dispatch_scan.cuh

Go to the documentation of this file.

/******************************************************************************
 * Copyright (c) 2011, Duane Merrill.  All rights reserved.
 * Copyright (c) 2011-2018, 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 <stdio.h>
#include <iterator>

#include "../../agent/agent_scan.cuh"
#include "../../thread/thread_operators.cuh"
#include "../../grid/grid_queue.cuh"
#include "../../util_arch.cuh"
#include "../../util_debug.cuh"
#include "../../util_device.cuh"
#include "../../util_namespace.cuh"

CUB_NS_PREFIX

namespace cub {


/******************************************************************************
 * Kernel entry points
 *****************************************************************************/

template <
    typename            ScanTileStateT>     
__global__ void DeviceScanInitKernel(
    ScanTileStateT      tile_state,         
    int                 num_tiles)          
{
    // Initialize tile status
    tile_state.InitializeStatus(num_tiles);
}

template <
    typename                ScanTileStateT,         
    typename                NumSelectedIteratorT>   
__global__ void DeviceCompactInitKernel(
    ScanTileStateT          tile_state,             
    int                     num_tiles,              
    NumSelectedIteratorT    d_num_selected_out)     
{
    // Initialize tile status
    tile_state.InitializeStatus(num_tiles);

    // Initialize d_num_selected_out
    if ((blockIdx.x == 0) && (threadIdx.x == 0))
        *d_num_selected_out = 0;
}


template <
    typename            ScanPolicyT,        
    typename            InputIteratorT,     
    typename            OutputIteratorT,    
    typename            ScanTileStateT,     
    typename            ScanOpT,            
    typename            InitValueT,         
    typename            OffsetT>            
__launch_bounds__ (int(ScanPolicyT::BLOCK_THREADS))
__global__ void DeviceScanKernel(
    InputIteratorT      d_in,               
    OutputIteratorT     d_out,              
    ScanTileStateT      tile_state,         
    int                 start_tile,         
    ScanOpT             scan_op,            
    InitValueT          init_value,         
    OffsetT             num_items)          
{
    // Thread block type for scanning input tiles
    typedef AgentScan<
        ScanPolicyT,
        InputIteratorT,
        OutputIteratorT,
        ScanOpT,
        InitValueT,
        OffsetT> AgentScanT;

    // Shared memory for AgentScan
    __shared__ typename AgentScanT::TempStorage temp_storage;

    // Process tiles
    AgentScanT(temp_storage, d_in, d_out, scan_op, init_value).ConsumeRange(
        num_items,
        tile_state,
        start_tile);
}




/******************************************************************************
 * Dispatch
 ******************************************************************************/


template <
    typename InputIteratorT,     
    typename OutputIteratorT,    
    typename ScanOpT,            
    typename InitValueT,          
    typename OffsetT>            
struct DispatchScan
{
    //---------------------------------------------------------------------
    // Constants and Types
    //---------------------------------------------------------------------

    enum
    {
        INIT_KERNEL_THREADS = 128
    };

    // The output value type
    typedef typename If<(Equals<typename std::iterator_traits<OutputIteratorT>::value_type, void>::VALUE),  // OutputT =  (if output iterator's value type is void) ?
        typename std::iterator_traits<InputIteratorT>::value_type,                                          // ... then the input iterator's value type,
        typename std::iterator_traits<OutputIteratorT>::value_type>::Type OutputT;                          // ... else the output iterator's value type

    // Tile status descriptor interface type
    typedef ScanTileState<OutputT> ScanTileStateT;


    //---------------------------------------------------------------------
    // Tuning policies
    //---------------------------------------------------------------------

    struct Policy600
    {
        typedef AgentScanPolicy<
            CUB_SCALED_GRANULARITIES(128, 15, OutputT),      
                BLOCK_LOAD_TRANSPOSE,
                LOAD_DEFAULT,
                BLOCK_STORE_TRANSPOSE,
                BLOCK_SCAN_WARP_SCANS>
            ScanPolicyT;
    };


    struct Policy520
    {
        // Titan X: 32.47B items/s @ 48M 32-bit T
        typedef AgentScanPolicy<
                CUB_SCALED_GRANULARITIES(128, 12, OutputT),      
                BLOCK_LOAD_DIRECT,
                LOAD_LDG,
                BLOCK_STORE_WARP_TRANSPOSE,
                BLOCK_SCAN_WARP_SCANS>
            ScanPolicyT;
    };


    struct Policy350
    {
        // GTX Titan: 29.5B items/s (232.4 GB/s) @ 48M 32-bit T
        typedef AgentScanPolicy<
                CUB_SCALED_GRANULARITIES(128, 12, OutputT),      
                BLOCK_LOAD_DIRECT,
                LOAD_LDG,
                BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED,
                BLOCK_SCAN_RAKING>
            ScanPolicyT;
    };

    struct Policy300
    {
        typedef AgentScanPolicy<
                CUB_SCALED_GRANULARITIES(256, 9, OutputT),      
                BLOCK_LOAD_WARP_TRANSPOSE,
                LOAD_DEFAULT,
                BLOCK_STORE_WARP_TRANSPOSE,
                BLOCK_SCAN_WARP_SCANS>
            ScanPolicyT;
    };

