-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpddistribute.c.all
2004 lines (1759 loc) · 62.5 KB
/
pddistribute.c.all
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
/*! \file
Copyright (c) 2003, The Regents of the University of California, through
Lawrence Berkeley National Laboratory (subject to receipt of any required
approvals from U.S. Dept. of Energy)
All rights reserved.
The source code is distributed under BSD license, see the file License.txt
at the top-level directory.
*/
/*! @file
* \brief Re-distribute A on the 2D process mesh.
* <pre>
* -- Distributed SuperLU routine (version 2.3) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley.
* October 15, 2008
* </pre>
*/
#include "superlu_ddefs.h"
#ifndef CACHELINE
#define CACHELINE 64 /* bytes, Xeon Phi KNL, Cori haswell, Edision */
#endif
/*! \brief
*
* <pre>
* Purpose
* =======
* Re-distribute A on the 2D process mesh.
*
* Arguments
* =========
*
* A (input) SuperMatrix*
* The distributed input matrix A of dimension (A->nrow, A->ncol).
* A may be overwritten by diag(R)*A*diag(C)*Pc^T.
* The type of A can be: Stype = SLU_NR_loc; Dtype = SLU_D; Mtype = SLU_GE.
*
* ScalePermstruct (input) ScalePermstruct_t*
* The data structure to store the scaling and permutation vectors
* describing the transformations performed to the original matrix A.
*
* Glu_freeable (input) *Glu_freeable_t
* The global structure describing the graph of L and U.
*
* grid (input) gridinfo_t*
* The 2D process mesh.
*
* colptr (output) int*
*
* rowind (output) int*
*
* a (output) double*
*
* Return value
* ============
* </pre>
*/
int_t
dReDistribute_A(SuperMatrix *A, ScalePermstruct_t *ScalePermstruct,
Glu_freeable_t *Glu_freeable, int_t *xsup, int_t *supno,
gridinfo_t *grid, int_t *colptr[], int_t *rowind[],
double *a[])
{
NRformat_loc *Astore;
int_t *perm_r; /* row permutation vector */
int_t *perm_c; /* column permutation vector */
int_t i, irow, fst_row, j, jcol, k, gbi, gbj, n, m_loc, jsize,nnz_tot;
int_t nnz_loc; /* number of local nonzeros */
int_t SendCnt; /* number of remote nonzeros to be sent */
int_t RecvCnt; /* number of remote nonzeros to be sent */
int_t *nnzToSend, *nnzToRecv, maxnnzToRecv;
int_t *ia, *ja, **ia_send, *index, *itemp;
int_t *ptr_to_send;
double *aij, **aij_send, *nzval, *dtemp;
double *nzval_a;
double asum,asum_tot;
int iam, it, p, procs, iam_g;
MPI_Request *send_req;
MPI_Status status;
/* ------------------------------------------------------------
INITIALIZATION.
------------------------------------------------------------*/
iam = grid->iam;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter dReDistribute_A()");
#endif
perm_r = ScalePermstruct->perm_r;
perm_c = ScalePermstruct->perm_c;
procs = grid->nprow * grid->npcol;
Astore = (NRformat_loc *) A->Store;
n = A->ncol;
m_loc = Astore->m_loc;
fst_row = Astore->fst_row;
nnzToRecv = intCalloc_dist(2*procs);
nnzToSend = nnzToRecv + procs;
/* ------------------------------------------------------------
COUNT THE NUMBER OF NONZEROS TO BE SENT TO EACH PROCESS,
THEN ALLOCATE SPACE.
THIS ACCOUNTS FOR THE FIRST PASS OF A.
