Blame view
tools/cub-1.8.0/cub/block/specializations/block_scan_raking.cuh
27.8 KB
8dcb6dfcb
first commit
|
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 |
/******************************************************************************
* Copyright (c) 2011, Duane Merrill. All rights reserved.
* Copyright (c) 2011-2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the
* names of its contributors may be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
* DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
******************************************************************************/
/**
* \file
* cub::BlockScanRaking provides variants of raking-based parallel prefix scan across a CUDA thread block.
*/
#pragma once
#include "../../util_ptx.cuh"
#include "../../util_arch.cuh"
#include "../../block/block_raking_layout.cuh"
#include "../../thread/thread_reduce.cuh"
#include "../../thread/thread_scan.cuh"
#include "../../warp/warp_scan.cuh"
#include "../../util_namespace.cuh"
/// Optional outer namespace(s)
CUB_NS_PREFIX
/// CUB namespace
namespace cub {
/**
* \brief BlockScanRaking provides variants of raking-based parallel prefix scan across a CUDA thread block.
*/
template <
typename T, ///< Data type being scanned
int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension
int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension
int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension
bool MEMOIZE, ///< Whether or not to buffer outer raking scan partials to incur fewer shared memory reads at the expense of higher register pressure
int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective
struct BlockScanRaking
{
//---------------------------------------------------------------------
// Types and constants
//---------------------------------------------------------------------
/// Constants
enum
{
/// The thread block size in threads
BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,
};
/// Layout type for padded thread block raking grid
typedef BlockRakingLayout<T, BLOCK_THREADS, PTX_ARCH> BlockRakingLayout;
/// Constants
enum
{
/// Number of raking threads
RAKING_THREADS = BlockRakingLayout::RAKING_THREADS,
/// Number of raking elements per warp synchronous raking thread
SEGMENT_LENGTH = BlockRakingLayout::SEGMENT_LENGTH,
/// Cooperative work can be entirely warp synchronous
WARP_SYNCHRONOUS = (BLOCK_THREADS == RAKING_THREADS),
};
/// WarpScan utility type
typedef WarpScan<T, RAKING_THREADS, PTX_ARCH> WarpScan;
/// Shared memory storage layout type
struct _TempStorage
{
typename WarpScan::TempStorage warp_scan; ///< Buffer for warp-synchronous scan
typename BlockRakingLayout::TempStorage raking_grid; ///< Padded thread block raking grid
T block_aggregate; ///< Block aggregate
};
/// Alias wrapper allowing storage to be unioned
struct TempStorage : Uninitialized<_TempStorage> {};
//---------------------------------------------------------------------
// Per-thread fields
//---------------------------------------------------------------------
// Thread fields
_TempStorage &temp_storage;
unsigned int linear_tid;
T cached_segment[SEGMENT_LENGTH];
//---------------------------------------------------------------------
// Utility methods
//---------------------------------------------------------------------
/// Templated reduction
template <int ITERATION, typename ScanOp>
__device__ __forceinline__ T GuardedReduce(
T* raking_ptr, ///< [in] Input array
ScanOp scan_op, ///< [in] Binary reduction operator
T raking_partial, ///< [in] Prefix to seed reduction with
Int2Type<ITERATION> /*iteration*/)
{
if ((BlockRakingLayout::UNGUARDED) || (((linear_tid * SEGMENT_LENGTH) + ITERATION) < BLOCK_THREADS))
{
T addend = raking_ptr[ITERATION];
raking_partial = scan_op(raking_partial, addend);
}
return GuardedReduce(raking_ptr, scan_op, raking_partial, Int2Type<ITERATION + 1>());
}
/// Templated reduction (base case)
template <typename ScanOp>
__device__ __forceinline__ T GuardedReduce(
T* /*raking_ptr*/, ///< [in] Input array
ScanOp /*scan_op*/, ///< [in] Binary reduction operator
T raking_partial, ///< [in] Prefix to seed reduction with
Int2Type<SEGMENT_LENGTH> /*iteration*/)
{
return raking_partial;
}
/// Templated copy
template <int ITERATION>
__device__ __forceinline__ void CopySegment(
T* out, ///< [out] Out array
