block_load.cuh

Go to the documentation of this file.

/******************************************************************************
 * 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 <iterator>

#include "block_exchange.cuh"
#include "../iterator/cache_modified_input_iterator.cuh"
#include "../util_ptx.cuh"
#include "../util_macro.cuh"
#include "../util_type.cuh"
#include "../util_namespace.cuh"

CUB_NS_PREFIX

namespace cub {

/******************************************************************/


template <
    typename        InputT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectBlocked(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD]) 
{
    InputIteratorT thread_itr = block_itr + (linear_tid * ITEMS_PER_THREAD);

    // Load directly in thread-blocked order
    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
    {
        items[ITEM] = thread_itr[ITEM];
    }
}


template <
    typename        InputT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectBlocked(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD], 
    int             valid_items)                
{
    InputIteratorT thread_itr = block_itr + (linear_tid * ITEMS_PER_THREAD);

    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
    {
        if ((linear_tid * ITEMS_PER_THREAD) + ITEM < valid_items)
        {
            items[ITEM] = thread_itr[ITEM];
        }
    }
}


template <
    typename        InputT,
    typename        DefaultT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectBlocked(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD], 
    int             valid_items,                
    DefaultT        oob_default)                
{
    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
        items[ITEM] = oob_default;

    LoadDirectBlocked(linear_tid, block_itr, items, valid_items);
}


#ifndef DOXYGEN_SHOULD_SKIP_THIS    // Do not document

template <
    CacheLoadModifier   MODIFIER,
    typename            T,
    int                 ITEMS_PER_THREAD>
__device__ __forceinline__ void InternalLoadDirectBlockedVectorized(
    int    linear_tid,                 
    T      *block_ptr,                 
    T      (&items)[ITEMS_PER_THREAD]) 
{
    // Biggest memory access word that T is a whole multiple of
    typedef typename UnitWord<T>::DeviceWord DeviceWord;

    enum
    {
        TOTAL_WORDS = sizeof(items) / sizeof(DeviceWord),

        VECTOR_SIZE = (TOTAL_WORDS % 4 == 0) ?
            4 :
            (TOTAL_WORDS % 2 == 0) ?
                2 :
                1,

        VECTORS_PER_THREAD = TOTAL_WORDS / VECTOR_SIZE,
    };

    // Vector type
    typedef typename CubVector<DeviceWord, VECTOR_SIZE>::Type Vector;

    // Vector items
    Vector vec_items[VECTORS_PER_THREAD];

    // Aliased input ptr
    Vector* vec_ptr = reinterpret_cast<Vector*>(block_ptr) + (linear_tid * VECTORS_PER_THREAD);

    // Load directly in thread-blocked order
    #pragma unroll
    for (int ITEM = 0; ITEM < VECTORS_PER_THREAD; ITEM++)
    {
        vec_items[ITEM] = ThreadLoad<MODIFIER>(vec_ptr + ITEM);
    }

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

#endif // DOXYGEN_SHOULD_SKIP_THIS


template <
    typename        T,
    int             ITEMS_PER_THREAD>
__device__ __forceinline__ void LoadDirectBlockedVectorized(
    int linear_tid,                 
    T   *block_ptr,                 
    T   (&items)[ITEMS_PER_THREAD]) 
{
    InternalLoadDirectBlockedVectorized<LOAD_DEFAULT>(linear_tid, block_ptr, items);
}


/******************************************************************/


template <
    int             BLOCK_THREADS,
    typename        InputT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectStriped(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD]) 
{
    InputIteratorT thread_itr = block_itr + linear_tid;

    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
    {
        items[ITEM] = thread_itr[ITEM * BLOCK_THREADS];
    }
}


template <
    int             BLOCK_THREADS,
    typename        InputT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectStriped(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD], 
    int             valid_items)                
{
    InputIteratorT thread_itr = block_itr + linear_tid;

    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
    {
        if (linear_tid + (ITEM * BLOCK_THREADS) < valid_items)
        {
            items[ITEM] = thread_itr[ITEM * BLOCK_THREADS];
        }
    }
}


template <
    int             BLOCK_THREADS,
    typename        InputT,
    typename        DefaultT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectStriped(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD], 
    int             valid_items,                
    DefaultT        oob_default)                
{
    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
        items[ITEM] = oob_default;

    LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_itr, items, valid_items);
}



/******************************************************************/


template <
    typename        InputT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectWarpStriped(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD]) 
{
    int tid                = linear_tid & (CUB_PTX_WARP_THREADS - 1);
    int wid                = linear_tid >> CUB_PTX_LOG_WARP_THREADS;
    int warp_offset        = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;

    InputIteratorT thread_itr = block_itr + warp_offset + tid ;

    // Load directly in warp-striped order
    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
    {
        items[ITEM] = thread_itr[(ITEM * CUB_PTX_WARP_THREADS)];
    }
}


template <
    typename        InputT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectWarpStriped(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD], 
    int             valid_items)                
{
    int tid                = linear_tid & (CUB_PTX_WARP_THREADS - 1);
    int wid                = linear_tid >> CUB_PTX_LOG_WARP_THREADS;
    int warp_offset        = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;

    InputIteratorT thread_itr = block_itr + warp_offset + tid ;

    // Load directly in warp-striped order
    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
    {
        if (warp_offset + tid + (ITEM * CUB_PTX_WARP_THREADS) < valid_items)
        {
            items[ITEM] = thread_itr[(ITEM * CUB_PTX_WARP_THREADS)];
        }
    }
}


template <
    typename        InputT,
    typename        DefaultT,
    int             ITEMS_PER_THREAD,
    typename        InputIteratorT>
__device__ __forceinline__ void LoadDirectWarpStriped(
    int             linear_tid,                 
    InputIteratorT  block_itr,                  
    InputT          (&items)[ITEMS_PER_THREAD], 
    int             valid_items,                
    DefaultT        oob_default)                
{
    // Load directly in warp-striped order
    #pragma unroll
    for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)
        items[ITEM] = oob_default;

    LoadDirectWarpStriped(linear_tid, block_itr, items, valid_items);
}



       // end group UtilIo



//-----------------------------------------------------------------------------
// Generic BlockLoad abstraction
//-----------------------------------------------------------------------------

enum BlockLoadAlgorithm
{
    BLOCK_LOAD_DIRECT,

    BLOCK_LOAD_VECTORIZE,

    BLOCK_LOAD_TRANSPOSE,


    BLOCK_LOAD_WARP_TRANSPOSE,


    BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED,
};


template <
    typename            InputT,
    int                 BLOCK_DIM_X,
    int                 ITEMS_PER_THREAD,
    BlockLoadAlgorithm  ALGORITHM           = BLOCK_LOAD_DIRECT,
    int                 BLOCK_DIM_Y         = 1,
    int                 BLOCK_DIM_Z         = 1,
    int                 PTX_ARCH            = CUB_PTX_ARCH>
class BlockLoad
{
private:

    /******************************************************************************
     * Constants and typed definitions
     ******************************************************************************/

    enum
    {
        BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,
    };


    /******************************************************************************
     * Algorithmic variants
     ******************************************************************************/

    template <BlockLoadAlgorithm _POLICY, int DUMMY>
    struct LoadInternal;


    template <int DUMMY>
    struct LoadInternal<BLOCK_LOAD_DIRECT, DUMMY>
    {
        typedef NullType TempStorage;

        int linear_tid;

        __device__ __forceinline__ LoadInternal(
            TempStorage &/*temp_storage*/,
            int linear_tid)
        :
            linear_tid(linear_tid)
        {}

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD])     
        {
            LoadDirectBlocked(linear_tid, block_itr, items);
        }

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items)                    
        {
            LoadDirectBlocked(linear_tid, block_itr, items, valid_items);
        }

        template <typename InputIteratorT, typename DefaultT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items,                    
            DefaultT        oob_default)                    
        {
            LoadDirectBlocked(linear_tid, block_itr, items, valid_items, oob_default);
        }

    };


    template <int DUMMY>
    struct LoadInternal<BLOCK_LOAD_VECTORIZE, DUMMY>
    {
        typedef NullType TempStorage;

        int linear_tid;

        __device__ __forceinline__ LoadInternal(
            TempStorage &/*temp_storage*/,
            int linear_tid)
        :
            linear_tid(linear_tid)
        {}

