Blame view
tools/cub-1.8.0/cub/agent/agent_histogram.cuh
32.5 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 |
/******************************************************************************
* 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::AgentHistogram implements a stateful abstraction of CUDA thread blocks for participating in device-wide histogram .
*/
#pragma once
#include <iterator>
#include "../util_type.cuh"
#include "../block/block_load.cuh"
#include "../grid/grid_queue.cuh"
#include "../iterator/cache_modified_input_iterator.cuh"
#include "../util_namespace.cuh"
/// Optional outer namespace(s)
CUB_NS_PREFIX
/// CUB namespace
namespace cub {
/******************************************************************************
* Tuning policy
******************************************************************************/
/**
*
*/
enum BlockHistogramMemoryPreference
{
GMEM,
SMEM,
BLEND
};
/**
* Parameterizable tuning policy type for AgentHistogram
*/
template <
int _BLOCK_THREADS, ///< Threads per thread block
int _PIXELS_PER_THREAD, ///< Pixels per thread (per tile of input)
BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use
CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements
bool _RLE_COMPRESS, ///< Whether to perform localized RLE to compress samples before histogramming
BlockHistogramMemoryPreference _MEM_PREFERENCE, ///< Whether to prefer privatized shared-memory bins (versus privatized global-memory bins)
bool _WORK_STEALING> ///< Whether to dequeue tiles from a global work queue
struct AgentHistogramPolicy
{
enum
{
BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block
PIXELS_PER_THREAD = _PIXELS_PER_THREAD, ///< Pixels per thread (per tile of input)
IS_RLE_COMPRESS = _RLE_COMPRESS, ///< Whether to perform localized RLE to compress samples before histogramming
MEM_PREFERENCE = _MEM_PREFERENCE, ///< Whether to prefer privatized shared-memory bins (versus privatized global-memory bins)
IS_WORK_STEALING = _WORK_STEALING, ///< Whether to dequeue tiles from a global work queue
};
static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use
static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements
};
/******************************************************************************
* Thread block abstractions
******************************************************************************/
/**
* \brief AgentHistogram implements a stateful abstraction of CUDA thread blocks for participating in device-wide histogram .
*/
template <
typename AgentHistogramPolicyT, ///< Parameterized AgentHistogramPolicy tuning policy type
int PRIVATIZED_SMEM_BINS, ///< Number of privatized shared-memory histogram bins of any channel. Zero indicates privatized counters to be maintained in device-accessible memory.
int NUM_CHANNELS, ///< Number of channels interleaved in the input data. Supports up to four channels.
int NUM_ACTIVE_CHANNELS, ///< Number of channels actively being histogrammed
typename SampleIteratorT, ///< Random-access input iterator type for reading samples
typename CounterT, ///< Integer type for counting sample occurrences per histogram bin
typename PrivatizedDecodeOpT, ///< The transform operator type for determining privatized counter indices from samples, one for each channel
typename OutputDecodeOpT, ///< The transform operator type for determining output bin-ids from privatized counter indices, one for each channel
typename OffsetT, ///< Signed integer type for global offsets
int PTX_ARCH = CUB_PTX_ARCH> ///< PTX compute capability
struct AgentHistogram
{
//---------------------------------------------------------------------
// Types and constants
//---------------------------------------------------------------------
/// The sample type of the input iterator
typedef typename std::iterator_traits<SampleIteratorT>::value_type SampleT;
/// The pixel type of SampleT
typedef typename CubVector<SampleT, NUM_CHANNELS>::Type PixelT;
/// The quad type of SampleT
typedef typename CubVector<SampleT, 4>::Type QuadT;
/// Constants
enum
{
BLOCK_THREADS = AgentHistogramPolicyT::BLOCK_THREADS,
PIXELS_PER_THREAD = AgentHistogramPolicyT::PIXELS_PER_THREAD,
SAMPLES_PER_THREAD = PIXELS_PER_THREAD * NUM_CHANNELS,
QUADS_PER_THREAD = SAMPLES_PER_THREAD / 4,
TILE_PIXELS = PIXELS_PER_THREAD * BLOCK_THREADS,
TILE_SAMPLES = SAMPLES_PER_THREAD * BLOCK_THREADS,
IS_RLE_COMPRESS = AgentHistogramPolicyT::IS_RLE_COMPRESS,
MEM_PREFERENCE = (PRIVATIZED_SMEM_BINS > 0) ?
