *
- * NB: Because we use the 2-item SIMD instruction HFMA2:
- *   - ThreadItemsX must be an even multiple of 2
- *   - ThreadItemsY must be an even multiple of 2
- *
- */
-template <
-    int ThreadItemsY,                   ///< Height in rows of thread tile in C
-    int ThreadItemsX,                   ///< Width in columns of thread tile in C
-    arch_family_t::kind_t ArchFamily>   ///< Architectural family enumerant
-struct thread_accumulator<
-    ThreadItemsY,
-    ThreadItemsX,
-    __half,                             ///< Multiplicand value type (matrices A and B)
-    __half,                             ///< Accumulator value type (matrix C and scalars)
-    2,                                  ///< Size in bytes of accum_t
-    ArchFamily>
-{
-protected:
-
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    /// Constants
-    enum
-    {
-        /// Height of thread tile in column-major uint32_t SIMD pairs along Y dimension
-        ThreadTilePairsY = divide_assert<ThreadItemsY, 2>::value,
-
-        /// Width of thread tile in column-major uint32_t SIMD pairs along X dimension
-        ThreadTilePairsX = ThreadItemsX,
-
-        /// Number of SIMD pairs in thread's slice of block-wide tile multiplicand A
-        ThreadPairsA = divide_assert<ThreadItemsY, 2>::value,
-
-        /// Number of SIMD pairs in thread's slice of block-wide tile multiplicand B
-        ThreadPairsB = divide_assert<ThreadItemsX, 2>::value,
-    };
-
-public:
-
-    //-------------------------------------------------------------------------
-    // Member types
-    //-------------------------------------------------------------------------
-
-    /// Dot-product vector type
-    typedef __half dp_vector_t;
-
-    /// Scratch storage layout
-    struct scratch_storage_t {};
-
-
-private:
-
-    //-------------------------------------------------------------------------
-    // Members
-    //-------------------------------------------------------------------------
-
-    /// Thread's tile of C accumulator pairs (the uint32_t SIMD pairs are
-    /// column-major, the 2D tile layout is also column-major)
-    uint32_t accumulator_pairs[ThreadTilePairsX][ThreadTilePairsY];
-
-
-    //-------------------------------------------------------------------------
-    // Utility methods
-    //-------------------------------------------------------------------------
-
-    /**
-     * Compute an HFMA2 MAD
-     */
-    inline __device__ void mad(
-        uint32_t &d,
-        const uint32_t &a,
-        const uint32_t &b,
-        const uint32_t &c)
-    {
-
-        asm volatile ("fma.rn.f16x2 %0, %1, %2, %3;\n"
-            : "=r"(d) : "r"(a), "r"(b), "r"(c));
-    }
-
-
-    /**
-     * Compute an HFMA2 MAD with replicated b.lo:
-     *   d{hi} = a{hi} * b{lo} + c{hi};
-     *   d{lo} = a{lo} * b{lo} + c{lo};
-     */
-    inline __device__ void mad_replicate_low(
-        uint32_t &d,
-        const uint32_t &a,
-        const uint32_t &b,
-        const uint32_t &c)
-    {
-        // Replicate low halves of b
-        uint32_t replicate;
-        asm volatile (
-            "{"
-            "  .reg .b16 b_low,b_high;\n"
-            "  mov.b32 {b_low,b_high}, %1;\n"
-            "  mov.b32 %0, {b_low,b_low};\n"
-            "}" : "=r"(replicate) : "r"(b));
-
-        mad(d, a, replicate, c);
-    }
-
-
-    /**
-     * Compute an HFMA2 MAD with replicated b.hi:
-     *   d{hi} = a{hi} * b{hi} + c{hi};
-     *   d{lo} = a{lo} * b{hi} + c{lo};
-     */
-    inline __device__ void mad_replicate_high(
-        uint32_t &d,
-        const uint32_t &a,
-        const uint32_t &b,
-        const uint32_t &c)
-    {
-        // Replicate high halves of b
-        uint32_t replicate;
-        asm volatile (
-            "{"
-            "  .reg .b16 b_low,b_high;\n"
-            "  mov.b32 {b_low,b_high}, %1;\n"
-            "  mov.b32 %0, {b_high,b_high};\n"
-            "}" : "=r"(replicate) : "r"(b));
-
-        mad(d, a, replicate, c);
-    }
-
-
-    /**
-     * Compute a multiply-add at accumulator SIMD-pair coordinates (pair_x, pair_y)
-     */
-    inline __device__
-    void mad_xy_even(
-        uint32_t (&pairs_tile_a)[ThreadPairsA],
-        uint32_t (&pairs_tile_b)[ThreadPairsB],
-        int pair_x,
-        int pair_y)
-    {
-        // Even column: use low half of the b pair
-        mad_replicate_low(
-            accumulator_pairs[pair_x][pair_y],
-            pairs_tile_a[pair_y],
-            pairs_tile_b[pair_x / 2],
-            accumulator_pairs[pair_x][pair_y]);
-    }
-
-
-    /**
-     * Compute a multiply-add at accumulator SIMD-pair coordinates (pair_x, pair_y)
-     */
-    inline __device__
-    void mad_xy_odd(
-        uint32_t (&pairs_tile_a)[ThreadPairsA],
-        uint32_t (&pairs_tile_b)[ThreadPairsB],
-        int pair_x,
-        int pair_y)
-    {
-        // Odd column: use high half of the b pair
-        mad_replicate_high(
-            accumulator_pairs[pair_x][pair_y],
-            pairs_tile_a[pair_y],
-            pairs_tile_b[pair_x / 2],
-            accumulator_pairs[pair_x][pair_y]);
-    }
-
-
-public:
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor
-    inline __device__
-    thread_accumulator(
-        scratch_storage_t &scratch)
-    {}
-
-
-    //-------------------------------------------------------------------------
-    // Accumulator API
-    //-------------------------------------------------------------------------
-
-    /**
-     * Zero-initialize thread accumulators.
-     */
-    inline __device__
-    void init()
-    {
-        #pragma unroll
-        for (int y = 0; y < ThreadTilePairsY; ++y)
-        {
-            #pragma unroll
-            for (int x = 0; x < ThreadTilePairsX; ++x)
-            {
-                accumulator_pairs[x][y] = 0;
-            }
-        }
-    }
-
-
-    /**
-     * Retrieve the accumulator at thread tile coordinates (x, y)
-     */
-    inline __device__
-    __half get(int x, int y)
-    {
-        // SIMD pairs are column-major
-        uint32_t pair = accumulator_pairs[x][y / 2];
-
-        return reinterpret_cast<__half (&)[2]>(pair)[y % 2];
-    }
-
-
-    /**
-     * \brief Compute the product of pairs_tile_a and pairs_tile_b and add the result to
-     * the tile of accumulators.
-     */
-    inline __device__
-    void multiply_accumulate(
-        dp_vector_t (&tile_a)[ThreadItemsY],
-        dp_vector_t (&tile_b)[ThreadItemsX])
-    {
-        typedef uint32_t pairs_tile_a_t[ThreadPairsA];
-        typedef uint32_t pairs_tile_b_t[ThreadPairsB];
-
-        // Alias slices in pairs
-        pairs_tile_a_t &pairs_tile_a = reinterpret_cast<pairs_tile_a_t&>(tile_a);
-        pairs_tile_b_t &pairs_tile_b = reinterpret_cast<pairs_tile_b_t&>(tile_b);
-
-        // Simply traverse the accumulator tile in column-major order
-        #pragma unroll
-        for (int x = 0; x < ThreadTilePairsX; ++x)
-        {
-            #pragma unroll
-            for (int y = 0; y < ThreadTilePairsY; ++y)
-            {
-                mad_xy_even(pairs_tile_a, pairs_tile_b, x, y);
-            }
-        }
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/thread_multiply_add.h

@@ -0,0 +1,96 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Template implementing matrix multiply-add operations on fragments.
+*/
+#pragma once
+
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Template performing matrix multiply-add operation within a thread
+template <typename ThreadGemmShape_,
+          typename ThreadsPerWarp_,
+          typename ScalarA_,
+          typename ScalarB_,
+          typename ScalarC_,
+          MatrixLayout::Kind kLayout_ = MatrixLayout::kColumnMajor>
+struct ThreadMultiplyAdd {
+  /// The shape of the instruction.
+  typedef Shape<1, 1, 1, 1> InstructionShape;
+  /// The shape of a thread-leveel matrix multiply accumulate.
+  typedef ThreadGemmShape_ ThreadGemmShape;
+  /// Aliased to "AccumulatorsPerThread" for compatibility. Expect to be renamed in CUTLASS v2.0
+  typedef ThreadGemmShape AccumulatorsPerThread;
+  /// The number of threads per warp.
+  typedef ThreadsPerWarp_ ThreadsPerWarp;
+  /// The number of accumulators per warp.
+  typedef typename ShapeMul<ThreadGemmShape, ThreadsPerWarp>::Shape AccumulatorsPerWarp;
+  /// The type for A.
+  typedef ScalarA_ ScalarA;
+  /// The fragment for A.
+  typedef Fragment<ScalarA, AccumulatorsPerThread::kW> FragmentA;
+  /// The type for B.
+  typedef ScalarB_ ScalarB;
+  /// The fragment for B.
+  typedef Fragment<ScalarB, AccumulatorsPerThread::kH> FragmentB;
+  /// The type for C and D.
+  typedef ScalarC_ ScalarC;
+  /// The accumulators.
+  typedef Fragment<ScalarC, AccumulatorsPerThread::kH * AccumulatorsPerThread::kW, 16> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE ThreadMultiplyAdd() {}
+
+  /// Multiply : d = a*b + c.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    if(kLayout_ == MatrixLayout::kColumnMajor) {
+      for (int j = 0; j < AccumulatorsPerThread::kH; ++j) {
+        for (int i = 0; i < AccumulatorsPerThread::kW; ++i) {
+          d[j * AccumulatorsPerThread::kW + i] = a[i] * b[j] + c[j * AccumulatorsPerThread::kW + i];
+        }
+      }
+    }
+    else {
+      for(int i = 0; i < AccumulatorsPerThread::kW; ++i) {
+        for(int j = 0; j < AccumulatorsPerThread::kH; ++j) {
+          d[i * AccumulatorsPerThread::kH + j] = a[i] * b[j] + c[i * AccumulatorsPerThread::kH + j];
+        }
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/threadblock_swizzle.h

@@ -0,0 +1,387 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defies functors for mapping blockIdx to partitions of the GEMM computation.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/gemm/gemm_coord.h"
+
+namespace cutlass {
+namespace gemm {
+
+struct swizzleDirection {
+  enum Kind { Boustrophedon, OneDirection };
+};
+// helper template function
+template <enum swizzleDirection::Kind>
+CUTLASS_DEVICE int getLinearIdx(int groups) {
+  // groupCols is not needed for OneDirection Swizzle
+  return blockIdx.y * gridDim.x + blockIdx.x;
+}
+template <>
+CUTLASS_DEVICE int getLinearIdx<swizzleDirection::Boustrophedon>(int groups) {
+  // reverse blockIdx.x for some columns
+  if ((blockIdx.y / groups) % 2 == 1)
+    return blockIdx.y * gridDim.x + (gridDim.x - blockIdx.x - 1);
+  else
+    return blockIdx.y * gridDim.x + blockIdx.x;
+}
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup IdentityBlockSwizzle Identity Block Swizzle
+@{
+    Block Swizzle provides the mapping logic between a block in the physical memory of Matrix C and
+Thread Block
+    Identiy Block Swizzle effective maps blocks in leading dimension order (column major) with
+thread block
+    in leading dimension order (blockIdx.x)
+    blockIdx.z is mapped with batch_count for batched GEMM
+@}
+*/
+struct IdentityBlockSwizzle {
+  /// Ctor. aka ColumnMajorBlockSwizzle<1>
+  CUTLASS_HOST_DEVICE IdentityBlockSwizzle() {}
+
+  /// Swizzle the block index.
+  CUTLASS_DEVICE dim3 swizzle() { return blockIdx; }
+
+  ///
+  CUTLASS_HOST_DEVICE dim3 get_grid_layout(GemmCoord const &problem_size,
+                                           Coord<3> const &OutputTile) {
+    /*OutputTile and problem_size are both in KNM order*/
+    dim3 grid;
+    grid.x = (problem_size.m() + OutputTile[2] - 1) / OutputTile[2];
+    grid.y = (problem_size.n() + OutputTile[1] - 1) / OutputTile[1];
+    grid.z = problem_size.batch();
+    return grid;
+  }
+
+  ///
+  CUTLASS_DEVICE Coord<3> get_threadblock_offset(Coord<3> const &OutputTile) {
+    dim3 block = swizzle();
+    Coord<3> threadblock_offset =
+        make_Coord(0, block.y * OutputTile[1], block.x * OutputTile[2]);
+    return threadblock_offset;
+  }
+
+  ///
+  CUTLASS_DEVICE int get_batch_id() {
+    dim3 block = swizzle();
+    return block.z;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*
+ColumnMajorBlockSwizzle<1, OneDirection> is equivalent with IdentityBlockSwizzle
+groupCols has the effect of controlling the schedulling of thread blocks
+settings with different groupCols can contribute to the overall performance by affecting L2 cache
+hit rate
+
+consider a regular thread block mapping btween matrix C and different thread blocks
+note that C is column major, and the leading dimension of thread block id is blockIdx.x
+
+let's look at an example where gridIdx.x = 6, gridIdx.y = 7, gridIdx.z = 1
+(blockIdx.x, blockIdx.y)
+mapping between threadblockID and C matrix:
+-------------------------------------------------------
+(0,0) | (0,1) | (0,2) | (0,3) | (0,4) | (0,5) | (0,6) |
+-------------------------------------------------------
+(1,0) | (1,1) | (1,2) | (1,3) | (1,4) | (1,5) | (1,6) |
+-------------------------------------------------------
+(2,0) | (2,1) | (2,2) | (2,3) | (2,4) | (2,5) | (2,6) |
+-------------------------------------------------------
+(3,0) | (3,1) | (3,2) | (3,3) | (3,4) | (3,5) | (3,6) |
+-------------------------------------------------------
+(4,0) | (4,1) | (4,2) | (4,3) | (4,4) | (4,5) | (4,6) |
+-------------------------------------------------------
+(5,0) | (5,1) | (5,2) | (5,3) | (5,4) | (5,5) | (5,6) |
+-------------------------------------------------------
+
+A ColumnMajorBlockSwizzle<1, OneDirection> will imply the above order where threadblocks are
+launched in a column major
+
+A ColumnMajorBlockSwizzle<2, OneDirection> swizzles things a little,
+-------------------------------------------------------
+(0,0) | (3,0) | (0,2) | (3,2) | (0,4) | (3,4) | (0,6) |
+-------------------------------------------------------
+(0,1) | (3,1) | (0,3) | (3,3) | (0,5) | (3,5) | (1,6) |
+-------------------------------------------------------
+(1,0) | (4,0) | (1,2) | (4,2) | (1,4) | (4,4) | (2,6) |
+-------------------------------------------------------
+(1,1) | (4,1) | (1,3) | (4,3) | (1,5) | (4,5) | (3,6) |
+-------------------------------------------------------
+(2,0) | (5,0) | (2,2) | (5,2) | (2,4) | (5,4) | (4,6) |
+-------------------------------------------------------
+(2,1) | (5,1) | (2,3) | (5,3) | (2,5) | (5,5) | (5,6) |
+-------------------------------------------------------
+
+so in memory, it would apprear that we work on 2 columns at a time rather than 1
+Note that the index here really represent how each block maps to memory
+
+A ColumnMajorBlockSwizzle<1, Boustrophedon> is similar to ColumnMajorBlockSwizzle<1, OneDirection>
+except that every column flips the ordering against the previous one
+-------------------------------------------------------
+(0,0) | (5,1) | (0,2) | (5,3) | (0,4) | (5,5) | (0,6) |
+-------------------------------------------------------
+(1,0) | (4,1) | (1,2) | (4,3) | (1,4) | (4,5) | (1,6) |
+-------------------------------------------------------
+(2,0) | (3,1) | (2,2) | (3,3) | (2,4) | (3,5) | (2,6) |
+-------------------------------------------------------
+(3,0) | (2,1) | (3,2) | (2,3) | (3,4) | (2,5) | (3,6) |
+-------------------------------------------------------
+(4,0) | (1,1) | (4,2) | (1,3) | (4,4) | (1,5) | (4,6) |
+-------------------------------------------------------
+(5,0) | (0,1) | (5,2) | (0,3) | (5,4) | (0,5) | (5,6) |
+-------------------------------------------------------
+
+similarily, A ColumnMajorBlockSwizzle<2, Boustrophedon> looks like
+-------------------------------------------------------
+(0,0) | (3,0) | (2,3) | (5,3) | (0,4) | (3,4) | (5,6) |
+-------------------------------------------------------
+(0,1) | (3,1) | (2,2) | (5,2) | (0,5) | (3,5) | (4,6) |
+-------------------------------------------------------
+(1,0) | (4,0) | (1,3) | (4,3) | (1,4) | (4,4) | (3,6) |
+-------------------------------------------------------
+(1,1) | (4,1) | (1,2) | (4,2) | (1,5) | (4,5) | (2,6) |
+-------------------------------------------------------
+(2,0) | (5,0) | (0,3) | (3,3) | (2,4) | (5,4) | (1,6) |
+-------------------------------------------------------
+(2,1) | (5,1) | (0,2) | (3,2) | (2,5) | (5,5) | (0,6) |
+-------------------------------------------------------
+
+*/
+
+template <int groupCols, enum swizzleDirection::Kind swDirection>
+struct ColumnMajorBlockSwizzle {
+  /// Ctor.
+  CUTLASS_HOST_DEVICE ColumnMajorBlockSwizzle() {}
+
+  /// Swizzle the block index.
+  CUTLASS_DEVICE dim3 swizzle() {
+    assert(gridDim.z == 1);
+    int linearIdx = getLinearIdx<swDirection>(groupCols);
+    dim3 swizzledBlockIdx;
+    int currGroupCols = groupCols;
+    int prevGroupCols = groupCols;
+
+    if ((gridDim.y % groupCols != 0) && ((blockIdx.y + (gridDim.y % groupCols)) >= gridDim.y)) {
+      // last colmuns if gridDim.y is not divisble by groupCols
+      currGroupCols = gridDim.y % groupCols;
+    }
+
+    swizzledBlockIdx.x = (linearIdx / currGroupCols) % gridDim.x;
+    swizzledBlockIdx.y =
+        linearIdx % currGroupCols + prevGroupCols * (linearIdx / (prevGroupCols * gridDim.x));
+    swizzledBlockIdx.z = blockIdx.z;
+
+    return swizzledBlockIdx;
+  }
+
+  ///
+  CUTLASS_HOST_DEVICE dim3 get_grid_layout(GemmCoord const &problem_size,
+                                           Coord<3> const &OutputTile) {
+    dim3 grid;
+    grid.x = (problem_size.m() + OutputTile[2] - 1) / OutputTile[2];
+    grid.y = (problem_size.n() + OutputTile[1] - 1) / OutputTile[1];
+    grid.z = problem_size.batch();
+    return grid;
+  }
+
+  ///
+  CUTLASS_DEVICE Coord<3> get_threadblock_offset(Coord<3> const &OutputTile) {
+    dim3 block = swizzle();
+    Coord<3> threadblock_offset =
+        make_Coord(0, block.y * OutputTile[1], block.x * OutputTile[2]);
+    return threadblock_offset;
+  }
+
+  ///
+  CUTLASS_DEVICE int get_batch_id() {
+    dim3 block = swizzle();
+    return block.z;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*
+
+consider a regular thread block mapping btween matrix C and different thread blocks
+note that C is column major, and the leading dimension of thread block id is blockIdx.x
+
+let's look at an example where gridIdx.x = 6, gridIdx.y = 7, gridIdx.z = 1
+(blockIdx.x, blockIdx.y)
+mapping between threadblockID and C matrix:
+-------------------------------------------------------
+(0,0) | (0,1) | (0,2) | (0,3) | (0,4) | (0,5) | (0,6) |
+-------------------------------------------------------
+(1,0) | (1,1) | (1,2) | (1,3) | (1,4) | (1,5) | (1,6) |
+-------------------------------------------------------
+(2,0) | (2,1) | (2,2) | (2,3) | (2,4) | (2,5) | (2,6) |
+-------------------------------------------------------
+(3,0) | (3,1) | (3,2) | (3,3) | (3,4) | (3,5) | (3,6) |
+-------------------------------------------------------
+(4,0) | (4,1) | (4,2) | (4,3) | (4,4) | (4,5) | (4,6) |
+-------------------------------------------------------
+(5,0) | (5,1) | (5,2) | (5,3) | (5,4) | (5,5) | (5,6) |
+-------------------------------------------------------
+
+A RowMajorBlockSwizzle<1, OneDirection> will effectively transpose the map
+
+-----------------------------------------------
+(0,0) | (1,0) | (2,0) | (3,0) | (4,0) | (5,0) |
+-----------------------------------------------
+(0,1) | (1,1) | (2,1) | (3,1) | (4,1) | (5,1) |
+-----------------------------------------------
+(0,2) | (1,2) | (2,2) | (3,2) | (4,2) | (5,2) |
+-----------------------------------------------
+(0,3) | (1,3) | (2,3) | (3,3) | (4,3) | (5,3) |
+-----------------------------------------------
+(0,4) | (1,4) | (2,4) | (3,4) | (4,4) | (5,4) |
+---------------------------------------------
+(0,5) | (1,5) | (2,5) | (3,5) | (4,5) | (5,5) |
+-----------------------------------------------
+(0,6) | (1,6) | (2,6) | (3,6) | (4,6) | (5,6) |
+-----------------------------------------------
+
+It would aprear in memory we are working on 1 row at a time
+
+A ColumnMajorBlockSwizzle<2, OneDirection> swizzles things a little bit more
+-----------------------------------------------
+(0,0) | (1,3) | (2,0) | (3,3) | (4,0) | (5,3) |
+-----------------------------------------------
+(1,0) | (0,4) | (3,0) | (2,4) | (5,0) | (4,4) |
+-----------------------------------------------
+(0,1) | (1,4) | (2,1) | (3,4) | (4,1) | (5,4) |
+-----------------------------------------------
+(1,1) | (0,5) | (3,1) | (2,5) | (5,1) | (4,5) |
+-----------------------------------------------
+(0,2) | (1,5) | (2,2) | (3,5) | (4,2) | (5,5) |
+---------------------------------------------
+(1,2) | (0,6) | (3,2) | (2,6) | (5,2) | (4,6) |
+-----------------------------------------------
+(0,3) | (1,6) | (2,3) | (3,6) | (4,3) | (5,6) |
+-----------------------------------------------
+
+so in memory, it would apprear that we work on 2 rows at a time rather than 1 row
+Note that the index here really represent how each block maps to memory
+
+A RowMajorBlockSwizzle<1, Boustrophedon> is similar to RowMajorBlockSwizzle<1, OneDirection>
+except that every column flips the ordering against the previous one
+
+-----------------------------------------------
+(0,0) | (1,6) | (2,0) | (3,6) | (4,0) | (5,6) |
+-----------------------------------------------
+(0,1) | (1,5) | (2,1) | (3,5) | (4,1) | (5,5) |
+-----------------------------------------------
+(0,2) | (1,4) | (2,2) | (3,4) | (4,2) | (5,4) |
+-----------------------------------------------
+(0,3) | (1,3) | (2,3) | (3,3) | (4,3) | (5,3) |
+-----------------------------------------------
+(0,4) | (1,2) | (2,4) | (3,2) | (4,4) | (5,2) |
+---------------------------------------------
+(0,5) | (1,1) | (2,5) | (3,1) | (4,5) | (5,1) |
+-----------------------------------------------
+(0,6) | (1,0) | (2,6) | (3,0) | (4,6) | (5,0) |
+-----------------------------------------------
+
+similarily, A RowMajorBlockSwizzle<2, Boustrophedon> looks like
+-----------------------------------------------
+(0,0) | (1,3) | (2,3) | (3,6) | (4,0) | (5,3) |
+-----------------------------------------------
+(1,0) | (0,4) | (3,2) | (2,6) | (5,0) | (4,4) |
+-----------------------------------------------
+(0,1) | (1,4) | (2,2) | (3,5) | (4,1) | (5,4) |
+-----------------------------------------------
+(1,1) | (0,5) | (3,1) | (2,5) | (5,1) | (4,5) |
+-----------------------------------------------
+(0,2) | (1,5) | (2,1) | (3,4) | (4,2) | (5,5) |
+---------------------------------------------
+(1,2) | (0,6) | (3,0) | (2,4) | (5,2) | (4,6) |
+-----------------------------------------------
+(0,3) | (1,6) | (2,0) | (3,3) | (4,3) | (5,6) |
+-----------------------------------------------
+
+*/
+
+template <int groupRows, enum swizzleDirection::Kind swDirection>
+struct RowMajorBlockSwizzle {
+  /// Ctor.
+  CUTLASS_HOST_DEVICE RowMajorBlockSwizzle() {}
+
+  /// Swizzle the block index.
+  CUTLASS_DEVICE dim3 swizzle() {
+    assert(gridDim.z == 1);
+    int linearIdx = getLinearIdx<swDirection>(groupRows);
+    dim3 swizzledBlockIdx;
+    int currGroupRows = groupRows;
+    int prevGroupRows = groupRows;
+
+    if ((gridDim.y % groupRows != 0) && ((blockIdx.y + (gridDim.y % groupRows)) >= gridDim.y)) {
+      // last columns
+      currGroupRows = gridDim.y % groupRows;
+    }
+
+    swizzledBlockIdx.x =
+        linearIdx % currGroupRows + prevGroupRows * (linearIdx / (prevGroupRows * gridDim.x));
+    swizzledBlockIdx.y = (linearIdx / currGroupRows) % gridDim.x;
+    swizzledBlockIdx.z = blockIdx.z;
+
+    return swizzledBlockIdx;
+  }
+
+  ///
+  CUTLASS_HOST_DEVICE dim3 get_grid_layout(GemmCoord const &problem_size,
+                                           Coord<3> const &OutputTile) {
+    dim3 grid;
+    grid.x = (problem_size.n() + OutputTile[1] - 1) / OutputTile[1];
+    grid.y = (problem_size.m() + OutputTile[2] - 1) / OutputTile[2];
+    grid.z = problem_size.batch();
+    return grid;
+  }
+
+  ///
+  CUTLASS_DEVICE Coord<3> get_threadblock_offset(Coord<3> const &OutputTile) {
+    dim3 block = swizzle();
+    Coord<3> threadblock_offset =
+        make_Coord(0, block.y * OutputTile[1], block.x * OutputTile[2]);
+    return threadblock_offset;
+  }
+
+  ///
+  CUTLASS_DEVICE int get_batch_id() {
+    dim3 block = swizzle();
+    return block.z;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/wmma_accumulator.h

@@ -1,215 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * Thread-level multiply-accumulate abstraction
- * (Volta 4B accum_t specialization)
- */
-
-#include <mma.h>
-
-#include "../util/util.h"
-#include "dp_accummulate.h"
-
-namespace cutlass {
-namespace gemm {
-
-
-/*!
- *\brief  matrix_layout to perform conversion between Cutlass types and WMMA types
- */
-template <matrix_transform_t::kind_t>
-struct matrix_layout;
-
-/// Maps matrix_transform_t::NonTranspose to nvcuda::wmma::mem_col_major
-template <>
-struct matrix_layout<matrix_transform_t::NonTranspose>
-{
-    /// Type tag in nvcuda::wmma namespace
-    typedef nvcuda::wmma::col_major tag;
-
-    /// Column major layout
-    static const nvcuda::wmma::layout_t kind = nvcuda::wmma::mem_col_major;
-
-    /// Cutlass matrix transform kind
-    static const matrix_transform_t::kind_t cutlass_kind = matrix_transform_t::NonTranspose;
-};
-
-/// Maps matrix_transform_t::NonTranspose to nvcuda::wmma::mem_row_major
-template <>
-struct matrix_layout<matrix_transform_t::Transpose>
-{
-    /// Type tag in nvcuda::wmma namespace
-    typedef nvcuda::wmma::row_major tag;
-
-    /// Column major layout
-    static const nvcuda::wmma::layout_t kind = nvcuda::wmma::mem_row_major;
-
-    /// Cutlass matrix transform kind
-    static const matrix_transform_t::kind_t cutlass_kind = matrix_transform_t::Transpose;
-};
-
-/*!
- * \brief Warp-synchronous matrix multiply-accumulate abstraction
- *
- * wmma_accumulator maps the CUDA WMMA API onto the GEMM structure
- */
-template <
-    int WarpItemsY,                         /// Number of rows of the warp's accumulator tile
-    int WarpItemsX,                         /// Number of columns of the warp's accumulator tile
-    int WmmaItemsY,                         /// Number of rows in a single WMMA operation
-    int WmmaItemsX,                         /// Number of columns in a single WMMA operation
-    int WmmaItemsK,                         /// Inner dimension of WMMA operation
-    typename value_a_t,                     /// Type of A operand
-    typename value_b_t,                     /// Type of B operand
-    typename accum_t,                       /// Type of source and destination accumulators
-    matrix_transform_t::kind_t TransformA,  /// Layout of A operand
-    matrix_transform_t::kind_t TransformB   /// Layout of B operand
->
-struct wmma_accumulator
-{
-public:
-
-    //-------------------------------------------------------------------------
-    // Constants and types
-    //-------------------------------------------------------------------------
-
-    enum
-    {
-        /// Number of WMMA blocks in warp row
-        WmmaBlocksX = divide_assert<WarpItemsX, WmmaItemsX>::value,
-
-        /// Number of WMMA blocks in a warp column
-        WmmaBlocksY = divide_assert<WarpItemsY, WmmaItemsY>::value,
-    };
-
-    /// Fragment type for matrix operand A
-    typedef nvcuda::wmma::fragment<
-            nvcuda::wmma::matrix_a,
-            WmmaItemsY,
-            WmmaItemsX,
-            WmmaItemsK,
-            value_a_t,
-            typename matrix_layout<TransformA>::tag>
-        fragment_a_t;
-
-    /// Fragment type for matrix operand B
-    typedef nvcuda::wmma::fragment<
-            nvcuda::wmma::matrix_b,
-            WmmaItemsY,
-            WmmaItemsX,
-            WmmaItemsK,
-            value_b_t,
-            typename matrix_layout<TransformB>::tag>
-        fragment_b_t;
-
-    /// Fragment type for accumulator
-    typedef nvcuda::wmma::fragment<
-            nvcuda::wmma::accumulator,
-            WmmaItemsY,
-            WmmaItemsX,
-            WmmaItemsK,
-            accum_t>
-        accumulator_t;
-
-    /// Scratch storage layout
-    struct scratch_storage_t
-    {
-        /// Initialization vector
-        uint4 zero_slab;
-    };
-
-public:
-    //-------------------------------------------------------------------------
-    // Data members
-    //-------------------------------------------------------------------------
-
-    /// Thread's tile of accumulators
-    accumulator_t accumulators[WmmaBlocksX][WmmaBlocksY];
-
-public:
-
-    //-------------------------------------------------------------------------
-    // Constructor API
-    //-------------------------------------------------------------------------
-
-    /// Constructor initializes accumulators to zero
-    inline __device__
-    wmma_accumulator()
-    {
-        init();
-    }
-
-
-    //-------------------------------------------------------------------------
-    // Accumulator API
-    //-------------------------------------------------------------------------
-
-    /**
-     * \brief Zero-initialize thread accumulators.
-     */
-    inline __device__
-    void init()
-    {
-        #pragma unroll
-        for (int x = 0; x < WmmaBlocksX; ++x)
-        {
-            #pragma unroll
-            for (int y = 0; y < WmmaBlocksY; ++y)
-            {
-                nvcuda::wmma::fill_fragment(accumulators[x][y], accum_t(0));
-            }
-        }
-    }
-
-    /**
-     * \brief Compute the product of tile_a and tile_b and add the result to
-     * the tile of accumulators.
-     */
-    inline __device__
-    void multiply_accumulate(
-        fragment_a_t (&tile_a)[WmmaBlocksY],
-        fragment_b_t (&tile_b)[WmmaBlocksX])
-    {
-        #pragma unroll
-        for (int x = 0; x < WmmaBlocksX; ++x)
-        {
-            #pragma unroll
-            for (int y = 0; y < WmmaBlocksY; ++y)
-            {
-                nvcuda::wmma::mma_sync(accumulators[x][y], tile_a[y], tile_b[x], accumulators[x][y]);
-            }
-        }
-    }
-};
-
-
-} // namespace gemm
-} // namespace cutlass

cutlass/gemm/wmma_gemm_epilogue_traits.h

@@ -0,0 +1,164 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines structural properties of WMMA GEMM's epilogue phase.
+*/
+#pragma once
+
+#include "cutlass/wmma_matrix.h"
+#ifdef CUTLASS_USE_WMMA_API
+
+#include "cutlass/convert.h"
+#include "cutlass/coord.h"
+#include "cutlass/gemm/gemm_global_stream.h"
+#include "cutlass/gemm/gemm_shared_stream.h"
+#include "cutlass/gemm/linear_scaling.h"
+#include "cutlass/gemm/wmma_gemm_global_tile.h"
+#include "cutlass/gemm/wmma_gemm_shared_tile.h"
+#include "cutlass/reshape_tile.h"
+#include "cutlass/tile_iterator.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename GemmConfig_, typename EpilogueFunctor_, typename Index_ = int>
+struct WmmaGemmEpilogueTraitsHelper {
+  /// The scalar.
+  typedef typename EpilogueFunctor_::Scalar Scalar;
+  /// The output tile.
+  typedef typename GemmConfig_::OutputTile OutputTile;
+
+  /// The number of WMMAs in the H dimension.
+  static int const kWmmasPerH =
+      GemmConfig_::AccumulatorsPerWarp::kH / GemmConfig_::InstructionShape::kH;
+  /// The number of iterations in the epilogue. That's the number of "horizontal" WMMAs.
+  typedef Shape<1, 1, kWmmasPerH> Iterations;
+  // The iteration strides in the H/W dimension.
+  typedef Shape<0, 0, 0> Delta;
+  /// The functor to do the math in the epilogue.
+  typedef EpilogueFunctor_ Functor;
+
+  /// The traits class to build the iterator to store to shared memory for D.
+  typedef WmmaGemmSharedStoreTileDTraits<
+      // The output layout.
+      MatrixLayout::kColumnMajor,
+      // The pointer is float.
+      typename Functor::Scalar,
+      // The output tile size.
+      typename GemmConfig_::OutputTile,
+      // The number of warps.
+      typename GemmConfig_::Warps,
+      // The shape of the instruction.
+      typename GemmConfig_::InstructionShape>
+      SharedStoreTileTraits;
+
+  typedef WmmaMatrix<GemmOperand::kC,
+                     MatrixLayout::kColumnMajor,
+                     Scalar,
+                     typename GemmConfig_::InstructionShape>
+      WmmaMatrix;
+
+  /// The iterator to store D to shared memory.
+  typedef TileStoreIterator<SharedStoreTileTraits,
+                            typename SharedStoreTileTraits::Scalar,
+                            IteratorAdvance::kH,
+                            MemorySpace::kShared,
+                            Index_,
+                            WmmaMatrix,
+                            FragmentElementType::kWmmaMatrix>
+      SharedStoreIteratorD;
+
+  /// The shared store transformer for D.
+  typedef Copy<typename SharedStoreIteratorD::Fragment> SharedStoreTransformerD;
+
+  /// The traits class to build the iterator to load from shared memory for D.
+  typedef WmmaGemmSharedLoadTileDTraits<
+      // The pointer.
+      typename Functor::Scalar,
+      // The tile size.
+      typename SharedStoreIteratorD::Tile,
+      // The number of threads.
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDS.
+      GemmConfig_::kScalarsPerLdsD>
+      SharedLoadTileTraits;
+
+  /// The iterator to load D from shared memory.
+  typedef TileLoadIterator<SharedLoadTileTraits,
+                           typename SharedLoadTileTraits::Scalar,
+                           IteratorAdvance::kH,
+                           MemorySpace::kShared>
+      SharedLoadIteratorD;
+
+  /// The stream to load D.
+  typedef SharedLoadStream<SharedLoadIteratorD> SharedLoadStreamD;
+
+  /// The traits class to build the iterator to load data from global memory for C^N.
+  typedef WmmaGemmGlobalIteratorCdTraits<
+      // The pointer is float const.
+      typename GemmConfig_::ScalarC const,
+      // The tile has size (N / Iterations)xM in GEMM's terminology.
+      Shape<1,
+            GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
+            GemmConfig_::OutputTile::kW>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerLdgC>
+      GlobalLoadTileTraits;
+
+  /// The iterator to load C.
+  typedef WmmaGemmGlobalIteratorCd<GlobalLoadTileTraits, Index_> GlobalLoadIteratorC;
+  /// The transformer for C.
+  typedef Copy<typename GlobalLoadIteratorC::Fragment> GlobalTransformerC;
+
+  /// The traits class to build the iterator to store data to global memory for D^N.
+  typedef WmmaGemmGlobalIteratorCdTraits<
+      // The pointer is float.
+      typename GemmConfig_::ScalarD,
+      // The tile has size (N / Iterations)xM in GEMM's terminology.
+      Shape<1,
+            GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
+            GemmConfig_::OutputTile::kW>,
+      // The threads are distributed as warps x 32 (the traits may reorganize).
+      Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
+      // The number of scalars per LDG (LDG.32 or LDG.128, etc).
+      GemmConfig_::kScalarsPerStgD>
+      GlobalStoreTileTraits;
+
+  /// The iterator to store D.
+  typedef WmmaGemmGlobalIteratorCd<GlobalStoreTileTraits, Index_> GlobalStoreIteratorD;
+  /// The transformer for D.
+  typedef Copy<typename GlobalStoreIteratorD::Fragment> GlobalTransformerD;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass
+
+#endif  // defined CUTLASS_USE_WMMA_API

