Files
cutlass/python/cutlass_cppgen/backend/gemm_operation.py
2025-09-18 14:26:57 -04:00

2146 lines
84 KiB
Python

#################################################################################################
#
# Copyright (c) 2017 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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.
#
#################################################################################################
from __future__ import annotations
import copy
import ctypes
import enum
from cutlass_cppgen.utils.lazy_import import lazy_import
cuda = lazy_import("cuda.cuda")
cudart = lazy_import("cuda.cudart")
from cutlass_library import SubstituteTemplate
import numpy as np
from cutlass_library import (
ComplexTransformTag,
DataType,
DataTypeNames,
DataTypeSize,
DataTypeTag,
EpilogueScheduleSuffixes,
EpilogueScheduleTag,
EpilogueScheduleType,
GemmKind,
GemmKindNames,
GemmUniversalMode,
KernelScheduleSuffixes,
KernelScheduleTag,
KernelScheduleType,
LayoutTag,
LayoutType,
MathOperation,
MathOperationTag,
OpcodeClass,
OpcodeClassNames,
OpcodeClassTag,
OperationKind,
ShortComplexLayoutNames,
ShortDataTypeNames,
ShortLayoutTypeNames,
SwizzlingFunctor,
SwizzlingFunctorTag,
TileSchedulerSuffixes,
TileSchedulerTag,
TileSchedulerType,
get_complex_from_real
)
from cutlass_cppgen.backend.arguments import ArgumentBase
from cutlass_cppgen.backend.c_types import (
GemmCoord_,
GemmCoordBatched_,
GenericMainloopArguments3x_,
StrideBatched_,
dim3_,
get_gemm_arguments,
get_gemm_arguments_3x,
get_gemm_arguments_streamk,
get_gemm_grouped_arguments,
get_mainloop_arguments_3x,
get_tile_scheduler_arguments_3x,
)
from cutlass_cppgen.backend.library import (
ApiVersion,
EmissionType,
SchedulerMode,
SchedulerModeTag,
TensorDescription,
TileDescription,
api_version,
)
from cutlass_cppgen.backend.memory_manager import device_mem_alloc, todevice
from cutlass_cppgen.backend.operation import ExecutableOperation, LaunchConfiguration
from cutlass_cppgen.backend.type_hint import GemmOperation, Tensor
from cutlass_cppgen.backend.utils.device import device_sm_count
from cutlass_cppgen.shape import GemmCoord, MatrixCoord
################################################################################
#
# Data structure modeling a GEMM operation
#
################################################################################
def leading_dimension(layout: LayoutType, shape: MatrixCoord) -> int:
"""
Returns the leading dimenson of a tensor with layout ``layout`` and shape ``shape``.
:param layout: layout of the tensor
:type layout: cutlass_cppgen.shape.LayoutType
:param shape: shape of the tensor
:type shape: cutlass_cppgen.shape.MatrixCoord
:return: leading dimension of the tensor
:rtype: int
"""
if layout == LayoutType.RowMajor:
return shape.column
elif layout == LayoutType.ColumnMajor:
return shape.row
def transpose_layout(layout: LayoutType) -> LayoutType:
if layout == LayoutType.ColumnMajor:
return LayoutType.RowMajor
elif layout == LayoutType.RowMajor:
return LayoutType.ColumnMajor
else:
raise ValueError(f"Unsupported Layout {layout}")
class GemmArguments2x(ArgumentBase):
"""
Argument wrapper for GEMM in CUTLASS 2. It encodes problem information and
user-provide tensors into the kernel's argument
:param operation: the GEMM operation to take the argument
:type operation: :class:`cutlass_cppgen.backend.GemmOperationUniversal` |
:class:`cutlass_cppgen.backend.GemmOperationGrouped`
:param problem_size: GEMM problem size gemm(M, N, K)
:type operation: :class:`cutlass_cppgen.shape.GemmCoord`
:param A: tensor A
:type A: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param B: tensor B
:type B: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param C: tensor C
:type C: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param D: tensor D
:type D: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param gemm_mode: GEMM mode
:type gemm_mode: :class:`cutlass_library.GemmUniversalMode`
:param output_op: output operator, optional
:type output_op: :class:`cutlass_cppgen.backend.LinearCombinationFunctorArguments`
:param stream: cuda stream, defaults to cuda.cuda.CUstream(0)
:type stream: :class:`cuda.cuda.CUstream`
"""
def __init__(self, operation, problem_size, A, B, C, D, gemm_mode=GemmUniversalMode.Gemm, **kwargs):
self.operation = operation
self.layout_A = operation.A.layout
self.layout_B = operation.B.layout
self.layout_C = operation.C.layout
self.element_A = operation.A.element
self.element_B = operation.B.element
self.element_C = operation.C.element
if operation.C.layout in [LayoutType.RowMajorInterleaved32, LayoutType.ColumnMajorInterleaved32]:
raise Exception("Interleaved layout not currently supported")
if hasattr(self.operation.epilogue_functor, "visitor") and operation.arch not in [90, 100, 101, 103]:
super().__init__(A, B, None, None, **kwargs)
else:
super().__init__(A, B, C, D, **kwargs)
if operation.switched:
self.problem_size = GemmCoord(problem_size.n, problem_size.m, problem_size.k)
self.ptr_A, self.ptr_B = self.ptr_B, self.ptr_A
else:
self.problem_size = problem_size
# If the number of elements in C = problem_size.n, C is treated as the bias
if hasattr(self, "tensor_c_numel"):
if self.tensor_c_numel == self.problem_size.n and self.problem_size.m != 1:
self.bias = True
self.lda = leading_dimension(self.layout_A, self.problem_size.mk)
self.ldb = leading_dimension(self.layout_B, self.problem_size.kn)
self.ldc = leading_dimension(self.layout_C, self.problem_size.mn)
self.ldd = self.ldc
if self.bias:
self.ldc = 0
if "output_op" in kwargs.keys() and gemm_mode != GemmUniversalMode.GemmSplitKParallel:
self.output_op = kwargs["output_op"]
else:
if self.operation.epilogue_functor.element_epilogue in [DataType.s8, DataType.s32, DataType.u8, DataType.u32]:
dtype = int
else:
dtype = float
self.output_op = self.operation.epilogue_type(dtype(1.0), dtype(0.0))
self.gemm_mode = gemm_mode
if gemm_mode in [GemmUniversalMode.Gemm, GemmUniversalMode.GemmSplitKParallel]:
if "split_k_slices" in kwargs.keys():
self.batch_count = kwargs["split_k_slices"]
else:
self.batch_count = 1
self.split_k_slices = self.batch_count
if gemm_mode in [GemmUniversalMode.Batched, GemmUniversalMode.Array]:
if "batch" in kwargs.keys():
self.batch_count = kwargs["batch"]
else:
self.batch_count = 1
if "batch_strides" in kwargs:
self.batched_stride_A = kwargs["batch_strides"]["A"]
self.batched_stride_B = kwargs["batch_strides"]["B"]
self.batched_stride_C = kwargs["batch_strides"]["C"]
self.batched_stride_D = kwargs["batch_strides"]["D"]
else:
self.batched_stride_A = self.problem_size.m * self.problem_size.k
self.batched_stride_B = self.problem_size.n * self.problem_size.k
self.batched_stride_C = self.problem_size.m * self.problem_size.n
self.batched_stride_D = self.problem_size.m * self.problem_size.n
if self.bias:
self.batched_stride_C = self.problem_size.n
if gemm_mode == GemmUniversalMode.Array:
self.ptr_A_array = []
self.ptr_B_array = []
self.ptr_C_array = []
self.ptr_D_array = []
ptr_A_addr = int(self.ptr_A)
ptr_B_addr = int(self.ptr_B)
ptr_C_addr = int(self.ptr_C)
ptr_D_addr = int(self.ptr_D)
stride_A = self.batched_stride_A * DataTypeSize[self.element_A] // 8
stride_B = self.batched_stride_B * DataTypeSize[self.element_B] // 8
stride_C = self.batched_stride_C * DataTypeSize[self.element_C] // 8
stride_D = self.batched_stride_D * DataTypeSize[self.element_C] // 8
for _ in range(self.batch_count):
self.ptr_A_array.append(ptr_A_addr)
self.ptr_B_array.append(ptr_B_addr)
self.ptr_C_array.append(ptr_C_addr)
self.ptr_D_array.append(ptr_D_addr)
ptr_A_addr += stride_A
ptr_B_addr += stride_B
ptr_C_addr += stride_C
ptr_D_addr += stride_D
self.ptr_A_array_buffer = todevice(self.ptr_A_array, dtype=np.int64)
self.ptr_B_array_buffer = todevice(self.ptr_B_array, dtype=np.int64)
self.ptr_C_array_buffer = todevice(self.ptr_C_array, dtype=np.int64)
self.ptr_D_array_buffer = todevice(self.ptr_D_array, dtype=np.int64)
if isinstance(self.operation, GemmOperationUniversal):
self.initialize()
def get_arguments(self):
problem_size_ = self.problem_size.ctype
grid_tiled_shape_ = GemmCoord(
self.grid_tiled_shape.x,
self.grid_tiled_shape.y,
self.grid_tiled_shape.z ).ctype
if self.gemm_mode == GemmUniversalMode.Array:
arguments = self.operation.argument_type(
# Arguments from UniversalArgumentsBase
self.gemm_mode,
problem_size_,
self.batch_count,
0,
# Remaining arguments
self.output_op,
int(self.ptr_A_array_buffer.ptr),
int(self.ptr_B_array_buffer.ptr),
int(self.ptr_C_array_buffer.ptr),
int(self.ptr_D_array_buffer.ptr),
0, 0, 0,
self.lda, self.ldb, self.ldc, self.ldd,
self.lda, self.ldb, self.ldc, self.ldd,
0, 0, 0
)
else:
arguments = self.operation.argument_type(
# Arguments from UniversalArgumentsBase
self.gemm_mode, problem_size_, self.batch_count, self.batched_stride_D,
# Remaining arguments
self.output_op,
int(self.ptr_A),
int(self.ptr_B),
int(self.ptr_C),
int(self.ptr_D),
self.batched_stride_A,
self.batched_stride_B,
self.batched_stride_C,
self.lda, self.ldb, self.ldc, self.ldd,
self.lda, self.ldb, self.ldc, self.ldd,
0, 0, 0
)
self.arguments = arguments, grid_tiled_shape_, self.gemm_k_size
def initialize(self):
launch_config = self.operation.rt_module.plan(self)
