-
Notifications
You must be signed in to change notification settings - Fork 0
/
cpu_spmm.cpp
1361 lines (1203 loc) · 45.6 KB
/
cpu_spmm.cpp
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
/******************************************************************************
* Copyright (c) 2011-2015, 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.
*
******************************************************************************/
/******************************************************************************
* How to build:
*
* VC++
* cl.exe mergebased_spmv.cpp /fp:strict /MT /O2 /openmp
*
* GCC (OMP is terrible)
* g++ mergebased_spmv.cpp -lm -ffloat-store -O3 -fopenmp
*
* Intel
* icpc mergebased_spmv.cpp -openmp -O3 -lrt -fno-alias -xHost -lnuma
* export KMP_AFFINITY=granularity=core,scatter
*
*
******************************************************************************/
//---------------------------------------------------------------------
// SpMV comparison tool
//---------------------------------------------------------------------
#include <omp.h>
#include <stdio.h>
#include <vector>
#include <algorithm>
#include <cstdio>
#include <fstream>
#include <sstream>
#include <iostream>
#include <limits>
#include <immintrin.h>
#include <mkl.h>
#include "sparse_matrix.h"
#include "utils.h"
//---------------------------------------------------------------------
// 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
int g_omp_threads = -1; // Number of openMP threads
int g_expected_calls = 1000000;
bool g_input_row_major = true;
bool g_output_row_major = true;
//---------------------------------------------------------------------
// Utility types
//---------------------------------------------------------------------
struct int2
{
int x;
int y;
};
/**
* Counting iterator
*/
template <
typename ValueType,
typename OffsetT = ptrdiff_t>
struct CountingInputIterator
{
// Required iterator traits
typedef CountingInputIterator self_type; ///< My own type
typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another
typedef ValueType value_type; ///< The type of the element the iterator can point to
typedef ValueType* pointer; ///< The type of a pointer to an element the iterator can point to
typedef ValueType reference; ///< The type of a reference to an element the iterator can point to
typedef std::random_access_iterator_tag iterator_category; ///< The iterator category
ValueType val;
/// Constructor
inline CountingInputIterator(
const ValueType &val) ///< Starting value for the iterator instance to report
:
val(val)
{}
/// Postfix increment
inline self_type operator++(int)
{
self_type retval = *this;
val++;
return retval;
}
/// Prefix increment
inline self_type operator++()
{
val++;
return *this;
}
/// Indirection
inline reference operator*() const
{
return val;
}
/// Addition
template <typename Distance>
inline self_type operator+(Distance n) const
{
self_type retval(val + n);
return retval;
}
/// Addition assignment
template <typename Distance>
inline self_type& operator+=(Distance n)
{
val += n;
return *this;
}
/// Subtraction
template <typename Distance>
inline self_type operator-(Distance n) const
{
self_type retval(val - n);
return retval;
}
/// Subtraction assignment
template <typename Distance>
inline self_type& operator-=(Distance n)
{
val -= n;
return *this;
}
/// Distance
inline difference_type operator-(self_type other) const
{
return val - other.val;
}
/// Array subscript
template <typename Distance>
inline reference operator[](Distance n) const
{
return val + n;
