CUTLASS 2.9 (#468)
This commit is contained in:
Andrew Kerr
GitHub
@ -1,24 +1,30 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
|
||||
* Copyright (c) 2017 - 2022 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:
|
||||
* * 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:
|
||||
*
|
||||
* 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
|
||||
* 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
@ -202,6 +208,19 @@ struct B2bGemm {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
// Determine if fusion sizes are valid
|
||||
if(problem_size_0.m() != problem_size_1.m())
|
||||
return Status::kErrorInvalidProblem;
|
||||
|
||||
if(problem_size_0.n() != problem_size_1.k())
|
||||
return Status::kErrorInvalidProblem;
|
||||
|
||||
if(problem_size_0.n() > B2bMma::Shape0::kN)
|
||||
return Status::kErrorInvalidProblem;
|
||||
|
||||
if(problem_size_1.n() > B2bMma::Shape1::kN)
|
||||
return Status::kErrorInvalidProblem;
|
||||
|
||||
return Status::kSuccess;
|
||||
}
|
||||
|
||||
|
||||
@ -1,24 +1,30 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
|
||||
* Copyright (c) 2017 - 2022 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:
|
||||
* * 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:
|
||||
*
|
||||
* 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
|
||||
* 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,749 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief
|
||||
Default kernel-level implicit GEMM convolution definitions combine threadblock-scoped
|
||||
matrix multiply-add with the appropriate threadblock-scoped epilogue.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/conv/kernel/default_conv2d.h"
|
||||
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_optimized.h"
|
||||
|
||||
#include "cutlass/transform/threadblock/predicated_vector_access_iterator.h"
|
||||
#include "cutlass/transform/threadblock/vector_iterator.h"
|
||||
#include "cutlass/transform/warp/vector_fragment_iterator.h"
|
||||
|
||||
#include "cutlass/gemm/warp/mma_tensor_op_fragment_iterator.h"
|
||||
|
||||
#include "kernel/default_b2b_conv2d_fprop.h"
|
||||
#include "kernel/b2b_implicit_gemm_convolution.h"
|
||||
#include "threadblock/b2b_implicit_gemm_pipelined.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace conv {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// OpClassTensorOp convolutions
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm
|
||||
/// and 2 stage pipeline.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::ColumnMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelined<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename detail::DefaultConvEpilogue<
|
||||
ArchTag,
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm and 2 stage
|
||||
/// pipeline with interleaved layout.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic,
|
||||
false
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelined<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm
|
||||
/// and 2 stage pipeline.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::ColumnMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelined<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename detail::DefaultConvEpilogue<
|
||||
ArchTag,
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm and 2 stage
|
||||
/// pipeline with interleaved layout.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelined<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace conv
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -0,0 +1,740 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief
|
||||
Default kernel-level implicit GEMM convolution definitions combine threadblock-scoped
|
||||
matrix multiply-add with the appropriate threadblock-scoped epilogue.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/conv/kernel/default_conv2d.h"
|
||||
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_optimized.h"
|
||||
|
||||
#include "cutlass/transform/threadblock/predicated_vector_access_iterator.h"
|
||||
#include "cutlass/transform/threadblock/vector_iterator.h"
|
||||
#include "cutlass/transform/warp/vector_fragment_iterator.h"
|
||||
|
||||
#include "cutlass/gemm/warp/mma_tensor_op_fragment_iterator.h"
|
||||
|
||||
#include "kernel/default_b2b_conv2d_fprop.h"
|
||||
#include "kernel/b2b_implicit_gemm_convolution.h"
|
||||
#include "threadblock/b2b_implicit_gemm_multistage.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace conv {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// OpClassTensorOp convolutions
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm and multistage
|
||||
/// pipeline.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::ColumnMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistage<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm and multistage
|
||||
/// pipeline with interleaved layout.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistage<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm and
|
||||
/// multistage pipeline.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::ColumnMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistage<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimzed IteratorAlgorithm and
|
||||
// multistage pipeline with interleaved layout.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
// Use fragment iterator for A operand
|
||||
using AccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorA1 =
|
||||
cutlass::gemm::warp::MmaTensorOpFragmentIterator<
|
||||
cutlass::MatrixShape<MmaCore1::WarpShape::kM, MmaCore1::InstructionShape::kK>, //warp shape
|
||||
cutlass::MatrixShape<MmaCore0::WarpShape::kM, MmaCore0::WarpShape::kN>, //accumulator shape
|
||||
MmaCore1::Shape::kK, //kBlocksColumn
|
||||
ElementAccumulator, ElementA, AccumulatorLayout, InstructionShape, EpilogueOutputOp0>;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape1::kM, WarpShape1::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Warp-level iterators to load scale and bias vectors
|
||||
