Blame view
tools/cub-1.8.0/experimental/spmv_compare.cu
29.9 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 |
/******************************************************************************
* 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 LIAeBILITY, 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.
*
******************************************************************************/
//---------------------------------------------------------------------
// SpMV comparison tool
//---------------------------------------------------------------------
#include <stdio.h>
#include <map>
#include <vector>
#include <algorithm>
#include <cstdio>
#include <fstream>
#include <cusparse.h>
#include "sparse_matrix.h"
// Ensure printing of CUDA runtime errors to console
#define CUB_STDERR
#include <cub/device/device_spmv.cuh>
#include <cub/util_allocator.cuh>
#include <cub/iterator/tex_ref_input_iterator.cuh>
#include <test/test_util.h>
using namespace cub;
//---------------------------------------------------------------------
// Globals, constants, and type declarations
//---------------------------------------------------------------------
bool g_quiet = false; // Whether to display stats in CSV format
bool g_verbose = false; // Whether to display output to console
bool g_verbose2 = false; // Whether to display input to console
CachingDeviceAllocator g_allocator(true); // Caching allocator for device memory
//---------------------------------------------------------------------
// SpMV verification
//---------------------------------------------------------------------
// Compute reference SpMV y = Ax
template <
typename ValueT,
typename OffsetT>
void SpmvGold(
CsrMatrix<ValueT, OffsetT>& a,
ValueT* vector_x,
ValueT* vector_y_in,
ValueT* vector_y_out,
ValueT alpha,
ValueT beta)
{
for (OffsetT row = 0; row < a.num_rows; ++row)
{
ValueT partial = beta * vector_y_in[row];
for (
OffsetT offset = a.row_offsets[row];
offset < a.row_offsets[row + 1];
++offset)
{
partial += alpha * a.values[offset] * vector_x[a.column_indices[offset]];
}
vector_y_out[row] = partial;
}
}
//---------------------------------------------------------------------
// GPU I/O proxy
//---------------------------------------------------------------------
/**
* Read every matrix nonzero value, read every corresponding vector value
*/
template <
int BLOCK_THREADS,
int ITEMS_PER_THREAD,
typename ValueT,
typename OffsetT,
typename VectorItr>
__launch_bounds__ (int(BLOCK_THREADS))
__global__ void NonZeroIoKernel(
SpmvParams<ValueT, OffsetT> params,
VectorItr d_vector_x)
{
enum
{
TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,
};
ValueT nonzero = 0.0;
int tile_idx = blockIdx.x;
OffsetT block_offset = tile_idx * TILE_ITEMS;
OffsetT column_indices[ITEMS_PER_THREAD];
ValueT values[ITEMS_PER_THREAD];
#pragma unroll
for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)
{
OffsetT nonzero_idx = block_offset + (ITEM * BLOCK_THREADS) + threadIdx.x;
OffsetT* ci = params.d_column_indices + nonzero_idx;
ValueT*a = params.d_values + nonzero_idx;
column_indices[ITEM] = (nonzero_idx < params.num_nonzeros) ? *ci : 0;
values[ITEM] = (nonzero_idx < params.num_nonzeros) ? *a : 0.0;
}
__syncthreads();
// Read vector
#pragma unroll
for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)
{
ValueT vector_value = ThreadLoad<LOAD_LDG>(params.d_vector_x + column_indices[ITEM]);
nonzero += vector_value * values[ITEM];
}
__syncthreads();
if (block_offset < params.num_rows)
{
#pragma unroll
for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)
{
OffsetT row_idx = block_offset + (ITEM * BLOCK_THREADS) + threadIdx.x;
if (row_idx < params.num_rows)
{
OffsetT row_end_offset = ThreadLoad<LOAD_DEFAULT>(params.d_row_end_offsets + row_idx);
