-
Notifications
You must be signed in to change notification settings - Fork 573
/
KokkosSparse_sptrsv_solve_impl.hpp
3466 lines (2948 loc) · 140 KB
/
KokkosSparse_sptrsv_solve_impl.hpp
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
/*
//@HEADER
// ************************************************************************
//
// Kokkos v. 3.0
// Copyright (2020) National Technology & Engineering
// Solutions of Sandia, LLC (NTESS).
//
// Under the terms of Contract DE-NA0003525 with NTESS,
// the U.S. Government retains certain rights in this software.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. 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.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY NTESS "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 NTESS OR THE
// CONTRIBUTORS 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.
//
// Questions? Contact Siva Rajamanickam (srajama@sandia.gov)
//
// ************************************************************************
//@HEADER
*/
#ifndef KOKKOSSPARSE_IMPL_SPTRSV_SOLVE_HPP_
#define KOKKOSSPARSE_IMPL_SPTRSV_SOLVE_HPP_
/// \file KokkosSparse_impl_sptrsv.hpp
/// \brief Implementation(s) of sparse triangular solve.
#include <KokkosKernels_config.h>
#include <Kokkos_ArithTraits.hpp>
#include <KokkosSparse_sptrsv_handle.hpp>
#include <KokkosSparse_spmv.hpp>
#include <KokkosSparse_CrsMatrix.hpp>
#ifdef KOKKOSKERNELS_ENABLE_SUPERNODAL_SPTRSV
// Enable supernodal sptrsv
#include "KokkosBlas2_gemv.hpp"
#include "KokkosBlas2_team_gemv.hpp"
#include "KokkosBlas3_trsm.hpp"
#include "KokkosSparse_spmv.hpp"
#include "KokkosBatched_Util.hpp"
#include "KokkosBatched_Gemv_Decl.hpp"
#include "KokkosBatched_Gemv_Team_Impl.hpp"
#include "KokkosBatched_Gemv_Serial_Impl.hpp"
#include "KokkosBatched_Trsm_Team_Impl.hpp"
#endif
//#define SERIAL_FOR_LOOP
#define KOKKOSKERNELS_SPTRSV_TRILVLSCHED
//#define KOKKOSPSTRSV_SOLVE_IMPL_PROFILE 1
#if defined(KOKKOS_ENABLE_CUDA) && defined(KOKKOSPSTRSV_SOLVE_IMPL_PROFILE)
#include "cuda_profiler_api.h"
#endif
namespace KokkosSparse {
namespace Impl {
namespace Experimental {
#if defined(KOKKOS_ENABLE_CUDA) && 10000 < CUDA_VERSION && defined(KOKKOSKERNELS_ENABLE_EXP_CUDAGRAPH)
#define KOKKOSKERNELS_SPTRSV_CUDAGRAPHSUPPORT
#endif
struct UnsortedTag {};
struct LargerCutoffTag {};
struct UnsortedLargerCutoffTag {};
template <class ViewType>
void print_view1d_solve(const ViewType dv, size_t range = 0) {
auto v = Kokkos::create_mirror_view(dv);
Kokkos::deep_copy(v, dv);
std::cout << "Output for view " << v.label() << std::endl;
range = range == 0 ? dv.extent(0) : range;
for (size_t i = 0; i < range; ++i) {
std::cout << "v(" << i << ") = " << v(i) << " , ";
}
std::cout << std::endl;
}
// Needed for cudagraphs
struct EmptyFunctor {
KOKKOS_INLINE_FUNCTION
void operator()(const int) const {}
};
// This functor unifies the lower and upper implementations, the hope is the "is_lowertri" check does not add noticable time on larger problems
template <class RowMapType, class EntriesType, class ValuesType, class LHSType, class RHSType, class NGBLType>
struct TriLvlSchedTP1SolverFunctor
{
typedef typename RowMapType::execution_space execution_space;
typedef Kokkos::TeamPolicy<execution_space> policy_type;
typedef typename policy_type::member_type member_type;
typedef typename EntriesType::non_const_value_type lno_t;
typedef typename ValuesType::non_const_value_type scalar_t;
RowMapType row_map;
EntriesType entries;
ValuesType values;
LHSType lhs;
RHSType rhs;
NGBLType nodes_grouped_by_level;
const bool is_lowertri;
long node_count; // like "block" offset into ngbl, my_league is the "local" offset
long dense_nrows;
