66d9cddc83
* New updates. * Minor profiler updates Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
266 lines
10 KiB
C++
266 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 TRMM 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/layout/matrix.h"
|
|
|
|
#include "cutlass/epilogue/threadblock/epilogue.h"
|
|
#include "cutlass/epilogue/thread/linear_combination.h"
|
|
|
|
#include "cutlass/gemm/gemm.h"
|
|
#include "cutlass/gemm/kernel/trmm_universal.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_core_sm75.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_core_sm70.h"
|
|
#include "cutlass/gemm/threadblock/default_multistage_mma_complex_core_sm80.h"
|
|
#include "cutlass/gemm/threadblock/default_mma.h"
|
|
#include "cutlass/gemm/threadblock/default_multistage_trmm_complex.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_core_simt.h"
|
|
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"
|
|
#include "cutlass/epilogue/threadblock/default_epilogue_complex_tensor_op.h"
|
|
#include "cutlass/epilogue/threadblock/default_epilogue_simt.h"
|
|
|
|
#include "cutlass/transform/threadblock/predicated_tile_iterator.h"
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
namespace gemm {
|
|
namespace kernel {
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <
|
|
/// Element type for A matrix operand
|
|
typename ElementA_,
|
|
/// Layout type for A matrix operand
|
|
typename LayoutA_,
|
|
/// Element type for B matrix operand
|
|
typename ElementB_,
|
|
/// Layout type for B matrix operand
|
|
typename LayoutB_,
|
|
/// Side Mode for the kernel
|
|
SideMode SideMode_,
|
|
/// Fill Mode for the triangular matrix
|
|
FillMode FillMode_,
|
|
/// Diag Type for the triangular matrix
|
|
DiagType DiagType_,
|
|
/// Element type for C and D matrix operands
|
|
typename ElementC_,
|
|
/// Layout type for C and D matrix operands
|
|
typename LayoutC_,
|
|
/// 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,
|
|
/// Complex elementwise transformation on A operand
|
|
ComplexTransform TransformA,
|
|
/// Complex elementwise transformation on B operand
|
|
ComplexTransform TransformB,
|
|
/// Multiply-add operator
|
|
// (arch::OpMultiplyAddComplex, arch::OpMultiplyGaussianComplex)
|
|
typename Operator,
|
|
/// If true, kernel is configured to support serial reduction in the epilogue
|
|
bool SplitKSerial
|
|
>
|
|
struct DefaultTrmmComplex;
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Partial specialization for Hopper Architecture
|
|
template <
|
|
/// Element type for A matrix operand
|
|
typename ElementA,
|
|
/// Layout type for A matrix operand
|
|
typename LayoutA,
|
|
/// Element type for B matrix operand
|
|
typename ElementB,
|
|
/// Layout type for B matrix operand
|
|
typename LayoutB,
|
|
/// Side Mode for the kernel
|
|
SideMode kSideMode,
|
|
/// Fill Mode for the triangular matrix
|
|
FillMode kFillMode,
|
|
/// Diag Type for the triangular matrix
|
|
DiagType kDiagType,
|
|
/// Element type for C and D matrix operands
|
|
typename ElementC,
|
|
/// Element type for internal accumulation
|
|
typename ElementAccumulator,
|
|
/// 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,
|
|
/// Complex elementwise transformation on A operand
|
|
ComplexTransform TransformA,
|
|
/// Complex elementwise transformation on B operand
|
|
ComplexTransform TransformB,
|
|
/// Multiply-add operator
|
|
// (arch::OpMultiplyAddComplex, arch::OpMultiplyGaussianComplex)
|
|
typename Operator,
|
|
/// If true, kernel is configured to support serial reduction in the epilogue
|
|
bool SplitKSerial
|
|
>
|
|
struct DefaultTrmmComplex<
|
|
ElementA, LayoutA, ElementB, LayoutB,
|
|
kSideMode, kFillMode, kDiagType,
|
|
ElementC, layout::RowMajor, ElementAccumulator, arch::OpClassTensorOp,
|
|
arch::Sm90, ThreadblockShape, WarpShape, InstructionShape,
|
|
EpilogueOutputOp, ThreadblockSwizzle, Stages, TransformA, TransformB, Operator, SplitKSerial> {
|
|
|
|
/// Define the threadblock-scoped matrix multiply-accumulate
|
|
using Mma = typename cutlass::gemm::threadblock::DefaultMultistageTrmmComplex<
|
|
ElementA, LayoutA, ElementB, LayoutB,
|
|
kSideMode, kFillMode, kDiagType,
|
|
ElementAccumulator,layout::RowMajor, arch::OpClassTensorOp, arch::Sm90, ThreadblockShape,
|
|
WarpShape, InstructionShape, Stages, TransformA, TransformB, Operator>::ThreadblockMma;
|
|
|
|
/// Define the epilogue
|
|
using Epilogue =
|
|
typename cutlass::epilogue::threadblock::DefaultEpilogueComplexTensorOp<
|
|
ThreadblockShape, typename Mma::Operator, 1, EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount, Operator>::Epilogue;
|
|
|
|
/// Define the kernel-level TRMM operator.
