cub/block/block_load.cuh

File members: cub/block/block_load.cuh

/******************************************************************************
 * Copyright (c) 2011, Duane Merrill.  All rights reserved.
 * Copyright (c) 2011-2016, 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/block/block_exchange.cuh>
#include <cub/iterator/cache_modified_input_iterator.cuh>
#include <cub/util_ptx.cuh>
#include <cub/util_type.cuh>

CUB_NAMESPACE_BEGIN

template <typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void
LoadDirectBlocked(int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
{
// Load directly in thread-blocked order
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = block_src_it[linear_tid * ITEMS_PER_THREAD + i];
  }
}

template <typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void LoadDirectBlocked(
  int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
{
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    const auto src_pos = linear_tid * ITEMS_PER_THREAD + i;
    if (src_pos < block_items_end)
    {
      dst_items[i] = block_src_it[src_pos];
    }
  }
}

template <typename T, typename DefaultT, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void LoadDirectBlocked(
  int linear_tid,
  RandomAccessIterator block_src_it,
  T (&dst_items)[ITEMS_PER_THREAD],
  int block_items_end,
  DefaultT oob_default)
{
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = oob_default;
  }

  LoadDirectBlocked(linear_tid, block_src_it, dst_items, block_items_end);
}

#ifndef _CCCL_DOXYGEN_INVOKED // Do not document

template <CacheLoadModifier MODIFIER, typename T, int ITEMS_PER_THREAD>
_CCCL_DEVICE _CCCL_FORCEINLINE void
InternalLoadDirectBlockedVectorized(int linear_tid, const T* block_src_ptr, T (&dst_items)[ITEMS_PER_THREAD])
{
  // Find biggest memory access word that T is a whole multiple of
  using device_word_t = typename UnitWord<T>::DeviceWord;
  _CCCL_DIAG_PUSH
#  if _CCCL_COMPILER(CLANG, >=, 10)
  _CCCL_DIAG_SUPPRESS_CLANG("-Wsizeof-array-div")
#  endif // _CCCL_COMPILER(CLANG, >=, 10)
  constexpr int total_words = static_cast<int>(sizeof(dst_items) / sizeof(device_word_t));
  _CCCL_DIAG_POP
  constexpr int vector_size        = (total_words % 4 == 0) ? 4 : (total_words % 2 == 0) ? 2 : 1;
  constexpr int vectors_per_thread = total_words / vector_size;
  using vector_t                   = typename CubVector<device_word_t, vector_size>::Type;

  // Load into an array of vectors in thread-blocked order
  vector_t vec_items[vectors_per_thread];
  const vector_t* vec_ptr = reinterpret_cast<const vector_t*>(block_src_ptr) + linear_tid * vectors_per_thread;
#  pragma unroll
  for (int i = 0; i < vectors_per_thread; i++)
  {
    vec_items[i] = ThreadLoad<MODIFIER>(vec_ptr + i);
  }

// Copy to destination
#  pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = *(reinterpret_cast<T*>(vec_items) + i);
  }
}

#endif // _CCCL_DOXYGEN_INVOKED

template <typename T, int ITEMS_PER_THREAD>
_CCCL_DEVICE _CCCL_FORCEINLINE void
LoadDirectBlockedVectorized(int linear_tid, T* block_src_ptr, T (&dst_items)[ITEMS_PER_THREAD])
{
  InternalLoadDirectBlockedVectorized<LOAD_DEFAULT>(linear_tid, block_src_ptr, dst_items);
}

template <int BLOCK_THREADS, typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void
LoadDirectStriped(int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
{
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = block_src_it[linear_tid + i * BLOCK_THREADS];
  }
}

namespace detail
{
template <int BLOCK_THREADS, typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator, typename TransformOpT>
_CCCL_DEVICE _CCCL_FORCEINLINE void load_transform_direct_striped(
  int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], TransformOpT transform_op)
{
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = transform_op(block_src_it[linear_tid + i * BLOCK_THREADS]);
  }
}
} // namespace detail

template <int BLOCK_THREADS, typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void LoadDirectStriped(
  int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
{
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    const auto src_pos = linear_tid + i * BLOCK_THREADS;
    if (src_pos < block_items_end)
    {
      dst_items[i] = block_src_it[src_pos];
    }
  }
}

template <int BLOCK_THREADS, typename T, typename DefaultT, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void LoadDirectStriped(
  int linear_tid,
  RandomAccessIterator block_src_it,
  T (&dst_items)[ITEMS_PER_THREAD],
  int block_items_end,
  DefaultT oob_default)
{
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = oob_default;
  }

  LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items, block_items_end);
}

template <typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void
LoadDirectWarpStriped(int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
{
  const int tid         = linear_tid & (CUB_PTX_WARP_THREADS - 1);
  const int wid         = linear_tid >> CUB_PTX_LOG_WARP_THREADS;
  const int warp_offset = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;

// Load directly in warp-striped order
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    new (&dst_items[i]) T(block_src_it[warp_offset + tid + (i * CUB_PTX_WARP_THREADS)]);
  }
}

template <typename T, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void LoadDirectWarpStriped(
  int linear_tid, RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
{
  const int tid         = linear_tid & (CUB_PTX_WARP_THREADS - 1);
  const int wid         = linear_tid >> CUB_PTX_LOG_WARP_THREADS;
  const int warp_offset = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;

// Load directly in warp-striped order
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    const auto src_pos = warp_offset + tid + (i * CUB_PTX_WARP_THREADS);
    if (src_pos < block_items_end)
    {
      new (&dst_items[i]) T(block_src_it[src_pos]);
    }
  }
}

template <typename T, typename DefaultT, int ITEMS_PER_THREAD, typename RandomAccessIterator>
_CCCL_DEVICE _CCCL_FORCEINLINE void LoadDirectWarpStriped(
  int linear_tid,
  RandomAccessIterator block_src_it,
  T (&dst_items)[ITEMS_PER_THREAD],
  int block_items_end,
  DefaultT oob_default)
{
// Load directly in warp-striped order
#pragma unroll
  for (int i = 0; i < ITEMS_PER_THREAD; i++)
  {
    dst_items[i] = oob_default;
  }

  LoadDirectWarpStriped(linear_tid, block_src_it, dst_items, block_items_end);
}

enum BlockLoadAlgorithm
{
  BLOCK_LOAD_DIRECT,

  BLOCK_LOAD_STRIPED,

  BLOCK_LOAD_VECTORIZE,

  BLOCK_LOAD_TRANSPOSE,

  BLOCK_LOAD_WARP_TRANSPOSE,

  BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED,
};

// The data type to read into (which must be convertible from the input iterator's value type).
template <typename T,
          int BLOCK_DIM_X,
          int ITEMS_PER_THREAD,
          BlockLoadAlgorithm ALGORITHM = BLOCK_LOAD_DIRECT,
          int BLOCK_DIM_Y              = 1,
          int BLOCK_DIM_Z              = 1,
          int LEGACY_PTX_ARCH          = 0>
class BlockLoad
{
  static constexpr int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z; // total threads in the block

  template <BlockLoadAlgorithm _POLICY, int DUMMY>
  struct LoadInternal; // helper to dispatch the load algorithm

  template <int DUMMY>
  struct LoadInternal<BLOCK_LOAD_DIRECT, DUMMY>
  {
    using TempStorage = NullType;
    int linear_tid;

    _CCCL_DEVICE _CCCL_FORCEINLINE LoadInternal(TempStorage& /*temp_storage*/, int linear_tid)
        : linear_tid(linear_tid)
    {}

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      LoadDirectBlocked(linear_tid, block_src_it, dst_items);
    }

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
    {
      LoadDirectBlocked(linear_tid, block_src_it, dst_items, block_items_end);
    }

    template <typename RandomAccessIterator, typename DefaultT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
    {
      LoadDirectBlocked(linear_tid, block_src_it, dst_items, block_items_end, oob_default);
    }
  };

  template <int DUMMY>
  struct LoadInternal<BLOCK_LOAD_STRIPED, DUMMY>
  {
    using TempStorage = NullType;
    int linear_tid;

