/
dbcsr_performance_multiply.F
721 lines (645 loc) · 32.3 KB
/
dbcsr_performance_multiply.F
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
!--------------------------------------------------------------------------------------------------!
! CP2K: A general program to perform molecular dynamics simulations !
! Copyright (C) 2000 - 2018 CP2K developers group !
!--------------------------------------------------------------------------------------------------!
! **************************************************************************************************
!> \brief Performance for DBCSR multiply
!> \author VW
!> \date 2010
!> \version 1.0
!>
!> <b>Modification history:</b>
!> - Created 2010
! **************************************************************************************************
MODULE dbcsr_performance_multiply
USE dbcsr_data_methods, ONLY: dbcsr_scalar,&
dbcsr_scalar_negative,&
dbcsr_scalar_one
USE dbcsr_dist_methods, ONLY: dbcsr_distribution_col_dist,&
dbcsr_distribution_new,&
dbcsr_distribution_release,&
dbcsr_distribution_row_dist
USE dbcsr_dist_operations, ONLY: dbcsr_dist_bin
USE dbcsr_dist_util, ONLY: dbcsr_checksum
USE dbcsr_io, ONLY: dbcsr_print
USE dbcsr_kinds, ONLY: int_8,&
real_4,&
real_8
USE dbcsr_machine, ONLY: m_walltime
USE dbcsr_message_passing, ONLY: mp_environ,&
mp_sum,&
mp_sync
USE dbcsr_methods, ONLY: &
dbcsr_col_block_offsets, dbcsr_col_block_sizes, dbcsr_distribution, dbcsr_get_data_type, &
dbcsr_get_matrix_type, dbcsr_name, dbcsr_nfullcols_total, dbcsr_nfullrows_total, &
dbcsr_release, dbcsr_row_block_offsets, dbcsr_row_block_sizes
USE dbcsr_multiply_api, ONLY: dbcsr_multiply
USE dbcsr_operations, ONLY: dbcsr_copy,&
dbcsr_scale
USE dbcsr_test_methods, ONLY: atoi,&
atol,&
ator,&
dbcsr_make_random_block_sizes,&
dbcsr_make_random_matrix,&
dbcsr_reset_randmat_seed
USE dbcsr_transformations, ONLY: dbcsr_redistribute
USE dbcsr_types, ONLY: &
dbcsr_conjugate_transpose, dbcsr_distribution_obj, dbcsr_mp_obj, dbcsr_no_transpose, &
dbcsr_scalar_type, dbcsr_transpose, dbcsr_type, dbcsr_type_antisymmetric, &
dbcsr_type_complex_4, dbcsr_type_complex_8, dbcsr_type_no_symmetry, dbcsr_type_real_4, &
dbcsr_type_real_8
USE dbcsr_work_operations, ONLY: dbcsr_create,&
dbcsr_finalize
#include "base/dbcsr_base_uses.f90"
!$ USE OMP_LIB, ONLY: omp_get_max_threads, omp_get_thread_num, omp_get_num_threads
IMPLICIT NONE
PRIVATE
PUBLIC :: dbcsr_perf_multiply
CHARACTER(len=*), PARAMETER, PRIVATE :: moduleN = 'dbcsr_performance_multiply'
CONTAINS
! **************************************************************************************************
!> \brief ...
!> \param group ...
!> \param mp_env ...
!> \param npdims ...
!> \param io_unit ...
!> \param narg ...
!> \param args ...
! **************************************************************************************************
SUBROUTINE dbcsr_perf_multiply(group, mp_env, npdims, io_unit, narg, args)
INTEGER :: group
TYPE(dbcsr_mp_obj), INTENT(IN) :: mp_env
INTEGER, DIMENSION(2), INTENT(in) :: npdims
INTEGER :: io_unit, narg
CHARACTER(len=*), DIMENSION(:), INTENT(IN) :: args
CHARACTER(len=*), PARAMETER :: routineN = 'dbcsr_perf_multiply', &
routineP = moduleN//':'//routineN
CHARACTER :: symmetries(3), trans(2)
INTEGER :: i, iblk, kblk_to_read, limits(6), &
matrix_sizes(3), mblk_to_read, &
nblk_to_read, nrep, TYPE
INTEGER, ALLOCATABLE, DIMENSION(:) :: bs_k, bs_m, bs_n
LOGICAL :: retain_sparsity
REAL(real_8) :: alpha(2), beta(2), sparsities(3)
!
