Alien-XGBoost

 view release on metacpan or  search on metacpan

xgboost/cub/cub/device/dispatch/dispatch_radix_sort.cuh  view on Meta::CPAN

        // Downsweep policies
        typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 16 / SCALE_FACTOR_4B),  BLOCK_LOAD_DIRECT, LOAD_LDG, true, BLOCK_SCAN_RAKING_MEMOIZE, RADIX_SORT_SCATTER_TWO_PHASE, PRIMARY_RADIX_BITS>   DownsweepPolicy;
        typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 16 / SCALE_FACTOR_4B),  BLOCK_LOAD_DIRECT, LOAD_LDG, true, BLOCK_SCAN_RAKING_MEMOIZE, RADIX_SORT_SCATTER_TWO_PHASE, ALT_RADIX_BITS>       AltDownsweepPolicy;

        // Upsweep policies
        typedef DownsweepPolicy UpsweepPolicy;
        typedef AltDownsweepPolicy AltUpsweepPolicy;

        // Single-tile policy
        typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 19 / SCALE_FACTOR_4B),  BLOCK_LOAD_DIRECT, LOAD_LDG, true, BLOCK_SCAN_WARP_SCANS, RADIX_SORT_SCATTER_TWO_PHASE, PRIMARY_RADIX_BITS> SingleTilePolicy;

        // Segmented policies
        typedef DownsweepPolicy     SegmentedPolicy;
        typedef AltDownsweepPolicy  AltSegmentedPolicy;
    };

    /// MaxPolicy
    typedef Policy620 MaxPolicy;
};



/******************************************************************************
 * Single-problem dispatch
 ******************************************************************************/

/**
 * Utility class for dispatching the appropriately-tuned kernels for device-wide radix sort
 */
template <
    bool     IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low
    typename KeyT,          ///< Key type
    typename ValueT,        ///< Value type
    typename OffsetT>       ///< Signed integer type for global offsets
struct DispatchRadixSort :
    DeviceRadixSortPolicy<KeyT, ValueT, OffsetT>
{
    //------------------------------------------------------------------------------
    // Constants
    //------------------------------------------------------------------------------

    enum
    {
        // Whether this is a keys-only (or key-value) sort
        KEYS_ONLY = (Equals<ValueT, NullType>::VALUE),
    };


    //------------------------------------------------------------------------------
    // Problem state
    //------------------------------------------------------------------------------

    void                    *d_temp_storage;        ///< [in] %Device-accessible allocation of temporary storage.  When NULL, the required allocation size is written to \p temp_storage_bytes and no work is done.
    size_t                  &temp_storage_bytes;    ///< [in,out] Reference to size in bytes of \p d_temp_storage allocation
    DoubleBuffer<KeyT>      &d_keys;                ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys
    DoubleBuffer<ValueT>    &d_values;              ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values
    OffsetT                 num_items;              ///< [in] Number of items to sort
    int                     begin_bit;              ///< [in] The beginning (least-significant) bit index needed for key comparison
    int                     end_bit;                ///< [in] The past-the-end (most-significant) bit index needed for key comparison
    cudaStream_t            stream;                 ///< [in] CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
    bool                    debug_synchronous;      ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors.  Also causes launch configurations to be printed to the console.  Default is \p false.
    int                     ptx_version;            ///< [in] PTX version
    bool                    is_overwrite_okay;      ///< [in] Whether is okay to overwrite source buffers


    //------------------------------------------------------------------------------
    // Constructor
    //------------------------------------------------------------------------------

    /// Constructor
    CUB_RUNTIME_FUNCTION __forceinline__
    DispatchRadixSort(
        void*                   d_temp_storage,
        size_t                  &temp_storage_bytes,
        DoubleBuffer<KeyT>      &d_keys,
        DoubleBuffer<ValueT>    &d_values,
        OffsetT                 num_items,
        int                     begin_bit,
        int                     end_bit,
        bool                    is_overwrite_okay,
        cudaStream_t            stream,
        bool                    debug_synchronous,
        int                     ptx_version)
    :
        d_temp_storage(d_temp_storage),
        temp_storage_bytes(temp_storage_bytes),
        d_keys(d_keys),
        d_values(d_values),
        num_items(num_items),
        begin_bit(begin_bit),
        end_bit(end_bit),
        stream(stream),
        debug_synchronous(debug_synchronous),
        ptx_version(ptx_version),
        is_overwrite_okay(is_overwrite_okay)
    {}