using FragmentIteratorA1ScaleBias = cutlass::transform::warp::VectorFragmentIterator<
|
||||
MatrixShape<1, IteratorAccumulatorScaleBias::Fragment::kElements>, ElementScaleBias,
|
||||
LayoutScaleBias, InstructionShape, kElementsPerAccess>;
|
||||
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistage<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
ThreadblockShape1,
|
||||
FragmentIteratorA1,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorA1ScaleBias,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace conv
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -0,0 +1,817 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief
|
||||
Default kernel-level implicit GEMM convolution definitions combine threadblock-scoped
|
||||
matrix multiply-add with the appropriate threadblock-scoped epilogue.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/conv/kernel/default_conv2d.h"
|
||||
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_optimized.h"
|
||||
|
||||
#include "cutlass/transform/threadblock/predicated_vector_access_iterator.h"
|
||||
#include "cutlass/transform/threadblock/vector_iterator.h"
|
||||
#include "cutlass/transform/warp/vector_fragment_iterator.h"
|
||||
|
||||
#include "cutlass/gemm/warp/mma_tensor_op_fragment_iterator.h"
|
||||
|
||||
#include "kernel/default_b2b_conv2d_fprop.h"
|
||||
#include "kernel/b2b_implicit_gemm_convolution.h"
|
||||
#include "threadblock/b2b_implicit_gemm_pipelined_smem_accumulator.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace conv {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm
|
||||
/// and 2 stage pipeline.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIterator<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelinedSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename detail::DefaultConvEpilogue<
|
||||
ArchTag,
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm and 2 stage
|
||||
/// pipeline with interleaved layout.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4; //For interleaved layout
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::ColumnMajorInterleaved<16>;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIteratorCanonical<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelinedSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm
|
||||
/// and 2 stage pipeline.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIterator<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelinedSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename detail::DefaultConvEpilogue<
|
||||
ArchTag,
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm and 2 stage
|
||||
/// pipeline with interleaved layout.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
2, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4; //For interleaved layout
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::TileIterator<
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::ColumnMajorInterleaved<16>;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIteratorCanonical<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmPipelinedSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace conv
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -0,0 +1,804 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief
|
||||
Default kernel-level implicit GEMM convolution definitions combine threadblock-scoped
|
||||
matrix multiply-add with the appropriate threadblock-scoped epilogue.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/conv/kernel/default_conv2d.h"
|
||||
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_analytic.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h"
|
||||
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_optimized.h"
|
||||
|
||||
#include "cutlass/transform/threadblock/predicated_vector_access_iterator.h"
|
||||
#include "cutlass/transform/threadblock/vector_iterator.h"
|
||||
#include "cutlass/transform/warp/vector_fragment_iterator.h"
|
||||
|
||||
#include "cutlass/gemm/warp/mma_tensor_op_fragment_iterator.h"
|
||||
|
||||
#include "kernel/default_b2b_conv2d_fprop.h"
|
||||
#include "kernel/b2b_implicit_gemm_convolution.h"
|
||||
#include "threadblock/b2b_implicit_gemm_multistage_smem_accumulator.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace conv {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm and multistage
|
||||
/// pipeline.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIterator<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistageSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Analytic IteratorAlgorithm and multistage
|
||||
/// pipeline with interleaved layout.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kAnalytic,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kK>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::ColumnMajorInterleaved<16>;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIteratorCanonical<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistageSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm and
|
||||
/// multistage pipeline.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename LayoutA,
|
||||
typename ElementB,
|
||||
typename LayoutB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::RowMajor,
|
||||
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA0 = typename MmaCore0::IteratorThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, LayoutA,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB0 = typename MmaCore0::IteratorThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 2;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB1 = typename MmaCore1::IteratorThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, LayoutB,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::RowMajor;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIterator<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistageSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines a kernel for Conv2dFprop specialization for Optimzed IteratorAlgorithm and
|
||||
// multistage pipeline with interleaved layout.
|
||||
/// Accumulator will be staged in shared memory.