! parsing
IF (narg .LT. 28) THEN
WRITE (*, *) "Input file format:"
WRITE (*, *) " dbcsr_multiply \\"
WRITE (*, *) " M N K \\"
WRITE (*, *) " SpA SpB SpC \\"
WRITE (*, *) " TrA TrB \\"
WRITE (*, *) " SymA SymB SymC \\"
WRITE (*, *) " data_type \\"
WRITE (*, *) " Re(alpha) Im(alpha) Re(beta) Im(beta) \\"
WRITE (*, *) " limRowL limRowU limColL limColU limKL limKU \\"
WRITE (*, *) " retain_sparsity nrep \\"
WRITE (*, *) " nmblksizes nnblksizes nkblksizes \\"
WRITE (*, *) " [mblksizes] [nblksizes] [kblksizes]"
DBCSR_ABORT("narg not correct")
ENDIF
matrix_sizes(1) = atoi(args(2))
matrix_sizes(2) = atoi(args(3))
matrix_sizes(3) = atoi(args(4))
sparsities(1) = ator(args(5))
sparsities(2) = ator(args(6))
sparsities(3) = ator(args(7))
trans(1) = args(8)
trans(2) = args(9)
symmetries(1) = args(10)
symmetries(2) = args(11)
symmetries(3) = args(12)
TYPE = atoi(args(13))
alpha(1) = ator(args(14))
alpha(2) = ator(args(15))
beta(1) = ator(args(16))
beta(2) = ator(args(17))
limits(1) = atoi(args(18))
limits(2) = atoi(args(19))
limits(3) = atoi(args(20))
limits(4) = atoi(args(21))
limits(5) = atoi(args(22))
limits(6) = atoi(args(23))
retain_sparsity = atol(args(24))
nrep = atoi(args(25))
mblk_to_read = atoi(args(26))
nblk_to_read = atoi(args(27))
kblk_to_read = atoi(args(28))
IF (narg < 28+2*(mblk_to_read+nblk_to_read+kblk_to_read)) &
DBCSR_ABORT("narg not correct")
ALLOCATE (bs_m(2*mblk_to_read), bs_n(2*nblk_to_read), bs_k(2*kblk_to_read))
i = 28
DO iblk = 1, mblk_to_read
i = i+1
bs_m(2*(iblk-1)+1) = atoi(args(i))
i = i+1
bs_m(2*(iblk-1)+2) = atoi(args(i))
ENDDO
DO iblk = 1, nblk_to_read
i = i+1
bs_n(2*(iblk-1)+1) = atoi(args(i))
i = i+1
bs_n(2*(iblk-1)+2) = atoi(args(i))
ENDDO
DO iblk = 1, kblk_to_read
i = i+1
bs_k(2*(iblk-1)+1) = atoi(args(i))
i = i+1
bs_k(2*(iblk-1)+2) = atoi(args(i))
ENDDO
!
! do checks here
!
! if the limits are not specified (i.e 0), we set them here
IF (limits(1) .EQ. 0) limits(1) = 1
IF (limits(2) .EQ. 0) limits(2) = matrix_sizes(1)
IF (limits(3) .EQ. 0) limits(3) = 1
IF (limits(4) .EQ. 0) limits(4) = matrix_sizes(2)
IF (limits(5) .EQ. 0) limits(5) = 1
IF (limits(6) .EQ. 0) limits(6) = matrix_sizes(3)
!
! lets go !
CALL dbcsr_perf_multiply_low(group, mp_env, npdims, io_unit, matrix_sizes, &
bs_m, bs_n, bs_k, sparsities, trans, symmetries, TYPE, &
alpha, beta, limits, retain_sparsity, nrep)
DEALLOCATE (bs_m, bs_n, bs_k)
END SUBROUTINE dbcsr_perf_multiply
! **************************************************************************************************
!> \brief Performs a variety of matrix multiplies of same matrices on different
!> processor grids
!> \param[in] mp_group MPI communicator
!> \param mp_env ...
!> \param npdims ...
!> \param[in] io_unit which unit to write to, if not negative
!> \param[in] matrix_sizes size of matrices to test
!> \param bs_m block sizes of the 3 dimensions
!> \param bs_n block sizes of the 3 dimensions
!> \param bs_k block sizes of the 3 dimensions
!> \param[in] sparsities sparsities of matrices to create
!> \param trans transposes of the two matrices
!> \param symmetries ...
!> \param data_type types of matrices to create
!> \param alpha_in alpha value to use in multiply
!> \param beta_in beta value to use in multiply
!> \param limits ...
