CUTLASS 2.2 (#96)
Adds support for NVIDIA Ampere Architecture features. CUDA 11 Toolkit recommended.
This commit is contained in:
Andrew Kerr
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* Copyright (c) 2017-2020, 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:
|
||||
@@ -38,6 +38,8 @@
|
||||
#include "cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm70.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm75.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm80.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_WMMA_ENABLED)
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_wmma.h"
|
||||
#endif //CUTLASS_ARCH_WMMA_ENABLED
|
||||
@@ -203,6 +205,58 @@ struct DefaultMma<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB,
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
/// Specialization for row-major output (OperatorClass TensorOp)
|
||||
template <
|
||||
/// Layout type for A matrix operand
|
||||
typename LayoutA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Layout type for B matrix operand
|
||||
typename LayoutB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Tag indicating architecture to tune for
|
||||
typename ArchTag,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape,
|
||||
/// Instruction-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator
|
||||
>
|
||||
struct DefaultMma<float, LayoutA, kAlignmentA, float, LayoutB,
|
||||
kAlignmentB, float, layout::RowMajor,
|
||||
arch::OpClassTensorOp, ArchTag, ThreadblockShape, WarpShape,
|
||||
InstructionShape, 2, Operator, false> {
|
||||
// Define the MmaCore components
|
||||
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape, WarpShape, InstructionShape, float, LayoutA, float,
|
||||
LayoutB, float, layout::RowMajor, arch::OpClassTensorOp, 2,
|
||||
arch::OpMultiplyAddFastF16>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using IteratorA =
|
||||
cutlass::transform::threadblock::PredicatedTileIterator<
|
||||
cutlass::MatrixShape<MmaCore::Shape::kM, MmaCore::Shape::kK>,
|
||||
float, LayoutA, 1, typename MmaCore::IteratorThreadMapA, kAlignmentA>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using IteratorB =
|
||||
cutlass::transform::threadblock::PredicatedTileIterator<
|
||||
cutlass::MatrixShape<MmaCore::Shape::kK, MmaCore::Shape::kN>,
|
||||
float, LayoutB, 0, typename MmaCore::IteratorThreadMapB, kAlignmentB>;
|
||||
|
||||
// Define the threadblock-scoped pipelined matrix multiply
|
||||
using ThreadblockMma = cutlass::gemm::threadblock::MmaPipelined<
|
||||
typename MmaCore::Shape, IteratorA, typename MmaCore::SmemIteratorA,
|
||||
IteratorB, typename MmaCore::SmemIteratorB, float,
|
||||
layout::RowMajor, typename MmaCore::MmaPolicy>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Specialization for column-major-interleaved output
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
@@ -271,6 +325,214 @@ struct DefaultMma<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB,
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Specialization for row-major output
|
||||
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 internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Tag indicating architecture to tune for
|
||||
typename ArchTag,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape,
|
||||
/// Instruction-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Number of stages used in the multistage mainloop
|
||||
int Stages,
|
||||
/// Operation perfomed by GEMM
|
||||
typename Operator
|
||||
>
|
||||
struct DefaultMma<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB,
|
||||
kAlignmentB, ElementAccumulator, layout::RowMajor,
|
||||
arch::OpClassSimt, ArchTag, ThreadblockShape, WarpShape,
|
||||
InstructionShape, Stages, Operator, false> {
|
||||
// Define the MmaCore components
|
||||
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape, WarpShape, InstructionShape, ElementA, LayoutA,
|
||||
ElementB, LayoutB, ElementAccumulator, layout::RowMajor, arch::OpClassSimt,
|
||||
Stages, Operator>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
||||
using AccessTypeA = cutlass::Array<ElementA, kAlignmentA>;
|
||||
using IteratorA =
|
||||
cutlass::transform::threadblock::PredicatedTileAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
||||
ElementA, LayoutA, 1, ThreadMapA, AccessTypeA>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
||||
using AccessTypeB = cutlass::Array<ElementB, kAlignmentB>;
|
||||
using IteratorB =
|
||||
cutlass::transform::threadblock::PredicatedTileAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
||||
ElementB, LayoutB, 0, ThreadMapB, AccessTypeB>;
|
||||
|
||||
// Define the threadblock-scoped multistage matrix multiply
|
||||
using ThreadblockMma = cutlass::gemm::threadblock::MmaMultistage<
|
||||
typename MmaCore::Shape, IteratorA, typename MmaCore::SmemIteratorA,
|
||||
MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
|
||||
MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor,
|
||||
typename MmaCore::MmaPolicy, Stages>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Specialization for row-major output (OperatorClass TensorOp)
