default_gemm_layernorm_mainloop_fusion.h

/***************************************************************************************************
 * Copyright (c) 2017 - 2025 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 
      Default kernel-level GEMM definitions combine threadblock-scoped matrix multiply-add with
      the appropriate threadblock-scoped epilogue.
  
      Note, CUTLASS epilogues universally target row-major outputs. Column-major outputs are
      accommodated by exchanging A and B operands and assuming transposed layouts. Partial
      specializations here choose 'device::GemmTransposed' to implement this functionality.
*/

#pragma once

#include "cutlass/cutlass.h"

#include "cutlass/layout/matrix.h"
#include "cutlass/numeric_types.h"
#include "cutlass/arch/wmma.h"

#include "cutlass/epilogue/threadblock/epilogue.h"
#include "cutlass/epilogue/thread/linear_combination.h"

#include "cutlass/gemm/gemm.h"
#include "cutlass/gemm/kernel/gemm_layernorm_mainloop_fusion.h"
#include "cutlass/gemm/threadblock/default_mma_layernorm_mainloop_fusion.h"
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"

#include "cutlass/epilogue/threadblock/default_epilogue_tensor_op.h"
#include "cutlass/transform/threadblock/predicated_tile_iterator.h"

namespace cutlass {
namespace gemm {
namespace kernel {

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

template <
    /// Element type for A matrix operand
    typename ElementA,
    /// Layout type for A matrix operand
    typename LayoutA,
    /// Access granularity of A matrix in units of elements
    int kAlignmentA,
    /// Element type for B matrix operand
    typename ElementB,
    /// Layout type for B matrix operand
    typename LayoutB,
    /// Access granularity of B matrix in units of elements
    int kAlignmentB,
    /// Element type for Scale/Bias vectors
    typename ElementScaleBias,
    /// Layout type for Scale/Bias vectors
    typename LayoutScaleBias,
    /// 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,
    /// Operator class tag
    typename OperatorClass,
    /// Tag indicating architecture to tune for
    typename ArchTag,
    /// Threadblock-level tile size (concept: GemmShape)
    typename ThreadblockShape,
    /// Warp-level tile size (concept: GemmShape)
    typename WarpShape,
    /// Warp-level tile size (concept: GemmShape)
    typename InstructionShape,
    /// Epilogue output operator
    typename EpilogueOutputOp,
    /// Threadblock-level swizzling operator
    typename ThreadblockSwizzle,
    /// Number of stages used in the pipelined mainloop
    int Stages,
    /// Operation performed by GEMM
    typename Operator,
    /// Use zfill or predicate for out-of-bound cp.async
    SharedMemoryClearOption SharedMemoryClear = SharedMemoryClearOption::kNone>
struct DefaultGemmLayernormMainloopFusion {

  /// Define the threadblock-scoped matrix multiply-accumulate
  using Mma = typename cutlass::gemm::threadblock::DefaultMmaLayernormMainloopFusion<
      ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
      ElementScaleBias, LayoutScaleBias, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, arch::Sm80,
      ThreadblockShape, WarpShape, InstructionShape, Stages,
      Operator, false, SharedMemoryClear>::ThreadblockMma;

  static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;

  /// Define the epilogue
  using Epilogue =
      typename cutlass::epilogue::threadblock::DefaultEpilogueTensorOp<
          ThreadblockShape, typename Mma::Operator, kPartitionsK, EpilogueOutputOp,
          EpilogueOutputOp::kCount>::Epilogue;

  /// Define the kernel-level GEMM operator.
  using GemmKernel = kernel::GemmLayernormMainloopFusion<Mma, Epilogue, ThreadblockSwizzle>;
};

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

}  // namespace kernel
}  // namespace gemm
}  // namespace cutlass

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