AgentHistogramPolicyT::MEM_PREFERENCE :
GMEM,
IS_WORK_STEALING = AgentHistogramPolicyT::IS_WORK_STEALING,
};
/// Cache load modifier for reading input elements
static const CacheLoadModifier LOAD_MODIFIER = AgentHistogramPolicyT::LOAD_MODIFIER;
/// Input iterator wrapper type (for applying cache modifier)
typedef typename If<IsPointer<SampleIteratorT>::VALUE,
CacheModifiedInputIterator<LOAD_MODIFIER, SampleT, OffsetT>, // Wrap the native input pointer with CacheModifiedInputIterator
SampleIteratorT>::Type // Directly use the supplied input iterator type
WrappedSampleIteratorT;
/// Pixel input iterator type (for applying cache modifier)
typedef CacheModifiedInputIterator<LOAD_MODIFIER, PixelT, OffsetT>
WrappedPixelIteratorT;
/// Qaud input iterator type (for applying cache modifier)
typedef CacheModifiedInputIterator<LOAD_MODIFIER, QuadT, OffsetT>
WrappedQuadIteratorT;
/// Parameterized BlockLoad type for samples
typedef BlockLoad<
SampleT,
BLOCK_THREADS,
SAMPLES_PER_THREAD,
AgentHistogramPolicyT::LOAD_ALGORITHM>
BlockLoadSampleT;
/// Parameterized BlockLoad type for pixels
typedef BlockLoad<
PixelT,
BLOCK_THREADS,
PIXELS_PER_THREAD,
AgentHistogramPolicyT::LOAD_ALGORITHM>
BlockLoadPixelT;
/// Parameterized BlockLoad type for quads
typedef BlockLoad<
QuadT,
BLOCK_THREADS,
QUADS_PER_THREAD,
AgentHistogramPolicyT::LOAD_ALGORITHM>
BlockLoadQuadT;
/// Shared memory type required by this thread block
struct _TempStorage
{
CounterT histograms[NUM_ACTIVE_CHANNELS][PRIVATIZED_SMEM_BINS + 1]; // Smem needed for block-privatized smem histogram (with 1 word of padding)
int tile_idx;
// Aliasable storage layout
union Aliasable
{
typename BlockLoadSampleT::TempStorage sample_load; // Smem needed for loading a tile of samples
typename BlockLoadPixelT::TempStorage pixel_load; // Smem needed for loading a tile of pixels
typename BlockLoadQuadT::TempStorage quad_load; // Smem needed for loading a tile of quads
} aliasable;
};
/// Temporary storage type (unionable)
struct TempStorage : Uninitialized<_TempStorage> {};
//---------------------------------------------------------------------
// Per-thread fields
//---------------------------------------------------------------------
/// Reference to temp_storage
_TempStorage &temp_storage;
/// Sample input iterator (with cache modifier applied, if possible)
WrappedSampleIteratorT d_wrapped_samples;
/// Native pointer for input samples (possibly NULL if unavailable)
SampleT* d_native_samples;
/// The number of output bins for each channel
int (&num_output_bins)[NUM_ACTIVE_CHANNELS];
/// The number of privatized bins for each channel
int (&num_privatized_bins)[NUM_ACTIVE_CHANNELS];
/// Reference to gmem privatized histograms for each channel
CounterT* d_privatized_histograms[NUM_ACTIVE_CHANNELS];
/// Reference to final output histograms (gmem)
CounterT* (&d_output_histograms)[NUM_ACTIVE_CHANNELS];
/// The transform operator for determining output bin-ids from privatized counter indices, one for each channel
OutputDecodeOpT (&output_decode_op)[NUM_ACTIVE_CHANNELS];
/// The transform operator for determining privatized counter indices from samples, one for each channel
PrivatizedDecodeOpT (&privatized_decode_op)[NUM_ACTIVE_CHANNELS];
/// Whether to prefer privatized smem counters vs privatized global counters
bool prefer_smem;
//---------------------------------------------------------------------
// Initialize privatized bin counters
//---------------------------------------------------------------------
// Initialize privatized bin counters
__device__ __forceinline__ void InitBinCounters(CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS])
{
// Initialize histogram bin counts to zeros
#pragma unroll