if ((row_end_offset >= 0) && (nonzero == nonzero))
params.d_vector_y[row_idx] = nonzero;
}
}
}
}
/**
* Run GPU I/O proxy
*/
template <
typename ValueT,
typename OffsetT>
float TestGpuCsrIoProxy(
SpmvParams<ValueT, OffsetT>& params,
int timing_iterations)
{
enum {
BLOCK_THREADS = 128,
ITEMS_PER_THREAD = 7,
TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD,
};
// size_t smem = 1024 * 16;
size_t smem = 1024 * 0;
unsigned int nonzero_blocks = (params.num_nonzeros + TILE_SIZE - 1) / TILE_SIZE;
unsigned int row_blocks = (params.num_rows + TILE_SIZE - 1) / TILE_SIZE;
unsigned int blocks = std::max(nonzero_blocks, row_blocks);
typedef TexRefInputIterator<ValueT, 1234, int> TexItr;
TexItr x_itr;
CubDebugExit(x_itr.BindTexture(params.d_vector_x));
// Get device ordinal
int device_ordinal;
CubDebugExit(cudaGetDevice(&device_ordinal));
// Get device SM version
int sm_version;
CubDebugExit(SmVersion(sm_version, device_ordinal));
void (*kernel)(SpmvParams<ValueT, OffsetT>, TexItr) = NonZeroIoKernel<BLOCK_THREADS, ITEMS_PER_THREAD>;
int spmv_sm_occupancy;
CubDebugExit(MaxSmOccupancy(spmv_sm_occupancy, kernel, BLOCK_THREADS, smem));
if (!g_quiet)
printf("NonZeroIoKernel<%d,%d><<<%d, %d>>>, sm occupancy %d
", BLOCK_THREADS, ITEMS_PER_THREAD, blocks, BLOCK_THREADS, spmv_sm_occupancy);
// Warmup
NonZeroIoKernel<BLOCK_THREADS, ITEMS_PER_THREAD><<<blocks, BLOCK_THREADS, smem>>>(params, x_itr);
// Check for failures
CubDebugExit(cudaPeekAtLastError());
CubDebugExit(SyncStream(0));
// Timing
GpuTimer timer;
float elapsed_millis = 0.0;
timer.Start();
for (int it = 0; it < timing_iterations; ++it)
{
NonZeroIoKernel<BLOCK_THREADS, ITEMS_PER_THREAD><<<blocks, BLOCK_THREADS, smem>>>(params, x_itr);
}
timer.Stop();
elapsed_millis += timer.ElapsedMillis();
CubDebugExit(x_itr.UnbindTexture());
return elapsed_millis / timing_iterations;
}
//---------------------------------------------------------------------
// cuSparse HybMV
//---------------------------------------------------------------------
/**
* Run cuSparse HYB SpMV (specialized for fp32)
*/
template <
typename OffsetT>
float TestCusparseHybmv(
float* vector_y_in,
float* reference_vector_y_out,
SpmvParams<float, OffsetT>& params,
int timing_iterations,
cusparseHandle_t cusparse)
{
CpuTimer cpu_timer;
cpu_timer.Start();
// Construct Hyb matrix
cusparseMatDescr_t mat_desc;
cusparseHybMat_t hyb_desc;
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&mat_desc));
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateHybMat(&hyb_desc));
cusparseStatus_t status = cusparseScsr2hyb(
cusparse,
params.num_rows, params.num_cols,
mat_desc,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
hyb_desc,
0,
CUSPARSE_HYB_PARTITION_AUTO);
AssertEquals(CUSPARSE_STATUS_SUCCESS, status);
cudaDeviceSynchronize();
cpu_timer.Stop();
float elapsed_millis = cpu_timer.ElapsedMillis();
printf("HYB setup ms, %.5f, ", elapsed_millis);
// Reset input/output vector y
CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));
// Warmup
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseShybmv(
cusparse,
CUSPARSE_OPERATION_NON_TRANSPOSE,
¶ms.alpha, mat_desc,
hyb_desc,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
if (!g_quiet)
{
int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);
printf("\t%s
", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Timing
elapsed_millis = 0.0;
GpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseShybmv(
cusparse,
CUSPARSE_OPERATION_NON_TRANSPOSE,
¶ms.alpha, mat_desc,
hyb_desc,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
}
timer.Stop();
elapsed_millis += timer.ElapsedMillis();
// Cleanup
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyHybMat(hyb_desc));
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(mat_desc));