!> \param retain_sparsity ...
!> \param nrep ...
! **************************************************************************************************
SUBROUTINE dbcsr_perf_multiply_low(mp_group, mp_env, npdims, io_unit, &
matrix_sizes, bs_m, bs_n, bs_k, sparsities, trans, symmetries, data_type, &
alpha_in, beta_in, limits, retain_sparsity, nrep)
INTEGER, INTENT(IN) :: mp_group
TYPE(dbcsr_mp_obj), INTENT(IN) :: mp_env
INTEGER, DIMENSION(2), INTENT(in) :: npdims
INTEGER, INTENT(IN) :: io_unit
INTEGER, DIMENSION(:), INTENT(in) :: matrix_sizes, bs_m, bs_n, bs_k
REAL(real_8), DIMENSION(3), INTENT(in) :: sparsities
CHARACTER, DIMENSION(2), INTENT(in) :: trans
CHARACTER, DIMENSION(3), INTENT(in) :: symmetries
INTEGER, INTENT(IN) :: data_type
REAL(real_8), DIMENSION(2), INTENT(in) :: alpha_in, beta_in
INTEGER, DIMENSION(6), INTENT(in) :: limits
LOGICAL, INTENT(in) :: retain_sparsity
INTEGER, INTENT(IN) :: nrep
CHARACTER(len=*), PARAMETER :: routineN = 'dbcsr_perf_multiply_low', &
routineP = moduleN//':'//routineN
CHARACTER :: a_symm, b_symm, c_symm, transa, transb
INTEGER :: handle, mynode, numnodes
INTEGER, DIMENSION(:), POINTER :: col_dist_a, col_dist_b, col_dist_c, my_sizes_k, &
my_sizes_m, my_sizes_n, row_dist_a, row_dist_b, row_dist_c, sizes_k, sizes_m, sizes_n
LOGICAL :: do_complex
LOGICAL, DIMENSION(2) :: trs
TYPE(dbcsr_distribution_obj) :: dist_a, dist_b, dist_c
TYPE(dbcsr_scalar_type) :: alpha, beta
TYPE(dbcsr_type) :: matrix_a, matrix_b, matrix_c
! ---------------------------------------------------------------------------
CALL timeset(routineN, handle)
NULLIFY (my_sizes_k, my_sizes_m, my_sizes_n, &
sizes_k, sizes_m, sizes_n)
!
! print
CALL mp_environ(numnodes, mynode, mp_group)
IF (io_unit .GT. 0) THEN
WRITE (io_unit, *) 'numnodes', numnodes
WRITE (io_unit, *) 'matrix_sizes', matrix_sizes
WRITE (io_unit, *) 'sparsities', sparsities
WRITE (io_unit, *) 'trans ', trans
WRITE (io_unit, *) 'symmetries ', symmetries
WRITE (io_unit, *) 'type ', data_type
WRITE (io_unit, *) 'alpha_in', alpha_in
WRITE (io_unit, *) 'beta_in', beta_in
WRITE (io_unit, *) 'limits', limits
WRITE (io_unit, *) 'retain_sparsity', retain_sparsity
WRITE (io_unit, *) 'nrep', nrep
WRITE (io_unit, *) 'bs_m', bs_m
WRITE (io_unit, *) 'bs_n', bs_n
WRITE (io_unit, *) 'bs_k', bs_k
ENDIF
!
CALL dbcsr_reset_randmat_seed()
!
a_symm = symmetries(1)
b_symm = symmetries(2)
c_symm = symmetries(3)
IF (a_symm .NE. dbcsr_type_no_symmetry .AND. matrix_sizes(1) .NE. matrix_sizes(3)) &
DBCSR_ABORT("")
IF (b_symm .NE. dbcsr_type_no_symmetry .AND. matrix_sizes(2) .NE. matrix_sizes(3)) &
DBCSR_ABORT("")
IF (c_symm .NE. dbcsr_type_no_symmetry .AND. matrix_sizes(1) .NE. matrix_sizes(2)) &
DBCSR_ABORT("")
do_complex = data_type .EQ. dbcsr_type_complex_4 .OR. data_type .EQ. dbcsr_type_complex_8
SELECT CASE (data_type)
CASE (dbcsr_type_real_4)
alpha = dbcsr_scalar(REAL(alpha_in(1), real_4))
beta = dbcsr_scalar(REAL(beta_in(1), real_4))
CASE (dbcsr_type_real_8)
alpha = dbcsr_scalar(REAL(alpha_in(1), real_8))
beta = dbcsr_scalar(REAL(beta_in(1), real_8))
CASE (dbcsr_type_complex_4)
alpha = dbcsr_scalar(CMPLX(alpha_in(1), alpha_in(2), real_4))
beta = dbcsr_scalar(CMPLX(beta_in(1), beta_in(2), real_4))
CASE (dbcsr_type_complex_8)
alpha = dbcsr_scalar(CMPLX(alpha_in(1), alpha_in(2), real_8))
beta = dbcsr_scalar(CMPLX(beta_in(1), beta_in(2), real_8))
END SELECT
transa = trans(1)
transb = trans(2)
!
