File size: 45,772 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 | /******************************************************************************
* Copyright (c) 2011, Duane Merrill. All rights reserved.
* Copyright (c) 2011-2022, 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::DeviceReduce provides device-wide, parallel operations for
* computing a reduction across a sequence of data items residing within
* device-accessible memory.
*/
#pragma once
#include <stdio.h>
#include <iterator>
#include <cub/agent/agent_reduce.cuh>
#include <cub/config.cuh>
#include <cub/grid/grid_even_share.cuh>
#include <cub/iterator/arg_index_input_iterator.cuh>
#include <cub/thread/thread_operators.cuh>
#include <cub/thread/thread_store.cuh>
#include <cub/util_debug.cuh>
#include <cub/util_deprecated.cuh>
#include <cub/util_device.cuh>
#include <thrust/system/cuda/detail/core/triple_chevron_launch.h>
CUB_NAMESPACE_BEGIN
namespace detail
{
namespace reduce
{
/**
* All cub::DeviceReduce::* algorithms are using the same implementation. Some of them, however,
* should use initial value only for empty problems. If this struct is used as initial value with
* one of the `DeviceReduce` algorithms, the `init` value wrapped by this struct will only be used
* for empty problems; it will not be incorporated into the aggregate of non-empty problems.
*/
template <class T>
struct empty_problem_init_t
{
T init;
__host__ __device__ operator T() const { return init; }
};
/**
* @brief Applies initial value to the block aggregate and stores the result to the output iterator.
*
* @param d_out Iterator to the output aggregate
* @param reduction_op Binary reduction functor
* @param init Initial value
* @param block_aggregate Aggregate value computed by the block
*/
template <class OutputIteratorT, class ReductionOpT, class InitT, class AccumT>
__host__ __device__ void finalize_and_store_aggregate(OutputIteratorT d_out,
ReductionOpT reduction_op,
InitT init,
AccumT block_aggregate)
{
*d_out = reduction_op(init, block_aggregate);
}
/**
* @brief Ignores initial value and stores the block aggregate to the output iterator.
*
* @param d_out Iterator to the output aggregate
* @param block_aggregate Aggregate value computed by the block
*/
template <class OutputIteratorT, class ReductionOpT, class InitT, class AccumT>
__host__ __device__ void finalize_and_store_aggregate(OutputIteratorT d_out,
ReductionOpT,
empty_problem_init_t<InitT>,
AccumT block_aggregate)
{
*d_out = block_aggregate;
}
} // namespace reduce
} // namespace detail
/******************************************************************************
* Kernel entry points
*****************************************************************************/
/**
* @brief Reduce region kernel entry point (multi-block). Computes privatized
* reductions, one per thread block.
*
* @tparam ChainedPolicyT
* Chained tuning policy
*
* @tparam InputIteratorT
* Random-access input iterator type for reading input items \iterator
*
* @tparam OffsetT
* Signed integer type for global offsets
*
* @tparam ReductionOpT
* Binary reduction functor type having member
* `auto operator()(const T &a, const U &b)`
*
* @tparam InitT
* Initial value type
*
* @tparam AccumT
* Accumulator type
*
* @param[in] d_in
* Pointer to the input sequence of data items
*
* @param[out] d_out
* Pointer to the output aggregate
*
* @param[in] num_items
* Total number of input data items
*
* @param[in] even_share
* Even-share descriptor for mapping an equal number of tiles onto each
* thread block
*
* @param[in] reduction_op
* Binary reduction functor
*/
template <typename ChainedPolicyT,
typename InputIteratorT,
typename OffsetT,
typename ReductionOpT,
typename AccumT>
__launch_bounds__(int(ChainedPolicyT::ActivePolicy::ReducePolicy::BLOCK_THREADS))
__global__ void DeviceReduceKernel(InputIteratorT d_in,
AccumT* d_out,
OffsetT num_items,
GridEvenShare<OffsetT> even_share,
ReductionOpT reduction_op)
{
// Thread block type for reducing input tiles
using AgentReduceT =
AgentReduce<typename ChainedPolicyT::ActivePolicy::ReducePolicy,
InputIteratorT,
AccumT*,
OffsetT,
ReductionOpT,
AccumT>;
// Shared memory storage
__shared__ typename AgentReduceT::TempStorage temp_storage;
// Consume input tiles
AccumT block_aggregate =
AgentReduceT(temp_storage, d_in, reduction_op).ConsumeTiles(even_share);
// Output result
if (threadIdx.x == 0)
{
detail::uninitialized_copy(d_out + blockIdx.x, block_aggregate);
}
}
/**
* @brief Reduce a single tile kernel entry point (single-block). Can be used
* to aggregate privatized thread block reductions from a previous
* multi-block reduction pass.
