default_rank_k_universal.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 Rank k  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.

  
*/

#pragma once

#include "cutlass/blas3.h"

#include "cutlass/complex.h"
#include "cutlass/layout/matrix.h"

#include "cutlass/gemm/kernel/rank_k_universal.h"
#include "cutlass/gemm/kernel/default_rank_k.h"
#include "cutlass/gemm/kernel/default_rank_k_complex.h"

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

namespace cutlass {
namespace gemm {
namespace kernel {

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

template <
    /// Element type for A matrix operand
    typename ElementA_,
    /// Layout type for A matrix operand
    typename LayoutA_,
    /// Complex elementwise transformation on A operand
    ComplexTransform TransformA,
    /// Access granularity of A matrix in units of elements
    int kAlignmentA,
    /// Element type for C and D matrix operands
    typename ElementC_,
    /// Layout type for C and D matrix operands
    typename LayoutC_,
    /// Fill Mode for C (kLower or kUpper)
    FillMode FillModeC_,
    /// 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,
    /// If true, kernel is configured to support serial reduction in the
    /// epilogue
    bool SplitKSerial,
    /// Operation performed by SYRK
    typename Operator,
    /// Blas3 computation mode (symmetric/hermitian)
    BlasMode BlasMode_ = BlasMode::kSymmetric,
    ///
    typename Enable = void
    >
struct DefaultRankKUniversal;

/////////////////////////////////////////////////////////////////////////////////////////////////
//
// Real-valued Rank k update kernels
//

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 C and D matrix operands
    typename ElementC,
    /// Layout type for C and D matrix operands
    typename LayoutC,
    /// Fill Mode for C (kLower or kUpper)
    FillMode FillModeC,
    /// 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,
    /// If true, kernel is configured to support serial reduction in the
    /// epilogue
    bool SplitKSerial,
    /// Operation performed by Rank2k
    typename Operator>
struct DefaultRankKUniversal<
  ElementA,
  LayoutA,
  ComplexTransform::kNone,   // transform A
  kAlignmentA,
  ElementC,
  LayoutC,
  FillModeC,
  ElementAccumulator,
  OperatorClass,
  ArchTag,
  ThreadblockShape,
  WarpShape,
  InstructionShape,
  EpilogueOutputOp,
  ThreadblockSwizzle,
  Stages,
  SplitKSerial,
  Operator,
  BlasMode::kSymmetric,
  typename platform::enable_if< ! cutlass::is_complex<ElementAccumulator>::value>::type
> {

  using DefaultRankKkernel = typename kernel::DefaultRankK<
    ElementA,
    LayoutA,
    kAlignmentA,
    ElementC,
    LayoutC,
    FillModeC,
    ElementAccumulator,
    OperatorClass,
    ArchTag,
    ThreadblockShape,
    WarpShape,
    InstructionShape,
    EpilogueOutputOp,
    ThreadblockSwizzle,
    Stages,
    SplitKSerial,
    Operator,
    BlasMode::kSymmetric
  >::RankKkernel;

    /// Define the kernel in terms of the default kernel
  using RankKkernel = kernel::RankKUniversal<
    typename DefaultRankKkernel::Mma,
    typename DefaultRankKkernel::Epilogue, 
    ThreadblockSwizzle,
    FillModeC
  >;
};

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

//
// Complex-valued Rank 2K update kernels
//
template <
    /// Element type for A matrix operand
    typename ElementA,
    /// Layout type for A matrix operand
    typename LayoutA,
    /// Complex elementwise transformation on A operand
    ComplexTransform TransformA,
    /// Access granularity of A matrix in units of elements
    int kAlignmentA,
    /// Element type for C and D matrix operands
    typename ElementC,
    /// Layout type for C and D matrix operands
    typename LayoutC,
    /// Fill Mode for C (kLower or kUpper)
    FillMode FillModeC,
    /// 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,
    /// If true, kernel is configured to support serial reduction in the
    /// epilogue
    bool SplitKSerial,
    /// Operation performed by SYRK
    typename Operator,
    // BlasMode
    BlasMode kBlasMode
  >

struct DefaultRankKUniversal<
  ElementA,
  LayoutA,
  TransformA,   
  kAlignmentA,
  ElementC,
  LayoutC,
  FillModeC,
  ElementAccumulator,
  OperatorClass,
  ArchTag,
  ThreadblockShape,
  WarpShape,
  InstructionShape,
  EpilogueOutputOp,
  ThreadblockSwizzle,
  Stages,
  SplitKSerial,
  Operator,
  kBlasMode,
  typename platform::enable_if<cutlass::is_complex<ElementAccumulator>::value>::type
> {

  using DefaultRankKkernel = typename kernel::DefaultRankKComplex<
    ElementA,
    LayoutA,
    ElementC,
    LayoutC,
    FillModeC,
    ElementAccumulator,
    OperatorClass,
    ArchTag,
    ThreadblockShape,
    WarpShape,
    InstructionShape,
    EpilogueOutputOp,
    ThreadblockSwizzle,
    Stages,
    TransformA,
    Operator,
    SplitKSerial,
    kBlasMode
  >::RankKkernel;

    /// Define the kernel in terms of the default kernel
  using RankKkernel = kernel::RankKUniversal<
    typename DefaultRankKkernel::Mma,
    typename DefaultRankKkernel::Epilogue, 
    ThreadblockSwizzle,
    FillModeC
  >;
};

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

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