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputT               *block_ptr,                     
            InputT               (&items)[ITEMS_PER_THREAD])     
        {
            InternalLoadDirectBlockedVectorized<LOAD_DEFAULT>(linear_tid, block_ptr, items);
        }

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            const InputT         *block_ptr,                     
            InputT               (&items)[ITEMS_PER_THREAD])     
        {
            InternalLoadDirectBlockedVectorized<LOAD_DEFAULT>(linear_tid, block_ptr, items);
        }

        template <
            CacheLoadModifier   MODIFIER,
            typename            ValueType,
            typename            OffsetT>
        __device__ __forceinline__ void Load(
            CacheModifiedInputIterator<MODIFIER, ValueType, OffsetT>    block_itr,                      
            InputT                                                     (&items)[ITEMS_PER_THREAD])     
        {
            InternalLoadDirectBlockedVectorized<MODIFIER>(linear_tid, block_itr.ptr, items);
        }

        template <typename _InputIteratorT>
        __device__ __forceinline__ void Load(
            _InputIteratorT   block_itr,                    
            InputT           (&items)[ITEMS_PER_THREAD])   
        {
            LoadDirectBlocked(linear_tid, block_itr, items);
        }

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items)                    
        {
            LoadDirectBlocked(linear_tid, block_itr, items, valid_items);
        }

        template <typename InputIteratorT, typename DefaultT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items,                    
            DefaultT          oob_default)                    
        {
            LoadDirectBlocked(linear_tid, block_itr, items, valid_items, oob_default);
        }

    };


    template <int DUMMY>
    struct LoadInternal<BLOCK_LOAD_TRANSPOSE, DUMMY>
    {
        // BlockExchange utility type for keys
        typedef BlockExchange<InputT, BLOCK_DIM_X, ITEMS_PER_THREAD, false, BLOCK_DIM_Y, BLOCK_DIM_Z, PTX_ARCH> BlockExchange;

        struct _TempStorage : BlockExchange::TempStorage
        {};

        struct TempStorage : Uninitialized<_TempStorage> {};

        _TempStorage &temp_storage;

        int linear_tid;

        __device__ __forceinline__ LoadInternal(
            TempStorage &temp_storage,
            int linear_tid)
        :
            temp_storage(temp_storage.Alias()),
            linear_tid(linear_tid)
        {}

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD])     
        {
            LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_itr, items);
            BlockExchange(temp_storage).StripedToBlocked(items, items);
        }

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items)                    
        {
            LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_itr, items, valid_items);
            BlockExchange(temp_storage).StripedToBlocked(items, items);
        }

        template <typename InputIteratorT, typename DefaultT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items,                    
            DefaultT        oob_default)                    
        {
            LoadDirectStriped<BLOCK_THREADS>(linear_tid, block_itr, items, valid_items, oob_default);
            BlockExchange(temp_storage).StripedToBlocked(items, items);
        }

    };


    template <int DUMMY>
    struct LoadInternal<BLOCK_LOAD_WARP_TRANSPOSE, DUMMY>
    {
        enum
        {
            WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH)
        };

        // Assert BLOCK_THREADS must be a multiple of WARP_THREADS
        CUB_STATIC_ASSERT((BLOCK_THREADS % WARP_THREADS == 0), "BLOCK_THREADS must be a multiple of WARP_THREADS");

        // BlockExchange utility type for keys
        typedef BlockExchange<InputT, BLOCK_DIM_X, ITEMS_PER_THREAD, false, BLOCK_DIM_Y, BLOCK_DIM_Z, PTX_ARCH> BlockExchange;

        struct _TempStorage : BlockExchange::TempStorage
        {};

        struct TempStorage : Uninitialized<_TempStorage> {};

        _TempStorage &temp_storage;

        int linear_tid;

        __device__ __forceinline__ LoadInternal(
            TempStorage &temp_storage,
            int linear_tid)
        :
            temp_storage(temp_storage.Alias()),
            linear_tid(linear_tid)
        {}

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD])     
        {
            LoadDirectWarpStriped(linear_tid, block_itr, items);
            BlockExchange(temp_storage).WarpStripedToBlocked(items, items);
        }