*
* @tparam ChainedPolicyT
* Chained tuning policy
*
* @tparam InputIteratorT
* Random-access input iterator type for reading input items \iterator
*
* @tparam OutputIteratorT
* Output iterator type for recording the reduced aggregate \iterator
*
* @tparam OffsetT
* Signed integer type for global offsets
*
* @tparam ReductionOpT
* Binary reduction functor type having member
* `T operator()(const T &a, const U &b)`
*
* @tparam InitT
* Initial value type
*
* @tparam AccumT
* Accumulator type
*
* @param[in] d_in
* Pointer to the input sequence of data items
*
* @param[out] d_out
* Pointer to the output aggregate
*
* @param[in] num_items
* Total number of input data items
*
* @param[in] reduction_op
* Binary reduction functor
*
* @param[in] init
* The initial value of the reduction
*/
template <typename ChainedPolicyT,
typename InputIteratorT,
typename OutputIteratorT,
typename OffsetT,
typename ReductionOpT,
typename InitT,
typename AccumT>
__launch_bounds__(int(ChainedPolicyT::ActivePolicy::SingleTilePolicy::BLOCK_THREADS), 1)
__global__ void DeviceReduceSingleTileKernel(InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_items,
ReductionOpT reduction_op,
InitT init)
{
// Thread block type for reducing input tiles
using AgentReduceT =
AgentReduce<typename ChainedPolicyT::ActivePolicy::SingleTilePolicy,
InputIteratorT,
OutputIteratorT,
OffsetT,
ReductionOpT,
AccumT>;
// Shared memory storage
__shared__ typename AgentReduceT::TempStorage temp_storage;
// Check if empty problem
if (num_items == 0)
{
if (threadIdx.x == 0)
{
*d_out = init;
}
return;
}
// Consume input tiles
AccumT block_aggregate = AgentReduceT(temp_storage, d_in, reduction_op)
.ConsumeRange(OffsetT(0), num_items);
// Output result
if (threadIdx.x == 0)
{
detail::reduce::finalize_and_store_aggregate(d_out, reduction_op, init, block_aggregate);
}
}
/// Normalize input iterator to segment offset
template <typename T, typename OffsetT, typename IteratorT>
__device__ __forceinline__ void NormalizeReductionOutput(T & /*val*/,
OffsetT /*base_offset*/,
IteratorT /*itr*/)
{}
/// Normalize input iterator to segment offset (specialized for arg-index)
template <typename KeyValuePairT,
typename OffsetT,
typename WrappedIteratorT,
typename OutputValueT>
__device__ __forceinline__ void NormalizeReductionOutput(
KeyValuePairT &val,
OffsetT base_offset,
ArgIndexInputIterator<WrappedIteratorT, OffsetT, OutputValueT> /*itr*/)
{
val.key -= base_offset;
}
/**
* Segmented reduction (one block per segment)
* @tparam ChainedPolicyT
* Chained tuning policy
*
* @tparam InputIteratorT
* Random-access input iterator type for reading input items \iterator
*
* @tparam OutputIteratorT
* Output iterator type for recording the reduced aggregate \iterator
*
* @tparam BeginOffsetIteratorT
* Random-access input iterator type for reading segment beginning offsets
* \iterator
*
* @tparam EndOffsetIteratorT
* Random-access input iterator type for reading segment ending offsets
* \iterator
*
* @tparam OffsetT
* Signed integer type for global offsets
*
* @tparam ReductionOpT
* Binary reduction functor type having member
* `T operator()(const T &a, const U &b)`
*
* @tparam InitT
* Initial value type
*
* @param[in] d_in
* Pointer to the input sequence of data items
*
* @param[out] d_out
* Pointer to the output aggregate
*
* @param[in] d_begin_offsets
* Random-access input iterator to the sequence of beginning offsets of
* length `num_segments`, such that `d_begin_offsets[i]` is the first element
* of the *i*<sup>th</sup> data segment in `d_keys_*` and `d_values_*`
*
* @param[in] d_end_offsets
* Random-access input iterator to the sequence of ending offsets of length
* `num_segments`, such that `d_end_offsets[i] - 1` is the last element of
* the *i*<sup>th</sup> data segment in `d_keys_*` and `d_values_*`.
