1928 lines
52 KiB
C++
1928 lines
52 KiB
C++
/***************************************************************************************************
|
|
* Copyright (c) 2017 - 2024 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_dgrad_output_gradient_tile_access_iterator_analytic.h"
|
|
#include "cutlass/conv/threadblock/conv2d_dgrad_output_gradient_tile_access_iterator_optimized.h"
|
|
#include "cutlass/conv/threadblock/conv2d_dgrad_filter_tile_access_iterator_analytic.h"
|
|
#include "cutlass/conv/threadblock/conv2d_dgrad_filter_tile_access_iterator_optimized.h"
|
|
#include "cutlass/conv/threadblock/conv2d_tile_iterator.h"
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
namespace conv {
|
|
namespace kernel {
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
/// Defines a kernel for Conv2dDgrad
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename OperatorClass,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
conv::IteratorAlgorithm IteratorAlgorithm = IteratorAlgorithm::kOptimized,
|
|
conv::StrideSupport StrideSupport = StrideSupport::kStrided,
|
|
/// Access granularity of A matrix in units of elements
|
|
int AlignmentA = 128 / cutlass::sizeof_bits<ElementA>::value,
|
|
/// Access granularity of B matrix in units of elements
|
|
int AlignmentB = 128 / cutlass::sizeof_bits<ElementB>::value
|
|
> struct DefaultConv2dDgrad;
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
// OpClassTensorOp convolutions
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Analytic IteratorAlgorithm Dgrad Strided and
|
|
// multistage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kStrided,
|
|
AccessTypeA
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kStrided,
|
|
AccessTypeB
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
static cutlass::arch::CacheOperation::Kind const CacheOpB =
|
|
((sizeof_bits<ElementB>::value * AlignmentB) == 128)
|
|
? cutlass::arch::CacheOperation::Global
|
|
: cutlass::arch::CacheOperation::Always;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
CacheOpB,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOpStridedDgrad<
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Analytic IteratorAlgorithm Dgrad Strided
|
|
// and 2 stage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kStrided,
|
|
AccessTypeA
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kStrided,
|
|
AccessTypeB
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename detail::DefaultConvEpilogueStridedDgrad<
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Analytic IteratorAlgorithm Dgrad Unity Strided
|
|
// and multistage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kUnity,
|
|
AccessTypeA
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kUnity,
|
|
AccessTypeB
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
static cutlass::arch::CacheOperation::Kind const CacheOpB =
|
|
((sizeof_bits<ElementB>::value * AlignmentB) == 128)
|
|
? cutlass::arch::CacheOperation::Global
|
|
: cutlass::arch::CacheOperation::Always;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
CacheOpB,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Analytic IteratorAlgorithm Dgrad Unity
|
|
// 2 stage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kUnity,
|
|
AccessTypeA
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kUnity,
|
|
AccessTypeB
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename detail::DefaultConvEpilogue<
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for optimized IteratorAlgorithm Dgrad Unity Strided
|
|
// and multistage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kUnity,
|
|
AccessTypeA
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kUnity,
|
|
AccessTypeB
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
static cutlass::arch::CacheOperation::Kind const CacheOpB =
|
|
((sizeof_bits<ElementB>::value * AlignmentB) == 128)
|
|
? cutlass::arch::CacheOperation::Global
|
|
: cutlass::arch::CacheOperation::Always;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
CacheOpB,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Optimized IteratorAlgorithm Dgrad Strided and
|
|
// multistage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kStrided,
|
|
AccessTypeA
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kStrided,
|
|
AccessTypeB
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
static cutlass::arch::CacheOperation::Kind const CacheOpB =
|
|
((sizeof_bits<ElementB>::value * AlignmentB) == 128)
|
|
? cutlass::arch::CacheOperation::Global
|
|
: cutlass::arch::CacheOperation::Always;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
CacheOpB,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOpStridedDgrad<
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Optimized IteratorAlgorithm Dgrad Strided
|
|
// and 2 stage pipeline.
