File size: 23,459 Bytes
0dc1b04
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494

/******************************************************************************
 * Copyright (c) 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
 * cub::DeviceSelect::UniqueByKey provides device-wide, parallel operations for selecting unique items by key from sequences of data items residing within device-accessible memory.
 */

#include <cub/agent/agent_unique_by_key.cuh>
#include <cub/device/dispatch/dispatch_scan.cuh>
#include <cub/util_deprecated.cuh>
#include <cub/util_macro.cuh>
#include <cub/util_math.cuh>

#include <iterator>

CUB_NAMESPACE_BEGIN

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

/**
 * Unique by key kernel entry point (multi-block)
 */
template <
    typename ChainedPolicyT,
    typename KeyInputIteratorT,                     ///< Random-access input iterator type for keys
    typename ValueInputIteratorT,                   ///< Random-access input iterator type for values
    typename KeyOutputIteratorT,                    ///< Random-access output iterator type for keys
    typename ValueOutputIteratorT,                  ///< Random-access output iterator type for values
    typename NumSelectedIteratorT,                  ///< Output iterator type for recording the number of items selected
    typename ScanTileStateT,                        ///< Tile status interface type
    typename EqualityOpT,                           ///< Equality operator type
    typename OffsetT>                               ///< Signed integer type for global offsets
__launch_bounds__ (int(ChainedPolicyT::ActivePolicy::UniqueByKeyPolicyT::BLOCK_THREADS))
__global__ void DeviceUniqueByKeySweepKernel(
    KeyInputIteratorT       d_keys_in,              ///< [in] Pointer to the input sequence of keys
    ValueInputIteratorT     d_values_in,            ///< [in] Pointer to the input sequence of values
    KeyOutputIteratorT      d_keys_out,             ///< [out] Pointer to the output sequence of selected data items
    ValueOutputIteratorT    d_values_out,           ///< [out] Pointer to the output sequence of selected data items
    NumSelectedIteratorT    d_num_selected_out,     ///< [out] Pointer to the total number of items selected (i.e., length of \p d_keys_out or \p d_values_out)
    ScanTileStateT          tile_state,             ///< [in] Tile status interface
    EqualityOpT             equality_op,            ///< [in] Equality operator
    OffsetT                 num_items,              ///< [in] Total number of input items (i.e., length of \p d_keys_in or \p d_values_in)
    int                     num_tiles)              ///< [in] Total number of tiles for the entire problem
{
    using AgentUniqueByKeyPolicyT = typename ChainedPolicyT::ActivePolicy::UniqueByKeyPolicyT;

    // Thread block type for selecting data from input tiles
    using AgentUniqueByKeyT = AgentUniqueByKey<AgentUniqueByKeyPolicyT,
                                               KeyInputIteratorT,
                                               ValueInputIteratorT,
                                               KeyOutputIteratorT,
                                               ValueOutputIteratorT,
                                               EqualityOpT,
                                               OffsetT>;

    // Shared memory for AgentUniqueByKey
    __shared__ typename AgentUniqueByKeyT::TempStorage temp_storage;

    // Process tiles
    AgentUniqueByKeyT(temp_storage, d_keys_in, d_values_in, d_keys_out, d_values_out, equality_op, num_items).ConsumeRange(
        num_tiles,
        tile_state,
        d_num_selected_out);
}


/******************************************************************************
 * Policy
 ******************************************************************************/

template <typename KeyInputIteratorT>
struct DeviceUniqueByKeyPolicy
{
    using KeyT = typename std::iterator_traits<KeyInputIteratorT>::value_type;

    // SM350
    struct Policy350 : ChainedPolicy<350, Policy350, Policy350> 
    {
        const static int INPUT_SIZE = sizeof(KeyT);
        enum
        {
            NOMINAL_4B_ITEMS_PER_THREAD = 9,
            ITEMS_PER_THREAD = Nominal4BItemsToItems<KeyT>(NOMINAL_4B_ITEMS_PER_THREAD),
        };

        using UniqueByKeyPolicyT = AgentUniqueByKeyPolicy<128,
                                                          ITEMS_PER_THREAD,
                                                          cub::BLOCK_LOAD_WARP_TRANSPOSE,
                                                          cub::LOAD_LDG,
                                                          cub::BLOCK_SCAN_WARP_SCANS,
                                                          detail::default_delay_constructor_t<int>>;
    };