------------------------------------------------------------*/
for (i = 0; i < m_loc; ++i) {
for (j = Astore->rowptr[i]; j < Astore->rowptr[i+1]; ++j) {
irow = perm_c[perm_r[i+fst_row]]; /* Row number in Pc*Pr*A */
jcol = Astore->colind[j];
gbi = BlockNum( irow );
gbj = BlockNum( jcol );
p = PNUM( PROW(gbi,grid), PCOL(gbj,grid), grid );
++nnzToSend[p];
}
}
/* All-to-all communication */
MPI_Alltoall( nnzToSend, 1, mpi_int_t, nnzToRecv, 1, mpi_int_t,
grid->comm);
maxnnzToRecv = 0;
nnz_loc = SendCnt = RecvCnt = 0;
for (p = 0; p < procs; ++p) {
if ( p != iam ) {
SendCnt += nnzToSend[p];
RecvCnt += nnzToRecv[p];
maxnnzToRecv = SUPERLU_MAX( nnzToRecv[p], maxnnzToRecv );
} else {
nnz_loc += nnzToRecv[p];
/*assert(nnzToSend[p] == nnzToRecv[p]);*/
}
}
k = nnz_loc + RecvCnt; /* Total nonzeros ended up in my process. */
/* Allocate space for storing the triplets after redistribution. */
if ( k ) { /* count can be zero. */
if ( !(ia = intMalloc_dist(2*k)) )
ABORT("Malloc fails for ia[].");
if ( !(aij = doubleMalloc_dist(k)) )
ABORT("Malloc fails for aij[].");
}
ja = ia + k;
/* Allocate temporary storage for sending/receiving the A triplets. */
if ( procs > 1 ) {
if ( !(send_req = (MPI_Request *)
SUPERLU_MALLOC(2*procs *sizeof(MPI_Request))) )
ABORT("Malloc fails for send_req[].");
if ( !(ia_send = (int_t **) SUPERLU_MALLOC(procs*sizeof(int_t*))) )
ABORT("Malloc fails for ia_send[].");
if ( !(aij_send = (double **)SUPERLU_MALLOC(procs*sizeof(double*))) )
ABORT("Malloc fails for aij_send[].");
if ( SendCnt ) { /* count can be zero */
if ( !(index = intMalloc_dist(2*SendCnt)) )
ABORT("Malloc fails for index[].");
if ( !(nzval = doubleMalloc_dist(SendCnt)) )
ABORT("Malloc fails for nzval[].");
}
if ( !(ptr_to_send = intCalloc_dist(procs)) )
ABORT("Malloc fails for ptr_to_send[].");
if ( maxnnzToRecv ) { /* count can be zero */
if ( !(itemp = intMalloc_dist(2*maxnnzToRecv)) )
ABORT("Malloc fails for itemp[].");
if ( !(dtemp = doubleMalloc_dist(maxnnzToRecv)) )
ABORT("Malloc fails for dtemp[].");
}
for (i = 0, j = 0, p = 0; p < procs; ++p) {
if ( p != iam ) {
ia_send[p] = &index[i];
i += 2 * nnzToSend[p]; /* ia/ja indices alternate */
aij_send[p] = &nzval[j];
j += nnzToSend[p];
}
}
} /* if procs > 1 */
if ( !(*colptr = intCalloc_dist(n+1)) )
ABORT("Malloc fails for *colptr[].");
/* ------------------------------------------------------------
LOAD THE ENTRIES OF A INTO THE (IA,JA,AIJ) STRUCTURES TO SEND.
THIS ACCOUNTS FOR THE SECOND PASS OF A.
------------------------------------------------------------*/
nnz_loc = 0; /* Reset the local nonzero count. */
nzval_a = Astore->nzval;
for (i = 0; i < m_loc; ++i) {
for (j = Astore->rowptr[i]; j < Astore->rowptr[i+1]; ++j) {
irow = perm_c[perm_r[i+fst_row]]; /* Row number in Pc*Pr*A */
jcol = Astore->colind[j];
gbi = BlockNum( irow );
gbj = BlockNum( jcol );
p = PNUM( PROW(gbi,grid), PCOL(gbj,grid), grid );
if ( p != iam ) { /* remote */
k = ptr_to_send[p];
ia_send[p][k] = irow;
ia_send[p][k + nnzToSend[p]] = jcol;
aij_send[p][k] = nzval_a[j];
++ptr_to_send[p];
} else { /* local */
ia[nnz_loc] = irow;
ja[nnz_loc] = jcol;
aij[nnz_loc] = nzval_a[j];
++nnz_loc;
++(*colptr)[jcol]; /* Count nonzeros in each column */
}
}
}
/* ------------------------------------------------------------
PERFORM REDISTRIBUTION. THIS INVOLVES ALL-TO-ALL COMMUNICATION.
NOTE: Can possibly use MPI_Alltoallv.