|
|
using TrmmKernel = kernel::TrmmUniversal<Mma, Epilogue, ThreadblockSwizzle, kSideMode, kFillMode, kDiagType>;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Partial specialization for Ampere Architecture
|
|
template <
|
|
/// Element type for A matrix operand
|
|
typename ElementA,
|
|
/// Layout type for A matrix operand
|
|
typename LayoutA,
|
|
/// Element type for B matrix operand
|
|
typename ElementB,
|
|
/// Layout type for B matrix operand
|
|
typename LayoutB,
|
|
/// Side Mode for the kernel
|
|
SideMode kSideMode,
|
|
/// Fill Mode for the triangular matrix
|
|
FillMode kFillMode,
|
|
/// Diag Type for the triangular matrix
|
|
DiagType kDiagType,
|
|
/// Element type for C and D matrix operands
|
|
typename ElementC,
|
|
/// Element type for internal accumulation
|
|
typename ElementAccumulator,
|
|
/// 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,
|
|
/// Complex elementwise transformation on A operand
|
|
ComplexTransform TransformA,
|
|
/// Complex elementwise transformation on B operand
|
|
ComplexTransform TransformB,
|
|
/// Multiply-add operator
|
|
// (arch::OpMultiplyAddComplex, arch::OpMultiplyGaussianComplex)
|
|
typename Operator,
|
|
/// If true, kernel is configured to support serial reduction in the epilogue
|
|
bool SplitKSerial
|
|
>
|
|
struct DefaultTrmmComplex<
|
|
ElementA, LayoutA, ElementB, LayoutB,
|
|
kSideMode, kFillMode, kDiagType,
|
|
ElementC, layout::RowMajor, ElementAccumulator, arch::OpClassTensorOp,
|
|
arch::Sm80, ThreadblockShape, WarpShape, InstructionShape,
|
|
EpilogueOutputOp, ThreadblockSwizzle, Stages, TransformA, TransformB, Operator, SplitKSerial> {
|
|
|
|
/// Define the threadblock-scoped matrix multiply-accumulate
|
|
using Mma = typename cutlass::gemm::threadblock::DefaultMultistageTrmmComplex<
|
|
ElementA, LayoutA, ElementB, LayoutB,
|
|
kSideMode, kFillMode, kDiagType,
|
|
ElementAccumulator,layout::RowMajor, arch::OpClassTensorOp, arch::Sm80, ThreadblockShape,
|
|
WarpShape, InstructionShape, Stages, TransformA, TransformB, Operator>::ThreadblockMma;
|
|
|
|
/// Define the epilogue
|
|
using Epilogue =
|
|
typename cutlass::epilogue::threadblock::DefaultEpilogueComplexTensorOp<
|
|
ThreadblockShape, typename Mma::Operator, 1, EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount, Operator>::Epilogue;
|
|
|
|
/// Define the kernel-level TRMM operator.
|
|
using TrmmKernel = kernel::TrmmUniversal<Mma, Epilogue, ThreadblockSwizzle, kSideMode, kFillMode, kDiagType>;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace kernel
|
|
} // namespace gemm
|
|
} // namespace cutlass
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|