TriLvlSchedTP1SolverFunctor(const RowMapType &row_map_, const EntriesType &entries_, const ValuesType &values_, LHSType &lhs_, const RHSType &rhs_, const NGBLType &nodes_grouped_by_level_, const bool is_lowertri_, long node_count_, long dense_nrows_ = 0) :
row_map(row_map_), entries(entries_), values(values_), lhs(lhs_), rhs(rhs_), nodes_grouped_by_level(nodes_grouped_by_level_), is_lowertri(is_lowertri_), node_count(node_count_), dense_nrows(dense_nrows_) {}
KOKKOS_INLINE_FUNCTION
void operator()( const member_type & team ) const {
auto my_league = team.league_rank(); // map to rowid
auto rowid = nodes_grouped_by_level(my_league + node_count);
auto my_rank = team.team_rank();
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
Kokkos::parallel_reduce( Kokkos::TeamThreadRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
}, diff );
team.team_barrier();
// At end, finalize rowid == colid
// only one thread should do this; can also use Kokkos::single
if ( my_rank == 0 )
{
// ASSUMPTION: sorted diagonal value located at eoffset - 1
lhs(rowid) = is_lowertri ? (rhs_rowid+diff)/values(eoffset-1) : (rhs_rowid+diff)/values(soffset);
}
}
KOKKOS_INLINE_FUNCTION
void operator()(const UnsortedTag&, const member_type & team) const {
auto my_league = team.league_rank(); // map to rowid
auto rowid = nodes_grouped_by_level(my_league + node_count);
auto my_rank = team.team_rank();
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
auto diag = -1;
Kokkos::parallel_reduce( Kokkos::TeamThreadRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
else {
diag = ptr;
}
}, diff );
team.team_barrier();
// At end, finalize rowid == colid
// only one thread should do this; can also use Kokkos::single
if ( my_rank == 0 )
{
lhs(rowid) = (rhs_rowid+diff)/values(diag);
}
}
};
template <class RowMapType, class EntriesType, class ValuesType, class LHSType, class RHSType, class NGBLType>
struct TriLvlSchedTP1SolverFunctorDiagValues
{
typedef typename RowMapType::execution_space execution_space;
typedef Kokkos::TeamPolicy<execution_space> policy_type;
typedef typename policy_type::member_type member_type;
typedef typename EntriesType::non_const_value_type lno_t;
typedef typename ValuesType::non_const_value_type scalar_t;
RowMapType row_map;
EntriesType entries;
ValuesType values;
LHSType lhs;
RHSType rhs;
NGBLType nodes_grouped_by_level;
ValuesType diagonal_values; // inserted according to rowid
const bool is_lowertri;
long node_count; // like "block" offset into ngbl, my_league is the "local" offset
long dense_nrows;
TriLvlSchedTP1SolverFunctorDiagValues( const RowMapType &row_map_, const EntriesType &entries_, const ValuesType &values_, LHSType &lhs_, const RHSType &rhs_, const NGBLType &nodes_grouped_by_level_, const ValuesType &diagonal_values_, const bool is_lowertri_, long node_count_, long dense_nrows_ = 0) :
row_map(row_map_), entries(entries_), values(values_), lhs(lhs_), rhs(rhs_), nodes_grouped_by_level(nodes_grouped_by_level_), diagonal_values(diagonal_values_), is_lowertri(is_lowertri_), node_count(node_count_), dense_nrows(dense_nrows_) {}
KOKKOS_INLINE_FUNCTION
void operator()( const member_type & team ) const {
auto my_league = team.league_rank(); // map to rowid
auto rowid = nodes_grouped_by_level(my_league + node_count);
auto my_rank = team.team_rank();
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
Kokkos::parallel_reduce( Kokkos::TeamThreadRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
}, diff );
team.team_barrier();
// At end, finalize rowid == colid
// only one thread should do this; can also use Kokkos::single
if ( my_rank == 0 )
{
//lhs(rowid) = is_lowertri ? (rhs_rowid+diff)/values(eoffset-1) : (rhs_rowid+diff)/values(soffset);
lhs(rowid) = (rhs_rowid+diff)/diagonal_values(rowid);
}
}
};