cutlass/gemm/wmma_gemm_global_tile.h

@@ -0,0 +1,167 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines tile iterator traits for loading thread block-level tile from global memory.
+*/
+#pragma once
+
+#include "cutlass/gemm/gemm_global_tile.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, typename Tile_, typename Threads_, int kAccessSize_>
+struct WmmaGemmGlobalIteratorCdTraits : public GemmGlobalTileTraits<GemmOperand::kC,
+                                                                    MatrixLayout::kColumnMajor,
+                                                                    Scalar_,
+                                                                    Tile_,
+                                                                    Threads_,
+                                                                    kAccessSize_> {
+  /// The base class.
+  typedef GemmGlobalTileTraits<GemmOperand::kC,
+                               MatrixLayout::kColumnMajor,
+                               Scalar_,
+                               Tile_,
+                               Threads_,
+                               kAccessSize_>
+      Base;
+
+  /// Override the strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Base::Delta::kW, Base::Delta::kC> Delta;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int thread_offset_h = threadIdx.x / Base::Threads::kW;
+      int thread_offset_w = threadIdx.x % Base::Threads::kW * Base::ThreadsDelta::kW;
+
+      return make_Coord(0, thread_offset_h, thread_offset_w, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename TileTraits_, typename Index_ = int>
+struct WmmaGemmGlobalIteratorCd : public GemmGlobalIteratorCd<TileTraits_, Index_> {
+  /// This class.
+  typedef WmmaGemmGlobalIteratorCd<TileTraits_, Index_> This_;
+  /// The traits.
+  typedef TileTraits_ Traits;
+  /// The base class.
+  typedef GemmGlobalIteratorCd<Traits, Index_> Base;
+  /// Override the strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Base::Delta::kW, Base::Delta::kC> ImmediateOffsetStrides;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = TileTraits_::kLayout;
+
+  /// The scalar.
+  typedef typename TileTraits_::Scalar Scalar;
+  /// The pointer.
+  typedef typename TileTraits_::Pointer Pointer;
+  /// The threads.
+  typedef typename TileTraits_::Threads Threads;
+  /// The index.
+  typedef Index_ Index;
+  /// The thread offset functor.
+  typedef typename TileTraits_::ThreadOffset ThreadOffset;
+  /// Base parameters.
+  typedef typename Base::Params BaseParams;
+
+  /// The params.
+  struct Params : public BaseParams {
+    /// Setup the params.
+    CUTLASS_HOST_DEVICE int initialize(Pointer pointer,
+                                       long long batch_stride,
+                                       Index ldm,
+                                       Index n,
+                                       Index epilogue_stride_w,
+                                       Index epilogue_delta_w) {
+      // The pointer.
+      BaseParams::pointer = pointer;
+      // Stride between GEMMs
+      BaseParams::stride_d = batch_stride;
+      // Setup the base stride. One "group of threads" per column.
+      BaseParams::stride_h = ldm;
+      // Each thread output 1 column per iteration. .
+      BaseParams::inc_h = ldm * TileTraits_::Threads::kH;
+      BaseParams::inc_advance = BaseParams::inc_h + epilogue_stride_w;
+
+      BaseParams::predicate_offset = n;
+      BaseParams::predicate_inc_h = TileTraits_::Threads::kH;
+      BaseParams::predicate_inc_advance = BaseParams::predicate_inc_h + epilogue_delta_w;
+
+      return 0;
+    }
+  };
+
+  /// Ctor.
+  CUTLASS_DEVICE WmmaGemmGlobalIteratorCd(Params const& params,
+                                          const Coord<3>& bounds,
+                                          const Coord<3>& block,
+                                          int const pointer_offset = 0,
+                                          int const pred_offset = 0,
+                                          ThreadOffset thread_offset_func = ThreadOffset())
+
+      : Base(params, bounds, block, pointer_offset, pred_offset, thread_offset_func) {}
+
+  /// Loads a single fragment element from memory
+  CUTLASS_DEVICE void load_element(
+      typename Base::AccessType& value, int d, int h, int w, int c) const {
+    Base::load_element(value, d, h, w, c);
+  }
+
+  /// Stores a single fragment element into memory
+  CUTLASS_DEVICE void store_element(
+      typename Base::AccessType const& value, int d, int h, int w, int c) {
+    int const offset =
+        ComputeOffsetFromStrides<typename Base::ImmediateOffsetStrides>::get(d, h, w, 0);
+    Store<Scalar,
+          Base::kAccessSize,
+          Base::kMemorySpace,
+          Base::kFragmentElementType,
+          typename Base::FragmentElement,
+          Base::Tile::kW>::store(value, Base::params.pointer, offset);
+  }
+
+ public:
+  template <typename Fragment>
+  CUTLASS_DEVICE void load_post_increment(Fragment& fragment) {
+    Base::load_post_increment(fragment);
+  }
+
+  template <typename Fragment>
+  CUTLASS_DEVICE void store_post_increment(Fragment& fragment) {
+    Base::store_post_increment(fragment);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass

cutlass/gemm/wmma_gemm_multiply_add.h

@@ -0,0 +1,355 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements warp-level matrix multiply-accumulate operation using CUDA WMMA API.
+*/
+#pragma once
+
+#include "cutlass/wmma_matrix.h"
+#ifdef CUTLASS_USE_WMMA_API
+#include "cutlass/fragment.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <MatrixLayout::Kind kLayoutA_,
+          typename ScalarA_,
+          MatrixLayout::Kind kLayoutB_,
+          typename ScalarB_,
+          MatrixLayout::Kind kLayoutC_,
+          typename ScalarC_,
+          typename WarpGemmShape_,
+          typename InstructionShape_>
+struct WmmaGemmMultiplyAdd {
+  /// The shape of the instruction.
+  typedef InstructionShape_ InstructionShape;
+  /// The number of threads per warp. That's a dummy configuration.
+  typedef Shape<1, InstructionShape_::kH, InstructionShape_::kW> ThreadsPerWarp;
+  /// Dimensions of the warp-level GEMM (K-by-N-by-M)
+  typedef WarpGemmShape_ WarpGemmShape;
+  /// Aliased for compatibility. Will be removed in CUTLASS v2.0
+  typedef WarpGemmShape_ AccumulatorsPerWarp;
+  /// The type for A.
+  typedef ScalarA_ ScalarA;
+  /// The type for B.
+  typedef ScalarB_ ScalarB;
+  /// The type for C and D.
+  typedef ScalarC_ ScalarC;
+  /// The number of iterations.
+  typedef typename ShapeDiv<AccumulatorsPerWarp, InstructionShape>::Shape Iterations;
+
+  /// The element for A.
+  typedef WmmaMatrix<GemmOperand::kA, kLayoutA_, ScalarA, InstructionShape> ElementA;
+  /// The fragment for A.
+  typedef Fragment<ElementA, Iterations::kW> FragmentA;
+
+  /// The element for B.
+  typedef WmmaMatrix<GemmOperand::kB, kLayoutB_, ScalarB, InstructionShape> ElementB;
+  /// The fragment for B.
+  typedef Fragment<ElementB, Iterations::kH> FragmentB;
+
+  /// The element for C.
+  typedef WmmaMatrix<GemmOperand::kC, kLayoutC_, ScalarC, InstructionShape> ElementC;
+  /// The fragment for C.
+  typedef Fragment<ElementC, Iterations::kH * Iterations::kW> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE WmmaGemmMultiplyAdd() {}
+
+  /// Multiply : d = a*b.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    for (int j = 0; j < Iterations::kH; ++j) {
+      for (int i = 0; i < Iterations::kW; ++i) {
+        // The input elements.
+        ElementA const& elt_a = a[i];
+        ElementB const& elt_b = b[j];
+        ElementC const& elt_c = c[j * Iterations::kW + i];
+
+        // The output element.
+        ElementC& elt_d = d[j * Iterations::kW + i];
+
+        // The wmma instruction.
+        nvcuda::wmma::mma_sync(elt_d, elt_a, elt_b, elt_c);
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#ifdef CUTLASS_USE_SUBBYTE_WMMA
+/// Specialization for WMMA GEMM with binary operands
+template<typename WarpGemmShape_>
+struct WmmaGemmMultiplyAdd <MatrixLayout::kRowMajor,
+                            Vector<bin1_t, 32>,
+                            MatrixLayout::kColumnMajor,
+                            Vector<bin1_t, 32>,
+                            MatrixLayout::kColumnMajor,
+                            int,
+                            WarpGemmShape_,
+                            Shape<128, 8, 8> >{
+  /// The shape of the instruction.
+  typedef Shape<128, 8, 8> InstructionShape;
+  /// The number of threads per warp. That's a dummy configuration.
+  typedef Shape<1, 4, 8> ThreadsPerWarp;
+  /// Dimensions of the warp-level GEMM (K-by-N-by-M)
+  typedef WarpGemmShape_ WarpGemmShape;
+  /// Aliased for compatibility. Will be removed in CUTLASS v2.0
+  typedef WarpGemmShape_ AccumulatorsPerWarp;
+  /// The type for A.
+  typedef Vector<bin1_t, 32> ScalarA;
+  /// The type for B.
+  typedef Vector<bin1_t, 32> ScalarB;
+  /// The type for C and D.
+  typedef int ScalarC;
+  /// The number of iterations.
+  typedef typename ShapeDiv<AccumulatorsPerWarp, InstructionShape>::Shape Iterations;
+
+  /// The element for A.
+  typedef WmmaMatrix<GemmOperand::kA,
+                     MatrixLayout::kRowMajor,
+                     Vector<bin1_t, 32>,
+                     InstructionShape> ElementA;
+  /// The fragment for A.
+  typedef Fragment<ElementA, Iterations::kW> FragmentA;
+
+  /// The element for B.
+  typedef WmmaMatrix<GemmOperand::kB,
+                     MatrixLayout::kColumnMajor,
+                     Vector<bin1_t, 32>,
+                     InstructionShape> ElementB;
+  /// The fragment for B.
+  typedef Fragment<ElementB, Iterations::kH> FragmentB;
+
+  /// The element for C.
+  typedef WmmaMatrix<GemmOperand::kC,
+                     MatrixLayout::kColumnMajor,
+                     int,
+                     InstructionShape> ElementC;
+  /// The fragment for C.
+  typedef Fragment<ElementC, Iterations::kH * Iterations::kW> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE WmmaGemmMultiplyAdd() {}
+
+  /// Multiply : d = a*b.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    for (int j = 0; j < Iterations::kH; ++j) {
+      for (int i = 0; i < Iterations::kW; ++i) {
+        // The input elements.
+        ElementA const& elt_a = a[i];
+        ElementB const& elt_b = b[j];
+        ElementC const& elt_c = c[j * Iterations::kW + i];
+
+        // The output element.
+        ElementC& elt_d = d[j * Iterations::kW + i];
+
+        // The wmma instruction.
+        nvcuda::wmma::bmma_sync(elt_d,
+                                elt_a,
+                                elt_b,
+                                elt_c,
+                                nvcuda::wmma::experimental::bmmaBitOpXOR,
+                                nvcuda::wmma::experimental::bmmaAccumulateOpPOPC);
+      }
+    }
+  }
+};
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#ifdef CUTLASS_USE_SUBBYTE_WMMA
+/// Specialization for WMMA GEMM with signed 4-bit integer operands
+template<typename WarpGemmShape_>
+struct WmmaGemmMultiplyAdd <MatrixLayout::kRowMajor,
+                            Vector<int4_t, 8>,
+                            MatrixLayout::kColumnMajor,
+                            Vector<int4_t, 8>,
+                            MatrixLayout::kColumnMajor,
+                            int,
+                            WarpGemmShape_,
+                            Shape<32, 8, 8> >{
+  /// The shape of the instruction.
+  typedef Shape<32, 8, 8> InstructionShape;
+  /// The number of threads per warp. That's a dummy configuration.
+  typedef Shape<1, 4, 8> ThreadsPerWarp;
+  /// Dimensions of the warp-level GEMM (K-by-N-by-M)
+  typedef WarpGemmShape_ WarpGemmShape;
+  /// Aliased for compatibility. Will be removed in CUTLASS v2.0
+  typedef WarpGemmShape_ AccumulatorsPerWarp;
+  /// The type for A.
+  typedef Vector<int4_t, 8> ScalarA;
+  /// The type for B.
+  typedef Vector<int4_t, 8> ScalarB;
+  /// The type for C and D.
+  typedef int ScalarC;
+  /// The number of iterations.
+  typedef typename ShapeDiv<AccumulatorsPerWarp, InstructionShape>::Shape Iterations;
+
+  /// The element for A.
+  typedef WmmaMatrix<GemmOperand::kA,
+                     MatrixLayout::kRowMajor,
+                     Vector<int4_t, 8>,
+                     InstructionShape> ElementA;
+  /// The fragment for A.
+  typedef Fragment<ElementA, Iterations::kW> FragmentA;
+
+  /// The element for B.
+  typedef WmmaMatrix<GemmOperand::kB,
+                     MatrixLayout::kColumnMajor,
+                     Vector<int4_t, 8>,
+                     InstructionShape> ElementB;
+  /// The fragment for B.
+  typedef Fragment<ElementB, Iterations::kH> FragmentB;
+
+  /// The element for C.
+  typedef WmmaMatrix<GemmOperand::kC,
+                     MatrixLayout::kColumnMajor,
+                     int,
+                     InstructionShape> ElementC;
+  /// The fragment for C.
+  typedef Fragment<ElementC, Iterations::kH * Iterations::kW> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE WmmaGemmMultiplyAdd() {}
+
+  /// Multiply : d = a*b.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    for (int j = 0; j < Iterations::kH; ++j) {
+      for (int i = 0; i < Iterations::kW; ++i) {
+        // The input elements.
+        ElementA const& elt_a = a[i];
+        ElementB const& elt_b = b[j];
+        ElementC const& elt_c = c[j * Iterations::kW + i];
+
+        // The output element.
+        ElementC& elt_d = d[j * Iterations::kW + i];
+
+        // The wmma instruction.
+        nvcuda::wmma::mma_sync(elt_d, elt_a, elt_b, elt_c);
+      }
+    }
+  }
+};
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#ifdef CUTLASS_USE_SUBBYTE_WMMA
+/// Specialization for WMMA GEMM with unsigned 4-bit integer operands
+template<typename WarpGemmShape_>
+struct WmmaGemmMultiplyAdd <MatrixLayout::kRowMajor,
+                            Vector<uint4_t, 8>,
+                            MatrixLayout::kColumnMajor,
+                            Vector<uint4_t, 8>,
+                            MatrixLayout::kColumnMajor,
+                            int,
+                            WarpGemmShape_,
+                            Shape<32, 8, 8> >{
+  /// The shape of the instruction.
+  typedef Shape<32, 8, 8> InstructionShape;
+  /// The number of threads per warp. That's a dummy configuration.
+  typedef Shape<1, 4, 8> ThreadsPerWarp;
+  /// Dimensions of the warp-level GEMM (K-by-N-by-M)
+  typedef WarpGemmShape_ WarpGemmShape;
+  /// Aliased for compatibility. Will be removed in CUTLASS v2.0
+  typedef WarpGemmShape_ AccumulatorsPerWarp;
+  /// The type for A.
+  typedef Vector<uint4_t, 8> ScalarA;
+  /// The type for B.
+  typedef Vector<uint4_t, 8> ScalarB;
+  /// The type for C and D.
+  typedef int ScalarC;
+  /// The number of iterations.
+  typedef typename ShapeDiv<AccumulatorsPerWarp, InstructionShape>::Shape Iterations;
+
+  /// The element for A.
+  typedef WmmaMatrix<GemmOperand::kA,
+                     MatrixLayout::kRowMajor,
+                     Vector<uint4_t, 8>,
+                     InstructionShape> ElementA;
+  /// The fragment for A.
+  typedef Fragment<ElementA, Iterations::kW> FragmentA;
+
+  /// The element for B.
+  typedef WmmaMatrix<GemmOperand::kB,
+                     MatrixLayout::kColumnMajor,
+                     Vector<uint4_t, 8>,
+                     InstructionShape> ElementB;
+  /// The fragment for B.
+  typedef Fragment<ElementB, Iterations::kH> FragmentB;
+
+  /// The element for C.
+  typedef WmmaMatrix<GemmOperand::kC,
+                     MatrixLayout::kColumnMajor,
+                     int,
+                     InstructionShape> ElementC;
+  /// The fragment for C.
+  typedef Fragment<ElementC, Iterations::kH * Iterations::kW> Accumulators;
+
+  /// Ctor.
+  CUTLASS_DEVICE WmmaGemmMultiplyAdd() {}
+
+  /// Multiply : d = a*b.
+  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
+                                   FragmentB const& b,
+                                   Accumulators const& c,
+                                   Accumulators& d) {
+    for (int j = 0; j < Iterations::kH; ++j) {
+      for (int i = 0; i < Iterations::kW; ++i) {
+        // The input elements.
+        ElementA const& elt_a = a[i];
+        ElementB const& elt_b = b[j];
+        ElementC const& elt_c = c[j * Iterations::kW + i];
+
+        // The output element.
+        ElementC& elt_d = d[j * Iterations::kW + i];
+
+        // The wmma instruction.
+        nvcuda::wmma::mma_sync(elt_d, elt_a, elt_b, elt_c);
+      }
+    }
+  }
+};
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass
+
+#endif  // defined CUTLASS_USE_WMMA_API

cutlass/gemm/wmma_gemm_shared_tile.h

@@ -0,0 +1,237 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines iterator traits for efficiently loading and storing fragment to and from shared
+      memory, specialized for WMMA GEMM.
+*/
+#pragma once
+
+#include "cutlass/wmma_matrix.h"
+#ifdef CUTLASS_USE_WMMA_API
+
+#include "cutlass/gemm/gemm_operand.h"
+#include "cutlass/reshape_tile.h"
+
+namespace cutlass {
+namespace gemm {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename Tile_,
+          typename Warps_,
+          int kWarpStride_,
+          typename Iterations_,
+          typename Delta_,
+          typename WmmaShape_>
+struct WmmaGemmSharedLoadTileATraits {
+  /// The operand.
+  static GemmOperand::Kind const kOperand = GemmOperand::kA;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = kLayout_;
+  /// The scalar.
+  typedef Scalar_ Scalar;
+  /// The pointer.
+  typedef Scalar const* Pointer;
+  /// The access size
+  static int const kAccessSize = 1;
+  /// The tile with skew.
+  typedef Tile_ Tile;
+  /// The number of warps.
+  typedef Warps_ Warps;
+  /// The warps strides.
+  static int const kWarpStride = kWarpStride_;
+  /// The number of iterations.
+  typedef Iterations_ Iterations;
+  /// The strides between iterations.
+  typedef Delta_ Delta;
+  /// The strides between iterations.
+  typedef Delta_ ImmediateOffsetStrides;
+  /// The shape of the WMMA instruction.
+  typedef WmmaShape_ WmmaShape;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+  /// ThreadOffset
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      // The warp id.
+      int const warp = threadIdx.x / kWarpSize;
+      // The offset.
+      int const offset = warp % Warps::kW * kWarpStride;
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename Tile_,
+          typename Warps_,
+          int kWarpStride_,
+          typename Iterations_,
+          typename Delta_,
+          typename WmmaShape_>
+struct WmmaGemmSharedLoadTileBTraits {
+  /// The operand.
+  static GemmOperand::Kind const kOperand = GemmOperand::kB;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = kLayout_;
+  /// The scalar.
+  typedef Scalar_ Scalar;
+  /// The pointer.
+  typedef Scalar const* Pointer;
+  /// The access size
+  static int const kAccessSize = 1;
+  /// The tile with skew.
+  typedef Tile_ Tile;
+  /// The number of warps.
+  typedef Warps_ Warps;
+  /// The warps strides.
+  static int const kWarpStride = kWarpStride_;
+  /// The number of iterations.
+  typedef Iterations_ Iterations;
+  /// The strides between iterations.
+  typedef Delta_ Delta;
+  /// The strides between iterations.
+  typedef Delta_ ImmediateOffsetStrides;
+  /// The shape of the WMMA instruction.
+  typedef WmmaShape_ WmmaShape;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+  /// ThreadOffset
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      // The warp id.
+      int const warp = threadIdx.x / kWarpSize;
+      // The offset.
+      int const offset = warp / Warps::kW * kWarpStride;
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename OutputTile_,
+          typename Warps_,
+          typename WmmaShape_,
+          int kSkew_ = 0>
+struct WmmaGemmSharedStoreTileDTraits {
+  /// The operand.
+  static GemmOperand::Kind const kOperand = GemmOperand::kC;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = kLayout_;
+  /// The scalar.
+  typedef Scalar_ Scalar;
+  // The access size
+  static int const kAccessSize = 1;
+  /// The pointer.
+  typedef Scalar* Pointer;
+  /// The number of warps.
+  typedef Warps_ Warps;
+  /// The shape of the WMMA instruction.
+  typedef WmmaShape_ WmmaShape;
+  /// The skew.
+  static int const kSkew = kSkew_;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+  /// The tile with skew.
+  typedef Shape<1, Warps_::kH * WmmaShape_::kH, OutputTile_::kW + kSkew_> Tile;
+  /// The number of iterations needed to store the tile.
+  typedef Shape<1, 1, OutputTile_::kW / Warps::kW / WmmaShape_::kW> Iterations;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Warps::kW * WmmaShape_::kW, 0> Delta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, 0, Warps::kW * WmmaShape_::kW, 0> ImmediateOffsetStrides;
+
+  /// ThreadOffset
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      // The warp id.
+      int const warp = threadIdx.x / kWarpSize;
+      // The starting column.
+      int const h = warp / Warps::kW * WmmaShape::kH;
+      // The w.
+      int const w = warp % Warps::kW * WmmaShape::kW;
+      // The offset.
+      int const offset = h * Tile::kW + w;
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, typename Tile_, typename Threads_, int kScalarsPerLds_>
+struct WmmaGemmSharedLoadTileDTraits {
+  /// The scalar.
+  typedef Scalar_ Scalar;
+  /// The pointer.
+  typedef Scalar const* Pointer;
+  /// The access size
+  static int const kAccessSize = kScalarsPerLds_;
+  /// The tile.
+  typedef typename ReshapeTile<Tile_, kScalarsPerLds_>::Tile Tile;
+  /// The threads.
+  typedef typename ReshapeThreads<Tile, Threads_>::Threads Threads;
+  /// The threads strides.
+  typedef Shape<1, Tile::kW * Tile::kC, Tile::kC> ThreadsStrides;
+  /// The memory space.
+  static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
+
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, Threads::kH * ShapeCount<Tile>::kWc, Threads::kW * kScalarsPerLds_> Delta;
+  /// The strides in each dimension between different loads/stores.
+  typedef Shape<0, Threads::kH * ShapeCount<Tile>::kWc, Threads::kW * kScalarsPerLds_>
+      ImmediateOffsetStrides;
+  /// The number of iterations needed to load/store the tile.
+  typedef Shape<1, Tile::kH / Threads::kH, Tile::kW / Threads::kW, Tile::kC / kScalarsPerLds_>
+      Iterations;
+
+  /// ThreadOffset
+  struct ThreadOffset {
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      // The offset.
+      int const offset = ComputeThreadOffsetFromStrides<Threads, ThreadsStrides>::get();
+      return make_Coord(0, 0, offset, 0);
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace gemm
+}  // namespace cutlass
+
+#endif  // defined CUTLASS_USE_WMMA_API

cutlass/iterator_access.h

@@ -0,0 +1,101 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Free functions for loading and storing to implementations of tile iteartor concepts.
+*/
+#pragma once
+
+#include "cutlass/load_store.h"
+#include "cutlass/predicate_vector.h"
+#include "cutlass/shape.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Used by convolution
+template <typename InputIterator, typename Fragment>
+CUTLASS_HOST_DEVICE void iterator_load(InputIterator &iterator, Fragment &fragment) {
+  typename InputIterator::FragmentIterator frag_iterator(fragment);
+  for (int d = 0; d < InputIterator::Iterations::kD; ++d) {
+    for (int h = 0; h < InputIterator::Iterations::kH; ++h) {
+      for (int w = 0; w < InputIterator::Iterations::kW; ++w) {
+        for (int c = 0; c < InputIterator::Iterations::kC; ++c) {
+          if (iterator.valid(d, h, w, c)) {
+            iterator.load_element(reinterpret_cast<typename InputIterator::AccessType &>(
+                                      frag_iterator.at(d, h, w, c)),
+                                  d,
+                                  h,
+                                  w,
+                                  c);
+          }
+        }
+        if (w < InputIterator::Iterations::kW - 1) {
+          iterator.inc_w();
+        }
+      }
+      if (h < InputIterator::Iterations::kH - 1) {
+        iterator.inc_h();
+      }
+    }
+    if (d < InputIterator::Iterations::kD - 1) {
+      iterator.inc_d();
+    }
+  }
+  iterator.inc_advance();
+}
+
+template <typename OutputIterator, typename Fragment>
+CUTLASS_HOST_DEVICE void iterator_store(OutputIterator &iterator, Fragment &fragment) {
+  typename OutputIterator::FragmentIterator frag_iterator(fragment);
+  for (int d = 0; d < OutputIterator::Iterations::kD; ++d) {
+    for (int h = 0; h < OutputIterator::Iterations::kH; ++h) {
+      for (int w = 0; w < OutputIterator::Iterations::kW; ++w) {
+        for (int c = 0; c < OutputIterator::Iterations::kC; ++c) {
+          if (iterator.valid(d, h, w, c)) {
+            iterator.store_element(reinterpret_cast<typename OutputIterator::AccessType &>(
+                                       frag_iterator.at(d, h, w, c)),
+                                   d,
+                                   h,
+                                   w,
+                                   c);
+          }
+        }
+        if (w < OutputIterator::Iterations::kW - 1) {
+          iterator.inc_w();
+        }
+      }
+      if (h < OutputIterator::Iterations::kH - 1) {
+        iterator.inc_h();
+      }
+    }
+    if (d < OutputIterator::Iterations::kD - 1) {
+      iterator.inc_d();
+    }
+  }
+  iterator.inc_advance();
+}
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/kernel_launch.h

@@ -0,0 +1,67 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *   * Redistributions of source code must retain the above copyright notice, this list of
+ *     conditions and the following disclaimer.
+ *   * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *     conditions and the following disclaimer in the documentation and/or other materials
+ *     provided with the distribution.
+ *   * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *     to endorse or promote products derived from this software without specific prior written
+ *     permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+  \brief Defines structures and helpers to launch CUDA kernels within CUTLASS.
+*/
+
+#pragma once
+
+#include "cutlass/cutlass.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Structure containing the basic launch configuration of a CUDA kernel.
+struct KernelLaunchConfiguration {
+
+  /// CUDA grid dimensions
+  dim3 grid;
+
+  /// CUDA threablock dimensions
+  dim3 block;
+
+  /// Bytes of dynamically allocated SMEM in addition to static SMEM
+  size_t dynamic_smem;
+
+  //
+  // Methods
+  //
+
+  /// Constructs a KernellaunchConfiguration object
+  CUTLASS_HOST_DEVICE
+  KernelLaunchConfiguration(
+    dim3 _grid = dim3(1,1,1),
+    dim3 _block = dim3(1,1,1),
+    size_t _dynamic_smem = 0
+  ):
+    grid(_grid),
+    block(_block),
+    dynamic_smem(_dynamic_smem) { }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/load_store.h

@@ -0,0 +1,379 @@
+/***************************************************************************************************
+ * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines abstractions for efficiently loading and storing vectors to memory.
+*/
+#pragma once
+
+#include "cutlass/vector.h"
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief Enum to specify which memory space data resides in.
+*/
+struct MemorySpace {
+  enum Kind {
+    kGeneric,  // Data accessed through pointer dereferencing
+    kShared,   // Data resides in shared memory
+    kGlobal    // Data resides in global memory
+  };
+};
+
+/// Specifies whether iterator storage fragment consists of Scalar values or WMMA matrix
+struct FragmentElementType {
+  enum Kind { kScalar, kWmmaMatrix };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          int kAccessSize,
+          MemorySpace::Kind Memory_,
+          FragmentElementType::Kind kFragmentElementType = FragmentElementType::kScalar,
+          typename FragmentElement_ = Scalar_,
+          int kStride = 1,
+          size_t size = (sizeof(Scalar_) * kAccessSize)>
+struct Load {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, Scalar_ const* pointer, int offset) {
+    dst = *reinterpret_cast<AccessType const*>(pointer + offset);
+  }
+
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Partial specialization for 16b loads
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_>
+struct Load<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, 1, 2> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, Scalar_ const* pointer, int offset) {
+    reinterpret_cast<uint16_t&>(dst) = reinterpret_cast<uint16_t const*>(&pointer[offset])[0];
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_, int kStride>
+struct Load<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, kStride, 4> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, Scalar_ const* pointer, int offset) {
+    dst.registers[0] = reinterpret_cast<uint32_t const*>(&pointer[offset])[0];
+  }
+
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_, int kStride>
+struct Load<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, kStride, 8> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, Scalar_ const* pointer, int offset) {
+    uint2 tmp = reinterpret_cast<uint2 const*>(&pointer[offset])[0];
+    dst.registers[0] = tmp.x;
+    dst.registers[1] = tmp.y;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <MemorySpace::Kind Memory_, int kStride>
+struct Load<double, 2, Memory_, FragmentElementType::kScalar, double, kStride, 16> {
+  /// The output type.
+  typedef typename Vectorize<double, 2>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, double const* pointer, int offset) {
+    double2 tmp = reinterpret_cast<double2 const*>(&pointer[offset])[0];
+    dst[0] = tmp.x;
+    dst[1] = tmp.y;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#if defined(__CUDACC_VERSION_MAJOR) && __CUDACC_VERSION_MAJOR < 10
+// WAR bug in NVCC where the upper and lower half of the register end up being the same
+template <MemorySpace::Kind Memory_, int kStride>
+struct Load<half, 8, Memory_, FragmentElementType::kScalar, half, kStride, 16> {
+  /// The output type.
+  typedef typename Vectorize<half, 8>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, half const* pointer, int offset) {
+    int2 tmp = reinterpret_cast<int2 const*>(&pointer[offset])[0];
+    dst.registers[0] = tmp.x;
+    dst.registers[1] = tmp.y;
+
+    tmp = reinterpret_cast<int2 const*>(&pointer[offset + 4])[0];
+    dst.registers[2] = tmp.x;
+    dst.registers[3] = tmp.y;
+  }
+};
+
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_, int kStride>
+struct Load<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, kStride, 16> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& dst, Scalar_ const* pointer, int offset) {
+    uint4 tmp = reinterpret_cast<uint4 const*>(&pointer[offset])[0];
+    dst.registers[0] = tmp.x;
+    dst.registers[1] = tmp.y;
+    dst.registers[2] = tmp.z;
+    dst.registers[3] = tmp.w;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          int kAccessSize,
+          MemorySpace::Kind Memory_,
+          FragmentElementType::Kind kFragmentElementType = FragmentElementType::kScalar,
+          typename FragmentElement_ = Scalar_,
+          int kStride = 1,
+          size_t size = (sizeof(Scalar_) * kAccessSize)>
+struct Store {
+  /// The output type.
+  typedef typename Vectorize<FragmentElement_, kAccessSize>::Type AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& src, Scalar_* pointer, int offset) {
+    pointer[offset] = *reinterpret_cast<Scalar_ const*>(&src);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_>
+struct Store<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, 1, 2> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& src, Scalar_* pointer, int offset) {
+    uint16_t* addr = reinterpret_cast<uint16_t*>(&pointer[offset]);
+    addr[0] = reinterpret_cast<uint16_t const&>(src);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_, int kStride>
+struct Store<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, kStride, 4> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& src, Scalar_* pointer, int offset) {
+    uint32_t* addr = reinterpret_cast<uint32_t*>(&pointer[offset]);
+    addr[0] = src.registers[0];
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_, int kStride>
+struct Store<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, kStride, 8> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& src, Scalar_* pointer, int offset) {
+    uint2* addr = reinterpret_cast<uint2*>(&pointer[offset]);
+    addr[0] = make_uint2(src.registers[0], src.registers[1]);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <MemorySpace::Kind Memory_, int kStride>
+struct Store<double, 2, Memory_, FragmentElementType::kScalar, double, kStride, 16> {
+  /// The output type.
+  typedef typename Vectorize<double, 2>::Type AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& src, double* pointer, int offset) {
+    double2* addr = reinterpret_cast<double2*>(&pointer[offset]);
+    addr[0] = make_double2(src[0], src[1]);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kAccessSize, MemorySpace::Kind Memory_, int kStride>
+struct Store<Scalar_, kAccessSize, Memory_, FragmentElementType::kScalar, Scalar_, kStride, 16> {
+  /// The output type.
+  typedef typename Vectorize<Scalar_, kAccessSize>::Type AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& src, Scalar_* pointer, int offset) {
+    uint4* addr = reinterpret_cast<uint4*>(&pointer[offset]);
+    addr[0] = make_uint4(src.registers[0], src.registers[1], src.registers[2], src.registers[3]);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_,
+          int kAccessSize,
+          MemorySpace::Kind Memory_,
+          typename FragmentElement_,
+          int kStride,
+          size_t size>
+struct Load<Scalar_,
+            kAccessSize,
+            Memory_,
+            FragmentElementType::kWmmaMatrix,
+            FragmentElement_,
+            kStride,
+            size> {
+  /// The output type.
+  typedef FragmentElement_ AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& value, Scalar_ const* pointer, int offset) {
+    value.load(&pointer[offset], kStride);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kAccessSize,
+          MemorySpace::Kind Memory_,
+          typename FragmentElement_,
+          int kStride,
+          size_t size>
+struct Load<Vector<bin1_t, 32>,
+            kAccessSize,
+            Memory_,
+            FragmentElementType::kWmmaMatrix,
+            FragmentElement_,
+            kStride,
+            size> {
+  /// The output type.
+  typedef FragmentElement_ AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& value, Vector<bin1_t, 32> const* pointer,
+                                       int offset) {
+    value.load(&pointer[offset], kStride * 32);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kAccessSize,
+          MemorySpace::Kind Memory_,
+          typename FragmentElement_,
+          int kStride,
+          size_t size>
+struct Load<Vector<int4_t, 8>,
+            kAccessSize,
+            Memory_,
+            FragmentElementType::kWmmaMatrix,
+            FragmentElement_,
+            kStride,
+            size> {
+  /// The output type.
+  typedef FragmentElement_ AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& value, Vector<int4_t, 8> const* pointer,
+                                       int offset) {
+    value.load(&pointer[offset], kStride * 8);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kAccessSize,
+          MemorySpace::Kind Memory_,
+          typename FragmentElement_,
+          int kStride,
+          size_t size>
+struct Load<Vector<uint4_t, 8>,
+            kAccessSize,
+            Memory_,
+            FragmentElementType::kWmmaMatrix,
+            FragmentElement_,
+            kStride,
+            size> {
+  /// The output type.
+  typedef FragmentElement_ AccessType;
+
+  /// The load function.
+  static CUTLASS_HOST_DEVICE void load(AccessType& value, Vector<uint4_t, 8> const* pointer,
+                                       int offset) {
+    value.load(&pointer[offset], kStride * 8);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+template <typename Scalar_,
+          int kAccessSize,
+          MemorySpace::Kind Memory_,
+          typename FragmentElement_,
+          int kStride,
+          size_t size>
+struct Store<Scalar_,
+             kAccessSize,
+             Memory_,
+             FragmentElementType::kWmmaMatrix,
+             FragmentElement_,
+             kStride,
+             size> {
+  /// The input type.
+  typedef FragmentElement_ AccessType;
+
+  /// The store function.
+  static CUTLASS_HOST_DEVICE void store(AccessType const& value, Scalar_* pointer, int offset) {
+    value.store(&pointer[offset], kStride);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/matrix_traits.h