* If `d_end_offsets[i] - 1 <= d_begin_offsets[i]`, the *i*<sup>th</sup> is
* considered empty.
*
* @param[in] num_segments
* The number of segments that comprise the sorting data
*
* @param[in] reduction_op
* Binary reduction functor
*
* @param[in] init
* The initial value of the reduction
*/
template <typename ChainedPolicyT,
typename InputIteratorT,
typename OutputIteratorT,
typename BeginOffsetIteratorT,
typename EndOffsetIteratorT,
typename OffsetT,
typename ReductionOpT,
typename InitT,
typename AccumT>
__launch_bounds__(int(ChainedPolicyT::ActivePolicy::ReducePolicy::BLOCK_THREADS))
__global__ void DeviceSegmentedReduceKernel(
InputIteratorT d_in,
OutputIteratorT d_out,
BeginOffsetIteratorT d_begin_offsets,
EndOffsetIteratorT d_end_offsets,
int /*num_segments*/,
ReductionOpT reduction_op,
InitT init)
{
// Thread block type for reducing input tiles
using AgentReduceT =
AgentReduce<typename ChainedPolicyT::ActivePolicy::ReducePolicy,
InputIteratorT,
OutputIteratorT,
OffsetT,
ReductionOpT,
AccumT>;
// Shared memory storage
__shared__ typename AgentReduceT::TempStorage temp_storage;
OffsetT segment_begin = d_begin_offsets[blockIdx.x];
OffsetT segment_end = d_end_offsets[blockIdx.x];
// Check if empty problem
if (segment_begin == segment_end)
{
if (threadIdx.x == 0)
{
*(d_out + blockIdx.x) = init;
}
return;
}
// Consume input tiles
AccumT block_aggregate = AgentReduceT(temp_storage, d_in, reduction_op)
.ConsumeRange(segment_begin, segment_end);
// Normalize as needed
NormalizeReductionOutput(block_aggregate, segment_begin, d_in);
if (threadIdx.x == 0)
{
detail::reduce::finalize_and_store_aggregate(d_out + blockIdx.x,
reduction_op,
init,
block_aggregate);
}
}
/******************************************************************************
* Policy
******************************************************************************/
/**
* @tparam AccumT
* Accumulator data type
*
* OffsetT
* Signed integer type for global offsets
*
* ReductionOpT
* Binary reduction functor type having member
* `auto operator()(const T &a, const U &b)`
*/
template <
typename AccumT,
typename OffsetT,
typename ReductionOpT>
struct DeviceReducePolicy
{
//---------------------------------------------------------------------------
// Architecture-specific tuning policies
//---------------------------------------------------------------------------
/// SM30
struct Policy300 : ChainedPolicy<300, Policy300, Policy300>
{
static constexpr int threads_per_block = 256;
static constexpr int items_per_thread = 20;
static constexpr int items_per_vec_load = 2;
// ReducePolicy (GTX670: 154.0 @ 48M 4B items)
using ReducePolicy = AgentReducePolicy<threads_per_block,
items_per_thread,
AccumT,
items_per_vec_load,
BLOCK_REDUCE_WARP_REDUCTIONS,
LOAD_DEFAULT>;
// SingleTilePolicy
using SingleTilePolicy = ReducePolicy;
// SegmentedReducePolicy
using SegmentedReducePolicy = ReducePolicy;
};
/// SM35
struct Policy350 : ChainedPolicy<350, Policy350, Policy300>
{
static constexpr int threads_per_block = 256;
static constexpr int items_per_thread = 20;
static constexpr int items_per_vec_load = 4;
// ReducePolicy (GTX Titan: 255.1 GB/s @ 48M 4B items; 228.7 GB/s @ 192M 1B
// items)
using ReducePolicy = AgentReducePolicy<threads_per_block,
items_per_thread,
AccumT,
items_per_vec_load,
BLOCK_REDUCE_WARP_REDUCTIONS,
LOAD_LDG>;
// SingleTilePolicy
using SingleTilePolicy = ReducePolicy;
// SegmentedReducePolicy
using SegmentedReducePolicy = ReducePolicy;
};
/// SM60
struct Policy600 : ChainedPolicy<600, Policy600, Policy350>
{
static constexpr int threads_per_block = 256;
static constexpr int items_per_thread = 16;
static constexpr int items_per_vec_load = 4;
// ReducePolicy (P100: 591 GB/s @ 64M 4B items; 583 GB/s @ 256M 1B items)
using ReducePolicy = AgentReducePolicy<threads_per_block,
items_per_thread,
AccumT,
items_per_vec_load,
BLOCK_REDUCE_WARP_REDUCTIONS,
LOAD_LDG>;
// SingleTilePolicy
using SingleTilePolicy = ReducePolicy;
// SegmentedReducePolicy
using SegmentedReducePolicy = ReducePolicy;
};
using MaxPolicy = Policy600;
};
/******************************************************************************
