Function naive_kernel

benchmark/benchmark_device_batch_memcpy.cpp:308–340

template<typename OffsetType, int32_t BlockSize>
__launch_bounds__(BlockSize) __global__
void naive_kernel(void** in_ptr, void** out_ptr, const OffsetType* sizes)
{
    using underlying_type              = unsigned char;
    constexpr int32_t items_per_thread = 4;
    constexpr int32_t tile_size        = items_per_thread * BlockSize;

    const int32_t buffer_id = rocprim::flat_block_id();
    auto          in        = reinterpret_cast<underlying_type*>(in_ptr[buffer_id]);
    auto          out       = reinterpret_cast<underlying_type*>(out_ptr[buffer_id]);

    const auto size             = sizes[buffer_id];
    const auto size_in_elements = size / sizeof(underlying_type);
    const auto tiles            = size_in_elements / tile_size;

    auto num_items_to_copy = size;

    for(size_t i = 0; i < tiles; ++i)
    {
        underlying_type data[items_per_thread];
        rocprim::block_load_direct_blocked(rocprim::flat_block_thread_id(),
                                           in,
                                           data,
                                           num_items_to_copy);
        rocprim::block_store_direct_blocked(rocprim::flat_block_thread_id(),
                                            out,
                                            data,
                                            num_items_to_copy);

        in += tile_size;
        out += tile_size;
        num_items_to_copy -= tile_size;
    }
}

template<typename ValueType,
         typename BufferSizeType,