# Get the host and device workspace
device_workspace_size = self.operation.rt_module.get_device_workspace_size(self)
if device_workspace_size > 0:
self.workspace_buffer = device_mem_alloc(device_workspace_size)
workspace_ptr = self.workspace_buffer.ptr
err, = cuda.cuMemsetD32(
workspace_ptr, 0, device_workspace_size // 4)
else:
workspace_ptr = None
device_workspace = 0
if workspace_ptr is not None and self.gemm_mode == GemmUniversalMode.GemmSplitKParallel:
# In GEMM splik-K parallel, the D pointer is redirected to the workspace
self.ptr_D = cuda.CUdeviceptr(workspace_ptr)
elif workspace_ptr is not None and self.gemm_mode == GemmUniversalMode.Gemm:
device_workspace = workspace_ptr
self.get_arguments()
arguments, grid_tiled_shape, gemm_k_size = self.arguments
res_arg = self.operation.rt_module.get_args(
ctypes.byref(arguments), ctypes.c_void_p(int(device_workspace)))
host_workspace = bytearray(res_arg.contents)
device_workspace = None
self.host_workspace = host_workspace
self.device_workspace = device_workspace
self.launch_config = launch_config
def sync(self, stream_sync=True):
super().sync(stream_sync)
if hasattr(self.output_op, "sync"):
self.output_op.sync()
class GemmArguments2xStreamK(GemmArguments2x):
"""
Argument wrapper for stream-K GEMMs in CUTLASS 2. It encodes problem information and
user-provide tensors into the kernel's argument
:param operation: the GEMM operation to take the argument
:type operation: :class:`cutlass_cppgen.backend.GemmOperationUniversal` |
:class:`cutlass_cppgen.backend.GemmOperationGrouped`
:param problem_size: GEMM problem size gemm(M, N, K)
:type operation: :class:`cutlass_cppgen.shape.GemmCoord`
:param A: tensor A
:type A: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param B: tensor B
:type B: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param C: tensor C
:type C: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param D: tensor D
:type D: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param gemm_mode: GEMM mode
:type gemm_mode: :class:`cutlass_library.GemmUniversalMode`
:param output_op: output operator, optional
:type output_op: :class:`cutlass_cppgen.backend.LinearCombinationFunctorArguments`
"""
def __init__(self, operation, problem_size, A, B, C, D, gemm_mode=GemmUniversalMode.Gemm, **kwargs):
if gemm_mode not in [GemmUniversalMode.Gemm, GemmUniversalMode.Batched]:
raise Exception(f"Unsupported GEMM mode {gemm_mode}.")
super().__init__(operation, problem_size, A, B, C, D, gemm_mode, **kwargs)
def get_arguments(self):
batch_stride_A = self.problem_size.m * self.problem_size.k
batch_stride_B = self.problem_size.k * self.problem_size.n
batch_stride_C = self.problem_size.m * self.problem_size.n
batch_stride_D = self.problem_size.m * self.problem_size.n
arguments = self.operation.argument_type(
self.gemm_mode,
GemmCoord_(self.problem_size.m, self.problem_size.n, self.problem_size.k),
self.batch_count,
self.output_op,
int(self.ptr_A),
int(self.ptr_B),
int(self.ptr_C),
int(self.ptr_D),
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_stride_D,
self.lda, self.ldb, self.ldc, self.ldd, # strides
self.lda, self.ldb, self.ldc, self.ldd,
-1, # avail_sms
)
return arguments
def initialize(self):
# Get the host and device workspace
device_workspace_size = self.operation.rt_module.get_device_workspace_size(
self,
device_sm_count(),
self.operation.rt_module.occupancy
)
if device_workspace_size > 0:
self.workspace_buffer = device_mem_alloc(device_workspace_size)
workspace_ptr = self.workspace_buffer.ptr
err, = cuda.cuMemsetD32(
workspace_ptr, 0, device_workspace_size // 4)
else:
workspace_ptr = None
device_workspace = 0
if workspace_ptr is not None and self.gemm_mode == GemmUniversalMode.GemmSplitKParallel:
# In GEMM splik-K parallel, the D pointer is redirected to the workspace
self.ptr_D = cuda.CUdeviceptr(workspace_ptr)
elif workspace_ptr is not None and self.gemm_mode == GemmUniversalMode.Gemm:
device_workspace = workspace_ptr
arguments = self.get_arguments()
res_arg = self.operation.rt_module.get_args(
ctypes.byref(arguments),
ctypes.c_void_p(int(device_workspace)),
device_sm_count(),
self.operation.rt_module.occupancy
)
host_workspace = bytearray(res_arg.contents)
grid = self.operation.rt_module.get_grid_shape(
ctypes.byref(arguments),
device_sm_count(),
self.operation.rt_module.occupancy
)
device_workspace = None
self.host_workspace = host_workspace
self.device_workspace = device_workspace
self.launch_config = LaunchConfiguration(
[grid.m, grid.n, grid.k],
[self.operation.rt_module.threads, 1, 1],
self.operation.rt_module.shared_memory_capacity
)
class GemmArguments3x(GemmArguments2x):
"""
Argument wrapper for GEMM in CUTLASS 3. It encodes problem information and
user-provide tensors into the kernel's argument
:param operation: the GEMM operation to take the argument
:type operation: :class:`cutlass_cppgen.backend.GemmOperationUniversal` |
:class:`cutlass_cppgen.backend.GemmOperationGrouped`
:param problem_size: GEMM problem size gemm(M, N, K)
:type operation: :class:`cutlass_cppgen.shape.GemmCoord`
:param A: tensor A
:type A: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param B: tensor B
:type B: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param C: tensor C
:type C: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param D: tensor D
:type D: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param gemm_mode: GEMM mode
:type gemm_mode: GemmUniversalMode
:param output_op: output operator, optional
:type output_op: :class:`cutlass_cppgen.backend.LinearCombinationFunctorArguments`
"""
def __init__(self, operation, problem_size, A, B, C, D, gemm_mode=GemmUniversalMode.Gemm, **kwargs):
if gemm_mode not in [GemmUniversalMode.Gemm, GemmUniversalMode.Batched]:
raise Exception(f"Unsupported GEMM mode {gemm_mode}.")
super().__init__(operation, problem_size, A, B, C, D, gemm_mode, **kwargs)
def get_arguments(self):
mainloop_args = get_mainloop_arguments_3x(
self.operation.tile_description.kernel_schedule,
self.operation.A.element,
self.operation.B.element,
self.operation.A.alignment,
self.operation.B.alignment
)
scheduler_args = get_tile_scheduler_arguments_3x(self.operation.tile_description.tile_scheduler)
uses_default_epilogue = self.operation.rt_module.uses_default_epilogue()
argument_type, epilogue_args, epilogue_type, hw_info = get_gemm_arguments_3x(
mainloop_args, self.operation.epilogue_functor, scheduler_args, uses_default_epilogue)
problem_size_ = GemmCoordBatched_(self.problem_size, self.batch_count)
if self.batch_count > 1:
bsA = self.batched_stride_A
bsB = self.batched_stride_B
bsC = self.batched_stride_C
bsD = self.batched_stride_D
else:
bsA = 0
bsB = 0
bsC = 0
bsD = 0
stride_A = StrideBatched_(self.lda, bsA)
stride_B = StrideBatched_(self.ldb, bsB)
stride_C = StrideBatched_(self.ldc, bsC)
stride_D = StrideBatched_(self.ldd, bsD)
# Superset of potential mainloop arguments
generic_args = GenericMainloopArguments3x_(
int(self.ptr_A),
stride_A,
int(self.ptr_B),
stride_B,
4 # mma_promotion_interval
)
# Set of mainloop arguments needed for this kernel
mainloop = mainloop_args.from_generic_mainloop_args(generic_args)
if not uses_default_epilogue and hasattr(self.output_op, "to_evt_params"):
self.output_op = self.output_op.to_evt_params()
epilogue = epilogue_args(
self.output_op,
int(self.ptr_C),
stride_C,
int(self.ptr_D),
stride_D,
)
# Set hardware info
hw_info_ = hw_info(
0, device_sm_count(), 0,
dim3_(0,0,0),
dim3_(0,0,0),
)
self.arguments = argument_type(
int(self.gemm_mode),
problem_size_,
mainloop,
epilogue,
hw_info_,
scheduler_args
)
return self.arguments
def initialize(self):
# Get the host and evice workspace
device_workspace_size = self.operation.rt_module.get_device_workspace_size(self)
if device_workspace_size > 0:
self.workspace_buffer = device_mem_alloc(device_workspace_size)
workspace_ptr = self.workspace_buffer.ptr
err, = cuda.cuMemsetD32(
workspace_ptr, 0, device_workspace_size // 4)
else:
workspace_ptr = None
device_workspace = 0
if workspace_ptr is not None and self.gemm_mode == GemmUniversalMode.GemmSplitKParallel:
# In GEMM splik-K parallel, the D pointer is redirected to the workspace
self.ptr_D = cuda.CUdeviceptr(workspace_ptr)
elif workspace_ptr is not None and self.gemm_mode == GemmUniversalMode.Gemm:
device_workspace = workspace_ptr
self.get_arguments()
res_arg = self.operation.rt_module.get_args(
ctypes.byref(self.arguments),
ctypes.c_void_p(int(device_workspace)),
)
host_workspace = bytearray(res_arg.contents)
grid = self.operation.rt_module.get_grid_shape(
ctypes.byref(self.arguments),
ctypes.c_void_p(int(device_workspace)),
)
block = self.operation.rt_module.get_block_shape()
device_workspace = None
self.host_workspace = host_workspace
self.device_workspace = device_workspace
self.launch_config = LaunchConfiguration(
[grid.x, grid.y, grid.z],
[block.x, block.y, block.z],
self.operation.rt_module.shared_memory_capacity,
)
def GemmArguments(operation, problem_size, A, B, C, D, gemm_mode=GemmUniversalMode.Gemm, **kwargs):
"""
Argument wrapper for GEMM in CUTLASS 2 or 3. It returns either 2x arguments
or 3x arguments depending on the `arch` field specified in `operation`.
