/***************************************************************************************************
 * Copyright (c) 2017 - 2023 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 Epilogue for threadblock scoped GEMMs using Tensor Ops.

  The epilogue rearranges the result of a matrix product through shared memory to match canonical
  tensor layouts in global memory. Epilogues support conversion and reduction operations.

  
*/

#pragma once

#include "cutlass/cutlass.h"
#include "cutlass/numeric_types.h"
#include "cutlass/array.h"
#include "cutlass/layout/matrix.h"
#include "cutlass/layout/tensor.h"
#include "cutlass/matrix_shape.h"
#include "cutlass/tensor_ref.h"
#include "cutlass/transform/pitch_linear_thread_map.h"
#include "cutlass/epilogue/threadblock/output_tile_thread_map.h"
#include "cutlass/arch/arch.h"
#include "cutlass/arch/memory.h"
#include "cutlass/epilogue/threadblock/predicated_tile_iterator_params.h"

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

namespace cutlass {

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

namespace epilogue {
namespace threadblock {

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

/// Tile iterator used to load and store output tile from global memory in epilogue.
///
/// Satisfies: ReadableTileIterator | PredicatedTileIterator | ForwardTileIterator
///
template <
  typename ThreadMap_,                     ///< Thread map (conept: OutputTileThreadMap)
  typename Element_,                        ///< Element data type
  BlasMode BlasMode_ = BlasMode::kGemm   ///< Tile Iterator for a Symmetric or Hermitian Kernel
>
class PredicatedTileIteratorBlas3 {
public:
  using ThreadMap = ThreadMap_;
  using Shape = typename ThreadMap::Shape;

  using Element = Element_;

  using Layout = layout::RowMajor;
  using TensorRef = TensorRef<Element, Layout>;
  using ConstTensorRef = typename TensorRef::ConstTensorRef;

  using Index = typename Layout::Index;
  using LongIndex = typename Layout::LongIndex;
  using TensorCoord = MatrixCoord;

  static BlasMode const kBlasMode = BlasMode_;

  static int const kElementsPerAccess = ThreadMap::kElementsPerAccess;
  static int const kThreads = ThreadMap::kThreads;
  static int const kIterations = ThreadMap::Count::kTile;

  static_assert( ThreadMap::Iterations::kRow > 0,"ThreadMap::Iterations::kRow must be > 0");
  static_assert( ThreadMap::Iterations::kGroup > 0,"ThreadMap::Iterations::kGroup must be > 0");
  static_assert( ThreadMap::Iterations::kCluster > 0,"ThreadMap::Iterations::kCluster must be > 0");
  static_assert( ThreadMap::Iterations::kColumn > 0,"ThreadMap::Iterations::kColumn must be > 0");

  /// Fragment object
  using Fragment = Array<
    Element, 
    ThreadMap::Iterations::kColumn * 
    ThreadMap::Iterations::kRow * 
    ThreadMap::Iterations::kGroup * 
    ThreadMap::Iterations::kCluster * ThreadMap::kElementsPerAccess>;

  /// Memory access size
  using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>;
  static_assert( AccessType::kElements == 1, "BLAS3 Epilogue must use AccessType::kElements as 1");

  //
  // Parameters struct
  //

  /// Uses a non-template class
  struct Params : PredicatedTileIteratorParams {

    CUTLASS_HOST_DEVICE
    Params() { }

    CUTLASS_HOST_DEVICE
    Params(Layout const &layout): 
      PredicatedTileIteratorParams(
        layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess,
        make_OutputTileThreadMapDesc<ThreadMap>()
      ) 
    {
        
    }
  };

  /// Mask object
  struct Mask {

    static int const kCount = ThreadMap::Iterations::kColumn;

    /// Predicate state
    bool predicates[kCount];

    //
    // Mask
    //
    CUTLASS_HOST_DEVICE
    Mask() {
      enable();
    }

    ///< Efficiently disables all accesses guarded by mask
    CUTLASS_HOST_DEVICE void clear() {
      CUTLASS_PRAGMA_UNROLL
      for (int i = 0; i < kCount; ++i) {
        predicates[i] = false;
      }
    }

    ///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask
    CUTLASS_DEVICE void enable() {
      CUTLASS_PRAGMA_UNROLL
      for (int i = 0; i < kCount; ++i) {
        predicates[i] = true;
      }
    }
  };

private:

  //
  // Data members
  //

  /// Parameters structure containing reference and precomputed state.
  PredicatedTileIteratorParams params_;