T* in, ///< [in] Input array
Int2Type<ITERATION> /*iteration*/)
{
out[ITERATION] = in[ITERATION];
CopySegment(out, in, Int2Type<ITERATION + 1>());
}
/// Templated copy (base case)
__device__ __forceinline__ void CopySegment(
T* /*out*/, ///< [out] Out array
T* /*in*/, ///< [in] Input array
Int2Type<SEGMENT_LENGTH> /*iteration*/)
{}
/// Performs upsweep raking reduction, returning the aggregate
template <typename ScanOp>
__device__ __forceinline__ T Upsweep(
ScanOp scan_op)
{
T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
// Read data into registers
CopySegment(cached_segment, smem_raking_ptr, Int2Type<0>());
T raking_partial = cached_segment[0];
return GuardedReduce(cached_segment, scan_op, raking_partial, Int2Type<1>());
}
/// Performs exclusive downsweep raking scan
template <typename ScanOp>
__device__ __forceinline__ void ExclusiveDownsweep(
ScanOp scan_op,
T raking_partial,
bool apply_prefix = true)
{
T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
// Read data back into registers
if (!MEMOIZE)
{
CopySegment(cached_segment, smem_raking_ptr, Int2Type<0>());
}
internal::ThreadScanExclusive(cached_segment, cached_segment, scan_op, raking_partial, apply_prefix);
// Write data back to smem
CopySegment(smem_raking_ptr, cached_segment, Int2Type<0>());
}
/// Performs inclusive downsweep raking scan
template <typename ScanOp>
__device__ __forceinline__ void InclusiveDownsweep(
ScanOp scan_op,
T raking_partial,
bool apply_prefix = true)
{
T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
// Read data back into registers
if (!MEMOIZE)
{
CopySegment(cached_segment, smem_raking_ptr, Int2Type<0>());
}
internal::ThreadScanInclusive(cached_segment, cached_segment, scan_op, raking_partial, apply_prefix);
// Write data back to smem
CopySegment(smem_raking_ptr, cached_segment, Int2Type<0>());
}
//---------------------------------------------------------------------
// Constructors
//---------------------------------------------------------------------
/// Constructor
__device__ __forceinline__ BlockScanRaking(
TempStorage &temp_storage)
:
temp_storage(temp_storage.Alias()),
linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))
{}
//---------------------------------------------------------------------
// Exclusive scans
//---------------------------------------------------------------------
/// Computes an exclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element. With no initial value, the output computed for <em>thread</em><sub>0</sub> is undefined.
template <typename ScanOp>
__device__ __forceinline__ void ExclusiveScan(
T input, ///< [in] Calling thread's input item
T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op) ///< [in] Binary scan operator
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
WarpScan(temp_storage.warp_scan).ExclusiveScan(input, exclusive_output, scan_op);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Warp-synchronous scan
T exclusive_partial;
WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, scan_op);
// Exclusive raking downsweep scan
ExclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));
}
CTA_SYNC();
// Grab thread prefix from shared memory
exclusive_output = *placement_ptr;
}
}
/// Computes an exclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element.
template <typename ScanOp>
__device__ __forceinline__ void ExclusiveScan(
T input, ///< [in] Calling thread's input items
T &output, ///< [out] Calling thread's output items (may be aliased to \p input)
const T &initial_value, ///< [in] Initial value to seed the exclusive scan
ScanOp scan_op) ///< [in] Binary scan operator
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
WarpScan(temp_storage.warp_scan).ExclusiveScan(input, output, initial_value, scan_op);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Exclusive Warp-synchronous scan
T exclusive_partial;
WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, initial_value, scan_op);
// Exclusive raking downsweep scan
ExclusiveDownsweep(scan_op, exclusive_partial);
}
CTA_SYNC();
// Grab exclusive partial from shared memory
output = *placement_ptr;
}
}
/// Computes an exclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \p block_aggregate of all inputs. With no initial value, the output computed for <em>thread</em><sub>0</sub> is undefined.