template <class RowMapType, class EntriesType, class ValuesType, class LHSType, class RHSType, class NGBLType>
struct TriLvlSchedTP2SolverFunctor
{
typedef typename RowMapType::execution_space execution_space;
typedef Kokkos::TeamPolicy<execution_space> policy_type;
typedef typename policy_type::member_type member_type;
typedef typename EntriesType::non_const_value_type lno_t;
typedef typename ValuesType::non_const_value_type scalar_t;
RowMapType row_map;
EntriesType entries;
ValuesType values;
LHSType lhs;
RHSType rhs;
NGBLType nodes_grouped_by_level;
const bool is_lowertri;
long node_count; // like "block" offset into ngbl, my_league is the "local" offset
long node_groups;
long dense_nrows;
TriLvlSchedTP2SolverFunctor(const RowMapType &row_map_, const EntriesType &entries_, const ValuesType &values_, LHSType &lhs_, const RHSType &rhs_, const NGBLType &nodes_grouped_by_level_, const bool is_lowertri_, long node_count_, long node_groups_ = 0, long dense_nrows_ = 0) :
row_map(row_map_), entries(entries_), values(values_), lhs(lhs_), rhs(rhs_), nodes_grouped_by_level(nodes_grouped_by_level_), is_lowertri(is_lowertri_), node_count(node_count_), node_groups(node_groups_), dense_nrows(dense_nrows_) {}
KOKKOS_INLINE_FUNCTION
void operator()(const member_type & team) const {
auto my_league = team.league_rank(); // map to rowid
size_t nrows = row_map.extent(0) - 1;
Kokkos::parallel_for( Kokkos::TeamThreadRange( team, 0, node_groups ), [&] ( const long ng ) {
auto rowid = nodes_grouped_by_level(node_count + my_league*node_groups + ng);
if ( size_t(rowid) < nrows ) {
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
Kokkos::parallel_reduce( Kokkos::ThreadVectorRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
}, diff );
// ASSUMPTION: sorted diagonal value located at eoffset - 1
lhs(rowid) = is_lowertri ? (rhs_rowid+diff)/values(eoffset-1) : (rhs_rowid+diff)/values(soffset);
} // end if
}); // end TeamThreadRange
team.team_barrier();
}
KOKKOS_INLINE_FUNCTION
void operator()(const UnsortedTag&, const member_type & team) const {
auto my_league = team.league_rank(); // map to rowid
size_t nrows = row_map.extent(0) - 1;
Kokkos::parallel_for( Kokkos::TeamThreadRange( team, 0, node_groups ), [&] ( const long ng ) {
auto rowid = nodes_grouped_by_level(node_count + my_league*node_groups + ng);
if ( size_t(rowid) < nrows ) {
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
auto diag = -1;
Kokkos::parallel_reduce( Kokkos::ThreadVectorRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
else {
diag = ptr;
}
}, diff );
lhs(rowid) = (rhs_rowid+diff)/values(diag);
} // end if
}); // end TeamThreadRange
team.team_barrier();
}
};
// Lower vs Upper Multi-block Functors
template <class RowMapType, class EntriesType, class ValuesType, class LHSType, class RHSType, class NGBLType>
struct LowerTriLvlSchedRPSolverFunctor
{
typedef typename EntriesType::non_const_value_type lno_t;
RowMapType row_map;
EntriesType entries;
ValuesType values;
LHSType lhs;
RHSType rhs;
NGBLType nodes_grouped_by_level;
LowerTriLvlSchedRPSolverFunctor( const RowMapType &row_map_, const EntriesType &entries_, const ValuesType &values_, LHSType &lhs_, const RHSType &rhs_, NGBLType nodes_grouped_by_level_ ) :
row_map(row_map_), entries(entries_), values(values_), lhs(lhs_), rhs(rhs_), nodes_grouped_by_level(nodes_grouped_by_level_) {}
KOKKOS_INLINE_FUNCTION
void operator()(const lno_t i) const {
auto rowid = nodes_grouped_by_level(i);
// Assuming indices are sorted per row, diag entry is final index in the list
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
for ( auto ptr = soffset; ptr < eoffset; ++ptr ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
rhs_rowid = rhs_rowid - val*lhs(colid);