:param operation: the GEMM operation to take the argument
:type operation: :class:`cutlass_cppgen.backend.GemmOperationUniversal` |
:class:`cutlass_cppgen.backend.GemmOperationGrouped`
:param problem_size: GEMM problem size gemm(M, N, K)
:type operation: :class:`cutlass_cppgen.shape.GemmCoord`
:param A: tensor A
:type A: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param B: tensor B
:type B: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param C: tensor C
:type C: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param D: tensor D
:type D: cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray
:param gemm_mode: GEMM mode
:type gemm_mode: :class:`cutlass_library.GemmUniversalMode`
:param output_op: output operator, optional
:type output_op: :class:`cutlass_cppgen.backend.LinearCombinationFunctorArguments`
"""
if operation.swizzling_functor == SwizzlingFunctor.StreamK:
if operation.api == ApiVersion.v3x:
raise Exception("Stream K is currently only supported in CUTLASS 2.x")
ArgClass = GemmArguments2xStreamK
else:
ArgClass = GemmArguments3x if operation.api == ApiVersion.v3x else GemmArguments2x
return ArgClass(operation, problem_size, A, B, C, D, gemm_mode, **kwargs)
class GemmGroupedArguments:
"""
Argument wrapper for GEMM Grouped. It encodes problem information and
user-provide tensors into the kernel's argument
:param operation: the GEMM Grouped operation to take the argument
:type operation: :class:`cutlass_cppgen.backend.GemmOperationGrouped`
:param problem_size: list of GEMM problem size gemm(M, N, K)
:type operation: list[:class:`cutlass_cppgen.shape.GemmCoord`]
:param A: list of tensor A
:type A: list[cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray]
:param B: list of tensor B
:type B: list[cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray]
:param C: list of tensor C
:type C: list[cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray]
:param D: list of tensor D
:type D: list[cuda.CUdeviceptr | numpy.ndarray | torch.Tensor | cupy.ndarray]
:param output_op: output operator, optional
:type output_op: :class:`cutlass_cppgen.backend.LinearCombinationFunctorArguments`
:param stream: cuda stream, defaults to cuda.cuda.CUstream(0)
:type stream: :class:`cuda.cuda.CUstream`
"""
def __init__(self, operation, problem_sizes, A, B, C, D, **kwargs):
# Get number of problems in the group
self.problem_count = len(problem_sizes)
# Check the input arguments
assert len(A) == self.problem_count
assert len(B) == self.problem_count
assert len(C) == self.problem_count
assert len(D) == self.problem_count
problem_size_host = []
self.ptr_A_host = []
self.ptr_B_host = []
self.ptr_C_host = []
self.ptr_D_host = []
lda_host = []
ldb_host = []
ldc_host = []
ldd_host = []
self.partitions = 1
self.operation = operation
# Get the threadblock
threadblock_shape = operation.tile_description.threadblock_shape
self.threadblock_shape = GemmCoord(
threadblock_shape[0],
threadblock_shape[1],
threadblock_shape[2],
)
self.threadblock_swizzle = operation.swizzling_functor
self.total_tiles = 0
self.gemm_arguments = []
self.stream = kwargs.get("stream", cuda.CUstream(0))
# Process the input arguments
for idx, problem_size in enumerate(problem_sizes):
M, N, K = problem_size.m, problem_size.n, problem_size.k
temp_argument = GemmArguments2x(
operation=operation,
problem_size=GemmCoord(M, N, K),
A=A[idx], B=B[idx], C=C[idx], D=D[idx])
self.gemm_arguments.append(temp_argument)
problem_size_host.append(
[temp_argument.problem_size.m,
temp_argument.problem_size.n,
temp_argument.problem_size.k]
)
self.ptr_A_host.append(int(temp_argument.ptr_A))
lda_host.append(temp_argument.lda)
self.ptr_B_host.append(int(temp_argument.ptr_B))
ldb_host.append(temp_argument.ldb)
self.ptr_C_host.append(int(temp_argument.ptr_C))
ldc_host.append(temp_argument.ldc)
self.ptr_D_host.append(int(temp_argument.ptr_D))
ldd_host.append(temp_argument.ldd)
# Get number of tiles
grid = self.operation.rt_module.get_grid_shape(
self.operation.rt_module.get_tiled_shape(
temp_argument.problem_size.ctype,
self.threadblock_shape.ctype,
temp_argument.batch_count
)
)
self.total_tiles += grid.x * grid.y * grid.z
self.problem_size_buffer = todevice(problem_size_host, np.int32)
self.ptr_A_buffer = todevice(self.ptr_A_host, np.int64)
self.ptr_B_buffer = todevice(self.ptr_B_host, np.int64)
self.ptr_C_buffer = todevice(self.ptr_C_host, np.int64)
self.ptr_D_buffer = todevice(self.ptr_D_host, np.int64)
self.lda_buffer = todevice(lda_host, np.int64)
self.ldb_buffer = todevice(ldb_host, np.int64)
self.ldc_buffer = todevice(ldc_host, np.int64)
self.ldd_buffer = todevice(ldd_host, np.int64)
if "output_op" in kwargs.keys():
self.alpha = kwargs["output_op"].alpha
self.beta = kwargs["output_op"].beta
else:
self.alpha = 1.0
self.beta = 0.0
if "output_op" in kwargs.keys():
self.output_op = kwargs["output_op"]
else:
self.output_op = self.operation.epilogue_type(1.0, 0.0)
# Get host problem size
self.host_problem_size_ptr = np.array(problem_size_host, dtype=np.int32).__array_interface__["data"][0]
self.arguments = self.get_arguments()
self.initialize()
def get_arguments(self):
return self.operation.argument_type(
self.problem_size_buffer.ptr,
self.problem_count,
self.total_tiles,
self.output_op,
self.ptr_A_buffer.ptr,
self.ptr_B_buffer.ptr,
self.ptr_C_buffer.ptr,
self.ptr_D_buffer.ptr,
self.lda_buffer.ptr,
self.ldb_buffer.ptr,
self.ldc_buffer.ptr,
self.ldd_buffer.ptr,
ctypes.c_void_p(int(self.host_problem_size_ptr)),
)
def initialize(self):
# Get launch configuration
launch_config = self.operation.rt_module.plan(self)
# Get the host and evice workspace
device_workspace_size = self.operation.rt_module.get_device_workspace_size(self)
if device_workspace_size > 0:
self.workspace_buffer = device_mem_alloc(device_workspace_size)
workspace_ptr = self.workspace_buffer.ptr
err, = cuda.cuMemsetD32(
workspace_ptr, 0, device_workspace_size // 4)
else:
workspace_ptr = None
if self.operation.precompute_mode == SchedulerMode.Host:
device_workspace_ptr = self.operation.rt_module.host_precompute(
self, self.operation.rt_module.get_workspace_size(self),)
else:
device_workspace_ptr = 0
result = self.operation.rt_module.get_args(
ctypes.byref(self.arguments),
self.total_tiles,
ctypes.c_void_p(int(device_workspace_ptr)),
)
host_workspace = bytearray(result.contents)
device_workspace = None
self.host_workspace = host_workspace
self.device_workspace = device_workspace
self.launch_config = launch_config
def sync(self):
err, = cudart.cudaDeviceSynchronize()
if err != cuda.CUresult.CUDA_SUCCESS:
raise RuntimeError("CUDA Error %s" % str(err))
for arg in self.gemm_arguments:
arg.sync(stream_sync=False)
################################################################################
# Base class for GEMM runtime module
################################################################################
class GemmRTbase(ExecutableOperation):
"""
GemmRT manages the CUTLASS runtime components
"""
KernelTemplate = r"""
extern "C"
__global__ void
${operation_name}(${operation_name}${operation_suffix}::Params params) {
// Dynamic shared memory base pointer
extern __shared__ int SharedStorageBase[];
// Declare pointer to dynamic shared memory.