}
/// Structure dereference
inline pointer operator->()
{
return &val;
}
/// Equal to
inline bool operator==(const self_type& rhs)
{
return (val == rhs.val);
}
/// Not equal to
inline bool operator!=(const self_type& rhs)
{
return (val != rhs.val);
}
/// ostream operator
friend std::ostream& operator<<(std::ostream& os, const self_type& itr)
{
os << "[" << itr.val << "]";
return os;
}
};
// mkl gemm
// void cblas_dgemm (const CBLAS_LAYOUT Layout, const CBLAS_TRANSPOSE transa, const CBLAS_TRANSPOSE transb, const MKL_INT m, const MKL_INT n, const MKL_INT k,
// const double alpha, const double *a, const MKL_INT lda, const double *b, const MKL_INT ldb, const double beta, double *c, const MKL_INT ldc);
// cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, 3, 2, 4, 1.0, a, 2, b, 4, 0.0, c, 4);
//---------------------------------------------------------------------
// MergePath Search
//---------------------------------------------------------------------
/**
* Computes the begin offsets into A and B for the specific diagonal
*/
template <
typename AIteratorT,
typename BIteratorT,
typename OffsetT,
typename CoordinateT>
inline void MergePathSearch(
OffsetT diagonal, ///< [in]The diagonal to search
AIteratorT a, ///< [in]List A
BIteratorT b, ///< [in]List B
OffsetT a_len, ///< [in]Length of A
OffsetT b_len, ///< [in]Length of B
CoordinateT& path_coordinate) ///< [out] (x,y) coordinate where diagonal intersects the merge path
{
OffsetT x_min = std::max(diagonal - b_len, 0);
OffsetT x_max = std::min(diagonal, a_len);
while (x_min < x_max)
{
OffsetT x_pivot = (x_min + x_max) >> 1;
if (a[x_pivot] <= b[diagonal - x_pivot - 1])
x_min = x_pivot + 1; // Contract range up A (down B)
else
x_max = x_pivot; // Contract range down A (up B)
}
path_coordinate.x = std::min(x_min, a_len);
path_coordinate.y = diagonal - x_min;
}
//---------------------------------------------------------------------
// 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;
}
}
//---------------------------------------------------------------------
// CPU normal omp SpMV
//---------------------------------------------------------------------
template <
typename ValueT,
typename OffsetT>
void OmpCsrSpmmT(
int num_threads,
CsrMatrix<ValueT, OffsetT>& a,
ValueT* __restrict vector_x,
ValueT* __restrict vector_y_out,
int num_vectors,
ValueT* __restrict vector_x_row_major)
{
int num_cols = a.num_cols;
int num_rows = a.num_rows;
int xt_index = 0;
if (!g_input_row_major)
{
#pragma omp parallel for schedule(static) num_threads(num_threads) private(xt_index)
for (int i=0; i<num_cols; i++){
for(int j=i; j<num_cols*num_vectors; j+=num_cols){
vector_x_row_major[xt_index] = vector_x[j];
xt_index += 1;
}
}
}
#pragma omp parallel for schedule(static) num_threads(num_threads)
for (OffsetT row = 0; row < a.num_rows; ++row)
{
ValueT partial[num_vectors] = {0.0};
for (
OffsetT offset = a.row_offsets[row];
offset < a.row_offsets[row + 1];
++offset)
{
ValueT val = a.values[offset];
int ind = a.column_indices[offset]*num_vectors;
for (int i=0; i<num_vectors; i++){
partial[i] += val * vector_x_row_major[ind + i];
}
}
if (g_output_row_major)
{
int ind = row*num_vectors;
for (int i=0; i<num_vectors; i++){
vector_y_out[ind + i] = partial[i];
}
}
else
{
for (int i=0; i<num_vectors; i++){
vector_y_out[row + i * num_rows] = partial[i];
}
}
}
}
template <
typename ValueT,
typename OffsetT>
void OmpCsrSymSpmm(
int num_threads,
CsrMatrix<ValueT, OffsetT>& a,
ValueT* __restrict vector_x,
ValueT* __restrict vector_y_out,
int num_vectors,
