66d9cddc83
* New updates. * Minor profiler updates Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
347 lines
10 KiB
C++
347 lines
10 KiB
C++
/***************************************************************************************************
|
|
* Copyright (c) 2017 - 2023 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 Rank 2k definitions combine threadblock-scoped matrix multiply-add with
|
|
the appropriate threadblock-scoped epilogue.
|
|
|
|
Note, CUTLASS epilogues universally target row-major outputs. Column-major outputs are
|
|
accommodated by exchanging A and B operands and assuming transposed layouts.
|
|
|
|
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "cutlass/blas3.h"
|
|
|
|
#include "cutlass/complex.h"
|
|
#include "cutlass/layout/matrix.h"
|
|
|
|
#include "cutlass/gemm/kernel/rank_2k_universal.h"
|
|
#include "cutlass/gemm/kernel/default_rank_2k.h"
|
|
#include "cutlass/gemm/kernel/default_rank_2k_complex.h"
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
namespace gemm {
|
|
namespace kernel {
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <
|
|
/// Element type for A matrix operand
|
|
typename ElementA_,
|
|
/// Layout type for A matrix operand
|
|
typename LayoutA_,
|
|
/// Complex elementwise transformation on A operand
|
|
ComplexTransform TransformA,
|
|
/// 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_,
|
|
/// Complex elementwise transformation on B operand
|
|
ComplexTransform TransformB,
|
|
/// Access granularity of B matrix in units of elements
|
|
int kAlignmentB,
|
|
/// Element type for C and D matrix operands
|
|
typename ElementC_,
|
|
/// Layout type for C and D matrix operands
|
|
typename LayoutC_,
|
|
/// Fill Mode for C (kLower or kUpper)
|
|
FillMode FillModeC_,
|
|
/// Element type for internal accumulation
|
|
typename ElementAccumulator,
|
|
/// Operator class tag
|
|
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,
|
|
/// Warp-level tile size (concept: GemmShape)
|
|
typename InstructionShape,
|
|
/// Epilogue output operator
|
|
typename EpilogueOutputOp,
|
|
/// Threadblock-level swizzling operator
|
|
typename ThreadblockSwizzle,
|
|
/// Number of stages used in the pipelined mainloop
|
|
int Stages,
|
|
/// If true, kernel is configured to support serial reduction in the
|
|
/// epilogue
|
|
bool SplitKSerial,
|
|
/// Operation performed by SYRK
|
|
typename Operator,
|
|
/// Blas3 computation mode (symmetric/hermitian)
|
|
BlasMode BlasMode_ = BlasMode::kSymmetric,
|
|
///
|
|
typename Enable = void
|
|
>
|
|
struct DefaultRank2KUniversal;
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
//
|
|
// Real-valued Rank 2k update kernels
|
|
//
|
|
|
|
template <
|
|
/// Element type for A matrix operand
|
|
typename ElementA,
|
|
/// Layout type for A matrix operand
|
|
typename LayoutA,
|
|
/// Access granularity of A matrix in units of elements
|
|
int kAlignmentA,
|
|
/// Element type for B matrix operand
|
|
typename ElementB,
|
|
/// Layout type for B matrix operand
|
|
typename LayoutB,
|
|
/// Access granularity of B matrix in units of elements
|
|
int kAlignmentB,
|
|
/// Element type for C and D matrix operands
|
|
typename ElementC,
|
|
/// Layout type for C and D matrix operands
|
|
typename LayoutC,
|
|
/// Fill Mode for C (kLower or kUpper)
|
|
FillMode FillModeC,
|
|
/// Element type for internal accumulation
|
|
typename ElementAccumulator,
|
|
/// Operator class tag
|
|
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,
|
|
/// Warp-level tile size (concept: GemmShape)
|
|
typename InstructionShape,
|
|
/// Epilogue output operator
|
|
typename EpilogueOutputOp,
|
|
/// Threadblock-level swizzling operator
|
|
typename ThreadblockSwizzle,
|
|
/// Number of stages used in the pipelined mainloop
|
|
int Stages,
|
|
/// If true, kernel is configured to support serial reduction in the
|
|
/// epilogue