|
||||
template <
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename LayoutC,
|
||||
typename ElementAccumulator,
|
||||
typename ArchTag,
|
||||
typename ThreadblockShape0,
|
||||
typename ThreadblockShape1,
|
||||
typename WarpShape0,
|
||||
typename WarpShape1,
|
||||
typename InstructionShape,
|
||||
typename EpilogueOutputOp0,
|
||||
typename EpilogueOutputOp1,
|
||||
typename ThreadblockSwizzle,
|
||||
int Stages,
|
||||
typename MathOperatorTag,
|
||||
int InterleavedK
|
||||
>
|
||||
struct DefaultB2bConv2dFprop <
|
||||
ElementA,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
ElementB,
|
||||
layout::TensorCxRSKx<InterleavedK>,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
ArchTag,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
Stages,
|
||||
MathOperatorTag,
|
||||
IteratorAlgorithm::kOptimized,
|
||||
true
|
||||
> {
|
||||
|
||||
// Define the core components from GEMM
|
||||
using MmaCore0 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape0, WarpShape0, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
using MmaCore1 = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape1, WarpShape1, InstructionShape, ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp,
|
||||
Stages, MathOperatorTag, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapA0 = typename MmaCore0::SmemThreadMapA;
|
||||
using IteratorA0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kK>,
|
||||
ElementA, layout::TensorNCxHWx<InterleavedK>,
|
||||
ThreadMapA0
|
||||
>;
|
||||
|
||||
using SmemIteratorA0 = typename MmaCore0::SmemIteratorA;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
// Note GEMM shared memory threadmap is used here because conv global memory
|
||||
// layout needs to be mapped to fprop which is similar to the crosswise
|
||||
// layout which is used by the interleaved GEMM shared memory threadmap.
|
||||
// The Interleaved GEMM global memory layout is similar to the congruous
|
||||
// layout.
|
||||
using ThreadMapB0 = typename MmaCore0::SmemThreadMapB;
|
||||
using IteratorB0 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kK, ThreadblockShape0::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB0
|
||||
>;
|
||||
|
||||
using SmemIteratorB0 = typename MmaCore0::SmemIteratorB;
|
||||
|
||||
/// Define iterators over tiles from scale/bias vectors
|
||||
using ElementScaleBias = typename EpilogueOutputOp0::ElementCompute;
|
||||
using LayoutScaleBias = layout::RowMajor; //vector layout doesn't really matter
|
||||
static int const kElementsPerAccess = 4;
|
||||
using IteratorAccumulatorScaleBias =
|
||||
cutlass::transform::threadblock::VectorIterator<
|
||||
cutlass::transform::threadblock::PredicatedVectorAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape0::kM, ThreadblockShape0::kN>,
|
||||
cutlass::MatrixShape<WarpShape0::kM, WarpShape0::kN>,
|
||||
ElementScaleBias, LayoutScaleBias, kElementsPerAccess>
|
||||
>;
|
||||
|
||||
using ThreadMapB1 = typename MmaCore1::SmemThreadMapB;
|
||||
using IteratorB1 =
|
||||
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorOptimized<
|
||||
cutlass::MatrixShape<ThreadblockShape1::kK, ThreadblockShape1::kN>,
|
||||
ElementB, layout::TensorCxRSKx<InterleavedK>,
|
||||
ThreadMapB1
|
||||
>;
|
||||
|
||||
using SmemIteratorB1 = typename MmaCore1::SmemIteratorB;
|
||||
|
||||
|
||||
// Warp-level GEMM components
|
||||
using WarpMmaTensorOp0 = typename MmaCore0::MmaTensorOp;
|
||||
using WarpMmaTensorOp1 = typename MmaCore1::MmaTensorOp;
|
||||
using MmaPolicy0 = typename MmaCore0::MmaPolicy;
|
||||
using MmaPolicy1 = typename MmaCore1::MmaPolicy;
|
||||
|
||||
// Use fragment iterator for the accumulator
|
||||
using SmemAccumulatorLayout = cutlass::layout::ColumnMajorInterleaved<16>;
|
||||
using FragmentIteratorAccumulator = cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
WarpShape0, InstructionShape,
|
||||
ElementAccumulator,
|
||||
typename WarpMmaTensorOp0::Policy::Operator::FragmentC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
|
||||
// Store Accumulator tiles to Shared Memory
|
||||
using SmemIteratorD0 =
|
||||
cutlass::epilogue::warp::TileIteratorTensorOp<
|
||||
WarpShape0,
|
||||
InstructionShape,
|
||||
ElementC,
|
||||
SmemAccumulatorLayout
|
||||
>;
|
||||
|
||||
static int const kThreadCount = 32;
|
||||
// load warp tile from Shared Memory accumulator
|
||||
using WarpIteratorA1 = cutlass::gemm::warp::MmaTensorOpMultiplicandTileIteratorCanonical<
|
||||
MatrixShape<WarpShape1::kM, InstructionShape::kK>, cutlass::gemm::Operand::kA,
|
||||
ElementA, SmemAccumulatorLayout,
|
||||
MatrixShape<InstructionShape::kM, InstructionShape::kK>,
|
||||
WarpMmaTensorOp1::Policy::OpDelta::kRow, kThreadCount>;
|
||||
|
||||
// Define the Mma
|
||||
using B2bMma = threadblock::B2bImplicitGemmMultistageSmemAccumulator<
|
||||
ThreadblockShape0,
|
||||
IteratorA0,
|
||||
SmemIteratorA0,
|
||||
arch::CacheOperation::Always,