ValueT* __restrict vector_x_row_major)
{
int num_cols = a.num_cols;
int num_rows = a.num_rows;
int xt_index = 0;
if (!g_input_row_major)
{
#pragma omp parallel for schedule(static) num_threads(num_threads) private(xt_index)
for (int i=0; i<num_cols; i++){
for(int j=i; j<num_cols*num_vectors; j+=num_cols){
vector_x_row_major[xt_index] = vector_x[j];
xt_index += 1;
}
}
}
#pragma omp parallel for schedule(static) num_threads(num_threads)
for (OffsetT row = 0; row < a.num_rows; ++row)
{
int ind = row*num_vectors;
for (
OffsetT offset = a.row_offsets[row];
offset < a.row_offsets[row + 1];
++offset)
{
ValueT val = a.values[offset];
int ind_y = a.column_indices[offset]*num_vectors;
for (int i=0; i<num_vectors; i++){
vector_y_out[ind_y + i] += val * vector_x_row_major[ind + i];
if (row == 366 && i ==0){
printf("%lf\n", val * vector_x_row_major[ind + i]);
}
}
}
}
}
/**
* Run OmpCsrSpmm
*/
template <
typename ValueT,
typename OffsetT>
float TestOmpCsrSpmmT(
CsrMatrix<ValueT, OffsetT>& a,
ValueT* vector_x,
ValueT* reference_vector_y_out,
ValueT* vector_y_out,
int timing_iterations,
float &setup_ms,
int num_vectors,
ValueT* vector_x_row_major)
{
setup_ms = 0.0;
if (g_omp_threads == -1)
g_omp_threads = omp_get_num_procs();
int num_threads = g_omp_threads;
if (!g_quiet)
printf("\tUsing %d threads on %d procs\n", g_omp_threads, omp_get_num_procs());
// Warmup/correctness
memset(vector_y_out, -1, sizeof(ValueT) * a.num_rows * num_vectors);
OmpCsrSpmmT(g_omp_threads, a, vector_x, vector_y_out, num_vectors, vector_x_row_major);
if (!g_quiet)
{
// Check answer
int compare = CompareResults(reference_vector_y_out, vector_y_out, a.num_rows, true);
printf("\t%s\n", compare ? "FAIL" : "PASS"); fflush(stdout);
if (g_input_row_major || g_output_row_major){
for(int i=0; i<a.num_rows; i++){
ValueT diff = reference_vector_y_out[i] - vector_y_out[i*num_vectors];
if(diff>0.000001 || diff<-0.0000001){
printf("%lf, %lf\n",reference_vector_y_out[i],vector_y_out[i*num_vectors]);
}
}
}
}
// Re-populate caches, etc.
for(int it = 0; it < timing_iterations; ++it)
{
OmpCsrSpmmT(g_omp_threads, a, vector_x, vector_y_out, num_vectors, vector_x_row_major);
}
// Timing
float elapsed_ms = 0.0;
CpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
OmpCsrSpmmT(g_omp_threads, a, vector_x, vector_y_out, num_vectors, vector_x_row_major);
}
timer.Stop();
elapsed_ms += timer.ElapsedMillis();
return elapsed_ms / timing_iterations;
}
template <
typename ValueT,
typename OffsetT>
float TestOmpCsrSymSpmm(
CsrMatrix<ValueT, OffsetT>& a,
ValueT* vector_x,
ValueT* reference_vector_y_out,
ValueT* vector_y_out,
int timing_iterations,
float &setup_ms,
int num_vectors,
ValueT* vector_x_row_major)
{
setup_ms = 0.0;
if (g_omp_threads == -1)
g_omp_threads = omp_get_num_procs();
int num_threads = g_omp_threads;
if (!g_quiet)
printf("\tUsing %d threads on %d procs\n", g_omp_threads, omp_get_num_procs());
// Warmup/correctness
memset(vector_y_out, -1, sizeof(ValueT) * a.num_rows * num_vectors);
OmpCsrSymSpmm(g_omp_threads, a, vector_x, vector_y_out, num_vectors, vector_x_row_major);
if (!g_quiet)
{
// Check answer
int compare = CompareResults(reference_vector_y_out, vector_y_out, a.num_rows, true);
printf("\t%s\n", compare ? "FAIL" : "PASS"); fflush(stdout);
if (g_input_row_major || g_output_row_major){
for(int i=0; i<a.num_rows; i++){
ValueT diff = reference_vector_y_out[i] - vector_y_out[i*num_vectors];
if(diff>0.000001 || diff<-0.0000001){
printf("%lf, %lf, %d\n",reference_vector_y_out[i],vector_y_out[i*num_vectors],i);
}
}
}
}
// Re-populate caches, etc.
