cub/block/block_adjacent_difference.cuh

File members: cub/block/block_adjacent_difference.cuh

/******************************************************************************
 * Copyright (c) 2011, Duane Merrill.  All rights reserved.
 * Copyright (c) 2011-2021, 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 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.
 *
 ******************************************************************************/

#pragma once

#include <cub/config.cuh>

#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
#  pragma GCC system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
#  pragma clang system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
#  pragma system_header
#endif // no system header

#include <cub/util_ptx.cuh>
#include <cub/util_type.cuh>

CUB_NAMESPACE_BEGIN

template <typename T, int BLOCK_DIM_X, int BLOCK_DIM_Y = 1, int BLOCK_DIM_Z = 1, int LEGACY_PTX_ARCH = 0>
class BlockAdjacentDifference
{
private:
  static constexpr int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z;

  struct _TempStorage
  {
    T first_items[BLOCK_THREADS];
    T last_items[BLOCK_THREADS];
  };

  _CCCL_DEVICE _CCCL_FORCEINLINE _TempStorage& PrivateStorage()
  {
    __shared__ _TempStorage private_storage;
    return private_storage;
  }

  template <typename FlagOp, bool HAS_PARAM = BinaryOpHasIdxParam<T, FlagOp>::value>
  struct ApplyOp
  {
    // Apply flag operator
    static _CCCL_DEVICE _CCCL_FORCEINLINE T FlagT(FlagOp flag_op, const T& a, const T& b, int idx)
    {
      return flag_op(b, a, idx);
    }
  };

  template <typename FlagOp>
  struct ApplyOp<FlagOp, false>
  {
    // Apply flag operator
    static _CCCL_DEVICE _CCCL_FORCEINLINE T FlagT(FlagOp flag_op, const T& a, const T& b, int /*idx*/)
    {
      return flag_op(b, a);
    }
  };

  struct Iterate
  {
    template <int ITEMS_PER_THREAD, typename FlagT, typename FlagOp>
    static _CCCL_DEVICE _CCCL_FORCEINLINE void FlagHeads(
      int linear_tid,
      FlagT (&flags)[ITEMS_PER_THREAD],
      T (&input)[ITEMS_PER_THREAD],
      T (&preds)[ITEMS_PER_THREAD],
      FlagOp flag_op)
    {
#pragma unroll
      for (int i = 1; i < ITEMS_PER_THREAD; ++i)
      {
        preds[i] = input[i - 1];
        flags[i] = ApplyOp<FlagOp>::FlagT(flag_op, preds[i], input[i], (linear_tid * ITEMS_PER_THREAD) + i);
      }
    }

    template <int ITEMS_PER_THREAD, typename FlagT, typename FlagOp>
    static _CCCL_DEVICE _CCCL_FORCEINLINE void
    FlagTails(int linear_tid, FlagT (&flags)[ITEMS_PER_THREAD], T (&input)[ITEMS_PER_THREAD], FlagOp flag_op)
    {
#pragma unroll
      for (int i = 0; i < ITEMS_PER_THREAD - 1; ++i)
      {
        flags[i] = ApplyOp<FlagOp>::FlagT(flag_op, input[i], input[i + 1], (linear_tid * ITEMS_PER_THREAD) + i + 1);
      }
    }
  };

  _TempStorage& temp_storage;

  unsigned int linear_tid;

public:
  struct TempStorage : Uninitialized<_TempStorage>
  {};

  _CCCL_DEVICE _CCCL_FORCEINLINE BlockAdjacentDifference()
      : temp_storage(PrivateStorage())
      , linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))
  {}

  _CCCL_DEVICE _CCCL_FORCEINLINE BlockAdjacentDifference(TempStorage& temp_storage)
      : temp_storage(temp_storage.Alias())
      , linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))
  {}

  template <int ITEMS_PER_THREAD, typename OutputType, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void
  SubtractLeft(T (&input)[ITEMS_PER_THREAD], OutputType (&output)[ITEMS_PER_THREAD], DifferenceOpT difference_op)
  {
    // Share last item
    temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];

    CTA_SYNC();

#pragma unroll
    for (int item = ITEMS_PER_THREAD - 1; item > 0; item--)
    {
      output[item] = difference_op(input[item], input[item - 1]);
    }

    if (linear_tid == 0)
    {
      output[0] = input[0];
    }
    else
    {
      output[0] = difference_op(input[0], temp_storage.last_items[linear_tid - 1]);
    }
  }

  template <int ITEMS_PER_THREAD, typename OutputT, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void SubtractLeft(
    T (&input)[ITEMS_PER_THREAD],
    OutputT (&output)[ITEMS_PER_THREAD],
    DifferenceOpT difference_op,
    T tile_predecessor_item)
  {
    // Share last item
    temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];

    CTA_SYNC();

#pragma unroll
    for (int item = ITEMS_PER_THREAD - 1; item > 0; item--)
    {
      output[item] = difference_op(input[item], input[item - 1]);
    }

    // Set flag for first thread-item
    if (linear_tid == 0)
    {
      output[0] = difference_op(input[0], tile_predecessor_item);
    }
    else
    {
      output[0] = difference_op(input[0], temp_storage.last_items[linear_tid - 1]);
    }
  }

  template <int ITEMS_PER_THREAD, typename OutputType, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void SubtractLeftPartialTile(
    T (&input)[ITEMS_PER_THREAD], OutputType (&output)[ITEMS_PER_THREAD], DifferenceOpT difference_op, int valid_items)
  {
    // Share last item
    temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];