! if C has a symmetry, we need special transpositions
IF (c_symm .NE. dbcsr_type_no_symmetry) THEN
IF (.NOT. (transa .EQ. dbcsr_no_transpose .AND. &
transb .EQ. dbcsr_transpose .OR. &
transa .EQ. dbcsr_transpose .AND. &
transb .EQ. dbcsr_no_transpose .OR. &
transa .EQ. dbcsr_no_transpose .AND. &
transb .EQ. dbcsr_conjugate_transpose .AND. &
.NOT. do_complex .OR. &
transa .EQ. dbcsr_conjugate_transpose .AND. &
transb .EQ. dbcsr_no_transpose .AND. &
.NOT. do_complex)) THEN
DBCSR_ABORT("")
ENDIF
ENDIF
!
! if C has symmetry and special limits
IF (c_symm .NE. dbcsr_type_no_symmetry) THEN
IF (limits(1) .NE. 1 .OR. limits(2) .NE. matrix_sizes(1) .OR. &
limits(3) .NE. 1 .OR. limits(4) .NE. matrix_sizes(2)) THEN
DBCSR_ABORT("")
ENDIF
ENDIF
!
! Create the row/column block sizes.
CALL dbcsr_make_random_block_sizes(sizes_m, matrix_sizes(1), bs_m)
CALL dbcsr_make_random_block_sizes(sizes_n, matrix_sizes(2), bs_n)
CALL dbcsr_make_random_block_sizes(sizes_k, matrix_sizes(3), bs_k)
!
! if we have symmetry the row and column block sizes hae to match
IF (c_symm .NE. dbcsr_type_no_symmetry .AND. a_symm .NE. dbcsr_type_no_symmetry .AND. &
b_symm .NE. dbcsr_type_no_symmetry) THEN
my_sizes_m => sizes_m; my_sizes_n => sizes_m; my_sizes_k => sizes_m
ELSEIF ((c_symm .EQ. dbcsr_type_no_symmetry .AND. a_symm .NE. dbcsr_type_no_symmetry .AND. &
b_symm .NE. dbcsr_type_no_symmetry) .OR. &
(c_symm .NE. dbcsr_type_no_symmetry .AND. a_symm .EQ. dbcsr_type_no_symmetry .AND. &
b_symm .NE. dbcsr_type_no_symmetry) .OR. &
(c_symm .NE. dbcsr_type_no_symmetry .AND. a_symm .NE. dbcsr_type_no_symmetry .AND. &
b_symm .EQ. dbcsr_type_no_symmetry)) THEN
my_sizes_m => sizes_m; my_sizes_n => sizes_m; my_sizes_k => sizes_m
ELSEIF (c_symm .EQ. dbcsr_type_no_symmetry .AND. a_symm .EQ. dbcsr_type_no_symmetry .AND. &
b_symm .NE. dbcsr_type_no_symmetry) THEN
my_sizes_m => sizes_m; my_sizes_n => sizes_n; my_sizes_k => sizes_n
ELSEIF (c_symm .EQ. dbcsr_type_no_symmetry .AND. a_symm .NE. dbcsr_type_no_symmetry .AND. &
b_symm .EQ. dbcsr_type_no_symmetry) THEN
my_sizes_m => sizes_m; my_sizes_n => sizes_n; my_sizes_k => sizes_m
ELSEIF (c_symm .NE. dbcsr_type_no_symmetry .AND. a_symm .EQ. dbcsr_type_no_symmetry .AND. &
b_symm .EQ. dbcsr_type_no_symmetry) THEN
my_sizes_m => sizes_m; my_sizes_n => sizes_m; my_sizes_k => sizes_k
ELSEIF (c_symm .EQ. dbcsr_type_no_symmetry .AND. a_symm .EQ. dbcsr_type_no_symmetry .AND. &
b_symm .EQ. dbcsr_type_no_symmetry) THEN
my_sizes_m => sizes_m; my_sizes_n => sizes_n; my_sizes_k => sizes_k
ELSE
DBCSR_ABORT("something wrong here...")