|
||||
IteratorB0,
|
||||
SmemIteratorB0,
|
||||
arch::CacheOperation::Global,
|
||||
IteratorAccumulatorScaleBias,
|
||||
FragmentIteratorAccumulator,
|
||||
SmemIteratorD0,
|
||||
ThreadblockShape1,
|
||||
WarpIteratorA1,
|
||||
IteratorB1,
|
||||
SmemIteratorB1,
|
||||
arch::CacheOperation::Global,
|
||||
EpilogueOutputOp0,
|
||||
MmaPolicy0,
|
||||
MmaPolicy1,
|
||||
Stages
|
||||
>;
|
||||
|
||||
// Define the epilogue
|
||||
using Epilogue = typename epilogue::threadblock::DefaultInterleavedConvEpilogue<
|
||||
ThreadblockShape1,
|
||||
WarpMmaTensorOp1,
|
||||
1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount,
|
||||
InterleavedK
|
||||
>::Epilogue;
|
||||
|
||||
// Define the kernel
|
||||
using Kernel = cutlass::conv::kernel::B2bImplicitGemmConvolution<
|
||||
B2bMma,
|
||||
Epilogue,
|
||||
ThreadblockSwizzle,
|
||||
conv::Operator::kFprop
|
||||
>;
|
||||
};
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace conv
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -1,24 +1,30 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
|
||||
* Copyright (c) 2017 - 2022 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:
|
||||
* * 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:
|
||||
*
|
||||
* 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
|
||||
* 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
@ -111,7 +117,9 @@ template <
|
||||
/// If true, kernel is configured to support serial reduction in the epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator
|
||||
typename Operator,
|
||||
/// Stage accumulator in shared memory
|
||||
bool SmemAccumulator = false
|
||||
>
|
||||
struct DefaultB2bGemm;
|
||||
|
||||
|
||||
@ -0,0 +1,397 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief
|
||||
Default kernel-level GEMM definitions combine threadblock-scoped matrix multiply-add with
|
||||
the appropriate threadblock-scoped epilogue.
|
||||
|
||||
Note, CUTLASS epilogues universally target row-major outputs. Column-major outputs are
|
||||
accommodated by exchanging A and B operands and assuming transposed layouts. Partial
|
||||
specializations here choose 'device::GemmTransposed' to implement this functionality.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/layout/matrix.h"
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/epilogue/threadblock/epilogue.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination.h"
|
||||
|
||||
#include "cutlass/gemm/gemm.h"
|
||||
#include "cutlass/gemm/kernel/gemm_pipelined.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm75.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm70.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm80.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_simt.h"
|
||||
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"
|
||||
#include "cutlass/epilogue/threadblock/default_epilogue_tensor_op.h"
|
||||
#include "cutlass/epilogue/threadblock/default_epilogue_volta_tensor_op.h"
|
||||
#include "cutlass/epilogue/threadblock/default_epilogue_simt.h"
|
||||
|
||||
#include "cutlass/transform/threadblock/predicated_tile_iterator.h"
|
||||
#include "cutlass/transform/threadblock/vector_iterator.h"
|
||||
#include "cutlass/transform/threadblock/predicated_vector_access_iterator.h"
|
||||
|
||||
#include "kernel/b2b_gemm.h"
|
||||
#include "threadblock/default_b2b_mma.h"
|
||||
#include "threadblock/default_b2b_mma_smem_accumulator.h"
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace gemm {
|
||||
namespace kernel {
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Partial specialization for Ampere Architecture
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Layout type for A matrix operand
|
||||
typename LayoutA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Layout type for B matrix operand
|
||||
typename LayoutB,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Element type for internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of stages used in the pipelined mainloop
|
||||
int Stages,
|
||||
/// If true, kernel is configured to support serial reduction in the
|
||||
/// epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator>
|
||||
struct DefaultB2bGemm<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB, ElementC,
|
||||
layout::RowMajor, ElementAccumulator, arch::OpClassTensorOp,
|
||||
arch::Sm80, ThreadblockShape0, ThreadblockShape1,
|
||||
WarpShape0, WarpShape1, InstructionShape,
|
||||
EpilogueOutputOp0, EpilogueOutputOp1, ThreadblockSwizzle, Stages, SplitKSerial,
|
||||
Operator, true> {
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
|
||||
ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, arch::Sm80,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, Stages, Operator, EpilogueOutputOp0, false, true>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue
|
||||
using Epilogue =
|
||||