    CTA_SYNC();

    if ((linear_tid + 1) * ITEMS_PER_THREAD <= valid_items)
    {
#pragma unroll
      for (int item = ITEMS_PER_THREAD - 1; item > 0; item--)
      {
        output[item] = difference_op(input[item], input[item - 1]);
      }
    }
    else
    {
#pragma unroll
      for (int item = ITEMS_PER_THREAD - 1; item > 0; item--)
      {
        const int idx = linear_tid * ITEMS_PER_THREAD + item;

        if (idx < valid_items)
        {
          output[item] = difference_op(input[item], input[item - 1]);
        }
        else
        {
          output[item] = input[item];
        }
      }
    }

    if (linear_tid == 0 || valid_items <= linear_tid * ITEMS_PER_THREAD)
    {
      output[0] = input[0];
    }
    else
    {
      output[0] = difference_op(input[0], temp_storage.last_items[linear_tid - 1]);
    }
  }

  template <int ITEMS_PER_THREAD, typename OutputType, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void SubtractLeftPartialTile(
    T (&input)[ITEMS_PER_THREAD],
    OutputType (&output)[ITEMS_PER_THREAD],
    DifferenceOpT difference_op,
    int valid_items,
    T tile_predecessor_item)
  {
    // Share last item
    temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];

    CTA_SYNC();

    if ((linear_tid + 1) * ITEMS_PER_THREAD <= valid_items)
    {
#pragma unroll
      for (int item = ITEMS_PER_THREAD - 1; item > 0; item--)
      {
        output[item] = difference_op(input[item], input[item - 1]);
      }
    }
    else
    {
#pragma unroll
      for (int item = ITEMS_PER_THREAD - 1; item > 0; item--)
      {
        const int idx = linear_tid * ITEMS_PER_THREAD + item;

        if (idx < valid_items)
        {
          output[item] = difference_op(input[item], input[item - 1]);
        }
        else
        {
          output[item] = input[item];
        }
      }
    }

    if (valid_items <= linear_tid * ITEMS_PER_THREAD)
    {
      output[0] = input[0];
    }
    else if (linear_tid == 0)
    {
      output[0] = difference_op(input[0], tile_predecessor_item);
    }
    else
    {
      output[0] = difference_op(input[0], temp_storage.last_items[linear_tid - 1]);
    }
  }

  template <int ITEMS_PER_THREAD, typename OutputT, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void
  SubtractRight(T (&input)[ITEMS_PER_THREAD], OutputT (&output)[ITEMS_PER_THREAD], DifferenceOpT difference_op)
  {
    // Share first item
    temp_storage.first_items[linear_tid] = input[0];

    CTA_SYNC();

#pragma unroll
    for (int item = 0; item < ITEMS_PER_THREAD - 1; item++)
    {
      output[item] = difference_op(input[item], input[item + 1]);
    }

    if (linear_tid == BLOCK_THREADS - 1)
    {
      output[ITEMS_PER_THREAD - 1] = input[ITEMS_PER_THREAD - 1];
    }
    else
    {
      output[ITEMS_PER_THREAD - 1] =
        difference_op(input[ITEMS_PER_THREAD - 1], temp_storage.first_items[linear_tid + 1]);
    }
  }

  template <int ITEMS_PER_THREAD, typename OutputT, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void SubtractRight(
    T (&input)[ITEMS_PER_THREAD],
    OutputT (&output)[ITEMS_PER_THREAD],
    DifferenceOpT difference_op,
    T tile_successor_item)
  {
    // Share first item
    temp_storage.first_items[linear_tid] = input[0];

    CTA_SYNC();

    // Set flag for last thread-item
    T successor_item = (linear_tid == BLOCK_THREADS - 1)
                       ? tile_successor_item // Last thread
                       : temp_storage.first_items[linear_tid + 1];

#pragma unroll
    for (int item = 0; item < ITEMS_PER_THREAD - 1; item++)
    {
      output[item] = difference_op(input[item], input[item + 1]);
    }

    output[ITEMS_PER_THREAD - 1] = difference_op(input[ITEMS_PER_THREAD - 1], successor_item);
  }

  template <int ITEMS_PER_THREAD, typename OutputT, typename DifferenceOpT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void SubtractRightPartialTile(
    T (&input)[ITEMS_PER_THREAD], OutputT (&output)[ITEMS_PER_THREAD], DifferenceOpT difference_op, int valid_items)
  {
    // Share first item
    temp_storage.first_items[linear_tid] = input[0];

    CTA_SYNC();

    if ((linear_tid + 1) * ITEMS_PER_THREAD < valid_items)
    {
#pragma unroll
      for (int item = 0; item < ITEMS_PER_THREAD - 1; item++)
      {
        output[item] = difference_op(input[item], input[item + 1]);
      }

      output[ITEMS_PER_THREAD - 1] =
        difference_op(input[ITEMS_PER_THREAD - 1], temp_storage.first_items[linear_tid + 1]);
    }
    else
    {
#pragma unroll
      for (int item = 0; item < ITEMS_PER_THREAD; item++)
      {
        const int idx = linear_tid * ITEMS_PER_THREAD + item;

        // Right element of input[valid_items - 1] is out of bounds.
        // According to the API it's copied into output array
        // without modification.
        if (idx < valid_items - 1)
        {
          output[item] = difference_op(input[item], input[item + 1]);
        }
        else
        {
          output[item] = input[item];
        }
      }
    }
  }
};

CUB_NAMESPACE_END