ENDIF
! Create the random matrices.
trs(1) = transa .EQ. dbcsr_no_transpose
trs(2) = transb .EQ. dbcsr_no_transpose
CALL dbcsr_dist_bin(row_dist_c, SIZE(sizes_m), npdims(1), &
sizes_m)
CALL dbcsr_dist_bin(col_dist_c, SIZE(sizes_n), npdims(2), &
sizes_n)
CALL dbcsr_distribution_new(dist_c, mp_env, row_dist_c, col_dist_c)
CALL dbcsr_make_random_matrix(matrix_c, sizes_m, sizes_n, "Matrix C", &
REAL(sparsities(3), real_8), &
mp_group, data_type=data_type, dist=dist_c)
CALL dbcsr_distribution_release(dist_c)
IF (trs(1)) THEN
CALL dbcsr_dist_bin(row_dist_a, SIZE(sizes_k), npdims(1), &
sizes_k)
CALL dbcsr_dist_bin(col_dist_a, SIZE(sizes_m), npdims(2), &
sizes_m)
CALL dbcsr_distribution_new(dist_a, mp_env, row_dist_a, col_dist_a)
CALL dbcsr_make_random_matrix(matrix_a, sizes_k, sizes_m, "Matrix A", &
REAL(sparsities(1), real_8), &
mp_group, data_type=data_type, dist=dist_a)
DEALLOCATE (row_dist_a, col_dist_a)
ELSE
CALL dbcsr_dist_bin(col_dist_a, SIZE(sizes_k), npdims(2), &
sizes_k)
CALL dbcsr_distribution_new(dist_a, mp_env, row_dist_c, col_dist_a)
CALL dbcsr_make_random_matrix(matrix_a, sizes_m, sizes_k, "Matrix A", &
REAL(sparsities(1), real_8), &
mp_group, data_type=data_type, dist=dist_a)
DEALLOCATE (col_dist_a)
ENDIF
CALL dbcsr_distribution_release(dist_a)
IF (trs(2)) THEN
CALL dbcsr_dist_bin(row_dist_b, SIZE(sizes_n), npdims(1), &
sizes_n)
CALL dbcsr_dist_bin(col_dist_b, SIZE(sizes_k), npdims(2), &
sizes_k)
CALL dbcsr_distribution_new(dist_b, mp_env, row_dist_b, col_dist_b)
CALL dbcsr_make_random_matrix(matrix_b, sizes_n, sizes_k, "Matrix B", &
REAL(sparsities(2), real_8), &
mp_group, data_type=data_type, dist=dist_b)
DEALLOCATE (row_dist_b, col_dist_b)
ELSE
CALL dbcsr_dist_bin(row_dist_b, SIZE(sizes_k), npdims(1), &
sizes_k)
CALL dbcsr_distribution_new(dist_b, mp_env, row_dist_b, col_dist_c)
CALL dbcsr_make_random_matrix(matrix_b, sizes_k, sizes_n, "Matrix B", &
REAL(sparsities(2), real_8), &
mp_group, data_type=data_type, dist=dist_b)
DEALLOCATE (row_dist_b)
ENDIF
CALL dbcsr_distribution_release(dist_b)
DEALLOCATE (row_dist_c, col_dist_c, sizes_m, sizes_n, sizes_k)
!
! if C has a symmetry, we build it accordingly, i.e. C=A*A and C=A*(-A)
IF (c_symm .NE. dbcsr_type_no_symmetry) THEN
CALL dbcsr_copy(matrix_b, matrix_a)
!print*, a_symm,b_symm,dbcsr_get_matrix_type(matrix_a),dbcsr_get_matrix_type(matrix_b)
IF (c_symm .EQ. dbcsr_type_antisymmetric) THEN
CALL dbcsr_scale(matrix_b, &
alpha_scalar=dbcsr_scalar_negative( &
dbcsr_scalar_one(data_type)))
ENDIF
ENDIF
!
! Prepare test parameters
CALL perf_multiply(mp_group, mp_env, io_unit, &
matrix_a, matrix_b, matrix_c, &
transa, transb, &
alpha, beta, &
limits, retain_sparsity, &
nrep)
!