@@ -0,0 +1,372 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines properties of matrices used to denote layout and operands to GEMM kernels.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// MatrixCoord wraps Coord<2, int> to provide a helper for accessing named dimensions. Classes
+/// expecting a coordinate in the rank=2 index space of a matrix should use MatrixCoord.
+struct MatrixCoord : public Coord<2, int> {
+
+  /// Integer-valued index
+  typedef int Index;
+
+  /// Base type is a Coord of rank=2
+  typedef Coord<2, Index> Base;
+
+  /// Rows dimension
+  static int const kRow = 0;
+
+  /// Columns dimension
+  static int const kColumn = 1;
+
+  //
+  // Methods
+  //
+
+  /// Default ctor
+  CUTLASS_HOST_DEVICE
+  MatrixCoord() { }
+
+  /// Constructs from Coord<2>
+  CUTLASS_HOST_DEVICE
+  MatrixCoord(Coord<2, Index> const &coord): Base(coord) { }
+
+  /// Helper to construct from a row and column
+  CUTLASS_HOST_DEVICE
+  MatrixCoord(Index row, Index column): Base(make_Coord(row, column)) { }
+
+  /// Returns the row of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & row() const { return this->at(kRow); }
+
+  /// Returns the row of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index & row() { return this->at(kRow); }
+
+  /// Returns the column of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & column() const { return this->at(kColumn); }
+
+  /// Returns the column of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index & column() { return this->at(kColumn); }
+
+  //
+  // Coord operators
+  //
+
+  /// Element-wise addition
+  CUTLASS_HOST_DEVICE
+  MatrixCoord operator+(Base const& b) const {
+    return MatrixCoord(Base::operator+(b));
+  }
+
+  /// Element-wise subtraction
+  CUTLASS_HOST_DEVICE
+  MatrixCoord operator-(Base const& b) const {
+    return MatrixCoord(Base::operator-(b));
+  }
+
+  /// Element-wise multiplication
+  CUTLASS_HOST_DEVICE
+  MatrixCoord operator*(Base const& b) const {
+    return MatrixCoord(Base::operator*(b));
+  }
+
+  /// Element-wise division
+  CUTLASS_HOST_DEVICE
+  MatrixCoord operator/(Base const& b) const {
+    return MatrixCoord(Base::operator/(b));
+  }
+
+  /// In-place addition
+  CUTLASS_HOST_DEVICE
+  MatrixCoord& operator+=(Base const& b) {
+    Base::operator+=(b);
+    return *this;
+  }
+
+  /// In-place subtraction
+  CUTLASS_HOST_DEVICE
+  MatrixCoord& operator-=(Base const& b) {
+    Base::operator-=(b);
+    return *this;
+  }
+
+  /// In-place multiplication
+  CUTLASS_HOST_DEVICE
+  MatrixCoord& operator*=(Base const& b) {
+    Base::operator*=(b);
+    return *this;
+  }
+
+  /// In-place division
+  CUTLASS_HOST_DEVICE
+  MatrixCoord& operator/=(Base const& b) {
+    Base::operator/=(b);
+    return *this;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Defines data layouts of various matrix formats usable by TensorRef and other classes.
+//
+// The following define classes satisfying the TensorRefMapFunc concept. These must support the
+// following operations, where func is an instance of type TensorRefMapFunc.
+//
+//   Coord<TensorRefMapFunc::kStorageRank> = func(Coord<kRank>);
+//
+// Though not required to be usable by TensorRef, each of the following also define a helper
+// function to map the "leading dimension" to an appropriate stride vector. Implementations
+// following this convention should also implement the following static method:
+//
+//   Coord<TensorRefMapFunc::kStorageRank> stride = TensorRefMapFunc::stride(leading_dim);
+//
+struct MatrixLayout {
+
+  /// Enumeration defining fundamental contiguous layouts.
+  enum Kind { kRowMajor, kColumnMajor };
+
+  //
+  // TensorRefMapFunc definitions for common layouts
+  //
+
+  /// Mapping function for row-major matrices
+  struct RowMajor {
+    static int const kStorageRank = 2;
+    /// Maps (i, j) to (i, j)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+      return coord;
+    }
+  };
+
+  /// Mapping function for column-major matrices
+  struct ColumnMajor {
+    static int const kStorageRank = 2;
+    /// Maps (i, j) to (j, i)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+      return make_Coord(coord.column(), coord.row());
+    }
+  };
+
+  /// Mapping function for interleaved matrices. Matrix is structured
+  /// as row-major arrangement of fixed-size columns.
+  template <int Interleave>
+  struct RowMajorInterleaved {
+
+    /// Rank of storage n-D array
+    static int const kStorageRank = 3;
+
+    /// Interleaving size
+    static int const kInterleave = Interleave;
+
+    /// Maps (row, col) to (row, col, row)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+      return make_Coord(
+        coord.row() / kInterleave,
+        coord.column(),
+        coord.row() % kInterleave
+      );
+    }
+
+    /// Helper to compute stride vector from leading dimension
+    CUTLASS_HOST_DEVICE
+    static Coord<kStorageRank> stride(int ldm) {
+      return make_Coord(
+        ldm * kInterleave,
+        kInterleave,
+        1
+      );
+    }
+  };
+
+  /// Mapping function for interleaved matrices. Matrix is structured
+  /// as column-major arrangement of fixed-size rows.
+  template <int Interleave>
+  struct ColumnMajorInterleaved {
+
+    /// Rank of storage n-D array
+    static int const kStorageRank = 3;
+
+    /// Interleaving size
+    static int const kInterleave = Interleave;
+
+    /// Maps (row, col) to (col, row, col)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+      return make_Coord(
+        coord.column() / kInterleave,
+        coord.row(),
+        coord.column() % kInterleave
+      );
+    }
+
+    /// Helper to compute stride vector from leading dimension
+    CUTLASS_HOST_DEVICE
+    static Coord<kStorageRank> stride(int ldm) {
+      return make_Coord(
+        ldm * kInterleave,
+        kInterleave,
+        1
+      );
+    }
+  };
+
+  /// Mapping function for scenario in which layout is row-major or column-major but this information
+  /// is only available at runtime.
+  struct ContiguousLayout {
+    /// Arbitrary storage rank
+    static int const kStorageRank = 3;
+
+    /// Dimension of rows
+    static int const kRow = 0;
+
+    /// Dimension of columns
+    static int const kColumn = 1;
+
+    /// Mapping function defined by runtime variable. Returns coordinates in n-D storage array
+    /// as (matrix row, matrix colum, 0)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+        return make_Coord(coord.row(), coord.column(), 0);
+    }
+
+    /// Helper to construct a stride vector based on contiguous matrix layout and leading dimension
+    CUTLASS_HOST_DEVICE
+    static Coord<kStorageRank> stride(MatrixLayout::Kind layout, int ldm) {
+      if (layout == MatrixLayout::kRowMajor) {
+        return make_Coord(ldm, 1, 1);
+      }
+      return make_Coord(1, ldm, 1);
+    }
+  };
+
+  /// Mapping function for block-linear matrices. Matrix is structured
+  /// as column-major arrangement of 2D tiles (that are column-major).
+  template <int BlockRows, int BlockColumns>
+  struct ColumnMajorBlockLinear {
+
+    /// Rank of storage n-D array
+    static int const kStorageRank = 4;
+
+    /// Interleaving size in rows dimension
+    static int const kBlockRows = BlockRows;
+
+    /// Interleaving size in columns dimension
+    static int const kBlockColumns = BlockColumns;
+
+    /// Maps (row, col) to (col, row, col, row)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+      return make_Coord(
+        coord.column() / kBlockColumns,
+        coord.row() / kBlockRows,
+        coord.column() % kBlockColumns,
+        coord.row() % kBlockRows
+      );
+    }
+
+    /// Helper to compute stride vector from leading dimension
+    CUTLASS_HOST_DEVICE
+    static Coord<kStorageRank> stride(int ldm) {
+      return make_Coord(
+        ldm * kBlockRows * kBlockColumns,
+        kBlockRows * kBlockColumns,
+        kBlockRows,
+        1
+      );
+    }
+  };
+
+  /// Mapping function for block-linear matrices. Matrix is structured
+  /// as row-major arrangement of 2D tiles (that are row-major)
+  template <int BlockRows, int BlockColumns>
+  struct RowMajorBlockLinear {
+
+    /// Rank of storage n-D array
+    static int const kStorageRank = 4;
+
+    /// Interleaving size in rows dimension
+    static int const kBlockRows = BlockRows;
+
+    /// Interleaving size in columns dimension
+    static int const kBlockColumns = BlockColumns;
+
+    /// Maps (row, col) to (row, col, row, col)
+    CUTLASS_HOST_DEVICE
+    Coord<kStorageRank> operator()(MatrixCoord const &coord) const {
+      return make_Coord(
+        coord.row() / kBlockRows,
+        coord.column() / kBlockColumns,
+        coord.row() % kBlockRows,
+        coord.column() % kBlockColumns
+      );
+    }
+
+    /// Helper to compute stride vector from leading dimension
+    CUTLASS_HOST_DEVICE
+    static Coord<kStorageRank> stride(int ldm) {
+      return make_Coord(
+        ldm * kBlockRows * kBlockColumns,
+        kBlockRows * kBlockColumns,
+        kBlockColumns,
+        1
+      );
+    }
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Gemm operand - D = A * B + C
+struct GemmOperand {
+  enum Kind { kA, kB, kC, kD };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Transformation applied to matrix operands
+struct MatrixTransform {
+  enum Kind {
+    kNone,       /// no operation
+    kConjugate,  /// conjugate
+  };
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/predicate_vector.h

@@ -0,0 +1,494 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines container classes and iterators for managing a statically sized vector
+      of boolean predicates.
+*/
+#pragma once
+
+#include <assert.h>
+#include <stdint.h>
+
+#include "cutlass/cutlass.h"
+#include "cutlass/shape.h"
+
+#include "cutlass/util/platform.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup predicate_vector_concept Predicate Vector Concept
+@{
+
+Implementations of \ref predicate_vector_concept contain an ordered set of boolean predicates which
+may be used as conditionals in other device-side operations. Both random access and iterators
+offering sequential access are provided.
+
+@par Predicate Vector
+   A \ref predicate_vector_concept satisfies the following expressions
+  - <b>at(int idx)</b> - returns the value of the indexed predicate
+  - <b>set(int idx, bool value)</b> - sets the value of the indexed predicate
+  - <b>begin()</b> - returns a \ref predicate_iterator_concept pointing to the first predicate
+
+@}
+*/
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup predicate_iterator_concept Predicate Iterator Concept
+@{
+
+Implementations of \ref predicate_iterator_concept enables accessing and traversing elements of a
+bit vector.
+
+@par Const Predicate Iterator
+  A const \ref predicate_iterator_concept satisfies the following expressions
+ - <b>++it</b> increments the iterator to the next predicate
+ - <b>*it</b> returns the value of the currently pointed-to predicate
+
+@par Mutable Predicate Iterator
+ A \ref predicate_iterator_concept that is non-const <b>also</b> satisfies the following expressions
+ - <b>it.set(bool value)</b> sets the value of the currently pointed-to predicate
+
+@}
+*/
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup predicate_tile_adapter Predicate Tile Adapter Concept
+@{
+
+Implementations of \ref predicate_tile_adapter provide a mapping between a the elements of a \ref
+tile_traits_concept and a \ref predicate_vector_concept.
+
+@par Predicate Tile Adapter
+  A \ref predicate_tile_adapter satisfies the following expressions
+ - <b>at(int d, int h, int w, int c)</b> - returns the value of a predicate corresponding to the
+   access (d, h, w, c) within the tile.
+
+@}
+*/
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Statically sized array of bits implementing @concept{predicate_vector_concept}.
+template <
+    /// Number of predicates conatined in predicate vector
+    int kPredicates_,
+    /// Number of predicates contained in each byte of internal storage
+    int kPredicatesPerByte_ = 4,
+    /// Location of first predicate within byte of internal storage
+    int kPredicateStart_ = 0>
+struct PredicateVector {
+  /// Number of bits stored by the PredicateVector
+  static int const kPredicates = kPredicates_;
+
+  /// Number of bits stored within each byte of the predicate bit vector
+  static int const kPredicatesPerByte = kPredicatesPerByte_;
+
+  /// First bit withing each byte containing predicates
+  static int const kPredicateStart = kPredicateStart_;
+
+  // Make sure no one tries to put more than 8 bits in a byte :)
+  static_assert(kPredicatesPerByte <= 8, "kPredicatesPerByte must fit within an actual byte");
+  // Make sure the "offsetted" bits fit in one byte.
+  static_assert(kPredicateStart + kPredicatesPerByte <= 8,
+                "The offsetted predicates must fit within an actual byte.");
+
+  /// Storage type of individual elements
+  typedef uint32_t Storage;
+
+  /// Number of bytes needed
+  static int const kBytes = (kPredicates + kPredicatesPerByte - 1) / kPredicatesPerByte;
+
+  /// Number of storage elements needed
+  static int const kWordCount = (kBytes + sizeof(Storage) - 1) / sizeof(Storage);
+
+ private:
+  //
+  // Data members
+  //
+
+  /// Words of bit vector
+  Storage storageData[kWordCount];
+
+  //
+  // Methods
+  //
+
+  /// Computes the word and bit corresponding to a logical predicate index
+  CUTLASS_HOST_DEVICE void computeStorageOffset(int &word, int &bit, int idx) const {
+    CUTLASS_ASSERT(idx < kPredicates);
+
+    int byte = (idx / kPredicatesPerByte);
+    int bit_offset = (idx % kPredicatesPerByte);
+
+    word = byte / sizeof(Storage);
+    int byte_offset = (byte % sizeof(Storage));
+
+    bit = byte_offset * 8 + bit_offset + kPredicateStart;
+  }
+
+  /// Accesses a given word with optional assertions
+  CUTLASS_HOST_DEVICE Storage &storage(int word) {
+    CUTLASS_ASSERT(word < kWordCount);
+    return storageData[word];
+  }
+
+  /// Accesses a given word with optional assertions
+  CUTLASS_HOST_DEVICE Storage const &storage(int word) const {
+    CUTLASS_ASSERT(word < kWordCount);
+    return storageData[word];
+  }
+
+ public:
+  //
+  // Iterator
+  //
+
+  /**
+  * @brief A const iterator implementing \ref predicate_iterator_concept enabling sequential
+  * read-only access to prediactes.
+  * @concept{predicate_iterator_concept}
+  */
+  class ConstIterator {
+    /// Reference to PredicateVector instance
+    PredicateVector const &vec_;
+
+    /// Index into PredicateVector
+    int bit_;
+
+   public:
+    /// Copy constructor
+    CUTLASS_HOST_DEVICE
+    ConstIterator(ConstIterator const &it) : vec_(it.vec_), bit_(it.bit_) {}
+
+    ///
+    CUTLASS_HOST_DEVICE
+    ConstIterator(PredicateVector const &_vec, int _start = 0) : vec_(_vec), bit_(_start) {}
+
+    /// Pre-increment
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator++() {
+      ++bit_;
+      return *this;
+    }
+
+    /// Pre-decrement
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator--() {
+      --bit_;
+      return *this;
+    }
+
+    /// Post-increment
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator++(int) {
+      ConstIterator ret(*this);
+      ret.bit_++;
+      return ret;
+    }
+
+    /// Post-decrement
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator--(int) {
+      ConstIterator ret(*this);
+      ret.bit_--;
+      return ret;
+    }
+
+    /// Returns true if iterators point to the same bit
+    CUTLASS_HOST_DEVICE
+    bool operator==(ConstIterator const &it) const { return bit_ == it.bit_; }
+
+    /// Returns false if iterators point to the same bit
+    CUTLASS_HOST_DEVICE
+    bool operator!=(ConstIterator const &it) const { return bit_ != it.bit_; }
+
+    /// Dereferences iterator
+    CUTLASS_HOST_DEVICE
+    bool operator*() const { return vec_[bit_]; }
+  };
+
+  /**
+  * @brief An iterator implementing \ref predicate_iterator_concept enabling sequential
+  * read and write access to predicates.
+  * @concept{predicate_iterator_concept}
+  */
+  class Iterator {
+    /// Reference to PredicateVector instance
+    PredicateVector &vec_;
+
+    /// Index into PredicateVector
+    int bit_;
+
+   public:
+    /// Copy constructor
+    CUTLASS_HOST_DEVICE
+    Iterator(Iterator const &it) : vec_(it.vec_), bit_(it.bit_) {}
+
+    /// Constructs an iterator from a PredicateVector
+    CUTLASS_HOST_DEVICE
+    Iterator(PredicateVector &_vec, int _start = 0) : vec_(_vec), bit_(_start) {}
+
+    /// Pre-increment
+    CUTLASS_HOST_DEVICE
+    Iterator &operator++() {
+      ++bit_;
+      return *this;
+    }
+
+    /// Pre-decrement
+    CUTLASS_HOST_DEVICE
+    Iterator &operator--() {
+      --bit_;
+      return *this;
+    }
+
+    /// Post-increment
+    CUTLASS_HOST_DEVICE
+    Iterator operator++(int) {
+      Iterator ret(*this);
+      ret.bit_++;
+      return ret;
+    }
+
+    /// Post-decrement
+    CUTLASS_HOST_DEVICE
+    Iterator operator--(int) {
+      Iterator ret(*this);
+      ret.bit_--;
+      return ret;
+    }
+
+    /// Returns true if iterators point to the same bit
+    CUTLASS_HOST_DEVICE
+    bool operator==(Iterator const &it) const { return bit_ == it.bit_; }
+
+    /// Returns false if iterators point to the same bit
+    CUTLASS_HOST_DEVICE
+    bool operator!=(Iterator const &it) const { return bit_ != it.bit_; }
+
+    /// Gets the bit at the pointed to location
+    CUTLASS_HOST_DEVICE
+    bool get() { return vec_[bit_]; }
+
+    /// Dereferences iterator
+    CUTLASS_HOST_DEVICE
+    bool operator*() const { return vec_[bit_]; }
+
+    /// Sets the bit at the pointed to location
+    CUTLASS_HOST_DEVICE
+    void set(bool value = true) { vec_.set(bit_, value); }
+  };
+
+  /// Iterator that always returns true
+  struct TrivialIterator {
+    /// Constructor
+    CUTLASS_HOST_DEVICE
+    TrivialIterator() {}
+
+    /// Copy constructor
+    CUTLASS_HOST_DEVICE
+    TrivialIterator(Iterator const &it) {}
+
+    /// Constructs an iterator from a PredicateVector
+    CUTLASS_HOST_DEVICE
+    TrivialIterator(PredicateVector const &_vec) {}
+
+    /// Pre-increment
+    CUTLASS_HOST_DEVICE
+    TrivialIterator &operator++() { return *this; }
+
+    /// Post-increment
+    CUTLASS_HOST_DEVICE
+    TrivialIterator operator++(int) { return *this; }
+
+    /// Dereferences iterator
+    CUTLASS_HOST_DEVICE
+    bool operator*() const { return true; }
+  };
+
+ public:
+  //
+  // Methods
+  //
+
+  /// Initialize the predicate vector
+  CUTLASS_HOST_DEVICE PredicateVector(bool value = true) { fill(value); }
+
+  /// Fills all predicates with a given value
+  CUTLASS_HOST_DEVICE void fill(bool value = true) {
+    Storage item = (value ? ~Storage(0) : Storage(0));
+
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < kWordCount; ++i) {
+      storage(i) = item;
+    }
+  }
+
+  /// Accesses a bit within the predicate vector.
+  CUTLASS_HOST_DEVICE bool operator[](int idx) const { return at(idx); }
+
+  /// Accesses a bit within the predicate vector.
+  CUTLASS_HOST_DEVICE bool at(int idx) const {
+    int bit, word;
+    computeStorageOffset(word, bit, idx);
+
+    return ((storage(word) >> bit) & 1);
+  }
+
+  /// Set a bit within the predicate vector.
+  CUTLASS_HOST_DEVICE void set(int idx, bool value = true) {
+    int bit, word;
+    computeStorageOffset(word, bit, idx);
+
+    Storage disable_mask = (~(Storage(1) << bit));
+    Storage enable_mask = (Storage(value) << bit);
+
+    storage(word) = ((storage(word) & disable_mask) | enable_mask);
+  }
+
+  /// Computes the intersection of two identical predicate vectors.
+  CUTLASS_HOST_DEVICE PredicateVector &operator&=(PredicateVector const &predicates) {
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < kWordCount; ++i) {
+      storage(i) = (storage(i) & predicates.storage(i));
+    }
+    return *this;
+  }
+
+  /// Computes the union of two identical predicate vectors.
+  CUTLASS_HOST_DEVICE PredicateVector &operator|=(PredicateVector const &predicates) {
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < kWordCount; ++i) {
+      storage(i) = (storage(i) | predicates.storage(i));
+    }
+    return *this;
+  }
+
+  /// Returns true if entire predicate array is zero.
+  CUTLASS_HOST_DEVICE bool is_zero() const {
+    Storage mask(0);
+    for (int byte = 0; byte < sizeof(Storage); ++byte) {
+      Storage byte_mask = (((1 << kPredicatesPerByte) - 1) << kPredicateStart);
+      mask |= (byte_mask << (byte * 8));
+    }
+    uint32_t result = 0;
+    for (int word = 0; word < kWordCount; ++word) {
+      result |= storage(word);
+    }
+    return result == 0;
+  }
+
+  /// Returns an iterator to the start of the bit vector
+  CUTLASS_DEVICE
+  Iterator begin() { return Iterator(*this); }
+
+  /// Returns an iterator
+  CUTLASS_DEVICE
+  Iterator end() { return Iterator(*this, kPredicates); }
+
+  /// Returns a ConstIterator
+  CUTLASS_DEVICE
+  ConstIterator const_begin() const { return ConstIterator(*this); }
+
+  /// Returns a ConstIterator
+  CUTLASS_DEVICE
+  ConstIterator const_end() const { return ConstIterator(*this, kPredicates); }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Always returns true predicate.
+struct TrivialPredicateTileAdapter {
+  /// Ctor.
+  CUTLASS_HOST_DEVICE TrivialPredicateTileAdapter() {}
+
+  /// The value at location (d, h, w, c).
+  CUTLASS_HOST_DEVICE bool at(int, int, int, int) const { return true; }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Adapter to enable random access to predicates via logical coordinate within a tile.
+template <typename PredicateVector_, typename Iterations_>
+struct PredicateTileAdapter {
+  /// The vector of predicates.
+  typedef PredicateVector_ PredicateVector;
+  /// The iterations.
+  typedef Iterations_ Iterations;
+
+ private:
+  /// The predicates.
+  PredicateVector &predicates;
+
+ public:
+  /// Ctor.
+  CUTLASS_DEVICE PredicateTileAdapter(PredicateVector &predicates_) : predicates(predicates_) {}
+
+  /// Get the value at location (d, h, w, c).
+  CUTLASS_DEVICE bool at(int d, int h, int w, int c) const {
+    int const bit = ComputeOffsetFromShape<Iterations>::get(d, h, w, c);
+    return predicates.at(bit);
+  }
+
+  /// Set the value at location (d, h, w, c).
+  CUTLASS_DEVICE void set(int d, int h, int w, int c, bool value) {
+    int const bit = ComputeOffsetFromShape<Iterations>::get(d, h, w, c);
+    predicates.set(bit, value);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Adapter to enable random access to predicates via logical coordinate within a tile.
+template <typename PredicateVector_, typename Iterations_>
+struct ConstPredicateTileAdapter {
+  /// The vector of predicates.
+  typedef PredicateVector_ PredicateVector;
+  /// The iterations.
+  typedef Iterations_ Iterations;
+
+ private:
+  /// The predicates.
+  PredicateVector const &predicates;
+
+ public:
+  /// Ctor.
+  CUTLASS_DEVICE ConstPredicateTileAdapter(PredicateVector const &predicates_)
+      : predicates(predicates_) {}
+
+  /// Get the value at location (d, h, w, c).
+  CUTLASS_DEVICE bool at(int d, int h, int w, int c) const {
+    int const bit = ComputeOffsetFromShape<Iterations>::get(d, h, w, c);
+    return predicates.at(bit);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/reshape_tile.h

@@ -0,0 +1,58 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a type for restructuring a tile.
+*/
+#pragma once
+
+#include "cutlass/shape.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// The following functor reshapes a tile of data. The goal is to have at least kAccessSize in
+// the inner-most dimension. If the user respects that constraint, there is nothing to be done. If
+// that's not the case, this functor will correct that and "extract" the right number of elements
+// from the next dimension.
+
+template <typename Tile_, int kAccessSize_, bool = (Tile_::kC < kAccessSize_)>
+struct ReshapeTile {
+  typedef Tile_ Tile;
+};
+
+template <typename Tile_, int kAccessSize_>
+struct ReshapeTile<Tile_, kAccessSize_, true> {
+  // Make sure the W dimension of the tile is large enough.
+  static_assert(Tile_::kW >= kAccessSize_, "The W dimension is too small");
+  // Make sure the dimension can be divided by the number of scalars.
+  static_assert(Tile_::kW % kAccessSize_ == 0, "Not supported");
+  // Collapse the W dimension.
+  typedef Shape<Tile_::kD, Tile_::kH, Tile_::kW / kAccessSize_, kAccessSize_> Tile;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/shape.h

@@ -0,0 +1,262 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines Shape implementing the Layout concept for representing a 4D hypercube of objects.
+*/
+#pragma once
+
+#include "cutlass/cutlass.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*!@defgroup layout_concept Layout Concept
+* @{
+* @par Implementations of \ref layout_concept are used to describe a cube with DxHxW elements and C
+scalars per element.
+ A HxW slice of a cube is called an image and a cube consists of D images.
+*
+* @par Notations
+*   Let Layout be an implementation of the \ref layout_concept.
+*
+* @par Valid Expressions
+* - <b>Layout::D</b> specifies the depth of a cube
+* - <b>Layout::H</b> specifies the height of a cube
+* - <b>Layout::W</b> specifies the height of a cube
+* - <b>Layout::C</b> specifies the number of channels of each element in a cube
+* - <b>Layout::W_c</b> specifies the number of scalars of each row in one image of a cube.
+* - <b>Layout::H_w</b> specifies the number of elements in an image slice.
+* - <b>Layout::H_w_c</b>_specifies the number of scalars in an image slice.
+* - <b>Layout::D_h_w</b> specifies the number of elements in a cube.
+* - <b>Layout::D_h_w_c</b> specifies the number of scalars in a cube.
+* - <b>Layout::Strides</b> is a \ref layout_concept specifying the strides.
+* @}
+*/
+
+/**
+* @brief A Shape implementing \ref layout_concept describing the dimensions of a cube.
+* @concept{layout_concept}
+*/
+template <int kD_ = 1, int kH_ = 1, int kW_ = 1, int kC_ = 1>
+struct Shape {
+  /// The depth of the cube.
+  static int const kD = kD_;
+  /// The height of the cube.
+  static int const kH = kH_;
+  /// The width of the cube.
+  static int const kW = kW_;
+  /// The number of scalars per element.
+  static int const kC = kC_;
+};
+
+/**
+* @brief Compute derived counted of a \ref layout_concept based class
+*/
+template <typename Shape>
+struct ShapeCount {
+  /// The number of elements per row.
+  static int const kWc = Shape::kW * Shape::kC;
+  /// The number of pixels per image.
+  static int const kHw = Shape::kH * Shape::kW;
+  /// The number of elements per image.
+  static int const kHwc = Shape::kH * kWc;
+  /// The number of pixels per cube.
+  static int const kDhw = Shape::kD * kHw;
+  /// The number of elements in the 4D space.
+  static int const kDhwc = Shape::kD * kHwc;
+  /// The number of elements in the 4D space.
+  static int const kCount = kDhwc;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, int kScale_>
+struct ShapeScale {
+  typedef Shape<A_::kD * kScale_, A_::kH * kScale_, A_::kW * kScale_, A_::kC * kScale_> Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeAdd {
+  typedef Shape<A_::kD + B_::kD, A_::kH + B_::kH, A_::kW + B_::kW, A_::kC + B_::kC> Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeSub {
+  typedef Shape<A_::kD - B_::kD, A_::kH - B_::kH, A_::kW - B_::kW, A_::kC - B_::kC> Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeMul {
+  typedef Shape<A_::kD * B_::kD, A_::kH * B_::kH, A_::kW * B_::kW, A_::kC * B_::kC> Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeDiv {
+  typedef Shape<A_::kD / B_::kD, A_::kH / B_::kH, A_::kW / B_::kW, A_::kC / B_::kC> Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeDivCeiling {
+  typedef Shape<(A_::kD + B_::kD - 1) / B_::kD,
+                (A_::kH + B_::kH - 1) / B_::kH,
+                (A_::kW + B_::kW - 1) / B_::kW,
+                (A_::kC + B_::kC - 1) / B_::kC>
+      Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeMax {
+  typedef Shape<(A_::kD > B_::kD ? A_::kD : B_::kD),
+                (A_::kH > B_::kH ? A_::kH : B_::kH),
+                (A_::kW > B_::kW ? A_::kW : B_::kW),
+                (A_::kC > B_::kC ? A_::kC : B_::kC)>
+      Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename A_, typename B_>
+struct ShapeMin {
+  typedef Shape<(A_::kD < B_::kD ? A_::kD : B_::kD),
+                (A_::kH < B_::kH ? A_::kH : B_::kH),
+                (A_::kW < B_::kW ? A_::kW : B_::kW),
+                (A_::kC < B_::kC ? A_::kC : B_::kC)>
+      Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Shape_, int elementsPerAccess>
+struct ShapeStrides {
+  typedef Shape<Shape_::kH * Shape_::kW * Shape_::kC,
+                Shape_::kW * Shape_::kC,
+                Shape_::kC,
+                elementsPerAccess>
+      Shape;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief Compute the offset for the given coordinates in a cube
+* @tparam A \ref layout_concept where each dimension of the cube specifies the corresponding stride.
+*/
+template <typename Shape_>
+struct ComputeOffsetFromShape {
+  static CUTLASS_HOST_DEVICE int get(int d, int h, int w, int c) {
+    // clang-format off
+    return d * Shape_::kH * Shape_::kW * Shape_::kC +
+           h * Shape_::kW * Shape_::kC +
+           w * Shape_::kC +
+           c;
+    // clang-format on
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief Compute the offset for the given coordinates in a cube
+* @tparam A \ref layout_concept where each dimension of the cube specifies the corresponding stride.
+*/
+template <typename Strides_>
+struct ComputeOffsetFromStrides {
+  static CUTLASS_HOST_DEVICE int get(int d, int h, int w, int c) {
+    return d * Strides_::kD + h * Strides_::kH + w * Strides_::kW + c * Strides_::kC;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief Decompose threadId.x into coordinate of a cube whose dimensions are specified by Threads_.
+* Afterwards compute the offset of those coordinates using Strides_
+* @tparam Threads_ The dimension of the cube the threadIdx.x value is mapped on
+* @tparam Strides_ The strides to use when compute the offsets based on the coordinates of the cube.
+*/
+template <typename Threads_, typename Strides_>
+struct ComputeThreadOffsetFromStrides {
+  static CUTLASS_DEVICE int get() {
+    // Decompose the thread index.
+    int c = threadIdx.x % Threads_::kC;
+    int w = threadIdx.x / Threads_::kC % Threads_::kW;
+    int h = threadIdx.x / Threads_::kC / Threads_::kW % Threads_::kH;
+    int d = threadIdx.x / Threads_::kC / Threads_::kW / Threads_::kH;
+
+    // Compute the offset.
+    return d * Strides_::kD + h * Strides_::kH + w * Strides_::kW + c * Strides_::kC;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+/**
+*@brief Specialization for D=1
+*/
+template <int T_h_, int T_w_, int T_c_, int S_h_, int S_w_, int S_c_>
+struct ComputeThreadOffsetFromStrides<Shape<1, T_h_, T_w_, T_c_>, Shape<1, S_h_, S_w_, S_c_> > {
+  static CUTLASS_DEVICE int get() {
+    // Decompose the thread index.
+    int c = threadIdx.x % T_c_;
+    int w = threadIdx.x / T_c_ % T_w_;
+    int h = threadIdx.x / T_c_ / T_w_ % T_h_;
+
+    // Compute the offset.
+    return h * S_h_ + w * S_w_ + c * S_c_;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+*@brief Specialization for D=1 and C=1
+*/
+template <int T_h_, int T_w_, int S_h_, int S_w_>
+struct ComputeThreadOffsetFromStrides<Shape<1, T_h_, T_w_, 1>, Shape<1, S_h_, S_w_, 1> > {
+  static CUTLASS_DEVICE int get() {
+    // Decompose the thread index.
+    int w = threadIdx.x % T_w_;
+    int h = threadIdx.x / T_w_;
+
+    // Compute the offset.
+    return h * S_h_ + w * S_w_;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/tensor_ref.h