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
{
for (int privatized_bin = threadIdx.x; privatized_bin < num_privatized_bins[CHANNEL]; privatized_bin += BLOCK_THREADS)
{
privatized_histograms[CHANNEL][privatized_bin] = 0;
}
}
// Barrier to make sure all threads are done updating counters
CTA_SYNC();
}
// Initialize privatized bin counters. Specialized for privatized shared-memory counters
__device__ __forceinline__ void InitSmemBinCounters()
{
CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS];
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
privatized_histograms[CHANNEL] = temp_storage.histograms[CHANNEL];
InitBinCounters(privatized_histograms);
}
// Initialize privatized bin counters. Specialized for privatized global-memory counters
__device__ __forceinline__ void InitGmemBinCounters()
{
InitBinCounters(d_privatized_histograms);
}
//---------------------------------------------------------------------
// Update final output histograms
//---------------------------------------------------------------------
// Update final output histograms from privatized histograms
__device__ __forceinline__ void StoreOutput(CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS])
{
// Barrier to make sure all threads are done updating counters
CTA_SYNC();
// Apply privatized bin counts to output bin counts
#pragma unroll
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
{
int channel_bins = num_privatized_bins[CHANNEL];
for (int privatized_bin = threadIdx.x;
privatized_bin < channel_bins;
privatized_bin += BLOCK_THREADS)
{
int output_bin = -1;
CounterT count = privatized_histograms[CHANNEL][privatized_bin];
bool is_valid = count > 0;
output_decode_op[CHANNEL].template BinSelect<LOAD_MODIFIER>((SampleT) privatized_bin, output_bin, is_valid);
if (output_bin >= 0)
{
atomicAdd(&d_output_histograms[CHANNEL][output_bin], count);
}
}
}
}
// Update final output histograms from privatized histograms. Specialized for privatized shared-memory counters
__device__ __forceinline__ void StoreSmemOutput()
{
CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS];
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
privatized_histograms[CHANNEL] = temp_storage.histograms[CHANNEL];
StoreOutput(privatized_histograms);
}
// Update final output histograms from privatized histograms. Specialized for privatized global-memory counters
__device__ __forceinline__ void StoreGmemOutput()
{
StoreOutput(d_privatized_histograms);
}
//---------------------------------------------------------------------
// Tile accumulation
//---------------------------------------------------------------------
// Accumulate pixels. Specialized for RLE compression.
__device__ __forceinline__ void AccumulatePixels(
SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],
bool is_valid[PIXELS_PER_THREAD],
CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS],
Int2Type<true> is_rle_compress)
{
#pragma unroll
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
{
// Bin pixels
int bins[PIXELS_PER_THREAD];
#pragma unroll
for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD; ++PIXEL)
{
bins[PIXEL] = -1;
privatized_decode_op[CHANNEL].template BinSelect<LOAD_MODIFIER>(samples[PIXEL][CHANNEL], bins[PIXEL], is_valid[PIXEL]);
}
CounterT accumulator = 1;
#pragma unroll
for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD - 1; ++PIXEL)
{
if (bins[PIXEL] != bins[PIXEL + 1])
{
if (bins[PIXEL] >= 0)
atomicAdd(privatized_histograms[CHANNEL] + bins[PIXEL], accumulator);
accumulator = 0;
}
accumulator++;
}
// Last pixel
if (bins[PIXELS_PER_THREAD - 1] >= 0)
atomicAdd(privatized_histograms[CHANNEL] + bins[PIXELS_PER_THREAD - 1], accumulator);
}
}
// Accumulate pixels. Specialized for individual accumulation of each pixel.