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kStrided,
|
|
AccessTypeA
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kStrided,
|
|
AccessTypeB
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename detail::DefaultConvEpilogueStridedDgrad<
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Optimized IteratorAlgorithm Dgrad Unity
|
|
// 2 stage pipeline
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using AccessTypeA = cutlass::AlignedArray<ElementA, AlignmentA>;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kUnity,
|
|
AccessTypeA
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kUnity,
|
|
AccessTypeB
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaTensorOp = typename MmaCore::MmaTensorOp;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename detail::DefaultConvEpilogue<
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
kPartitionsK,
|
|
EpilogueOutputOp
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
// OpClassSimt convolutions
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
/// Defines a kernel for Conv2dDgrad specialization for Analytic IteratorAlgorithm,
|
|
/// multi-stage pipeline, and FFMA-based mainloop for SM80
|
|
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
conv::StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
conv::StrideSupport::kUnity
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
conv::StrideSupport::kUnity
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
arch::CacheOperation::Always,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimt<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
conv::StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
conv::StrideSupport::kStrided
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
conv::StrideSupport::kStrided
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
arch::CacheOperation::Always,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimtStridedDgrad<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Optimized IteratorAlgorithm,
|
|
/// multi-stage pipeline, and FFMA-based mainloop for SM80
|
|
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kUnity
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kUnity
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
arch::CacheOperation::Always,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimt<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
int Stages,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
conv::StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
conv::StrideSupport::kStrided
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
conv::StrideSupport::kStrided
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
arch::CacheOperation::Always,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimtStridedDgrad<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
|
|
};
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Analytic IteratorAlgorithm,
|
|
/// 2 stage pipeline, and FFMA-based mainloop for SM50
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
conv::StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
conv::StrideSupport::kUnity
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
conv::StrideSupport::kUnity
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimt<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
|
|
};
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic,
|
|
conv::StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
conv::StrideSupport::kStrided
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
conv::StrideSupport::kStrided
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimtStridedDgrad<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dDgrad specialization for Optimized IteratorAlgorithm,
|
|
/// 2 stage pipeline, and FFMA-based mainloop for SM50
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
StrideSupport::kUnity,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
StrideSupport::kUnity
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
StrideSupport::kUnity
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimt<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
template <
|
|
typename ElementA,
|
|
typename LayoutA,
|
|
typename ElementB,
|
|
typename LayoutB,
|
|
typename ElementC,
|
|
typename LayoutC,
|
|
typename ElementAccumulator,
|
|
typename ArchTag,
|
|
typename ThreadblockShape,
|
|
typename WarpShape,
|
|
typename InstructionShape,
|
|
typename EpilogueOutputOp,
|
|
typename ThreadblockSwizzle,
|
|
typename MathOperatorTag,
|
|
int AlignmentA,
|
|
int AlignmentB
|
|
>
|
|
struct DefaultConv2dDgrad <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassSimt,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized,
|
|
conv::StrideSupport::kStrided,
|
|
AlignmentA,
|
|
AlignmentB
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::RowMajor, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradOutputGradientTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA,
|
|
conv::StrideSupport::kStrided
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::TileIteratorStridedDgrad<
|
|
cutlass::conv::threadblock::Conv2dDgradFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB,
|
|
conv::StrideSupport::kStrided
|
|
>
|
|
>;
|
|
|
|
using SmemIteratorB = typename MmaCore::SmemIteratorB;
|
|
|
|
// Warp-level GEMM components
|
|
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
|
|
using MmaPolicy = typename MmaCore::MmaPolicy;
|
|
|
|
// Define the Mma
|
|
using Mma = threadblock::ImplicitGemmPipelined<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
ElementC,
|
|
LayoutC,
|
|
MmaPolicy
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimtStridedDgrad<
|
|
ThreadblockShape,
|
|
WarpMmaSimtOp,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolutionStridedDgrad<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kDgrad
|
|
>;
|
|
|
|
};
|
|
|
|
} // namespace kernel
|
|
} // namespace conv
|
|
} // namespace cutlass
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|