    struct Policy200
    {
        // GTX 580: 20.3B items/s (162.3 GB/s) @ 48M 32-bit T
        typedef AgentScanPolicy<
                CUB_SCALED_GRANULARITIES(128, 12, OutputT),      
                BLOCK_LOAD_WARP_TRANSPOSE,
                LOAD_DEFAULT,
                BLOCK_STORE_WARP_TRANSPOSE,
                BLOCK_SCAN_WARP_SCANS>
            ScanPolicyT;
    };

    struct Policy130
    {
        typedef AgentScanPolicy<
                CUB_SCALED_GRANULARITIES(96, 21, OutputT),      
                BLOCK_LOAD_WARP_TRANSPOSE,
                LOAD_DEFAULT,
                BLOCK_STORE_WARP_TRANSPOSE,
                BLOCK_SCAN_RAKING_MEMOIZE>
            ScanPolicyT;
    };

    struct Policy100
    {
        typedef AgentScanPolicy<
                CUB_SCALED_GRANULARITIES(64, 9, OutputT),      
                BLOCK_LOAD_WARP_TRANSPOSE,
                LOAD_DEFAULT,
                BLOCK_STORE_WARP_TRANSPOSE,
                BLOCK_SCAN_WARP_SCANS>
            ScanPolicyT;
    };


    //---------------------------------------------------------------------
    // Tuning policies of current PTX compiler pass
    //---------------------------------------------------------------------

#if (CUB_PTX_ARCH >= 600)
    typedef Policy600 PtxPolicy;

#elif (CUB_PTX_ARCH >= 520)
    typedef Policy520 PtxPolicy;

#elif (CUB_PTX_ARCH >= 350)
    typedef Policy350 PtxPolicy;

#elif (CUB_PTX_ARCH >= 300)
    typedef Policy300 PtxPolicy;

#elif (CUB_PTX_ARCH >= 200)
    typedef Policy200 PtxPolicy;

#elif (CUB_PTX_ARCH >= 130)
    typedef Policy130 PtxPolicy;

#else
    typedef Policy100 PtxPolicy;

#endif

    // "Opaque" policies (whose parameterizations aren't reflected in the type signature)
    struct PtxAgentScanPolicy : PtxPolicy::ScanPolicyT {};


    //---------------------------------------------------------------------
    // Utilities
    //---------------------------------------------------------------------

    template <typename KernelConfig>
    CUB_RUNTIME_FUNCTION __forceinline__
    static void InitConfigs(
        int             ptx_version,
        KernelConfig    &scan_kernel_config)
    {
    #if (CUB_PTX_ARCH > 0)
        (void)ptx_version;

        // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy
        scan_kernel_config.template Init<PtxAgentScanPolicy>();

    #else

        // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version
        if (ptx_version >= 600)
        {
            scan_kernel_config.template Init<typename Policy600::ScanPolicyT>();
        }
        else if (ptx_version >= 520)
        {
            scan_kernel_config.template Init<typename Policy520::ScanPolicyT>();
        }
        else if (ptx_version >= 350)
        {
            scan_kernel_config.template Init<typename Policy350::ScanPolicyT>();
        }
        else if (ptx_version >= 300)
        {
            scan_kernel_config.template Init<typename Policy300::ScanPolicyT>();
        }
        else if (ptx_version >= 200)
        {
            scan_kernel_config.template Init<typename Policy200::ScanPolicyT>();
        }
        else if (ptx_version >= 130)
        {
            scan_kernel_config.template Init<typename Policy130::ScanPolicyT>();
        }
        else
        {
            scan_kernel_config.template Init<typename Policy100::ScanPolicyT>();
        }

    #endif
    }


    struct KernelConfig
    {
        int block_threads;
        int items_per_thread;
        int tile_items;

        template <typename PolicyT>
        CUB_RUNTIME_FUNCTION __forceinline__
        void Init()
        {
            block_threads       = PolicyT::BLOCK_THREADS;
            items_per_thread    = PolicyT::ITEMS_PER_THREAD;
            tile_items          = block_threads * items_per_thread;
        }
    };


    //---------------------------------------------------------------------
    // Dispatch entrypoints
    //---------------------------------------------------------------------

    template <
        typename            ScanInitKernelPtrT,     
        typename            ScanSweepKernelPtrT>    
    CUB_RUNTIME_FUNCTION __forceinline__
    static cudaError_t Dispatch(
        void*               d_temp_storage,         
        size_t&             temp_storage_bytes,     
        InputIteratorT      d_in,                   
        OutputIteratorT     d_out,                  
        ScanOpT             scan_op,                
        InitValueT          init_value,             
        OffsetT             num_items,              
        cudaStream_t        stream,                 
        bool                debug_synchronous,      
        int                 /*ptx_version*/,        
        ScanInitKernelPtrT  init_kernel,            
        ScanSweepKernelPtrT scan_kernel,            
        KernelConfig        scan_kernel_config)     
    {