}
else {
lhs(rowid) = rhs_rowid/val;
}
} // end for ptr
}
KOKKOS_INLINE_FUNCTION
void operator()(const UnsortedTag&, const lno_t i) const {
auto rowid = nodes_grouped_by_level(i);
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
auto diag = -1;
for ( auto ptr = soffset; ptr < eoffset; ++ptr ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
rhs_rowid = rhs_rowid - val*lhs(colid);
}
else {
diag = ptr;
}
} // end for ptr
lhs(rowid) = rhs_rowid/values(diag);
}
};
template <class RowMapType, class EntriesType, class ValuesType, class LHSType, class RHSType, class NGBLType>
struct LowerTriLvlSchedTP1SolverFunctor
{
typedef typename RowMapType::execution_space execution_space;
typedef Kokkos::TeamPolicy<execution_space> policy_type;
typedef typename policy_type::member_type member_type;
typedef typename EntriesType::non_const_value_type lno_t;
typedef typename ValuesType::non_const_value_type scalar_t;
RowMapType row_map;
EntriesType entries;
ValuesType values;
LHSType lhs;
RHSType rhs;
NGBLType nodes_grouped_by_level;
long node_count; // like "block" offset into ngbl, my_league is the "local" offset
long node_groups;
LowerTriLvlSchedTP1SolverFunctor( const RowMapType &row_map_, const EntriesType &entries_, const ValuesType &values_, LHSType &lhs_, const RHSType &rhs_, const NGBLType &nodes_grouped_by_level_, long node_count_, long node_groups_ = 0) :
row_map(row_map_), entries(entries_), values(values_), lhs(lhs_), rhs(rhs_), nodes_grouped_by_level(nodes_grouped_by_level_), node_count(node_count_), node_groups(node_groups_) {}
KOKKOS_INLINE_FUNCTION
void operator()( const member_type & team ) const {
auto my_league = team.league_rank(); // map to rowid
auto rowid = nodes_grouped_by_level(my_league + node_count);
auto my_rank = team.team_rank();
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
Kokkos::parallel_reduce( Kokkos::TeamThreadRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
}, diff );
team.team_barrier();
// At end, finalize rowid == colid
// only one thread should do this; can also use Kokkos::single
if ( my_rank == 0 )
{
// ASSUMPTION: sorted diagonal value located at eoffset - 1
lhs(rowid) = (rhs_rowid+diff)/values(eoffset-1);
}
}
KOKKOS_INLINE_FUNCTION
void operator()(const UnsortedTag&, const member_type & team) const {
auto my_league = team.league_rank(); // map to rowid
auto rowid = nodes_grouped_by_level(my_league + node_count);
auto my_rank = team.team_rank();
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
auto diag = -1;
Kokkos::parallel_reduce( Kokkos::TeamThreadRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
else {
diag = ptr;
}
}, diff );
team.team_barrier();
// At end, finalize rowid == colid
// only one thread should do this; can also use Kokkos::single
if ( my_rank == 0 )
{
lhs(rowid) = (rhs_rowid+diff)/values(diag);
}
}
};
// FIXME CUDA: This algorithm not working with all integral type combos
// In any case, this serves as a skeleton for 3-level hierarchical parallelism for alg dev
template <class RowMapType, class EntriesType, class ValuesType, class LHSType, class RHSType, class NGBLType>
struct LowerTriLvlSchedTP2SolverFunctor
{
typedef typename RowMapType::execution_space execution_space;
typedef Kokkos::TeamPolicy<execution_space> policy_type;
typedef typename policy_type::member_type member_type;
typedef typename EntriesType::non_const_value_type lno_t;
typedef typename ValuesType::non_const_value_type scalar_t;
RowMapType row_map;
EntriesType entries;
ValuesType values;
LHSType lhs;
RHSType rhs;
NGBLType nodes_grouped_by_level;
long node_count; // like "block" offset into ngbl, my_league is the "local" offset
long node_groups;