for(int it = 0; it < timing_iterations; ++it)
{
OmpCsrSymSpmm(g_omp_threads, a, vector_x, vector_y_out, num_vectors, vector_x_row_major);
}
// Timing
float elapsed_ms = 0.0;
CpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
OmpCsrSymSpmm(g_omp_threads, a, vector_x, vector_y_out, num_vectors, vector_x_row_major);
}
timer.Stop();
elapsed_ms += timer.ElapsedMillis();
return elapsed_ms / timing_iterations;
}
//---------------------------------------------------------------------
// MKL SpMV
//---------------------------------------------------------------------
/**
* MKL CPU SpMV (specialized for fp32)
*/
template <typename OffsetT>
void MKLCsrmm(
int num_threads,
CsrMatrix<float, OffsetT>& a,
OffsetT* __restrict row_end_offsets, ///< Merge list A (row end-offsets)
OffsetT* __restrict column_indices,
float* __restrict values,
float* __restrict vector_x,
float* __restrict vector_y_out,
int num_vectors)
{
struct matrix_descr A_descr;
A_descr.type = SPARSE_MATRIX_TYPE_GENERAL;
sparse_matrix_t csrA;
mkl_sparse_s_create_csr( &csrA, SPARSE_INDEX_BASE_ZERO, a.num_rows, a.num_cols, a.row_offsets, row_end_offsets, a.column_indices, a.values);
if (g_input_row_major && g_output_row_major)
{
mkl_sparse_s_mm(SPARSE_OPERATION_NON_TRANSPOSE, 1.0, csrA, A_descr, SPARSE_LAYOUT_ROW_MAJOR, vector_x, num_vectors, num_vectors, 0.0, vector_y_out, num_vectors);
}
else
{
mkl_sparse_s_mm(SPARSE_OPERATION_NON_TRANSPOSE, 1.0, csrA, A_descr, SPARSE_LAYOUT_COLUMN_MAJOR, vector_x, num_vectors, a.num_cols, 0.0, vector_y_out, a.num_rows);
}
}
/**
* MKL CPU SpMV (specialized for fp64)
*/
template <typename OffsetT>
void MKLCsrmm(
int num_threads,
CsrMatrix<double, OffsetT>& a,
OffsetT* __restrict row_end_offsets, ///< Merge list A (row end-offsets)
OffsetT* __restrict column_indices,
double* __restrict values,
double* __restrict vector_x,
double* __restrict vector_y_out,
int num_vectors)
{
struct matrix_descr A_descr;
A_descr.type = SPARSE_MATRIX_TYPE_GENERAL;
sparse_matrix_t csrA;
mkl_sparse_d_create_csr( &csrA, SPARSE_INDEX_BASE_ZERO, a.num_rows, a.num_cols, a.row_offsets, row_end_offsets, a.column_indices, a.values);
if (g_input_row_major && g_output_row_major)
{
mkl_sparse_d_mm(SPARSE_OPERATION_NON_TRANSPOSE, 1.0, csrA, A_descr, SPARSE_LAYOUT_ROW_MAJOR, vector_x, num_vectors, num_vectors, 0.0, vector_y_out, num_vectors);
}
else
{
mkl_sparse_d_mm(SPARSE_OPERATION_NON_TRANSPOSE, 1.0, csrA, A_descr, SPARSE_LAYOUT_COLUMN_MAJOR, vector_x, num_vectors, a.num_cols, 0.0, vector_y_out, a.num_rows);
}
}
/**
* Run MKL CsrMV
*/
template <
typename ValueT,
typename OffsetT>
float TestMKLCsrmm(
CsrMatrix<ValueT, OffsetT>& a,
ValueT* vector_x,
ValueT* reference_vector_y_out,
ValueT* vector_y_out,
int timing_iterations,
float &setup_ms,
int num_vectors)
{
setup_ms = 0.0;
// Warmup/correctness
memset(vector_y_out, -1, sizeof(ValueT) * a.num_rows * num_vectors);
MKLCsrmm(g_omp_threads, a, a.row_offsets + 1, a.column_indices, a.values, vector_x, vector_y_out, num_vectors);
if (!g_quiet)
{
// Check answer
int compare = CompareResults(reference_vector_y_out, vector_y_out, a.num_rows, true);
printf("\t%s\n", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Re-populate caches, etc.