${operation_name}${operation_suffix}::SharedStorage *shared_storage =
reinterpret_cast<${operation_name}${operation_suffix}::SharedStorage *>(SharedStorageBase);
${operation_name}${operation_suffix}::invoke(params, *shared_storage);
}
"""
def __init__(self, operation: "GemmOperation"):
super().__init__(operation)
self.operation = operation
threadblock_shape = operation.tile_description.threadblock_shape
self.threadblock_shape = GemmCoord(
threadblock_shape[0], threadblock_shape[1], threadblock_shape[2])
self.threadblock_swizzle = operation.swizzling_functor
# Threads per threadblock
self.threads = operation.tile_description.num_threads
def emit(self):
return self.emitter.emit(self.operation)
def can_implement(self, configuration, arguments):
raise NotImplementedError()
def get_host_workspace_size(self, arguments):
raise NotImplementedError()
def get_device_workspace_size(self, arguments):
return 0
def initialize(self):
err, = cuda.cuFuncSetAttribute(
self.kernel,
attrib=cuda.CUfunction_attribute.CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
value=self.shared_memory_capacity)
if err != cuda.CUresult.CUDA_SUCCESS:
raise RuntimeError(
f"CUDA error on call to cuFuncSetAttribute: {cuda.cuGetErrorString(err)[1]}"
)
################################################################################
# Runtime module for GEMM Universal
################################################################################
class GemmRTUniversal(GemmRTbase):
"""
GemmRTUniversal manages the CUTLASS runtime components
"""
HostTemplate = r"""
extern "C" {
// Get the size of params in bytes
int ${operation_name}_get_param_size(){
return sizeof(${operation_name}${operation_suffix}::Params);
}
// Get the size of dynamic shared memory in bytes
int ${operation_name}_shared_memory_size() {
return int(sizeof(${operation_name}${operation_suffix}::SharedStorage));
}
// Get the params as byte array
char* ${operation_name}_get_params(${operation_name}_base::Arguments* argument, int* workspace){
${operation_name}_base::Params* params;
params = new ${operation_name}_base::Params(*argument,
-1, // SM count. Only used for stream-K
-1 // Occupancy. Only used for stream-K
);
// Semaphore holds the pointer to the workspace in the Params struct
params->semaphore = workspace;
char *bytes = ((char*)(params));
char *output = new char[sizeof(${operation_name}_base::Params)];
for (unsigned int i = 0; i < sizeof(${operation_name}_base::Params); i ++)
output[i] = bytes[i];
return output;
}
cutlass::gemm::GemmCoord ${operation_name}_get_tiled_shape(
cutlass::gemm::GemmCoord problem_size, cutlass::gemm::GemmCoord tile_size, int split_k_slices) {
return ${operation_name}_base::ThreadblockSwizzle::get_tiled_shape(
problem_size, tile_size, split_k_slices);
}
dim3 ${operation_name}_get_grid_shape(cutlass::gemm::GemmCoord tiled_shape) {
return ${operation_name}_base::ThreadblockSwizzle::get_grid_shape(tiled_shape);
}
}
"""
def __init__(self, operation):
super(GemmRTUniversal, self).__init__(operation)
self.extra_funcs = {
"get_tiled_shape": GemmCoord_,
"get_grid_shape": dim3_,
}
self.emitter = EmitGemmUniversalInstance(
"_type", operation.direct_store)
self.argument_type, self.epilogue_type = get_gemm_arguments(operation.epilogue_functor)
self.argtype = [
ctypes.POINTER(self.argument_type),
ctypes.POINTER(GemmCoord_), ctypes.c_int, ctypes.c_void_p
]
def plan(self, arguments):
grid = self.get_tiled_shape(
arguments.problem_size.ctype,
self.threadblock_shape.ctype,
arguments.batch_count
)
gemm_k_size = arguments.problem_size.k
if arguments.gemm_mode in [GemmUniversalMode.Gemm, GemmUniversalMode.GemmSplitKParallel]:
alignk = max(max(128 // DataTypeSize[self.operation.A.element],
128 // DataTypeSize[self.operation.B.element]), 1)
gemm_k_size = (((arguments.problem_size.k + arguments.batch_count - 1) //
arguments.batch_count + alignk - 1) // alignk) * alignk
if gemm_k_size:
grid_z = (arguments.problem_size.k + gemm_k_size - 1) // gemm_k_size
grid = GemmCoord(grid.m, grid.n, grid_z).ctype
arguments.grid_tiled_shape = dim3_(grid.m, grid.n, grid.k)
grid = self.get_grid_shape(grid)
arguments.gemm_k_size = gemm_k_size
return LaunchConfiguration(
[grid.x, grid.y, grid.z],
[self.threads, 1, 1],
self.shared_memory_capacity)
def get_device_workspace_size(self, arguments: GemmArguments):
workspace_bytes = 0
if arguments.gemm_mode == GemmUniversalMode.GemmSplitKParallel:
workspace_bytes = (DataTypeSize[arguments.operation.C.element]
* arguments.batched_stride_D * arguments.grid_tiled_shape.z // 8)
elif (arguments.gemm_mode == GemmUniversalMode.Gemm and
arguments.split_k_slices > 1):
workspace_bytes = 4 * arguments.grid_tiled_shape.x * arguments.grid_tiled_shape.y
return workspace_bytes
class GemmRTUniversalStreamK(GemmRTUniversal):
"""
Manages the CUTLASS runtime components for 2.x stream K kernels
"""
HostTemplate = r"""
extern "C" {
// Get the size of params in bytes
int ${operation_name}_get_param_size(){
return sizeof(${operation_name}${operation_suffix}::Params);
}
// Get the size of dynamic shared memory in bytes
int ${operation_name}_shared_memory_size() {
return int(sizeof(${operation_name}${operation_suffix}::SharedStorage));
}
using GemmType = ${operation_name}_base;
// Get the params as byte array
char* ${operation_name}_get_params(GemmType::Arguments* argument, int* workspace,
int sm_count, int occupancy) {
GemmType::Params* params;
params = new GemmType::Params(*argument, sm_count, occupancy);
params->init_workspace(workspace);
char *bytes = ((char*)(params));
char *output = new char[sizeof(GemmType::Params)];
for (unsigned int i = 0; i < sizeof(GemmType::Params); i ++)
output[i] = bytes[i];
return output;
}
dim3 ${operation_name}_get_grid_shape(GemmType::Arguments* args, int device_sms, int sm_occupancy) {
typename GemmType::Params params(*args, device_sms, sm_occupancy);
return params.get_grid_dims();
}
uint64_t ${operation_name}_get_kernel_workspace_size(GemmType::Arguments* args, int device_sms, int sm_occupancy) {
typename GemmType::Params params(*args, device_sms, sm_occupancy);
return params.get_workspace_size();
}
}
"""
def __init__(self, operation: "GemmOperation"):
super(GemmRTUniversalStreamK, self).__init__(operation)
self.extra_funcs = {
"get_grid_shape": GemmCoord_,
"get_kernel_workspace_size": ctypes.c_uint64,
}
self._occupancy = None
self.argument_type, self.epilogue_type = get_gemm_arguments_streamk(operation.epilogue_functor)
@property
def occupancy(self):
if self._occupancy is None:
err, self._occupancy = cuda.cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(
self.kernel, self.threads, self.shared_memory_capacity,
cuda.CUoccupancy_flags.CU_OCCUPANCY_DISABLE_CACHING_OVERRIDE)
if err != cuda.CUresult.CUDA_SUCCESS:
raise RuntimeError(
"CUDA error on call to cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags: "
f"{cuda.cuGetErrorString(err)[1]}")
return self._occupancy
def get_device_workspace_size(self, arguments: GemmArguments2xStreamK, device_sms: int, sm_occupancy: int):
return self.get_kernel_workspace_size(ctypes.byref(arguments.get_arguments()), device_sms, sm_occupancy)
################################################################################
# Runtime module for GEMM Universal within CUTLASS 3
################################################################################
class GemmRTUniversal3x(GemmRTUniversal):
"""
Manages the CUTLASS runtime components for 3.x kernels
"""
KernelTemplate = r"""
using Operator = ${operation_name}${operation_suffix};
extern "C"
__global__ __launch_bounds__(Operator::MaxThreadsPerBlock, Operator::MinBlocksPerMultiprocessor)
void ${operation_name}(__grid_constant__ typename Operator::Params const params) {
// Dynamic shared memory base pointer
extern __shared__ char smem[];
// Declare pointer to dynamic shared memory.