LowerTriLvlSchedTP2SolverFunctor(const RowMapType &row_map_, const EntriesType &entries_, const ValuesType &values_, LHSType &lhs_, const RHSType &rhs_, const NGBLType &nodes_grouped_by_level_, long node_count_, long node_groups_ = 0) :
row_map(row_map_), entries(entries_), values(values_), lhs(lhs_), rhs(rhs_), nodes_grouped_by_level(nodes_grouped_by_level_), node_count(node_count_), node_groups(node_groups_) {}
KOKKOS_INLINE_FUNCTION
void operator()(const member_type & team) const {
auto my_league = team.league_rank(); // map to rowid
size_t nrows = row_map.extent(0) - 1;
Kokkos::parallel_for( Kokkos::TeamThreadRange( team, 0, node_groups ), [&] ( const long ng ) {
auto rowid = nodes_grouped_by_level(node_count + my_league*node_groups + ng);
if ( size_t(rowid) < nrows ) {
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
Kokkos::parallel_reduce( Kokkos::ThreadVectorRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
}, diff );
// ASSUMPTION: sorted diagonal value located at eoffset - 1
lhs(rowid) = (rhs_rowid+diff)/values(eoffset-1);
} // end if
}); // end TeamThreadRange
team.team_barrier();
}
KOKKOS_INLINE_FUNCTION
void operator()(const UnsortedTag&, const member_type & team) const {
auto my_league = team.league_rank(); // map to rowid
size_t nrows = row_map.extent(0) - 1;
Kokkos::parallel_for( Kokkos::TeamThreadRange( team, 0, node_groups ), [&] ( const long ng ) {
auto rowid = nodes_grouped_by_level(node_count + my_league*node_groups + ng);
if ( size_t(rowid) < nrows ) {
auto soffset = row_map(rowid);
auto eoffset = row_map(rowid+1);
auto rhs_rowid = rhs(rowid);
scalar_t diff = scalar_t(0.0);
auto diag = -1;
Kokkos::parallel_reduce( Kokkos::ThreadVectorRange( team, soffset, eoffset ), [&] ( const long ptr, scalar_t &tdiff ) {
auto colid = entries(ptr);
auto val = values(ptr);
if ( colid != rowid ) {
tdiff = tdiff - val*lhs(colid);
}
else {
diag = ptr;
}
}, diff );
// ASSUMPTION: sorted diagonal value located at eoffset - 1
lhs(rowid) = (rhs_rowid+diff)/values(diag);
} // end if
}); // end TeamThreadRange
team.team_barrier();
}
};
#if defined(KOKKOSKERNELS_ENABLE_SUPERNODAL_SPTRSV)
// -----------------------------------------------------------
// Helper functors for Lower-triangular solve with SpMV
template <class TriSolveHandle, class LHSType, class NGBLType>
struct SparseTriSupernodalSpMVFunctor
{
//using execution_space = typename LHSType::execution_space;
//using memory_space = typename execution_space::memory_space;
using execution_space = typename TriSolveHandle::HandleExecSpace;
using memory_space = typename TriSolveHandle::HandleTempMemorySpace;
using policy_type = Kokkos::TeamPolicy<execution_space>;
using member_type = typename policy_type::member_type;
using scalar_t = typename LHSType::non_const_value_type;
using work_view_t = typename Kokkos::View<scalar_t*, Kokkos::Device<execution_space, memory_space>>;
int flag;
long node_count;
NGBLType nodes_grouped_by_level;
const int *supercols;
const int *workoffset;
LHSType X;
work_view_t work;
// constructor
SparseTriSupernodalSpMVFunctor (int flag_,
long node_count_,
const NGBLType &nodes_grouped_by_level_,
const int *supercols_,
const int *workoffset_,
LHSType &X_,
work_view_t work_) :
flag(flag_), node_count(node_count_), nodes_grouped_by_level(nodes_grouped_by_level_), supercols(supercols_),
workoffset(workoffset_), X(X_), work(work_) {
}
// operator
KOKKOS_INLINE_FUNCTION
void operator()(const member_type & team) const {
const int league_rank = team.league_rank(); // batch id
const int team_size = team.team_size ();
const int team_rank = team.team_rank ();
const scalar_t zero (0.0);
auto s = nodes_grouped_by_level (node_count + league_rank);
// copy vector elements for the diagonal to input vector (work)