return elapsed_millis / timing_iterations;
}
/**
* Run cuSparse HYB SpMV (specialized for fp64)
*/
template <
typename OffsetT>
float TestCusparseHybmv(
double* vector_y_in,
double* reference_vector_y_out,
SpmvParams<double, OffsetT>& params,
int timing_iterations,
cusparseHandle_t cusparse)
{
CpuTimer cpu_timer;
cpu_timer.Start();
// Construct Hyb matrix
cusparseMatDescr_t mat_desc;
cusparseHybMat_t hyb_desc;
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&mat_desc));
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateHybMat(&hyb_desc));
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDcsr2hyb(
cusparse,
params.num_rows, params.num_cols,
mat_desc,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
hyb_desc,
0,
CUSPARSE_HYB_PARTITION_AUTO));
cudaDeviceSynchronize();
cpu_timer.Stop();
float elapsed_millis = cpu_timer.ElapsedMillis();
printf("HYB setup ms, %.5f, ", elapsed_millis);
// Reset input/output vector y
CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));
// Warmup
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDhybmv(
cusparse,
CUSPARSE_OPERATION_NON_TRANSPOSE,
¶ms.alpha, mat_desc,
hyb_desc,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
if (!g_quiet)
{
int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);
printf("\t%s
", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Timing
elapsed_millis = 0.0;
GpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDhybmv(
cusparse,
CUSPARSE_OPERATION_NON_TRANSPOSE,
¶ms.alpha, mat_desc,
hyb_desc,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
}
timer.Stop();
elapsed_millis += timer.ElapsedMillis();
// Cleanup
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyHybMat(hyb_desc));
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(mat_desc));
return elapsed_millis / timing_iterations;
}
//---------------------------------------------------------------------
// cuSparse CsrMV
//---------------------------------------------------------------------
/**
* Run cuSparse SpMV (specialized for fp32)
*/
template <
typename OffsetT>
float TestCusparseCsrmv(
float* vector_y_in,
float* reference_vector_y_out,
SpmvParams<float, OffsetT>& params,
int timing_iterations,
cusparseHandle_t cusparse)
{
cusparseMatDescr_t desc;
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&desc));
// Reset input/output vector y
CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));
// Warmup
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseScsrmv(
cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,
params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
if (!g_quiet)
{
int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);
printf("\t%s
", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Timing
float elapsed_millis = 0.0;
GpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseScsrmv(
cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,
params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
}
timer.Stop();
elapsed_millis += timer.ElapsedMillis();
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(desc));
return elapsed_millis / timing_iterations;
}
/**
* Run cuSparse SpMV (specialized for fp64)
*/
template <
typename OffsetT>
float TestCusparseCsrmv(
double* vector_y_in,
double* reference_vector_y_out,
SpmvParams<double, OffsetT>& params,
int timing_iterations,
cusparseHandle_t cusparse)
{
cusparseMatDescr_t desc;
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&desc));
// Reset input/output vector y
CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));
// Warmup
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDcsrmv(
cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,
params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
if (!g_quiet)
