162 lines
5.3 KiB
C++
162 lines
5.3 KiB
C++
/***************************************************************************************************
|
|
* 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:
|
|
* * Redistributions of source code must retain the above copyright notice, this list of
|
|
* conditions and the following disclaimer.
|
|
* * 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.
|
|
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 Epilogue for threadblock scoped GEMMs using Tensor Ops.
|
|
|
|
The epilogue rearranges the result of a matrix product through shared memory to match canonical
|
|
tensor layouts in global memory. Epilogues support conversion and reduction operations.
|
|
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "cutlass/cutlass.h"
|
|
#include "cutlass/numeric_types.h"
|
|
#include "cutlass/array.h"
|
|
|
|
#include "cutlass/gemm/gemm.h"
|
|
|
|
#include "cutlass/epilogue/threadblock/default_epilogue_tensor_op.h"
|
|
#include "cutlass/epilogue/threadblock/default_epilogue_volta_tensor_op.h"
|
|
#include "cutlass/epilogue/threadblock/epilogue.h"
|
|
#include "cutlass/epilogue/threadblock/epilogue_with_reduction.h"
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
namespace epilogue {
|
|
namespace threadblock {
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines sensible defaults for epilogues for TensorOps.
|
|
template <
|
|
typename Shape,
|
|
typename WarpMmaTensorOp,
|
|
int PartitionsK,
|
|
typename ElementOutput,
|
|
typename OutputOp,
|
|
typename ReductionOp,
|
|
int ElementsPerAccess
|
|
>
|
|
struct DefaultEpilogueWithReductionTensorOp {
|
|
|
|
/// Use defaults related to the existing epilogue
|
|
using Base = DefaultEpilogueTensorOp<
|
|
Shape,
|
|
WarpMmaTensorOp,
|
|
PartitionsK,
|
|
OutputOp,
|
|
ElementsPerAccess
|
|
>;
|
|
|
|
/// Additional tensor tile iterator
|
|
using TensorTileIterator = cutlass::epilogue::threadblock::PredicatedTileIterator<
|
|
typename Base::OutputTileThreadMap,
|
|
typename OutputOp::ElementTensor
|
|
>;
|
|
|
|
using OutputTileIterator = cutlass::epilogue::threadblock::PredicatedTileIterator<
|
|
typename Base::OutputTileThreadMap,
|
|
ElementOutput
|
|
>;
|
|
|
|
/// Define the epilogue
|
|
using Epilogue = EpilogueWithReduction<
|
|
Shape,
|
|
WarpMmaTensorOp,
|
|
PartitionsK,
|
|
OutputTileIterator,
|
|
TensorTileIterator,
|
|
typename WarpMmaTensorOp::ElementC,
|
|
typename Base::AccumulatorFragmentIterator,
|
|
typename Base::WarpTileIterator,
|
|
typename Base::SharedLoadIterator,
|
|
typename Base::OutputOp,
|
|
ReductionOp,
|
|
typename Base::Padding
|
|
>;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Defines sensible defaults for epilogues for TensorOps.
|
|
template <
|
|
typename Shape,
|
|
typename WarpMmaTensorOp,
|
|
int PartitionsK,
|
|
typename ElementOutput,
|
|
typename OutputOp,
|
|
typename ReductionOp,
|
|
int ElementsPerAccess
|
|
>
|
|
struct DefaultEpilogueWithReductionVoltaTensorOp {
|
|
|
|
/// Use defaults related to the existing epilogue
|
|
using Base = DefaultEpilogueVoltaTensorOp<
|
|
Shape,
|
|
WarpMmaTensorOp,
|
|
PartitionsK,
|
|
OutputOp,
|
|
ElementsPerAccess
|
|
>;
|
|
|
|
/// Additional tensor tile iterator
|
|
using TensorTileIterator = cutlass::epilogue::threadblock::PredicatedTileIterator<
|
|
typename Base::OutputTileThreadMap,
|
|
typename OutputOp::ElementTensor
|
|
>;
|
|
|
|
using OutputTileIterator = cutlass::epilogue::threadblock::PredicatedTileIterator<
|
|
typename Base::OutputTileThreadMap,
|
|
ElementOutput
|
|
>;
|
|
|
|
/// Define the epilogue
|
|
using Epilogue = EpilogueWithReduction<
|
|
Shape,
|
|
WarpMmaTensorOp,
|
|
PartitionsK,
|
|
OutputTileIterator,
|
|
TensorTileIterator,
|
|
typename WarpMmaTensorOp::ElementC,
|
|
typename Base::AccumulatorFragmentIterator,
|
|
typename Base::WarpTileIterator,
|
|
typename Base::SharedLoadIterator,
|
|
typename Base::OutputOp,
|
|
ReductionOp,
|
|
typename Base::Padding
|
|
>;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace threadblock
|
|
} // namespace epilogue
|
|
} // namespace cutlass
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|