// and zero out the corresponding elements in output (X)
int w1 = workoffset[s];
int j1 = supercols[s];
// number of columns in the s-th supernode column
int nscol = supercols[s+1] - j1 ;
if (flag == -2) {
// copy X to work
for (int j = team_rank; j < nscol; j += team_size) {
work (w1 + j) = X (j1 + j);
}
} else if (flag == -1) {
// copy work to X
for (int j = team_rank; j < nscol; j += team_size) {
X (j1 + j) = work (w1 + j);
}
} else if (flag == 1) {
for (int j = team_rank; j < nscol; j += team_size) {
work (w1 + j) = X (j1 + j);
X (j1 + j) = zero;
}
} else {
// reinitialize work to zero
for (int j = team_rank; j < nscol; j += team_size) {
work (w1 + j) = zero;
}
}
team.team_barrier ();
}
};
// -----------------------------------------------------------
// Functor for Lower-triangular solve
template <class TriSolveHandle, class ColptrView, class RowindType, class ValuesType, class LHSType, class NGBLType>
struct LowerTriSupernodalFunctor
{
using execution_space = typename TriSolveHandle::HandleExecSpace;
using memory_space = typename TriSolveHandle::HandleTempMemorySpace;
using policy_type = Kokkos::TeamPolicy<execution_space>;
using member_type = typename policy_type::member_type;
using scalar_t = typename ValuesType::non_const_value_type;
using integer_view_t = Kokkos::View<int*, memory_space>;
using work_view_t = typename Kokkos::View<scalar_t*, Kokkos::Device<execution_space, memory_space>>;
using range_type = Kokkos::pair<int, int>;
const bool unit_diagonal;
const bool invert_diagonal;
const bool invert_offdiagonal;
const int *supercols;
ColptrView colptr;
RowindType rowind;
ValuesType values;
int level;
integer_view_t kernel_type;
integer_view_t diag_kernel_type;
LHSType X;
work_view_t work; // needed with gemv for update&scatter
integer_view_t work_offset;
NGBLType nodes_grouped_by_level;
long node_count;
// constructor
LowerTriSupernodalFunctor (// supernode info
const bool unit_diagonal_,
const bool invert_diagonal_,
const bool invert_offdiagonal_,
const int *supercols_,
// L in CSC
const ColptrView &colptr_,
const RowindType &rowind_,
const ValuesType &values_,
// options to pick kernel type
int level_,
integer_view_t &kernel_type_,
integer_view_t &diag_kernel_type_,
// right-hand-side (input), solution (output)
LHSType &X_,
// workspace
work_view_t work_,
integer_view_t &work_offset_,
//
const NGBLType &nodes_grouped_by_level_,
long node_count_) :
unit_diagonal(unit_diagonal_), invert_diagonal(invert_diagonal_), invert_offdiagonal(invert_offdiagonal_),
supercols(supercols_), colptr(colptr_), rowind(rowind_), values(values_),
level(level_), kernel_type(kernel_type_), diag_kernel_type(diag_kernel_type_),
X(X_), work(work_), work_offset(work_offset_),
nodes_grouped_by_level(nodes_grouped_by_level_), node_count(node_count_) {
}
// operator
KOKKOS_INLINE_FUNCTION
void operator()(const member_type & team) const {
/* ---------------------------------------------------------------------- */
/* get inputs */
/* ---------------------------------------------------------------------- */
const int league_rank = team.league_rank(); // batch id
const int team_size = team.team_size ();
const int team_rank = team.team_rank ();
const scalar_t zero (0.0);
const scalar_t one (1.0);
auto s = nodes_grouped_by_level (node_count + league_rank);
// supernodal column size
const int j1 = supercols[s];
const int j2 = supercols[s+1];
// > number of columns in the s-th supernode column
const int nscol = j2 - j1;
// "total" number of rows in all the supernodes (diagonal+off-diagonal)
const int i1 = colptr (j1);
const int nsrow = colptr (j1+1) - i1;
// create a view for the s-th supernocal column
scalar_t *dataL = const_cast<scalar_t*> (values.data ());