template <typename ScanOp>
__device__ __forceinline__ void ExclusiveScan(
T input, ///< [in] Calling thread's input item
T &output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op, ///< [in] Binary scan operator
T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
WarpScan(temp_storage.warp_scan).ExclusiveScan(input, output, scan_op, block_aggregate);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
// Raking upsweep reduction across shared partials
T upsweep_partial= Upsweep(scan_op);
// Warp-synchronous scan
T inclusive_partial;
T exclusive_partial;
WarpScan(temp_storage.warp_scan).Scan(upsweep_partial, inclusive_partial, exclusive_partial, scan_op);
// Exclusive raking downsweep scan
ExclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));
// Broadcast aggregate to all threads
if (linear_tid == RAKING_THREADS - 1)
temp_storage.block_aggregate = inclusive_partial;
}
CTA_SYNC();
// Grab thread prefix from shared memory
output = *placement_ptr;
// Retrieve block aggregate
block_aggregate = temp_storage.block_aggregate;
}
}
/// Computes an exclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \p block_aggregate of all inputs.
template <typename ScanOp>
__device__ __forceinline__ void ExclusiveScan(
T input, ///< [in] Calling thread's input items
T &output, ///< [out] Calling thread's output items (may be aliased to \p input)
const T &initial_value, ///< [in] Initial value to seed the exclusive scan
ScanOp scan_op, ///< [in] Binary scan operator
T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
WarpScan(temp_storage.warp_scan).ExclusiveScan(input, output, initial_value, scan_op, block_aggregate);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Warp-synchronous scan
T exclusive_partial;
WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, initial_value, scan_op, block_aggregate);
// Exclusive raking downsweep scan
ExclusiveDownsweep(scan_op, exclusive_partial);
// Broadcast aggregate to other threads
if (linear_tid == 0)
temp_storage.block_aggregate = block_aggregate;
}
CTA_SYNC();
// Grab exclusive partial from shared memory
output = *placement_ptr;
// Retrieve block aggregate
block_aggregate = temp_storage.block_aggregate;
}
}
/// Computes an exclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element. the call-back functor \p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by <em>lane</em><sub>0</sub> in that warp is used as the "seed" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \p block_aggregate of all inputs.
template <
typename ScanOp,
typename BlockPrefixCallbackOp>
__device__ __forceinline__ void ExclusiveScan(
T input, ///< [in] Calling thread's input item
T &output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op, ///< [in] Binary scan operator
BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] <b>[<em>warp</em><sub>0</sub> only]</b> Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
T block_aggregate;
WarpScan warp_scan(temp_storage.warp_scan);
warp_scan.ExclusiveScan(input, output, scan_op, block_aggregate);
// Obtain warp-wide prefix in lane0, then broadcast to other lanes
T block_prefix = block_prefix_callback_op(block_aggregate);
block_prefix = warp_scan.Broadcast(block_prefix, 0);
output = scan_op(block_prefix, output);
if (linear_tid == 0)
output = block_prefix;
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
WarpScan warp_scan(temp_storage.warp_scan);
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Warp-synchronous scan
T exclusive_partial, block_aggregate;
warp_scan.ExclusiveScan(upsweep_partial, exclusive_partial, scan_op, block_aggregate);
// Obtain block-wide prefix in lane0, then broadcast to other lanes
T block_prefix = block_prefix_callback_op(block_aggregate);
block_prefix = warp_scan.Broadcast(block_prefix, 0);
// Update prefix with warpscan exclusive partial
T downsweep_prefix = scan_op(block_prefix, exclusive_partial);
if (linear_tid == 0)
downsweep_prefix = block_prefix;
// Exclusive raking downsweep scan
ExclusiveDownsweep(scan_op, downsweep_prefix);
}
CTA_SYNC();
// Grab thread prefix from shared memory
output = *placement_ptr;
}
}
//---------------------------------------------------------------------
// Inclusive scans
//---------------------------------------------------------------------
/// Computes an inclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element.