  /// Byte-level pointer
  uint8_t *byte_pointer_;

  /// Fill Mode for a tile on diagonal of a symmetric kernel
  cutlass::FillMode fill_mode;

  /// Array of boolean values to contain steady-state predicates
  Mask mask_;

  /// Extent of the matrix tile in rows
  Index extent_row_;

  /// A thread's starting row position (assuming steady-state predicates have been computed)
  Index thread_start_row_;

  /// Internal state counter
  int state_[3];

  /// Starting address of the matrix  
  size_t matrix_start_addr; 
 
  static_assert((kBlasMode == BlasMode::kSymmetric || kBlasMode == BlasMode::kHermitian), 
    "Unsupported blas3 mode.");

private:

  //
  // Methods
  //

public:

  //
  // Methods
  //

  /// Constructor
  CUTLASS_DEVICE
  PredicatedTileIteratorBlas3(
    PredicatedTileIteratorParams const & params,
    Element *pointer,
    TensorCoord extent,
    int thread_idx,
    TensorCoord threadblock_offset
    , cutlass::FillMode fill_mode
  ): 
    params_(params), fill_mode(fill_mode)
  {

    TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset;

    extent_row_ = extent.row();
    thread_start_row_ = thread_offset.row();

    // Initialize predicates
    CUTLASS_PRAGMA_UNROLL
    for (int c = 0; c < ThreadMap::Iterations::kColumn; ++c) {

      mask_.predicates[c] = ((thread_offset.column() 
        + ThreadMap::Delta::kColumn * c) < extent.column());
    }

    // Check Symmetric kernel modes (Lower and Upper - for diagonal CTAs, None for rest CTAs)
    if ((kBlasMode == BlasMode::kSymmetric || kBlasMode == BlasMode::kHermitian) && 
        fill_mode == cutlass::FillMode::kInvalid) {
      arch::device_breakpoint();
    }

    // Starting address of the matrix
    matrix_start_addr =  reinterpret_cast<size_t>(pointer); 

    // Initialize pointer
    byte_pointer_ = reinterpret_cast<uint8_t *>(pointer) + 
      LongIndex(thread_offset.row()) * LongIndex(params_.stride) + 
      LongIndex(thread_offset.column()) * sizeof(AccessType) / kElementsPerAccess;

    // Initialize internal state counter
    state_[0] = state_[1] = state_[2] = 0;
  }

  /// Adds a pointer offset in units of Element
  CUTLASS_HOST_DEVICE
  void add_pointer_offset(LongIndex pointer_offset) {
    byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8;
  }

  /// Loads a fragment from memory
  CUTLASS_DEVICE
  void load_with_byte_offset(Fragment &frag, int64_t byte_offset) {

    uint8_t *byte_pointer = byte_pointer_;
    AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);

    CUTLASS_PRAGMA_UNROLL
    for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {

      CUTLASS_PRAGMA_UNROLL
      for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {

        CUTLASS_PRAGMA_UNROLL
        for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {

          int frag_row_idx = 
            (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));

          int row_offset = row * ThreadMap::Delta::kRow 
            + group * ThreadMap::Delta::kGroup 
            + cluster * ThreadMap::Delta::kCluster;

          bool row_guard = ((row_offset + thread_start_row_) < extent_row_);

          AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);

          CUTLASS_PRAGMA_UNROLL
          for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {

            bool guard = row_guard && mask_.predicates[column];

            cutlass::arch::global_load<
              AccessType, 
              sizeof(AccessType)
            >(
                frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
                         column],
                (void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
                                        kElementsPerAccess],
                guard);
          }

          if (row + 1 < ThreadMap::Iterations::kRow) {
            byte_pointer += params_.increment_row;
          }
        }

        if (group + 1 < ThreadMap::Iterations::kGroup) {
          byte_pointer += params_.increment_group;
        }
      }

      if (cluster + 1 < ThreadMap::Iterations::kCluster) {
        byte_pointer += params_.increment_cluster;
      }
    }
  }

  /// Loads a fragment on the diagonal of a symmetric kernel to memory 
  CUTLASS_DEVICE
  void load_symmetric_with_byte_offset(Fragment &frag, int64_t byte_offset) {

    uint8_t *byte_pointer = byte_pointer_;
    AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);

    bool isLowerMode = (fill_mode == cutlass::FillMode::kLower) ? true : false;