! cleanup
CALL dbcsr_release(matrix_a)
CALL dbcsr_release(matrix_b)
CALL dbcsr_release(matrix_c)
CALL timestop(handle)
END SUBROUTINE dbcsr_perf_multiply_low
! **************************************************************************************************
!> \brief Performs a variety of matrix multiplies of same matrices on different
!> processor grids
!> \param[in] mp_group MPI communicator
!> \param mp_env ...
!> \param[in] io_unit which unit to write to, if not negative
!> \param matrix_a matrices to multiply
!> \param matrix_b matrices to multiply
!> \param matrix_c matrices to multiply
!> \param transa ...
!> \param transb ...
!> \param alpha ...
!> \param beta ...
!> \param limits ...
!> \param retain_sparsity ...
!> \param nrep ...
! **************************************************************************************************
SUBROUTINE perf_multiply(mp_group, mp_env, io_unit, &
matrix_a, matrix_b, matrix_c, &
transa, transb, alpha, beta, limits, retain_sparsity, &
nrep)
INTEGER, INTENT(IN) :: mp_group
TYPE(dbcsr_mp_obj), INTENT(IN) :: mp_env
INTEGER, INTENT(IN) :: io_unit
TYPE(dbcsr_type), INTENT(in) :: matrix_a, matrix_b, matrix_c
CHARACTER, INTENT(in) :: transa, transb
TYPE(dbcsr_scalar_type), INTENT(in) :: alpha, beta
INTEGER, DIMENSION(6), INTENT(in) :: limits
LOGICAL, INTENT(in) :: retain_sparsity
INTEGER, INTENT(IN) :: nrep
CHARACTER(len=*), PARAMETER :: routineN = 'perf_multiply', routineP = moduleN//':'//routineN
INTEGER :: c_a, c_b, c_c, handle, irep, mynode, &
nthreads, numnodes, r_a, r_b, r_c
INTEGER(int_8) :: flop
INTEGER(int_8), ALLOCATABLE, DIMENSION(:) :: flop_sum
INTEGER(int_8), ALLOCATABLE, DIMENSION(:, :) :: flops
INTEGER, DIMENSION(:), POINTER :: blk_offsets, col_dist_a, col_dist_b, &
col_dist_c, row_dist_a, row_dist_b, &
row_dist_c
REAL(real_8) :: chksum_a, chksum_b, chksum_c_in, &
chksum_c_out, std_all, std_t, t1, t2
REAL(real_8), ALLOCATABLE, DIMENSION(:) :: flops_all, flops_node, flops_thread, &
load_imb, t, t_max, t_min
REAL(real_8), ALLOCATABLE, DIMENSION(:, :) :: times
TYPE(dbcsr_distribution_obj) :: dist_a, dist_b, dist_c
TYPE(dbcsr_type) :: m_a, m_b, m_c, m_c_orig
! ---------------------------------------------------------------------------
CALL timeset(routineN, handle)
CALL mp_environ(numnodes, mynode, mp_group)
nthreads = 1
!$ nthreads = OMP_GET_MAX_THREADS()
ALLOCATE (times(0:numnodes-1, nrep), flops(0:numnodes-1, nrep), t_max(nrep), &
flops_node(nrep), flops_thread(nrep), flops_all(nrep), flop_sum(nrep), &
t_min(nrep), t(nrep), load_imb(nrep))
times(:, :) = 0.0_real_8
t_max(:) = 0.0_real_8
t_min(:) = 0.0_real_8
t(:) = 0.0_real_8
flops_node(:) = 0.0_real_8
flops_thread(:) = 0.0_real_8
flops_all(:) = 0.0_real_8
flops(:, :) = 0
flop_sum(:) = 0
load_imb(:) = 0.0_real_8
! Row & column distributions
row_dist_c => dbcsr_distribution_row_dist(dbcsr_distribution(matrix_c))
col_dist_c => dbcsr_distribution_col_dist(dbcsr_distribution(matrix_c))
row_dist_a => row_dist_c