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items)                    
        {
            LoadDirectWarpStriped(linear_tid, block_itr, items, valid_items);
            BlockExchange(temp_storage).WarpStripedToBlocked(items, items);
        }


        template <typename InputIteratorT, typename DefaultT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items,                    
            DefaultT        oob_default)                    
        {
            LoadDirectWarpStriped(linear_tid, block_itr, items, valid_items, oob_default);
            BlockExchange(temp_storage).WarpStripedToBlocked(items, items);
        }
    };


    template <int DUMMY>
    struct LoadInternal<BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED, DUMMY>
    {
        enum
        {
            WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH)
        };

        // Assert BLOCK_THREADS must be a multiple of WARP_THREADS
        CUB_STATIC_ASSERT((BLOCK_THREADS % WARP_THREADS == 0), "BLOCK_THREADS must be a multiple of WARP_THREADS");

        // BlockExchange utility type for keys
        typedef BlockExchange<InputT, BLOCK_DIM_X, ITEMS_PER_THREAD, true, BLOCK_DIM_Y, BLOCK_DIM_Z, PTX_ARCH> BlockExchange;

        struct _TempStorage : BlockExchange::TempStorage
        {};

        struct TempStorage : Uninitialized<_TempStorage> {};

        _TempStorage &temp_storage;

        int linear_tid;

        __device__ __forceinline__ LoadInternal(
            TempStorage &temp_storage,
            int linear_tid)
        :
            temp_storage(temp_storage.Alias()),
            linear_tid(linear_tid)
        {}

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD])     
        {
            LoadDirectWarpStriped(linear_tid, block_itr, items);
            BlockExchange(temp_storage).WarpStripedToBlocked(items, items);
        }

        template <typename InputIteratorT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items)                    
        {
            LoadDirectWarpStriped(linear_tid, block_itr, items, valid_items);
            BlockExchange(temp_storage).WarpStripedToBlocked(items, items);
        }


        template <typename InputIteratorT, typename DefaultT>
        __device__ __forceinline__ void Load(
            InputIteratorT  block_itr,                      
            InputT          (&items)[ITEMS_PER_THREAD],     
            int             valid_items,                    
            DefaultT        oob_default)                    
        {
            LoadDirectWarpStriped(linear_tid, block_itr, items, valid_items, oob_default);
            BlockExchange(temp_storage).WarpStripedToBlocked(items, items);
        }
    };


    /******************************************************************************
     * Type definitions
     ******************************************************************************/

    typedef LoadInternal<ALGORITHM, 0> InternalLoad;


    typedef typename InternalLoad::TempStorage _TempStorage;


    /******************************************************************************
     * Utility methods
     ******************************************************************************/

    __device__ __forceinline__ _TempStorage& PrivateStorage()
    {
        __shared__ _TempStorage private_storage;
        return private_storage;
    }


    /******************************************************************************
     * Thread fields
     ******************************************************************************/

    _TempStorage &temp_storage;

    int linear_tid;

public:

    struct TempStorage : Uninitialized<_TempStorage> {};


    /******************************************************************/

    __device__ __forceinline__ BlockLoad()
    :
        temp_storage(PrivateStorage()),
        linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))
    {}


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




    /******************************************************************/


    template <typename InputIteratorT>
    __device__ __forceinline__ void Load(
        InputIteratorT  block_itr,                  
        InputT          (&items)[ITEMS_PER_THREAD]) 
    {
        InternalLoad(temp_storage, linear_tid).Load(block_itr, items);
    }


    template <typename InputIteratorT>
    __device__ __forceinline__ void Load(
        InputIteratorT  block_itr,                  
        InputT          (&items)[ITEMS_PER_THREAD], 
        int             valid_items)                
    {
        InternalLoad(temp_storage, linear_tid).Load(block_itr, items, valid_items);
    }


    template <typename InputIteratorT, typename DefaultT>
    __device__ __forceinline__ void Load(
        InputIteratorT  block_itr,                  
        InputT          (&items)[ITEMS_PER_THREAD], 
        int             valid_items,                
        DefaultT        oob_default)                
    {
        InternalLoad(temp_storage, linear_tid).Load(block_itr, items, valid_items, oob_default);
    }



};


}               // CUB namespace
CUB_NS_POSTFIX  // Optional outer namespace(s)