2e07c4cc2f
CUTLASS 2.7 Mainloop fusion for GEMM: summation over A or B Strided DGRAD (optimized iterators) Half-precision GELU_taylor activation functions Use these when accumulation and epilogue compute types are all cutlass::half_t Tuning and bug fixes to fused GEMM + GEMM example Support for smaller than 128b aligned Convolutions: see examples Caching of results to accelerate Convolution unit tests Can be enabled or disabled by running cmake .. -DCUTLASS_TEST_ENABLE_CACHED_RESULTS=OFF Corrections and bug fixes reported by the CUTLASS community Thank you for filing these issues! authored-by: Haicheng Wu haichengw@nvidia.com, Manish Gupta manigupta@nvidia.com, Dustyn Blasig dblasig@nvidia.com, Andrew Kerr akerr@nvidia.com
510 lines
14 KiB
C++
510 lines
14 KiB
C++
/***************************************************************************************************
|
|
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
|
|
*
|
|
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
|
* provided that the following conditions are met:
|
|
* * Redistributions of source code must retain the above copyright notice, this list of
|
|
* conditions and the following disclaimer.
|
|
* * Redistributions in binary form must reproduce the above copyright notice, this list of
|
|
* conditions and the following disclaimer in the documentation and/or other materials
|
|
* provided with the distribution.
|
|
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
|
|
* to endorse or promote products derived from this software without specific prior written
|
|
* permission.
|
|
*
|
|
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
|
|
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
|
|
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
|
|
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
|
|
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
|
|
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
|
|
* STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
*
|
|
**************************************************************************************************/
|
|
|
|
/*! \file
|
|
\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/conv3d_fprop_activation_tile_access_iterator_optimized.h"
|
|
#include "cutlass/conv/threadblock/conv3d_fprop_filter_tile_access_iterator_optimized.h"
|
|
|
|
|
|
#include "cutlass/conv/threadblock/conv3d_fprop_activation_tile_access_iterator_analytic.h"
|
|
#include "cutlass/conv/threadblock/conv3d_fprop_filter_tile_access_iterator_analytic.h"
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
namespace conv {
|
|
namespace kernel {
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
/// Defines a kernel for Conv2dFprop
|
|
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
|
|
> struct DefaultConv3dFprop;
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv3dFprop specialization for Analytic Iterator Algorithm
|
|
/// 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
|
|
>
|
|
struct DefaultConv3dFprop <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv3dFpropActivationTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA
|
|
>
|
|
>;
|
|
|
|
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::Conv3dFpropFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB
|
|
>
|
|
>;
|
|
|
|
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
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename detail::DefaultConvEpilogue<
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
1,
|
|
EpilogueOutputOp
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kFprop,
|
|
Conv3dProblemSize
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv2dFprop specialzation for Analytic IteratorAlgorithm 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
|
|
>
|
|
struct DefaultConv3dFprop <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kAnalytic
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv3dFpropActivationTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
ThreadMapA
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv3dFpropFilterTileAccessIteratorAnalytic<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
ThreadMapB
|
|
>;
|
|
|
|
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::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
arch::CacheOperation::Global,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
1,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kFprop,
|
|
Conv3dProblemSize
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv3dFprop specialization for Optimized Iterator Algorithm
|
|
/// 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
|
|
>
|
|
struct DefaultConv3dFprop <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
2,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
2, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::TileIterator<
|
|
cutlass::conv::threadblock::Conv3dFpropActivationTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
LayoutA,
|
|
ThreadMapA
|
|
>
|
|
>;
|
|
|
|
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::Conv3dFpropFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
LayoutB,
|
|
ThreadMapB
|
|
>
|
|
>;
|
|
|
|
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
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename detail::DefaultConvEpilogue<
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
1,
|
|
EpilogueOutputOp
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kFprop,
|
|
Conv3dProblemSize
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines a kernel for Conv3dFprop specialzation for Optimized IteratorAlgorithm 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
|
|
>
|
|
struct DefaultConv3dFprop <
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
arch::OpClassTensorOp,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
MathOperatorTag,
|
|
IteratorAlgorithm::kOptimized
|
|
> {
|
|
|
|
// Define the core components from GEMM
|
|
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
|
ThreadblockShape, WarpShape, InstructionShape, ElementA, layout::RowMajor,
|
|
ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
|
Stages, MathOperatorTag>;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
|
using IteratorA =
|
|
cutlass::conv::threadblock::Conv3dFpropActivationTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA,
|
|
LayoutA,
|
|
ThreadMapA
|
|
>;
|
|
|
|
using SmemIteratorA = typename MmaCore::SmemIteratorA;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
|
|
|
using IteratorB =
|
|
cutlass::conv::threadblock::Conv3dFpropFilterTileAccessIteratorOptimized<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB,
|
|
LayoutB,
|
|
ThreadMapB
|
|
>;
|
|
|
|
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::ImplicitGemmMultistage<
|
|
ThreadblockShape,
|
|
IteratorA,
|
|
SmemIteratorA,
|
|
arch::CacheOperation::Always,
|
|
IteratorB,
|
|
SmemIteratorB,
|
|
arch::CacheOperation::Global,
|
|
MmaPolicy,
|
|
Stages
|
|
>;
|
|
|
|
// Define the epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpilogueTensorOp<
|
|
ThreadblockShape,
|
|
WarpMmaTensorOp,
|
|
1,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
// Define the kernel
|
|
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle,
|
|
conv::Operator::kFprop,
|
|
Conv3dProblemSize
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace kernel
|
|
} // namespace conv
|
|
} // namespace cutlass
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|