    //------------------------------------------------------------------------------
    // Small-problem (single tile) invocation
    //------------------------------------------------------------------------------

    /// Invoke a single block to sort in-core
    template <
        typename                ActivePolicyT,          ///< Umbrella policy active for the target device
        typename                SingleTileKernelT>      ///< Function type of cub::DeviceRadixSortSingleTileKernel
    CUB_RUNTIME_FUNCTION __forceinline__
    cudaError_t InvokeSingleTile(
        SingleTileKernelT       single_tile_kernel)     ///< [in] Kernel function pointer to parameterization of cub::DeviceRadixSortSingleTileKernel
    {
#ifndef CUB_RUNTIME_ENABLED
        (void)single_tile_kernel;
        // Kernel launch not supported from this device
        return CubDebug(cudaErrorNotSupported );
#else
        cudaError error = cudaSuccess;
        do
        {
            // Return if the caller is simply requesting the size of the storage allocation
            if (d_temp_storage == NULL)
            {
                temp_storage_bytes = 1;
                break;
            }

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

            // Log single_tile_kernel configuration
            if (debug_synchronous)
                _CubLog("Invoking single_tile_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\n",
                    1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, (long long) stream,
                    ActivePolicyT::SingleTilePolicy::ITEMS_PER_THREAD, 1, begin_bit, ActivePolicyT::SingleTilePolicy::RADIX_BITS);

            // Invoke upsweep_kernel with same grid size as downsweep_kernel
            single_tile_kernel<<<1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, 0, stream>>>(
                d_keys.Current(),
                d_keys.Alternate(),
                d_values.Current(),
                d_values.Alternate(),
                num_items,
                begin_bit,
                end_bit);

            // 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;

            // Update selector
            d_keys.selector ^= 1;
            d_values.selector ^= 1;
        }
        while (0);

        return error;

#endif // CUB_RUNTIME_ENABLED
    }


    //------------------------------------------------------------------------------
    // Normal problem size invocation
    //------------------------------------------------------------------------------

    /**
     * Invoke a three-kernel sorting pass at the current bit.
     */
    template <typename PassConfigT>
    CUB_RUNTIME_FUNCTION __forceinline__
    cudaError_t InvokePass(
        const KeyT      *d_keys_in,
        KeyT            *d_keys_out,
        const ValueT    *d_values_in,
        ValueT          *d_values_out,
        OffsetT         *d_spine,
        int             spine_length,
        int             &current_bit,
        PassConfigT     &pass_config)
    {
        cudaError error = cudaSuccess;
        do
        {
            int pass_bits = CUB_MIN(pass_config.radix_bits, (end_bit - current_bit));

            // Log upsweep_kernel configuration
            if (debug_synchronous)
                _CubLog("Invoking upsweep_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\n",
                pass_config.even_share.grid_size, pass_config.upsweep_config.block_threads, (long long) stream,
                pass_config.upsweep_config.items_per_thread, pass_config.upsweep_config.sm_occupancy, current_bit, pass_bits);

            // Invoke upsweep_kernel with same grid size as downsweep_kernel
            pass_config.upsweep_kernel<<<pass_config.even_share.grid_size, pass_config.upsweep_config.block_threads, 0, stream>>>(
                d_keys_in,
                d_spine,
                num_items,
                current_bit,
                pass_bits,
                pass_config.even_share);

            // 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;

            // Log scan_kernel configuration
            if (debug_synchronous) _CubLog("Invoking scan_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread\n",
                1, pass_config.scan_config.block_threads, (long long) stream, pass_config.scan_config.items_per_thread);

            // Invoke scan_kernel
            pass_config.scan_kernel<<<1, pass_config.scan_config.block_threads, 0, stream>>>(
                d_spine,
                spine_length);

            // 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;

            // Log downsweep_kernel configuration
            if (debug_synchronous) _CubLog("Invoking downsweep_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\n",
                pass_config.even_share.grid_size, pass_config.downsweep_config.block_threads, (long long) stream,
                pass_config.downsweep_config.items_per_thread, pass_config.downsweep_config.sm_occupancy);