    _CCCL_DEVICE _CCCL_FORCEINLINE LoadInternal(TempStorage& /*temp_storage*/, int linear_tid)
        : linear_tid(linear_tid)
    {}

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items);
    }

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
    {
      LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items, block_items_end);
    }

    template <typename RandomAccessIterator, typename DefaultT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
    {
      LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items, block_items_end, oob_default);
    }
  };

  template <int DUMMY>
  struct LoadInternal<BLOCK_LOAD_VECTORIZE, DUMMY>
  {
    using TempStorage = NullType;
    int linear_tid;

    _CCCL_DEVICE _CCCL_FORCEINLINE LoadInternal(TempStorage& /*temp_storage*/, int linear_tid)
        : linear_tid(linear_tid)
    {}

    // attempts vectorization (pointer)
    template <typename>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(const T* block_ptr, T (&dst_items)[ITEMS_PER_THREAD])
    {
      InternalLoadDirectBlockedVectorized<LOAD_DEFAULT>(linear_tid, block_ptr, dst_items);
    }

    // any other iterator, no vectorization
    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      LoadDirectBlocked(linear_tid, block_src_it, dst_items);
    }

    // attempts vectorization (cache modified iterator)
    template <CacheLoadModifier MODIFIER, typename ValueType, typename OffsetT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(CacheModifiedInputIterator<MODIFIER, ValueType, OffsetT> block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      InternalLoadDirectBlockedVectorized<MODIFIER>(linear_tid, block_src_it.ptr, dst_items);
    }

    // skips vectorization
    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
    {
      LoadDirectBlocked(linear_tid, block_src_it, dst_items, block_items_end);
    }

    // skips vectorization
    template <typename RandomAccessIterator, typename DefaultT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
    {
      LoadDirectBlocked(linear_tid, block_src_it, dst_items, block_items_end, oob_default);
    }
  };

  template <int DUMMY>
  struct LoadInternal<BLOCK_LOAD_TRANSPOSE, DUMMY>
  {
    using BlockExchange = BlockExchange<T, BLOCK_DIM_X, ITEMS_PER_THREAD, false, BLOCK_DIM_Y, BLOCK_DIM_Z>;
    using _TempStorage  = typename BlockExchange::TempStorage;
    using TempStorage   = Uninitialized<_TempStorage>;

    _TempStorage& temp_storage;
    int linear_tid;

    _CCCL_DEVICE _CCCL_FORCEINLINE LoadInternal(TempStorage& temp_storage, int linear_tid)
        : temp_storage(temp_storage.Alias())
        , linear_tid(linear_tid)
    {}

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items);
      BlockExchange(temp_storage).StripedToBlocked(dst_items, dst_items);
    }

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
    {
      LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items, block_items_end);
      BlockExchange(temp_storage).StripedToBlocked(dst_items, dst_items);
    }

    template <typename RandomAccessIterator, typename DefaultT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
    {
      LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_src_it, dst_items, block_items_end, oob_default);
      BlockExchange(temp_storage).StripedToBlocked(dst_items, dst_items);
    }
  };

  template <int DUMMY>
  struct LoadInternal<BLOCK_LOAD_WARP_TRANSPOSE, DUMMY>
  {
    static constexpr int WARP_THREADS = CUB_WARP_THREADS(0);
    static_assert(BLOCK_THREADS % WARP_THREADS == 0, "BLOCK_THREADS must be a multiple of WARP_THREADS");

    using BlockExchange = BlockExchange<T, BLOCK_DIM_X, ITEMS_PER_THREAD, false, BLOCK_DIM_Y, BLOCK_DIM_Z>;
    using _TempStorage  = typename BlockExchange::TempStorage;
    using TempStorage   = Uninitialized<_TempStorage>;

    _TempStorage& temp_storage;
    int linear_tid;

    _CCCL_DEVICE _CCCL_FORCEINLINE LoadInternal(TempStorage& temp_storage, int linear_tid)
        : temp_storage(temp_storage.Alias())
        , linear_tid(linear_tid)
    {}