{
int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);
printf("\t%s
", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Timing
float elapsed_millis = 0.0;
GpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDcsrmv(
cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,
params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, ¶ms.beta, params.d_vector_y));
}
timer.Stop();
elapsed_millis += timer.ElapsedMillis();
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(desc));
return elapsed_millis / timing_iterations;
}
//---------------------------------------------------------------------
// GPU Merge-based SpMV
//---------------------------------------------------------------------
/**
* Run CUB SpMV
*/
template <
typename ValueT,
typename OffsetT>
float TestGpuMergeCsrmv(
ValueT* vector_y_in,
ValueT* reference_vector_y_out,
SpmvParams<ValueT, OffsetT>& params,
int timing_iterations)
{
// Allocate temporary storage
size_t temp_storage_bytes = 0;
void *d_temp_storage = NULL;
// Get amount of temporary storage needed
CubDebugExit(DeviceSpmv::CsrMV(
d_temp_storage, temp_storage_bytes,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, params.d_vector_y,
params.num_rows, params.num_cols, params.num_nonzeros,
// params.alpha, params.beta,
(cudaStream_t) 0, false));
// Allocate
CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));
// Reset input/output vector y
CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(ValueT) * params.num_rows, cudaMemcpyHostToDevice));
// Warmup
CubDebugExit(DeviceSpmv::CsrMV(
d_temp_storage, temp_storage_bytes,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, params.d_vector_y,
params.num_rows, params.num_cols, params.num_nonzeros,
// params.alpha, params.beta,
(cudaStream_t) 0, !g_quiet));
if (!g_quiet)
{
int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);
printf("\t%s
", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Timing
GpuTimer timer;
float elapsed_millis = 0.0;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
CubDebugExit(DeviceSpmv::CsrMV(
d_temp_storage, temp_storage_bytes,
params.d_values, params.d_row_end_offsets, params.d_column_indices,
params.d_vector_x, params.d_vector_y,
params.num_rows, params.num_cols, params.num_nonzeros,
// params.alpha, params.beta,
(cudaStream_t) 0, false));
}
timer.Stop();
elapsed_millis += timer.ElapsedMillis();
return elapsed_millis / timing_iterations;
}
//---------------------------------------------------------------------
// Test generation
//---------------------------------------------------------------------
/**
* Display perf
*/
template <typename ValueT, typename OffsetT>
void DisplayPerf(
float device_giga_bandwidth,
double avg_millis,
CsrMatrix<ValueT, OffsetT>& csr_matrix)
{
double nz_throughput, effective_bandwidth;
size_t total_bytes = (csr_matrix.num_nonzeros * (sizeof(ValueT) * 2 + sizeof(OffsetT))) +
(csr_matrix.num_rows) * (sizeof(OffsetT) + sizeof(ValueT));
nz_throughput = double(csr_matrix.num_nonzeros) / avg_millis / 1.0e6;
effective_bandwidth = double(total_bytes) / avg_millis / 1.0e6;
if (!g_quiet)
printf("fp%d: %.4f avg ms, %.5f gflops, %.3lf effective GB/s (%.2f%% peak)
",
sizeof(ValueT) * 8,
avg_millis,
2 * nz_throughput,
effective_bandwidth,
effective_bandwidth / device_giga_bandwidth * 100);
else
printf("%.5f, %.6f, %.3lf, %.2f%%, ",
avg_millis,
2 * nz_throughput,
effective_bandwidth,
effective_bandwidth / device_giga_bandwidth * 100);
fflush(stdout);
}
/**
* Run tests
*/
template <
typename ValueT,
typename OffsetT>
void RunTest(
bool rcm_relabel,
ValueT alpha,
ValueT beta,
CooMatrix<ValueT, OffsetT>& coo_matrix,
int timing_iterations,
CommandLineArgs& args)
{
// Adaptive timing iterations: run 16 billion nonzeros through
