353 lines
11 KiB
C++
353 lines
11 KiB
C++
/***************************************************************************************************
|
|
* Copyright (c) 2017 - 2024 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 GEMM 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. Partial
|
|
specializations here choose 'device::GemmTransposed' to implement this functionality.
|
|
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "cutlass/cutlass.h"
|
|
|
|
#include "cutlass/complex.h"
|
|
#include "cutlass/layout/matrix.h"
|
|
#include "cutlass/numeric_types.h"
|
|
|
|
#include "cutlass/gemm/kernel/gemm_planar_complex.h"
|
|
#include "cutlass/gemm/kernel/gemm_planar_complex_array.h"
|
|
#include "cutlass/gemm/kernel/default_gemm.h"
|
|
#include "cutlass/gemm/kernel/default_gemm_complex.h"
|
|
|
|
#include "cutlass/epilogue/threadblock/default_epilogue_planar_complex.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_planar_complex_pipelined.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_planar_complex_multistage.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,
|
|
/// 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,
|
|
/// Math operation performed by GEMM (e.g. arch::OpMultiplyAdd)
|
|
typename Operator,
|
|
/// Conditional enabling to switch between stages
|
|
typename Enable = void
|
|
>
|
|
struct DefaultGemmPlanarComplexUniversal;
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Partial specialization for pipelined mainloop
|
|
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,
|
|
/// 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,
|
|
/// Operation performed by GEMM
|
|
typename Operator
|
|
>
|
|
struct DefaultGemmPlanarComplexUniversal<
|
|
ElementA,
|
|
LayoutA,
|
|
TransformA,
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
TransformB,
|
|
kAlignmentB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
Operator,
|
|
typename platform::enable_if<(Stages <= 2)>::type
|
|
> {
|
|
|
|
/// Define planar complex valued variants instead
|
|
using Mma = typename gemm::threadblock::DefaultMmaPlanarComplexPipelined<
|
|
ElementA,
|
|
LayoutA,
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
kAlignmentB,
|
|
ElementAccumulator,
|
|
LayoutC,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
Stages,
|
|
TransformA,
|
|
TransformB,
|
|
Operator
|
|
>::ThreadblockMma;
|
|
|
|
/// Planar complex epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpiloguePlanarComplex<
|
|
ThreadblockShape,
|
|
typename Mma::Policy::Operator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape::kK / WarpShape::kK,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
/// Define the kernel in terms of the default kernel
|
|
using GemmKernel = kernel::GemmPlanarComplex<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle
|
|
>;
|
|
|
|
// Array variant
|
|
using GemmArrayKernel = kernel::GemmPlanarComplexArray<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Partial specialization for multiple pipeline stages.
|
|
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,
|
|
/// 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,
|
|
/// Operation performed by GEMM
|
|
typename Operator
|
|
>
|
|
struct DefaultGemmPlanarComplexUniversal<
|
|
ElementA,
|
|
LayoutA,
|
|
TransformA,
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
TransformB,
|
|
kAlignmentB,
|
|
ElementC,
|
|
LayoutC,
|
|
ElementAccumulator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
EpilogueOutputOp,
|
|
ThreadblockSwizzle,
|
|
Stages,
|
|
Operator,
|
|
typename platform::enable_if<(Stages > 2)>::type
|
|
> {
|
|
|
|
/// Define planar complex valued variants instead
|
|
using Mma = typename gemm::threadblock::DefaultMmaPlanarComplexMultistage<
|
|
ElementA,
|
|
LayoutA,
|
|
kAlignmentA,
|
|
ElementB,
|
|
LayoutB,
|
|
kAlignmentB,
|
|
ElementAccumulator,
|
|
LayoutC,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape,
|
|
WarpShape,
|
|
InstructionShape,
|
|
Stages,
|
|
TransformA,
|
|
TransformB,
|
|
Operator
|
|
>::ThreadblockMma;
|
|
|
|
/// Planar complex epilogue
|
|
using Epilogue = typename epilogue::threadblock::DefaultEpiloguePlanarComplex<
|
|
ThreadblockShape,
|
|
typename Mma::Policy::Operator,
|
|
OperatorClass,
|
|
ArchTag,
|
|
ThreadblockShape::kK / WarpShape::kK,
|
|
EpilogueOutputOp,
|
|
EpilogueOutputOp::kCount
|
|
>::Epilogue;
|
|
|
|
/// Define the kernel in terms of the default kernel
|
|
using GemmKernel = kernel::GemmPlanarComplex<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle
|
|
>;
|
|
|
|
// Array variant
|
|
using GemmArrayKernel = kernel::GemmPlanarComplexArray<
|
|
Mma,
|
|
Epilogue,
|
|
ThreadblockSwizzle
|
|
>;
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace kernel
|
|
} // namespace gemm
|
|
} // namespace cutlass
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|