* Single-problem dispatch
*****************************************************************************/
/**
* @brief Utility class for dispatching the appropriately-tuned kernels for
* device-wide reduction
*
* @tparam InputIteratorT
* Random-access input iterator type for reading input items \iterator
*
* @tparam OutputIteratorT
* Output iterator type for recording the reduced aggregate \iterator
*
* @tparam OffsetT
* Signed integer type for global offsets
*
* @tparam ReductionOpT
* Binary reduction functor type having member
* `auto operator()(const T &a, const U &b)`
*
* @tparam InitT
* Initial value type
*/
template <
typename InputIteratorT,
typename OutputIteratorT,
typename OffsetT,
typename ReductionOpT,
typename InitT =
cub::detail::non_void_value_t<
OutputIteratorT,
cub::detail::value_t<InputIteratorT>>,
typename AccumT =
detail::accumulator_t<
ReductionOpT,
InitT,
cub::detail::value_t<InputIteratorT>>,
typename SelectedPolicy = DeviceReducePolicy<AccumT, OffsetT, ReductionOpT>>
struct DispatchReduce : SelectedPolicy
{
//---------------------------------------------------------------------------
// Problem state
//---------------------------------------------------------------------------
/// Device-accessible allocation of temporary storage. When `nullptr`, the
/// required allocation size is written to `temp_storage_bytes` and no work
/// is done.
void *d_temp_storage;
/// Reference to size in bytes of `d_temp_storage` allocation
size_t &temp_storage_bytes;
/// Pointer to the input sequence of data items
InputIteratorT d_in;
/// Pointer to the output aggregate
OutputIteratorT d_out;
/// Total number of input items (i.e., length of `d_in`)
OffsetT num_items;
/// Binary reduction functor
ReductionOpT reduction_op;
/// The initial value of the reduction
InitT init;
/// CUDA stream to launch kernels within. Default is stream<sub>0</sub>.
cudaStream_t stream;
int ptx_version;
//---------------------------------------------------------------------------
// Constructor
//---------------------------------------------------------------------------
/// Constructor
CUB_RUNTIME_FUNCTION __forceinline__ DispatchReduce(void *d_temp_storage,
size_t &temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_items,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream,
int ptx_version)
: d_temp_storage(d_temp_storage)
, temp_storage_bytes(temp_storage_bytes)
, d_in(d_in)
, d_out(d_out)
, num_items(num_items)
, reduction_op(reduction_op)
, init(init)
, stream(stream)
, ptx_version(ptx_version)
{}
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED
CUB_RUNTIME_FUNCTION __forceinline__
DispatchReduce(void* d_temp_storage,
size_t &temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_items,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream,
bool debug_synchronous,
int ptx_version)
: d_temp_storage(d_temp_storage)
, temp_storage_bytes(temp_storage_bytes)
, d_in(d_in)
, d_out(d_out)
, num_items(num_items)
, reduction_op(reduction_op)
, init(init)
, stream(stream)
, ptx_version(ptx_version)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
}
//---------------------------------------------------------------------------
// Small-problem (single tile) invocation
//---------------------------------------------------------------------------
/**
* @brief Invoke a single block block to reduce in-core
*
* @tparam ActivePolicyT
* Umbrella policy active for the target device
*
* @tparam SingleTileKernelT
* Function type of cub::DeviceReduceSingleTileKernel
*
* @param[in] single_tile_kernel
* Kernel function pointer to parameterization of
* cub::DeviceReduceSingleTileKernel
*/
template <typename ActivePolicyT, typename SingleTileKernelT>
CUB_RUNTIME_FUNCTION __forceinline__ cudaError_t
InvokeSingleTile(SingleTileKernelT single_tile_kernel)
{
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;
}
// Log single_reduce_sweep_kernel configuration
#ifdef CUB_DETAIL_DEBUG_ENABLE_LOG
_CubLog("Invoking DeviceReduceSingleTileKernel<<<1, %d, 0, %lld>>>(), "
"%d items per thread\n",
ActivePolicyT::SingleTilePolicy::BLOCK_THREADS,