__device__ __forceinline__ void AccumulatePixels(
SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],
bool is_valid[PIXELS_PER_THREAD],
CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS],
Int2Type<false> is_rle_compress)
{
#pragma unroll
for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD; ++PIXEL)
{
#pragma unroll
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
{
int bin = -1;
privatized_decode_op[CHANNEL].template BinSelect<LOAD_MODIFIER>(samples[PIXEL][CHANNEL], bin, is_valid[PIXEL]);
if (bin >= 0)
atomicAdd(privatized_histograms[CHANNEL] + bin, 1);
}
}
}
/**
* Accumulate pixel, specialized for smem privatized histogram
*/
__device__ __forceinline__ void AccumulateSmemPixels(
SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],
bool is_valid[PIXELS_PER_THREAD])
{
CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS];
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
privatized_histograms[CHANNEL] = temp_storage.histograms[CHANNEL];
AccumulatePixels(samples, is_valid, privatized_histograms, Int2Type<IS_RLE_COMPRESS>());
}
/**
* Accumulate pixel, specialized for gmem privatized histogram
*/
__device__ __forceinline__ void AccumulateGmemPixels(
SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],
bool is_valid[PIXELS_PER_THREAD])
{
AccumulatePixels(samples, is_valid, d_privatized_histograms, Int2Type<IS_RLE_COMPRESS>());
}
//---------------------------------------------------------------------
// Tile loading
//---------------------------------------------------------------------
// Load full, aligned tile using pixel iterator (multi-channel)
template <int _NUM_ACTIVE_CHANNELS>
__device__ __forceinline__ void LoadFullAlignedTile(
OffsetT block_offset,
int valid_samples,
SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],
Int2Type<_NUM_ACTIVE_CHANNELS> num_active_channels)
{
typedef PixelT AliasedPixels[PIXELS_PER_THREAD];
WrappedPixelIteratorT d_wrapped_pixels((PixelT*) (d_native_samples + block_offset));
// Load using a wrapped pixel iterator
BlockLoadPixelT(temp_storage.aliasable.pixel_load).Load(
d_wrapped_pixels,
reinterpret_cast<AliasedPixels&>(samples));
}
// Load full, aligned tile using quad iterator (single-channel)
__device__ __forceinline__ void LoadFullAlignedTile(
OffsetT block_offset,
int valid_samples,
SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],
Int2Type<1> num_active_channels)
{
typedef QuadT AliasedQuads[QUADS_PER_THREAD];
WrappedQuadIteratorT d_wrapped_quads((QuadT*) (d_native_samples + block_offset));
// Load using a wrapped quad iterator
BlockLoadQuadT(temp_storage.aliasable.quad_load).Load(
d_wrapped_quads,
reinterpret_cast<AliasedQuads&>(samples));
}
// Load full, aligned tile
__device__ __forceinline__ void LoadTile(
OffsetT block_offset,
int valid_samples,
SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],
Int2Type<true> is_full_tile,
Int2Type<true> is_aligned)
{
LoadFullAlignedTile(block_offset, valid_samples, samples, Int2Type<NUM_ACTIVE_CHANNELS>());
}
// Load full, mis-aligned tile using sample iterator
__device__ __forceinline__ void LoadTile(
OffsetT block_offset,
int valid_samples,
SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],
Int2Type<true> is_full_tile,
Int2Type<false> is_aligned)
{
typedef SampleT AliasedSamples[SAMPLES_PER_THREAD];
// Load using sample iterator
BlockLoadSampleT(temp_storage.aliasable.sample_load).Load(
d_wrapped_samples + block_offset,
reinterpret_cast<AliasedSamples&>(samples));
}
// Load partially-full, aligned tile using the pixel iterator
__device__ __forceinline__ void LoadTile(
OffsetT block_offset,
int valid_samples,
SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],
Int2Type<false> is_full_tile,
Int2Type<true> is_aligned)
{
typedef PixelT AliasedPixels[PIXELS_PER_THREAD];
WrappedPixelIteratorT d_wrapped_pixels((PixelT*) (d_native_samples + block_offset));
int valid_pixels = valid_samples / NUM_CHANNELS;
// Load using a wrapped pixel iterator
BlockLoadPixelT(temp_storage.aliasable.pixel_load).Load(
d_wrapped_pixels,
reinterpret_cast<AliasedPixels&>(samples),
valid_pixels);
}
// Load partially-full, mis-aligned tile using sample iterator
__device__ __forceinline__ void LoadTile(
OffsetT block_offset,