|
|
bool SplitKSerial,
|
|
/// Operation performed by Rank2k
|
|
typename Operator>
|
|
struct DefaultRank2KUniversal<
|
|
ElementA,
|
|
LayoutA,
|
|
ComplexTransform::kNone, // transform A
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
ComplexTransform::kNone, // transform B
|
|
kAlignmentB,
|
|
ElementC,
|
|
LayoutC,
|
|
FillModeC,
|
|
ElementAccumulator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
SplitKSerial,
|
|
Operator,
|
|
BlasMode::kSymmetric,
|
|
typename std::enable_if< ! cutlass::is_complex<ElementAccumulator>::value>::type
|
|
> {
|
|
|
|
using DefaultRank2Kkernel = typename kernel::DefaultRank2K<
|
|
ElementA,
|
|
LayoutA,
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
kAlignmentB,
|
|
ElementC,
|
|
LayoutC,
|
|
FillModeC,
|
|
ElementAccumulator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
SplitKSerial,
|
|
Operator,
|
|
BlasMode::kSymmetric
|
|
>::Rank2Kkernel;
|
|
|
|
/// Define the kernel in terms of the default kernel
|
|
using Rank2Kkernel = kernel::Rank2KUniversal<
|
|
typename DefaultRank2Kkernel::Mma1,
|
|
typename DefaultRank2Kkernel::Mma2,
|
|
typename DefaultRank2Kkernel::Epilogue,
|
|
ThreadblockSwizzle,
|
|
FillModeC,
|
|
BlasMode::kSymmetric
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
//
|
|
// Complex-valued Rank 2K update kernels
|
|
//
|
|
|
|
template <
|
|
/// Element type for A matrix operand
|
|
typename ElementA,
|
|
/// Layout type for A matrix operand
|
|
typename LayoutA,
|
|
/// Complex elementwise transformation on A operand
|
|
ComplexTransform TransformA,
|
|
/// 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,
|
|
/// Complex elementwise transformation on B operand
|
|
ComplexTransform TransformB,
|
|
/// Access granularity of B matrix in units of elements
|
|
int kAlignmentB,
|
|
/// Element type for C and D matrix operands
|
|
typename ElementC,
|
|
/// Layout type for C and D matrix operands
|
|
typename LayoutC,
|
|
/// Fill Mode for C (kLower or kUpper)
|
|
FillMode FillModeC,
|
|
/// Element type for internal accumulation
|
|
typename ElementAccumulator,
|
|
/// Operator class tag
|
|
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,
|
|
/// Warp-level tile size (concept: GemmShape)
|
|
typename InstructionShape,
|
|
/// Epilogue output operator
|
|
typename EpilogueOutputOp,
|
|
/// Threadblock-level swizzling operator
|
|
typename ThreadblockSwizzle,
|
|
/// Number of stages used in the pipelined mainloop
|
|
int Stages,
|
|
/// If true, kernel is configured to support serial reduction in the
|
|
/// epilogue
|
|
bool SplitKSerial,
|
|
/// Operation performed by SYRK
|
|
typename Operator,
|
|
// BlasMode
|
|
BlasMode kBlasMode
|
|
>
|
|
|
|
struct DefaultRank2KUniversal<
|
|
ElementA,
|
|
LayoutA,
|
|
TransformA,
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
TransformB,
|
|
kAlignmentB,
|
|
ElementC,
|
|
LayoutC,
|
|
FillModeC,
|
|
ElementAccumulator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
SplitKSerial,
|
|
Operator,
|
|
kBlasMode,
|
|
typename std::enable_if<cutlass::is_complex<ElementAccumulator>::value>::type
|
|
> {
|
|
|
|
using DefaultRank2Kkernel = typename kernel::DefaultRank2KComplex<
|
|
ElementA,
|
|
LayoutA,
|
|
ElementB,
|
|
LayoutB,
|
|
ElementC,
|
|
LayoutC,
|
|
FillModeC,
|
|
ElementAccumulator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
TransformA,
|
|
TransformB,
|
|
Operator,
|
|
SplitKSerial,
|
|
kBlasMode
|
|
>::Rank2Kkernel;
|
|
|
|
/// Define the kernel in terms of the default kernel
|
|
using Rank2Kkernel = kernel::Rank2KUniversal<
|
|
typename DefaultRank2Kkernel::Mma1,
|
|
typename DefaultRank2Kkernel::Mma2,
|
|
typename DefaultRank2Kkernel::Epilogue,
|
|
ThreadblockSwizzle,
|
|
FillModeC,
|
|
kBlasMode
|
|
>;
|
|
};
|
|
|
|
} // namespace kernel
|
|
} // namespace gemm
|
|
} // namespace cutlass
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|