(long long)stream,
ActivePolicyT::SingleTilePolicy::ITEMS_PER_THREAD);
#endif
// Invoke single_reduce_sweep_kernel
THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron(
1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, 0, stream)
.doit(single_tile_kernel, d_in, d_out, num_items, reduction_op, init);
// 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;
}
} while (0);
return error;
}
//---------------------------------------------------------------------------
// Normal problem size invocation (two-pass)
//---------------------------------------------------------------------------
/**
* @brief Invoke two-passes to reduce
* @tparam ActivePolicyT
* Umbrella policy active for the target device
*
* @tparam ReduceKernelT
* Function type of cub::DeviceReduceKernel
*
* @tparam SingleTileKernelT
* Function type of cub::DeviceReduceSingleTileKernel
*
* @param[in] reduce_kernel
* Kernel function pointer to parameterization of cub::DeviceReduceKernel
*
* @param[in] single_tile_kernel
* Kernel function pointer to parameterization of
* cub::DeviceReduceSingleTileKernel
*/
template <typename ActivePolicyT,
typename ReduceKernelT,
typename SingleTileKernelT>
CUB_RUNTIME_FUNCTION __forceinline__ cudaError_t
InvokePasses(ReduceKernelT reduce_kernel,
SingleTileKernelT single_tile_kernel)
{
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;
}
// Init regular kernel configuration
KernelConfig reduce_config;
if (CubDebug(
error = reduce_config.Init<typename ActivePolicyT::ReducePolicy>(
reduce_kernel)))
{
break;
}
int reduce_device_occupancy = reduce_config.sm_occupancy * sm_count;
// Even-share work distribution
int max_blocks = reduce_device_occupancy * CUB_SUBSCRIPTION_FACTOR(0);
GridEvenShare<OffsetT> even_share;
even_share.DispatchInit(num_items, max_blocks, reduce_config.tile_size);
// Temporary storage allocation requirements
void *allocations[1] = {};
size_t allocation_sizes[1] = {
max_blocks * sizeof(AccumT) // bytes needed for privatized block
// reductions
};
// Alias the temporary allocations from the single storage blob (or
// compute the necessary size of the blob)
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
return cudaSuccess;
}
// Alias the allocation for the privatized per-block reductions
AccumT *d_block_reductions = (AccumT *)allocations[0];
// Get grid size for device_reduce_sweep_kernel
int reduce_grid_size = even_share.grid_size;
// Log device_reduce_sweep_kernel configuration
#ifdef CUB_DETAIL_DEBUG_ENABLE_LOG
_CubLog("Invoking DeviceReduceKernel<<<%d, %d, 0, %lld>>>(), %d items "
"per thread, %d SM occupancy\n",
reduce_grid_size,
ActivePolicyT::ReducePolicy::BLOCK_THREADS,
(long long)stream,
ActivePolicyT::ReducePolicy::ITEMS_PER_THREAD,
reduce_config.sm_occupancy);
#endif
// Invoke DeviceReduceKernel
THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron(
reduce_grid_size,
ActivePolicyT::ReducePolicy::BLOCK_THREADS,
0,
stream)
.doit(reduce_kernel,
d_in,
d_block_reductions,
num_items,
even_share,
reduction_op);
// 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;
}
// Log single_reduce_sweep_kernel configuration
#ifdef CUB_DETAIL_DEBUG_ENABLE_LOG
_CubLog("Invoking DeviceReduceSingleTileKernel<<<1, %d, 0, %lld>>>(), "
"%d items per thread\n",
ActivePolicyT::SingleTilePolicy::BLOCK_THREADS,
(long long)stream,
ActivePolicyT::SingleTilePolicy::ITEMS_PER_THREAD);
#endif
// Invoke DeviceReduceSingleTileKernel
THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron(
1,
ActivePolicyT::SingleTilePolicy::BLOCK_THREADS,
0,
stream)
.doit(single_tile_kernel,
d_block_reductions,
d_out,
reduce_grid_size, // triple_chevron is not type safe, make sure to use int
reduction_op,
init);
// 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;
}
} while (0);
return error;
}
//---------------------------------------------------------------------------
// Chained policy invocation
//---------------------------------------------------------------------------
/// Invocation
template <typename ActivePolicyT>
CUB_RUNTIME_FUNCTION __forceinline__ cudaError_t Invoke()
{
typedef typename ActivePolicyT::SingleTilePolicy SingleTilePolicyT;