#ifndef CUB_RUNTIME_ENABLED
        (void)d_temp_storage;
        (void)temp_storage_bytes;
        (void)d_in;
        (void)d_out;
        (void)scan_op;
        (void)init_value;
        (void)num_items;
        (void)stream;
        (void)debug_synchronous;
        (void)init_kernel;
        (void)scan_kernel;
        (void)scan_kernel_config;

        // Kernel launch not supported from this device
        return CubDebug(cudaErrorNotSupported);

#else
        cudaError error = cudaSuccess;
        do
        {
            // Get device ordinal
            int device_ordinal;
            if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;

            // Get SM count
            int sm_count;
            if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;

            // Number of input tiles
            int tile_size = scan_kernel_config.block_threads * scan_kernel_config.items_per_thread;
            int num_tiles = (num_items + tile_size - 1) / tile_size;

            // Specify temporary storage allocation requirements
            size_t  allocation_sizes[1];
            if (CubDebug(error = ScanTileStateT::AllocationSize(num_tiles, allocation_sizes[0]))) break;    // bytes needed for tile status descriptors

            // Compute allocation pointers into the single storage blob (or compute the necessary size of the blob)
            void* allocations[1];
            if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;
            if (d_temp_storage == NULL)
            {
                // Return if the caller is simply requesting the size of the storage allocation
                break;
            }

            // Return if empty problem
            if (num_items == 0)
                break;

            // Construct the tile status interface
            ScanTileStateT tile_state;
            if (CubDebug(error = tile_state.Init(num_tiles, allocations[0], allocation_sizes[0]))) break;

            // Log init_kernel configuration
            int init_grid_size = (num_tiles + INIT_KERNEL_THREADS - 1) / INIT_KERNEL_THREADS;
            if (debug_synchronous) _CubLog("Invoking init_kernel<<<%d, %d, 0, %lld>>>()\n", init_grid_size, INIT_KERNEL_THREADS, (long long) stream);

            // Invoke init_kernel to initialize tile descriptors
            init_kernel<<<init_grid_size, INIT_KERNEL_THREADS, 0, stream>>>(
                tile_state,
                num_tiles);

            // Check for failure to launch
            if (CubDebug(error = cudaPeekAtLastError())) break;

            // Sync the stream if specified to flush runtime errors
            if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;

            // Get SM occupancy for scan_kernel
            int scan_sm_occupancy;
            if (CubDebug(error = MaxSmOccupancy(
                scan_sm_occupancy,            // out
                scan_kernel,
                scan_kernel_config.block_threads))) break;

            // Get max x-dimension of grid
            int max_dim_x;
            if (CubDebug(error = cudaDeviceGetAttribute(&max_dim_x, cudaDevAttrMaxGridDimX, device_ordinal))) break;;

            // Run grids in epochs (in case number of tiles exceeds max x-dimension
            int scan_grid_size = CUB_MIN(num_tiles, max_dim_x);
            for (int start_tile = 0; start_tile < num_tiles; start_tile += scan_grid_size)
            {
                // Log scan_kernel configuration
                if (debug_synchronous) _CubLog("Invoking %d scan_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\n",
                    start_tile, scan_grid_size, scan_kernel_config.block_threads, (long long) stream, scan_kernel_config.items_per_thread, scan_sm_occupancy);

                // Invoke scan_kernel
                scan_kernel<<<scan_grid_size, scan_kernel_config.block_threads, 0, stream>>>(
                    d_in,
                    d_out,
                    tile_state,
                    start_tile,
                    scan_op,
                    init_value,
                    num_items);

                // Check for failure to launch
                if (CubDebug(error = cudaPeekAtLastError())) break;

                // Sync the stream if specified to flush runtime errors
                if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;
            }
        }
        while (0);

        return error;

#endif  // CUB_RUNTIME_ENABLED
    }


    CUB_RUNTIME_FUNCTION __forceinline__
    static cudaError_t Dispatch(
        void*           d_temp_storage,         
        size_t&         temp_storage_bytes,     
        InputIteratorT  d_in,                   
        OutputIteratorT d_out,                  
        ScanOpT         scan_op,                
        InitValueT      init_value,             
        OffsetT         num_items,              
        cudaStream_t    stream,                 
        bool            debug_synchronous)      
    {
        cudaError error = cudaSuccess;
        do
        {
            // Get PTX version
            int ptx_version;
            if (CubDebug(error = PtxVersion(ptx_version))) break;

            // Get kernel kernel dispatch configurations
            KernelConfig scan_kernel_config;
            InitConfigs(ptx_version, scan_kernel_config);

            // Dispatch
            if (CubDebug(error = Dispatch(
                d_temp_storage,
                temp_storage_bytes,
                d_in,
                d_out,
                scan_op,
                init_value,
                num_items,
                stream,
                debug_synchronous,
                ptx_version,
                DeviceScanInitKernel<ScanTileStateT>,
                DeviceScanKernel<PtxAgentScanPolicy, InputIteratorT, OutputIteratorT, ScanTileStateT, ScanOpT, InitValueT, OffsetT>,
                scan_kernel_config))) break;
        }
        while (0);

        return error;
    }
};



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