            // Invoke downsweep_kernel
            pass_config.downsweep_kernel<<<pass_config.even_share.grid_size, pass_config.downsweep_config.block_threads, 0, stream>>>(
                d_keys_in,
                d_keys_out,
                d_values_in,
                d_values_out,
                d_spine,
                num_items,
                current_bit,
                pass_bits,
                pass_config.even_share);

            // 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;

            // Update current bit
            current_bit += pass_bits;
        }
        while (0);

        return error;
    }



    /// Pass configuration structure
    template <
        typename UpsweepKernelT,
        typename ScanKernelT,
        typename DownsweepKernelT>
    struct PassConfig
    {
        UpsweepKernelT          upsweep_kernel;
        KernelConfig            upsweep_config;
        ScanKernelT             scan_kernel;
        KernelConfig            scan_config;
        DownsweepKernelT        downsweep_kernel;
        KernelConfig            downsweep_config;
        int                     radix_bits;
        int                     radix_digits;
        int                     max_downsweep_grid_size;
        GridEvenShare<OffsetT>  even_share;

        /// Initialize pass configuration
        template <
            typename UpsweepPolicyT,
            typename ScanPolicyT,
            typename DownsweepPolicyT>
        CUB_RUNTIME_FUNCTION __forceinline__
        cudaError_t InitPassConfig(
            UpsweepKernelT      upsweep_kernel,
            ScanKernelT         scan_kernel,
            DownsweepKernelT    downsweep_kernel,
            int                 ptx_version,
            int                 sm_count,
            int                 num_items)
        {
            cudaError error = cudaSuccess;
            do
            {
                this->upsweep_kernel    = upsweep_kernel;
                this->scan_kernel       = scan_kernel;
                this->downsweep_kernel  = downsweep_kernel;
                radix_bits              = DownsweepPolicyT::RADIX_BITS;
                radix_digits            = 1 << radix_bits;

                if (CubDebug(error = upsweep_config.Init<UpsweepPolicyT>(upsweep_kernel))) break;
                if (CubDebug(error = scan_config.Init<ScanPolicyT>(scan_kernel))) break;
                if (CubDebug(error = downsweep_config.Init<DownsweepPolicyT>(downsweep_kernel))) break;

                max_downsweep_grid_size = (downsweep_config.sm_occupancy * sm_count) * CUB_SUBSCRIPTION_FACTOR(ptx_version);

                even_share = GridEvenShare<OffsetT>(
                    num_items,

xgboost/cub/cub/device/dispatch/dispatch_radix_sort.cuh  view on Meta::CPAN

            d_values.selector = (d_values.selector + num_passes) & 1;
        }
        while (0);

        return error;

#endif // CUB_RUNTIME_ENABLED
    }


    //------------------------------------------------------------------------------
    // Chained policy invocation
    //------------------------------------------------------------------------------

    /// Invocation
    template <typename ActivePolicyT>
    CUB_RUNTIME_FUNCTION __forceinline__
    cudaError_t Invoke()
    {
        typedef typename DispatchRadixSort::MaxPolicy       MaxPolicyT;
        typedef typename ActivePolicyT::SingleTilePolicy    SingleTilePolicyT;

        // Force kernel code-generation in all compiler passes
        if (num_items <= (SingleTilePolicyT::BLOCK_THREADS * SingleTilePolicyT::ITEMS_PER_THREAD))
        {
            // Small, single tile size
            return InvokeSingleTile<ActivePolicyT>(
                DeviceRadixSortSingleTileKernel<MaxPolicyT, IS_DESCENDING, KeyT, ValueT, OffsetT>);
        }
        else
        {
            // Regular size
            return InvokePasses<ActivePolicyT>(
                DeviceRadixSortUpsweepKernel<   MaxPolicyT, false,   IS_DESCENDING, KeyT, OffsetT>,
                DeviceRadixSortUpsweepKernel<   MaxPolicyT, true,    IS_DESCENDING, KeyT, OffsetT>,
                RadixSortScanBinsKernel<        MaxPolicyT, OffsetT>,
                DeviceRadixSortDownsweepKernel< MaxPolicyT, false,   IS_DESCENDING, KeyT, ValueT, OffsetT>,
                DeviceRadixSortDownsweepKernel< MaxPolicyT, true,    IS_DESCENDING, KeyT, ValueT, OffsetT>);
        }
    }