typedef typename DispatchReduce::MaxPolicy MaxPolicyT;
// 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>(
DeviceReduceSingleTileKernel<MaxPolicyT,
InputIteratorT,
OutputIteratorT,
OffsetT,
ReductionOpT,
InitT,
AccumT>);
}
else
{
// Regular size
return InvokePasses<ActivePolicyT>(
DeviceReduceKernel<typename DispatchReduce::MaxPolicy,
InputIteratorT,
OffsetT,
ReductionOpT,
AccumT>,
DeviceReduceSingleTileKernel<MaxPolicyT,
AccumT *,
OutputIteratorT,
int, // Always used with int offsets
ReductionOpT,
InitT,
AccumT>);
}
}
//---------------------------------------------------------------------------
// Dispatch entrypoints
//---------------------------------------------------------------------------
/**
* @brief Internal dispatch routine for computing a device-wide reduction
*
* @param[in] d_temp_storage
* Device-accessible allocation of temporary storage. When `nullptr`, the
* required allocation size is written to `temp_storage_bytes` and no work
* is done.
*
* @param[in,out] temp_storage_bytes
* Reference to size in bytes of `d_temp_storage` allocation
*
* @param[in] d_in
* Pointer to the input sequence of data items
*
* @param[out] d_out
* Pointer to the output aggregate
*
* @param[in] num_items
* Total number of input items (i.e., length of `d_in`)
*
* @param[in] reduction_op
* Binary reduction functor
*
* @param[in] init
* The initial value of the reduction
*
* @param[in] stream
* **[optional]** CUDA stream to launch kernels within.
* Default is stream<sub>0</sub>.
*/
CUB_RUNTIME_FUNCTION __forceinline__ static cudaError_t
Dispatch(void *d_temp_storage,
size_t &temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_items,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream)
{
typedef typename DispatchReduce::MaxPolicy MaxPolicyT;
cudaError error = cudaSuccess;
do
{
// Get PTX version
int ptx_version = 0;
if (CubDebug(error = PtxVersion(ptx_version)))
{
break;
}
// Create dispatch functor
DispatchReduce dispatch(d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
num_items,
reduction_op,
init,
stream,
ptx_version);
// 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,
InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_items,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return Dispatch(d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
num_items,
reduction_op,
init,
stream);
}
};
/******************************************************************************
* Segmented dispatch
*****************************************************************************/
/**
* @brief Utility class for dispatching the appropriately-tuned kernels for
* device-wide reduction
*
* @tparam InputIteratorT
* Random-access input iterator type for reading input items \iterator
*
* @tparam OutputIteratorT
* Output iterator type for recording the reduced aggregate \iterator
*
* @tparam BeginOffsetIteratorT
* Random-access input iterator type for reading segment beginning offsets
* \iterator
*
* @tparam EndOffsetIteratorT
* Random-access input iterator type for reading segment ending offsets
* \iterator
*
* @tparam OffsetT
* Signed integer type for global offsets
*
* @tparam ReductionOpT
* Binary reduction functor type having member
* `auto operator()(const T &a, const U &b)`
*
* @tparam InitT
* value type
*/
template <
typename InputIteratorT,
typename OutputIteratorT,
typename BeginOffsetIteratorT,
typename EndOffsetIteratorT,
typename OffsetT,
typename ReductionOpT,
typename InitT =
cub::detail::non_void_value_t<
OutputIteratorT,
cub::detail::value_t<InputIteratorT>>,
typename AccumT =
detail::accumulator_t<
ReductionOpT,
InitT,
cub::detail::value_t<InputIteratorT>>,
typename SelectedPolicy = DeviceReducePolicy<AccumT, OffsetT, ReductionOpT>>
struct DispatchSegmentedReduce : SelectedPolicy
{
//---------------------------------------------------------------------------
// Problem state
//---------------------------------------------------------------------------
/// Device-accessible allocation of temporary storage. When `nullptr`, the
/// required allocation size is written to `temp_storage_bytes` and no work
/// is done.
