view release on metacpan or  search on metacpan

xgboost/cub/cub/warp/warp_reduce.cuh  view on Meta::CPAN

/******************************************************************************
 * 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.
 *
 ******************************************************************************/

/**
 * \file
 * The cub::WarpReduce class provides [<em>collective</em>](index.html#sec0) methods for computing a parallel reduction of items partitioned across a CUDA thread warp.
 */

#pragma once

#include "specializations/warp_reduce_shfl.cuh"
#include "specializations/warp_reduce_smem.cuh"
#include "../thread/thread_operators.cuh"
#include "../util_arch.cuh"
#include "../util_type.cuh"
#include "../util_namespace.cuh"

/// Optional outer namespace(s)
CUB_NS_PREFIX

/// CUB namespace
namespace cub {


/**
 * \addtogroup WarpModule
 * @{
 */

/**
 * \brief The WarpReduce class provides [<em>collective</em>](index.html#sec0) methods for computing a parallel reduction of items partitioned across a CUDA thread warp. ![](warp_reduce_logo.png)
 *
 * \tparam T                        The reduction input/output element type
 * \tparam LOGICAL_WARP_THREADS     <b>[optional]</b> The number of threads per "logical" warp (may be less than the number of hardware warp threads).  Default is the warp size of the targeted CUDA compute-capability (e.g., 32 threads for SM20).
 * \tparam PTX_ARCH                 <b>[optional]</b> \ptxversion
 *
 * \par Overview
 * - A <a href="http://en.wikipedia.org/wiki/Reduce_(higher-order_function)"><em>reduction</em></a> (or <em>fold</em>)
 *   uses a binary combining operator to compute a single aggregate from a list of input elements.
 * - Supports "logical" warps smaller than the physical warp size (e.g., logical warps of 8 threads)
 * - The number of entrant threads must be an multiple of \p LOGICAL_WARP_THREADS
 *
 * \par Performance Considerations
 * - Uses special instructions when applicable (e.g., warp \p SHFL instructions)
 * - Uses synchronization-free communication between warp lanes when applicable
 * - Incurs zero bank conflicts for most types
 * - Computation is slightly more efficient (i.e., having lower instruction overhead) for:
 *     - Summation (<b><em>vs.</em></b> generic reduction)
 *     - The architecture's warp size is a whole multiple of \p LOGICAL_WARP_THREADS
 *
 * \par Simple Examples
 * \warpcollective{WarpReduce}
 * \par
 * The code snippet below illustrates four concurrent warp sum reductions within a block of
 * 128 threads (one per each of the 32-thread warps).
 * \par
 * \code
 * #include <cub/cub.cuh>
 *
 * __global__ void ExampleKernel(...)
 * {
 *     // Specialize WarpReduce for type int
 *     typedef cub::WarpReduce<int> WarpReduce;
 *
 *     // Allocate WarpReduce shared memory for 4 warps
 *     __shared__ typename WarpReduce::TempStorage temp_storage[4];
 *
 *     // Obtain one input item per thread
 *     int thread_data = ...
 *
 *     // Return the warp-wide sums to each lane0 (threads 0, 32, 64, and 96)
 *     int warp_id = threadIdx.x / 32;
 *     int aggregate = WarpReduce(temp_storage[warp_id]).Sum(thread_data);
 *
 * \endcode
 * \par
 * Suppose the set of input \p thread_data across the block of threads is <tt>{0, 1, 2, 3, ..., 127}</tt>.
 * The corresponding output \p aggregate in threads 0, 32, 64, and 96 will \p 496, \p 1520,
 * \p 2544, and \p 3568, respectively (and is undefined in other threads).
 *
 * \par
 * The code snippet below illustrates a single warp sum reduction within a block of
 * 128 threads.
 * \par
 * \code
 * #include <cub/cub.cuh>
 *
 * __global__ void ExampleKernel(...)
 * {
 *     // Specialize WarpReduce for type int
 *     typedef cub::WarpReduce<int> WarpReduce;