template <typename ScanOp>
__device__ __forceinline__ void InclusiveScan(
T input, ///< [in] Calling thread's input item
T &output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op) ///< [in] Binary scan operator
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
WarpScan(temp_storage.warp_scan).InclusiveScan(input, output, scan_op);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Exclusive Warp-synchronous scan
T exclusive_partial;
WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, scan_op);
// Inclusive raking downsweep scan
InclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));
}
CTA_SYNC();
// Grab thread prefix from shared memory
output = *placement_ptr;
}
}
/// Computes an inclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \p block_aggregate of all inputs.
template <typename ScanOp>
__device__ __forceinline__ void InclusiveScan(
T input, ///< [in] Calling thread's input item
T &output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op, ///< [in] Binary scan operator
T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
WarpScan(temp_storage.warp_scan).InclusiveScan(input, output, scan_op, block_aggregate);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Warp-synchronous scan
T inclusive_partial;
T exclusive_partial;
WarpScan(temp_storage.warp_scan).Scan(upsweep_partial, inclusive_partial, exclusive_partial, scan_op);
// Inclusive raking downsweep scan
InclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));
// Broadcast aggregate to all threads
if (linear_tid == RAKING_THREADS - 1)
temp_storage.block_aggregate = inclusive_partial;
}
CTA_SYNC();
// Grab thread prefix from shared memory
output = *placement_ptr;
// Retrieve block aggregate
block_aggregate = temp_storage.block_aggregate;
}
}
/// Computes an inclusive thread block-wide prefix scan using the specified binary \p scan_op functor. Each thread contributes one input element. the call-back functor \p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by <em>lane</em><sub>0</sub> in that warp is used as the "seed" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \p block_aggregate of all inputs.
template <
typename ScanOp,
typename BlockPrefixCallbackOp>
__device__ __forceinline__ void InclusiveScan(
T input, ///< [in] Calling thread's input item
T &output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op, ///< [in] Binary scan operator
BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] <b>[<em>warp</em><sub>0</sub> only]</b> Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.
{
if (WARP_SYNCHRONOUS)
{
// Short-circuit directly to warp-synchronous scan
T block_aggregate;
WarpScan warp_scan(temp_storage.warp_scan);
warp_scan.InclusiveScan(input, output, scan_op, block_aggregate);
// Obtain warp-wide prefix in lane0, then broadcast to other lanes
T block_prefix = block_prefix_callback_op(block_aggregate);
block_prefix = warp_scan.Broadcast(block_prefix, 0);
// Update prefix with exclusive warpscan partial
output = scan_op(block_prefix, output);
}
else
{
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
*placement_ptr = input;
CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS)
{
WarpScan warp_scan(temp_storage.warp_scan);
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Warp-synchronous scan
T exclusive_partial, block_aggregate;
warp_scan.ExclusiveScan(upsweep_partial, exclusive_partial, scan_op, block_aggregate);
// Obtain block-wide prefix in lane0, then broadcast to other lanes
T block_prefix = block_prefix_callback_op(block_aggregate);
block_prefix = warp_scan.Broadcast(block_prefix, 0);
// Update prefix with warpscan exclusive partial
T downsweep_prefix = scan_op(block_prefix, exclusive_partial);
if (linear_tid == 0)
downsweep_prefix = block_prefix;
// Inclusive raking downsweep scan
InclusiveDownsweep(scan_op, downsweep_prefix);
}
CTA_SYNC();
// Grab thread prefix from shared memory
output = *placement_ptr;
}
}
};
} // CUB namespace
CUB_NS_POSTFIX // Optional outer namespace(s)
|