    // SM520
    struct Policy520 : ChainedPolicy<520, Policy520, Policy350>
    {
        const static int INPUT_SIZE = sizeof(KeyT);
        enum
        {
            NOMINAL_4B_ITEMS_PER_THREAD = 11,
            ITEMS_PER_THREAD = Nominal4BItemsToItems<KeyT>(NOMINAL_4B_ITEMS_PER_THREAD),
        };

        using UniqueByKeyPolicyT = AgentUniqueByKeyPolicy<64,
                                                          ITEMS_PER_THREAD,
                                                          cub::BLOCK_LOAD_WARP_TRANSPOSE,
                                                          cub::LOAD_LDG,
                                                          cub::BLOCK_SCAN_WARP_SCANS,
                                                          detail::default_delay_constructor_t<int>>;
    };

    /// MaxPolicy
    using MaxPolicy = Policy520;
};


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

/**
 * Utility class for dispatching the appropriately-tuned kernels for DeviceSelect
 */
template <
    typename KeyInputIteratorT,                 ///< Random-access input iterator type for keys
    typename ValueInputIteratorT,               ///< Random-access input iterator type for values
    typename KeyOutputIteratorT,                ///< Random-access output iterator type for keys
    typename ValueOutputIteratorT,              ///< Random-access output iterator type for values
    typename NumSelectedIteratorT,              ///< Output iterator type for recording the number of items selected
    typename EqualityOpT,                       ///< Equality operator type
    typename OffsetT,                           ///< Signed integer type for global offsets
    typename SelectedPolicy = DeviceUniqueByKeyPolicy<KeyInputIteratorT>>
struct DispatchUniqueByKey: SelectedPolicy
{
    /******************************************************************************
     * Types and constants
     ******************************************************************************/

    enum
    {
        INIT_KERNEL_THREADS = 128,
    };

    // The input key and value type
    using KeyT = typename std::iterator_traits<KeyInputIteratorT>::value_type;
    using ValueT = typename std::iterator_traits<ValueInputIteratorT>::value_type;

    // Tile status descriptor interface type
    using ScanTileStateT = ScanTileState<OffsetT>;


    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
    KeyInputIteratorT       d_keys_in;                  ///< [in] Pointer to the input sequence of keys
    ValueInputIteratorT     d_values_in;                ///< [in] Pointer to the input sequence of values
    KeyOutputIteratorT      d_keys_out;                 ///< [out] Pointer to the output sequence of selected data items
    ValueOutputIteratorT    d_values_out;               ///< [out] Pointer to the output sequence of selected data items
    NumSelectedIteratorT    d_num_selected_out;         ///< [out] Pointer to the total number of items selected (i.e., length of \p d_keys_out or \p d_values_out)
    EqualityOpT             equality_op;                ///< [in] Equality operator
    OffsetT                 num_items;                  ///< [in] Total number of input items (i.e., length of \p d_keys_in or \p d_values_in)
    cudaStream_t            stream;                     ///< [in] <b>[optional]</b> CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.

    CUB_RUNTIME_FUNCTION __forceinline__
    DispatchUniqueByKey(
        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
        KeyInputIteratorT       d_keys_in,              ///< [in] Pointer to the input sequence of keys
        ValueInputIteratorT     d_values_in,            ///< [in] Pointer to the input sequence of values
        KeyOutputIteratorT      d_keys_out,             ///< [out] Pointer to the output sequence of selected data items
        ValueOutputIteratorT    d_values_out,           ///< [out] Pointer to the output sequence of selected data items
        NumSelectedIteratorT    d_num_selected_out,     ///< [out] Pointer to the total number of items selected (i.e., length of \p d_keys_out or \p d_values_out)
        EqualityOpT             equality_op,            ///< [in] Equality operator
        OffsetT                 num_items,              ///< [in] Total number of input items (i.e., length of \p d_keys_in or \p d_values_in)
        cudaStream_t            stream                  ///< [in] <b>[optional]</b> CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
    ):
        d_temp_storage(d_temp_storage),
        temp_storage_bytes(temp_storage_bytes),
        d_keys_in(d_keys_in),
        d_values_in(d_values_in),
        d_keys_out(d_keys_out),
        d_values_out(d_values_out),
        d_num_selected_out(d_num_selected_out),
        equality_op(equality_op),
        num_items(num_items),
        stream(stream)
    {}

    CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED
    CUB_RUNTIME_FUNCTION __forceinline__
    DispatchUniqueByKey(
        void*                   d_temp_storage,     
        size_t&                 temp_storage_bytes,  
        KeyInputIteratorT       d_keys_in,            
        ValueInputIteratorT     d_values_in,           
        KeyOutputIteratorT      d_keys_out,             
        ValueOutputIteratorT    d_values_out,         
        NumSelectedIteratorT    d_num_selected_out,    
        EqualityOpT             equality_op,            
        OffsetT                 num_items,           
        cudaStream_t            stream,               
        bool                    debug_synchronous
    ):
        d_temp_storage(d_temp_storage),
        temp_storage_bytes(temp_storage_bytes),
        d_keys_in(d_keys_in),
        d_values_in(d_values_in),
        d_keys_out(d_keys_out),
        d_values_out(d_values_out),
        d_num_selected_out(d_num_selected_out),
        equality_op(equality_op),
        num_items(num_items),
        stream(stream)
    {
      CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
    }


    /******************************************************************************
     * Dispatch entrypoints
     ******************************************************************************/

    template <typename ActivePolicyT, typename InitKernel, typename ScanKernel>
    CUB_RUNTIME_FUNCTION __host__  __forceinline__
    cudaError_t Invoke(InitKernel init_kernel, ScanKernel scan_kernel)
    {
        using Policy = typename ActivePolicyT::UniqueByKeyPolicyT;
        using UniqueByKeyAgentT = AgentUniqueByKey<Policy,
                                                   KeyInputIteratorT,
                                                   ValueInputIteratorT,
                                                   KeyOutputIteratorT,
                                                   ValueOutputIteratorT,
                                                   EqualityOpT,
                                                   OffsetT>;

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

            // Number of input tiles
            int tile_size = Policy::BLOCK_THREADS * Policy::ITEMS_PER_THREAD;
            int num_tiles = static_cast<int>(cub::DivideAndRoundUp(num_items, tile_size));

            // Size of virtual shared memory
            int max_shmem = 0;
            if (CubDebug(
                error = cudaDeviceGetAttribute(&max_shmem,
                                               cudaDevAttrMaxSharedMemoryPerBlock,
                                               device_ordinal)))
            {
                break;
            }
            std::size_t vshmem_size = detail::VshmemSize(max_shmem, sizeof(typename UniqueByKeyAgentT::TempStorage), num_tiles);

            // Specify temporary storage allocation requirements
            size_t allocation_sizes[2] = {0, vshmem_size};
            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[2] = {NULL, NULL};
            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;
            }

            // 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
            num_tiles = CUB_MAX(1, num_tiles);
            int init_grid_size = cub::DivideAndRoundUp(num_tiles, INIT_KERNEL_THREADS);

            #ifdef CUB_DETAIL_DEBUG_ENABLE_LOG
            _CubLog("Invoking init_kernel<<<%d, %d, 0, %lld>>>()\n", init_grid_size, INIT_KERNEL_THREADS, (long long) stream);
            #endif

            // Invoke init_kernel to initialize tile descriptors
            THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron(
                init_grid_size, INIT_KERNEL_THREADS, 0, stream
            ).doit(init_kernel, tile_state, num_tiles, d_num_selected_out);

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

            // Sync the stream if specified to flush runtime errors
            error = detail::DebugSyncStream(stream);
            if (CubDebug(error))
            {
              break;
            }

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

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

            // Get grid size for scanning tiles
            dim3 scan_grid_size;
            scan_grid_size.z = 1;
            scan_grid_size.y = cub::DivideAndRoundUp(num_tiles, max_dim_x);
            scan_grid_size.x = CUB_MIN(num_tiles, max_dim_x);

            // Log select_if_kernel configuration
            #ifdef CUB_DETAIL_DEBUG_ENABLE_LOG
            {
              // Get SM occupancy for unique_by_key_kernel
              int scan_sm_occupancy;
              if (CubDebug(error = MaxSmOccupancy(scan_sm_occupancy, // out
                                                  scan_kernel,
                                                  Policy::BLOCK_THREADS)))
              {
                break;
              }