typename cutlass::epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1, typename B2bMma::Operator1, kPartitionsK1, EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Partial specialization for Turing Architecture
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Layout type for A matrix operand
|
||||
typename LayoutA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Layout type for B matrix operand
|
||||
typename LayoutB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Element type for internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// If true, kernel is configured to support serial reduction in the epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator
|
||||
>
|
||||
struct DefaultB2bGemm<
|
||||
ElementA, LayoutA, kAlignmentA,
|
||||
ElementB, LayoutB, kAlignmentB,
|
||||
ElementC, layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
arch::Sm75,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
SplitKSerial,
|
||||
Operator,
|
||||
true
|
||||
> {
|
||||
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
kAlignmentA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
kAlignmentB,
|
||||
ElementAccumulator,
|
||||
layout::RowMajor,
|
||||
arch::OpClassTensorOp,
|
||||
arch::Sm75,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
2,
|
||||
Operator,
|
||||
EpilogueOutputOp0,
|
||||
false,
|
||||
true
|
||||
>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue
|
||||
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1,
|
||||
typename B2bMma::Operator1,
|
||||
kPartitionsK1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount
|
||||
>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
|
||||
/// Partial specialization for Ampere Integer Matrix Multiply Interleaved layout
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of stages used in the pipelined mainloop
|
||||
int Stages,
|
||||
/// Number of Interleaved k
|
||||
int InterleavedK,
|
||||
/// If true, kernel is configured to support serial reduction in the
|
||||
/// epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator>
|
||||
struct DefaultB2bGemm<
|
||||
ElementA, layout::ColumnMajorInterleaved<InterleavedK>, kAlignmentA,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>, kAlignmentB,
|
||||
ElementC, layout::ColumnMajorInterleaved<InterleavedK>, int32_t,
|
||||
arch::OpClassTensorOp, arch::Sm80,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, EpilogueOutputOp0, EpilogueOutputOp1,
|
||||
ThreadblockSwizzle, Stages,
|
||||
SplitKSerial, Operator, true> {
|
||||
using LayoutA = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
using LayoutB = layout::RowMajorInterleaved<InterleavedK>;
|
||||
using LayoutC = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
|
||||
using ElementAccumulator = int32_t;
|
||||
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp, arch::Sm80,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, Stages, Operator, EpilogueOutputOp0,
|
||||
true, true>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue
|
||||
using Epilogue = typename cutlass::epilogue::threadblock::
|
||||
DefaultInterleavedEpilogueTensorOp<
|
||||
ThreadblockShape1, typename B2bMma::Operator1, kPartitionsK1, EpilogueOutputOp1,
|
||||
64 / sizeof_bits<ElementC>::value, InterleavedK>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
/// Partial specialization for Turing Integer Tensor Core Interleaved layout
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of Interleaved k
|
||||
int InterleavedK,
|
||||
/// If true, kernel is configured to support serial reduction in the
|
||||
/// epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator>
|
||||
struct DefaultB2bGemm<ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
kAlignmentA, ElementB,
|
||||
layout::RowMajorInterleaved<InterleavedK>, kAlignmentB,
|
||||
ElementC, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
int32_t, arch::OpClassTensorOp, arch::Sm75,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, EpilogueOutputOp0, EpilogueOutputOp1,
|
||||
ThreadblockSwizzle, 2, SplitKSerial, Operator, true> {
|
||||
using LayoutA = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
using LayoutB = layout::RowMajorInterleaved<InterleavedK>;
|
||||
using LayoutC = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
|
||||
using ElementAccumulator = int32_t;
|
||||
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp, arch::Sm75,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, 2, Operator, EpilogueOutputOp0, true, true>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue for the 2nd Gemm
|
||||
using Epilogue = typename cutlass::epilogue::threadblock::
|
||||
DefaultInterleavedEpilogueTensorOp<
|
||||
ThreadblockShape1, typename B2bMma::Operator1, kPartitionsK1, EpilogueOutputOp1,
|
||||
64 / sizeof_bits<ElementC>::value, InterleavedK>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace gemm
|
||||
} // namespace cutlass
|
||||
Reference in New Issue
Block a user