@@ -0,0 +1,639 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a structure containing strides, bounds, and a pointer to tensor data.
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/cutlass.h"
+#include "cutlass/vector.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Default mapping function from coordinates in a tensor's index space into the n-D array held
+/// in memory. Assumes StorageRank = Rank
+template <int Rank>
+struct IdentityTensorMapFunc {
+  static int const kStorageRank = Rank;
+  CUTLASS_HOST_DEVICE
+  Coord<Rank> operator()(Coord<Rank> const &coord) const {
+    return coord;
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/* \brief Structure modeling a pointer and stride into a tensor.
+
+  A tensor consists of an index space with Rank_ dimensions. It is stored in memory modeled
+  as an n-D array, where n = StorageRank_. A mapping function maps the logical coordinates of the
+  tensor's index space into the n-D array, and a stride vector maps the n-D array to linear memory.
+
+  CUTLASS requires the n-D array's least significant, "fastest changing" dimension to
+  be contiguous in memory. It therefore has a stride of 1 and is not stored. Construction is offered
+  from vectors of full StorageRank and of the 'compact' rank, though it is in error to construct
+  with the least significant stride != 1.
+
+  The requirement that the least significant dimension be consecutive enables numerous optimizations
+  and assumptions about vectorizing memory accesses throughout CUTLASS. It also matches various
+  BLAS conventions in which only the "leading dimension" or most significant stride of a rank=2
+  matrix is provided.
+
+  This does affect the ability of constructing arbitrary "sparse" 2-D matrices in memory where all
+  stride elements are > 1. This can be overcome by defining a custom mapping function and a
+  StorageRank of 3 or more.
+
+
+  Examples:
+
+  (These examples use helpers for matrix layouts defined in cutlass/matrix_traits.h)
+
+  1. Column-major matrix may be represented as a rank=2 tensor:
+
+    TensorRef<float, 2, MatrixLayout::ColumnMajor> A(ptr_A, make_Coord(ldm, 1));
+
+  2. Row-major matrix may be represented as a rank=2 tensor:
+
+    TensorRef<float, 2, MatrixLayout::RowMajor> B(ptr_A, ldm);
+
+  3. An interleaved matrix may be represented as a rank=2 tensor:
+
+    TensorRef<int8_t, 2, MatrixLayout::ColumnMajorInterleaved<32> > C;
+
+  4. Defining a sparse matrix with arbitrary strides in each dimension
+
+    struct ContiguousLayout {
+
+      /// Arbitrary storage rank
+      static int const kStorageRank = 3;
+
+      /// Mapping function defined by runtime stride configuration
+      CUTLASS_HOST_DEVICE
+      Coord<3> operator()(MatrixCoord const &coord) const {
+          return make_Coord(coord.row(), coord.column(), 0);
+      }
+    };
+
+    typedef TensorRef<float, 2, ContiguousLayout> ContiguousTensorRef;
+
+    // Construct the TensorRef object from a pair of stride values
+    ContiguousTensorRef D(ptr_D, make_Coord(row_stride, column_stride));
+
+
+  5. A helper exists to define a TensorRef for a contiguous matrix whose layout
+     is not known at compile time.
+
+    MatrixLayout::Kind layout;   // Could be MatrixLayout::kRowMajor or MatrixLayout::kColumnMajor
+    int ldm;                     // leading dimension
+
+    ContiguousTensorRef E(ptr_E, ContiguousLayout::stride(layout, ldm));
+
+*/
+template <
+  /// Data type of element stored within tensor
+  typename Storage_,
+  /// Rank of logical tensor
+  int Rank_,
+  /// Maps a Coord<Rank_> in the logical tensor index space to the internal n-D array
+  typename MapFunc_ = IdentityTensorMapFunc<Rank_>,
+  /// Rank of internal n-D array
+  int StorageRank_ = MapFunc_::kStorageRank,
+  /// Index type used for coordinates
+  typename Index_ = int,
+  /// Index type used for offsets and pointer differences
+  typename LongIndex_ = long long
+>
+class TensorRef {
+ public:
+  /// Data type of individual access
+  typedef Storage_ Storage;
+
+  /// Logical rank of tensor index space
+  static int const kRank = Rank_;
+
+  /// Mapping function from logical coordinate to internal n-D array
+  typedef MapFunc_ MapFunc;
+
+  /// Rank of internal storage
+  static int const kStorageRank = StorageRank_;
+
+  /// Index type
+  typedef Index_ Index;
+
+  /// Typically, strides in memory can be very large
+  typedef LongIndex_ LongIndex;
+
+  /// Coordinate in logical tensor space
+  typedef Coord<kRank> TensorCoord;
+
+  /// Coordinate in storage n-D array
+  typedef Coord<kStorageRank> StorageCoord;
+
+  /// Stride vector in storage coordinage space - assumes least significant stride
+  /// is 1 and does not store it.
+  typedef Coord<kStorageRank - 1> StrideVector;
+
+  /// Tensor reference to of constant value
+  typedef TensorRef<
+    typename platform::remove_const<Storage>::type const,
+    Rank_,
+    MapFunc_,
+    StorageRank_,
+    Index_,
+    LongIndex_> ConstTensorRef;
+
+  /// Require at least rank=1. Mathematically, a rank=0 tensor would be considered to be a
+  /// scalar, but degenerate cases such as these are difficult to accommodate without
+  /// extensive C++ metaprogramming or support for zero-length arrays.
+  static_assert(kRank > 0, "Cannot define a zero-rank TensorRef");
+
+  //
+  // Definitions included for backwards compatibility - to be removed in next major release
+  //
+
+  /// Coordinate in logical tensor space
+  typedef TensorCoord Coord_t;
+
+  /// Logical rank of tensor index space
+  static int const Rank = kRank;
+
+ private:
+
+  /// Pointer
+  Storage* ptr_;
+
+  /// Stride vector - fastest-changing stride assumed to be 1 and not stored
+  StrideVector stride_;
+
+  /// Maps a logical coordinate to an n-D array's tensor space
+  MapFunc coord_map_;
+
+ public:
+
+  //
+  // Methods
+  //
+
+  /// Helper for 1-D memory. All higher ranks are projected onto the fastest changing rank.
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage *ptr = nullptr): ptr_(ptr) {
+    for (int i = 0; i < kStorageRank - 1; ++i) {
+      stride_[i] = 1;
+    }
+  }
+
+  /// Helper to construct from a pointer and single stride element for 2-D pitch linear memory.
+  // Higher ranks are projected onto the fastest-changing rank.
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage* ptr, Index ldm) {
+    ptr_ = ptr;
+    for (int i = 0; i < kStorageRank - 1; ++i) {
+      stride_[i] = ldm;
+    }
+  }
+
+  /// Constructs from a single pointer and stride vector
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage* ptr, StrideVector const& stride) : ptr_(ptr), stride_(stride) {
+
+  }
+
+  /// Constructs from a pointer and a stride vector of size kRank. If fastest changing
+  /// stride is not 1, construction fails and subsequent calls to good() will return false.
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage* ptr, StorageCoord const& stride) {
+    // Fastest-changing stride must be one
+    if (stride.at(kStorageRank - 1) == 1) {
+      ptr_ = ptr;
+      for (int i = 0; i < kStorageRank - 1; ++i) {
+        stride_[i] = stride[i];
+      }
+    }
+    else {
+      // Fastest-chaning stride must be 1.
+      reset();
+    }
+  }
+
+  /// Enables conversion from TensorRef of non-const type
+  CUTLASS_HOST_DEVICE
+  TensorRef(
+    TensorRef<
+      typename platform::remove_const<Storage>::type,
+      kRank,
+      MapFunc,
+      kStorageRank,
+      Index,
+      LongIndex> const &ref
+  ):
+    ptr_(ref.data()) {
+    for (int i = 0; i < kStorageRank - 1; ++i) {
+      stride_[i] = ref.stride(i);
+    }
+  }
+
+  /// Returns a reference to constant-valued tensor
+  CUTLASS_HOST_DEVICE
+  ConstTensorRef const_ref() const {
+    return ConstTensorRef(*this);
+  }
+
+  /// Updates only the pointer
+  CUTLASS_HOST_DEVICE
+  void reset(Storage* ptr = nullptr) {
+    ptr_ = ptr;
+  }
+
+  /// Updates the pointer, stride, and location within a TensorRef
+  CUTLASS_HOST_DEVICE
+  void reset(Storage* ptr, StorageCoord const & stride) {
+    // Fastest-changing stride must be one
+    if (stride.at(kStorageRank - 1) == 1) {
+      ptr_ = ptr;
+      for (int i = 0; i < kStorageRank - 1; ++i) {
+        stride_[i] = stride[i];
+      }
+    }
+    else {
+      // Fastest-changing stride must be 1 - this is an error.
+      reset();
+    }
+  }
+
+  /// Returns true if the TensorRef may be safely accessed
+  CUTLASS_HOST_DEVICE
+  bool good() const {
+    return ptr_ != nullptr;
+  }
+
+  /// Returns the pointer to referenced data
+  CUTLASS_HOST_DEVICE
+  Storage * data() const { return ptr_; }
+
+  /// Returns the stride of the tensor
+  CUTLASS_HOST_DEVICE
+  StorageCoord stride() const {
+    StorageCoord ld;
+    for (int i = 0; i < kStorageRank - 1; ++i) {
+      ld[i] = stride_[i];
+    }
+    ld[kStorageRank - 1] = 1;
+    return ld;
+  }
+
+  /// Returns the stride of the tensor in the given dimension
+  CUTLASS_HOST_DEVICE
+  Index stride(int dim) const {
+    // fastest-changing stride assumbed to be 1
+    if (dim + 1 >= kStorageRank) {
+      return 1;
+    }
+    return stride_.at(dim);
+  }
+
+  /// Returns the maximum stride element as the 'leading dimension'
+  CUTLASS_HOST_DEVICE
+  Index leading_dim(int idx = 0) const { return stride(idx); }
+
+  /// Maps a logical coordinate to an n-D array in memory
+  CUTLASS_HOST_DEVICE
+  StorageCoord map(TensorCoord const &coord) const {
+    return coord_map_(coord);
+  }
+
+  /// Computes the offset of an index from the origin of the tensor
+  CUTLASS_HOST_DEVICE
+  LongIndex offset(TensorCoord const& coord) const {
+    return stride().template dot<LongIndex>(map(coord));
+  }
+
+  /// Returns a reference to the element at a given Coord
+  CUTLASS_HOST_DEVICE
+  Storage& at(TensorCoord const& coord) const {
+    return ptr_[offset(coord)];
+  }
+
+  /// Returns a reference to the element at a given linear index
+  CUTLASS_HOST_DEVICE
+  Storage& at(LongIndex idx) const { return ptr_[idx]; }
+
+  /// Returns a reference to the element at a given Coord
+  CUTLASS_HOST_DEVICE
+  Storage& operator[](TensorCoord const& coord) const {
+    return ptr_[offset(coord)];
+  }
+
+  /// Returns a reference to the element at a given linear index
+  CUTLASS_HOST_DEVICE
+  Storage& operator[](LongIndex idx) const { return ptr_[idx]; }
+
+  /// Adds an offset to each pointer
+  CUTLASS_HOST_DEVICE
+  TensorRef & add_pointer_offset(LongIndex delta) {
+    ptr_ += delta;
+    return *this;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef operator+(TensorCoord const& b) const {
+    TensorRef result(*this);
+    result.add_pointer_offset(offset(b));
+    return result;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef& operator+=(TensorCoord const& b) {
+    add_pointer_offset(offset(b));
+    return *this;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef operator-(TensorCoord const& b) const {
+    TensorRef result(*this);
+    result.add_pointer_offset(-offset(b));
+    return result;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef& operator-=(TensorCoord const& b) {
+    add_pointer_offset(-offset(b));
+    return *this;
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+//
+// Partial specializations to handle degenerate cases.
+//
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Specialization for rank=1 case with no internal StrideVector
+template <
+  /// Data type of element stored within tensor
+  typename Storage_,
+  /// Rank of logical tensor
+  int Rank_,
+  /// Maps a Coord<Rank_> in the logical tensor index space to the internal n-D array
+  typename MapFunc_,
+  /// Index type used for coordinates
+  typename Index_,
+  /// Index type used for offsets and pointer differences
+  typename LongIndex_
+>
+class TensorRef<Storage_, Rank_, MapFunc_, 1, Index_, LongIndex_> {
+ public:
+  /// Data type of individual access
+  typedef Storage_ Storage;
+
+  /// Logical rank of tensor index space
+  static int const kRank = Rank_;
+
+  /// Mapping function from logical coordinate to internal n-D array
+  typedef MapFunc_ MapFunc;
+
+  /// Rank of internal storage
+  static int const kStorageRank = 1;
+
+  /// Index type
+  typedef Index_ Index;
+
+  /// Typically, strides in memory can be very large
+  typedef LongIndex_ LongIndex;
+
+  /// Coordinate in logical tensor space
+  typedef Coord<kRank> TensorCoord;
+
+  /// Coordinate in storage n-D array
+  typedef Coord<kStorageRank> StorageCoord;
+
+  /// Stride vector in storage coordinage space - assumes least significant stride
+  /// is 1 and does not store it.
+  struct StrideVector { };
+
+  /// Tensor reference to of constant value
+  typedef TensorRef<
+    typename platform::remove_const<Storage>::type const,
+    Rank_,
+    MapFunc_,
+    kStorageRank,
+    Index_,
+    LongIndex_> ConstTensorRef;
+
+  //
+  // Definitions included for backwards compatibility - to be removed in next major release
+  //
+
+  /// Coordinate in logical tensor space
+  typedef TensorCoord Coord_t;
+
+  /// Logical rank of tensor index space
+  static int const Rank = kRank;
+
+ private:
+
+  /// Pointer
+  Storage* ptr_;
+
+  /// Maps a logical coordinate to an n-D array's tensor space
+  MapFunc coord_map_;
+
+ public:
+
+  //
+  // Methods
+  //
+
+  /// Helper for 1-D memory. All higher ranks are projected onto the fastest changing rank.
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage *ptr = nullptr): ptr_(ptr) { }
+
+  /// Constructs from a single pointer and stride vector
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage* ptr, StrideVector const& stride) : ptr_(ptr) {
+
+  }
+
+  /// Constructs from a pointer and a stride vector of size kRank. If fastest changing
+  /// stride is not 1, construction fails and subsequent calls to good() will return false.
+  CUTLASS_HOST_DEVICE
+  TensorRef(Storage* ptr, StorageCoord const& stride) {
+    // Fastest-changing stride must be one
+    if (stride.at(kStorageRank - 1) == 1) {
+      ptr_ = ptr;
+    }
+    else {
+      // Fastest-chaning stride must be 1.
+      reset();
+    }
+  }
+
+  /// Enables conversion from TensorRef of non-const type
+  CUTLASS_HOST_DEVICE
+  TensorRef(
+    TensorRef<
+      typename platform::remove_const<Storage>::type,
+      kRank,
+      MapFunc,
+      kStorageRank,
+      Index,
+      LongIndex> const &ref
+  ):
+    ptr_(ref.data()) {
+  }
+
+  /// Returns a reference to constant-valued tensor
+  CUTLASS_HOST_DEVICE
+  ConstTensorRef const_ref() const {
+    return ConstTensorRef(*this);
+  }
+
+  /// Updates only the pointer
+  CUTLASS_HOST_DEVICE
+  void reset(Storage* ptr = nullptr) {
+    ptr_ = ptr;
+  }
+
+  /// Updates the pointer, stride, and location within a TensorRef
+  CUTLASS_HOST_DEVICE
+  void reset(Storage* ptr, StorageCoord const & stride) {
+    // Fastest-changing stride must be one
+    if (stride.at(kStorageRank - 1) == 1) {
+      ptr_ = ptr;
+    }
+    else {
+      // Fastest-changing stride must be 1 - this is an error.
+      reset();
+    }
+  }
+
+  /// Returns true if the TensorRef may be safely accessed
+  CUTLASS_HOST_DEVICE
+  bool good() const {
+    return ptr_ != nullptr;
+  }
+
+  /// Returns the pointer to referenced data
+  CUTLASS_HOST_DEVICE
+  Storage * data() const { return ptr_; }
+
+  /// Returns the stride of the tensor
+  CUTLASS_HOST_DEVICE
+  StorageCoord stride() const {
+    StorageCoord ld;
+    ld[kStorageRank - 1] = 1;
+    return ld;
+  }
+
+  /// Returns the stride of the tensor in the given dimension
+  CUTLASS_HOST_DEVICE
+  Index stride(int dim) const {
+    // fastest-changing stride assumbed to be 1
+    return 1;
+  }
+
+  /// Returns the maximum stride element as the 'leading dimension'
+  CUTLASS_HOST_DEVICE
+  Index leading_dim(int idx = 0) const { return 1; }
+
+  /// Maps a logical coordinate to an n-D array in memory
+  CUTLASS_HOST_DEVICE
+  StorageCoord map(TensorCoord const &coord) const {
+    return coord_map_(coord);
+  }
+
+  /// Computes the offset of an index from the origin of the tensor
+  CUTLASS_HOST_DEVICE
+  LongIndex offset(TensorCoord const& coord) const {
+    return stride().template dot<LongIndex>(map(coord));
+  }
+
+  /// Returns a reference to the element at a given Coord
+  CUTLASS_HOST_DEVICE
+  Storage& at(TensorCoord const& coord) const {
+    return ptr_[offset(coord)];
+  }
+
+  /// Returns a reference to the element at a given linear index
+  CUTLASS_HOST_DEVICE
+  Storage& at(LongIndex idx) const { return ptr_[idx]; }
+
+  /// Returns a reference to the element at a given Coord
+  CUTLASS_HOST_DEVICE
+  Storage& operator[](TensorCoord const& coord) const {
+    return ptr_[offset(coord)];
+  }
+
+  /// Returns a reference to the element at a given linear index
+  CUTLASS_HOST_DEVICE
+  Storage& operator[](LongIndex idx) const { return ptr_[idx]; }
+
+  /// Adds an offset to each pointer
+  CUTLASS_HOST_DEVICE
+  TensorRef & add_pointer_offset(LongIndex delta) {
+    ptr_ += delta;
+    return *this;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef operator+(TensorCoord const& b) const {
+    TensorRef result(*this);
+    result.add_pointer_offset(offset(b));
+    return result;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef& operator+=(TensorCoord const& b) {
+    add_pointer_offset(offset(b));
+    return *this;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef operator-(TensorCoord const& b) const {
+    TensorRef result(*this);
+    result.add_pointer_offset(-offset(b));
+    return result;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorRef& operator-=(TensorCoord const& b) {
+    add_pointer_offset(-offset(b));
+    return *this;
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/tensor_ref_collection.h

@@ -0,0 +1,420 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Introduces TensorRefCollection concept and defines TensorRefBatch and TensorRefArray. 
+*/
+
+#pragma once
+
+#include "cutlass/tensor_ref.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+//
+// TensorRefCollection is a concept for storing a logical collection of TensorRef objects. Classes
+// satisfying the TensorRefCollection concept must support the following:
+//
+//   // Define storage type
+//   typedef typename TensorRefCollection::Storage Storage;
+//
+//   // Define a type for offsets in memory
+//   typedef typename TensorRefCollection::LongIndex LongIndex;
+//
+//   // Define a ConstIterator type satisfying TensorRefIterator
+//   typedef typename TensorRefCollection::ConstIterator TensorRefIterator;
+//
+//   // Implement a begin() method.
+//   TensorRefIterator iterator = collection.begin();
+//
+//
+// TensorRefIterator is a concept for accessing an element in a TensorRefCollection. Classes
+// satisfying the TensorRefIterator concept must support the following:
+//
+//   // Define a TensorRef type accessed by the iterator
+//   typedef typename TensorRefIterator::TensorRef TensorRef;
+//
+//   // Access the TensorRef
+//   TensorRef ref = *iterator;
+//
+//   // Pre-increment and post-increment
+//   ++iterator;
+//   iterator++;
+//
+//   // Pre-decrement and post-decrement
+//   --iterator;
+//   iterator--;
+//
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// This satisfies TensorRefCollection and stores a collection of TensorRef objects that
+/// have identical strides. TensorRef objects are separated by a linear stride.
+template <
+  /// Data type of element stored within tensor
+  typename Storage_,
+  /// Rank of logical tensor
+  int Rank_,
+  /// Maps a Coord<Rank_> in the logical tensor index space to the internal n-D array
+  typename MapFunc_ = IdentityTensorMapFunc<Rank_>,
+  /// Rank of internal n-D array
+  int StorageRank_ = MapFunc_::kStorageRank,
+  /// Index type used for coordinates
+  typename Index_ = int,
+  /// Index type used for offsets and pointer differences
+  typename LongIndex_ = long long
+>
+struct TensorRefBatchStrided: 
+  public TensorRef<Storage_, Rank_, MapFunc_, StorageRank_, Index_, LongIndex_> {
+
+  //
+  // Type definitions
+  //
+
+  /// Underlying TensorRef type
+  typedef TensorRef<Storage_, Rank_, MapFunc_, StorageRank_, Index_, LongIndex_> Base;
+
+  /// Storage type
+  typedef typename Base::Storage Storage;
+
+  /// Index type
+  typedef Index_ Index;
+
+  /// Typically, strides in memory can be very large
+  typedef LongIndex_ LongIndex;
+
+  /// Coordinate in logical tensor space
+  typedef Coord<kRank> TensorCoord;
+
+  /// Tensor reference implied by the TensorRefBatchStrided
+  typedef Base TensorRef;
+
+  /// Constant iterator over tensors implied by TensorRefBatchStrided
+  class ConstIterator {
+  public:
+    /// TensorRef returned by the iterator
+    typedef Base TensorRef;
+
+  private:
+
+    /// Reference to the parent TensorBatchRef object
+    TensorRefBatchStrided const &ref_;
+
+    /// Offset from the base TensorRef pointer 
+    LongIndex offset_;
+
+  public:
+
+    /// Constructs a ConstIterator from a parent TensorRefBatchStrided
+    CUTLASS_HOST_DEVICE
+    ConstIterator(
+      TensorRefBatchStrided const &ref, 
+      LongIndex offset = 0): ref_(ref), offset_(offset) { }
+
+    /// Obtains a TensorRef pointed to by the iterator
+    CUTLASS_HOST_DEVICE
+    TensorRef *operator() const {
+      TensorRef ref(ref_);
+      ref.add_pointer_offset(offset_);
+      return ref;
+    }
+
+    /// Advances the iterator to point to the next tensor
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator++() {
+      offset_ += ref_.tensor_stride;
+      return *this;
+    }
+
+    /// Advances the iterator to point to the next tensor
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator++(int) {
+      ConstIterator ret(*this);
+      offset_ += ref_.tensor_stride;
+      return ret;
+    }
+
+    /// Returns an iterator advanced by (idx) amount
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator+(Index idx) {
+      return ConstIterator(ref, offset_ + ref_.tensor_stride * idx);
+    }
+
+    /// Advances this iterator by (idx) and returns a reference to self
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator+=(Index idx) {
+      offset_ += ref_.tensor_stride * idx;
+      return *this;
+    }
+
+    /// Moves to the previous tensor
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator--() {
+      offset_ -= ref_.tensor_stride;
+      return *this;
+    }
+
+    /// Moves to the previous tensor
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator--(int) {
+      ConstIterator ret(*this);
+      offset_ -= ref_.tensor_stride;
+      return ret;
+    }
+
+    /// Returns an iterator moved forward by (idx) amount
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator-(Index idx) {
+      return ConstIterator(ref_, offset_ - ref_.tensor_stride * idx);
+    }
+
+    /// Moves this iterator by (idx) and returns a reference to self
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator-=(Index idx) {
+      offset_ -= ref_.tensor_stride * idx;
+      return *this;
+    }
+
+    /// Returns the difference in offset between two iterators
+    CUTLASS_HOST_DEVICE
+    Stride operator-(ConstIterator const &it) {
+      return offset_ - it.offset_;
+    }
+  };
+
+  //
+  // Data members
+  //
+
+  /// Stride between tensors
+  LongIndex tensor_stride;
+
+  //
+  // Methods
+  //
+
+  // Default ctor
+  CUTLASS_HOST_DEVICE
+  TensorRefBatchStrided(): tensor_stride(0) { }
+
+  // Constructs form a tensor reference and 
+  CUTLASS_HOST_DEVICE
+  TensorRefBatchStrided(TensorRef const &ref, LongIndex _tensor_stride = 0): 
+    TensorRef(ref), 
+    tensor_stride(_tensor_stride) { }
+
+  /// Gets the pointer offset
+  CUTLASS_HOST_DEVICE
+  LongIndex get_pointer_offset(Index idx) const {
+    return idx * tensor_stride;
+  }
+
+  // Returns a reference
+  CUTLASS_HOST_DEVICE
+  TensorRef at(Index idx) const {
+    TensorRef ref(*this);
+    ref.add_pointer_offset(get_pointer_offset(idx));
+    return ref;
+  }
+
+  /// Returns an iterator
+  CUTLASS_HOST_DEVICE
+  ConstIterator begin() {
+    return ConstIterator(*this);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// This satisfies TensorRefCollection and stores a collection of TensorRef objects. This is a
+/// structure of arrays in that the individual members of the TensorRef are held in distinct arrays.
+///
+/// Note, TensorRef maps a logical coordinate space to an n-D array with rank kStorageRank. It
+/// maintains a stride vector of similar rank, but the least significant rank is defined to be 1.
+///
+/// The least significant stride of 1 is not stored, and therefore the number of stride arrays is 
+/// kStorageRank - 1.
+template <
+  /// Data type of element stored within tensor
+  typename Storage_,
+  /// Rank of logical tensor
+  int Rank_,
+  /// Maps a Coord<Rank_> in the logical tensor index space to the internal n-D array
+  typename MapFunc_ = IdentityTensorMapFunc<Rank_>,
+  /// Rank of internal n-D array
+  int StorageRank_ = MapFunc_::kStorageRank,
+  /// Index type used for coordinates
+  typename Index_ = int,
+  /// Index type used for offsets and pointer differences
+  typename LongIndex_ = long long
+>
+struct TensorRefArray {
+  //
+  // Type definitions
+  //
+
+  /// TensorRef type obtained from the TensorRefArray
+  typedef TensorRef<Storage_, Rank_, MapFunc_, StorageRank_, Index_, LongIndex_> TensorRef;
+
+  /// Element pointed to by the TensorRef
+  typedef Storage_ Storage;
+
+  /// Index type
+  typedef Index_ Index;
+
+  /// Typically, strides in memory can be very large
+  typedef LongIndex_ LongIndex;
+
+  /// Rank of the stride vector
+  static int const kStorageRank = TensorRef::kStorageRank;
+
+  /// TensorRefIterator over TensorRef objects in TensorRefArray 
+  class ConstIterator {
+  public:
+
+    /// TensorRef returned by the iterator
+    typedef Base TensorRef;
+
+  private:
+    /// Reference to the TensorRefArray
+    TensorRefArray const &ref_;
+
+    /// Index into TensorRefArray
+    int idx_;
+
+  public:
+
+    /// Constructs a ConstIterator over the TensorRef objects
+    CUTLASS_HOST_DEVICE
+    ConstIterator(TensorArrayRef const &ref, int idx = 0): ref_(ref), idx_(idx) { }
+
+    /// Obtains a TensorRef pointed to by this iterator
+    CUTLASS_HOST_DEVICE
+    TensorRef *operator() const {
+      return ref_.reference(idx_);
+    }
+
+    /// Advances to next TensorRef
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator++() {
+      ++idx_;
+      return *this;
+    }
+
+    /// Advances to next TensorRef
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator++(int) {
+      ConstIterator ret(*this);
+      idx_ ++;
+      return ret;
+    }
+
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator+(Index idx) {
+      return ConstIterator(ref_, idx_ + idx);
+    }
+
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator+=(Index idx) {
+      idx_ += idx;
+      return *this;
+    }
+
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator--() {
+      --idx_;
+      return *this;
+    }
+
+    /// Advances to next TensorRef
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator--(int) {
+      ConstIterator ret(*this);
+      --idx_;
+      return ret;
+    }
+
+    CUTLASS_HOST_DEVICE
+    ConstIterator &operator-=(Index idx) {
+      idx_ -= idx;
+      return *this;
+    }
+
+    CUTLASS_HOST_DEVICE
+    ConstIterator operator-(Index idx) {
+      return ConstIterator(ref_, idx_ + idx);
+    }
+  };
+
+  //
+  // Data members
+  //
+
+  /// Base addresses
+  Storage **pointers;
+
+  /// Array of strides
+  Index *strides[kStorageRank - 1];
+
+  //
+  // Methods
+  //
+
+  // Default ctor
+  CUTLASS_HOST_DEVICE
+  TensorArrayRef() { }
+
+  // Construct from pointers to arrays to strides
+  CUTLASS_HOST_DEVICE
+  TensorArrayRef(
+    Storage **_pointers,
+    Index _strides[kStorageRank - 1]): pointers(_pointers) { 
+
+    // Copy pointers to strides arrays
+    for (int i = 0; i < kStorageRank - 1; ++i) {
+      strides[i] = _strides[i];
+    }
+  }
+
+  // Returns a TensorRef at the given index in the collection
+  CUTLASS_HOST_DEVICE
+  TensorRef at(Index idx) const {
+    Coord<kStorageRank - 1, Index> stride;
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < kStorageRank - 1; ++i) {
+      stride[i] = stride_[idx][i];
+    }
+    return TensorRef(pointers[idx], stride);
+  }
+
+  /// Returns an TesnorRefIterator over the TensorRef objects in this collection
+  CUTLASS_HOST_DEVICE
+  ConstIterator begin() {
+    return ConstIterator(*this);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/tensor_view.h

@@ -0,0 +1,266 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a structure containing strides and a pointer to tensor data.
+
+    TensorView is derived from TensorRef and contributes bounds to the tensor's index space. Thus,
+    it is a complete mathematical object and may be used in tensor algorithms. It is decoupled from
+    data storage and is therefore lightweight and may be embedded in larger tensor objects or
+    memory structures.
+
+    See cutlass/tensor_ref.h for more details about the mapping of the logical tensor index space to
+    linear memory.
+*/
+
+#pragma once
+
+#include <cmath>
+
+#include "cutlass/cutlass.h"
+#include "cutlass/tensor_ref.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Defines a view into a logical tensor
+template <
+  /// Data type of element stored within tensor
+  typename Storage_,
+  /// Rank of logical tensor
+  int Rank_ = 4,
+  /// Maps a Coord<Rank_> in the logical tensor index space to the internal n-D array
+  typename MapFunc_ = IdentityTensorMapFunc<Rank_>,
+  /// Rank of internal n-D array
+  int StorageRank_ = MapFunc_::kStorageRank,
+  /// Index type used for coordinates
+  typename Index_ = int,
+  /// Index type used for offsets and pointer differences
+  typename LongIndex_ = long long
+>
+class TensorView : public TensorRef<Storage_, Rank_, MapFunc_, StorageRank_, Index_, LongIndex_> {
+ public:
+  /// Base tensor reference
+  typedef TensorRef<Storage_, Rank_, MapFunc_, StorageRank_, Index_, LongIndex_> Base;
+
+  /// Tensor reference to of constant value
+  typedef TensorRef<
+    typename platform::remove_const<Storage_>::type const,
+    Rank_,
+    MapFunc_,
+    StorageRank_,
+    Index_,
+    LongIndex_> ConstTensorRef;
+
+  /// Base tensor reference
+  typedef Base TensorRef;
+
+  /// Storage type
+  typedef typename Base::Storage Storage;
+
+  /// Index type
+  typedef typename Base::Index Index;
+
+  /// Coordinate in logical tensor space
+  typedef typename TensorRef::TensorCoord TensorCoord;
+
+  /// Coordinate in storage n-D array
+  typedef typename TensorRef::StorageCoord StorageCoord;
+
+  /// Stride vector in storage coordinate space
+  /// Least significant stride is = 1 and not stored
+  typedef typename TensorRef::StrideVector StrideVector;
+
+  /// TensorView of constant value
+  typedef TensorView<
+    typename platform::remove_const<Storage>::type const,
+    Rank_,
+    MapFunc_,
+    StorageRank_,
+    Index_,
+    LongIndex_> ConstTensorView;
+
+  //
+  // Definitions included for backwards compatibility - to be removed in next major release
+  //
+
+  /// Coordinate in logical tensor space
+  typedef TensorCoord Coord_t;
+
+  /// Logical rank of tensor index space
+  static int const Rank = Base::kRank;
+
+  /// Type used to compute the offset of an element to the base of a tensor
+  typedef typename Base::LongIndex Offset_t;
+
+  /// Base class
+  typedef TensorRef TensorRef_t;
+
+  /// TensorRef to const-valued type
+  typedef typename TensorRef::ConstTensorRef ConstTensorRef_t;
+
+ private:
+  //
+  // Data members
+  //
+
+  /// Dimensions of coordinate (independent of stride)
+  TensorCoord size_;
+
+ public:
+  //
+  // Device and Host Methods
+  //
+
+  /// Default constructor
+  CUTLASS_HOST_DEVICE
+  TensorView() {}
+
+  /// Constructs a TensorView from a TensorRef and size
+  CUTLASS_HOST_DEVICE
+  TensorView(Base const& _ref, TensorCoord const& _size) : Base(_ref), size_(_size) {}
+
+  /// Constructs a TensorView from a pointer, a stride vector, and size
+  CUTLASS_HOST_DEVICE
+  TensorView(
+    Storage *ptr,
+    StrideVector const &stride,
+    TensorCoord const& size
+  ):
+    Base(ptr, stride), size_(size) {}
+
+  /// Constructs a TensorView from a pointer, a stride vector, and size
+  CUTLASS_HOST_DEVICE
+  TensorView(
+    Storage *ptr,
+    StorageCoord const &stride,
+    TensorCoord const& size
+  ):
+    Base(ptr, stride), size_(size) {}
+
+  /// Updates the reference and size of a Tensor_view object
+  CUTLASS_HOST_DEVICE
+  void reset(Base const& _ref = Base(), TensorCoord const& _size = TensorCoord()) {
+    Base::operator=(_ref);
+    size_ = _size;
+  }
+
+  /// Accesses the size
+  CUTLASS_HOST_DEVICE
+  TensorCoord const& size() const { return size_; }
+
+  /// Accesses the size
+  CUTLASS_HOST_DEVICE
+  Index size(int dim) const { return size_.at(dim); }
+
+  /// Assigns the Tensor_view
+  CUTLASS_HOST_DEVICE
+  TensorView& operator=(TensorView const& _tensor) {
+    Base::operator=(_tensor);
+    size_ = _tensor.size_;
+    return *this;
+  }
+
+  /// Determines whether a location is within a tensor
+  CUTLASS_HOST_DEVICE
+  bool contains(TensorCoord const& coord) const {
+    CUTLASS_PRAGMA_UNROLL
+    for (int dim = 0; dim < Rank_; ++dim) {
+      if (coord[dim] >= size_[dim]) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  /// Returns a TensorRef pointing to the first element of the tensor.
+  CUTLASS_HOST_DEVICE
+  TensorRef ref() const {
+    return TensorRef(*this);
+  }
+
+  /// Returns a TensorRef pointing to the first element of the tensor.
+  CUTLASS_HOST_DEVICE
+  ConstTensorRef const_ref() const {
+    return ConstTensorRef(*this);
+  }
+
+  /// Returns a Tensor_view given location and size quantities
+  CUTLASS_HOST_DEVICE
+  TensorView subview(TensorCoord const& location, TensorCoord size) const {
+    return TensorView((*this) + location, size.clamp(size_ - location));
+  }
+
+  /// Returns the number of scalar elements needed to store tensor
+  CUTLASS_HOST_DEVICE
+  size_t capacity() const {
+    int max_rank = 0;
+
+    StorageCoord mapped_size(this->map(size()));
+
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < Base::kStorageRank; ++i) {
+      if (!i ||
+        this->stride(i) * mapped_size[i] > this->stride(max_rank) * mapped_size[max_rank]) {
+        max_rank = i;
+      }
+    }
+    return this->stride(max_rank) * mapped_size[max_rank];
+  }
+
+  /// Returns a TensorView offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorView operator+(TensorCoord const& b) const {
+    TensorView result(*this);
+    result.add_pointer_offset(this->offset(b));
+    return result;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorView& operator+=(TensorCoord const& b) {
+    this->add_pointer_offset(this->offset(b));
+    return *this;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorView operator-(TensorCoord const& b) const {
+    TensorRef result(*this);
+    result.add_pointer_offset(-this->offset(b));
+    return result;
+  }
+
+  /// Returns a TensorRef offset by a given amount
+  CUTLASS_HOST_DEVICE
+  TensorView& operator-=(TensorCoord const& b) {
+    this->add_pointer_offset(-this->offset(b));
+    return *this;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/tile_allocation.h