Operator op;
op(params, smem);
}
"""
HostTemplate = r"""
extern "C" {
// Get the size of params in bytes
int ${operation_name}_get_param_size(){
return sizeof(${operation_name}${operation_suffix}::Params);
}
// Get the size of dynamic shared memory in bytes
int ${operation_name}_shared_memory_size() {
return ${operation_name}${operation_suffix}::SharedStorageSize;
}
using GemmType = ${operation_name}_base;
bool ${operation_name}_uses_default_epilogue() {
return std::is_same_v<GemmType::CollectiveEpilogue::DispatchPolicy, cutlass::gemm::EpilogueDefault>;
}
// Get the workspace size
uint64_t ${operation_name}_get_kernel_workspace_size(GemmType::Arguments* argument) {
return GemmType::get_workspace_size(*argument);
}
// Get the params as byte array
char* ${operation_name}_get_params(GemmType::Arguments* argument, int* workspace){
GemmType::Params params = GemmType::to_underlying_arguments(*argument, workspace);
char *bytes = ((char*)(&params));
char *output = new char[sizeof(GemmType::Params)];
for (unsigned int i = 0; i < sizeof(GemmType::Params); i ++)
output[i] = bytes[i];
return output;
}
// Get the total number of blocks for a persistent kernel
uint64_t ${operation_name}_get_persistent_tiled_blk_shape_mnl(GemmType::ProblemShape problem) {
auto problem_shape_MNKL = append<4>(problem, Int<1>{});
auto [problem_blocks_m, problem_blocks_n, problem_blocks_l] =
cutlass::gemm::kernel::detail::PersistentTileSchedulerSm90::get_tiled_cta_shape_mnl(
problem_shape_MNKL, GemmType::TileShape{}, GemmType::DispatchPolicy::ClusterShape{});
return problem_blocks_m * problem_blocks_n * problem_blocks_l;
}
// Get the grid shape
dim3 ${operation_name}_get_grid_shape(GemmType::Arguments* args, int* workspace) {
auto tmp_params = GemmType::to_underlying_arguments(*args, workspace);
return GemmType::get_grid_shape(tmp_params);
}
// Get the block shape
dim3 ${operation_name}_get_block_shape() {
return GemmType::get_block_shape();
}
}
"""
def __init__(self, operation):
super(GemmRTUniversal3x, self).__init__(operation)
self.extra_funcs = {
"get_grid_shape": dim3_,
"get_block_shape": dim3_,
"get_persistent_tiled_blk_shape_mnl": ctypes.c_uint64,
"get_kernel_workspace_size": ctypes.c_uint64,
"uses_default_epilogue": ctypes.c_bool,
}
self.emitter = EmitGemmUniversalInstance3x("_type")
def get_device_workspace_size(self, arguments: GemmArguments3x):
return self.get_kernel_workspace_size(ctypes.byref(arguments.get_arguments()))
class EmitGemmUniversalInstance3x:
"""Responsible for emitting a CUTLASS 3 template definition"""
def __init__(self, operation_suffix=""):
self.operation_suffix = operation_suffix
self.includes = [
"cutlass/cutlass.h",
"cute/tensor.hpp",
"cute/atom/mma_atom.hpp",
"cutlass/numeric_types.h",
"cutlass/gemm/collective/collective_builder.hpp",
"cutlass/gemm/kernel/sm90_tile_scheduler.hpp",
"cutlass/gemm/kernel/gemm_universal.hpp",
"cutlass/epilogue/collective/collective_builder.hpp",
"cutlass/epilogue/collective/default_epilogue.hpp",
"cutlass/epilogue/thread/linear_combination.h"
]
self.gemm_template_kernel = """
using namespace cute;
using CollectiveEpilogue =
typename cutlass::epilogue::collective::CollectiveBuilder<
${arch}, ${opcode_class},
cute::Shape<cute::_${threadblock_shape_m}, cute::_${threadblock_shape_n}, cute::_${threadblock_shape_k}>,
cute::Shape<cute::_${cluster_m},cute::_${cluster_n},cute::_${cluster_k}>,
cutlass::epilogue::collective::EpilogueTileAuto,
${element_accumulator}, ${element_epilogue},
${element_c}, ${layout_c}, ${align_c},
${element_d}, ${layout_d}, ${align_d},
${epilogue_schedule}
>::CollectiveOp;
using CollectiveMainloop =
typename cutlass::gemm::collective::CollectiveBuilder<
${arch}, ${opcode_class},
${element_a}, ${layout_a}, ${align_a},
${element_b}, ${layout_b}, ${align_b},
${element_accumulator},
cute::Shape<cute::_${threadblock_shape_m}, cute::_${threadblock_shape_n}, cute::_${threadblock_shape_k}>,
cute::Shape<cute::_${cluster_m},cute::_${cluster_n},cute::_${cluster_k}>,
${stage_count_type},
${kernel_schedule}
>::CollectiveOp;
// Gemm operator ${operation_name}
using ${operation_name}_base = cutlass::gemm::kernel::GemmUniversal<
Shape<int,int,int,int>,
CollectiveMainloop,
CollectiveEpilogue,
${tile_scheduler}
>;
// Define named type
struct ${operation_name}${operation_suffix} :
public ${operation_name}_base { };
"""
self.gemm_template_kernel_visitor = """
using namespace cute;
${callback_decl}
using CollectiveEpilogue =
typename cutlass::epilogue::collective::CollectiveBuilder<
${arch}, ${opcode_class},
cute::Shape<cute::_${threadblock_shape_m}, cute::_${threadblock_shape_n}, cute::_${threadblock_shape_k}>,
cute::Shape<cute::_${cluster_m},cute::_${cluster_n},cute::_${cluster_k}>,
cutlass::epilogue::collective::EpilogueTileAuto,
${element_accumulator}, ${element_epilogue},
ElementC, StrideC, ${align_c},
ElementD, StrideD, ${align_d},
${epilogue_schedule},
${callback_name}
>::CollectiveOp;
using CollectiveMainloop =
typename cutlass::gemm::collective::CollectiveBuilder<
${arch}, ${opcode_class},
${element_a}, ${layout_a}, ${align_a},
${element_b}, ${layout_b}, ${align_b},
${element_accumulator},
cute::Shape<cute::_${threadblock_shape_m}, cute::_${threadblock_shape_n}, cute::_${threadblock_shape_k}>,
cute::Shape<cute::_${cluster_m},cute::_${cluster_n},cute::_${cluster_k}>,
${stage_count_type},
${kernel_schedule}
>::CollectiveOp;
// Gemm operator ${operation_name}
using ${operation_name}_base = cutlass::gemm::kernel::GemmUniversal<
Shape<int,int,int,int>,
CollectiveMainloop,
CollectiveEpilogue,
${tile_scheduler}
>;
// Define named type
struct ${operation_name}${operation_suffix} :
public ${operation_name}_base { };
"""
self.gemm_template_device = self.gemm_template_kernel + """
// Define device-level operator
using DeviceKernel = cutlass::gemm::device::GemmUniversalAdapter<${operation_name}${operation_suffix}>;
"""
def emit(self, operation):
# Support built-in epilogue functors or user-defined functions
if operation.tile_description.stages is None or operation.tile_description.stages == 0:
stage_count_type = "cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>"
else:
stage_count_type = "_" + str(operation.tile_description.stages)
if operation.emission_type == EmissionType.Kernel:
gemm_template = self.gemm_template_kernel
else:
gemm_template = self.gemm_template_device
kschedule = KernelScheduleType.ScheduleAuto
eschedule = EpilogueScheduleType.ScheduleAuto
tschedule = TileSchedulerType.Default
if operation.tile_description.kernel_schedule is not None:
kschedule = operation.tile_description.kernel_schedule
if operation.tile_description.epilogue_schedule is not None:
eschedule = operation.tile_description.epilogue_schedule
if operation.tile_description.tile_scheduler is not None:
tschedule = operation.tile_description.tile_scheduler
emit_tile_m, emit_tile_n, emit_tile_k = operation.tile_description.blackwell_threadblock_shape
values = {
"operation_name": operation.procedural_name(),
"operation_suffix": self.operation_suffix,
"element_a": DataTypeTag[operation.A.element],
"layout_a": LayoutTag[operation.A.layout],
"element_b": DataTypeTag[operation.B.element],
"layout_b": LayoutTag[operation.B.layout],
"element_c": DataTypeTag[operation.C.element],
"layout_c": LayoutTag[operation.C.layout],
"element_d": DataTypeTag[operation.epilogue_functor.element_output],
"layout_d": LayoutTag[operation.C.layout],
"element_accumulator": DataTypeTag[operation.accumulator_type()],
"element_epilogue": DataTypeTag[operation.epilogue_functor.element_epilogue],
"opcode_class": OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
"arch": "cutlass::arch::Sm%d" % operation.arch,
"threadblock_shape_m": str(emit_tile_m),
"threadblock_shape_n": str(emit_tile_n),
"threadblock_shape_k": str(emit_tile_k),
"cluster_m": str(operation.tile_description.cluster_shape[0]),
"cluster_n": str(operation.tile_description.cluster_shape[1]),
"cluster_k": str(operation.tile_description.cluster_shape[2]),
"align_a": str(operation.A.alignment),
"align_b": str(operation.B.alignment),
"align_c": str(operation.C.alignment),
"align_d": str(operation.C.alignment),
"stage_count_type": stage_count_type,
"kernel_schedule": KernelScheduleTag[kschedule],
"epilogue_schedule": EpilogueScheduleTag[eschedule],
"tile_scheduler": TileSchedulerTag[tschedule]
}
if hasattr(operation.epilogue_functor, "visitor"):
callback_name, callback_decl = operation.epilogue_functor.emit(operation)
values["callback_name"] = callback_name
values["callback_decl"] = callback_decl
return SubstituteTemplate(self.gemm_template_kernel_visitor, values)
else:
values["epilogue_functor"] = operation.epilogue_functor.emit()
return SubstituteTemplate(gemm_template, values)
###################################################################################################
# Runtime module for GEMM Grouped
###################################################################################################
class GemmRTGrouped(GemmRTbase):
"""
GemmRTGrouped manages the CUTLASS runtime components
"""
KernelTemplate = r"""
extern "C"
__global__ void
${operation_name}(${operation_name}${operation_suffix}::Params params) {
// Dynamic shared memory base pointer
extern __shared__ int SharedStorageBase[];
// Declare pointer to dynamic shared memory.