if (timing_iterations == -1)
timing_iterations = std::min(50000ull, std::max(100ull, ((16ull << 30) / coo_matrix.num_nonzeros)));
if (!g_quiet)
printf("\t%d timing iterations
", timing_iterations);
// Convert to CSR
CsrMatrix<ValueT, OffsetT> csr_matrix;
csr_matrix.FromCoo(coo_matrix);
if (!args.CheckCmdLineFlag("csrmv"))
coo_matrix.Clear();
// Relabel
if (rcm_relabel)
{
if (!g_quiet)
{
csr_matrix.Stats().Display();
printf("
");
csr_matrix.DisplayHistogram();
printf("
");
if (g_verbose2)
csr_matrix.Display();
printf("
");
}
RcmRelabel(csr_matrix, !g_quiet);
if (!g_quiet) printf("
");
}
// Display matrix info
csr_matrix.Stats().Display(!g_quiet);
if (!g_quiet)
{
printf("
");
csr_matrix.DisplayHistogram();
printf("
");
if (g_verbose2)
csr_matrix.Display();
printf("
");
}
fflush(stdout);
// Allocate input and output vectors
ValueT* vector_x = new ValueT[csr_matrix.num_cols];
ValueT* vector_y_in = new ValueT[csr_matrix.num_rows];
ValueT* vector_y_out = new ValueT[csr_matrix.num_rows];
for (int col = 0; col < csr_matrix.num_cols; ++col)
vector_x[col] = 1.0;
for (int row = 0; row < csr_matrix.num_rows; ++row)
vector_y_in[row] = 1.0;
// Compute reference answer
SpmvGold(csr_matrix, vector_x, vector_y_in, vector_y_out, alpha, beta);
float avg_millis;
if (g_quiet) {
printf("%s, %s, ", args.deviceProp.name, (sizeof(ValueT) > 4) ? "fp64" : "fp32"); fflush(stdout);
}
// Get GPU device bandwidth (GB/s)
float device_giga_bandwidth = args.device_giga_bandwidth;
// Allocate and initialize GPU problem
SpmvParams<ValueT, OffsetT> params;
CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_values, sizeof(ValueT) * csr_matrix.num_nonzeros));
CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_row_end_offsets, sizeof(OffsetT) * (csr_matrix.num_rows + 1)));
CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_column_indices, sizeof(OffsetT) * csr_matrix.num_nonzeros));
CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_vector_x, sizeof(ValueT) * csr_matrix.num_cols));
CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_vector_y, sizeof(ValueT) * csr_matrix.num_rows));
params.num_rows = csr_matrix.num_rows;
params.num_cols = csr_matrix.num_cols;
params.num_nonzeros = csr_matrix.num_nonzeros;
params.alpha = alpha;
params.beta = beta;
CubDebugExit(cudaMemcpy(params.d_values, csr_matrix.values, sizeof(ValueT) * csr_matrix.num_nonzeros, cudaMemcpyHostToDevice));
CubDebugExit(cudaMemcpy(params.d_row_end_offsets, csr_matrix.row_offsets, sizeof(OffsetT) * (csr_matrix.num_rows + 1), cudaMemcpyHostToDevice));
CubDebugExit(cudaMemcpy(params.d_column_indices, csr_matrix.column_indices, sizeof(OffsetT) * csr_matrix.num_nonzeros, cudaMemcpyHostToDevice));
CubDebugExit(cudaMemcpy(params.d_vector_x, vector_x, sizeof(ValueT) * csr_matrix.num_cols, cudaMemcpyHostToDevice));
if (!g_quiet) printf("
");
printf("GPU CSR I/O Prox, "); fflush(stdout);
avg_millis = TestGpuCsrIoProxy(params, timing_iterations);
DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);
if (args.CheckCmdLineFlag("csrmv"))
{
if (!g_quiet) printf("
");
printf("CUB, "); fflush(stdout);
avg_millis = TestGpuMergeCsrmv(vector_y_in, vector_y_out, params, timing_iterations);
DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);
}
// Initialize cuSparse
cusparseHandle_t cusparse;
AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreate(&cusparse));
if (args.CheckCmdLineFlag("csrmv"))
{
if (!g_quiet) printf("
");
printf("Cusparse CsrMV, "); fflush(stdout);
avg_millis = TestCusparseCsrmv(vector_y_in, vector_y_out, params, timing_iterations, cusparse);
DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);
}
if (args.CheckCmdLineFlag("hybmv"))
{
if (!g_quiet) printf("
");
printf("Cusparse HybMV, "); fflush(stdout);