@@ -0,0 +1,143 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a fragment based on a Shape<> template.
+*/
+#pragma once
+
+#include "cutlass/shape.h"
+#include "cutlass/fragment.h"
+#include "cutlass/tensor_ref.h"
+#include "cutlass/zip_tensor_ref.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Class for storing a tile in memory and accessing it through a tensor ref
+template <typename Scalar_, typename Shape_>
+struct TileAllocation {
+  //
+  // Type definitions
+  //
+
+  /// Scalar element
+  typedef Scalar_ Scalar;
+
+  /// The actual storage (may differ from the scalar type)
+  typedef typename StorageType<sizeof(Scalar)>::Type Storage;
+
+  /// Size of the allocation in units of scalars
+  typedef Shape_ Shape;
+
+  /// Strides
+  typedef typename ShapeStrides<Shape, 1>::Shape Strides;
+
+  /// Defines the tensor reference for this allocation
+  typedef TensorRef<Scalar const, 4> ConstTensorRef;
+
+  /// Defines the tensor reference for this allocation
+  typedef TensorRef<Scalar, 4> TensorRef;
+
+  //
+  // Data members
+  //
+
+  /// Storage
+  Storage storage[Shape::kD][Shape::kH][Shape::kW][Shape::kC];
+
+  //
+  // Methods
+  //
+
+  /// Returns a pointer to the raw data
+  CUTLASS_DEVICE
+  Scalar *data() { return reinterpret_cast<Scalar *>(&storage[0][0][0][0]); }
+
+  /// Returns a const pointer to the raw data
+  CUTLASS_DEVICE
+  Scalar const *data() const { return reinterpret_cast<Scalar const *>(&storage[0][0][0][0]); }
+
+  /// Returns a TensorRef object pointing to the data
+  CUTLASS_DEVICE
+  TensorRef reference() {
+    return TensorRef(data(), make_Coord(Strides::kD, Strides::kH, Strides::kW, Strides::kC));
+  }
+
+  /// Returns a TensorRef object pointing to the data
+  CUTLASS_DEVICE
+  ConstTensorRef reference() const {
+    return ConstTensorRef(data(), make_Coord(Strides::kD, Strides::kH, Strides::kW, Strides::kC));
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Manages a pair of tile allocations as if they are one allocation
+template <typename First_, typename Second_>
+struct ZipTileAllocation {
+  //
+  // Type definitions
+  //
+
+  /// First tensor allocation
+  typedef First_ First;
+
+  /// Second tensor allocation
+  typedef Second_ Second;
+
+  /// Defines the tensor reference for this allocation
+  typedef ZipTensorRef<typename First::TensorRef, typename Second::TensorRef> TensorRef;
+
+  /// Defines the tensor reference for this allocation
+  typedef ZipTensorRef<typename First::ConstTensorRef, typename Second::ConstTensorRef>
+      ConstTensorRef;
+
+  //
+  // Data members
+  //
+
+  /// First tensor allocation
+  First first;
+
+  /// Second tensor allocation
+  Second second;
+
+  //
+  // Methods
+  //
+
+  /// Returns a TensorRef object pointing to the data
+  CUTLASS_DEVICE
+  TensorRef reference() { return TensorRef(first.reference(), second.reference()); }
+
+  /// Returns a TensorRef object pointing to the data
+  CUTLASS_DEVICE
+  ConstTensorRef reference() const { return ConstTensorRef(first.reference(), second.reference()); }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/tile_coord.h

@@ -0,0 +1,194 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a coordinate used for the CUTLASS 4-D tile structure. 
+*/
+
+#pragma once
+
+#include "cutlass/coord.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// TileCoord wraps Coord<4, int> to provide a helper for accessing named dimensions. Classes
+/// expecting a coordinate in the rank=4 index space of a CUTLASS tile structure should use TileCoord. 
+template <typename Index_ = int>
+struct TileCoord : public Coord<4, Index_> {
+  
+  /// Index type
+  typedef Index_ Index;
+
+  /// Underlying Coord<4>
+  typedef Coord<4, Index> Base;
+
+  /// D dimension
+  static int kD = 0;
+
+  /// H dimension
+  static int kH = 1;
+
+  /// W dimension
+  static int kW = 2;
+
+  /// C dimension
+  static int kC = 3;
+
+  //
+  // Methods
+  //
+
+  /// Default ctor
+  CUTLASS_HOST_DEVICE
+  TileCoord() { }
+
+  /// Constructs from Coord<3> and infers coord[kC] = 0
+  CUTLASS_HOST_DEVICE
+  TileCoord(Coord<3, Index> const &coord): 
+    Base(make_Coord(coord[0], coord[1], coord[2], 0)) { }
+
+  /// Constructs from Coord<4>
+  CUTLASS_HOST_DEVICE
+  TileCoord(Coord<4, Index> const &coord): Base(coord) { }
+
+  /// Constructs from an array of coordinate elements
+  CUTLASS_HOST_DEVICE
+  TileCoord(Index coord[4]): Base(coord) { }
+  
+  /// Helper to construct from a row and column
+  CUTLASS_HOST_DEVICE
+  TileCoord(Index d, Index h, Index w, Index c): Base(make_Coord(d, h, w, c)) { }
+
+  /// Returns the D element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & d() const { return this->at(kD); }
+
+  /// Returns the D element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index & d() { return this->at(kD); }
+
+  /// Returns the H element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & h() const { return this->at(kH); }
+
+  /// Returns the H element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index & h() { return this->at(kH); }
+
+  /// Returns the W element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & w() const { return this->at(kW); }
+
+  /// Returns the W element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index & w() { return this->at(kW); }
+
+  /// Returns the Celement of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index const & c() const { return this->at(kC); }
+
+  /// Returns the C element of the coordinate
+  CUTLASS_HOST_DEVICE
+  Index & c() { return this->at(kC); }
+
+  /// Gets H and W dimensions as a Coord<2>
+  CUTLASS_HOST_DEVICE
+  Coord<2> hw() const {
+    return make_Coord(h(), w());
+  }
+
+  /// Gets H, W, and C dimensions as a Coord<3>
+  CUTLASS_HOST_DEVICE
+  Coord<3> hwc() const {
+    return make_Coord(h(), w(), c());
+  }
+
+  /// Gets D, H, and W dimensions as a Coord<3>
+  CUTLASS_HOST_DEVICE
+  Coord<3> dhw() const {
+    return make_Coord(d(), h(), w());
+  }
+
+  //
+  // Coord operators
+  //
+
+  /// Element-wise addition
+  CUTLASS_HOST_DEVICE
+  TileCoord operator+(Base const& b) const {
+    return TileCoord(Base::operator+(b));
+  }
+
+  /// Element-wise subtraction
+  CUTLASS_HOST_DEVICE
+  TileCoord operator-(Base const& b) const {
+    return TileCoord(Base::operator-(b));
+  }
+
+  /// Element-wise multiplication
+  CUTLASS_HOST_DEVICE
+  TileCoord operator*(Base const& b) const {
+    return TileCoord(Base::operator*(b));
+  }
+
+  /// Element-wise division
+  CUTLASS_HOST_DEVICE
+  TileCoord operator/(Base const& b) const {
+    return TileCoord(Base::operator/(b));
+  }
+
+  /// In-place addition
+  CUTLASS_HOST_DEVICE
+  TileCoord& operator+=(Base const& b) {
+    Base::operator+=(b);
+    return *this;
+  }
+
+  /// In-place subtraction
+  CUTLASS_HOST_DEVICE
+  TileCoord& operator-=(Base const& b) {
+    Base::operator-=(b);
+    return *this;
+  }
+
+  /// In-place multiplication
+  CUTLASS_HOST_DEVICE
+  TileCoord& operator*=(Base const& b) {
+    Base::operator*=(b);
+    return *this;
+  }
+
+  /// In-place division
+  CUTLASS_HOST_DEVICE
+  TileCoord& operator/=(Base const& b) {
+    Base::operator/=(b);
+    return *this;
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass

cutlass/tile_stream.h

@@ -0,0 +1,378 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Implements the tile stream concept, composing an iterator with a transformation. Offers
+      split-phase semantics, separating the initiation of an asynchronous memory operation with a
+      fence forcing it to complete.
+*/
+#pragma once
+
+// clang-format off
+
+#include "cutlass/convert.h"
+#include "cutlass/tile_iterator.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Generic stream for loading and transforming fragments
+template <typename Iterator_, typename Transformer_ = Copy<typename Iterator_::Fragment> >
+struct TileLoadStream {
+  //
+  // Type definitions
+  //
+
+  /// TileLoadIterator
+  typedef Iterator_ Iterator;
+
+  /// Transformer
+  typedef Transformer_ Transformer;
+
+  /// Fragment fetched from source memory
+  typedef typename Iterator::Fragment Fragment;
+
+  /// Output fragment from transformer
+  typedef typename Transformer::OutputFragment TransformedFragment;
+
+  /// Tensor reference expected by the stream
+  typedef typename Iterator::TensorRef TensorRef;
+
+  /// Empty predicate vector struct
+  struct PredicateVector {};
+
+  /// Index type
+  typedef typename Iterator::Index Index;
+
+  /// Parameters object used to construct generic load stream
+  struct Params {
+    /// Parameters to the iterator
+    typename Iterator::Params iterator;
+
+    //
+    // Methods
+    //
+
+    /// Default constructor
+    CUTLASS_HOST_DEVICE
+    Params() {}
+
+    /// Constructor with iterator params
+    CUTLASS_HOST_DEVICE
+    Params(typename Iterator::Params const &_iterator) : iterator(_iterator) {}
+  };
+
+  //
+  // Data members
+  //
+
+  /// Iterator to load tiles
+  Iterator iterator;
+
+  /// Fragment loaded via iterator
+  Fragment fetched_fragment;
+
+  /// Transformation applied to fragments
+  Transformer transformer;
+
+  /// Transformed fragment from transformer
+  TransformedFragment transformed_fragment;
+
+  //
+  // Methods
+  //
+
+  /// Ctor
+  CUTLASS_DEVICE
+  TileLoadStream(Params const &_params, TensorRef const &_ref)
+      : iterator(_params.iterator, _ref) {}
+
+  /// Ctor
+  CUTLASS_DEVICE
+  TileLoadStream(Params const &_params,
+    Coord<3> const &threadblock_offset = make_Coord(0, 0, 0)
+  ): iterator(_params.iterator, threadblock_offset) { }
+
+  /// Loads a tile and increments the iterator
+  CUTLASS_DEVICE
+  void copy() { iterator.load_post_increment(fetched_fragment); }
+
+  /// Commits the fetched fragment and applies a transformation
+  CUTLASS_DEVICE
+  void commit() { transformer.transform(fetched_fragment, transformed_fragment); }
+
+  /// Accesses the loaded, transformed fragment
+  CUTLASS_DEVICE
+  Fragment &intermediate_fragment() { return fetched_fragment; }
+
+  /// Accesses the loaded, transformed fragment
+  CUTLASS_DEVICE
+  TransformedFragment &fragment() { return transformed_fragment; }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Generic stream for transforming and storing fragments
+template <typename Iterator_, typename Transformer_ = Copy<typename Iterator_::Fragment> >
+struct TileStoreStream {
+  //
+  // Type definitions
+  //
+
+  /// TileLoadIterator
+  typedef Iterator_ Iterator;
+
+  /// Transformer
+  typedef Transformer_ Transformer;
+
+  /// Source fragment
+  typedef typename Transformer::InputFragment Fragment;
+
+  /// Transformed fragment, compatible with Iterator::Fragment
+  typedef typename Transformer::OutputFragment TransformedFragment;
+
+  /// Tensor reference expected by the underlying iterator
+  typedef typename Iterator::TensorRef TensorRef;
+
+  /// Empty predicate vector struct
+  struct PredicateVector {};
+
+  /// Index type
+  typedef typename Iterator::Index Index;
+
+  /// Parameters used to construct the stream
+  struct Params {
+    /// Parameters to the iterator
+    typename Iterator::Params iterator;
+
+    //
+    // Methods
+    //
+
+    /// Default constructor
+    CUTLASS_HOST_DEVICE
+    Params() {}
+
+    /// Constructor with iterator params
+    CUTLASS_HOST_DEVICE
+    Params(typename Iterator::Params const &_iterator) : iterator(_iterator) {}
+  };
+
+  //
+  // Data members
+  //
+
+  /// Iterator to store tiles
+  Iterator iterator;
+
+  /// Transformation applied to inputs
+  Transformer transformer;
+
+  /// Source fragment
+  Fragment source_fragment;
+
+  /// Transformed fragment from transformer
+  TransformedFragment transformed_fragment;
+
+  //
+  // Methods
+  //
+
+  /// Ctor
+  CUTLASS_DEVICE
+  TileStoreStream(Params const &_params, TensorRef const &_ref)
+      : iterator(_params.iterator, _ref) {}
+
+  /// Ctor
+  CUTLASS_DEVICE
+  TileStoreStream(Params const &_params,
+                  Coord<3> const &threadblock_offset = make_Coord(0, 0, 0)
+  ): iterator(_params.iterator, threadblock_offset) { }
+
+  /// Stores a fragment and increments the iterator
+  CUTLASS_DEVICE
+  void copy() {
+
+    transformer.transform(source_fragment, transformed_fragment);
+    iterator.store_post_increment(transformed_fragment);
+  }
+
+  /// Stores a fragment and increments the iterator
+  CUTLASS_DEVICE
+  void copy(Fragment const &frag) {
+    source_fragment = frag;
+    copy();
+  }
+
+  /// Commits the store operation
+  CUTLASS_DEVICE
+  void commit() {}
+
+  /// Accesses the transformed fragment
+  CUTLASS_DEVICE
+  Fragment &fragment() { return source_fragment; }
+
+  /// Accesses the fragment after trasnforming
+  CUTLASS_DEVICE
+  TransformedFragment &intermediate_fragment() { return transformed_fragment; }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Generic stream for loading and transforming fragments
+template <typename Iterator_,
+          typename PredicateFunctor_ =
+              RegularTilePredicateFunctor<typename Iterator_::Traits::Delta>,
+          typename Transformer_ = Copy<typename Iterator_::Fragment> >
+struct PredicatedTileLoadStream : public TileLoadStream<Iterator_, Transformer_> {
+  //
+  // Type definitions
+  //
+
+  typedef TileLoadStream<Iterator_, Transformer_> Base;
+
+  /// TileLoadIterator
+  typedef Iterator_ Iterator;
+
+  /// Predicate functor
+  typedef PredicateFunctor_ PredicateFunctor;
+
+  /// Transformer
+  typedef Transformer_ Transformer;
+
+  /// Fragment fetched from source memory
+  typedef typename Base::Fragment Fragment;
+
+  /// Output fragment from transformer
+  typedef typename Base::TransformedFragment TransformedFragment;
+
+  /// Parameters object used to construct generic load stream
+  typedef typename Base::Params Params;
+
+  //
+  // Data members
+  //
+
+  /// Predicates
+  typename Iterator::PredicateVector predicates;
+
+  //
+  // Methods
+  //
+
+  /// Ctor
+  CUTLASS_DEVICE
+  PredicatedTileLoadStream(Params const &_params,
+                          Coord<3> const &bounds,
+                          Coord<3> const &threadblock_offset = make_Coord(0, 0, 0))
+      : Base(_params, threadblock_offset) {
+    this->iterator.initialize_predicates(
+        predicates.begin(), PredicateFunctor(bounds), threadblock_offset);
+  }
+
+  /// Loads a tile and increments the iterator
+  CUTLASS_DEVICE
+  void copy() { this->iterator.load_post_increment(this->fetched_fragment, predicates.begin()); }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Generic stream for transforming and storing fragments
+template <typename Iterator_,
+          typename PredicateFunctor_ =
+              RegularTilePredicateFunctor<typename Iterator_::Traits::Delta>,
+          typename Transformer_ = Copy<typename Iterator_::Fragment> >
+struct PredicatedTileStoreStream : public TileStoreStream<Iterator_, Transformer_> {
+  //
+  // Type definitions
+  //
+
+  typedef TileStoreStream<Iterator_, Transformer_> Base;
+
+  /// TileLoadIterator
+  typedef Iterator_ Iterator;
+
+  /// Predicate functor
+  typedef PredicateFunctor_ PredicateFunctor;
+
+  /// Transformer
+  typedef Transformer_ Transformer;
+
+  /// Fragment fetched from source memory
+  typedef typename Base::Fragment Fragment;
+
+  /// Output fragment from transformer
+  typedef typename Base::TransformedFragment TransformedFragment;
+
+  /// Parameters object used to construct generic load stream
+  typedef typename Base::Params Params;
+
+  //
+  // Data members
+  //
+
+  /// Predicates
+  typename Iterator::PredicateVector predicates;
+
+  //
+  // Methods
+  //
+
+  /// Ctor
+  CUTLASS_DEVICE
+  PredicatedTileStoreStream(Params const &_params,
+                           Coord<3> const &bounds,
+                           Coord<3> const &threadblock_offset = make_Coord(0, 0, 0))
+      : Base(_params, threadblock_offset) {
+    this->iterator.initialize_predicates(
+        predicates.begin(), PredicateFunctor(bounds), threadblock_offset);
+  }
+
+  /// Stores the fragment and increments the iterator
+  CUTLASS_DEVICE
+  void copy() {
+    this->transformer.transform(this->source_fragment, this->transformed_fragment);
+    this->iterator.store_post_increment(this->transformed_fragment, predicates.begin());
+  }
+
+  /// Stores the fragment and increments the iterator
+  CUTLASS_DEVICE
+  void copy(Fragment const &frag) {
+    this->source_fragment = frag;
+    copy();
+  }
+
+  /// Commits the store operation
+  CUTLASS_DEVICE
+  void commit() {}
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass
+
+// clang-format on

cutlass/tile_traits_standard.h

@@ -0,0 +1,240 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines tile traits for several tile partitioning arrangements of threads expected to
+      achieve efficient streaming performance.
+*/
+#pragma once
+
+#include "cutlass/tile_iterator.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Basic thread offset function computed from a thread shape
+template <typename ThreadShape>
+struct TiledThreadOffset {
+  /// Computes the logical coordinate from thread shape
+  CUTLASS_HOST_DEVICE
+  Coord<4> operator()() const {
+    Coord<4> thread_offset;
+
+    int index = threadIdx.x;
+
+    thread_offset[3] = (index % ThreadShape::kC);
+    index = (index / ThreadShape::kC);
+
+    thread_offset[2] = (index % ThreadShape::kW);
+    index = (index / ThreadShape::kW);
+
+    thread_offset[1] = (index % ThreadShape::kH);
+    index = (index / ThreadShape::kH);
+
+    thread_offset[0] = index;
+
+    return thread_offset;
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Tiling in which the number of threads is greater than the
+/// contiguous dimension of the tile.
+template <typename Tile_, int Threads>
+struct TileTraitsStrideMajor {
+  /// Shape of tile
+  typedef Tile_ Tile;
+
+  /// Number of participating threads
+  static int const kThreads = Threads;
+
+  // Static assertions
+  static_assert(!(ShapeCount<Tile>::kDhw % kThreads),
+                "Tiling undefined if elements not divisible by threads.");
+
+  static_assert(Tile::kW <= kThreads,
+                "This specialization assumes there are more threads than the contiguous dimension "
+                "of the tile.");
+
+  /// Shape of threads
+  typedef Shape<1, kThreads / Tile::kW, Tile::kW, 1> ThreadShape;
+
+  /// Delta along each dimension
+  typedef Shape<1, ThreadShape::kH, 1, 1> Delta;
+
+  /// Number of iterations
+  typedef Shape<1, Tile::kH / ThreadShape::kH, 1, 1> Iterations;
+
+  /// Computes the initial offset
+  typedef TiledThreadOffset<ThreadShape> ThreadOffset;
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Tiling in which the number of threads is fewer than the tile size
+/// in the contiguous dimension.
+template <typename Tile_, int Threads>
+struct TileTraitsContiguousMajor {
+  /// Shape of tile
+  typedef Tile_ Tile;
+
+  /// Number of participating threads
+  static int const kThreads = Threads;
+
+  // Static assertions
+  static_assert(Tile::kW >= kThreads,
+                "This specialization assumes there are more threads than the contiguous dimension "
+                "of the tile.");
+
+  static_assert(!(ShapeCount<Tile>::kDhw % kThreads),
+                "Tiling undefined if elements not divisible by threads.");
+
+  static_assert(!(Tile::kW % kThreads),
+                "The contiguous size of the tile must be divisible by the number of threads.");
+
+  /// Thread shape
+  typedef Shape<1, 1, kThreads> ThreadShape;
+
+  /// Delta between each thread's access
+  typedef Shape<1, 1, kThreads> Delta;
+
+  /// Number of iterations
+  typedef Shape<1, Tile::kH, Tile::kW / kThreads> Iterations;
+
+  /// Computes the initial offset
+  typedef TiledThreadOffset<ThreadShape> ThreadOffset;
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Tiling in which warps rake across the contiguous dimension
+template <typename Tile_, int Threads>
+struct TileTraitsWarpRake {
+  /// Shape of tile
+  typedef Tile_ Tile;
+
+  /// Number of participating threads
+  static int const kThreads = Threads;
+
+  /// Hard-coded warp size
+  static int const kWarpSize = 32;
+
+  /// Number of participating warps
+  static int const kWarpCount = kThreads / kWarpSize;
+
+  // Static assertions
+  static_assert(!(ShapeCount<Tile>::kDhw % kThreads),
+                "Tiling undefined if elements not divisible by threads.");
+
+  static_assert(!(kThreads % kWarpSize), "Number of threads must be divisible by the warp size.");
+
+  static_assert(!(Tile::kW % kWarpSize), "Contiguous dimension must be divisible by the warp size");
+
+  /// Warps strip-mined across strided dimension
+  static int const kWarpsStrided = __NV_STD_MIN(kWarpCount, Tile::kH);
+
+  /// Warps stripmined contiguous dimension
+  static int const kWarpsContiguous = kWarpCount / kWarpsStrided;
+
+  /// Arrangement of threads
+  typedef Shape<1, kWarpsStrided, kWarpsContiguous * kWarpSize> ThreadShape;
+
+  /// The same warp rakes along the contiguous dimension
+  typedef Shape<1, kWarpsStrided, kWarpSize> Delta;
+
+  /// Number of iterations
+  typedef Shape<1, Tile::kH / Delta::kH, Tile::kW / ThreadShape::kW> Iterations;
+
+  /// Computes the thread offset in (H, W) based on thread ID
+  struct ThreadOffset {
+    /// Basic thread offset function computed from a thread shape
+    CUTLASS_HOST_DEVICE
+    Coord<4> operator()() const {
+      int tid = threadIdx.x;
+      int warp = (tid / kWarpSize);
+      int lane = (tid % kWarpSize);
+
+      static int const kWarpSpanContiguous = kWarpSize * Iterations::kW;
+
+      int warp_w = (warp % kWarpsContiguous);
+      int warp_h = (warp / kWarpsContiguous);
+
+      return make_Coord(0, warp_h, lane + kWarpSpanContiguous * warp_w, 0);
+    }
+  };
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Chooses 'best' shape to enable warp raking along contiguous dimension if possible.
+template <typename Tile_, int Threads>
+struct TileTraitsStandard {
+  /// Shape of tile
+  typedef Tile_ Tile;
+
+  /// Number of participating threads
+  static int const kThreads = Threads;
+
+  /// Hard-coded warp size
+  static int const kWarpSize = 32;
+
+  /// Number of participating warps
+  static int const kWarpCount = kThreads / kWarpSize;
+
+  /// By default, do not do scalar loads
+  static int const kAccessSize = 1;
+
+  // Static assertions
+  static_assert(!(ShapeCount<Tile>::kDhw % kThreads),
+                "Tiling undefined if elements not divisible by threads.");
+
+  /// Choose the stride-major contiguous tiling if the contiguous dimension is
+  /// smaller than the warp size. Otherwise, if it is divisible by the warp size,
+  /// choose the warp rake arrangement.
+  typedef typename platform::conditional <
+      Tile::kW<kWarpSize,
+               TileTraitsStrideMajor<Tile, Threads>,
+               typename platform::conditional<!(Tile::kW % kWarpSize),
+                                              TileTraitsWarpRake<Tile, Threads>,
+                                              TileTraitsContiguousMajor<Tile, Threads> >::type>::
+          type Traits;
+
+  /// Delta between accesses
+  typedef typename Traits::Delta Delta;
+
+  /// Delta between each thread's access
+  typedef Shape<0, 0, 0, 0> ImmediateOffsetStrides;
+
+  /// Number of accesses
+  typedef typename Traits::Iterations Iterations;
+
+  /// Thread offset functor
+  typedef typename Traits::ThreadOffset ThreadOffset;
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/util/complex.h

@@ -0,0 +1,457 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+#pragma once
+
+#include <cuComplex.h>
+#include "cutlass/cutlass.h"
+#include <iosfwd>
+
+namespace cutlass {
+namespace platform {
+
+//////////////////////////////////////////////////////////////////////////////////////////////////
+
+//
+// Accessors for CUDA complex types
+//
+
+/// Returns the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+float const &real(cuFloatComplex const &z) { return z.x; }
+
+/// Returns the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+float &real(cuFloatComplex &z) { return z.x; }
+
+/// Returns the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+double const &real(cuDoubleComplex const &z) { return z.x; }
+
+/// Returns the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+double &real(cuDoubleComplex &z) { return z.x; }
+
+/// Returns the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+float const &imag(cuFloatComplex const &z) { return z.y; }
+
+/// Returns the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+float &imag(cuFloatComplex &z) { return z.y; }
+
+/// Returns the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+double const &imag(cuDoubleComplex const &z) { return z.y; }
+
+/// Returns the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+CUTLASS_HOST_DEVICE
+double &imag(cuDoubleComplex &z) { return z.y; }
+
+//////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Class for representing and manipulating complex numbers with conversions from built-in CUDA
+/// complex types.
+template <typename T>
+class complex {
+ public:
+  /// Type alias for scalar type
+  typedef T value_type;
+
+ private:
+  //
+  // Data members
+  //
+
+  /// Real part
+  T _real;
+
+  /// Imaginary part
+  T _imag;
+
+ public:
+//
+// Methods
+//
+
+/// Constructor
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  complex(T r = T(0), T i = T(0)) : _real(r), _imag(i) {}
+
+/// Conversion from cuFloatComplex
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  complex(cuFloatComplex const &z) : _real(platform::real(z)), _imag(platform::imag(z)) {}
+
+/// Conversion from cuDoubleComplex
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  complex(cuDoubleComplex const &z) : _real(platform::real(z)), _imag(platform::imag(z)) {}
+
+/// Accesses the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  T const &real() const { return _real; }
+
+/// Accesses the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  T &real() { return _real; }
+
+/// Accesses the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  T const &imag() const { return _imag; }
+
+/// Accesses the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  T &imag() { return _imag; }
+
+/// Converts to cuFloatComplex
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  operator cuFloatComplex() const { return make_cuFloatComplex(real(), imag()); }
+
+/// Converts to cuDoubleComplex
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+  CUTLASS_HOST_DEVICE
+  operator cuDoubleComplex() const { return make_cuDoubleComplex(real(), imag()); }
+};
+
+//
+// Accessors for complex template
+//
+
+/// Returns the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T const &real(complex<T> const &z) {
+  return z.real();
+}
+
+/// Returns the real part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T &real(complex<T> &z) {
+  return z.real();
+}
+
+/// Returns the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T const &imag(complex<T> const &z) {
+  return z.imag();
+}
+
+/// Returns the imaginary part of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T &imag(complex<T> &z) {
+  return z.imag();
+}
+
+//
+// Output operators
+//
+
+template <typename T>
+std::ostream &operator<<(std::ostream &out, complex<T> const &z) {
+  T _r = real(z);
+  T _i = imag(z);
+  return out << _r << "+i" << _i;
+}
+
+//
+// Non-member operators defined for complex types
+//
+
+/// Equality operator
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE bool operator==(complex<T> const &lhs, complex<T> const &rhs) {
+  return real(lhs) == (rhs) && imag(lhs) == imag(rhs);
+}
+
+/// Inequality operator
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE bool operator!=(complex<T> const &lhs, complex<T> const &rhs) {
+  return !(lhs == rhs);
+}
+
+/// Addition
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator+(complex<T> const &lhs, complex<T> const &rhs) {
+  return complex<T>(real(lhs) + real(rhs), imag(lhs) + imag(rhs));
+}
+
+/// Subtraction
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator-(complex<T> const &lhs, complex<T> const &rhs) {
+  return complex<T>(real(lhs) - real(rhs), imag(lhs) - imag(rhs));
+}
+
+/// Multiplication
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator*(complex<T> const &lhs, complex<T> const &rhs) {
+  return complex<T>(real(lhs) * real(rhs) - imag(lhs) * imag(rhs),
+                    real(lhs) * imag(rhs) + imag(lhs) * real(rhs));
+}
+
+/// Scalar Multiplication
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator*(complex<T> const &lhs, T const &s) {
+  return complex<T>(real(lhs) * s, imag(lhs) * s);
+}
+
+/// Scalar Multiplication
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator*(T const &s, complex<T> const &rhs) {
+  return complex<T>(s * real(rhs), s * imag(rhs));
+}
+
+/// Division
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator/(complex<T> const &lhs, complex<T> const &rhs) {
+  T d = (real(rhs) * (rhs) + imag(rhs) * imag(rhs));
+
+  return complex<T>((real(lhs) * (rhs) + imag(lhs) * imag(rhs)) / d,
+                    (imag(lhs) * (rhs)-real(lhs) * imag(rhs)) / d);
+}
+
+/// Scalar Division
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator/(complex<T> const &lhs, T const &s) {
+  return complex<T>(real(lhs) / s, imag(lhs) / s);
+}
+
+/// Scalar divided by complex
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> operator/(T const &s, complex<T> const &rhs) {
+  T d = (real(rhs) * (rhs) + imag(rhs) * imag(rhs));
+
+  return complex<T>((s * (rhs)) / d, -(s * imag(rhs)) / d);
+}
+
+/// Addition
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> &operator+=(complex<T> &lhs, complex<T> const &rhs) {
+  lhs = (lhs + rhs);
+  return lhs;
+}
+
+/// Subtraction
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> &operator-=(complex<T> &lhs, complex<T> const &rhs) {
+  lhs = (lhs - rhs);
+  return lhs;
+}
+
+/// Multiplication
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> &operator*=(complex<T> &lhs, complex<T> const &rhs) {
+  lhs = (lhs * rhs);
+  return lhs;
+}
+
+/// Scalar multiplication
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> &operator*=(complex<T> &lhs, T s) {
+  lhs = (lhs * s);
+  return lhs;
+}
+
+/// Division
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> &operator/=(complex<T> &lhs, complex<T> const &rhs) {
+  lhs = (lhs / rhs);
+  return lhs;
+}
+
+//
+// Non-member functions defined for complex numbers
+//
+
+/// Returns the magnitude of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T abs(complex<T> const &z) {
+  return sqrt(norm(z));
+}
+
+/// Returns the magnitude of the complex number
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T arg(complex<T> const &z) {
+  return atan2(imag(z), real(z));
+}
+
+/// Returns the squared magnitude
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE T norm(complex<T> const &z) {
+  return real(z) * real(z) + imag(z) * imag(z);
+}
+
+/// Returns the complex conjugate
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> conj(complex<T> const &z) {
+  return complex<T>(real(z), -imag(z));
+}
+
+/// Projects the complex number z onto the Riemann sphere
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> proj(complex<T> const &z) {
+  T d = real(z) * real(z) + imag(z) * imag(z) + T(1);
+  return complex<T>((T(2) * real(z)) / d, (T(2) * imag(z)) / d);
+}
+
+/// Returns a complex number with magnitude r and phase theta
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> polar(T const &r, T const &theta = T()) {
+  return complex<T>(r * cos(theta), r * sin(theta));
+}
+
+/// Computes the complex exponential of z.
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> exp(complex<T> const &z) {
+  return complex<T>(real(z) * cos(imag(z)), real(z) * sin(imag(z)));
+}
+
+/// Computes the complex exponential of z.
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> log(complex<T> const &z) {
+  return complex<T>(log(abs(z)), arg(z));
+}
+
+/// Computes the complex exponential of z.
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> log10(complex<T> const &z) {
+  return log(z) / T(log(T(10)));
+}
+
+/// Computes the square root of complex number z
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> sqrt(complex<T> const &z) {
+  return sqrt(T(2)) / T(2) *
+         complex<T>(sqrt(sqrt(norm(z)) + real(z)),
+                    (imag(z) < 0 ? T(-1) : T(1)) * sqrt(sqrt(norm(z)) - real(z)));
+}
+
+/// Computes the cosine of complex z.
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> cos(complex<T> const &z) {
+  return (exp(z) + exp(-z)) / T(2);
+}
+
+/// Computes the sin of complex z.
+#pragma hd_warning_disable  // Suppresses warnings when attempting to instantiate complex<T> with a
+                            // host-only type
+template <typename T>
+CUTLASS_HOST_DEVICE complex<T> sin(complex<T> const &z) {
+  return (exp(-z) - exp(z)) * complex<T>(T(0), T(1) / T(2));
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace platform
+}  // namespace cutlass

cutlass/util/cutlass_math.h

@@ -0,0 +1,165 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+/**
+ * \file
+ * \brief Math utilities
+ */
+
+#include "cutlass/util/platform.h"
+
+namespace cutlass {
+
+/******************************************************************************
+ * Static math utilities
+ ******************************************************************************/
+
+/**
+ * Statically determine if N is a power-of-two
+ */
+template <int N>
+struct is_pow2 : platform::integral_constant<bool, (N & (N - 1)) == 0> {};
+
+/**
+ * Statically determine log2(N), rounded down
+ */
+template <int N, int CurrentVal = N, int Count = 0>
+struct log2_down {
+  /// Static logarithm value
+  enum { value = log2_down<N, (CurrentVal >> 1), Count + 1>::value };
+};
+
+// Base case
+template <int N, int Count>
+struct log2_down<N, 1, Count> {
+  enum { value = Count };
+};
+
+/**
+ * Statically determine log2(N), rounded up
+ */
+template <int N, int CurrentVal = N, int Count = 0>
+struct log2_up {
+  /// Static logarithm value
+  enum { value = log2_up<N, (CurrentVal >> 1), Count + 1>::value };
+};
+
+// Base case
+template <int N, int Count>
+struct log2_up<N, 1, Count> {
+  enum { value = ((1 << Count) < N) ? Count + 1 : Count };
+};
+
+/**
+ * Statically estimate sqrt(N) to the nearest power-of-two
+ */
+template <int N>
+struct sqrt_est {
+  enum { value = 1 << (log2_up<N>::value / 2) };
+};
+
+/**
+ * For performing a constant-division with a compile-time assertion that the
+ * Divisor evenly-divides the Dividend.
+ */
+template <int Dividend, int Divisor>
+struct divide_assert {
+  enum { value = Dividend / Divisor };
+
+  static_assert((Dividend % Divisor == 0), "Not an even multiple");
+};
+
+/******************************************************************************
+ * Rounding
+ ******************************************************************************/
+
+/**
+ * Round dividend up to the nearest multiple of divisor
+ */
+template <typename dividend_t, typename divisor_t>
+CUTLASS_HOST_DEVICE dividend_t round_nearest(dividend_t dividend, divisor_t divisor) {
+  return ((dividend + divisor - 1) / divisor) * divisor;
+}
+
+/**
+ * Greatest common divisor
+ */
+template <typename value_t>
+CUTLASS_HOST_DEVICE value_t gcd(value_t a, value_t b) {
+  for (;;) {
+    if (a == 0) return b;
+    b %= a;
+    if (b == 0) return a;
+    a %= b;
+  }
+}
+
+/**
+ * Least common multiple
+ */
+template <typename value_t>
+CUTLASS_HOST_DEVICE value_t lcm(value_t a, value_t b) {
+  value_t temp = gcd(a, b);
+
+  return temp ? (a / temp * b) : 0;
+}
+
+/**
+ * log2 computation, what's the
+ * difference between the below codes and
+ * log2_up/down codes?
+ */
+template <typename value_t>
+CUTLASS_HOST_DEVICE value_t clz(value_t x) {
+  for (int i = 31; i >= 0; --i) {
+    if ((1 << i) & x) return 31 - i;
+  }
+  return 32;
+}
+
+template <typename value_t>
+CUTLASS_HOST_DEVICE value_t find_log2(value_t x) {
+  int a = 31 - clz(x);
+  a += (x & (x - 1)) != 0;  // Round up, add 1 if not a power of 2.
+  return a;
+}
+
+/******************************************************************************
+ * Min/Max
+ ******************************************************************************/
+
+template <int A, int B>
+struct Min {
+  static int const kValue = (A < B) ? A : B;
+};
+
+template <int A, int B>
+struct Max {
+  static int const kValue = (A > B) ? A : B;
+};
+
+}  // namespace cutlass

cutlass/util/debug.h

@@ -1,29 +1,27 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
  *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
  *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  *
- ******************************************************************************/
+ **************************************************************************************************/
 