int valid_samples,
SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],
Int2Type<false> is_full_tile,
Int2Type<false> is_aligned)
{
typedef SampleT AliasedSamples[SAMPLES_PER_THREAD];
BlockLoadSampleT(temp_storage.aliasable.sample_load).Load(
d_wrapped_samples + block_offset,
reinterpret_cast<AliasedSamples&>(samples),
valid_samples);
}
//---------------------------------------------------------------------
// Tile processing
//---------------------------------------------------------------------
// Consume a tile of data samples
template <
bool IS_ALIGNED, // Whether the tile offset is aligned (quad-aligned for single-channel, pixel-aligned for multi-channel)
bool IS_FULL_TILE> // Whether the tile is full
__device__ __forceinline__ void ConsumeTile(OffsetT block_offset, int valid_samples)
{
SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS];
bool is_valid[PIXELS_PER_THREAD];
// Load tile
LoadTile(
block_offset,
valid_samples,
samples,
Int2Type<IS_FULL_TILE>(),
Int2Type<IS_ALIGNED>());
// Set valid flags
#pragma unroll
for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD; ++PIXEL)
is_valid[PIXEL] = IS_FULL_TILE || (((threadIdx.x * PIXELS_PER_THREAD + PIXEL) * NUM_CHANNELS) < valid_samples);
// Accumulate samples
#if CUB_PTX_ARCH >= 120
if (prefer_smem)
AccumulateSmemPixels(samples, is_valid);
else
AccumulateGmemPixels(samples, is_valid);
#else
AccumulateGmemPixels(samples, is_valid);
#endif
}
// Consume row tiles. Specialized for work-stealing from queue
template <bool IS_ALIGNED>
__device__ __forceinline__ void ConsumeTiles(
OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest
OffsetT num_rows, ///< The number of rows in the region of interest
OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest
int tiles_per_row, ///< Number of image tiles per row
GridQueue<int> tile_queue,
Int2Type<true> is_work_stealing)
{
int num_tiles = num_rows * tiles_per_row;
int tile_idx = (blockIdx.y * gridDim.x) + blockIdx.x;
OffsetT num_even_share_tiles = gridDim.x * gridDim.y;
while (tile_idx < num_tiles)
{
int row = tile_idx / tiles_per_row;
int col = tile_idx - (row * tiles_per_row);
OffsetT row_offset = row * row_stride_samples;
OffsetT col_offset = (col * TILE_SAMPLES);
OffsetT tile_offset = row_offset + col_offset;
if (col == tiles_per_row - 1)
{
// Consume a partially-full tile at the end of the row
OffsetT num_remaining = (num_row_pixels * NUM_CHANNELS) - col_offset;
ConsumeTile<IS_ALIGNED, false>(tile_offset, num_remaining);
}
else
{
// Consume full tile
ConsumeTile<IS_ALIGNED, true>(tile_offset, TILE_SAMPLES);
}
CTA_SYNC();
// Get next tile
if (threadIdx.x == 0)
temp_storage.tile_idx = tile_queue.Drain(1) + num_even_share_tiles;
CTA_SYNC();
tile_idx = temp_storage.tile_idx;
}
}
// Consume row tiles. Specialized for even-share (striped across thread blocks)
template <bool IS_ALIGNED>
__device__ __forceinline__ void ConsumeTiles(
OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest
OffsetT num_rows, ///< The number of rows in the region of interest
OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest
int tiles_per_row, ///< Number of image tiles per row
GridQueue<int> tile_queue,
Int2Type<false> is_work_stealing)
{
for (int row = blockIdx.y; row < num_rows; row += gridDim.y)
{
OffsetT row_begin = row * row_stride_samples;
OffsetT row_end = row_begin + (num_row_pixels * NUM_CHANNELS);
OffsetT tile_offset = row_begin + (blockIdx.x * TILE_SAMPLES);
while (tile_offset < row_end)
{
OffsetT num_remaining = row_end - tile_offset;
if (num_remaining < TILE_SAMPLES)
{
// Consume partial tile
ConsumeTile<IS_ALIGNED, false>(tile_offset, num_remaining);
break;
}
// Consume full tile
ConsumeTile<IS_ALIGNED, true>(tile_offset, TILE_SAMPLES);
tile_offset += gridDim.x * TILE_SAMPLES;
}
}
}
//---------------------------------------------------------------------
// Parameter extraction
//---------------------------------------------------------------------
// Return a native pixel pointer (specialized for CacheModifiedInputIterator types)
template <
CacheLoadModifier _MODIFIER,
typename _ValueT,
typename _OffsetT>
__device__ __forceinline__ SampleT* NativePointer(CacheModifiedInputIterator<_MODIFIER, _ValueT, _OffsetT> itr)
{
