default_mma_core_sm50.h

Go to the documentation of this file.

/***************************************************************************************************
 * Copyright (c) 2017-2019, 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 TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 **************************************************************************************************/
#pragma once

#include "cutlass/cutlass.h"
#include "cutlass/array.h"

#include "cutlass/numeric_types.h"
#include "cutlass/matrix_shape.h"

#include "cutlass/layout/matrix.h"
#include "cutlass/transform/pitch_linear_thread_map.h"
#include "cutlass/transform/threadblock/regular_tile_iterator.h"

#include "cutlass/gemm/warp/mma_simt.h"
#include "cutlass/gemm/threadblock/default_mma_core.h"


namespace cutlass {
namespace gemm {
namespace threadblock {


template <
    typename Shape_,
    typename WarpShape_,
    typename ElementA_,
    typename ElementB_,
    typename ElementC_,
    typename LayoutC_,
    typename Operator_>
struct DefaultMmaCore<Shape_, WarpShape_, GemmShape<1, 1, 1>, ElementA_,
                      layout::ColumnMajor, ElementB_, layout::RowMajor,
                      ElementC_, LayoutC_, arch::OpClassSimt, 2, Operator_,
                      > {
  using Shape = Shape_;
  using WarpShape = WarpShape_;
  using InstructionShape = InstructionShape_;
  using ElementA = ElementA_;
  using LayoutA = layout::ColumnMajor;
  using ElementB = ElementB_;
  using LayoutB = layout::RowMajor;
  using ElementC = ElementC_;
  using LayoutC = LayoutC_;
  using OperatorClass = arch::OpClassSimt;

  using WarpCount = GemmShape<
    Shape::kM / WarpShape::kM,
    Shape::kN / WarpShape::kN,
    Shape::kK / WarpShape::kK
  >;

  // Divisility requirements
  static_assert(
    !(Shape::kM % WarpShape::kM) &&
    !(Shape::kN % WarpShape::kN),
    "Threadblock-scoped GEMM should be divisible by warp-scoped GEMM size."
  );

  static int const kWarpSize = warp::WarpSize<arch::OpClassTensorOp>::value;

  static int const kThreads = WarpCount::kCount * kWarpSize;

  //
  // Shared memory layouts
  //

  using SmemLayoutA = layout::ColumnMajor;

  using SmemLayoutB = layout::RowMajor;

  //
  // Iterators to write to shared memory
  //

  using IteratorThreadMapA = transform::PitchLinearStripminedThreadMap<
    layout::PitchLinearShape<Shape::kM, Shape::kK>,
    kThreads,
    1
  >;

  using SmemIteratorA = transform::threadblock::RegularTileIterator<
    MatrixShape<Shape::kM, Shape::kK>, 
    ElementA, 
    SmemLayoutA,
    1,
    IteratorThreadMapA
  >;

  using IteratorThreadMapB = transform::PitchLinearStripminedThreadMap<
    layout::PitchLinearShape<Shape::kN, Shape::kK>,
    kThreads,
    1
  >;

  using SmemIteratorB = transform::threadblock::RegularTileIterator<
    MatrixShape<Shape::kK, Shape::kN>, 
    ElementB, 
    SmemLayoutB,
    0,
    IteratorThreadMapB
  >;

  //
  // Warp-level matrix multiply operator
  //

  // Define the warp-level tensor op
  using WarpMma = cutlass::gemm::warp::MmaSimt<
    WarpShape,
    ElementA, 
    SmemLayoutA, 
    ElementB, 
    SmemLayoutB,
    ElementC, 
    LayoutC,
    warp::MmaSimtPolicy<
      MatrixShape<4, 8>,
      layout::RowMajorInterleaved<2>,
      GemmShape<
        128 / sizeof_bits<ElementA>::value,
        128 / sizeof_bits<ElementB>::value,
        1>
      >
    >
  >;

  using MmaPolicy = MmaPolicy<
    WarpMma,
    MatrixShape<0, 0>,
    MatrixShape<0, 0>,
    WarpCount::kK
  >;
};


} // namespace threadblock
} // namespace gemm
} // namespace cutlass