void *d_temp_storage;
/// Reference to size in bytes of `d_temp_storage` allocation
size_t &temp_storage_bytes;
/// Pointer to the input sequence of data items
InputIteratorT d_in;
/// Pointer to the output aggregate
OutputIteratorT d_out;
/// The number of segments that comprise the sorting data
OffsetT num_segments;
/// Random-access input iterator to the sequence of beginning offsets of
/// length `num_segments`, such that `d_begin_offsets[i]` is the first
/// element of the *i*<sup>th</sup> data segment in `d_keys_*` and
/// `d_values_*`
BeginOffsetIteratorT d_begin_offsets;
/// Random-access input iterator to the sequence of ending offsets of length
/// `num_segments`, such that `d_end_offsets[i] - 1` is the last element of
/// the *i*<sup>th</sup> data segment in `d_keys_*` and `d_values_*`.
/// If `d_end_offsets[i] - 1 <= d_begin_offsets[i]`, the *i*<sup>th</sup> is
/// considered empty.
EndOffsetIteratorT d_end_offsets;
/// Binary reduction functor
ReductionOpT reduction_op;
/// The initial value of the reduction
InitT init;
/// CUDA stream to launch kernels within. Default is stream<sub>0</sub>.
cudaStream_t stream;
int ptx_version;
//---------------------------------------------------------------------------
// Constructor
//---------------------------------------------------------------------------
/// Constructor
CUB_RUNTIME_FUNCTION __forceinline__
DispatchSegmentedReduce(void *d_temp_storage,
size_t &temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_segments,
BeginOffsetIteratorT d_begin_offsets,
EndOffsetIteratorT d_end_offsets,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream,
int ptx_version)
: d_temp_storage(d_temp_storage)
, temp_storage_bytes(temp_storage_bytes)
, d_in(d_in)
, d_out(d_out)
, num_segments(num_segments)
, d_begin_offsets(d_begin_offsets)
, d_end_offsets(d_end_offsets)
, reduction_op(reduction_op)
, init(init)
, stream(stream)
, ptx_version(ptx_version)
{}
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED
CUB_RUNTIME_FUNCTION __forceinline__
DispatchSegmentedReduce(void *d_temp_storage,
size_t &temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
OffsetT num_segments,
BeginOffsetIteratorT d_begin_offsets,
EndOffsetIteratorT d_end_offsets,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream,
bool debug_synchronous,
int ptx_version)
: d_temp_storage(d_temp_storage)
, temp_storage_bytes(temp_storage_bytes)
, d_in(d_in)
, d_out(d_out)
, num_segments(num_segments)
, d_begin_offsets(d_begin_offsets)
, d_end_offsets(d_end_offsets)
, reduction_op(reduction_op)
, init(init)
, stream(stream)
, ptx_version(ptx_version)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
}
//---------------------------------------------------------------------------
// Chained policy invocation
//---------------------------------------------------------------------------
/**
* @brief Invocation
*
* @tparam ActivePolicyT
* Umbrella policy active for the target device
*
* @tparam DeviceSegmentedReduceKernelT
* Function type of cub::DeviceSegmentedReduceKernel
*
* @param[in] segmented_reduce_kernel
* Kernel function pointer to parameterization of
* cub::DeviceSegmentedReduceKernel
*/
template <typename ActivePolicyT, typename DeviceSegmentedReduceKernelT>
CUB_RUNTIME_FUNCTION __forceinline__ cudaError_t
InvokePasses(DeviceSegmentedReduceKernelT segmented_reduce_kernel)
{
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;
return cudaSuccess;
}
// Init kernel configuration
KernelConfig segmented_reduce_config;
if (CubDebug(
error = segmented_reduce_config
.Init<typename ActivePolicyT::SegmentedReducePolicy>(
segmented_reduce_kernel)))
{
break;
}
// Log device_reduce_sweep_kernel configuration
#ifdef CUB_DETAIL_DEBUG_ENABLE_LOG
_CubLog("Invoking SegmentedDeviceReduceKernel<<<%d, %d, 0, %lld>>>(), "
