/***************************************************************************************************
 * Copyright (c) 2017-2021, 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.
 *
 **************************************************************************************************/

/***************************************************************************************************
* Copyright (c) 2021 Iluvatar CoreX. All rights reserved.
* Copyright Declaration: This software, including all of its code and documentation,
* except for the third-party software it contains, is a copyrighted work of Shanghai Iluvatar CoreX
* Semiconductor Co., Ltd. and its affiliates ("Iluvatar CoreX") in accordance with the PRC Copyright
* Law and relevant international treaties, and all rights contained therein are enjoyed by Iluvatar
* CoreX. No user of this software shall have any right, ownership or interest in this software and
* any use of this software shall be in compliance with the terms and conditions of the End User
* License Agreement.
 **************************************************************************************************/


/*! \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/mma.h"

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

#include "cutlass/gemm/gemm.h"
#include "cutlass/gemm/kernel/gemm.h"
#include "cutlass/gemm/kernel/gemm_pipelined.h"
#include "cutlass/gemm/threadblock/default_mma.h"
#include "cutlass/gemm/threadblock/default_mma_core_simt.h"
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"

#include "cutlass/epilogue/threadblock/default_epilogue_simt.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 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,
    /// If true, kernel is configured to support serial reduction in the
    /// epilogue
    bool SplitKSerial,
    /// Operation performed by GEMM
    typename Operator>
struct DefaultGemm;

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

/// Partial specialization for SIMT
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 A matrix in units of elements
    int kAlignmentB,
    /// Element type for C and D matrix operands
    typename ElementC,
    /// Element type for internal accumulation
    typename ElementAccumulator,
    /// Tag indicating architecture to tune for
    typename ArchTag,
    /// Threadblock-level tile size (concept: GemmShape)
    typename ThreadblockShape,
    /// Warp-level tile size (concept: GemmShape)
    typename WarpShape,
    /// Epilogue output operator
    typename EpilogueOutputOp,
    /// Threadblock-level swizzling operator
    typename ThreadblockSwizzle,
    /// If true, kernel is configured to support serial reduction in the epilogue
    bool SplitKSerial,
    /// Operation performed by GEMM
    typename Operator
  >
struct DefaultGemm<
    ElementA,
    LayoutA,
    kAlignmentA,
    ElementB,
    LayoutB,
    kAlignmentB,
    ElementC,
    layout::RowMajor,
    ElementAccumulator,
    arch::OpClassSimt,
    ArchTag,
    ThreadblockShape,
    WarpShape,
    GemmShape<1, 1, 1>,
    EpilogueOutputOp,
    ThreadblockSwizzle,
    2,
    SplitKSerial,
    Operator> {
  /// Define the threadblock-scoped matrix multiply-accumulate
  using Mma = typename cutlass::gemm::threadblock::DefaultMma<
      ElementA,
      LayoutA,
      kAlignmentA,
      ElementB,
      LayoutB,
      kAlignmentB,
      ElementAccumulator,
      layout::RowMajor,
      arch::OpClassSimt,
      arch::Sm50,
      ThreadblockShape,
      WarpShape,
      GemmShape<1, 1, 1>,
      2,
      Operator>::ThreadblockMma;

  static int const kEpilogueElementsPerAccess = EpilogueOutputOp::kCount;
  static_assert(kEpilogueElementsPerAccess == 1, "simt epilogue must operate on scalars");

  /// Define the epilogue
  using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
      ThreadblockShape,
      typename Mma::Operator,
      EpilogueOutputOp,
      kEpilogueElementsPerAccess
      >::Epilogue;

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

////////////////////////////////////////////////////////////////////////////////
/// Partial specialization for SIMT DP4A

template <
    /// Layout type for A matrix operand
    typename LayoutA,
    /// Access granularity of A matrix in units of elements
    int kAlignmentA,
    /// Layout type for B matrix operand
    typename LayoutB,
    /// Access granularity of A matrix in units of elements
    int kAlignmentB,
    /// Layout type for C matrix operand
    typename LayoutC,
    /// Element type for C and D matrix operands
    typename ElementC,
    /// Tag indicating architecture to tune for
    typename ArchTag,
    /// Element type for internal accumulation
    typename ElementAccumulator,
    /// Threadblock-level tile size (concept: GemmShape)
    typename ThreadblockShape,
    /// Warp-level tile size (concept: GemmShape)
    typename WarpShape,
    /// Epilogue output operator
    typename EpilogueOutputOp,
    /// Threadblock-level swizzling operator
    typename ThreadblockSwizzle,
    /// If true, kernel is configured to support serial reduction in the
    /// epilogue
    bool SplitKSerial,
    /// Operation performed by GEMM
    typename Operator>
struct DefaultGemm<int8_t, LayoutA, kAlignmentA, int8_t, LayoutB, kAlignmentB,
                   ElementC, LayoutC, ElementAccumulator, arch::OpClassSimt,
                   ArchTag, ThreadblockShape, WarpShape, GemmShape<1, 1, 4>,
                   EpilogueOutputOp, ThreadblockSwizzle, 2, SplitKSerial,
                   Operator> {
  using InstructionShape = GemmShape<1, 1, 4>;
  using ElementA = int8_t;
  using ElementB = int8_t;

  using OperatorClass =  arch::OpClassSimt;
  /// Define the threadblock-scoped matrix multiply-accumulate
  using Mma = typename cutlass::gemm::threadblock::DefaultMma<ElementA,
      LayoutA,
      kAlignmentA,
      ElementB,
      LayoutB,
      kAlignmentB,
      ElementAccumulator,
      LayoutC,
      arch::OpClassSimt,
      arch::Sm50,
      ThreadblockShape,
      WarpShape,
      InstructionShape,
      2,
      Operator,
      false
      >::ThreadblockMma;

  static int const kEpilogueElementsPerAccess = EpilogueOutputOp::kCount;
  static_assert(kEpilogueElementsPerAccess == 1, "simt epilogue must operate on scalars");

  /// Define the epilogue
  using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
      ThreadblockShape,
      typename Mma::Operator,
      EpilogueOutputOp,
      kEpilogueElementsPerAccess
      >::Epilogue;

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


////////////////////////////////////////////////////////////////////////////////
/// Partial specialization for BigIsland 1.0 tensor op architecture
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 C and D matrix operands
  typename ElementC,
  /// Element type for internal accumulation
  typename ElementAccumulator,
  /// Tag indicating architecture to tune for
  typename ArchTag,
  /// Threadblock-level tile size (concept: GemmShape)
  typename ThreadblockShape,
  /// Warp-level tile size (concept: GemmShape)
  typename WarpShape,
  /// Instrcution shape
  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 GEMM
  typename Operator
>
struct DefaultGemm<
  ElementA, LayoutA, kAlignmentA,
  ElementB, LayoutB, kAlignmentB,
  ElementC, layout::RowMajor,
  ElementAccumulator,
  arch::OpClassTensorOp,
  ArchTag,
  ThreadblockShape,
  WarpShape,
  InstructionShape,
  EpilogueOutputOp,
  ThreadblockSwizzle,
  Stages,
  SplitKSerial,
  Operator
> {

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

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

  /// FIXME(Peter Han): Probably DefaultEpiloguesTensorOp should be used here, let's see
  static const int kEpilougeElementsPerAccess = EpilogueOutputOp::kCount;

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

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


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

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