col_dist_a => dbcsr_distribution_col_dist(dbcsr_distribution(matrix_a))
row_dist_b => dbcsr_distribution_row_dist(dbcsr_distribution(matrix_b))
col_dist_b => col_dist_c
CALL dbcsr_distribution_new(dist_a, mp_env, row_dist_a, col_dist_a)
CALL dbcsr_distribution_new(dist_b, mp_env, row_dist_b, col_dist_b)
CALL dbcsr_distribution_new(dist_c, mp_env, row_dist_c, col_dist_c)
! Redistribute the matrices
! A
CALL dbcsr_create(m_a, "Test for "//TRIM(dbcsr_name(matrix_a)), &
dist_a, dbcsr_get_matrix_type(matrix_a), &
dbcsr_row_block_sizes(matrix_a), &
dbcsr_col_block_sizes(matrix_a), &
data_type=dbcsr_get_data_type(matrix_a))
CALL dbcsr_distribution_release(dist_a)
CALL dbcsr_redistribute(matrix_a, m_a)
! B
CALL dbcsr_create(m_b, "Test for "//TRIM(dbcsr_name(matrix_b)), &
dist_b, dbcsr_get_matrix_type(matrix_b), &
dbcsr_row_block_sizes(matrix_b), &
dbcsr_col_block_sizes(matrix_b), &
data_type=dbcsr_get_data_type(matrix_b))
CALL dbcsr_distribution_release(dist_b)
CALL dbcsr_redistribute(matrix_b, m_b)
! C orig
CALL dbcsr_create(m_c_orig, "Test for "//TRIM(dbcsr_name(matrix_c)), &
dist_c, dbcsr_get_matrix_type(matrix_c), &
dbcsr_row_block_sizes(matrix_c), &
dbcsr_col_block_sizes(matrix_c), &
data_type=dbcsr_get_data_type(matrix_c))
CALL dbcsr_distribution_release(dist_c)
CALL dbcsr_redistribute(matrix_c, m_c_orig)
! C
CALL dbcsr_create(m_c, "Test for "//TRIM(dbcsr_name(matrix_c)), &
dist_c, dbcsr_get_matrix_type(matrix_c), &
dbcsr_row_block_sizes(matrix_c), &
dbcsr_col_block_sizes(matrix_c), &
data_type=dbcsr_get_data_type(matrix_c))
CALL dbcsr_finalize(m_c)
IF (.FALSE.) THEN
blk_offsets => dbcsr_row_block_offsets(matrix_c)
WRITE (*, *) 'row_block_offsets(matrix_c)', blk_offsets
blk_offsets => dbcsr_col_block_offsets(matrix_c)
WRITE (*, *) 'col_block_offsets(matrix_c)', blk_offsets
ENDIF
IF (.FALSE.) THEN
CALL dbcsr_print(m_c, matlab_format=.FALSE., variable_name='c_in_')
CALL dbcsr_print(m_a, matlab_format=.FALSE., variable_name='a_')
CALL dbcsr_print(m_b, matlab_format=.FALSE., variable_name='b_')
CALL dbcsr_print(m_c, matlab_format=.FALSE., variable_name='c_out_')
ENDIF
r_a = dbcsr_nfullrows_total(m_a)
c_a = dbcsr_nfullcols_total(m_a)
r_b = dbcsr_nfullrows_total(m_b)
c_b = dbcsr_nfullcols_total(m_b)
r_c = dbcsr_nfullrows_total(m_c_orig)
c_c = dbcsr_nfullcols_total(m_c_orig)
chksum_a = dbcsr_checksum(m_a)
chksum_b = dbcsr_checksum(m_b)
chksum_c_in = dbcsr_checksum(m_c_orig)
!
!
DO irep = 1, nrep
!
! set the C matrix
CALL dbcsr_copy(m_c, m_c_orig)
!
! Perform multiply
CALL mp_sync(mp_group)
t1 = m_walltime()
flop = 0
CALL dbcsr_multiply(transa, transb, alpha, &
m_a, m_b, beta, m_c, &
first_row=limits(1), &
last_row=limits(2), &
first_column=limits(3), &
last_column=limits(4), &
first_k=limits(5), &
last_k=limits(6), &
retain_sparsity=retain_sparsity, &
flop=flop)
t2 = m_walltime()
times(mynode, irep) = t2-t1
flops(mynode, irep) = flop
ENDDO
chksum_c_out = dbcsr_checksum(m_c)
CALL mp_sum(times, 0, mp_group)
CALL mp_sum(flops, 0, mp_group)
!