avg_millis = TestCusparseHybmv(vector_y_in, vector_y_out, params, timing_iterations, cusparse);
DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);
}
// Cleanup
if (params.d_values) CubDebugExit(g_allocator.DeviceFree(params.d_values));
if (params.d_row_end_offsets) CubDebugExit(g_allocator.DeviceFree(params.d_row_end_offsets));
if (params.d_column_indices) CubDebugExit(g_allocator.DeviceFree(params.d_column_indices));
if (params.d_vector_x) CubDebugExit(g_allocator.DeviceFree(params.d_vector_x));
if (params.d_vector_y) CubDebugExit(g_allocator.DeviceFree(params.d_vector_y));
if (vector_x) delete[] vector_x;
if (vector_y_in) delete[] vector_y_in;
if (vector_y_out) delete[] vector_y_out;
}
/**
* Run tests
*/
template <
typename ValueT,
typename OffsetT>
void RunTests(
bool rcm_relabel,
ValueT alpha,
ValueT beta,
const std::string& mtx_filename,
int grid2d,
int grid3d,
int wheel,
int dense,
int timing_iterations,
CommandLineArgs& args)
{
// Initialize matrix in COO form
CooMatrix<ValueT, OffsetT> coo_matrix;
if (!mtx_filename.empty())
{
// Parse matrix market file
printf("%s, ", mtx_filename.c_str()); fflush(stdout);
coo_matrix.InitMarket(mtx_filename, 1.0, !g_quiet);
if ((coo_matrix.num_rows == 1) || (coo_matrix.num_cols == 1) || (coo_matrix.num_nonzeros == 1))
{
if (!g_quiet) printf("Trivial dataset
");
exit(0);
}
}
else if (grid2d > 0)
{
// Generate 2D lattice
printf("grid2d_%d, ", grid2d); fflush(stdout);
coo_matrix.InitGrid2d(grid2d, false);
}
else if (grid3d > 0)
{
// Generate 3D lattice
printf("grid3d_%d, ", grid3d); fflush(stdout);
coo_matrix.InitGrid3d(grid3d, false);
}
else if (wheel > 0)
{
// Generate wheel graph
printf("wheel_%d, ", grid2d); fflush(stdout);
coo_matrix.InitWheel(wheel);
}
else if (dense > 0)
{
// Generate dense graph
OffsetT size = 1 << 24; // 16M nnz
args.GetCmdLineArgument("size", size);
OffsetT rows = size / dense;
printf("dense_%d_x_%d, ", rows, dense); fflush(stdout);
coo_matrix.InitDense(rows, dense);
}
else
{
fprintf(stderr, "No graph type specified.
");
exit(1);
}
RunTest(
rcm_relabel,
alpha,
beta,
coo_matrix,
timing_iterations,
args);
}
/**
* Main
*/
int main(int argc, char **argv)
{
// Initialize command line
CommandLineArgs args(argc, argv);
if (args.CheckCmdLineFlag("help"))
{
printf(
"%s "
"[--csrmv | --hybmv | --bsrmv ] "
"[--device=<device-id>] "
"[--quiet] "
"[--v] "
"[--i=<timing iterations>] "
"[--fp64] "
"[--rcm] "
"[--alpha=<alpha scalar (default: 1.0)>] "
"[--beta=<beta scalar (default: 0.0)>] "
"
\t"
"--mtx=<matrix market file> "
"
\t"
"--dense=<cols>"
"
\t"
"--grid2d=<width>"
"
\t"
"--grid3d=<width>"
"
\t"
"--wheel=<spokes>"
"
", argv[0]);
exit(0);
}
bool fp64;
bool rcm_relabel;
std::string mtx_filename;
int grid2d = -1;
int grid3d = -1;
int wheel = -1;
int dense = -1;
int timing_iterations = -1;
float alpha = 1.0;
float beta = 0.0;
g_verbose = args.CheckCmdLineFlag("v");
g_verbose2 = args.CheckCmdLineFlag("v2");
g_quiet = args.CheckCmdLineFlag("quiet");
fp64 = args.CheckCmdLineFlag("fp64");
rcm_relabel = args.CheckCmdLineFlag("rcm");
args.GetCmdLineArgument("i", timing_iterations);
args.GetCmdLineArgument("mtx", mtx_filename);
args.GetCmdLineArgument("grid2d", grid2d);
args.GetCmdLineArgument("grid3d", grid3d);
args.GetCmdLineArgument("wheel", wheel);
args.GetCmdLineArgument("dense", dense);
args.GetCmdLineArgument("alpha", alpha);
args.GetCmdLineArgument("beta", beta);
// Initialize device
CubDebugExit(args.DeviceInit());
// Run test(s)
if (fp64)
{
RunTests<double, int>(rcm_relabel, alpha, beta, mtx_filename, grid2d, grid3d, wheel, dense, timing_iterations, args);
}
else
{
RunTests<float, int>(rcm_relabel, alpha, beta, mtx_filename, grid2d, grid3d, wheel, dense, timing_iterations, args);
}
CubDebugExit(cudaDeviceSynchronize());
printf("
");
return 0;
}
|