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

    /**
     * Internal dispatch routine
     */
    CUB_RUNTIME_FUNCTION __forceinline__
    static cudaError_t Dispatch(
        void*                   d_temp_storage,         ///< [in] %Device-accessible allocation of temporary storage.  When NULL, the required allocation size is written to \p temp_storage_bytes and no work is done.
        size_t                  &temp_storage_bytes,    ///< [in,out] Reference to size in bytes of \p d_temp_storage allocation
        DoubleBuffer<KeyT>      &d_keys,                ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys
        DoubleBuffer<ValueT>    &d_values,              ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values
        OffsetT                 num_items,              ///< [in] Number of items to sort
        int                     begin_bit,              ///< [in] The beginning (least-significant) bit index needed for key comparison
        int                     end_bit,                ///< [in] The past-the-end (most-significant) bit index needed for key comparison
        bool                    is_overwrite_okay,      ///< [in] Whether is okay to overwrite source buffers
        cudaStream_t            stream,                 ///< [in] CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
        bool                    debug_synchronous)      ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors.  Also causes launch configurations to be printed to the console.  Default is \p false.
    {
        typedef typename DispatchRadixSort::MaxPolicy MaxPolicyT;

        cudaError_t error;
        do {
            // Get PTX version
            int ptx_version;
            if (CubDebug(error = PtxVersion(ptx_version))) break;

            // Create dispatch functor
            DispatchRadixSort dispatch(
                d_temp_storage, temp_storage_bytes,
                d_keys, d_values,
                num_items, begin_bit, end_bit, is_overwrite_okay,
                stream, debug_synchronous, ptx_version);

            // Dispatch to chained policy
            if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;

        } while (0);

        return error;
    }
};




/******************************************************************************
 * Segmented dispatch
 ******************************************************************************/

/**
 * Utility class for dispatching the appropriately-tuned kernels for segmented device-wide radix sort
 */
template <
    bool     IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low
    typename KeyT,          ///< Key type
    typename ValueT,        ///< Value type
    typename OffsetT>       ///< Signed integer type for global offsets
struct DispatchSegmentedRadixSort :
    DeviceRadixSortPolicy<KeyT, ValueT, OffsetT>
{
    //------------------------------------------------------------------------------
    // Constants
    //------------------------------------------------------------------------------

    enum
    {
        // Whether this is a keys-only (or key-value) sort
        KEYS_ONLY = (Equals<ValueT, NullType>::VALUE),
    };


    //------------------------------------------------------------------------------
    // Parameter members
    //------------------------------------------------------------------------------

    void                    *d_temp_storage;        ///< [in] %Device-accessible allocation of temporary storage.  When NULL, the required allocation size is written to \p temp_storage_bytes and no work is done.
    size_t                  &temp_storage_bytes;    ///< [in,out] Reference to size in bytes of \p d_temp_storage allocation
    DoubleBuffer<KeyT>      &d_keys;                ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys
    DoubleBuffer<ValueT>    &d_values;              ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values
    OffsetT                 num_items;              ///< [in] Number of items to sort
    OffsetT                 num_segments;           ///< [in] The number of segments that comprise the sorting data
    const OffsetT           *d_begin_offsets;       ///< [in] %Device-accessible pointer to the sequence of beginning offsets of length \p num_segments, such that <tt>d_begin_offsets[i]</tt> is the first element of the <em>i</em><sup>th</sup> data se...
    const OffsetT           *d_end_offsets;         ///< [in] %Device-accessible pointer to the sequence of ending offsets of length \p num_segments, such that <tt>d_end_offsets[i]-1</tt> is the last element of the <em>i</em><sup>th</sup> data segmen...
    int                     begin_bit;              ///< [in] The beginning (least-significant) bit index needed for key comparison
    int                     end_bit;                ///< [in] The past-the-end (most-significant) bit index needed for key comparison
    cudaStream_t            stream;                 ///< [in] CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
    bool                    debug_synchronous;      ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors.  Also causes launch configurations to be printed to the console.  Default is \p false.
    int                     ptx_version;            ///< [in] PTX version
    bool                    is_overwrite_okay;      ///< [in] Whether is okay to overwrite source buffers


    //------------------------------------------------------------------------------
    // Constructors
    //------------------------------------------------------------------------------

