wmma_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 <assert.h>
#include "cutlass/layout/matrix.h"

namespace cutlass {
namespace arch {


//
// WMMA template structure defines nvcuda::wmma::fragments and static assert for
// wmma native instruction sizes supported for half
//
template <
typename Shape_, 
typename LayoutA_, 
typename LayoutB_,
typename ElementC_,
typename LayoutC_>
struct Wmma<
  Shape_,                                   
  cutlass::half_t,                          
  LayoutA_,                                 
  cutlass::half_t,                          
  LayoutB_,                                 
  ElementC_,                                
  LayoutC_,                                 
  cutlass::arch::OpMultiplyAdd              
> {

#if defined(CUTLASS_ARCH_WMMA_SM70_ENABLED)
  using Shape = Shape_;
  using ElementA = cutlass::half_t;
  using LayoutA = LayoutA_;
  using ElementB = cutlass::half_t;
  using LayoutB = LayoutB_;
  using ElementC = ElementC_;
  using LayoutC = LayoutC_;
  using Operator = cutlass::arch::OpMultiplyAdd;

  // check supported wmma shape for the given multiplicand data types
  static_assert(
    platform::is_same<cutlass::gemm::GemmShape<16, 16, 16>, Shape>::value ||
    platform::is_same<cutlass::gemm::GemmShape< 8, 32, 16>, Shape>::value ||
    platform::is_same<cutlass::gemm::GemmShape<32,  8, 16>, Shape>::value,
    "Supported list of wmma operator shape for f16 multiplicands are: 16x16x16, 8x328x16, and 32x8x16");

  // check supported wmma output data type for the given multiplicand data types
  static_assert(
    platform::is_same<cutlass::half_t, ElementC>::value || platform::is_same<float, ElementC>::value,
    "Supported of wmma output data type for f16 multiplicands are: f16 and f32");

  // Wmma Fragment
  using FragmentA = nvcuda::wmma::fragment<
          nvcuda::wmma::matrix_a,
          Shape::kM,
          Shape::kN,
          Shape::kK,
          typename CutlassToWmmaDataType<ElementA>::Type,
          typename CutlassToWmmaLayout<LayoutA>::Layout>;

  using FragmentB = nvcuda::wmma::fragment<
          nvcuda::wmma::matrix_b,
          Shape::kM,
          Shape::kN,
          Shape::kK,
          typename CutlassToWmmaDataType<ElementB>::Type,
          typename CutlassToWmmaLayout<LayoutB>::Layout>;

  using FragmentC = nvcuda::wmma::fragment<
          nvcuda::wmma::accumulator,
          Shape::kM,
          Shape::kN,
          Shape::kK,
          typename CutlassToWmmaDataType<ElementC>::Type>;

  CUTLASS_DEVICE
  void operator()(
    FragmentC &D, 
    FragmentA const &A, 
    FragmentB const &B, 
    FragmentC const &C) const {
    
      nvcuda::wmma::mma_sync(D, A, B, C);
  }
#else
    static_assert(false, "wmma.mma.sync for floating point multiplicands is avialable only for SM70 and beyond");
#endif

};

} // namespace arch
} // namespace cutlass