              _CubLog("Invoking unique_by_key_kernel<<<{%d,%d,%d}, %d, 0, "
                      "%lld>>>(), %d items per thread, %d SM occupancy\n",
                      scan_grid_size.x,
                      scan_grid_size.y,
                      scan_grid_size.z,
                      Policy::BLOCK_THREADS,
                      (long long)stream,
                      Policy::ITEMS_PER_THREAD,
                      scan_sm_occupancy);
            }
            #endif

            // Invoke select_if_kernel
            error = THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron(
                scan_grid_size, Policy::BLOCK_THREADS, 0, stream
            ).doit(scan_kernel,
                   d_keys_in,
                   d_values_in,
                   d_keys_out,
                   d_values_out,
                   d_num_selected_out,
                   tile_state,
                   equality_op,
                   num_items,
                   num_tiles);

            // Check for failure to launch
            if (CubDebug(error))
            {
              break;
            }

            // Sync the stream if specified to flush runtime errors
            error = detail::DebugSyncStream(stream);
            if (CubDebug(error))
            {
              break;
            }
        }
        while(0);

        return error;
    }

    template <typename ActivePolicyT>
    CUB_RUNTIME_FUNCTION __host__  __forceinline__
    cudaError_t Invoke()
    {
        using MaxPolicyT = typename DispatchUniqueByKey::MaxPolicy;

        // Ensure kernels are instantiated.
        return Invoke<ActivePolicyT>(
            DeviceCompactInitKernel<ScanTileStateT, NumSelectedIteratorT>,
            DeviceUniqueByKeySweepKernel<
                MaxPolicyT,
                KeyInputIteratorT,
                ValueInputIteratorT,
                KeyOutputIteratorT,
                ValueOutputIteratorT,
                NumSelectedIteratorT,
                ScanTileStateT,
                EqualityOpT,
                OffsetT>
        );
    }


    /**
    * 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
        KeyInputIteratorT       d_keys_in,              ///< [in] Pointer to the input sequence of keys
        ValueInputIteratorT     d_values_in,            ///< [in] Pointer to the input sequence of values
        KeyOutputIteratorT      d_keys_out,             ///< [out] Pointer to the output sequence of selected data items
        ValueOutputIteratorT    d_values_out,           ///< [out] Pointer to the output sequence of selected data items
        NumSelectedIteratorT    d_num_selected_out,     ///< [out] Pointer to the total number of items selected (i.e., length of \p d_keys_out or \p d_values_out)
        EqualityOpT             equality_op,            ///< [in] Equality operator
        OffsetT                 num_items,              ///< [in] Total number of input items (i.e., the length of \p d_in)
        cudaStream_t            stream)                 ///< [in] <b>[optional]</b> CUDA stream to launch kernels within.  Default is stream<sub>0</sub>.
    {
        using MaxPolicyT = typename DispatchUniqueByKey::MaxPolicy;

        cudaError_t error;
        do
        {
            // Get PTX version
            int ptx_version = 0;
            if (CubDebug(error = PtxVersion(ptx_version))) break;

            // Create dispatch functor
            DispatchUniqueByKey dispatch(
                d_temp_storage,
                temp_storage_bytes,
                d_keys_in,
                d_values_in,
                d_keys_out,
                d_values_out,
                d_num_selected_out,
                equality_op,
                num_items,
                stream);

            // Dispatch to chained policy
            if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;
        }
        while (0);

        return error;
    }

    CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED
    CUB_RUNTIME_FUNCTION __forceinline__
    static cudaError_t Dispatch(
        void*                   d_temp_storage,         
        size_t                  &temp_storage_bytes,    
        KeyInputIteratorT       d_keys_in,             
        ValueInputIteratorT     d_values_in,            
        KeyOutputIteratorT      d_keys_out,            
        ValueOutputIteratorT    d_values_out,         
        NumSelectedIteratorT    d_num_selected_out,  
        EqualityOpT             equality_op,        
        OffsetT                 num_items,              
        cudaStream_t            stream,                
        bool                    debug_synchronous)    
    {
      CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG

      return Dispatch(d_temp_storage,
                      temp_storage_bytes,
                      d_keys_in,
                      d_values_in,
                      d_keys_out,
                      d_values_out,
                      d_num_selected_out,
                      equality_op,
                      num_items,
                      stream);
    }
};

CUB_NAMESPACE_END