${operation_name}${operation_suffix}::SharedStorage *shared_storage =
reinterpret_cast<${operation_name}${operation_suffix}::SharedStorage *>(SharedStorageBase);
${operation_name}${operation_suffix} op;
op(params, *shared_storage);
}
"""
HostTemplate = r"""
extern "C" {
// precompute scheduling information
char * ${operation_name}_precompute(${operation_name}_base::Arguments const &args, int tile_count, size_t workspace_bytes) {
char* host_workspace = new char[workspace_bytes];
${operation_name}_base::ProblemVisitor::host_precompute(
args.host_problem_sizes,
args.problem_count,
args.threadblock_count,
(void*)host_workspace
);
return host_workspace;
}
// Get the size of params in bytes
int ${operation_name}_get_param_size(){
return sizeof(${operation_name}${operation_suffix}::Params);
}
// Get the size of dynamic shared memory in bytes
int ${operation_name}_shared_memory_size() {
return int(sizeof(${operation_name}${operation_suffix}::SharedStorage));
}
// Get the params as byte array
char* ${operation_name}_get_params(${operation_name}_base::Arguments* argument, int tile_count, void* workspace=nullptr){
${operation_name}_base::Params* params;
params = new ${operation_name}_base::Params(*argument, workspace, tile_count);
char *bytes = ((char*)(params));
char *output = new char[sizeof(${operation_name}_base::Params)];
for (unsigned int i = 0; i < sizeof(${operation_name}_base::Params); i ++)
output[i] = bytes[i];
return output;
}
cutlass::gemm::GemmCoord ${operation_name}_get_tiled_shape(
cutlass::gemm::GemmCoord problem_size, cutlass::gemm::GemmCoord tile_size, int split_k_slices) {
return ${operation_name}_base::ThreadblockSwizzle::get_tiled_shape(
problem_size, tile_size, split_k_slices);
}
dim3 ${operation_name}_get_grid_shape(cutlass::gemm::GemmCoord tiled_shape) {
return ${operation_name}_base::ThreadblockSwizzle::get_grid_shape(tiled_shape);
}
}
"""
def __init__(self, operation: "GemmOperation"):
super(GemmRTGrouped, self).__init__(operation)
self.extra_funcs = {
"precompute": None,
"get_tiled_shape": GemmCoord_,
"get_grid_shape": dim3_,
}
self.emitter = EmitGemmGroupedInstance("_type")
self.argument_type, self.epilogue_type = get_gemm_grouped_arguments(operation.epilogue_functor)
self.argtype = [ctypes.POINTER(self.argument_type), ctypes.c_int, ctypes.c_void_p]
def host_precompute(self, arguments, workspace_bytes):
self.precompute.argtype = [
self.argtype[0], ctypes.c_int, ctypes.c_longlong]
self.precompute.restype = ctypes.POINTER(ctypes.c_byte * workspace_bytes)
problem_info = self.precompute(
ctypes.byref(arguments.arguments),
arguments.total_tiles,
workspace_bytes)
problem_info_array = bytearray(problem_info.contents)
# copy to device memory
return todevice(problem_info_array).ptr
def plan(self, arguments):
return LaunchConfiguration(
[arguments.total_tiles, 1, 1],
[self.threads, 1, 1],
self.shared_memory_capacity,
)
def get_workspace_size(self, arguments):
if self.operation.precompute_mode == SchedulerMode.Device:
return 0
elif self.operation.precompute_mode == SchedulerMode.Host:
total_tiles = arguments.total_tiles
entries_per_block = 1
return 8 * entries_per_block * total_tiles # three int32_t
################################################################################
# Runtime module for GEMM and grouped GEMM
################################################################################
class GemmOperationBase:
"""
CUTLASS GEMM operation
"""
def __init__(
self, gemm_kind, arch, tile_description: TileDescription,
A: TensorDescription, B: TensorDescription, C: TensorDescription,
epilogue_functor, swizzling_functor=SwizzlingFunctor.Identity1,
api=ApiVersion.v2x, emission_type=EmissionType.Kernel, **kwargs):
self.operation_kind: OperationKind = OperationKind.Gemm
self.arch: int = arch
self.tile_description: TileDescription = tile_description
self.gemm_kind: GemmKind = gemm_kind
self.api = api
self.prefix = "3x" if self.api == ApiVersion.v3x else ""
self.emission_type = emission_type
# Optionally swap the TensorDescriptions for operands A and B and transpose their
# layouts. This is needed to mimic the transpose performed by device::GemmUniversal.
# The code below uses deep copy to avoid overwritting the original TensorDescription
self.switched = (self.api != ApiVersion.v3x and
self.emission_type == EmissionType.Kernel and
C.layout == LayoutType.ColumnMajor)
self.A, self.B, self.C = GemmOperationBase.get_operands(A, B, C, self.switched)
self.epilogue_functor = epilogue_functor
self.swizzling_functor = swizzling_functor
if "direct_store" in kwargs:
self.direct_store = kwargs["direct_store"]
else:
self.direct_store = False
@staticmethod
def get_operands(A: TensorDescription, B: TensorDescription, C: TensorDescription, swap: bool):
"""
Makes copies of A, B, and C, and possibly transposes their order. If ``swap`` is set,
A and B are swapped, and the layout of A, B, and C are transposed.
:param A: description of operand A
:type A: TensorDescription
:param B: description of operand B
:type B: TensorDescription
:param C: description of operand C
:type C: TensorDescription
:return: descriptions of operands A, B, and C
:rtype: tuple[TileDescription]
"""
if swap:
A_out = copy.deepcopy(B)
B_out = copy.deepcopy(A)
C_out = copy.deepcopy(C)
A_out.layout = transpose_layout(A_out.layout)
B_out.layout = transpose_layout(B_out.layout)
C_out.layout = transpose_layout(C_out.layout)
else:
A_out = copy.deepcopy(A)
B_out = copy.deepcopy(B)
C_out = copy.deepcopy(C)
return A_out, B_out, C_out
def run(self, arguments: GemmArguments) -> cuda.CUresult:
"""
Configure and launch the cuda kernel with input arguments
"""
if self.emission_type == EmissionType.Device:
raise Exception('Running a kernel via PyCUTLASS is only enabled with emission type "Kernel"')
err = self.rt_module.run(
arguments.host_workspace,
arguments.device_workspace,
arguments.launch_config,
arguments.stream
)
if err != cuda.CUresult.CUDA_SUCCESS:
raise RuntimeError("CUDA Error %s" % str(err))
return err
def is_complex(self):
complex_operators = [
MathOperation.multiply_add_complex,
MathOperation.multiply_add_complex_gaussian,
MathOperation.multiply_add_complex_fast_f32,
]
return self.tile_description.math_instruction.math_operation in complex_operators
def is_planar_complex(self):
return self.gemm_kind in (GemmKind.PlanarComplex, GemmKind.PlanarComplexArray)
def accumulator_type(self):
accum = self.tile_description.math_instruction.element_accumulator
if self.is_complex():
return get_complex_from_real(accum)
return accum
def short_math_name(self):
if self.tile_description.math_instruction.math_operation == MathOperation.multiply_add_complex_gaussian:
return "g%s" % ShortDataTypeNames[self.accumulator_type()]
return ShortDataTypeNames[self.accumulator_type()]
def core_name(self):
"""The basic operation kind is prefixed with a letter indicating the accumulation type."""
inst_shape = ""
inst_operation = ""
intermediate_type = ""
math_operations_map = {
MathOperation.xor_popc: "xor",
}
if (self.tile_description.math_instruction.opcode_class == OpcodeClass.TensorOp or
self.tile_description.math_instruction.opcode_class == OpcodeClass.WmmaTensorOp):
math_op = self.tile_description.math_instruction.math_operation
math_op_string = math_operations_map[math_op] if math_op in math_operations_map.keys() else ""
if self.tile_description.math_instruction.instruction_shape is not None:
if self.api == ApiVersion.v3x and self.arch >= 90:
inst_shape = "%dx%dx%d" % tuple(
self.tile_description.math_instruction.instruction_shape)
else:
inst_shape = "%d%d%d" % tuple(
self.tile_description.math_instruction.instruction_shape)
else:
inst_shape = "Default"
inst_shape += math_op_string
if (self.tile_description.math_instruction.element_a != self.A.element and
self.tile_description.math_instruction.element_a != self.tile_description.math_instruction.element_accumulator):
intermediate_type = DataTypeNames[self.tile_description.math_instruction.element_a]
return "%s%s%s%s" % (self.short_math_name(), inst_shape, intermediate_type, GemmKindNames[self.gemm_kind])
def extended_name(self):
"""Append data types if they differ from compute type."""