    /// Constructor
    CUB_RUNTIME_FUNCTION __forceinline__
    DispatchSegmentedRadixSort(
        void*                   d_temp_storage,
        size_t                  &temp_storage_bytes,
        DoubleBuffer<KeyT>      &d_keys,
        DoubleBuffer<ValueT>    &d_values,
        OffsetT                 num_items,
        OffsetT                 num_segments,
        const OffsetT           *d_begin_offsets,
        const OffsetT           *d_end_offsets,
        int                     begin_bit,
        int                     end_bit,
        bool                    is_overwrite_okay,
        cudaStream_t            stream,
        bool                    debug_synchronous,
        int                     ptx_version)
    :
        d_temp_storage(d_temp_storage),
        temp_storage_bytes(temp_storage_bytes),
        d_keys(d_keys),
        d_values(d_values),
        num_items(num_items),
        num_segments(num_segments),
        d_begin_offsets(d_begin_offsets),
        d_end_offsets(d_end_offsets),
        begin_bit(begin_bit),
        end_bit(end_bit),
        is_overwrite_okay(is_overwrite_okay),
        stream(stream),
        debug_synchronous(debug_synchronous),
        ptx_version(ptx_version)
    {}


    //------------------------------------------------------------------------------
    // Multi-segment invocation
    //------------------------------------------------------------------------------

    /// Invoke a three-kernel sorting pass at the current bit.
    template <typename PassConfigT>
    CUB_RUNTIME_FUNCTION __forceinline__
    cudaError_t InvokePass(
        const KeyT      *d_keys_in,
        KeyT            *d_keys_out,
        const ValueT    *d_values_in,
        ValueT          *d_values_out,
        int             &current_bit,
        PassConfigT     &pass_config)
    {
        cudaError error = cudaSuccess;
        do
        {
            int pass_bits = CUB_MIN(pass_config.radix_bits, (end_bit - current_bit));

            // Log kernel configuration
            if (debug_synchronous)
                _CubLog("Invoking segmented_kernels<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\n",
                    num_segments, pass_config.segmented_config.block_threads, (long long) stream,
                pass_config.segmented_config.items_per_thread, pass_config.segmented_config.sm_occupancy, current_bit, pass_bits);

            pass_config.segmented_kernel<<<num_segments, pass_config.segmented_config.block_threads, 0, stream>>>(
                d_keys_in, d_keys_out,
                d_values_in,  d_values_out,
                d_begin_offsets, d_end_offsets, num_segments,
                current_bit, pass_bits);

            // 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;

            // Update current bit
            current_bit += pass_bits;
        }
        while (0);

        return error;
    }


    /// PassConfig data structure
    template <typename SegmentedKernelT>
    struct PassConfig
    {
        SegmentedKernelT    segmented_kernel;
        KernelConfig        segmented_config;
        int                 radix_bits;
        int                 radix_digits;

        /// Initialize pass configuration
        template <typename SegmentedPolicyT>
        CUB_RUNTIME_FUNCTION __forceinline__
        cudaError_t InitPassConfig(SegmentedKernelT segmented_kernel)
        {
            this->segmented_kernel  = segmented_kernel;
            this->radix_bits        = SegmentedPolicyT::RADIX_BITS;
            this->radix_digits      = 1 << radix_bits;

            return CubDebug(segmented_config.Init<SegmentedPolicyT>(segmented_kernel));
        }
    };


    /// Invocation (run multiple digit passes)
    template <
        typename                ActivePolicyT,          ///< Umbrella policy active for the target device
        typename                SegmentedKernelT>       ///< Function type of cub::DeviceSegmentedRadixSortKernel
    CUB_RUNTIME_FUNCTION __forceinline__
    cudaError_t InvokePasses(
        SegmentedKernelT     segmented_kernel,          ///< [in] Kernel function pointer to parameterization of cub::DeviceSegmentedRadixSortKernel
        SegmentedKernelT     alt_segmented_kernel)      ///< [in] Alternate kernel function pointer to parameterization of cub::DeviceSegmentedRadixSortKernel
    {
#ifndef CUB_RUNTIME_ENABLED
      (void)segmented_kernel;
      (void)alt_segmented_kernel;

