CUTLASS 2.5
This commit is contained in:
Andrew Kerr
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -34,6 +34,10 @@
|
||||
#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"
|
||||
|
||||
@@ -68,6 +72,113 @@ template <
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// 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 <
|
||||
@@ -173,6 +284,114 @@ struct DefaultConv3dFprop <
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// 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
|
||||
|
||||
Reference in New Issue
Block a user