mma_tensor_op.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/memory_sm75.h"
#include "cutlass/arch/mma_sm75.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.h"

namespace cutlass {
namespace gemm {
namespace warp {


template <
  typename Shape_,
  typename ElementA_,
  typename LayoutA_,
  typename ElementB_,
  typename LayoutB_,
  typename ElementC_,
  typename LayoutC_,
  typename Policy_,
  int PartitionsK_ = 1,
  bool AccumulatorsInRowMajor = false,
  int PartitionsN_ = 1,
  typename Enable = bool
>
class MmaTensorOp {
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;

  static int const kPartitionsK = PartitionsK_;

  static int const kPartitionsN = PartitionsN_;

public:

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

  using FragmentA = typename IteratorA::Fragment;

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

  using FragmentB = typename IteratorB::Fragment;

  using IteratorC = MmaTensorOpAccumulatorTileIterator<
     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<
    Shape::kM / Policy::Operator::Shape::kM,
    (Shape::kN / Policy::Operator::Shape::kN / kPartitionsN > 0) ?
     Shape::kN / Policy::Operator::Shape::kN / kPartitionsN :
      1
  >;

public:

  typename Policy::Operator mma;

public:

  //
  // Methods
  //

  CUTLASS_DEVICE
  MmaTensorOp() {}

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

    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);

      // The offset of multilicand B for current partition
      const int n_off = partitionN_idx * FragmentB::kElements / MmaOperandB::kElements / kPartitionsN;
      // Serpentine visitation order maximizing reuse of Rb
      CUTLASS_PRAGMA_UNROLL
      for (int n = 0; n < MmaIterations::kColumn; ++n) {

        CUTLASS_PRAGMA_UNROLL
        for (int m = 0; m < MmaIterations::kRow; ++m) {

          int m_serpentine = ((n % 2) ? (MmaIterations::kRow - 1 - m) : m);

          if (AccumulatorsInRowMajor) {  // matrix B is reordered
            mma(
              ptr_D[n + m_serpentine * MmaIterations::kColumn],
              ptr_A[m_serpentine],
              ptr_B[n],
              ptr_D[n + m_serpentine * MmaIterations::kColumn]);
          } else {
            mma(
              ptr_D[m_serpentine + (n + n_off) * MmaIterations::kRow],
              ptr_A[m_serpentine],
              ptr_B[n + n_off],
              ptr_D[m_serpentine + (n + n_off) * MmaIterations::kRow]);
          }
        }
      }
  }
};


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