Kokkos::View<scalar_t**, Kokkos::LayoutLeft, memory_space, Kokkos::MemoryUnmanaged> viewL (&dataL[i1], nsrow, nscol);
// extract part of the solution, corresponding to the diagonal block
auto Xj = Kokkos::subview (X, range_type(j1, j2));
// workspace
const int workoffset = work_offset (s);
auto Z = Kokkos::subview (work, range_type(workoffset+nscol, workoffset+nsrow));
if (diag_kernel_type (level) != 3) { // not a device-level TRSM-solve
if (invert_offdiagonal) {
// combined TRSM solve with diagonal + GEMV update with off-diagonal
auto Y = Kokkos::subview (work, range_type(workoffset, workoffset+nsrow)); // needed for gemv instead of trmv/trsv
auto Ljj = Kokkos::subview (viewL, range_type (0, nsrow), Kokkos::ALL ());
KokkosBatched::TeamGemv<member_type,
KokkosBatched::Trans::NoTranspose,
KokkosBatched::Algo::Gemv::Unblocked>
::invoke(team, one, Ljj, Xj, zero, Y);
team.team_barrier ();
for (int ii = team_rank; ii < nscol; ii += team_size) {
Xj(ii) = Y(ii);
}
team.team_barrier ();
} else {
/* TRSM with diagonal block */
// extract diagonal and off-diagonal blocks of L
auto Ljj = Kokkos::subview (viewL, range_type (0, nscol), Kokkos::ALL ());
if (invert_diagonal) {
// workspace
auto Y = Kokkos::subview (work, range_type(workoffset, workoffset+nscol)); // needed for gemv instead of trmv/trsv
for (int ii = team_rank; ii < nscol; ii += team_size) {
Y(ii) = Xj(ii);
}
team.team_barrier ();
// calling team-level "Unblocked" gemv on small-size diagonal in KokkosBatched
KokkosBatched::TeamGemv<member_type,
KokkosBatched::Trans::NoTranspose,
KokkosBatched::Algo::Gemv::Unblocked>
::invoke(team, one, Ljj, Y, zero, Xj);
} else {
Kokkos::View<scalar_t**, Kokkos::LayoutLeft, memory_space, Kokkos::MemoryUnmanaged> Xjj (Xj.data (), nscol, 1);
if (unit_diagonal) {
KokkosBatched::TeamTrsm<member_type,
KokkosBatched::Side::Left,
KokkosBatched::Uplo::Lower,
KokkosBatched::Trans::NoTranspose,
KokkosBatched::Diag::Unit,
KokkosBatched::Algo::Trsm::Unblocked>
::invoke(team, one, Ljj, Xjj);
} else {
KokkosBatched::TeamTrsm<member_type,
KokkosBatched::Side::Left,
KokkosBatched::Uplo::Lower,
KokkosBatched::Trans::NoTranspose,
KokkosBatched::Diag::NonUnit,
KokkosBatched::Algo::Trsm::Unblocked>
::invoke(team, one, Ljj, Xjj);
}
}
team.team_barrier ();
/* GEMM to update with off diagonal blocks */
auto Lij = Kokkos::subview (viewL, range_type (nscol, nsrow), Kokkos::ALL ());
KokkosBatched::TeamGemv<member_type,
KokkosBatched::Trans::NoTranspose,
KokkosBatched::Algo::Gemv::Unblocked>
::invoke(team, one, Lij, Xj, zero, Z);
team.team_barrier();
}
}
/* scatter vectors back into X */
int i2 = i1 + nscol; // offset into rowind
int nsrow2 = nsrow - nscol; // "total" number of rows in all the off-diagonal supernodes
Kokkos::View<scalar_t*, memory_space, Kokkos::MemoryTraits<Kokkos::Unmanaged | Kokkos::Atomic> > Xatomic(X.data(), X.extent(0));
for (int ii = team_rank; ii < nsrow2; ii += team_size) {
int i = rowind (i2 + ii);
Xatomic (i) -= Z (ii);
}
team.team_barrier();
}
};
// -----------------------------------------------------------
// Functor for Upper-triangular solve in CSR
template <class TriSolveHandle, class ColptrType, class RowindType, class ValuesType, class LHSType, class NGBLType>
struct UpperTriSupernodalFunctor
{
using execution_space = typename TriSolveHandle::HandleExecSpace;
using memory_space = typename TriSolveHandle::HandleTempMemorySpace;
using policy_type = Kokkos::TeamPolicy<execution_space>;
using member_type = typename policy_type::member_type;
using scalar_t = typename ValuesType::non_const_value_type;
using integer_view_t = Kokkos::View<int*, memory_space>;
using work_view_t = typename Kokkos::View<scalar_t*, Kokkos::Device<execution_space, memory_space>>;
using SupernodeView = typename Kokkos::View<scalar_t**, Kokkos::LayoutLeft,