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

        cudaError error = cudaSuccess;
        do
        {
            // Init regular and alternate kernel configurations
            PassConfig<SegmentedKernelT> pass_config, alt_pass_config;
            if ((error = pass_config.template       InitPassConfig<typename ActivePolicyT::SegmentedPolicy>(segmented_kernel))) break;
            if ((error = alt_pass_config.template   InitPassConfig<typename ActivePolicyT::AltSegmentedPolicy>(alt_segmented_kernel))) break;

            // Temporary storage allocation requirements
            void* allocations[2];

xgboost/cub/cub/device/dispatch/dispatch_radix_sort.cuh  view on Meta::CPAN


                // Invert selectors and update current bit
                d_keys_remaining_passes.selector ^= 1;
                d_values_remaining_passes.selector ^= 1;
            }

            // Update selector
            if (!is_overwrite_okay) {
                num_passes = 1; // Sorted data always ends up in the other vector
            }

            d_keys.selector = (d_keys.selector + num_passes) & 1;
            d_values.selector = (d_values.selector + num_passes) & 1;
        }
        while (0);

        return error;

#endif // CUB_RUNTIME_ENABLED
    }


    //------------------------------------------------------------------------------
    // Chained policy invocation
    //------------------------------------------------------------------------------

    /// Invocation
    template <typename ActivePolicyT>
    CUB_RUNTIME_FUNCTION __forceinline__
    cudaError_t Invoke()
    {
        typedef typename DispatchSegmentedRadixSort::MaxPolicy MaxPolicyT;

        // Force kernel code-generation in all compiler passes
        return InvokePasses<ActivePolicyT>(
            DeviceSegmentedRadixSortKernel<MaxPolicyT, false,   IS_DESCENDING, KeyT, ValueT, OffsetT>,
            DeviceSegmentedRadixSortKernel<MaxPolicyT, true,    IS_DESCENDING, KeyT, ValueT, OffsetT>);
    }


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


    /// Internal dispatch routine
    CUB_RUNTIME_FUNCTION __forceinline__
    static cudaError_t Dispatch(
        void*                   d_temp_storage,         ///< [in] %Device-accessible allocation of temporary storage.  When NULL, the required allocation size is written to \p temp_storage_bytes and no work is done.
        size_t                  &temp_storage_bytes,    ///< [in,out] Reference to size in bytes of \p d_temp_storage allocation
        DoubleBuffer<KeyT>      &d_keys,                ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys
        DoubleBuffer<ValueT>    &d_values,              ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values
        int                     num_items,              ///< [in] Number of items to sort
        int                     num_segments,           ///< [in] The number of segments that comprise the sorting data
        const int               *d_begin_offsets,       ///< [in] %Device-accessible pointer to the sequence of beginning offsets of length \p num_segments, such that <tt>d_begin_offsets[i]</tt> is the first element of the <em>i</em><sup>th</sup> dat...
        const int               *d_end_offsets,         ///< [in] %Device-accessible pointer to the sequence of ending offsets of length \p num_segments, such that <tt>d_end_offsets[i]-1</tt> is the last element of the <em>i</em><sup>th</sup> data se...
        int                     begin_bit,              ///< [in] The beginning (least-significant) bit index needed for key comparison
        int                     end_bit,                ///< [in] The past-the-end (most-significant) bit index needed for key comparison
        bool                    is_overwrite_okay,      ///< [in] Whether is okay to overwrite source buffers
        cudaStream_t            stream,                 ///< [in] CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
        bool                    debug_synchronous)      ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors.  Also causes launch configurations to be printed to the console.  Default is \p false.
    {
        typedef typename DispatchSegmentedRadixSort::MaxPolicy MaxPolicyT;

        cudaError_t error;
        do {
            // Get PTX version
            int ptx_version;
            if (CubDebug(error = PtxVersion(ptx_version))) break;

            // Create dispatch functor
            DispatchSegmentedRadixSort dispatch(
                d_temp_storage, temp_storage_bytes,
                d_keys, d_values,
                num_items, num_segments, d_begin_offsets, d_end_offsets,
                begin_bit, end_bit, is_overwrite_okay,
                stream, debug_synchronous, ptx_version);

            // Dispatch to chained policy
            if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;

        } while (0);

        return error;
    }
};


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




( run in 1.809 second using v1.01-cache-2.11-cpan-39bf76dae61 )