 #pragma once
 
@@ -44,87 +42,81 @@ namespace cutlass {
  * Formats and prints the given message to stdout
  */
 #if !defined(CUDA_LOG)
-    #if !defined(__CUDA_ARCH__)
-        #define CUDA_LOG(format, ...) printf(format, __VA_ARGS__)
-    #else
-inline __host__ __device__ unsigned get_threadidx_x() { return threadIdx.x; }
-inline __host__ __device__ unsigned get_threadidx_y() { return threadIdx.y; }
-inline __host__ __device__ unsigned get_threadidx_z() { return threadIdx.z; }
-inline __host__ __device__ unsigned get_blockidx_x() { return blockIdx.x; }
-inline __host__ __device__ unsigned get_blockidx_y() { return blockIdx.y; }
-inline __host__ __device__ unsigned get_blockidx_z() { return blockIdx.z; }
-        #define CUDA_LOG(format, ...)                                          \
-            printf("[block (%d,%d,%d), thread (%d,%d,%d)]: " format,           \
-                   get_blockidx_x(), get_blockidx_y(), get_blockidx_z(),       \
-                   get_threadidx_x(), get_threadidx_y(), get_threadidx_z(),    \
-                   __VA_ARGS__);
-    #endif
+#if !defined(__CUDA_ARCH__)
+#define CUDA_LOG(format, ...) printf(format, __VA_ARGS__)
+#else
+#define CUDA_LOG(format, ...)                              \
+  printf("[block (%d,%d,%d), thread (%d,%d,%d)]: " format, \
+         blockIdx.x,                                       \
+         blockIdx.y,                                       \
+         blockIdx.z,                                       \
+         threadIdx.x,                                      \
+         threadIdx.y,                                      \
+         threadIdx.z,                                      \
+         __VA_ARGS__);
+#endif
 #endif
-
 
 /**
  * Formats and prints the given message to stdout only if DEBUG is defined
  */
 #if !defined(CUDA_LOG_DEBUG)
-    #ifdef DEBUG
-        #define CUDA_LOG_DEBUG(format, ...) CUDA_LOG(format, __VA_ARGS__)
-    #else
-        #define CUDA_LOG_DEBUG(format, ...)
-    #endif
+#ifdef DEBUG
+#define CUDA_LOG_DEBUG(format, ...) CUDA_LOG(format, __VA_ARGS__)
+#else
+#define CUDA_LOG_DEBUG(format, ...)
+#endif
 #endif
 
-
 /**
- * \brief The corresponding error message is printed to \p stderr (or \p stdout in device code) along with the supplied source context.
+ * \brief The corresponding error message is printed to \p stderr (or \p stdout in device code)
+ * along with the supplied source context.
  *
  * \return The CUDA error.
  */
-__host__ __device__ inline cudaError_t cuda_perror_impl(
-    cudaError_t     error,
-    const char*     filename,
-    int             line)
-{
-    (void)filename;
-    (void)line;
-    if (error)
-    {
+__host__ CUTLASS_DEVICE cudaError_t cuda_perror_impl(cudaError_t error,
+                                                     const char* filename,
+                                                     int line) {
+  (void)filename;
+  (void)line;
+  if (error) {
 #if !defined(__CUDA_ARCH__)
-        fprintf(stderr, "CUDA error %d [%s, %d]: %s\n", error, filename, line, cudaGetErrorString(error));
-        fflush(stderr);
+    fprintf(
+        stderr, "CUDA error %d [%s, %d]: %s\n", error, filename, line, cudaGetErrorString(error));
+    fflush(stderr);
 #else
-        printf("CUDA error %d [%s, %d]\n", error, filename, line);
+    printf("CUDA error %d [%s, %d]\n", error, filename, line);
 #endif
-    }
-    return error;
+  }
+  return error;
 }
 
-
 /**
  * \brief Perror macro
  */
 #ifndef CUDA_PERROR
-    #define CUDA_PERROR(e) cuda_perror_impl((cudaError_t) (e), __FILE__, __LINE__)
+#define CUDA_PERROR(e) cuda_perror_impl((cudaError_t)(e), __FILE__, __LINE__)
 #endif
 
-
 /**
  * \brief Perror macro with exit
  */
 #ifndef CUDA_PERROR_EXIT
-    #define CUDA_PERROR_EXIT(e) if (cuda_perror_impl((cudaError_t) (e), __FILE__, __LINE__)) { exit(1); }
+#define CUDA_PERROR_EXIT(e)                                     \
+  if (cuda_perror_impl((cudaError_t)(e), __FILE__, __LINE__)) { \
+    exit(1);                                                    \
+  }
 #endif
 
-
 /**
  * \brief Perror macro only if DEBUG is defined
  */
 #ifndef CUDA_PERROR_DEBUG
-    #ifdef DEBUG
-        #define CUDA_PERROR_DEBUG(e) CUDA_PERROR(e)
-    #else
-        #define CUDA_PERROR_DEBUG(e) (e)
-    #endif
+#ifdef DEBUG
+#define CUDA_PERROR_DEBUG(e) CUDA_PERROR(e)
+#else
+#define CUDA_PERROR_DEBUG(e) (e)
+#endif
 #endif
 
-
-} // namespace cutlass
+}  // namespace cutlass

cutlass/util/device_introspection.h

@@ -1,224 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief Utilities for device introspection
- */
-
-#include "debug.h"
-#include "nv_std.h"
-#include "printable.h"
-
-namespace cutlass {
-
-
-/******************************************************************************
- * math_operation_class_t
- *
- * Enumeration to select the appropriate math operation
- *
- * The assumption is multiple math operations may be used to compute GEMM
- * for a given selection of operand and accumulator types.
- *
- ******************************************************************************/
-
-/// Math operation
-enum class math_operation_class_t
-{
-    scalar,                     // scalar (and vector) multiply-accumulate operations
-    matrix                      // Volta tensor operations
-};
-
-/******************************************************************************
- * arch_family_t
- ******************************************************************************/
-
-/**
- * \brief Enumeration of NVIDIA GPU architectural families
- */
-struct arch_family_t
-{
-    /// \brief Enumerants
-    enum kind_t
-    {
-        Unsupported     = 0,
-        Kepler      = 3,
-        Maxwell     = 5,
-        Volta       = 7,
-    };
-
-    /// Enumerant value
-    kind_t kind;
-
-    /// Default constructor
-    arch_family_t() : kind(Unsupported) {}
-
-    /// Copy constructor
-    arch_family_t(const kind_t &other_kind) : kind(other_kind) {}
-
-    /// Cast to kind_t
-    operator kind_t() const { return kind; }
-
-    /// Returns the instance as a string
-    __host__ __device__ inline
-    char const* to_string() const
-    {
-        switch (kind)
-        {
-            case Kepler:    return "Kepler";
-            case Maxwell:   return "Maxwell";
-            case Volta:     return "Volta";
-            case Unsupported:
-            default: return "Unsupported";
-        }
-    }
-
-    /// Insert the formatted instance into the output stream
-    void print(std::ostream& out) const { out << to_string(); }
-
-};
-
-
-/**
- * Macro for architecture targeted by the current compiler pass
- */
-#if defined(__CUDA_ARCH__)
-    #define CUTLASS_ARCH __CUDA_ARCH__
-#else
-    #define CUTLASS_ARCH 0
-#endif
-
-
-/**
- * Macro for architecture family targeted by the current compiler pass
- */
-#define CUTLASS_ARCH_FAMILY                         \
-    (                                               \
-        (CUTLASS_ARCH < 300) ?                      \
-            arch_family_t::Unsupported :            \
-            (CUTLASS_ARCH < 500) ?                  \
-                arch_family_t::Kepler :             \
-                (CUTLASS_ARCH < 700) ?              \
-                    arch_family_t::Maxwell :        \
-                    arch_family_t::Volta            \
-    )
-
-
-
-
-/******************************************************************************
- * Device introspection
- ******************************************************************************/
-
-/**
- * Empty kernel for querying PTX manifest metadata (e.g., version) for the current device
- */
-template <typename T>
-__global__ void empty_kernel(void) { }
-
-
-
-/**
- * \brief Retrieves the PTX version that will be used on the current device (major * 100 + minor * 10)
- */
-cudaError_t ptx_version(int &version)
-{
-    struct Dummy
-    {
-        /// Type definition of the empty_kernel kernel entry point
-        typedef void (*EmptyKernelPtr)();
-
-        /// Force empty_kernel<void> to be generated if this class is used
-        EmptyKernelPtr Empty()
-        {
-            return empty_kernel<void>;
-        }
-    };
-
-    cudaError_t error = cudaSuccess;
-    do
-    {
-        cudaFuncAttributes empty_kernel_attrs;
-        if (CUDA_PERROR_DEBUG(error = cudaFuncGetAttributes(&empty_kernel_attrs, empty_kernel<void>))) break;
-        version = empty_kernel_attrs.ptxVersion * 10;
-    }
-    while (0);
-
-    return error;
-}
-
-
-/**
- * \brief Retrieves the SM version (major * 100 + minor * 10) for the current device
- */
-cudaError_t get_sm_version(int &sm_version)
-{
-    cudaError_t error = cudaSuccess;
-
-    // Get device ordinal
-    int device_ordinal;
-    if (CUDA_PERROR_DEBUG(error = cudaGetDevice(&device_ordinal)))
-        return error;
-
-    // Fill in SM version
-    int major, minor;
-    if (CUDA_PERROR_DEBUG(error = cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, device_ordinal)))
-        return error;
-    if (CUDA_PERROR_DEBUG(error = cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, device_ordinal)))
-        return error;
-    sm_version = major * 100 + minor * 10;
-
-    return error;
-}
-
-
-/**
- * \brief Retrieves the count for the current device
- */
-cudaError_t get_sm_count(int &sm_count)
-{
-    cudaError_t error = cudaSuccess;
-
-    // Get device ordinal
-    int device_ordinal;
-    if (CUDA_PERROR_DEBUG(error = cudaGetDevice(&device_ordinal)))
-        return error;
-
-    // Get SM count
-    if (CUDA_PERROR_DEBUG(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal)))
-        return error;
-
-    return error;
-}
-
-
-} // namespace cutlass
-
-

cutlass/util/io_intrinsics.h

@@ -1,492 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief I/O device intrinsics
- */
-
-#include <stdint.h>
-#include <cuda_fp16.h>
-
-#include "nv_std.h"
-#include "math.h"
-
-namespace cutlass {
-
-
-
-
-/******************************************************************************
- * io_vector
- ******************************************************************************/
-
-/**
- * Base aligned storage for IO vector
- */
-template <typename value_t, int VectorItems, int AlignBytes> struct io_vector_base;
-template <typename value_t, int VectorItems> struct __align__(1) io_vector_base<value_t, VectorItems, 1> { value_t buff[VectorItems]; };
-template <typename value_t, int VectorItems> struct __align__(2) io_vector_base<value_t, VectorItems, 2> { value_t buff[VectorItems]; };
-template <typename value_t, int VectorItems> struct __align__(4) io_vector_base<value_t, VectorItems, 4> { value_t buff[VectorItems]; };
-template <typename value_t, int VectorItems> struct __align__(8) io_vector_base<value_t, VectorItems, 8> { value_t buff[VectorItems]; };
-template <typename value_t, int VectorItems> struct __align__(16) io_vector_base<value_t, VectorItems, 16> { value_t buff[VectorItems]; };
-
-
-/**
- * \brief Aligned vector type for coarsening data movement instructions
- *
- * Exposes the member constant \p VectorItems, the actual number of component
- * values comprising the io_vector
- */
-template <
-    typename value_t,                                                           ///< Component value type
-    int MaxVectorItems,                                                         ///< Maximum allowable component values
-    int MaxAlignBytes                                                           ///< Maximum allowable alignment
-            = __NV_STD_MIN(16, MaxVectorItems * sizeof(value_t)),
-    int AlignBytes                                                              ///< Actual alignment
-            = __NV_STD_MIN(sizeof(value_t) * MaxVectorItems, MaxAlignBytes),
-    int VectorItems                                                             ///< Actual number of component values
-            = divide_assert<AlignBytes, sizeof(value_t)>::value,
-    bool MustAlias                                                              ///< Whether we need to alias during loads/stores
-            = (VectorItems > 4)>
-struct io_vector;
-
-
-/**
- * IO vector (specialization for VectorItems <= 4)
- */
-template <
-    typename value_t,
-    int MaxVectorItems,
-    int MaxAlignBytes,
-    int _AlignBytes,
-    int _VectorItems>
-struct io_vector <
-    value_t,
-    MaxVectorItems,
-    MaxAlignBytes,
-    _AlignBytes,
-    _VectorItems,
-    false>
-:
-    io_vector_base<value_t, _VectorItems, _AlignBytes>
-{
-    enum
-    {
-        VectorItems = _VectorItems,
-        AlignBytes = _AlignBytes
-    };
-
-    static_assert(is_pow2<AlignBytes>::value, "I/O vector alignment must be a power-of-two.");
-    static_assert((AlignBytes <= 16), "I/O vector alignment must <= 16B.");
-
-    inline __device__
-    void load(const io_vector *ptr)
-    {
-        *this = *ptr;
-    }
-
-    inline __device__
-    void load(const value_t *ptr)
-    {
-        *this = *reinterpret_cast<const io_vector*>(ptr);
-    }
-
-
-    inline __device__
-    void store(io_vector *ptr) const
-    {
-        *ptr = *this;
-    }
-
-    inline __device__
-    void store(value_t *ptr) const
-    {
-        *reinterpret_cast<io_vector*>(ptr) = *this;
-    }
-};
-
-
-/**
- * IO vector (specialization for VectorItems > 4)
- *
- * NB: Workaround for NVCC not generating 128-bit loads/stores for aligned
- * structures having component types < 32b
- */
-template <
-    typename value_t,
-    int MaxVectorItems,
-    int MaxAlignBytes,
-    int _AlignBytes,
-    int _VectorItems>
-struct io_vector <
-    value_t,
-    MaxVectorItems,
-    MaxAlignBytes,
-    _AlignBytes,
-    _VectorItems,
-    true>
-:
-    io_vector_base<value_t, _VectorItems, _AlignBytes>
-{
-    enum
-    {
-        VectorItems = _VectorItems,
-        AlignBytes = _AlignBytes
-    };
-
-    static_assert(is_pow2<AlignBytes>::value, "I/O vector alignment must be a power-of-two.");
-    static_assert((AlignBytes <= 16), "I/O vector alignment must <= 16B.");
-
-    typedef typename nv_std::conditional<(AlignBytes == 8),
-            uint2,  // Use 8B load
-            uint4>  // Use 16B load
-        ::type align_t;
-
-    inline __device__
-    void load(const io_vector *ptr)
-    {
-        *reinterpret_cast<align_t*>(this) = *reinterpret_cast<const align_t*>(ptr);
-    }
-
-    inline __device__
-    void load(const value_t *ptr)
-    {
-        *reinterpret_cast<align_t*>(this) = *reinterpret_cast<const align_t*>(ptr);
-    }
-
-
-    inline __device__
-    void store(io_vector *ptr) const
-    {
-        *reinterpret_cast<align_t*>(ptr) = *reinterpret_cast<const align_t*>(this);
-    }
-
-    inline __device__
-    void store(value_t *ptr) const
-    {
-        *reinterpret_cast<align_t*>(ptr) = *reinterpret_cast<const align_t*>(this);
-    }
-
-};
-
-
-
-
-
-
-
-/******************************************************************************
- * Macro expansions for vector loads
- ******************************************************************************/
-
-/**
- * Define vector-4 LD specialization for the given load modifier
- */
-#define CUTLASS_LD_V4(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint) \
-    template <typename ptr_t>                                                                   \
-    inline __device__                                                                           \
-    void f_name(                                                                                \
-        value_t (&dest)[4],                                                                     \
-        ptr_t ptr)                                                                              \
-    {                                                                                           \
-        asm volatile ("ld."#load_modifier".v4."#ptx_type" {%0, %1, %2, %3}, [%4];\n"            \
-            :                                                                                   \
-                "="#val_constraint(dest[0]),                                                    \
-                "="#val_constraint(dest[1]),                                                    \
-                "="#val_constraint(dest[2]),                                                    \
-                "="#val_constraint(dest[3])                                                     \
-            :                                                                                   \
-                #ptr_constraint(ptr));                                                          \
-    }
-
-/**
- * Define vector-2 LD specialization for the given load modifier
- */
-#define CUTLASS_LD_V2(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint) \
-    template <typename ptr_t>                                                                   \
-    inline __device__                                                                           \
-    void f_name(                                                                                \
-        value_t (&dest)[2],                                                                     \
-        ptr_t ptr)                                                                              \
-    {                                                                                           \
-        asm volatile ("ld."#load_modifier".v2."#ptx_type" {%0, %1}, [%2];\n"                    \
-            :                                                                                   \
-                "="#val_constraint(dest[0]),                                                    \
-                "="#val_constraint(dest[1])                                                     \
-            :                                                                                   \
-                #ptr_constraint(ptr));                                                          \
-    }
-
-
-/**
- * Define vector-1 LD specialization for the given load modifier
- */
-#define CUTLASS_LD_V1(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint) \
-    template <typename ptr_t>                                                                   \
-    inline __device__                                                                           \
-    void f_name(                                                                                \
-        value_t (&dest)[1],                                                                     \
-        ptr_t ptr)                                                                              \
-    {                                                                                           \
-        asm volatile ("ld."#load_modifier"."#ptx_type" %0, [%1];\n"                             \
-            :                                                                                   \
-                "="#val_constraint(dest[0])                                                     \
-            :                                                                                   \
-                #ptr_constraint(ptr));                                                          \
-    }
-
-
-/**
- * Define powers-of-two vector LD specializations
- */
-#define CUTLASS_LD_ALL(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)    \
-    CUTLASS_LD_V4(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)         \
-    CUTLASS_LD_V2(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)         \
-    CUTLASS_LD_V1(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)
-
-
-/******************************************************************************
- * Macro expansions for vector stores
- ******************************************************************************/
-
-/**
- * Define vector-4 ST specialization for the given load modifier
- */
-#define CUTLASS_ST_V4(f_name, value_t, store_modifier, ptx_type, val_constraint, ptr_constraint)    \
-    template <typename ptr_t>                                                                       \
-    inline __device__                                                                               \
-    void f_name(                                                                                    \
-        ptr_t ptr,                                                                                  \
-        const value_t (&src)[4])                                                                    \
-    {                                                                                               \
-        asm volatile ("st."#store_modifier".v4."#ptx_type" [%0], {%1, %2, %3, %4};\n"               \
-            : :                                                                                     \
-                #ptr_constraint(ptr),                                                               \
-                #val_constraint(src[0]),                                                            \
-                #val_constraint(src[1]),                                                            \
-                #val_constraint(src[2]),                                                            \
-                #val_constraint(src[3]));                                                           \
-    }
-
-
-/**
- * Define vector-2 ST specialization for the given load modifier
- */
-#define CUTLASS_ST_V2(f_name, value_t, store_modifier, ptx_type, val_constraint, ptr_constraint)    \
-    template <typename ptr_t>                                                                       \
-    inline __device__                                                                               \
-    void f_name(                                                                                    \
-        ptr_t ptr,                                                                                  \
-        const value_t (&src)[2])                                                                    \
-    {                                                                                               \
-        asm volatile ("st."#store_modifier".v2."#ptx_type" [%0], {%1, %2};\n"                       \
-            : :                                                                                     \
-                #ptr_constraint(ptr),                                                               \
-                #val_constraint(src[0]),                                                            \
-                #val_constraint(src[1]));                                                           \
-    }
-
-/**
- * Define vector-1 ST specialization for the given load modifier
- */
-#define CUTLASS_ST_V1(f_name, value_t, store_modifier, ptx_type, val_constraint, ptr_constraint)    \
-    template <typename ptr_t>                                                                       \
-    inline __device__                                                                               \
-    void f_name(                                                                                    \
-        ptr_t ptr,                                                                                  \
-        const value_t (&src)[1])                                                                    \
-    {                                                                                               \
-        asm volatile ("st."#store_modifier"."#ptx_type" [%0], %1;\n"                                \
-            : :                                                                                     \
-                #ptr_constraint(ptr),                                                               \
-                #val_constraint(src[0]));                                                           \
-    }
-
-
-/**
- * Define powers-of-two vector LD specializations
- */
-#define CUTLASS_ST_ALL(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)    \
-    CUTLASS_ST_V4(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)         \
-    CUTLASS_ST_V2(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)         \
-    CUTLASS_ST_V1(f_name, value_t, load_modifier, ptx_type, val_constraint, ptr_constraint)
-
-
-
-/******************************************************************************
- * Macro expansions for vector IO
- ******************************************************************************/
-
-/**
- * Define global and shared LD specializations
- */
-#define CUTLASS_IO(value_t, ptx_type, val_constraint)                                       \
-    CUTLASS_LD_ALL(ldg_cg_internal, value_t, global.cg, ptx_type, val_constraint, l)        \
-    CUTLASS_ST_ALL(stg_cg_internal, value_t, global.cg, ptx_type, val_constraint, l)
-
-
-// Define IO for useful types
-CUTLASS_IO(double,     f64, d)
-CUTLASS_IO(float,      f32, f)
-CUTLASS_IO(int64_t,    b64, l)
-CUTLASS_IO(int32_t,    b32, r)
-CUTLASS_IO(int16_t,    b16, h)
-
-
-// Macro cleanup
-#undef CUTLASS_IO
-#undef CUTLASS_LD_ALL
-#undef CUTLASS_LD_V4
-#undef CUTLASS_LD_V2
-#undef CUTLASS_LD_V1
-#undef CUTLASS_ST_ALL
-#undef CUTLASS_ST_V4
-#undef CUTLASS_ST_V2
-#undef CUTLASS_ST_V1
-
-
-/******************************************************************************
- * I/O cast types
- ******************************************************************************/
-
-/// Provides the type for which to reinterpret-cast a given vector
-template <
-    typename value_t,
-    int IoVecDim,
-    int ValueBytes = sizeof(value_t)>
-struct io_cast
-{
-    typedef value_t type[IoVecDim];
-};
-
-
-/// Provides the type for which to reinterpret-cast a vector of 1B types
-template <
-    typename value_t,
-    int IoVecDim>
-struct io_cast<value_t, IoVecDim, 1>
-{
-    typedef typename nv_std::conditional<
-            (IoVecDim < 2),
-            int8_t[1],                                 // Use 8b load
-            typename nv_std::conditional<
-                (IoVecDim < 4),
-                int16_t[1],                            // Use 16b load
-                int32_t[IoVecDim / 4]>::type>::type    // Use up to 128b load
-        type;
-};
-
-
-/// Provides the type for which to reinterpret-cast a vector of 2B types
-template <
-    typename value_t,
-    int IoVecDim>
-struct io_cast<value_t, IoVecDim, 2>
-{
-    typedef typename nv_std::conditional<
-            (IoVecDim < 2),
-            int16_t[1],                                // Use 16b load
-            int32_t[IoVecDim / 2]>::type               // Use up to 128b load
-        type;
-};
-
-
-
-/******************************************************************************
- * ldg_cg intrinsics
- ******************************************************************************/
-
-/// Load from global (cache-global modifier)
-template <typename value_t, typename ptr_t>
-inline __device__
-void ldg_cg(
-    value_t &dest,
-    ptr_t d_in)
-{
-    // Cast dest to a different array type if necessary
-    ldg_cg_internal(
-        reinterpret_cast<typename io_cast<value_t, 1>::type &>(dest),
-        d_in);
-}
-
-/// Load from global (cache-global modifier)
-template <typename value_t, int IoVecDim, typename ptr_t>
-inline __device__
-void ldg_cg(
-    value_t (&dest)[IoVecDim],
-    ptr_t d_in)
-{
-    static_assert(is_pow2<IoVecDim>::value, "I/O vectors must be a power-of-two.");
-
-    // Cast dest to a different array type if necessary
-    ldg_cg_internal(
-        reinterpret_cast<typename io_cast<value_t, IoVecDim>::type &>(dest),
-        d_in);
-}
-
-
-/******************************************************************************
- * stg_cg intrinsics
- ******************************************************************************/
-
-/// Store to global (cache-global modifier)
-template <typename ptr_t, typename value_t>
-inline __device__
-void stg_cg(
-    ptr_t dest,
-    const value_t &src)
-{
-    // Cast src to a different array type if necessary
-    stg_cg_internal(
-        dest,
-        reinterpret_cast<const typename io_cast<value_t, 1>::type &>(src));
-}
-
-/// Store to global (cache-global modifier)
-template <typename ptr_t, int IoVecDim, typename value_t>
-inline __device__
-void stg_cg(
-    ptr_t dest,
-    const value_t (&src)[IoVecDim])
-{
-    static_assert(is_pow2<IoVecDim>::value, "I/O vectors must be a power-of-two.");
-
-    // Cast src to a different array type if necessary
-    stg_cg_internal(
-        dest,
-        reinterpret_cast<const typename io_cast<value_t, IoVecDim>::type &>(src));
-}
-
-
-
-
-
-} // namespace cutlass
-

cutlass/util/math.h

@@ -1,167 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief Math utilities
- */
-
-#include "nv_std.h"
-
-namespace cutlass {
-
-
-/******************************************************************************
- * Static math utilities
- ******************************************************************************/
-
-/**
- * Statically determine if N is a power-of-two
- */
-template <int N>
-struct is_pow2 : nv_std::integral_constant<bool, (N & (N - 1)) == 0>
-{};
-
-
-
-
-
-/**
- * Statically determine log2(N), rounded down
- */
-template <int N, int CurrentVal = N, int Count = 0>
-struct log2_down
-{
-    /// Static logarithm value
-    enum { value = log2_down<N, (CurrentVal >> 1), Count + 1>::value };
-};
-
-// Base case
-template <int N, int Count>
-struct log2_down<N, 1, Count>
-{
-    enum { value = Count };
-};
-
-
-
-
-/**
- * Statically determine log2(N), rounded up
- */
-template <int N, int CurrentVal = N, int Count = 0>
-struct log2_up
-{
-    /// Static logarithm value
-    enum { value = log2_up<N, (CurrentVal >> 1), Count + 1>::value };
-};
-
-// Base case
-template <int N, int Count>
-struct log2_up<N, 1, Count>
-{
-    enum { value = ((1 << Count) < N) ? Count + 1 : Count };
-};
-
-
-
-/**
- * Statically estimate sqrt(N) to the nearest power-of-two
- */
-template <int N>
-struct sqrt_est
-{
-    enum { value = 1 << (log2_up<N>::value / 2) };
-};
-
-
-
-/**
- * For performing a constant-division with a compile-time assertion that the
- * Divisor evenly-divides the Dividend.
- */
-template <int Dividend, int Divisor>
-struct divide_assert
-{
-    enum { value =  Dividend / Divisor};
-
-    static_assert((Dividend % Divisor == 0), "Not an even multiple");
-};
-
-
-
-
-
-/******************************************************************************
- * Rounding
- ******************************************************************************/
-
-/**
- * Round dividend up to the nearest multiple of divisor
- */
-template <typename dividend_t, typename divisor_t>
-inline __host__ __device__
-dividend_t round_nearest(dividend_t dividend, divisor_t divisor)
-{
-    return ((dividend + divisor - 1) / divisor) * divisor;
-}
-
-
-/**
- * Greatest common divisor
- */
-template <typename value_t>
-inline __host__ __device__
-value_t gcd(value_t a, value_t b)
-{
-    for (;;)
-    {
-        if (a == 0) return b;
-        b %= a;
-        if (b == 0) return a;
-        a %= b;
-    }
-}
-
-
-/**
- * Least common multiple
- */
-template <typename value_t>
-inline __host__ __device__
-value_t lcm(value_t a, value_t b)
-{
-    value_t temp = gcd(a, b);
-
-    return temp ? (a / temp * b) : 0;
-}
-
-
-} // namespace cutlass
-

cutlass/util/matrix_transform.h

@@ -1,102 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief Enumeration of dense matrix view transformations
- */
-
-#include "printable.h"
-
-namespace cutlass {
-
-
-/******************************************************************************
- * matrix_transform_t
- ******************************************************************************/
-
-/**
- * \brief Enumeration of dense matrix view transformations
- *
- * These enumerators (and corresponding tag types) describe which view
- * transformation needs to be applied prior to operation upon a given dense
- * matrix. Its values correspond to Fortran characters 'n' (non-transpose),
- * 't'(transpose) and 'c'(conjugate transpose) that are often
- * used as parameters to legacy BLAS implementations
- */
-struct matrix_transform_t : printable_t
-{
-    /// \brief Enumerants (same as CUBLAS)
-    enum kind_t
-    {
-        /// Invalid view
-        Invalid             = -1,
-
-        /// Non-transpose view
-        NonTranspose        = 0,
-
-        /// Transpose view
-        Transpose           = 1,
-
-        /// Conjugate transpose view
-        ConjugateTranpose   = 2,
-    };
-
-    /// Enumerant value
-    kind_t kind;
-
-    /// Default constructor
-    matrix_transform_t() : kind(Invalid) {}
-
-    /// Copy constructor
-    matrix_transform_t(const kind_t &other_kind) : kind(other_kind) {}
-
-    /// Cast to kind_t
-    operator kind_t() const { return kind; }
-
-    /// Returns the instance as a string
-    __host__ __device__ inline
-    char const* to_string() const
-    {
-        switch (kind)
-        {
-            case NonTranspose:      return "NonTranspose";
-            case Transpose:         return "Transpose";
-            case ConjugateTranpose: return "ConjugateTranpose";
-            default:                return "Invalid";
-        }
-    }
-
-    /// Insert the formatted instance into the output stream
-    void print(std::ostream& out) const { out << to_string(); }
-
-};
-
-
-} // namespace cutlass

cutlass/util/numeric_types.h

@@ -0,0 +1,47 @@
+/***************************************************************************************************
+ * Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*!
+    \file
+    \brief
+*/
+#pragma once
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+//
+// Definitions for 1-bit binary and 4-bit integer types
+//
+
+struct bin1_t {};  // 1-bit binary type
+
+struct int4_t {};  // 4-bit signed integer type
+
+struct uint4_t {};  // 4-bit unsigned integer type
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/util/nv_std.h