for(int it = 0; it < timing_iterations; ++it)
{
MKLCsrmm(g_omp_threads, a, a.row_offsets + 1, a.column_indices, a.values, vector_x, vector_y_out, num_vectors);
}
// Timing
float elapsed_ms = 0.0;
CpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
MKLCsrmm(g_omp_threads, a, a.row_offsets + 1, a.column_indices, a.values, vector_x, vector_y_out, num_vectors);
}
timer.Stop();
elapsed_ms += timer.ElapsedMillis();
return elapsed_ms / timing_iterations;
}
//---------------------------------------------------------------------
// CPU merge-based SpMV
//---------------------------------------------------------------------
/**
* OpenMP CPU merge-based SpMV
*/
template <
typename ValueT,
typename OffsetT>
void OmpMergeCsrmm(
int num_threads,
CsrMatrix<ValueT, OffsetT>& a,
OffsetT* __restrict row_end_offsets, ///< Merge list A (row end-offsets)
OffsetT* __restrict column_indices,
ValueT* __restrict values,
ValueT* __restrict vector_x,
ValueT* __restrict vector_y_out,
int num_vectors,
ValueT* __restrict vector_x_row_major)
{
// Temporary storage for inter-thread fix-up after load-balanced work
OffsetT row_carry_out[256]; // The last row-id each worked on by each thread when it finished its path segment
ValueT* value_carry_out[256]; // The running total within each thread when it finished its path segment
int num_cols = a.num_cols;
int num_rows = a.num_rows;
int xt_index = 0;
if (!g_input_row_major)
{
#pragma omp parallel for schedule(static) num_threads(num_threads)
for (int i=0; i<num_cols; i++){
for(int j=i; j<num_cols*num_vectors; j+=num_cols){
vector_x_row_major[xt_index] = vector_x[j];
xt_index += 1;
}
}
}
#pragma omp parallel for schedule(static) num_threads(num_threads)
for (int tid = 0; tid < num_threads; tid++)
{
// Merge list B (NZ indices)
CountingInputIterator<OffsetT> nonzero_indices(0);
OffsetT num_merge_items = a.num_rows + a.num_nonzeros; // Merge path total length
OffsetT items_per_thread = (num_merge_items + num_threads - 1) / num_threads; // Merge items per thread
// Find starting and ending MergePath coordinates (row-idx, nonzero-idx) for each thread
int2 thread_coord;
int2 thread_coord_end;
int start_diagonal = std::min(items_per_thread * tid, num_merge_items);
int end_diagonal = std::min(start_diagonal + items_per_thread, num_merge_items);
MergePathSearch(start_diagonal, row_end_offsets, nonzero_indices, a.num_rows, a.num_nonzeros, thread_coord);
MergePathSearch(end_diagonal, row_end_offsets, nonzero_indices, a.num_rows, a.num_nonzeros, thread_coord_end);
// Consume whole rows
ValueT running_total[num_vectors] = {0.0};
ValueT val;
int ind;
ValueT* tmp;
for (; thread_coord.x < thread_coord_end.x; ++thread_coord.x)
{
for (; thread_coord.y < row_end_offsets[thread_coord.x]; ++thread_coord.y)
{
val = values[thread_coord.y];
ind = column_indices[thread_coord.y]*num_vectors;
tmp = vector_x_row_major+ind;
for (int i=0; i<num_vectors; i++){
running_total[i] += val * tmp[i];
}
}
if (g_output_row_major)
{
ind = thread_coord.x*num_vectors;
tmp = vector_y_out+ind;
for (int i=0; i<num_vectors; i++){
tmp[i] = running_total[i];
running_total[i] = 0.0;
}
}
else
{
for (int i=0; i<num_vectors; i++){
vector_y_out[thread_coord.x + i * num_rows] = running_total[i];
running_total[i] = 0.0;
}
}
}
// Consume partial portion of thread's last row
for (; thread_coord.y < thread_coord_end.y; ++thread_coord.y)
{
val = values[thread_coord.y];