    CUTLASS_PRAGMA_UNROLL
    for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {

      CUTLASS_PRAGMA_UNROLL
      for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {

        CUTLASS_PRAGMA_UNROLL
        for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {

          int frag_row_idx = 
            (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));

          int row_offset = row * ThreadMap::Delta::kRow 
            + group * ThreadMap::Delta::kGroup 
            + cluster * ThreadMap::Delta::kCluster;

          bool row_guard = ((row_offset + thread_start_row_) < extent_row_);

          AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);

          // Offset of row from beginning of the matrix per thread
          size_t row_start_offset = (size_t)memory_pointer - matrix_start_addr;

          // Absolute row index
          int row_index = int(row_start_offset/params_.stride);

          CUTLASS_PRAGMA_UNROLL
          for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {

            bool guard = row_guard && mask_.predicates[column];

            // Offset of column from beginning of row per thread     
            size_t col_start_offset = row_start_offset + 
                        (column * ThreadMap::Delta::kColumn / kElementsPerAccess) * sizeof(AccessType);

            // Absolute column index
            size_t col_index = (col_start_offset%params_.stride)/sizeof(AccessType);
            guard = guard && ( (isLowerMode && row_index >= col_index) ||
                               (!isLowerMode && row_index <= col_index) );

            cutlass::arch::global_load<
              AccessType, 
              sizeof(AccessType)
            >(
                frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
                         column],
                (void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
                                        kElementsPerAccess],
                guard);

            // The imaginary parts of the diagonal elements of a complex element are assumed and set to zero
            if (guard && kBlasMode == BlasMode::kHermitian && cutlass::is_complex<Element>::value) {
              Element *scalar_ptr = reinterpret_cast<Element *>(frag_ptr);

              if (row_index == col_index) {
                scalar_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column] = 
                  real(scalar_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column]);
              }
            }
          }

          if (row + 1 < ThreadMap::Iterations::kRow) {
            byte_pointer += params_.increment_row;
          }
        }

        if (group + 1 < ThreadMap::Iterations::kGroup) {
          byte_pointer += params_.increment_group;
        }
      }

      if (cluster + 1 < ThreadMap::Iterations::kCluster) {
        byte_pointer += params_.increment_cluster;
      }
    }
  }

  /// Loads a fragment from memory
  CUTLASS_DEVICE
  void load(Fragment &frag) {
    
    if (fill_mode == cutlass::FillMode::kNone) {
      load_with_byte_offset(frag, 0);
    }
    else {
      load_symmetric_with_byte_offset(frag, 0);
    }
  }

  /// Stores a fragment to memory
  CUTLASS_DEVICE
  void store_with_byte_offset(Fragment const &frag, int64_t byte_offset) {
    uint8_t *byte_pointer = byte_pointer_;
    AccessType const *frag_ptr = reinterpret_cast<AccessType const *>(&frag);

    CUTLASS_PRAGMA_UNROLL
    for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {

      CUTLASS_PRAGMA_UNROLL
      for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {

        CUTLASS_PRAGMA_UNROLL
        for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {

          int frag_row_idx = 
            (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));

          int row_offset = row * ThreadMap::Delta::kRow 
            + group * ThreadMap::Delta::kGroup 
            + cluster * ThreadMap::Delta::kCluster;

          bool row_guard = ((row_offset + thread_start_row_) < extent_row_);

          AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);

          CUTLASS_PRAGMA_UNROLL
          for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {

            bool guard = row_guard && mask_.predicates[column];

            cutlass::arch::global_store<AccessType, sizeof(AccessType)>(
                frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column],
                (void *)&memory_pointer[column * ThreadMap::Delta::kColumn / kElementsPerAccess],
                guard);
          }

          if (row + 1 < ThreadMap::Iterations::kRow) {
            byte_pointer += params_.increment_row;
          }
        }

        if (group + 1 < ThreadMap::Iterations::kGroup) {
          byte_pointer += params_.increment_group;
        }
      }

      if (cluster + 1 < ThreadMap::Iterations::kCluster) {
        byte_pointer += params_.increment_cluster;
      }
    }
  }

  /// Stores a fragment on the diagonal of a symmetric kernel to memory 
  CUTLASS_DEVICE
  void store_symmetric_with_byte_offset(Fragment const &frag, int64_t byte_offset) {
    uint8_t *byte_pointer = byte_pointer_;
    AccessType const *frag_ptr = reinterpret_cast<AccessType const *>(&frag);

    bool isLowerMode = (fill_mode == cutlass::FillMode::kLower) ? true : false;