@@ -1,705 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief C++ features that may be otherwise unimplemented for CUDA device functions.
- *
- * This file has three components:
- *
- *   (1) Macros:
- *       - Empty macro defines for C++ keywords not supported by the current
- *         version of C++. These simply allow compilation to proceed (but do
- *         not provide the added semantics).
- *           - \p noexcept
- *           - \p constexpr
- *           - \p nullptr
- *           - \p static_assert
- *
- *       - Macro functions that we need in constant expressions because the
- *         C++ equivalents require constexpr compiler support.  These are
- *         prefixed with \p __NV_STD_*
- *           - \p __NV_STD_MAX
- *           - \p __NV_STD_MIN
- *
- *   (2) Re-implementations of STL functions and types:
- *       - C++ features that need the \p __device__ annotation.  These are
- *         placed into the \p nv_std namespace.
- *           - \p plus
- *           - \p less
- *           - \p greater
- *           - \p min
- *           - \p max
- *           - \p methods on std::pair (==, !=, <, <=, >, >=, and make_pair())
- *
- *   (3) Stop-gap implementations of unsupported STL functions and types:
- *       - STL functions and types defined by C++ 11/14/17/etc. that are not
- *         provided by the current version of C++. These are placed into the
- *         \p nv_std namespace
- *           - \p integral_constant
- *           - \p nullptr_t
- *           - \p true_type
- *           - \p false_type
- *           - \p bool_constant
- *           - \p enable_if
- *           - \p conditional
- *           - \p is_same
- *           - \p is_base_of
- *           - \p remove_const
- *           - \p remove_volatile
- *           - \p remove_cv
- *           - \p is_volatile
- *           - \p is_pointer
- *           - \p is_void
- *           - \p is_integral
- *           - \p is_floating_point
- *           - \p is_arithmetic
- *           - \p is_fundamental
- *           - \p is_trivially_copyable
- *           - \p alignment_of
- *           - \p aligned_storage
- *
- *   (4) Functions and types that are STL-like (but aren't in the STL):
- *           - \p TODO: min and max functors?
- *
- * The idea is that, as we drop support for older compilers, we can simply #define
- * the \p __NV_STD_XYZ macros and \p nv_std namespace to alias their C++
- * counterparts (or trivially find-and-replace their occurrences in code text).
- */
-
-
-//-----------------------------------------------------------------------------
-// Include STL files that nv_std provides functionality for
-//-----------------------------------------------------------------------------
-
-#include <cstddef>          // nullptr_t
-#include <algorithm>        // Minimum/maximum operations
-#include <functional>       // Arithmetic operations
-#include <utility>          // For methods on std::pair
-#if (!defined(_MSC_VER) && (__cplusplus >= 201103L)) || (defined(_MSC_VER) && (_MS_VER >= 1500))
-    #include <type_traits>  // For integral constants, conditional metaprogramming, and type traits
-#endif
-
-
-
-/******************************************************************************
- * Macros
- ******************************************************************************/
-//-----------------------------------------------------------------------------
-// Keywords
-//-----------------------------------------------------------------------------
-
-/// noexcept, constexpr
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1900))
-    #ifndef noexcept
-        #define noexcept
-    #endif
-    #ifndef constexpr
-        #define constexpr
-    #endif
-#endif
-
-/// nullptr
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1310 ))
-    #ifndef nullptr
-        #define nullptr 0
-    #endif
-#endif
-
-/// static_assert
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1600 ))
-    #ifndef static_assert
-        #define __nv_std_cat_(a, b) a ## b
-        #define __nv_std_cat(a, b) __nv_std_cat_(a, b)
-        #define static_assert(__e, __m) typedef int __nv_std_cat(AsSeRt, __LINE__)[(__e) ? 1 : -1]
-    #endif
-#endif
-
-
-//-----------------------------------------------------------------------------
-// Functions
-//-----------------------------------------------------------------------------
-
-/// Select maximum(a, b)
-#ifndef __NV_STD_MAX
-    #define __NV_STD_MAX(a, b) (((b) > (a)) ? (b) : (a))
-#endif
-
-/// Select minimum(a, b)
-#ifndef __NV_STD_MIN
-    #define __NV_STD_MIN(a, b) (((b) < (a)) ? (b) : (a))
-#endif
-
-
-
-
-/******************************************************************************
- * Re-implementations
- ******************************************************************************/
-
-namespace nv_std {
-
-    //-----------------------------------------------------------------------------
-    // Arithmetic operations, comparisons <functional>
-    //-----------------------------------------------------------------------------
-
-    /// nv_std::plus
-    template <typename T>
-    struct plus
-    {
-        inline __host__ __device__
-        constexpr T operator()(const T &lhs, const T &rhs) const
-        {
-            return lhs + rhs;
-        }
-    };
-
-
-    /// std::less
-    template <typename T>
-    struct less
-    {
-        inline __host__ __device__
-        constexpr bool operator()(const T &lhs, const T &rhs) const
-        {
-            return lhs < rhs;
-        }
-    };
-
-    /// std::greater
-    template <typename T>
-    struct greater
-    {
-        inline __host__ __device__
-        constexpr bool operator()(const T &lhs, const T &rhs) const
-        {
-            return lhs > rhs;
-        }
-    };
-
-
-    //-----------------------------------------------------------------------------
-    // Minimum/maximum operations <algorithm>
-    //-----------------------------------------------------------------------------
-
-    /// std::min
-    template <typename T>
-    inline __host__ __device__
-    constexpr const T& min(
-        const T& a,
-        const T& b)
-    {
-        return (b < a) ? b : a;
-    }
-
-    /// std::max
-    template <typename T>
-    inline __host__ __device__
-    constexpr const T& max(
-        const T& a,
-        const T& b)
-    {
-        return (a < b) ? b : a;
-    }
-
-
-    //-----------------------------------------------------------------------------
-    // Methods on std::pair
-    //-----------------------------------------------------------------------------
-
-    using std::pair;
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    constexpr bool operator==( const pair<T1,T2>& lhs, const pair<T1,T2>& rhs )
-    {
-        return (lhs.first == rhs.first) && (lhs.second == rhs.second);
-    }
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    constexpr bool operator!=( const pair<T1,T2>& lhs, const pair<T1,T2>& rhs )
-    {
-        return (lhs.first != rhs.first) && (lhs.second != rhs.second);
-    }
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    constexpr bool operator<( const pair<T1,T2>& lhs, const pair<T1,T2>& rhs )
-    {
-        return (lhs.first < rhs.first) ?
-            true :
-            (rhs.first < lhs.first) ?
-                false :
-                (lhs.second < rhs.second);
-    }
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    constexpr bool operator<=( const pair<T1,T2>& lhs, const pair<T1,T2>& rhs )
-    {
-        return !(rhs < lhs);
-    }
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    constexpr bool operator>( const pair<T1,T2>& lhs, const pair<T1,T2>& rhs )
-    {
-        return (rhs < lhs);
-    }
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    constexpr bool operator>=( const pair<T1,T2>& lhs, const pair<T1,T2>& rhs )
-    {
-        return !(lhs < rhs);
-    }
-
-    template< class T1, class T2 >
-    inline __host__ __device__
-    std::pair<T1,T2> make_pair( T1 t, T2 u )
-    {
-        std::pair<T1,T2> retval;
-        retval.first = t;
-        retval.second = u;
-        return retval;
-    }
-
-}   // namespace nv_std
-
-
-
-/******************************************************************************
- * Implementations of C++ 11/14/17/... STL features
- ******************************************************************************/
-
-namespace nv_std {
-
-//-----------------------------------------------------------------------------
-// Integral constant helper types <type_traits>
-//-----------------------------------------------------------------------------
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
-
-    /// std::integral_constant
-    template <typename value_t, value_t V>
-    struct integral_constant;
-
-    /// std::integral_constant
-    template <typename value_t, value_t V>
-    struct integral_constant
-    {
-        static const value_t value = V;
-
-        typedef value_t                           value_type;
-        typedef integral_constant<value_t, V>     type;
-
-        inline __host__ __device__ operator value_type() const
-        {
-             return value;
-        }
-
-        inline __host__ __device__ const value_type operator()() const
-        {
-            return value;
-        }
-    };
-
-
-#else
-
-    using std::integral_constant;
-    using std::pair;
-
-#endif
-
-    /// The type used as a compile-time boolean with true value.
-    typedef integral_constant<bool, true>   true_type;
-
-    /// The type used as a compile-time boolean with false value.
-    typedef integral_constant<bool, false>  false_type;
-
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201402L)) || (defined(_MSC_VER) && (_MSC_VER < 1900))
-
-    /// std::bool_constant
-    template <bool V>
-    struct bool_constant : nv_std::integral_constant<bool, V>
-    {};
-
-#else
-
-    using std::bool_constant;
-
-#endif
-
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1700))
-
-    /// std::nullptr_t
-    struct nullptr_t {};
-
-#else
-
-    using std::nullptr_t;
-
-#endif
-
-
-
-    //-----------------------------------------------------------------------------
-    // Conditional metaprogramming <type_traits>
-    //-----------------------------------------------------------------------------
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1600))
-
-    /// std::enable_if (true specialization)
-    template<bool C, typename T = void>
-    struct enable_if {
-      typedef T type;
-    };
-
-    /// std::enable_if (false specialization)
-    template<typename T>
-    struct enable_if<false, T> { };
-
-
-    /// std::conditional (true specialization)
-    template<bool B, class T, class F>
-    struct conditional { typedef T type; };
-
-    /// std::conditional (false specialization)
-    template<class T, class F>
-    struct conditional<false, T, F> { typedef F type; };
-
-#else
-
-    using std::enable_if;
-    using std::conditional;
-
-#endif
-
-
-
-    //-----------------------------------------------------------------------------
-    // Const/volatility specifiers <type_traits>
-    //-----------------------------------------------------------------------------
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
-
-    /// std::remove_const (non-const specialization)
-    template <typename T> struct remove_const                   { typedef T type; };
-
-    /// std::remove_const (const specialization)
-    template <typename T> struct remove_const<const T>          { typedef T type; };
-
-
-
-    /// std::remove_volatile (non-volatile specialization)
-    template <typename T> struct remove_volatile                { typedef T type; };
-
-    /// std::remove_volatile (volatile specialization)
-    template <typename T> struct remove_volatile<volatile T>    { typedef T type; };
-
-
-
-    /// std::remove_cv
-    template <typename T>
-    struct remove_cv {
-        typedef typename remove_volatile<typename remove_const<T>::type>::type type;
-    };
-
-#else
-
-    using std::remove_const;
-    using std::remove_volatile;
-    using std::remove_cv;
-
-#endif
-
-
-    //-----------------------------------------------------------------------------
-    // Type relationships <type_traits>
-    //-----------------------------------------------------------------------------
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
-
-    /// std::is_same (false specialization)
-    template <typename A, typename B>
-    struct is_same : false_type
-    {};
-
-    /// std::is_same (true specialization)
-    template <typename A>
-    struct is_same<A, A> : true_type
-    {};
-
-
-    /// Helper for std::is_base_of
-    template<typename BaseT, typename DerivedT>
-    struct is_base_of_helper
-    {
-        typedef char (&yes)[1];
-        typedef char (&no)[2];
-
-        template<typename B, typename D>
-        struct dummy
-        {
-            operator B*() const;
-            operator D*();
-        };
-
-        template<typename T>
-        static yes check(DerivedT*, T);
-
-        static no check(BaseT*, int);
-
-        static const bool value = sizeof(check(dummy<BaseT, DerivedT>(), int())) == sizeof(yes);
-    };
-
-    /// std::is_base_of
-    template <typename BaseT, typename DerivedT>
-    struct is_base_of : integral_constant<
-        bool,
-        (is_base_of_helper<typename remove_cv<BaseT>::type, typename remove_cv<DerivedT>::type>::value) ||
-            (is_same<typename remove_cv<BaseT>::type, typename remove_cv<DerivedT>::type>::value)>
-    {};
-
-
-#else
-
-    using std::is_same;
-    using std::is_base_of;
-
-#endif
-
-
-
-    //-----------------------------------------------------------------------------
-    // Type properties <type_traits>
-    //-----------------------------------------------------------------------------
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
-
-    /// std::is_volatile
-    template <typename T> struct is_volatile             : false_type {};
-    template <typename T> struct is_volatile<volatile T> : true_type {};
-
-
-    /// Helper for std::is_pointer (false specialization)
-    template <typename T> struct is_pointer_helper        : false_type {};
-
-    /// Helper for std::is_pointer (true specialization)
-    template <typename T> struct is_pointer_helper<T*>    : true_type {};
-
-    /// std::is_pointer
-    template <typename T> struct is_pointer : is_pointer_helper<typename remove_cv<T>::type> {};
-
-
-
-    /// std::is_void
-    template <typename T>
-    struct is_void : is_same<void, typename remove_cv<T>::type>
-    {};
-
-
-
-    /// std::is_integral
-    template <typename T> struct is_integral            : false_type {};
-    template <> struct is_integral<char>                : true_type {};
-    template <> struct is_integral<signed char>         : true_type {};
-    template <> struct is_integral<unsigned char>       : true_type {};
-    template <> struct is_integral<short>               : true_type {};
-    template <> struct is_integral<unsigned short>      : true_type {};
-    template <> struct is_integral<int>                 : true_type {};
-    template <> struct is_integral<unsigned int>        : true_type {};
-    template <> struct is_integral<long>                : true_type {};
-    template <> struct is_integral<unsigned long>       : true_type {};
-    template <> struct is_integral<long long>           : true_type {};
-    template <> struct is_integral<unsigned long long>  : true_type {};
-    template <typename T> struct is_integral<volatile T>        : is_integral<T> {};
-    template <typename T> struct is_integral<const T>           : is_integral<T> {};
-    template <typename T> struct is_integral<const volatile T>  : is_integral<T> {};
-
-
-
-    /// std::is_floating_point
-    template <typename T>
-    struct is_floating_point : integral_constant<
-        bool,
-        (is_same<float, typename remove_cv<T>::type>::value ||
-            is_same<double, typename remove_cv<T>::type>::value)>
-    {};
-
-
-
-    /// std::is_arithmetic
-    template <typename T>
-    struct is_arithmetic :
-        integral_constant<bool, (is_integral<T>::value || is_floating_point<T>::value)>
-    {};
-
-
-    /// std::is_fundamental
-    template <typename T>
-    struct is_fundamental : integral_constant<
-        bool, (is_arithmetic<T>::value ||
-            is_void<T>::value ||
-            is_same<nullptr_t, typename remove_cv<T>::type>::value)>
-    {};
-
-
-
-
-#else
-
-    using std::is_volatile;
-    using std::is_pointer;
-    using std::is_void;
-    using std::is_integral;
-    using std::is_floating_point;
-    using std::is_arithmetic;
-    using std::is_fundamental;
-
-#endif
-
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) ||      \
-    (defined(_MSC_VER) && (_MSC_VER < 1800)) ||             \
-    (defined(__GNUG__) && (__GNUC__ < 5))
-
-    /**
-     * std::is_trivially_copyable
-     *
-     * This implementation only evaluates true if T is fundamental or pointer
-     *
-     * Without help from partial template specializations provided by the user for
-     * a specific class or struct, this trait will never report that the specified
-     * class or struct  is trivially-copyable ; this is always safe,
-     * if possibly sub-optimal.
-     */
-    template <typename T>
-    struct is_trivially_copyable :
-        integral_constant<bool, (is_fundamental<T>::value || is_pointer<T>::value)>
-    {};
-
-#else
-
-    using std::is_trivially_copyable;
-
-#endif
-
-
-
-
-    //-----------------------------------------------------------------------------
-    // Alignment and layout utilities
-    //-----------------------------------------------------------------------------
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
-
-
-    /// std::alignment_of
-    template <typename value_t>
-    struct alignment_of
-    {
-        struct pad
-        {
-            value_t val;
-            char    byte;
-        };
-
-        enum
-        {
-            value = sizeof(pad) - sizeof(value_t)
-        };
-    };
-
-#else
-
-    template <typename value_t>
-    struct alignment_of : std::alignment_of<value_t> {};
-
-#endif
-
-    /* 16B specializations where 32-bit Win32 host compiler disagrees with device compiler */
-    template <> struct alignment_of<int4>                 { enum { value = 16 }; };
-    template <> struct alignment_of<uint4>                { enum { value = 16 }; };
-    template <> struct alignment_of<float4>               { enum { value = 16 }; };
-    template <> struct alignment_of<long4>                { enum { value = 16 }; };
-    template <> struct alignment_of<ulong4>               { enum { value = 16 }; };
-    template <> struct alignment_of<longlong2>            { enum { value = 16 }; };
-    template <> struct alignment_of<ulonglong2>           { enum { value = 16 }; };
-    template <> struct alignment_of<double2>              { enum { value = 16 }; };
-    template <> struct alignment_of<longlong4>            { enum { value = 16 }; };
-    template <> struct alignment_of<ulonglong4>           { enum { value = 16 }; };
-    template <> struct alignment_of<double4>              { enum { value = 16 }; };
-
-    // Specializations for volatile/const qualified types
-    template <typename value_t> struct alignment_of<volatile value_t>       : alignment_of<value_t> {};
-    template <typename value_t> struct alignment_of<const value_t>          : alignment_of<value_t> {};
-    template <typename value_t> struct alignment_of<const volatile value_t> : alignment_of<value_t> {};
-
-
-
-#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1800))
-
-    template<size_t Align> struct aligned_chunk;
-    template<> struct __align__(1) aligned_chunk<1> { uint8_t buff; };
-    template<> struct __align__(2) aligned_chunk<2> { uint16_t buff; };
-    template<> struct __align__(4) aligned_chunk<4> { uint32_t buff; };
-    template<> struct __align__(8) aligned_chunk<8> { uint32_t buff[2]; };
-    template<> struct __align__(16) aligned_chunk<16> { uint32_t buff[4]; };
-    template<> struct __align__(32) aligned_chunk<32> { uint32_t buff[8]; };
-    template<> struct __align__(64) aligned_chunk<64> { uint32_t buff[16]; };
-    template<> struct __align__(128) aligned_chunk<128> { uint32_t buff[32]; };
-    template<> struct __align__(256) aligned_chunk<256> { uint32_t buff[64]; };
-    template<> struct __align__(512) aligned_chunk<512> { uint32_t buff[128]; };
-    template<> struct __align__(1024) aligned_chunk<1024> { uint32_t buff[256]; };
-    template<> struct __align__(2048) aligned_chunk<2048> { uint32_t buff[512]; };
-    template<> struct __align__(4096) aligned_chunk<4096> { uint32_t buff[1024]; };
-
-    /// std::aligned_storage
-    template <size_t Len, size_t Align>
-    struct aligned_storage
-    {
-        typedef aligned_chunk<Align> type[Len / sizeof(aligned_chunk<Align>)];
-    };
-
-#else
-
-    using std::aligned_storage;
-
-#endif
-
-
-
-
-}; // namespace nv_std
-

cutlass/util/platform.h

@@ -0,0 +1,809 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+/**
+ * \file
+ * \brief C++ features that may be otherwise unimplemented for CUDA device functions.
+ *
+ * This file has three components:
+ *
+ *   (1) Macros:
+ *       - Empty macro defines for C++ keywords not supported by the current
+ *         version of C++. These simply allow compilation to proceed (but do
+ *         not provide the added semantics).
+ *           - \p noexcept
+ *           - \p constexpr
+ *           - \p nullptr
+ *           - \p static_assert
+ *
+ *       - Macro functions that we need in constant expressions because the
+ *         C++ equivalents require constexpr compiler support.  These are
+ *         prefixed with \p __NV_STD_*
+ *           - \p __NV_STD_MAX
+ *           - \p __NV_STD_MIN
+ *
+ *   (2) Re-implementations of STL functions and types:
+ *       - C++ features that need the \p __device__ annotation.  These are
+ *         placed into the \p platform namespace.
+ *           - \p plus
+ *           - \p less
+ *           - \p greater
+ *           - \p min
+ *           - \p max
+ *           - \p methods on std::pair (==, !=, <, <=, >, >=, and make_pair())
+ *
+ *   (3) Stop-gap implementations of unsupported STL functions and types:
+ *       - STL functions and types defined by C++ 11/14/17/etc. that are not
+ *         provided by the current version of C++. These are placed into the
+ *         \p platform namespace
+ *           - \p integral_constant
+ *           - \p nullptr_t
+ *           - \p true_type
+ *           - \p false_type
+ *           - \p bool_constant
+ *           - \p enable_if
+ *           - \p conditional
+ *           - \p is_same
+ *           - \p is_base_of
+ *           - \p remove_const
+ *           - \p remove_volatile
+ *           - \p remove_cv
+ *           - \p is_volatile
+ *           - \p is_pointer
+ *           - \p is_void
+ *           - \p is_integral
+ *           - \p is_floating_point
+ *           - \p is_arithmetic
+ *           - \p is_fundamental
+ *           - \p is_trivially_copyable
+ *           - \p alignment_of
+ *           - \p aligned_storage
+ *
+ *   (4) Functions and types that are STL-like (but aren't in the STL):
+ *           - \p TODO: min and max functors?
+ *
+ * The idea is that, as we drop support for older compilers, we can simply #define
+ * the \p __NV_STD_XYZ macros and \p platform namespace to alias their C++
+ * counterparts (or trivially find-and-replace their occurrences in code text).
+ */
+
+//-----------------------------------------------------------------------------
+// Dependencies
+//-----------------------------------------------------------------------------
+
+#include <stdint.h>
+
+#if !defined(__CUDACC_RTC__)
+//-----------------------------------------------------------------------------
+// Include STL files that platform provides functionality for
+//-----------------------------------------------------------------------------
+
+#include <algorithm>   // Minimum/maximum operations
+#include <cstddef>     // nullptr_t
+#include <functional>  // Arithmetic operations
+#include <utility>     // For methods on std::pair
+#if (!defined(_MSC_VER) && (__cplusplus >= 201103L)) || (defined(_MSC_VER) && (_MS_VER >= 1500))
+#include <type_traits>  // For integral constants, conditional metaprogramming, and type traits
+#endif
+
+#include "cutlass/cutlass.h"
+
+#endif
+
+//-----------------------------------------------------------------------------
+// OS
+//-----------------------------------------------------------------------------
+#if defined(WIN32) || defined(_WIN32) || defined(__WIN32) && !defined(__CYGWIN__)
+#define CUTLASS_OS_WINDOWS
+#endif
+
+/******************************************************************************
+ * Macros
+ ******************************************************************************/
+//-----------------------------------------------------------------------------
+// Keywords
+//-----------------------------------------------------------------------------
+
+/// noexcept, constexpr
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1900))
+#ifndef noexcept
+#define noexcept
+#endif
+#ifndef constexpr
+#define constexpr
+#endif
+#endif
+
+/// nullptr
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1310))
+#ifndef nullptr
+#define nullptr 0
+#endif
+#endif
+
+/// static_assert
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+#ifndef static_assert
+#define __platform_cat_(a, b) a##b
+#define __platform_cat(a, b) __platform_cat_(a, b)
+#define static_assert(__e, __m) typedef int __platform_cat(AsSeRt, __LINE__)[(__e) ? 1 : -1]
+#endif
+#endif
+
+//-----------------------------------------------------------------------------
+// Functions
+//-----------------------------------------------------------------------------
+
+/// Select maximum(a, b)
+#ifndef __NV_STD_MAX
+#define __NV_STD_MAX(a, b) (((b) > (a)) ? (b) : (a))
+#endif
+
+/// Select minimum(a, b)
+#ifndef __NV_STD_MIN
+#define __NV_STD_MIN(a, b) (((b) < (a)) ? (b) : (a))
+#endif
+
+/******************************************************************************
+ * Re-implementations
+ ******************************************************************************/
+namespace cutlass {
+namespace platform {
+
+//-----------------------------------------------------------------------------
+// Arithmetic operations, comparisons <functional>
+//-----------------------------------------------------------------------------
+
+/// platform::plus
+template <typename T>
+struct plus {
+  CUTLASS_HOST_DEVICE constexpr T operator()(const T& lhs, const T& rhs) const { return lhs + rhs; }
+};
+
+/// std::less
+template <typename T>
+struct less {
+  CUTLASS_HOST_DEVICE constexpr bool operator()(const T& lhs, const T& rhs) const {
+    return lhs < rhs;
+  }
+};
+
+/// std::greater
+template <typename T>
+struct greater {
+  CUTLASS_HOST_DEVICE constexpr bool operator()(const T& lhs, const T& rhs) const {
+    return lhs > rhs;
+  }
+};
+
+//-----------------------------------------------------------------------------
+// Minimum/maximum operations <algorithm>
+//-----------------------------------------------------------------------------
+
+/// std::min
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr const T& min(const T& a, const T& b) {
+  return (b < a) ? b : a;
+}
+
+/// std::max
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr const T& max(const T& a, const T& b) {
+  return (a < b) ? b : a;
+}
+
+#if !defined(__CUDACC_RTC__)
+//-----------------------------------------------------------------------------
+// Methods on std::pair
+//-----------------------------------------------------------------------------
+
+using std::pair;
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE constexpr bool operator==(const pair<T1, T2>& lhs, const pair<T1, T2>& rhs) {
+  return (lhs.first == rhs.first) && (lhs.second == rhs.second);
+}
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE constexpr bool operator!=(const pair<T1, T2>& lhs, const pair<T1, T2>& rhs) {
+  return (lhs.first != rhs.first) && (lhs.second != rhs.second);
+}
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE constexpr bool operator<(const pair<T1, T2>& lhs, const pair<T1, T2>& rhs) {
+  return (lhs.first < rhs.first) ? true : (rhs.first < lhs.first) ? false
+                                                                  : (lhs.second < rhs.second);
+}
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE constexpr bool operator<=(const pair<T1, T2>& lhs, const pair<T1, T2>& rhs) {
+  return !(rhs < lhs);
+}
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE constexpr bool operator>(const pair<T1, T2>& lhs, const pair<T1, T2>& rhs) {
+  return (rhs < lhs);
+}
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE constexpr bool operator>=(const pair<T1, T2>& lhs, const pair<T1, T2>& rhs) {
+  return !(lhs < rhs);
+}
+
+template <class T1, class T2>
+CUTLASS_HOST_DEVICE std::pair<T1, T2> make_pair(T1 t, T2 u) {
+  std::pair<T1, T2> retval;
+  retval.first = t;
+  retval.second = u;
+  return retval;
+}
+#endif
+
+}  // namespace platform
+
+/******************************************************************************
+ * Implementations of C++ 11/14/17/... STL features
+ ******************************************************************************/
+
+namespace platform {
+
+//-----------------------------------------------------------------------------
+// Integral constant helper types <type_traits>
+//-----------------------------------------------------------------------------
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
+
+/// std::integral_constant
+template <typename value_t, value_t V>
+struct integral_constant;
+
+/// std::integral_constant
+template <typename value_t, value_t V>
+struct integral_constant {
+  static const value_t value = V;
+
+  typedef value_t value_type;
+  typedef integral_constant<value_t, V> type;
+
+  CUTLASS_HOST_DEVICE operator value_type() const { return value; }
+
+  CUTLASS_HOST_DEVICE const value_type operator()() const { return value; }
+};
+
+#else
+
+using std::integral_constant;
+using std::pair;
+
+#endif
+
+/// The type used as a compile-time boolean with true value.
+typedef integral_constant<bool, true> true_type;
+
+/// The type used as a compile-time boolean with false value.
+typedef integral_constant<bool, false> false_type;
+
+#if (!defined(_MSC_VER) && (__cplusplus <= 201402L)) || (defined(_MSC_VER) && (_MSC_VER < 1900))
+
+/// std::bool_constant
+template <bool V>
+struct bool_constant : platform::integral_constant<bool, V> {};
+
+#else
+
+using std::bool_constant;
+
+#endif
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1700))
+
+/// std::nullptr_t
+struct nullptr_t {};
+
+#else
+
+using std::nullptr_t;
+
+#endif
+
+//-----------------------------------------------------------------------------
+// Conditional metaprogramming <type_traits>
+//-----------------------------------------------------------------------------
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+
+/// std::enable_if (true specialization)
+template <bool C, typename T = void>
+struct enable_if {
+  typedef T type;
+};
+
+/// std::enable_if (false specialization)
+template <typename T>
+struct enable_if<false, T> {};
+
+/// std::conditional (true specialization)
+template <bool B, class T, class F>
+struct conditional {
+  typedef T type;
+};
+
+/// std::conditional (false specialization)
+template <class T, class F>
+struct conditional<false, T, F> {
+  typedef F type;
+};
+
+#else
+
+using std::enable_if;
+using std::conditional;
+
+#endif
+
+//-----------------------------------------------------------------------------
+// Const/volatility specifiers <type_traits>
+//-----------------------------------------------------------------------------
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
+
+/// std::remove_const (non-const specialization)
+template <typename T>
+struct remove_const {
+  typedef T type;
+};
+
+/// std::remove_const (const specialization)
+template <typename T>
+struct remove_const<const T> {
+  typedef T type;
+};
+
+/// std::remove_volatile (non-volatile specialization)
+template <typename T>
+struct remove_volatile {
+  typedef T type;
+};
+
+/// std::remove_volatile (volatile specialization)
+template <typename T>
+struct remove_volatile<volatile T> {
+  typedef T type;
+};
+
+/// std::remove_cv
+template <typename T>
+struct remove_cv {
+  typedef typename remove_volatile<typename remove_const<T>::type>::type type;
+};
+
+#else
+
+using std::remove_const;
+using std::remove_volatile;
+using std::remove_cv;
+
+#endif
+
+//-----------------------------------------------------------------------------
+// Type relationships <type_traits>
+//-----------------------------------------------------------------------------
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
+
+/// std::is_same (false specialization)
+template <typename A, typename B>
+struct is_same : false_type {};
+
+/// std::is_same (true specialization)
+template <typename A>
+struct is_same<A, A> : true_type {};
+
+/// Helper for std::is_base_of
+template <typename BaseT, typename DerivedT>
+struct is_base_of_helper {
+  typedef char (&yes)[1];
+  typedef char (&no)[2];
+
+  template <typename B, typename D>
+  struct dummy {
+    CUTLASS_HOST_DEVICE operator B*() const;
+    CUTLASS_HOST_DEVICE operator D*();
+  };
+
+  template <typename T>
+  CUTLASS_HOST_DEVICE static yes check(DerivedT*, T);
+
+  CUTLASS_HOST_DEVICE static no check(BaseT*, int);
+
+  static const bool value = sizeof(check(dummy<BaseT, DerivedT>(), int())) == sizeof(yes);
+};
+
+/// std::is_base_of
+template <typename BaseT, typename DerivedT>
+struct is_base_of
+    : integral_constant<bool,
+                        (is_base_of_helper<typename remove_cv<BaseT>::type,
+                                           typename remove_cv<DerivedT>::type>::value) ||
+                            (is_same<typename remove_cv<BaseT>::type,
+                                     typename remove_cv<DerivedT>::type>::value)> {};
+
+#else
+
+using std::is_same;
+using std::is_base_of;
+
+#endif
+
+//-----------------------------------------------------------------------------
+// Type properties <type_traits>
+//-----------------------------------------------------------------------------
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
+
+/// std::is_volatile
+template <typename T>
+struct is_volatile : false_type {};
+template <typename T>
+struct is_volatile<volatile T> : true_type {};
+
+/// Helper for std::is_pointer (false specialization)
+template <typename T>
+struct is_pointer_helper : false_type {};
+
+/// Helper for std::is_pointer (true specialization)
+template <typename T>
+struct is_pointer_helper<T*> : true_type {};
+
+/// std::is_pointer
+template <typename T>
+struct is_pointer : is_pointer_helper<typename remove_cv<T>::type> {};
+
+/// std::is_void
+template <typename T>
+struct is_void : is_same<void, typename remove_cv<T>::type> {};
+
+/// std::is_integral
+template <typename T>
+struct is_integral : false_type {};
+template <>
+struct is_integral<char> : true_type {};
+template <>
+struct is_integral<signed char> : true_type {};
+template <>
+struct is_integral<unsigned char> : true_type {};
+template <>
+struct is_integral<short> : true_type {};
+template <>
+struct is_integral<unsigned short> : true_type {};
+template <>
+struct is_integral<int> : true_type {};
+template <>
+struct is_integral<unsigned int> : true_type {};
+template <>
+struct is_integral<long> : true_type {};
+template <>
+struct is_integral<unsigned long> : true_type {};
+template <>
+struct is_integral<long long> : true_type {};
+template <>
+struct is_integral<unsigned long long> : true_type {};
+template <typename T>
+struct is_integral<volatile T> : is_integral<T> {};
+template <typename T>
+struct is_integral<const T> : is_integral<T> {};
+template <typename T>
+struct is_integral<const volatile T> : is_integral<T> {};
+
+/// std::is_floating_point
+template <typename T>
+struct is_floating_point
+    : integral_constant<bool,
+                        (is_same<float, typename remove_cv<T>::type>::value ||
+                         is_same<double, typename remove_cv<T>::type>::value)> {};
+
+/// std::is_arithmetic
+template <typename T>
+struct is_arithmetic
+    : integral_constant<bool, (is_integral<T>::value || is_floating_point<T>::value)> {};
+
+/// std::is_fundamental
+template <typename T>
+struct is_fundamental
+    : integral_constant<bool,
+                        (is_arithmetic<T>::value || is_void<T>::value ||
+                         is_same<nullptr_t, typename remove_cv<T>::type>::value)> {};
+
+#else
+
+using std::is_volatile;
+using std::is_pointer;
+using std::is_void;
+using std::is_integral;
+using std::is_floating_point;
+using std::is_arithmetic;
+using std::is_fundamental;
+
+#endif
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1800)) || \
+    (defined(__GNUG__) && (__GNUC__ < 5))
+
+/**
+     * std::is_trivially_copyable
+     *
+     * This implementation only evaluates true if T is fundamental or pointer
+     *
+     * Without help from partial template specializations provided by the user for
+     * a specific class or struct, this trait will never report that the specified
+     * class or struct  is trivially-copyable ; this is always safe,
+     * if possibly sub-optimal.
+     */
+template <typename T>
+struct is_trivially_copyable
+    : integral_constant<bool, (is_fundamental<T>::value || is_pointer<T>::value)> {};
+
+#else
+
+using std::is_trivially_copyable;
+
+#endif
+
+//-----------------------------------------------------------------------------
+// Alignment and layout utilities
+//-----------------------------------------------------------------------------
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1500))
+
+/// std::alignment_of
+template <typename value_t>
+struct alignment_of {
+  struct pad {
+    value_t val;
+    char byte;
+  };
+
+  enum { value = sizeof(pad) - sizeof(value_t) };
+};
+
+#else
+
+template <typename value_t>
+struct alignment_of : std::alignment_of<value_t> {};
+
+#endif
+
+/* 16B specializations where 32-bit Win32 host compiler disagrees with device compiler */
+template <>
+struct alignment_of<int4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<uint4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<float4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<long4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<ulong4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<longlong2> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<ulonglong2> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<double2> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<longlong4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<ulonglong4> {
+  enum { value = 16 };
+};
+template <>
+struct alignment_of<double4> {
+  enum { value = 16 };
+};
+
+// Specializations for volatile/const qualified types
+template <typename value_t>
+struct alignment_of<volatile value_t> : alignment_of<value_t> {};
+template <typename value_t>
+struct alignment_of<const value_t> : alignment_of<value_t> {};
+template <typename value_t>
+struct alignment_of<const volatile value_t> : alignment_of<value_t> {};
+
+#if (!defined(_MSC_VER) && (__cplusplus < 201103L)) || (defined(_MSC_VER) && (_MSC_VER < 1800))
+
+template <size_t Align>
+struct aligned_chunk;
+template <>
+struct __align__(1) aligned_chunk<1> {
+  uint8_t buff;
+};
+template <>
+struct __align__(2) aligned_chunk<2> {
+  uint16_t buff;
+};
+template <>
+struct __align__(4) aligned_chunk<4> {
+  uint32_t buff;
+};
+template <>
+struct __align__(8) aligned_chunk<8> {
+  uint32_t buff[2];
+};
+template <>
+struct __align__(16) aligned_chunk<16> {
+  uint32_t buff[4];
+};
+template <>
+struct __align__(32) aligned_chunk<32> {
+  uint32_t buff[8];
+};
+template <>
+struct __align__(64) aligned_chunk<64> {
+  uint32_t buff[16];
+};
+template <>
+struct __align__(128) aligned_chunk<128> {
+  uint32_t buff[32];
+};
+template <>
+struct __align__(256) aligned_chunk<256> {
+  uint32_t buff[64];
+};
+template <>
+struct __align__(512) aligned_chunk<512> {
+  uint32_t buff[128];
+};
+template <>
+struct __align__(1024) aligned_chunk<1024> {
+  uint32_t buff[256];
+};
+template <>
+struct __align__(2048) aligned_chunk<2048> {
+  uint32_t buff[512];
+};
+template <>
+struct __align__(4096) aligned_chunk<4096> {
+  uint32_t buff[1024];
+};
+
+/// std::aligned_storage
+template <size_t Len, size_t Align>
+struct aligned_storage {
+  typedef aligned_chunk<Align> type[Len / sizeof(aligned_chunk<Align>)];
+};
+
+#else
+
+using std::aligned_storage;
+
+#endif
+
+#if !defined(__CUDACC_RTC__)
+/// Default deleter
+template <typename T>
+struct default_delete {
+  void operator()(T* ptr) const { delete ptr; }
+};
+
+/// Partial specialization for deleting array types
+template <typename T>
+struct default_delete<T[]> {
+  void operator()(T* ptr) const { delete[] ptr; }
+};
+
+/// std::unique_ptr
+template <class T, class Deleter = default_delete<T> >
+class unique_ptr {
+ public:
+  typedef T* pointer;
+  typedef T element_type;
+  typedef Deleter deleter_type;
+
+ private:
+  /// Pointer to memory
+  pointer _ptr;
+
+  /// Deleter
+  deleter_type _deleter;
+
+ public:
+  unique_ptr() : _ptr(nullptr) {}
+  unique_ptr(pointer p) : _ptr(p) {}
+
+  ~unique_ptr() {
+    if (_ptr) {
+      _deleter(_ptr);
+    }
+  }
+  /// Returns a pointer to the managed object or nullptr if no object is owned.
+  pointer get() const noexcept { return _ptr; }
+
+  /// Releases ownership of the managed object, if any
+  pointer release() noexcept {
+    pointer p(_ptr);
+    _ptr = nullptr;
+    return p;
+  }
+
+  /// Replaces the managed object, deleting the old object.
+  void reset(pointer p = pointer()) noexcept {
+    pointer old_ptr = _ptr;
+    _ptr = p;
+    if (old_ptr != nullptr) {
+      get_deleter()(old_ptr);
+    }
+  }
+
+  /// Swaps the managed objects with *this and another unique_ptr
+  void swap(unique_ptr& other) noexcept { std::swap(_ptr, other._ptr); }
+
+  /// Returns the deleter object
+  Deleter& get_deleter() noexcept { return _deleter; }
+
+  /// Returns the deleter object
+  Deleter const& get_deleter() const noexcept { return _deleter; }
+
+  /// Checks whether an object is owned
+  operator bool() const noexcept { return _ptr != nullptr; }
+
+  /// Dereferences the unique_ptr
+  T& operator*() const { return *_ptr; }
+
+  /// Returns a pointer to the managed object
+  pointer operator->() const noexcept { return _ptr; }
+
+  /// Array access to managed object
+  T& operator[](size_t i) const { return _ptr[i]; }
+};
+
+/// Specializes the swap algorithm
+template <typename T, typename Deleter>
+void swap(unique_ptr<T, Deleter>& lhs, unique_ptr<T, Deleter>& rhs) noexcept {
+  lhs.swap(rhs);
+}
+#endif
+
+};  // namespace platform
+};  // namespace cutlass

cutlass/util/printable.h

@@ -1,72 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief Pure virtual base class for printable types
- */
-
-#include <iostream>
-
-
-namespace cutlass {
-
-
-/******************************************************************************
- * printable_t
- ******************************************************************************/
-
-/**
- * Pure virtual base class for printable types
- */
-struct printable_t
-{
-    /// Returns the instance as a string
-    __host__ __device__ inline
-    virtual char const* to_string() const = 0;
-
-    /// Insert the formatted instance into the output stream
-    virtual void print(std::ostream& out) const = 0;
-
-    /// Destructor
-    virtual ~printable_t() {}
-};
-
-
-/// Insert the formatted \p printable into the output stream
-std::ostream& operator<<(
-    std::ostream& out,
-    printable_t const& printable)
-{
-    printable.print(out);
-    return out;
-}
-
-
-} // namespace cutlass

cutlass/util/util.h

@@ -1,82 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in the
- *       documentation and/or other materials provided with the distribution.
- *     * Neither the name of the NVIDIA CORPORATION nor the
- *       names of its contributors may be used to endorse or promote products
- *       derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
- * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
- * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
- * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
- * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
- * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
- * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- ******************************************************************************/
-
-#pragma once
-
-/**
- * \file
- * \brief Umbrella header file for utilities
- */
-
-#include "debug.h"
-#include "device_introspection.h"
-#include "io_intrinsics.h"
-#include "math.h"
-#include "nv_std.h"
-#include "printable.h"
-#include "matrix_transform.h"
-
-
-
-namespace cutlass {
-
-
-/******************************************************************************
- * int_constant
- ******************************************************************************/
-
-/**
- * Shorthand for nv_std::integral_constant of int32_t type
- */
-template <int V>
-struct int_constant : nv_std::integral_constant<int32_t, V>
-{};
-
-
-/******************************************************************************
- * Uninitialized
- ******************************************************************************/
-
-/**
- * \brief A storage-backing wrapper that allows types with non-trivial constructors to be aliased in unions
- */
-template <typename T>
-struct __align__(16) uninitialized
-{
-    /// Backing storage
-    uint8_t storage[sizeof(T)];
-
-    /// Alias
-    __host__ __device__ __forceinline__ T& alias()
-    {
-        return reinterpret_cast<T&>(*this);
-    }
-};
-
-
-
-} // namespace cutlass