if self.is_complex():
extended_name = "${core_name}"
else:
if (self.C.element != self.tile_description.math_instruction.element_accumulator and
self.A.element != self.tile_description.math_instruction.element_accumulator):
extended_name = "${element_c}_${core_name}_${element_a}"
elif (self.C.element == self.tile_description.math_instruction.element_accumulator and
self.A.element != self.tile_description.math_instruction.element_accumulator):
extended_name = "${core_name}_${element_a}"
else:
extended_name = "${core_name}"
extended_name = SubstituteTemplate(extended_name, {
"element_a": DataTypeNames[self.A.element],
"element_c": DataTypeNames[self.C.element],
"core_name": self.core_name(),
})
return extended_name
def extended_name_3x(self):
"""Generates a string representing the MMA atom. Assumes accumulator type is C type."""
extended_name = "{core_name}_{element_a}_{element_b}_{element_acc}_{element_c}_{element_d}".format(
element_a=DataTypeNames[self.A.element],
element_b=DataTypeNames[self.B.element],
element_acc=DataTypeNames[self.accumulator_type()],
element_c=DataTypeNames[self.C.element],
element_d=DataTypeNames[self.epilogue_functor.element_output],
core_name=self.core_name())
return extended_name
def layout_name(self):
if self.is_complex() or self.is_planar_complex():
return "%s%s" % (
ShortComplexLayoutNames[(self.A.layout, self.A.complex_transform)],
ShortComplexLayoutNames[(self.B.layout, self.B.complex_transform)]
)
return "%s%s" % (ShortLayoutTypeNames[self.A.layout], ShortLayoutTypeNames[self.B.layout])
# Generates a short string representing the ABC layout tags (e.g. ntn or tnn)
def layout_name_3x(self):
if self.is_complex() or self.is_planar_complex():
return "{}{}{}".format(
ShortComplexLayoutNames[(self.A.layout, self.A.complex_transform)],
ShortComplexLayoutNames[(self.B.layout, self.B.complex_transform)],
ShortComplexLayoutNames[(self.C.layout, self.C.complex_transform)])
else:
return "{}{}{}".format(
ShortLayoutTypeNames[self.A.layout],
ShortLayoutTypeNames[self.B.layout],
ShortLayoutTypeNames[self.C.layout])
# Generates a short string representing underlying kernel schedule type
def kernel_schedule_name_3x(self):
if self.tile_description.kernel_schedule is None:
return KernelScheduleSuffixes[KernelScheduleType.ScheduleAuto]
else:
return KernelScheduleSuffixes[self.tile_description.kernel_schedule]
# Generates a short string representing underlying epilogue schedule type
def epilogue_schedule_name_3x(self):
if self.tile_description.epilogue_schedule is None:
return EpilogueScheduleSuffixes[EpilogueScheduleType.ScheduleAuto]
else:
return EpilogueScheduleSuffixes[self.tile_description.epilogue_schedule]
def procedural_name(self):
"""The full procedural name indicates architecture, extended name, tile size, and layout."""
opcode_class_name = OpcodeClassNames[self.tile_description.math_instruction.opcode_class]
if self.api == ApiVersion.v3x and self.arch >= 90:
kernel_name_template = "cutlass{p}_sm{ar}_{op}_{ex}_{tbm}x{tbn}x{tbk}_{cm}x{cn}x{ck}_{l}_{s}_align{al}{k}{e}"
return kernel_name_template.format(
p=self.prefix,
ar=self.arch,
op=opcode_class_name,
ex=self.extended_name_3x(),
tbm=self.tile_description.threadblock_shape[0],
tbn=self.tile_description.threadblock_shape[1],
tbk=self.tile_description.threadblock_shape[2],
cm=self.tile_description.cluster_shape[0],
cn=self.tile_description.cluster_shape[1],
ck=self.tile_description.cluster_shape[2],
l=self.tile_description.stages,
s=self.layout_name_3x(),
al=str(self.A.alignment),
k=self.kernel_schedule_name_3x(),
e=self.epilogue_schedule_name_3x()
)
else:
threadblock = self.tile_description.procedural_name_2x()
return "cutlass{p}_{op}_{ex}_{tb}_{l}_align{a}".format(
p=self.prefix,
op=opcode_class_name,
ex=self.extended_name(),
tb=threadblock,
l=self.layout_name(),
a=str(self.A.alignment)
)
def configuration_name(self):
"""The full procedural name indicates architecture, extended name, tile size, and layout."""
return self.procedural_name()
class GemmOperationUniversal(GemmOperationBase):
def __init__(self, arch, tile_description: TileDescription, A: TensorDescription, B, C,
epilogue_functor, swizzling_functor=SwizzlingFunctor.Identity1, **kwargs):
api = api_version(arch, tile_description.math_instruction.opcode_class, A.element)
super(GemmOperationUniversal, self).__init__(GemmKind.Universal, arch, tile_description,
A, B, C, epilogue_functor, swizzling_functor,
api=api, **kwargs, )
if api == ApiVersion.v3x:
if swizzling_functor == SwizzlingFunctor.StreamK:
raise Exception("Stream K swizzle functor is currently only supported for CUTLASS 2.x kernels")
self.rt_module = GemmRTUniversal3x(self)
else:
if swizzling_functor == SwizzlingFunctor.StreamK:
self.rt_module = GemmRTUniversalStreamK(self)
else:
self.rt_module = GemmRTUniversal(self)
self.argument_type = self.rt_module.argument_type
self.epilogue_type = self.rt_module.epilogue_type
def device_op(self):
"""
Returns a new GemmOperationUniversal object that is constructed with emission type
``EmissionType.Device``. Since the device-emitted kernel does not require swapping,
any swappng performed by the kernel-emitted operation is reversed.
:return: operation ready for device-level code emission
:rtype: GemmUniversalOperation
"""
A, B, C = GemmOperationBase.get_operands(self.A, self.B, self.C, self.switched)
return GemmOperationUniversal(self.arch, self.tile_description, A, B, C,
self.epilogue_functor, self.swizzling_functor,
emission_type=EmissionType.Device, direct_store=self.direct_store)
class GemmOperationGrouped(GemmOperationBase):
def __init__(self, arch, tile_description: TileDescription, A: TensorDescription, B, C,
epilogue_functor, swizzling_functor=SwizzlingFunctor.Identity1, **kwargs):
super(GemmOperationGrouped, self).__init__(GemmKind.Grouped, arch, tile_description,
A, B, C, epilogue_functor, swizzling_functor, **kwargs)
assert "precompute_mode" in kwargs.keys(), "missing keyword arguement 'precompute_mode'."
self.precompute_mode = kwargs["precompute_mode"]
self.rt_module = GemmRTGrouped(self)
self.argument_type = self.rt_module.argument_type
self.epilogue_type = self.rt_module.epilogue_type
def device_op(self):
"""
Returns a new GemmOperationGrouped object that is constructed with emission type
``EmissionType.Device``. Since the device-emitted kernel does not require swapping,
any swappng performed by the kernel-emitted operation is reversed.