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      LoadDirectWarpStriped(linear_tid, block_src_it, dst_items);
      BlockExchange(temp_storage).WarpStripedToBlocked(dst_items, dst_items);
    }

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
    {
      LoadDirectWarpStriped(linear_tid, block_src_it, dst_items, block_items_end);
      BlockExchange(temp_storage).WarpStripedToBlocked(dst_items, dst_items);
    }

    template <typename RandomAccessIterator, typename DefaultT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
    {
      LoadDirectWarpStriped(linear_tid, block_src_it, dst_items, block_items_end, oob_default);
      BlockExchange(temp_storage).WarpStripedToBlocked(dst_items, dst_items);
    }
  };

  template <int DUMMY>
  struct LoadInternal<BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED, DUMMY>
  {
    static constexpr int WARP_THREADS = CUB_WARP_THREADS(0);
    static_assert(BLOCK_THREADS % WARP_THREADS == 0, "BLOCK_THREADS must be a multiple of WARP_THREADS");

    using BlockExchange = BlockExchange<T, BLOCK_DIM_X, ITEMS_PER_THREAD, true, BLOCK_DIM_Y, BLOCK_DIM_Z>;
    using _TempStorage  = typename BlockExchange::TempStorage;
    using TempStorage   = Uninitialized<_TempStorage>;

    _TempStorage& temp_storage;
    int linear_tid;

    _CCCL_DEVICE _CCCL_FORCEINLINE LoadInternal(TempStorage& temp_storage, int linear_tid)
        : temp_storage(temp_storage.Alias())
        , linear_tid(linear_tid)
    {}

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
    {
      LoadDirectWarpStriped(linear_tid, block_src_it, dst_items);
      BlockExchange(temp_storage).WarpStripedToBlocked(dst_items, dst_items);
    }

    template <typename RandomAccessIterator>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
    {
      LoadDirectWarpStriped(linear_tid, block_src_it, dst_items, block_items_end);
      BlockExchange(temp_storage).WarpStripedToBlocked(dst_items, dst_items);
    }

    template <typename RandomAccessIterator, typename DefaultT>
    _CCCL_DEVICE _CCCL_FORCEINLINE void
    Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
    {
      LoadDirectWarpStriped(linear_tid, block_src_it, dst_items, block_items_end, oob_default);
      BlockExchange(temp_storage).WarpStripedToBlocked(dst_items, dst_items);
    }
  };

  using InternalLoad = LoadInternal<ALGORITHM, 0>; // load implementation to use
  using _TempStorage = typename InternalLoad::TempStorage;

  // Internal storage allocator
  _CCCL_DEVICE _CCCL_FORCEINLINE _TempStorage& PrivateStorage()
  {
    __shared__ _TempStorage private_storage;
    return private_storage;
  }

  _TempStorage& temp_storage;
  int linear_tid;

public:
  using TempStorage = Uninitialized<_TempStorage>;

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

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

  template <typename RandomAccessIterator>
  _CCCL_DEVICE _CCCL_FORCEINLINE void Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD])
  {
    InternalLoad(temp_storage, linear_tid).Load(block_src_it, dst_items);
  }

  template <typename RandomAccessIterator>
  _CCCL_DEVICE _CCCL_FORCEINLINE void
  Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end)
  {
    InternalLoad(temp_storage, linear_tid).Load(block_src_it, dst_items, block_items_end);
  }

  template <typename RandomAccessIterator, typename DefaultT>
  _CCCL_DEVICE _CCCL_FORCEINLINE void
  Load(RandomAccessIterator block_src_it, T (&dst_items)[ITEMS_PER_THREAD], int block_items_end, DefaultT oob_default)
  {
    InternalLoad(temp_storage, linear_tid).Load(block_src_it, dst_items, block_items_end, oob_default);
  }

};

template <class Policy, class It, class T = cub::detail::value_t<It>>
struct BlockLoadType
{
  using type = cub::BlockLoad<T, Policy::BLOCK_THREADS, Policy::ITEMS_PER_THREAD, Policy::LOAD_ALGORITHM>;
};

CUB_NAMESPACE_END