mma_tensor_op_sm70.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/arch/mma.h"

#include "cutlass/gemm/gemm.h"
#include "cutlass/gemm/warp/mma.h"

#include "cutlass/gemm/warp/mma_tensor_op_policy.h"
#include "cutlass/gemm/warp/mma_tensor_op_tile_iterator_sm70.h"


namespace cutlass {
namespace gemm {
namespace warp {


template <
  typename Shape_,
  typename ElementA_,
  typename LayoutA_,
  typename ElementB_,
  typename LayoutB_,
  typename ElementC_,
  typename LayoutC_,
  typename Policy_,
  typename Enable = bool
>
class MmaVoltaTensorOp {
public:
  using Shape = Shape_;

  using ElementA = ElementA_;

  using LayoutA = LayoutA_;

  using ElementB = ElementB_;

  using LayoutB = LayoutB_;

  using ElementC = ElementC_;

  using LayoutC = LayoutC_;

  using Policy = Policy_;

  using OperatorClass = arch::OpClassTensorOp;

  static int const kThreadCount = 32;

  using InterleavedTileShape = GemmShape<32, 32, 4>;

  static_assert(!(Shape::kM % InterleavedTileShape::kM) &&
                !(Shape::kN % InterleavedTileShape::kN),
                "Shape must be a multiple of InterleavedTileShape.");
public:

  using IteratorA = MmaVoltaTensorOpMultiplicandTileIterator<
    MatrixShape<Shape::kM, Shape::kK>,
    Operand::kA,
    ElementA,
    LayoutA,
    MatrixShape<
      Policy::Operator::Shape::kM,
      Policy::Operator::Shape::kK
    >,
    Policy::OpDelta::kRow,
    kThreadCount
  >;

  using FragmentA = typename IteratorA::Fragment;

  using IteratorB = MmaVoltaTensorOpMultiplicandTileIterator<
    MatrixShape<Shape::kK, Shape::kN>,
    Operand::kB,
    ElementB,
    LayoutB,
    MatrixShape<
      Policy::Operator::Shape::kK,
      Policy::Operator::Shape::kN
    >,
    Policy::OpDelta::kRow,
    kThreadCount
  >;

  using FragmentB = typename IteratorB::Fragment;

  using IteratorC = MmaVoltaTensorOpAccumulatorTileIterator<
    MatrixShape<Shape::kM, Shape::kN>,
    ElementC,
    LayoutC,
    typename Policy::Operator::Shape,
    typename Policy::OpDelta
  >;

  using FragmentC = typename IteratorC::Fragment;

private:

  static_assert(
    !(Shape::kM % Policy::Operator::Shape::kM) && 
    !(Shape::kN % Policy::Operator::Shape::kN),
    "Shape of warp-level Mma must be divisible by operator shape.");

  using MmaIterations = MatrixShape<
    InterleavedTileShape::kM / Policy::Operator::Shape::kM,
    InterleavedTileShape::kN / Policy::Operator::Shape::kN
  >;
  using TileIterations = MatrixShape<
    Shape::kM / InterleavedTileShape::kM,
    Shape::kN / InterleavedTileShape::kN
  >;

  // Whether matrix B is reordered
  bool reorder_B_;

public:

  typename Policy::Operator mma;

public:

  //
  // Methods
  //
  
  CUTLASS_DEVICE
  MmaVoltaTensorOp() {}

  CUTLASS_DEVICE
  void operator()(
    FragmentC &D, 
    FragmentA const &A, 
    FragmentB const &B, 
    FragmentC const &C,
    int const &partitionN_idx = 0)  {

    using MmaOperandA = typename Policy::Operator::FragmentA;
    using MmaOperandB = typename Policy::Operator::FragmentB;
    using MmaOperandC = typename Policy::Operator::FragmentC;

    D = C;

    MmaOperandA const *ptr_A = reinterpret_cast<MmaOperandA const *>(&A);
    MmaOperandB const *ptr_B = reinterpret_cast<MmaOperandB const *>(&B);
    MmaOperandC *ptr_D = reinterpret_cast<MmaOperandC *>(&D);

    CUTLASS_PRAGMA_UNROLL
    for (int outer_col = 0; outer_col < TileIterations::kColumn; ++outer_col) {
      CUTLASS_PRAGMA_UNROLL
      for (int inner_col = 0; inner_col < MmaIterations::kColumn; ++inner_col) {
        CUTLASS_PRAGMA_UNROLL
        for (int outer_row = 0; outer_row < TileIterations::kRow; ++outer_row) {
          CUTLASS_PRAGMA_UNROLL

          for (int inner_row = 0; inner_row < MmaIterations::kRow; ++inner_row) {
      
            int op_col = inner_col + MmaIterations::kColumn * outer_col;

            // Column-major serpentine sequence to maximize reuse of A operand.
            int inner_row_serp = inner_row;
            int outer_row_serp = outer_row;
            if (op_col & 1) {
              inner_row_serp = MmaIterations::kRow - inner_row - 1;
              outer_row_serp = TileIterations::kRow - outer_row - 1;
            }
            int op_row = inner_row_serp + MmaIterations::kRow * outer_row_serp;
            int op_idx = inner_row_serp + MmaIterations::kRow * 
                         (inner_col + MmaIterations::kColumn * 
                          (outer_row_serp + TileIterations::kRow * outer_col));
            mma(
              ptr_D[op_idx],
              ptr_A[op_row],
              ptr_B[op_col],
              ptr_D[op_idx]);

          }
        }
      }
    }
  }
};


} // namespace warp
} // namespace gemm
} // namespace cutlass