memory_space, Kokkos::MemoryUnmanaged>;
using range_type = Kokkos::pair<int, int>;
bool invert_diagonal;
const int *supercols;
ColptrType colptr;
RowindType rowind;
ValuesType values;
int level;
integer_view_t kernel_type;
integer_view_t diag_kernel_type;
LHSType X;
work_view_t work; // needed with gemv for update&scatter
integer_view_t work_offset;
NGBLType nodes_grouped_by_level;
long node_count;
// constructor
UpperTriSupernodalFunctor (// supernode info
bool invert_diagonal_,
const int *supercols_,
// U in CSR
const ColptrType &colptr_,
const RowindType &rowind_,
const ValuesType &values_,
// options to pick kernel type
int level_,
integer_view_t &kernel_type_,
integer_view_t &diag_kernel_type_,
// right-hand-side (input), solution (output)
LHSType &X_,
// workspace
work_view_t &work_,
integer_view_t &work_offset_,
//
const NGBLType &nodes_grouped_by_level_,
long node_count_) :
invert_diagonal(invert_diagonal_), supercols(supercols_),
colptr(colptr_), rowind(rowind_), values(values_),
level(level_), kernel_type(kernel_type_), diag_kernel_type(diag_kernel_type_),
X(X_), work(work_), work_offset(work_offset_),
nodes_grouped_by_level(nodes_grouped_by_level_), node_count(node_count_) {
}
// operator
KOKKOS_INLINE_FUNCTION
void operator()(const member_type & team) const {
/* ---------------------------------------------------------------------- */
/* get inputs */
/* ---------------------------------------------------------------------- */
const int league_rank = team.league_rank(); // batch id
const int team_size = team.team_size ();
const int team_rank = team.team_rank ();
const scalar_t zero (0.0);
const scalar_t one (1.0);
auto s = nodes_grouped_by_level (node_count + league_rank);
// number of columns in the s-th supernode column
int j1 = supercols[s];
int j2 = supercols[s+1];
int nscol = j2 - j1;
// "total" number of rows in all the supernodes (diagonal+off-diagonal)
int i1 = colptr (j1);
int nsrow = colptr (j1+1) - i1 ;
// create a view of the s-th supernocal row of U
scalar_t *dataU = const_cast<scalar_t*> (values.data ());
SupernodeView viewU (&dataU[i1], nsrow, nscol);
// extract part of solution, corresponding to the diagonal block U(s, s)
auto Xj = Kokkos::subview (X, range_type(j1, j2));
// workspaces
int workoffset = work_offset (s);
// "total" number of rows in all the off-diagonal supernodes
int nsrow2 = nsrow - nscol;
/* gather vector into Z */
int i2 = i1 + nscol; // offset into rowind
auto Z = Kokkos::subview(work, range_type(workoffset+nscol, workoffset+nsrow)); // needed with gemv for update&scatter
for (int ii = team_rank; ii < nsrow2 ; ii += team_size) {
int i = rowind (i2 + ii);
Z (ii) = X (i);
}
team.team_barrier();
/* GEMM to update with off diagonal blocks, Xj = -Uij^T * Z */
if (diag_kernel_type (level) != 3) {
// not device-level GEMV-udpate
auto Uij = Kokkos::subview (viewU, range_type (nscol, nsrow), Kokkos::ALL ());
KokkosBatched::TeamGemv<member_type,
KokkosBatched::Trans::Transpose,
KokkosBatched::Algo::Gemv::Unblocked>
::invoke(team, -one, Uij, Z, one, Xj);
team.team_barrier();
/* TRSM with diagonal block */
// extract diagonal and off-diagonal blocks of U
auto Ujj = Kokkos::subview (viewU, range_type (0, nscol), Kokkos::ALL ());
if (invert_diagonal) {
// workspace
auto Y = Kokkos::subview (work, range_type(workoffset, workoffset+nscol)); // needed for gemv instead of trmv/trsv
for (int ii = team_rank; ii < nscol; ii += team_size) {
Y (ii) = Xj (ii);
}
team.team_barrier();
// caling team-level kernel in KokkosBatched on a small-size diagonal
KokkosBatched::TeamGemv<member_type,
KokkosBatched::Trans::Transpose,
KokkosBatched::Algo::Gemv::Unblocked>
::invoke(team, one, Ujj, Y, zero, Xj);