ind = column_indices[thread_coord.y]*num_vectors;
tmp = vector_x_row_major+ind;
for (int i=0; i<num_vectors; i++){
running_total[i] += val * tmp[i];
}
}
// Save carry-outs
row_carry_out[tid] = thread_coord_end.x;
value_carry_out[tid] = running_total;
}
// Carry-out fix-up (rows spanning multiple threads)
int ind;
ValueT* tmp;
if (g_output_row_major)
{
for (int tid = 0; tid < num_threads - 1; ++tid)
{
if (row_carry_out[tid] < a.num_rows){
ind = row_carry_out[tid]*num_vectors;
tmp = vector_y_out+ind;
for (int i=0; i<num_vectors; i++){
tmp[i] += value_carry_out[tid][i];
}
}
}
}
else
{
for (int tid = 0; tid < num_threads - 1; ++tid)
{
if (row_carry_out[tid] < a.num_rows){
for (int i=0; i<num_vectors; i++){
vector_y_out[row_carry_out[tid] + i * num_rows] += value_carry_out[tid][i];
}
}
}
}
}
/**
* Run OmpMergeCsrmm
*/
template <
typename ValueT,
typename OffsetT>
float TestOmpMergeCsrmm(
CsrMatrix<ValueT, OffsetT>& a,
ValueT* vector_x,
ValueT* reference_vector_y_out,
ValueT* vector_y_out,
int timing_iterations,
float &setup_ms,
int num_vectors,
ValueT* vector_x_row_major)
{
setup_ms = 0.0;
if (g_omp_threads == -1)
g_omp_threads = omp_get_num_procs();
int num_threads = g_omp_threads;
if (!g_quiet)
printf("\tUsing %d threads on %d procs\n", g_omp_threads, omp_get_num_procs());
// Warmup/correctness
memset(vector_y_out, -1, sizeof(ValueT) * a.num_rows * num_vectors);
OmpMergeCsrmm(g_omp_threads, a, a.row_offsets + 1, a.column_indices, a.values, vector_x, vector_y_out, num_vectors, vector_x_row_major);
if (!g_quiet)
{
// Check answer
int compare = CompareResults(reference_vector_y_out, vector_y_out, a.num_rows, true);
printf("\t%s\n", compare ? "FAIL" : "PASS"); fflush(stdout);
}
// Re-populate caches, etc.
for(int it = 0; it < timing_iterations; ++it)
{
OmpMergeCsrmm(g_omp_threads, a, a.row_offsets + 1, a.column_indices, a.values, vector_x, vector_y_out, num_vectors, vector_x_row_major);
}
// Timing
float elapsed_ms = 0.0;
CpuTimer timer;
timer.Start();
for(int it = 0; it < timing_iterations; ++it)
{
OmpMergeCsrmm(g_omp_threads, a, a.row_offsets + 1, a.column_indices, a.values, vector_x, vector_y_out, num_vectors, vector_x_row_major);
}
timer.Stop();
elapsed_ms += timer.ElapsedMillis();
return elapsed_ms / timing_iterations;
}
template <
typename AIteratorT,
typename BIteratorT,
typename OffsetT,
typename CoordinateT>
inline void RowPathSearch(
AIteratorT a, ///< [in]List A
BIteratorT b, ///< [in]List B
OffsetT a_len, ///< [in]Length of A
OffsetT b_len, ///< [in]Length of B
CoordinateT& path_coordinate) ///< [out] (x,y) coordinate where diagonal intersects the merge path
{
OffsetT x_min = 0;
OffsetT x_max = a_len;
while (x_min < x_max)
{
OffsetT x_pivot = (x_min + x_max) >> 1;
if (a[x_pivot] <= b[path_coordinate.y - 1])
x_min = x_pivot + 1; // Contract range up A (down B)
else
x_max = x_pivot; // Contract range down A (up B)
}
path_coordinate.x = std::min(x_min, a_len);
}
inline void mkl_axpy(int n, double a, double* x, double*y){
cblas_daxpy(n, a, x, 1, y, 1);
}
inline void mkl_axpy(int n, float a, float* x, float*y){
cblas_saxpy(n, a, x, 1, y, 1);
}
/**
* OpenMP CPU row-based SpMM
*/
template <
typename ValueT,
typename OffsetT>
void OmpNonzeroSplitCsrmm(
int num_threads,
CsrMatrix<ValueT, OffsetT>& a,
OffsetT* __restrict row_end_offsets, ///< Merge list A (row end-offsets)
OffsetT* __restrict column_indices,