"%d items per thread, %d SM occupancy\n",
num_segments,
ActivePolicyT::SegmentedReducePolicy::BLOCK_THREADS,
(long long)stream,
ActivePolicyT::SegmentedReducePolicy::ITEMS_PER_THREAD,
segmented_reduce_config.sm_occupancy);
#endif
// Invoke DeviceReduceKernel
THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron(
num_segments,
ActivePolicyT::SegmentedReducePolicy::BLOCK_THREADS,
0,
stream)
.doit(segmented_reduce_kernel,
d_in,
d_out,
d_begin_offsets,
d_end_offsets,
num_segments,
reduction_op,
init);
// 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;
}
} while (0);
return error;
}
/// Invocation
template <typename ActivePolicyT>
CUB_RUNTIME_FUNCTION __forceinline__ cudaError_t Invoke()
{
typedef typename DispatchSegmentedReduce::MaxPolicy MaxPolicyT;
// Force kernel code-generation in all compiler passes
return InvokePasses<ActivePolicyT>(
DeviceSegmentedReduceKernel<MaxPolicyT,
InputIteratorT,
OutputIteratorT,
BeginOffsetIteratorT,
EndOffsetIteratorT,
OffsetT,
ReductionOpT,
InitT,
AccumT>);
}
//---------------------------------------------------------------------------
// Dispatch entrypoints
//---------------------------------------------------------------------------
/**
* @brief Internal dispatch routine for computing a device-wide reduction
*
* @param[in] d_temp_storage
* Device-accessible allocation of temporary storage. When `nullptr`, the
* required allocation size is written to `temp_storage_bytes` and no work
* is done.
*
* @param[in,out] temp_storage_bytes
* Reference to size in bytes of `d_temp_storage` allocation
*
* @param[in] d_in
* Pointer to the input sequence of data items
*
* @param[out] d_out
* Pointer to the output aggregate
*
* @param[in] num_segments
* The number of segments that comprise the sorting data
*
* @param[in] d_begin_offsets
* Random-access input iterator to the sequence of beginning offsets of
* length `num_segments`, such that `d_begin_offsets[i]` is the first
* element of the *i*<sup>th</sup> data segment in `d_keys_*` and
* `d_values_*`
*
* @param[in] d_end_offsets
* Random-access input iterator to the sequence of ending offsets of length
* `num_segments`, such that `d_end_offsets[i] - 1` is the last element of
* the *i*<sup>th</sup> data segment in `d_keys_*` and `d_values_*`.
* If `d_end_offsets[i] - 1 <= d_begin_offsets[i]`, the *i*<sup>th</sup> is
* considered empty.
*
* @param[in] reduction_op
* Binary reduction functor
*
* @param[in] init
* The initial value of the reduction
*
* @param[in] stream
* **[optional]** CUDA stream to launch kernels within.
* Default is stream<sub>0</sub>.
*/
CUB_RUNTIME_FUNCTION __forceinline__ static cudaError_t
Dispatch(void *d_temp_storage,
size_t &temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_segments,
BeginOffsetIteratorT d_begin_offsets,
EndOffsetIteratorT d_end_offsets,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream)
{
typedef typename DispatchSegmentedReduce::MaxPolicy MaxPolicyT;
if (num_segments <= 0)
{
return cudaSuccess;
}
cudaError error = cudaSuccess;
do
{
// Get PTX version
int ptx_version = 0;
if (CubDebug(error = PtxVersion(ptx_version)))
{
break;
}
// Create dispatch functor
DispatchSegmentedReduce dispatch(d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
num_segments,
d_begin_offsets,
d_end_offsets,
reduction_op,
init,
stream,
ptx_version);
// 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,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_segments,
BeginOffsetIteratorT d_begin_offsets,
EndOffsetIteratorT d_end_offsets,
ReductionOpT reduction_op,
InitT init,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return Dispatch(d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
num_segments,
d_begin_offsets,
d_end_offsets,
reduction_op,
init,
stream);
}
};
CUB_NAMESPACE_END
|