!
t_max(:) = MAXVAL(times, DIM=1)
t_min(:) = MINVAL(times, DIM=1)
t(:) = SUM(times, DIM=1)/REAL(numnodes, real_8)
flop_sum(:) = SUM(flops, DIM=1)
t_max(:) = MAX(t_max(:), 0.001_real_8)
flops_all(:) = REAL(flop_sum(:), KIND=real_8)/t_max(:) !* 1.0e-9_real_8
flops_node(:) = flops_all(:)/REAL(numnodes, real_8)
flops_thread(:) = flops_node(:)/REAL(nthreads, real_8)
load_imb(:) = t_max(:)-t(:)/REAL(numnodes, real_8)
std_t = 1.0_real_8
std_all = 1.0_real_8
IF (io_unit .GT. 0) THEN
WRITE (io_unit, *) REPEAT("*", 80)
WRITE (io_unit, *) " -- PERF dbcsr_multiply (", transa, ", ", transb, &
", ", dbcsr_get_data_type(m_a), &
", ", dbcsr_get_matrix_type(m_a), &
", ", dbcsr_get_matrix_type(m_b), &
", ", dbcsr_get_matrix_type(m_c), &
")"
WRITE (io_unit, '(T4,3(A,I6,A,I6),A)') &
"matrix sizes A(", r_a, " x", c_a, "), B(", r_b, " x", c_b, ") and C(", r_c, " x", c_c, ")"
WRITE (io_unit, '(T4,A,I5)') 'numnodes = ', numnodes
WRITE (io_unit, '(T4,A,I5)') 'nthreads = ', nthreads
WRITE (io_unit, '(T4,A,E26.15)') 'checksum(A) = ', chksum_a
WRITE (io_unit, '(T4,A,E26.15)') 'checksum(B) = ', chksum_b
WRITE (io_unit, '(T4,A,E26.15)') 'checksum(C_in) = ', chksum_c_in
WRITE (io_unit, '(T4,A,E26.15)') 'checksum(C_out) = ', chksum_c_out
WRITE (io_unit, *)
WRITE (io_unit, *)
WRITE (io_unit, '(T4,A)') " mean std minmin maxmax"
WRITE (io_unit, '(T4,A,4EN12.2,A)') "time = ", mean(t), std(t), &
MINVAL(times), MAXVAL(times), ' seconds'
WRITE (io_unit, '(T4,A,4EN12.2,A)') "perf total = ", mean(flops_all), std(flops_all), &
MINVAL(flops_all), MAXVAL(flops_all), ' FLOPS'
WRITE (io_unit, '(T4,A,4EN12.2,A)') "perf per node = ", mean(flops_node), std(flops_node), &
MINVAL(flops_node), MAXVAL(flops_node), ' FLOPS'
WRITE (io_unit, '(T4,A,4EN12.2,A)') "perf per thread = ", mean(flops_thread), std(flops_thread), &
MINVAL(flops_thread), MAXVAL(flops_thread), ' FLOPS'
WRITE (io_unit, '(T4,A,4E12.2,A)') "load imbalance = ", mean(load_imb), std(load_imb), &
MINVAL(load_imb), MAXVAL(load_imb), ''
WRITE (io_unit, '(T4,A,4E12.2,A)') "rel load imbal = ", mean(load_imb/t_max), std(load_imb/t_max), &
MINVAL(load_imb/t_max), MAXVAL(load_imb/t_max), ''
WRITE (io_unit, *) REPEAT("*", 80)
ENDIF
CALL dbcsr_release(m_a)
CALL dbcsr_release(m_b)
CALL dbcsr_release(m_c)
CALL dbcsr_release(m_c_orig)
DEALLOCATE (times, flops, t_max, flops_node, flops_thread, flops_all, &
flop_sum, t_min, t, load_imb)
CALL timestop(handle)
END SUBROUTINE perf_multiply
! **************************************************************************************************
!> \brief ...
!> \param v ...
!> \return ...
! **************************************************************************************************
FUNCTION mean(v)
REAL(real_8), DIMENSION(:) :: v
REAL(real_8) :: mean
INTEGER :: i, n
mean = 0.0_real_8
n = SIZE(v, 1)
DO i = 1, n
mean = mean+v(i)
ENDDO
mean = mean/REAL(n, real_8)
END FUNCTION mean
! **************************************************************************************************
!> \brief ...
!> \param v ...
!> \return ...
! **************************************************************************************************
FUNCTION std(v)
REAL(real_8), DIMENSION(:) :: v
REAL(real_8) :: std
INTEGER :: i, n
REAL(real_8) :: mn
mn = mean(v)
std = 0.0_real_8
n = SIZE(v, 1)
DO i = 1, n
std = std+(v(i)-mn)**2
ENDDO
std = SQRT(std)/REAL(n, real_8)
END FUNCTION std
END MODULE dbcsr_performance_multiply