ValueT* __restrict values,
ValueT* __restrict vector_x,
ValueT* __restrict vector_y_out,
int num_vectors,
ValueT* __restrict vector_x_row_major)
{
// Temporary storage for inter-thread fix-up after load-balanced work
OffsetT row_carry_out[256]; // The last row-id each worked on by each thread when it finished its path segment
ValueT* value_carry_out[256]; // The running total within each thread when it finished its path segment
int num_cols = a.num_cols;
int num_rows = a.num_rows;
int xt_index = 0;
if (!g_input_row_major)
{
#pragma omp parallel for schedule(static) num_threads(num_threads)
for (int i=0; i<num_cols; i++){
for(int j=i; j<num_cols*num_vectors; j+=num_cols){
vector_x_row_major[xt_index] = vector_x[j];
xt_index += 1;
}
}
}
#pragma omp parallel for schedule(static) num_threads(num_threads)
for (int tid = 0; tid < num_threads; tid++)
{
// Merge list B (NZ indices)
CountingInputIterator<OffsetT> nonzero_indices(0);
OffsetT num_nonzeros = a.num_nonzeros;
OffsetT items_per_thread = (num_nonzeros + num_threads - 1) / num_threads;
// Find starting and ending MergePath coordinates (row-idx, nonzero-idx) for each thread
int2 thread_coord;
int2 thread_coord_end;
thread_coord.y = std::min(items_per_thread * tid, num_nonzeros);
thread_coord_end.y = std::min(thread_coord.y + items_per_thread, num_nonzeros);
RowPathSearch(row_end_offsets, nonzero_indices, a.num_rows, a.num_nonzeros, thread_coord);
RowPathSearch(row_end_offsets, nonzero_indices, a.num_rows, a.num_nonzeros, thread_coord_end);
// Consume whole rows
ValueT running_total[num_vectors] = {0.0};
ValueT val;
ValueT* tmp;
int ind;
for (; thread_coord.x < thread_coord_end.x; ++thread_coord.x)
{
for (; thread_coord.y < row_end_offsets[thread_coord.x]; ++thread_coord.y)
{
val = values[thread_coord.y];
ind = column_indices[thread_coord.y]*num_vectors;
tmp = vector_x_row_major+ind;
for (int i=0; i<num_vectors; i+=1){
running_total[i] += val * tmp[i];
}
}
if (g_output_row_major)
{
ind = thread_coord.x*num_vectors;
tmp = vector_y_out+ind;
for (int i=0; i<num_vectors; i++){
tmp[i] = running_total[i];
running_total[i] = 0.0;
}
}
else
{
for (int i=0; i<num_vectors; i++){
vector_y_out[thread_coord.x + i * num_rows] = running_total[i];
running_total[i] = 0.0;
}
}
}
// Consume partial portion of thread's last row
for (; thread_coord.y < thread_coord_end.y; ++thread_coord.y)
{
val = values[thread_coord.y];
ind = column_indices[thread_coord.y]*num_vectors;
tmp = vector_x_row_major+ind;
for (int i=0; i<num_vectors; i++){
running_total[i] += val * tmp[i];
}
}
// Save carry-outs
row_carry_out[tid] = thread_coord_end.x;
value_carry_out[tid] = running_total;
}
// Carry-out fix-up (rows spanning multiple threads)
int ind;
ValueT* tmp;
if (g_output_row_major)
{
for (int tid = 0; tid < num_threads - 1; ++tid)
{
if (row_carry_out[tid] < a.num_rows){
ind = row_carry_out[tid]*num_vectors;
tmp = vector_y_out+ind;
for (int i=0; i<num_vectors; i++){
tmp[i] += value_carry_out[tid][i];
}
}
}
}
else
{
for (int tid = 0; tid < num_threads - 1; ++tid)
{
if (row_carry_out[tid] < a.num_rows){
for (int i=0; i<num_vectors; i++){
vector_y_out[row_carry_out[tid] + i * num_rows] += value_carry_out[tid][i];
}
}
}
}
}
template <
typename ValueT,
typename OffsetT>
float TestOmpNonzeroSplitCsrmm(
CsrMatrix<ValueT, OffsetT>& a,
ValueT* vector_x,
ValueT* reference_vector_y_out,
ValueT* vector_y_out,
int timing_iterations,