|
||||
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 internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Tag indicating architecture to tune for
|
||||
typename ArchTag,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape,
|
||||
/// Instruction-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Number of stages used in the multistage mainloop
|
||||
int Stages,
|
||||
/// Operation perfomed by GEMM
|
||||
typename Operator
|
||||
>
|
||||
struct DefaultMma<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB,
|
||||
kAlignmentB, ElementAccumulator, layout::RowMajor,
|
||||
arch::OpClassTensorOp, ArchTag, ThreadblockShape, WarpShape,
|
||||
InstructionShape, Stages, Operator, false> {
|
||||
static cutlass::arch::CacheOperation::Kind const CacheOpA =
|
||||
((sizeof_bits<ElementA>::value * kAlignmentA) == 128)
|
||||
? cutlass::arch::CacheOperation::Global
|
||||
: cutlass::arch::CacheOperation::Always;
|
||||
|
||||
static cutlass::arch::CacheOperation::Kind const CacheOpB =
|
||||
((sizeof_bits<ElementB>::value * kAlignmentB) == 128)
|
||||
? cutlass::arch::CacheOperation::Global
|
||||
: cutlass::arch::CacheOperation::Always;
|
||||
|
||||
// Define the MmaCore components
|
||||
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape, WarpShape, InstructionShape, ElementA, LayoutA,
|
||||
ElementB, LayoutB, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp,
|
||||
Stages, Operator, false, CacheOpA, CacheOpB>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
||||
using AccessTypeA = cutlass::Array<ElementA, kAlignmentA>;
|
||||
using IteratorA =
|
||||
cutlass::transform::threadblock::PredicatedTileAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
||||
ElementA, LayoutA, 1, ThreadMapA, AccessTypeA>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
||||
using AccessTypeB = cutlass::Array<ElementB, kAlignmentB>;
|
||||
using IteratorB =
|
||||
cutlass::transform::threadblock::PredicatedTileAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
||||
ElementB, LayoutB, 0, ThreadMapB, AccessTypeB>;
|
||||
|
||||
// Define the threadblock-scoped multistage matrix multiply
|
||||
using ThreadblockMma = cutlass::gemm::threadblock::MmaMultistage<
|
||||
typename MmaCore::Shape, IteratorA, typename MmaCore::SmemIteratorA,
|
||||
MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
|
||||
MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor,
|
||||
typename MmaCore::MmaPolicy, Stages>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Specialization for column-major-interleaved output
|
||||
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 internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Tag indicating architecture to tune for
|
||||
typename OperatorClass,
|
||||
/// Tag indicating architecture to tune for
|
||||
typename ArchTag,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape,
|
||||
/// Instruction-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Number of stages used in the multistage mainloop
|
||||
int Stages,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator,
|
||||
/// Number of Interleaved K
|
||||
int InterleavedK>
|
||||
struct DefaultMma<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB,
|
||||
kAlignmentB, ElementAccumulator,
|
||||
layout::ColumnMajorInterleaved<InterleavedK>, OperatorClass,
|
||||
ArchTag, ThreadblockShape, WarpShape, InstructionShape,
|
||||
Stages, Operator, true> {
|
||||
// Define the MmaCore components
|
||||
using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<
|
||||
ThreadblockShape, WarpShape, InstructionShape, ElementA, LayoutA,
|
||||
ElementB, LayoutB, ElementAccumulator,
|
||||
layout::ColumnMajorInterleaved<InterleavedK>, OperatorClass, Stages,
|
||||
Operator, true>;
|
||||
|
||||
// Define iterators over tiles from the A operand
|
||||
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
|
||||
using AccessTypeA = cutlass::Array<ElementA, kAlignmentA>;
|
||||
using IteratorA =
|
||||
cutlass::transform::threadblock::PredicatedTileAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
||||
ElementA, LayoutA, 1, ThreadMapA, AccessTypeA>;
|
||||
|
||||
// Define iterators over tiles from the B operand
|
||||
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
|
||||
using AccessTypeB = cutlass::Array<ElementB, kAlignmentB>;
|
||||
using IteratorB =
|
||||
cutlass::transform::threadblock::PredicatedTileAccessIterator<
|
||||
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
||||
ElementB, LayoutB, 0, ThreadMapB, AccessTypeB>;
|
||||
|
||||
// Define the threadblock-scoped multistage matrix multiply
|
||||
using ThreadblockMma = cutlass::gemm::threadblock::MmaMultistage<
|
||||
typename MmaCore::Shape, IteratorA, typename MmaCore::SmemIteratorA,
|
||||
MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
|
||||
MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor,
|
||||
typename MmaCore::MmaPolicy, Stages>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Specialization for SIMT IDP4A Kernels
|
||||
|
||||
Reference in New Issue
Block a user