:return: operation ready for device-level code emission
:rtype: GemmOperationGrouped
"""
A, B, C = GemmOperationBase.get_operands(self.A, self.B, self.C, self.switched)
return GemmOperationGrouped(
self.arch, self.tile_description, A, B, C, self.epilogue_functor,
self.swizzling_functor, emission_type=EmissionType.Device,
direct_store=self.direct_store, precompute_mode=self.precompute_mode, )
###################################################################################################
#
# Emits single instances of a CUTLASS device-wide operator
#
###################################################################################################
class EmitGemmUniversalInstance:
"""Responsible for emitting a CUTLASS template definition"""
def __init__(
self,
operation_suffix="",
direct_store=False
):
self.operation_suffix = operation_suffix
self.direct_store = direct_store
self.includes = [
"cutlass/cutlass.h",
"cutlass/gemm_coord.h",
"cutlass/numeric_types.h",
"cutlass/arch/arch.h",
"cutlass/arch/mma.h",
"cutlass/layout/matrix.h",
"cutlass/gemm/device/gemm.h",
"cutlass/gemm/device/gemm_universal_adapter.h",
"cutlass/gemm/kernel/default_gemm_universal.h",
]
if self.direct_store:
self.includes.append(
"cutlass/epilogue/threadblock/default_epilogue_direct_store.h"
)
self.gemm_template_kernel = """
// Gemm operator ${operation_name}
using ${operation_name}_base =
typename cutlass::gemm::kernel::DefaultGemmUniversal<
${element_a}, ${layout_a}, ${transform_a}, ${align_a},
${element_b}, ${layout_b}, ${transform_b}, ${align_b},
${element_c}, ${layout_c},
${element_accumulator},
${opcode_class},
${arch},
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
${epilogue_functor},
${swizzling_functor},
${stages},
${math_operation}
>::GemmKernel;
// Define named type
struct ${operation_name}${operation_suffix} :
public ${operation_name}_base { };
"""
self.gemm_template_device = """
// Gemm operator ${operation_name}
using DeviceKernel =
typename cutlass::gemm::device::GemmUniversal<
// Data type and layout of operand A
${element_a}, ${layout_a},
// Data type and layout of operand B
${element_b}, ${layout_b},
// Data type and layout of operand C
${element_c}, ${layout_c},
// Data type of accumulator
${element_accumulator},
// Class of operation
${opcode_class},
// Compute capability of the target kernel
${arch},
// Threadblock tile shape
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
// Warp tile shape
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
// Instruction shape
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
// Epilogue functor
${epilogue_functor},
// Swizzling function
${swizzling_functor},
// Number of pipeline stages
${stages},
// Alignment of operands A and B
${align_a}, ${align_b},
// Type of math operation
${math_operation},
// Complex transform types of operands A and B
${transform_a}, ${transform_b}
>;
"""
self.gemm_template_direct_store = """
// Gemm operator ${operation_name}
using ${operation_name}_default =
typename cutlass::gemm::kernel::DefaultGemmUniversal<
${element_a}, ${layout_a}, ${transform_a}, ${align_a},
${element_b}, ${layout_b}, ${transform_b}, ${align_b},
${element_c}, ${layout_c},
${element_accumulator},
${opcode_class},
${arch},
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
${epilogue_functor},
${swizzling_functor},
${stages},
${math_operation}
>::GemmKernel;
using ${operation_name}_base =
cutlass::gemm::kernel::GemmUniversal<
${operation_name}_default::Mma,
cutlass::epilogue::threadblock::DefaultEpilogueDirectStore<
${operation_name}_default::Epilogue
>::Epilogue,
${operation_name}_default::ThreadblockSwizzle
>;
// Define named type
struct ${operation_name}${operation_suffix} :
public ${operation_name}_base { };
"""
self.gemm_template_kernel_visitor = """
using OutputTileThreadMap = cutlass::epilogue::threadblock::OutputTileThreadLayout<
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
${element_c},
${align_c},
${epilogue_stages} /* epilogue stages */
>;
${callback_decl}
// Gemm operator ${operation_name}
using ${operation_name}_base =
typename cutlass::gemm::kernel::DefaultGemmWithVisitor<
${element_a}, ${layout_a}, ${transform_a}, ${align_a},
${element_b}, ${layout_b}, ${transform_b}, ${align_b},
${element_c}, ${layout_c}, ${align_c},
${element_accumulator},
${element_epilogue},
${opcode_class},
${arch},
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
${callback_name},
${swizzling_functor},
${stages},
${math_operation},
${epilogue_stages} /* epilogue stages */
>::GemmKernel;
// Define named type
struct ${operation_name}${operation_suffix} :
public ${operation_name}_base { };
"""
def instance_template(self):
return """
${compile_guard_start}
manifest.append(new ${gemm_kind}<
cutlass::gemm::device::GemmUniversalAdapter<${operation_name}>
>("${operation_name}"));
${compile_guard_end}
"""
def emit(self, operation):
threadblock_shape = operation.tile_description.threadblock_shape
warp_count = operation.tile_description.warp_count
warp_shape = [threadblock_shape[idx] // warp_count[idx] for idx in range(3)]
instance_layout_A, instance_layout_B, instance_layout_C = \
(operation.A.layout, operation.B.layout, operation.C.layout)
if operation.emission_type == EmissionType.Kernel:
if self.direct_store:
gemm_template = self.gemm_template_direct_store
else:
gemm_template = self.gemm_template_kernel
else:
gemm_template = self.gemm_template_device
values = {
"operation_name": operation.procedural_name(),
"operation_suffix": self.operation_suffix,
"element_a": DataTypeTag[operation.A.element],
"layout_a": LayoutTag[instance_layout_A],
"element_b": DataTypeTag[operation.B.element],
"layout_b": LayoutTag[instance_layout_B],
"element_c": DataTypeTag[operation.C.element],
"layout_c": LayoutTag[instance_layout_C],
"element_accumulator": DataTypeTag[operation.accumulator_type()],
"opcode_class": OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
"arch": "cutlass::arch::Sm%d" % operation.arch,
"threadblock_shape_m": str(operation.tile_description.threadblock_shape[0]),
"threadblock_shape_n": str(operation.tile_description.threadblock_shape[1]),
"threadblock_shape_k": str(operation.tile_description.threadblock_shape[2]),
"warp_shape_m": str(warp_shape[0]),
"warp_shape_n": str(warp_shape[1]),
"warp_shape_k": str(warp_shape[2]),
"instruction_shape_m": str(operation.tile_description.math_instruction.instruction_shape[0]),
"instruction_shape_n": str(operation.tile_description.math_instruction.instruction_shape[1]),
"instruction_shape_k": str(operation.tile_description.math_instruction.instruction_shape[2]),
"swizzling_functor": SwizzlingFunctorTag[operation.swizzling_functor],
"stages": str(operation.tile_description.stages),
"align_a": str(operation.A.alignment),
"align_b": str(operation.B.alignment),
"transform_a": ComplexTransformTag[operation.A.complex_transform],
"transform_b": ComplexTransformTag[operation.B.complex_transform],
"math_operation": MathOperationTag[operation.tile_description.math_instruction.math_operation],
}
if hasattr(operation.epilogue_functor, "visitor"):
self.includes += [
"cutlass/epilogue/threadblock/fusion/visitors.hpp",
"cutlass/gemm/kernel/default_gemm_universal_with_visitor.h"
]
callback_name, callback_decl = operation.epilogue_functor.emit(operation)
values["callback_name"] = callback_name
values["callback_decl"] = callback_decl
values["align_c"] = str(operation.C.alignment)
values["element_epilogue"] = DataTypeTag[operation.epilogue_functor.element_epilogue]
if hasattr(operation.epilogue_functor, "epilogue_stages"):
epilogue_stages = operation.epilogue_functor.epilogue_stages
else:
epilogue_stages = 1
values["epilogue_stages"] = str(epilogue_stages)
return SubstituteTemplate(self.gemm_template_kernel_visitor, values)
else:
values["epilogue_functor"] = operation.epilogue_functor.emit()
return SubstituteTemplate(gemm_template, values)
class EmitGemmGroupedInstance:
"""Responsible for emitting a CUTLASS template definition"""
def __init__(self, operation_suffix=""):
self.operation_suffix = operation_suffix
self.includes = [
"cutlass/cutlass.h",
"cutlass/numeric_types.h",
"cutlass/arch/arch.h",
"cutlass/arch/mma.h",
"cutlass/layout/matrix.h",
"cutlass/gemm/kernel/gemm_grouped.h",
"cutlass/gemm/kernel/default_gemm_grouped.h",
]
self.gemm_template_kernel = """
// Gemm operator ${operation_name}
using ${operation_name}_base =
typename cutlass::gemm::kernel::DefaultGemmGrouped<
${element_a}, ${layout_a}, ${transform_a}, ${align_a},
${element_b}, ${layout_b}, ${transform_b}, ${align_b},
${element_c}, ${layout_c},
${element_accumulator},
${opcode_class},
${arch},
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
${epilogue_functor},
${swizzling_functor},
${stages},
${precompute_mode},
${math_operation}
>::GemmKernel;
// Define named type
struct ${operation_name}${operation_suffix} :
public ${operation_name}_base { };
"""
self.gemm_template_device = (
self.gemm_template_kernel
+ """
using DeviceKernel = cutlass::gemm::device::GemmGrouped<${operation_name}_base>;
"""
)
def instance_template(self):
return """
${compile_guard_start}
manifest.append(new ${gemm_kind}<
cutlass::gemm::device::GemmGrouped<${operation_name}>
>("${operation_name}"));
${compile_guard_end}
"""
def emit(self, operation):
threadblock_shape = operation.tile_description.threadblock_shape
warp_count = operation.tile_description.warp_count
warp_shape = [threadblock_shape[idx] // warp_count[idx] for idx in range(3)]
instance_layout_A, instance_layout_B, instance_layout_C = \
(operation.A.layout, operation.B.layout, operation.C.layout)
# Support built-in epilogue functors or user-defined functions
epilogue_functor = operation.epilogue_functor.emit()
values = {
"operation_name": operation.procedural_name(),
"operation_suffix": self.operation_suffix,
"element_a": DataTypeTag[operation.A.element],
"layout_a": LayoutTag[instance_layout_A],
"element_b": DataTypeTag[operation.B.element],
"layout_b": LayoutTag[instance_layout_B],
"element_c": DataTypeTag[operation.C.element],
"layout_c": LayoutTag[instance_layout_C],
"element_accumulator": DataTypeTag[operation.accumulator_type()],
"opcode_class": OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
"arch": "cutlass::arch::Sm%d" % operation.arch,
"threadblock_shape_m": str(operation.tile_description.threadblock_shape[0]),
"threadblock_shape_n": str(operation.tile_description.threadblock_shape[1]),
"threadblock_shape_k": str(operation.tile_description.threadblock_shape[2]),
"warp_shape_m": str(warp_shape[0]),
"warp_shape_n": str(warp_shape[1]),
"warp_shape_k": str(warp_shape[2]),
"instruction_shape_m": str(operation.tile_description.math_instruction.instruction_shape[0]),
"instruction_shape_n": str(operation.tile_description.math_instruction.instruction_shape[1]),
"instruction_shape_k": str(operation.tile_description.math_instruction.instruction_shape[2]),
"epilogue_functor": epilogue_functor,
"swizzling_functor": SwizzlingFunctorTag[operation.swizzling_functor],
"stages": str(operation.tile_description.stages),
"align_a": str(operation.A.alignment),
"align_b": str(operation.B.alignment),
"transform_a": ComplexTransformTag[operation.A.complex_transform],
"transform_b": ComplexTransformTag[operation.B.complex_transform],
"precompute_mode": SchedulerModeTag[operation.precompute_mode],
"math_operation": MathOperationTag[operation.tile_description.math_instruction.math_operation],
}
if operation.emission_type == EmissionType.Kernel:
gemm_template = self.gemm_template_kernel
else:
gemm_template = self.gemm_template_device
return SubstituteTemplate(gemm_template, values)