    CUTLASS_PRAGMA_UNROLL
    for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {

      CUTLASS_PRAGMA_UNROLL
      for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {

        CUTLASS_PRAGMA_UNROLL
        for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {

          int frag_row_idx = 
            (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));

          int row_offset = row * ThreadMap::Delta::kRow 
            + group * ThreadMap::Delta::kGroup 
            + cluster * ThreadMap::Delta::kCluster;

          bool row_guard = ((row_offset + thread_start_row_) < extent_row_);

          AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);

          // Offset of row from beginning of the matrix per thread
          size_t row_start_offset = (size_t)memory_pointer - matrix_start_addr;

          // Absolute row index
          int row_index = int(row_start_offset/params_.stride);

          CUTLASS_PRAGMA_UNROLL
          for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {

            bool guard = row_guard && mask_.predicates[column];

            // Offset of column from beginning of row per thread     
            size_t col_start_offset = row_start_offset + 
                        (column * ThreadMap::Delta::kColumn / kElementsPerAccess) * sizeof(AccessType);

            // Absolute column index
            size_t col_index = (col_start_offset%params_.stride)/sizeof(AccessType);

            guard = guard && ( (isLowerMode && row_index >= col_index) ||
                               (!isLowerMode && row_index <= col_index) );

            // The imaginary parts of the diagonal elements of a complex element are assumed and set to zero
            if (guard && kBlasMode == BlasMode::kHermitian && cutlass::is_complex<Element>::value) {

              AccessType *frag_ptr_modify = const_cast<AccessType *>(frag_ptr);
              Element *scalar_ptr = reinterpret_cast<Element *>(frag_ptr_modify);

              if (row_index == col_index) {
                scalar_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column] = 
                  real(scalar_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column]);
              }
            }

            cutlass::arch::global_store<AccessType, sizeof(AccessType)>(
                frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
                         column],
                (void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
                                        kElementsPerAccess],
                guard);
          }

          if (row + 1 < ThreadMap::Iterations::kRow) {
            byte_pointer += params_.increment_row;
          }
        }

        if (group + 1 < ThreadMap::Iterations::kGroup) {
          byte_pointer += params_.increment_group;
        }
      }

      if (cluster + 1 < ThreadMap::Iterations::kCluster) {
        byte_pointer += params_.increment_cluster;
      }
    }
  }

  /// Stores a fragment to memory
  CUTLASS_DEVICE
  void store(Fragment const &frag) {
    
    if (fill_mode == cutlass::FillMode::kNone) {
      store_with_byte_offset(frag, 0);
    }
    else {
      store_symmetric_with_byte_offset(frag, 0); 
    }

  }

  /// Advances to the next position to load or store
  CUTLASS_HOST_DEVICE
  PredicatedTileIteratorBlas3 &operator++() {

    ++state_[0];
    byte_pointer_ += params_.advance_row;
    thread_start_row_ += ThreadMap::Shape::kRow;
    
    if (state_[0] == ThreadMap::Count::kRow) {

      state_[0] = 0;
      ++state_[1];
      byte_pointer_ += params_.advance_group;

      thread_start_row_ += (ThreadMap::Shape::kGroup - 1) * 
        ThreadMap::Shape::kRow * ThreadMap::Count::kRow;

      if (state_[1] == ThreadMap::Count::kGroup) {

        state_[1] = 0;
        ++state_[2];
        byte_pointer_ += params_.advance_cluster;

        thread_start_row_ += ThreadMap::Count::kGroup * 
          ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow;

        if (state_[2] == ThreadMap::Count::kCluster) {
          state_[2] = 0;
          byte_pointer_ += params_.advance_tile;
        }
      }
    }

    return *this;
  }

  ///< Efficiently disables all accesses guarded by mask
  CUTLASS_DEVICE void clear_mask() {
    mask_.clear();
  }

  ///< Efficiently enables all accesses guarded by mask
  CUTLASS_DEVICE void enable_mask() {
    mask_.enable();
  }

  ///< Sets the mask
  CUTLASS_DEVICE void get_mask(Mask &mask) {
    mask = mask_;
  }

  ///< Sets the mask
  CUTLASS_DEVICE void set_mask(Mask const &mask) {
    mask_ = mask;
  }
};

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

} // namespace threadblock
} // namespace epilogue
} // namespace cutlass

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