cutlass/include/cutlass/gemm/device/gemm_sparse_universal_with_absmax.h

/***************************************************************************************************
 * Copyright (c) 2024 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: BSD-3-Clause
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 * list of conditions and the following disclaimer.
 *
 * 2. 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.
 *
 * 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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
*/

#pragma once

#include "cutlass/arch/mma.h"
#include "cutlass/cutlass.h"
#include "cutlass/numeric_types.h"
#include "cutlass/arch/arch.h"
#include "cutlass/device_kernel.h"

#include "cutlass/gemm/gemm.h"
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"
#include "cutlass/gemm/kernel/gemm_sparse_universal.h"

#include "cutlass/gemm/kernel/default_gemm_sparse_universal_with_absmax.h"
#include "cutlass/gemm/device/default_gemm_configuration.h"
#include "cutlass/gemm/device/gemm_universal_base.h"

#include "cutlass/layout/permute.h"

////////////////////////////////////////////////////////////////////////////////

namespace cutlass {
namespace gemm {
namespace device {

/////////////////////////////////////////////////////////////////////////////////////////////////

template <
    /// Element type for A matrix operand
    typename ElementA_,
    /// Layout type for A matrix operand
    typename LayoutA_,
    /// Element type for B matrix operand
    typename ElementB_,
    /// Layout type for B matrix operand
    typename LayoutB_,
    /// Element type for C and D matrix operands
    typename ElementC_,
    /// Layout type for C and D matrix operands
    typename LayoutC_,
    /// Element type for internal accumulation
    typename ElementAccumulator_ = ElementC_,
    /// Operator class tag
    typename OperatorClass_ = arch::OpClassTensorOp,
    /// Tag indicating architecture to tune for.  This is the minimum SM that
    /// supports the intended feature. The device kernel can be built
    /// targeting any SM larger than this number.
    typename ArchTag_ = arch::Sm80,
    /// Threadblock-level tile size (concept: GemmShape)
    typename ThreadblockShape_ = typename DefaultGemmConfiguration<
        OperatorClass_, ArchTag_, ElementA_, ElementB_, ElementC_,
        ElementAccumulator_>::ThreadblockShape,
    /// Warp-level tile size (concept: GemmShape)
    typename WarpShape_ = typename DefaultGemmConfiguration<
        OperatorClass_, ArchTag_, ElementA_, ElementB_, ElementC_,
        ElementAccumulator_>::WarpShape,
    /// Instruction-level tile size (concept: GemmShape)
    typename InstructionShape_ = typename DefaultGemmConfiguration<
        OperatorClass_, ArchTag_, ElementA_, ElementB_, ElementC_,
        ElementAccumulator_>::InstructionShape,
    /// Epilogue output operator
    typename EpilogueOutputOp_ = typename DefaultGemmConfiguration<
        OperatorClass_, ArchTag_, ElementA_, ElementB_, ElementC_,
        ElementAccumulator_>::EpilogueOutputOp,
    /// Threadblock-level swizzling operator
    typename ThreadblockSwizzle_ = threadblock::GemmIdentityThreadblockSwizzle<>,
    /// Number of stages used in the pipelined mainloop
    int Stages =
        DefaultGemmConfiguration<OperatorClass_, ArchTag_, ElementA_, ElementB_,
                                 ElementC_, ElementAccumulator_>::kStages,
    /// Access granularity of A matrix in units of elements
    int AlignmentA =
        DefaultGemmConfiguration<OperatorClass_, ArchTag_, ElementA_, ElementB_,
                                 ElementC_, ElementAccumulator_>::kAlignmentA,
    /// Access granularity of B matrix in units of elements
    int AlignmentB =
        DefaultGemmConfiguration<OperatorClass_, ArchTag_, ElementA_, ElementB_,
                                 ElementC_, ElementAccumulator_>::kAlignmentB,
    /// Operation performed by GEMM
    typename Operator_ = typename DefaultGemmConfiguration<
        OperatorClass_, ArchTag_, ElementA_, ElementB_, ElementC_,
        ElementAccumulator_>::Operator>
class GemmSparseUniversalWithAbsmax :
  public GemmUniversalBase<
    typename kernel::DefaultGemmSparseUniversalWithAbsmax<
      ElementA_,
      LayoutA_,
      AlignmentA,
      ElementB_,
      LayoutB_,
      AlignmentB,
      ElementC_,
      LayoutC_,
      ElementAccumulator_,
      OperatorClass_,
      ArchTag_,
      ThreadblockShape_,
      WarpShape_,
      InstructionShape_,
      EpilogueOutputOp_,
      ThreadblockSwizzle_,
      Stages,
      Operator_
    >::GemmKernel
  > {

 public:

  static_assert((platform::is_same<LayoutC_, layout::RowMajor>::value),
             "Epilogue of Ada sparse GEMM must be row major for now.");

  using ElementAccumulator = ElementAccumulator_;
  using OperatorClass = OperatorClass_;
  using ArchTag = ArchTag_;
  using ThreadblockShape = ThreadblockShape_;
  using WarpShape = WarpShape_;
  using InstructionShape = InstructionShape_;
  using EpilogueOutputOp = EpilogueOutputOp_;
  using ThreadblockSwizzle = ThreadblockSwizzle_;
  using Operator = Operator_;
  static int const kStages = Stages;
  static int const kAlignmentA = AlignmentA;
  static int const kAlignmentB = AlignmentB;
  static int const kAlignmentC = EpilogueOutputOp::kCount;

  using Base = GemmUniversalBase<
    typename kernel::DefaultGemmSparseUniversalWithAbsmax<
      ElementA_,
      LayoutA_,
      AlignmentA,
      ElementB_,
      LayoutB_,
      AlignmentB,
      ElementC_,
      LayoutC_,
      ElementAccumulator_,
      OperatorClass_,
      ArchTag_,
      ThreadblockShape_,
      WarpShape_,
      InstructionShape_,
      EpilogueOutputOp_,
      ThreadblockSwizzle_,
      Stages,
      Operator_
    >::GemmKernel
  >;

  using Arguments = typename Base::Arguments;
  using GemmKernel = typename Base::GemmKernel;

  using ElementE = typename GemmKernel::ElementE;

  using LayoutE = typename GemmKernel::LayoutE;

  static int const kAlignmentE = 128 / sizeof_bits<ElementE>::value;

  static int const kSparse = GemmKernel::kSparse;
  static int const kMetaSizeInBits = GemmKernel::kMetaSizeInBits;
  static int const kElementsPerElementE = GemmKernel::kElementsPerElementE;
};

////////////////////////////////////////////////////////////////////////////////

} // namespace device
} // namespace gemm
} // namespace cutlass

////////////////////////////////////////////////////////////////////////////////