return itr.ptr;
}
// Return a native pixel pointer (specialized for other types)
template <typename IteratorT>
__device__ __forceinline__ SampleT* NativePointer(IteratorT itr)
{
return NULL;
}
//---------------------------------------------------------------------
// Interface
//---------------------------------------------------------------------
/**
* Constructor
*/
__device__ __forceinline__ AgentHistogram(
TempStorage &temp_storage, ///< Reference to temp_storage
SampleIteratorT d_samples, ///< Input data to reduce
int (&num_output_bins)[NUM_ACTIVE_CHANNELS], ///< The number bins per final output histogram
int (&num_privatized_bins)[NUM_ACTIVE_CHANNELS], ///< The number bins per privatized histogram
CounterT* (&d_output_histograms)[NUM_ACTIVE_CHANNELS], ///< Reference to final output histograms
CounterT* (&d_privatized_histograms)[NUM_ACTIVE_CHANNELS], ///< Reference to privatized histograms
OutputDecodeOpT (&output_decode_op)[NUM_ACTIVE_CHANNELS], ///< The transform operator for determining output bin-ids from privatized counter indices, one for each channel
PrivatizedDecodeOpT (&privatized_decode_op)[NUM_ACTIVE_CHANNELS]) ///< The transform operator for determining privatized counter indices from samples, one for each channel
:
temp_storage(temp_storage.Alias()),
d_wrapped_samples(d_samples),
num_output_bins(num_output_bins),
num_privatized_bins(num_privatized_bins),
d_output_histograms(d_output_histograms),
privatized_decode_op(privatized_decode_op),
output_decode_op(output_decode_op),
d_native_samples(NativePointer(d_wrapped_samples)),
prefer_smem((MEM_PREFERENCE == SMEM) ?
true : // prefer smem privatized histograms
(MEM_PREFERENCE == GMEM) ?
false : // prefer gmem privatized histograms
blockIdx.x & 1) // prefer blended privatized histograms
{
int blockId = (blockIdx.y * gridDim.x) + blockIdx.x;
// Initialize the locations of this block's privatized histograms
for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)
this->d_privatized_histograms[CHANNEL] = d_privatized_histograms[CHANNEL] + (blockId * num_privatized_bins[CHANNEL]);
}
/**
* Consume image
*/
__device__ __forceinline__ void ConsumeTiles(
OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest
OffsetT num_rows, ///< The number of rows in the region of interest
OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest
int tiles_per_row, ///< Number of image tiles per row
GridQueue<int> tile_queue) ///< Queue descriptor for assigning tiles of work to thread blocks
{
// Check whether all row starting offsets are quad-aligned (in single-channel) or pixel-aligned (in multi-channel)
int quad_mask = AlignBytes<QuadT>::ALIGN_BYTES - 1;
int pixel_mask = AlignBytes<PixelT>::ALIGN_BYTES - 1;
size_t row_bytes = sizeof(SampleT) * row_stride_samples;
bool quad_aligned_rows = (NUM_CHANNELS == 1) && (SAMPLES_PER_THREAD % 4 == 0) && // Single channel
((size_t(d_native_samples) & quad_mask) == 0) && // ptr is quad-aligned
((num_rows == 1) || ((row_bytes & quad_mask) == 0)); // number of row-samples is a multiple of the alignment of the quad
bool pixel_aligned_rows = (NUM_CHANNELS > 1) && // Multi channel
((size_t(d_native_samples) & pixel_mask) == 0) && // ptr is pixel-aligned
((row_bytes & pixel_mask) == 0); // number of row-samples is a multiple of the alignment of the pixel
// Whether rows are aligned and can be vectorized
if ((d_native_samples != NULL) && (quad_aligned_rows || pixel_aligned_rows))
ConsumeTiles<true>(num_row_pixels, num_rows, row_stride_samples, tiles_per_row, tile_queue, Int2Type<IS_WORK_STEALING>());
else
ConsumeTiles<false>(num_row_pixels, num_rows, row_stride_samples, tiles_per_row, tile_queue, Int2Type<IS_WORK_STEALING>());
}
/**
* Initialize privatized bin counters. Specialized for privatized shared-memory counters
*/
__device__ __forceinline__ void InitBinCounters()
{
if (prefer_smem)
InitSmemBinCounters();
else
InitGmemBinCounters();
}
/**
* Store privatized histogram to device-accessible memory. Specialized for privatized shared-memory counters
*/
__device__ __forceinline__ void StoreOutput()
{
if (prefer_smem)
StoreSmemOutput();
else
StoreGmemOutput();
}
};
} // CUB namespace
CUB_NS_POSTFIX // Optional outer namespace(s)
|