cutlass/vector.h

@@ -0,0 +1,379 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a 1D vector of elements held in the registers of each thread.
+*/
+#pragma once
+
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
+#include <cuda_fp16.h>
+#endif
+
+#include "cutlass/util/numeric_types.h"
+#include "cutlass/util/platform.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <size_t kAlignment_>
+struct AlignedStruct {};
+
+template <>
+struct __align__(1) AlignedStruct<1>{};
+template <>
+struct __align__(2) AlignedStruct<2>{};
+template <>
+struct __align__(4) AlignedStruct<4>{};
+template <>
+struct __align__(8) AlignedStruct<8>{};
+template <>
+struct __align__(16) AlignedStruct<16>{};
+template <>
+struct __align__(32) AlignedStruct<32>{};
+template <>
+struct __align__(64) AlignedStruct<64>{};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kLanes_>
+union Vector {
+  /// The scalar type.
+  typedef Scalar_ Scalar;
+
+  /// The number of elements in the vector.
+  enum { kLanes = kLanes_ };
+  /// The size of the vector.
+  enum { kVectorSize = kLanes * (int)sizeof(Scalar) };
+  /// The number of registers needed to store the vector.
+  enum { kRegisters = kVectorSize < 4 ? 1 : kVectorSize / 4 };
+
+  // Make sure that the vector type makes sense.
+  static_assert(kVectorSize <= 16, "Vector type is too large");
+
+  /// The aligned storage to make sure we have good alignment.
+  AlignedStruct<kVectorSize> aligned_;
+  /// The associated array of scalars.
+  Scalar scalars[kLanes];
+  /// The data in registers.
+  uint32_t registers[kRegisters];
+
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE Scalar const& operator[](uint32_t i) const { return scalars[i]; }
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE Scalar& operator[](uint32_t i) { return scalars[i]; }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <>
+union Vector<half, 1> {
+  /// The scalar type.
+  typedef half Scalar;
+
+  /// The number of elements in the vector.
+  enum { kLanes = 1 };
+  /// The size of the vector.
+  enum { kVectorSize = kLanes * (int)sizeof(Scalar) };
+  /// The number of registers needed to store the vector.
+  enum { kRegisters = kVectorSize < 4 ? 1 : kVectorSize / 4 };
+
+  // Make sure that the vector type makes sense.
+  static_assert(kVectorSize <= 16, "Vector type is too large");
+
+  /// The aligned storage to make sure we have good alignment.
+  AlignedStruct<kVectorSize> aligned_;
+  /// The associated array of scalars.
+  uint16_t scalars[kLanes];
+
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE Scalar const& operator[](uint32_t i) const {
+    return reinterpret_cast<Scalar const&>(scalars[i]);
+  }
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE Scalar& operator[](uint32_t i) {
+      return reinterpret_cast<Scalar&>(scalars[i]);
+  }
+};
+
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
+
+template <int kLanes_>
+union Vector<half, kLanes_> {
+  /// The scalar type.
+  typedef half Scalar;
+
+  /// The number of elements in the vector.
+  enum { kLanes = kLanes_ };
+  /// The size of the vector.
+  enum { kVectorSize = kLanes * (int)sizeof(Scalar) };
+  /// The number of registers needed to store the vector.
+  enum { kRegisters = kVectorSize < 4 ? 1 : kVectorSize / 4 };
+
+  // Make sure that the vector type makes sense.
+  static_assert(kVectorSize <= size_t(16), "Vector type is too large");
+
+  /// The aligned storage to make sure we have good alignment.
+  AlignedStruct<kVectorSize> aligned_;
+  /// The associated array of scalars.
+  uint16_t scalars[kLanes];
+  /// The data in registers.
+  uint32_t registers[kRegisters];
+
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE Scalar const& operator[](uint32_t i) const {
+    return reinterpret_cast<Scalar const&>(scalars[i]);
+  }
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE Scalar& operator[](uint32_t i) {
+      return reinterpret_cast<Scalar&>(scalars[i]);
+  }
+};
+
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Vector definition for 1-bit binary datatype
+template <int kLanes_>
+union Vector<bin1_t, kLanes_> {
+  /// The scalar type.
+  typedef bin1_t Scalar;
+
+  /// The number of elements in the vector.
+  enum { kLanes = kLanes_ };
+  /// The size of the vector.
+  enum { kVectorSize = kLanes / 8 };
+  /// The number of registers needed to store the vector.
+  enum { kRegisters = kVectorSize < 4 ? 1 : kVectorSize / 4 };
+
+  static_assert((kLanes >= 8) && !(kLanes % 8),
+                "May only construct vectors of bin1_t that are multiples of 8 bits.");
+
+  /// The aligned storage to make sure we have good alignment.
+  AlignedStruct<kVectorSize> aligned_;
+  /// The data in registers.
+  uint32_t registers[kRegisters];
+
+  /// Default Constructor
+  CUTLASS_HOST_DEVICE
+  Vector() {}
+  /// Constructor to convert from uint32_t type
+  CUTLASS_HOST_DEVICE Vector(uint32_t value) { registers[0] = value; }
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE bool operator[](uint32_t i) const {
+    return ( (registers[i / 32] & (1 << (i % 32))) != 0 );
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Vector definition for 4-bit signed integer datatype
+template <int kLanes_>
+union Vector<int4_t, kLanes_> {
+  /// The scalar type.
+  typedef int4_t Scalar;
+
+  /// The number of elements in the vector.
+  enum { kLanes = kLanes_ };
+  /// The size of the vector.
+  enum { kVectorSize = kLanes / 2 };
+  /// The number of registers needed to store the vector.
+  enum { kRegisters = kVectorSize < 4 ? 1 : kVectorSize / 4 };
+
+  static_assert((kLanes >= 2) && !(kLanes % 2),
+   "May only construct vectors of int4_t that are multiples of 8 bits.");
+
+  /// The aligned storage to make sure we have good alignment.
+  AlignedStruct<kVectorSize> aligned_;
+  /// The data in registers.
+  uint32_t registers[kRegisters];
+
+  /// Default Constructor
+  CUTLASS_HOST_DEVICE
+  Vector() {}
+  /// Constructor to convert from uint32_t type
+  CUTLASS_HOST_DEVICE Vector(uint32_t value) { registers[0] = value; }
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE int operator[](uint32_t i) const {
+    return (registers[i / 8] >> (i % 8 * 4) & 0x0f)
+              - 16 * (registers[i / 8] >> (i % 8 * 4 + 3) & 0x01);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Vector definition for 4-bit unsigned integer datatype
+template <int kLanes_>
+union Vector<uint4_t, kLanes_> {
+  /// The scalar type.
+  typedef uint4_t Scalar;
+
+  /// The number of elements in the vector.
+  enum { kLanes = kLanes_ };
+  /// The size of the vector.
+  enum { kVectorSize = kLanes / 2 };
+  /// The number of registers needed to store the vector.
+  enum { kRegisters = kVectorSize < 4 ? 1 : kVectorSize / 4 };
+
+  static_assert((kLanes >= 2) && !(kLanes % 2),
+    "May only construct vectors of uint4_t that are multiples of 8 bits.");
+
+  /// The aligned storage to make sure we have good alignment.
+  AlignedStruct<kVectorSize> aligned_;
+  /// The data in registers.
+  uint32_t registers[kRegisters];
+
+  /// Default Constructor
+  CUTLASS_HOST_DEVICE
+  Vector() {}
+  /// Constructor to convert from uint32_t type
+  CUTLASS_HOST_DEVICE Vector(uint32_t value) { registers[0] = value; }
+  /// Accessor to the ith lane.
+  CUTLASS_HOST_DEVICE int operator[](uint32_t i) const {
+    return registers[i / 8] >> (i % 8 * 4) & 0x0f;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_>
+CUTLASS_HOST_DEVICE void make_zero(Scalar_& x) {
+  x = Scalar_(0);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Element_, int kLanes_ = 1>
+struct Vectorize {
+  typedef Vector<Element_, kLanes_> Type;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kLanes_>
+struct Vectorize<Vector<bin1_t, 32>, kLanes_> {
+  typedef Vector<bin1_t, kLanes_ * 32> Type;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kLanes_>
+struct Vectorize<Vector<int4_t, 8>, kLanes_> {
+  typedef Vector<int4_t, kLanes_ * 8> Type;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int kLanes_>
+struct Vectorize<Vector<uint4_t, 8>, kLanes_> {
+  typedef Vector<uint4_t, kLanes_ * 8> Type;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Scalar_, int kLanes_>
+CUTLASS_HOST_DEVICE void make_zero(Vector<Scalar_, kLanes_>& vec) {
+  for (int i = 0; i < Vector<Scalar_, kLanes_>::kRegisters; ++i) {
+    vec.registers[i] = 0;
+  }
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+//
+// cutlass::Extent similar to std::extent but applicable to CUTLASS types
+//
+
+/// Returns the extent of a scalar or vector
+template <typename T>
+struct Extent {
+  static size_t const kValue = 1;
+};
+
+/// Returns the number of lanes of a vector if need be
+template <typename T, int Lanes>
+struct Extent<Vector<T, Lanes> > {
+  static size_t const kValue = Lanes;
+};
+
+/// Returns the number of lanes of a vector if need be
+template <typename T, int Lanes>
+struct Extent<Vector<T, Lanes> const> {
+  static size_t const kValue = Lanes;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Traits describing properties of vectors and scalar-as-vectors
+template <typename T>
+struct VectorTraits {
+  /// Scalar type
+  typedef T Scalar;
+
+  /// Number of lanes of vector
+  static int const kLanes = 1;
+
+  /// True if the type is actually a cutlass::Vector, otherwise false
+  static bool const IsVector = false;
+
+  /// Type that is always a vector
+  typedef Vector<T, 1> Vector;
+};
+
+/// Partial specialization for actual cutlass::Vector
+template <typename T, int Lanes>
+struct VectorTraits<Vector<T, Lanes> > {
+  /// Scalar type
+  typedef T Scalar;
+
+  /// Number of lanes of vector
+  static int const kLanes = Lanes;
+
+  /// Type is actually a cutlass::Vector
+  static bool const IsVector = true;
+
+  /// Type that is always a Vector
+  typedef Vector<T, Lanes> Vector;
+};
+
+/// Partial specialization for actual cutlass::Vector
+template <typename T, int Lanes>
+struct VectorTraits<Vector<T, Lanes> const> {
+  /// Scalar type
+  typedef T Scalar;
+
+  /// Number of lanes of vector
+  static int const kLanes = Lanes;
+
+  /// Type is actually a cutlass::Vector
+  static bool const IsVector = true;
+
+  /// Type that is always a Vector
+  typedef Vector<T, Lanes> Vector;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/wmma_matrix.h

@@ -0,0 +1,236 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Abstractions for loading and storing matrices using the CUDA WMMA API.
+*/
+#pragma once
+
+#if defined(__CUDACC__) && (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 700)
+#define CUTLASS_USE_WMMA_API
+
+#if defined(__CUDACC__) && (__CUDACC_VER_MAJOR__ >= 10) && (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 750)
+#define CUTLASS_USE_SUBBYTE_WMMA
+#endif
+
+#include "stdio.h"
+
+#if __CUDACC_VER_MAJOR__ >= 10
+#include <mma.h>
+#else
+#include <crt/mma.h>
+#endif
+#include "cutlass/fragment.h"
+#include "cutlass/matrix_traits.h"
+#include "cutlass/shape.h"
+#include "cutlass/vector.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Statically maps cutlass::MatrixLayout => nvcuda::wmma layout tags
+template <MatrixLayout::Kind kLayout_>
+struct WmmaLayout {
+  typedef nvcuda::wmma::col_major Layout;
+};
+
+/// Statically maps cutlass::MatrixLayout => nvcuda::wmma layout tags
+template <>
+struct WmmaLayout<MatrixLayout::kRowMajor> {
+  typedef nvcuda::wmma::row_major Layout;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Statically maps cutlass types to nvcuda::wmma datatypes
+template <typename Type_>
+struct WmmaDataType{
+  typedef Type_ Type;
+};
+
+#ifdef CUTLASS_USE_SUBBYTE_WMMA
+/// Statically maps cutlass::Vector<bin1_t, 32> to nvcuda::wmma::experimental::precision::b1
+template<>
+struct WmmaDataType<Vector<bin1_t, 32> > {
+  typedef nvcuda::wmma::experimental::precision::b1 Type;
+};
+
+/// Statically maps cutlass::Vector<int4_t, 8> to nvcuda::wmma::experimental::precision::s4
+template<>
+struct WmmaDataType<Vector<int4_t, 8> > {
+  typedef nvcuda::wmma::experimental::precision::s4 Type;
+};
+
+/// Statically maps cutlass::Vector<uint4_t, 8> to nvcuda::wmma::experimental::precision::u4
+template<>
+struct WmmaDataType<Vector<uint4_t, 8> > {
+  typedef nvcuda::wmma::experimental::precision::u4 Type;
+};
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Adapter to nvcuda::wmma fragment load and store operations
+template <GemmOperand::Kind kOperand_,
+          MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename WmmaShape_>
+struct WmmaMatrix {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Adapter to nvcuda::wmma fragment accessors for A operand
+template <MatrixLayout::Kind kLayout_, typename Scalar_, typename WmmaShape_>
+struct WmmaMatrix<GemmOperand::kA, kLayout_, Scalar_, WmmaShape_>
+    : public nvcuda::wmma::fragment<
+          /// The nvcuda::wmma operand name.
+          nvcuda::wmma::matrix_a,
+          /// The dimensions.
+          WmmaShape_::kW,
+          WmmaShape_::kH,
+          WmmaShape_::kD,
+          /// The scalar.
+          typename WmmaDataType<Scalar_>::Type,
+          /// The layout.
+          typename WmmaLayout<kLayout_>::Layout> {
+  /// This type.
+  typedef WmmaMatrix<GemmOperand::kA, kLayout_, Scalar_, WmmaShape_> This_;
+
+  /// Fill-in the element.
+  CUTLASS_DEVICE This_& operator=(Scalar_ const& x) {
+    nvcuda::wmma::fill_fragment(*this, x);
+    return *this;
+  }
+
+  /// Load from memory.
+  CUTLASS_DEVICE void load(Scalar_ const* pointer, int const stride) {
+    nvcuda::wmma::load_matrix_sync(*this, pointer, stride);
+  }
+
+  /// Store to memory.
+  CUTLASS_DEVICE void store(Scalar_* pointer, int const stride) const {
+    nvcuda::wmma::store_matrix_sync(pointer, *this, stride);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Adapter to nvcuda::wmma fragment accessors for B operand
+template <MatrixLayout::Kind kLayout_, typename Scalar_, typename WmmaShape_>
+struct WmmaMatrix<GemmOperand::kB, kLayout_, Scalar_, WmmaShape_>
+    : public nvcuda::wmma::fragment<
+          /// The nvcuda::wmma operand name.
+          nvcuda::wmma::matrix_b,
+          /// The dimensions.
+          WmmaShape_::kW,
+          WmmaShape_::kH,
+          WmmaShape_::kD,
+          /// The scalar.
+          typename WmmaDataType<Scalar_>::Type,
+          /// The layout.
+          typename WmmaLayout<kLayout_>::Layout> {
+  /// This type.
+  typedef WmmaMatrix<GemmOperand::kB, kLayout_, Scalar_, WmmaShape_> This_;
+
+  /// Fill-in the element.
+  CUTLASS_DEVICE This_& operator=(Scalar_ const& x) {
+    nvcuda::wmma::fill_fragment(*this, x);
+    return *this;
+  }
+
+  /// Load from memory.
+  CUTLASS_DEVICE void load(Scalar_ const* pointer, int const stride) {
+    nvcuda::wmma::load_matrix_sync(*this, pointer, stride);
+  }
+
+  /// Store to memory.
+  CUTLASS_DEVICE void store(Scalar_* pointer, int const stride) const {
+    nvcuda::wmma::store_matrix_sync(pointer, *this, stride);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Adapter to nvcuda::wmma fragment accessors for C operand
+template <MatrixLayout::Kind kLayout_, typename Scalar_, typename WmmaShape_>
+struct WmmaMatrix<GemmOperand::kC, kLayout_, Scalar_, WmmaShape_>
+    : public nvcuda::wmma::fragment<
+          /// The nvcuda::wmma operand name.
+          nvcuda::wmma::accumulator,
+          /// The dimensions.
+          WmmaShape_::kW,
+          WmmaShape_::kH,
+          WmmaShape_::kD,
+          /// The scalar.
+          Scalar_> {
+  /// This type.
+  typedef WmmaMatrix<GemmOperand::kC, kLayout_, Scalar_, WmmaShape_> This_;
+  /// The layout.
+  static MatrixLayout::Kind const kLayout = kLayout_;
+
+  /// Fill-in the element.
+  CUTLASS_DEVICE This_& operator=(Scalar_ const& x) {
+    nvcuda::wmma::fill_fragment(*this, x);
+    return *this;
+  }
+
+  /// Load from memory.
+  CUTLASS_DEVICE void load(Scalar_ const* pointer, int const stride) {
+    bool const kIsRowMajor = kLayout == MatrixLayout::kRowMajor;
+    nvcuda::wmma::load_matrix_sync(
+        *this,
+        pointer,
+        stride,
+        kIsRowMajor ? nvcuda::wmma::mem_row_major : nvcuda::wmma::mem_col_major);
+  }
+
+  /// Store to memory.
+  CUTLASS_DEVICE void store(Scalar_* pointer, int const stride) const {
+    bool const kIsRowMajor = kLayout == MatrixLayout::kRowMajor;
+    nvcuda::wmma::store_matrix_sync(
+        pointer,
+        *this,
+        stride,
+        kIsRowMajor ? nvcuda::wmma::mem_row_major : nvcuda::wmma::mem_col_major);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// WmmaMatrix cannot be used in a Union and thus in cannot be used in our Vector implementation.
+// The only use of WmmaMatrix in in combination with Vectorize has kLanes == 1. Due to this it is
+// safe to keep the Vector->Scalar conversion for WmmaMatrix.
+template <GemmOperand::Kind kOperand_,
+          MatrixLayout::Kind kLayout_,
+          typename Scalar_,
+          typename WmmaShape_>
+struct Vectorize<WmmaMatrix<kOperand_, kLayout_, Scalar_, WmmaShape_>, 1> {
+  typedef WmmaMatrix<kOperand_, kLayout_, Scalar_, WmmaShape_> Type;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+}
+
+#endif  // defined CUTLASS_USE_WMMA_API

cutlass/zip_fragment.h

@@ -0,0 +1,150 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Models a pair of fragments
+*/
+#pragma once
+
+#include <assert.h>
+
+#include "cutlass/cutlass.h"
+#include "cutlass/shape.h"
+#include "cutlass/util/cutlass_math.h"
+#include "cutlass/vector.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/**
+* @brief A template defining \ref fragment_concept
+* @concept{fragment_concept}
+*/
+template <typename First_, typename Second_>
+struct ZipFragment {
+  /// First fragment object
+  typedef First_ First;
+
+  /// Second fragment object
+  typedef Second_ Second;
+
+  /// This class.
+  typedef ZipFragment<First, Second> This_;
+
+  //
+  // Data members
+  //
+
+  /// First fragment object
+  First first;
+
+  /// Second fragment object
+  Second second;
+
+  //
+  // Methods
+  //
+
+  /// Default ctor
+  CUTLASS_DEVICE
+  ZipFragment() { }
+
+  /// Copy ctor
+  CUTLASS_DEVICE
+  ZipFragment(First const &_first, Second const &_second): first(_first), second(_second) { }
+
+  /// Clear a fragment.
+  CUTLASS_DEVICE void clear() {
+    first.clear();
+    second.clear();
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Helper to construct a ZipFragment object
+template <typename First, typename Second>
+CUTLASS_HOST_DEVICE
+ZipFragment<First, Second> make_ZipFragment(First const &first, Second const &second) {
+  return ZipFragment<First, Second>(first, second);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Zips two convert operations
+template <typename First_, typename Second_>
+struct ZipConvert {
+  /// First convert operator
+  typedef First_ First;
+
+  /// Second convert operator
+  typedef Second_ Second;
+
+  /// Defines the input zip fragment
+  typedef ZipFragment<typename First::InputFragment, typename Second::InputFragment> InputFragment;
+
+  /// Defines the output zip fragment
+  typedef ZipFragment<typename First::OutputFragment, typename Second::OutputFragment>
+      OutputFragment;
+
+  //
+  //
+  //
+
+  /// First transformer
+  First first;
+
+  /// Second transformer
+  Second second;
+
+  //
+  //
+  //
+
+  /// Ctor.
+  CUTLASS_DEVICE ZipConvert() {}
+
+  /// Ctor.
+  CUTLASS_DEVICE ZipConvert(First const &_first, Second const &_second): first(_first), second(_second) { }
+
+  /// Transform a fragment.
+  CUTLASS_DEVICE void transform(InputFragment const& src, OutputFragment& dst) {
+    first.transform(src.first, dst.first);
+    second.transform(src.second, dst.second);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Helper to construct a ZipConvert object
+template <typename First, typename Second>
+CUTLASS_HOST_DEVICE
+ZipConvert<First, Second> make_ZipConvert(First const &first, Second const &second) {
+  return ZipConvert<First, Second>(first, second);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/zip_tensor_ref.h

@@ -0,0 +1,77 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+    \brief Defines a structure containing a pair of TensorRef-like objects
+*/
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/tensor_ref.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename First_, typename Second_>
+struct ZipTensorRef {
+  /// First tensor ref
+  typedef First_ First;
+
+  /// Second tensor ref
+  typedef Second_ Second;
+
+  //
+  // Data members
+  //
+
+  /// First TensorRef
+  First first;
+
+  /// Second TensorRef
+  Second second;
+
+  //
+  // Methods
+  //
+
+  CUTLASS_HOST_DEVICE
+  ZipTensorRef() {}
+
+  CUTLASS_HOST_DEVICE
+  ZipTensorRef(First const& _first, Second const& _second) : first(_first), second(_second) {}
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Constructs a ZipTensorRef
+template <typename First, typename Second>
+CUTLASS_HOST_DEVICE
+ZipTensorRef<First, Second> make_ZipTensorRef(First const &first, Second const &second) {
+  return ZipTensorRef<First, Second>(first, second);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

cutlass/zip_tile_iterator.h

@@ -0,0 +1,287 @@
+/***************************************************************************************************
+ * Copyright (c) 2017-2018, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification, are permitted
+ * provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright notice, this list of
+ *       conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright notice, this list of
+ *       conditions and the following disclaimer in the documentation and/or other materials
+ *       provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
+ *       to endorse or promote products derived from this software without specific prior written
+ *       permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
+ * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+/*! \file
+    \brief Constructs an iterator that owns two tile iterator instances
+*/
+
+#pragma once
+
+#include "cutlass/coord.h"
+#include "cutlass/zip_tensor_ref.h"
+#include "cutlass/zip_fragment.h"
+
+namespace cutlass {
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Constructs an iterator from a pair of iterators
+template <typename First_, typename Second_>
+class ZipTileIterator {
+ public:
+  /// First iterator type
+  typedef First_ First;
+
+  /// Second iterator type
+  typedef Second_ Second;
+
+  /// Params object
+  struct Params {
+    /// Parameters of first iterator
+    typename First::Params first;
+
+    /// Parameters of second iterator
+    typename Second::Params second;
+
+    /// Constructs a parameters object
+    CUTLASS_HOST_DEVICE
+    Params() {}
+
+    /// Constructs a parameters object
+    CUTLASS_HOST_DEVICE
+    Params(typename First::Params const &_first, typename Second::Params const &_second)
+        : first(_first), second(_second) {}
+  };
+
+  /// Fragment type
+  typedef ZipFragment<typename First::Fragment, typename Second::Fragment> Fragment;
+
+  /// Predicate vector
+  typedef typename First::PredicateVector PredicateVector;
+
+  /// Index type
+  typedef typename First::Index Index;
+
+  /// Tensor reference
+  typedef ZipTensorRef<
+    typename First::TensorRef,
+    typename Second::TensorRef> TensorRef;
+
+  //
+  // Data members
+  //
+
+  /// First iterator
+  First first;
+
+  /// Second iterator
+  Second second;
+
+  //
+  // Methods
+  //
+
+  /// Default constructor
+  CUTLASS_DEVICE
+  ZipTileIterator() {}
+
+  /// Constructs a zip iterator from params
+  CUTLASS_DEVICE
+  ZipTileIterator(Params const &_params, Coord<3> const &threadblock_offset = make_Coord(0, 0, 0))
+      : first(_params.first, threadblock_offset), second(_params.second, threadblock_offset) {}
+
+  /// Constructs a zip iterator from iterator instances
+  CUTLASS_DEVICE
+  ZipTileIterator(First const &_first, Second const &_second) : first(_first), second(_second) {}
+
+  /// Constructs a zip iterator from iterator instances
+  CUTLASS_DEVICE
+  ZipTileIterator(TensorRef const &ref) : first(ref.first), second(ref.second) {}
+
+  /// Constructs a zip iterator from iterator instances
+  CUTLASS_DEVICE
+  ZipTileIterator(Params const &_params, TensorRef const &ref):
+    first(_params.first, ref.first), second(_params.second, ref.second) {}
+
+  //
+  // Predicate initialization
+  //
+
+  /// Initializes a predicate vector using a RegularTilePredicateFunctor
+  template <
+      /// Predicate iterator
+      typename PredicateIterator>
+  CUTLASS_HOST_DEVICE void initialize_predicates(PredicateIterator predicate_it,
+                                                 Coord<3> const &bounds,
+                                                 Coord<3> const &block_offset = make_Coord(0,
+                                                                                           0,
+                                                                                           0)) {
+    first.initialize_predicates(predicate_it, bounds, block_offset);
+  }
+
+  /// Initializes a predicate vector using an arbitrary predicate functor
+  template <
+      /// Predicate iterator
+      typename PredicateIterator,
+      /// Functor computing predicates
+      typename PredicateFunctor>
+  CUTLASS_HOST_DEVICE void initialize_predicates(PredicateIterator predicate_it,
+                                                 PredicateFunctor const &functor,
+                                                 Coord<3> const &block_offset) {
+    first.initialize_predicates(predicate_it, functor, block_offset);
+  }
+
+  //
+  // No predicates
+  //
+
+  /// Loads a fragment and increments without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void load_post_increment(Fragment &fragment) {
+    first.load_post_increment(fragment.first);
+    second.load_post_increment(fragment.second);
+  }
+
+  /// Loads a fragment and increments without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void load_post_increment(Fragment &fragment,
+                            Coord<4> const &offset) {
+    first.load_post_increment(fragment.first, offset);
+    second.load_post_increment(fragment.second, offset);
+  }
+
+  /// Loads a fragment without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void load(Fragment &fragment) const {
+    first.load(fragment.first);
+    second.load(fragment.second);
+  }
+
+  /// Loads a fragment without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void load(Fragment &fragment,
+                            Coord<4> const &offset) const {
+    first.load(fragment.first, offset);
+    second.load(fragment.second, offset);
+  }
+
+  /// Stores a fragment and increments without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void store_post_increment(Fragment const &fragment) {
+    first.store_post_increment(fragment.first);
+    second.store_post_increment(fragment.second);
+  }
+
+  /// Stores a fragment and increments without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void store_post_increment(Fragment const &fragment,
+                            Coord<4> const &offset) {
+    first.store_post_increment(fragment.first, offset);
+    second.store_post_increment(fragment.second, offset);
+  }
+
+  /// Stores a fragment without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void store(Fragment const &fragment) const {
+    first.store(fragment.first);
+    second.store(fragment.second);
+  }
+
+  /// Stores a fragment without predicates
+  template <typename Fragment>
+  CUTLASS_DEVICE void store(Fragment const &fragment,
+                            Coord<4> const &offset) const {
+    first.store(fragment.first, offset);
+    second.store(fragment.second, offset);
+  }
+
+  //
+  // With predication
+  //
+
+  /// Loads a fragment and increments, using predicates
+  template <typename Fragment, typename PredicateIterator>
+  CUTLASS_DEVICE void load_post_increment(Fragment &fragment, PredicateIterator pred_it) {
+    first.load_post_increment(fragment.first, pred_it);
+    second.load_post_increment(fragment.second, pred_it);
+  }
+
+  /// Loads a fragment with predicates
+  template <typename Fragment, typename PredicateIterator>
+  CUTLASS_DEVICE void load(Fragment &fragment, PredicateIterator pred_it) const {
+    first.load(fragment.first, pred_it);
+    second.load(fragment.second, pred_it);
+  }
+
+  /// Loads a fragment and increments, using predicates
+  template <typename Fragment, typename PredicateIterator>
+  CUTLASS_DEVICE void store_post_increment(Fragment const &fragment, PredicateIterator pred_it) {
+    first.store_post_increment(fragment.first, pred_it);
+    second.store_post_increment(fragment.second, pred_it);
+  }
+
+  /// Loads a fragment with predicates
+  template <typename Fragment, typename PredicateIterator>
+  CUTLASS_DEVICE void store(Fragment const &fragment, PredicateIterator pred_it) const {
+    first.store(fragment.first, pred_it);
+    second.store(fragment.second, pred_it);
+  }
+
+  //
+  // Advances the iterators
+  //
+
+  /// Increments store iterator to next tile
+  CUTLASS_DEVICE ZipTileIterator &increment(int count = 1) {
+    first.increment(count);
+    second.increment(count);
+    return *this;
+  }
+
+  /// Increments to next tile
+  CUTLASS_DEVICE ZipTileIterator &operator++() { return increment(); }
+
+  CUTLASS_DEVICE ZipTileIterator &operator+=(int count) { return increment(count); }
+
+  /// Adds a vector offset to the underlying iterators
+  CUTLASS_DEVICE ZipTileIterator &operator+=(Coord<3> const &offset) {
+    first += offset;
+    second += offset;
+    return *this;
+  }
+
+  /// Increments store iterator to previous tile
+  CUTLASS_DEVICE ZipTileIterator &decrement(int count = 1) {
+    first.decrement(count);
+    second.decrement(count);
+    return *this;
+  }
+
+  /// Increments to subsequent tile
+  CUTLASS_DEVICE ZipTileIterator &operator--() { return decrement(); }
+
+  /// Decrements to previous tile
+  CUTLASS_DEVICE ZipTileIterator &operator-=(int count) { return decrement(count); }
+
+  /// Adds an offset to both iterators
+  CUTLASS_DEVICE void add_pointer_offset(Index offset) {
+    first.add_pointer_offset(offset);
+    second.add_pointer_offset(offset);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namspace cutlass

cutlass_test/.gitignore

@@ -1,7 +0,0 @@
-/bin/
-/gemm-GPU.csv
-/gemm-REF.csv
-/a.csv
-/b.csv
-/gp100_schmoo/
-/ignore/

cutlass_test/Makefile

@@ -1,180 +0,0 @@
-#/******************************************************************************
-# * Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.
-# *
-# * Redistribution and use in source and binary forms, with or without
-# * modification, are permitted provided that the following conditions are met:
-# *     * Redistributions of source code must retain the above copyright
-# *       notice, this list of conditions and the following disclaimer.
-# *     * Redistributions in binary form must reproduce the above copyright
-# *       notice, this list of conditions and the following disclaimer in the
-# *       documentation and/or other materials provided with the distribution.
-# *     * Neither the name of the NVIDIA CORPORATION nor the
-# *       names of its contributors may be used to endorse or promote products
-# *       derived from this software without specific prior written permission.
-# *
-# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
-# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
-# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
-# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
-# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
-# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
-# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
-# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-# *
-# ******************************************************************************/
-
-
-#-------------------------------------------------------------------------------
-#
-# Makefile usage
-#
-# make <target> sm=<XX[,YY,ZZ,..]> [transpose=<nn*|nt|tn|tt>] [verbose=<0*|1>] [keep=<0*|1>]
-#
-#   * : default
-#
-#-------------------------------------------------------------------------------
-
-TEST_DIR := $(dir $(lastword $(MAKEFILE_LIST)))
-
-include ../common.mk
-
-
-#-------------------------------------------------------------------------------
-# Commandline Options
-#-------------------------------------------------------------------------------
-
-ifdef transpose
-	TRANSPOSE := $(transpose)
-else
-    TRANSPOSE := nn
-endif
-
-# If defined, GEMMs only compiled with specified alignment restrictions on A and B
-# matrices. Otherwise, kernels are compiled for all feasible alignment options, and
-# the appropriate kernel is selected.
-ifdef alignment
-	DEFINES += -DGEMM_ALIGNMENT=$(alignment)
-endif
-
-# If defined as false, ragged handling can be disabled.
-ifdef ragged
-	DEFINES += -DGEMM_RAGGED=$(ragged)
-endif
-
-#-------------------------------------------------------------------------------
-# Include and Library paths
-#-------------------------------------------------------------------------------
-
-INC += -I$(TEST_DIR)
-INC += -I$(BASE_DIR)
-
-LIBS += -lcublas
-
-#-------------------------------------------------------------------------------
-# Preprocessor definitions
-#-------------------------------------------------------------------------------
-
-ifeq (nt, $(TRANSPOSE))
-	DEFINES += -DTRANSPOSE_B
-else ifeq (tn, $(TRANSPOSE))
-	DEFINES += -DTRANSPOSE_A
-
-else ifeq (tt, $(TRANSPOSE))
-	DEFINES += -DTRANSPOSE_A
-	DEFINES += -DTRANSPOSE_B
-endif
-
-NVCCFLAGS += -std=c++11
-
-
-#-------------------------------------------------------------------------------
-# Dependency Lists
-#-------------------------------------------------------------------------------
-
-DEPS := $(call rwildcard, $(BASE_DIR),*.h) \
-        $(call rwildcard, $(BASE_DIR)cgl,*.h) \
-        $(BASE_DIR)common.mk \
-        $(TEST_DIR)Makefile
-
-
-ALL :=  sgemm \
-        dgemm \
-        hgemm \
-        igemm
-
-
-#-------------------------------------------------------------------------------
-# make default
-#-------------------------------------------------------------------------------
-
-default:
-
-
-#-------------------------------------------------------------------------------
-# make clean
-#-------------------------------------------------------------------------------
-
-clean :
-	rm -f bin/*
-	rm -f *.i* *.cubin *.cu.c *.cudafe* *.fatbin.c *.ptx *.hash *.cu.cpp *.o *.obj* *dlink.* *.res *.fatbin *.module_id
-
-
-#-------------------------------------------------------------------------------
-# make all
-#-------------------------------------------------------------------------------
-
-all : $(ALL)
-
-
-#-------------------------------------------------------------------------------
-# make sgemm
-#-------------------------------------------------------------------------------
-
-sgemm: bin/sgemm_$(TRANSPOSE)_$(BIN_SUFFIX)
-
-bin/sgemm_$(TRANSPOSE)_$(BIN_SUFFIX) : gemm.cu $(DEPS)
-	mkdir -p bin
-	$(NVCC) -DTEST_SGEMM $(DEFINES) $(SM_TARGETS) -o $@ gemm.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBINC) $(LIBS)
-
-#-------------------------------------------------------------------------------
-# make dgemm
-#-------------------------------------------------------------------------------
-
-dgemm: bin/dgemm_$(TRANSPOSE)_$(BIN_SUFFIX)
-
-bin/dgemm_$(TRANSPOSE)_$(BIN_SUFFIX) : gemm.cu $(DEPS)
-	mkdir -p bin
-	$(NVCC) -DTEST_DGEMM $(DEFINES) $(SM_TARGETS) -o $@ gemm.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBINC) $(LIBS)
-
-#-------------------------------------------------------------------------------
-# make hgemm
-#-------------------------------------------------------------------------------
-
-hgemm: bin/hgemm_$(TRANSPOSE)_$(BIN_SUFFIX)
-
-bin/hgemm_$(TRANSPOSE)_$(BIN_SUFFIX) : gemm.cu $(DEPS)
-	mkdir -p bin
-	$(NVCC) -DTEST_HGEMM $(DEFINES) $(SM_TARGETS) -o $@ gemm.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBINC) $(LIBS)
-
-#-------------------------------------------------------------------------------
-# make igemm
-#-------------------------------------------------------------------------------
-
-igemm: bin/igemm_$(TRANSPOSE)_$(BIN_SUFFIX)
-
-bin/igemm_$(TRANSPOSE)_$(BIN_SUFFIX) : gemm.cu $(DEPS)
-	mkdir -p bin
-	$(NVCC) -DTEST_IGEMM $(DEFINES) $(SM_TARGETS) -o $@ gemm.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBINC) $(LIBS)
-
-#-------------------------------------------------------------------------------
-# make wgemm
-#-------------------------------------------------------------------------------
-
-wgemm: bin/wgemm_$(TRANSPOSE)_$(BIN_SUFFIX)
-
-bin/wgemm_$(TRANSPOSE)_$(BIN_SUFFIX) : gemm.cu $(DEPS)
-	mkdir -p bin
-	$(NVCC) -DTEST_WGEMM -DWMMA $(DEFINES) $(SM_TARGETS) -o $@ gemm.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBINC) $(LIBS)
-