------------------------------------------------------------*/
for (p = 0; p < procs; ++p) {
if ( p != iam ) {
it = 2*nnzToSend[p];
MPI_Isend( ia_send[p], it, mpi_int_t,
p, iam, grid->comm, &send_req[p] );
it = nnzToSend[p];
MPI_Isend( aij_send[p], it, MPI_DOUBLE,
p, iam+procs, grid->comm, &send_req[procs+p] );
}
}
for (p = 0; p < procs; ++p) {
if ( p != iam ) {
it = 2*nnzToRecv[p];
MPI_Recv( itemp, it, mpi_int_t, p, p, grid->comm, &status );
it = nnzToRecv[p];
MPI_Recv( dtemp, it, MPI_DOUBLE, p, p+procs,
grid->comm, &status );
for (i = 0; i < nnzToRecv[p]; ++i) {
ia[nnz_loc] = itemp[i];
jcol = itemp[i + nnzToRecv[p]];
/*assert(jcol<n);*/
ja[nnz_loc] = jcol;
aij[nnz_loc] = dtemp[i];
++nnz_loc;
++(*colptr)[jcol]; /* Count nonzeros in each column */
}
}
}
for (p = 0; p < procs; ++p) {
if ( p != iam ) {
MPI_Wait( &send_req[p], &status);
MPI_Wait( &send_req[procs+p], &status);
}
}
/* ------------------------------------------------------------
DEALLOCATE TEMPORARY STORAGE
------------------------------------------------------------*/
SUPERLU_FREE(nnzToRecv);
if ( procs > 1 ) {
SUPERLU_FREE(send_req);
SUPERLU_FREE(ia_send);
SUPERLU_FREE(aij_send);
if ( SendCnt ) {
SUPERLU_FREE(index);
SUPERLU_FREE(nzval);
}
SUPERLU_FREE(ptr_to_send);
if ( maxnnzToRecv ) {
SUPERLU_FREE(itemp);
SUPERLU_FREE(dtemp);
}
}
/* ------------------------------------------------------------
CONVERT THE TRIPLET FORMAT INTO THE CCS FORMAT.
------------------------------------------------------------*/
if ( nnz_loc ) { /* nnz_loc can be zero */
if ( !(*rowind = intMalloc_dist(nnz_loc)) )
ABORT("Malloc fails for *rowind[].");
if ( !(*a = doubleMalloc_dist(nnz_loc)) )
ABORT("Malloc fails for *a[].");
}
/* Initialize the array of column pointers */
k = 0;
jsize = (*colptr)[0];
(*colptr)[0] = 0;
for (j = 1; j < n; ++j) {
k += jsize;
jsize = (*colptr)[j];
(*colptr)[j] = k;
}
/* Copy the triplets into the column oriented storage */
for (i = 0; i < nnz_loc; ++i) {
j = ja[i];
k = (*colptr)[j];
(*rowind)[k] = ia[i];
(*a)[k] = aij[i];
++(*colptr)[j];
}
/* Reset the column pointers to the beginning of each column */
for (j = n; j > 0; --j) (*colptr)[j] = (*colptr)[j-1];
(*colptr)[0] = 0;
if ( nnz_loc ) {
SUPERLU_FREE(ia);
SUPERLU_FREE(aij);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Exit dReDistribute_A()");
#endif
return 0;
} /* dReDistribute_A */
#ifdef oneside
int* BufSize;
int* BufSize_rd;
#endif
float
pddistribute(fact_t fact, int_t n, SuperMatrix *A,
ScalePermstruct_t *ScalePermstruct,
Glu_freeable_t *Glu_freeable, LUstruct_t *LUstruct,
gridinfo_t *grid)
/*
* -- Distributed SuperLU routine (version 2.0) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley.
* March 15, 2003
*
*
* Purpose
* =======
* Distribute the matrix onto the 2D process mesh.
*
* Arguments
* =========
*
* fact (input) fact_t
* Specifies whether or not the L and U structures will be re-used.
* = SamePattern_SameRowPerm: L and U structures are input, and
* unchanged on exit.
* = DOFACT or SamePattern: L and U structures are computed and output.
*
* n (input) int
* Dimension of the matrix.
*
* A (input) SuperMatrix*
* The distributed input matrix A of dimension (A->nrow, A->ncol).
* A may be overwritten by diag(R)*A*diag(C)*Pc^T. The type of A can be:
* Stype = SLU_NR_loc; Dtype = SLU_D; Mtype = SLU_GE.
*
* ScalePermstruct (input) ScalePermstruct_t*
* The data structure to store the scaling and permutation vectors
* describing the transformations performed to the original matrix A.
*
* Glu_freeable (input) *Glu_freeable_t
* The global structure describing the graph of L and U.
*
* LUstruct (input) LUstruct_t*
* Data structures for L and U factors.
*
* grid (input) gridinfo_t*
* The 2D process mesh.
*
* Return value
* ============
* > 0, working storage required (in bytes).
*
*/
{
Glu_persist_t *Glu_persist = LUstruct->Glu_persist;
LocalLU_t *Llu = LUstruct->Llu;
int_t bnnz, fsupc, fsupc1, i, ii, irow, istart, j, ib, jb, jj, k, k1,
len, len1, nsupc;
int_t lib; /* local block row number */
int_t nlb; /* local block rows*/
int_t ljb; /* local block column number */
int_t nrbl; /* number of L blocks in current block column */
int_t nrbu; /* number of U blocks in current block column */
int_t gb; /* global block number; 0 < gb <= nsuper */
int_t lb; /* local block number; 0 < lb <= ceil(NSUPERS/Pr) */
int_t ub,gik,iklrow,fnz;
int iam, jbrow, kcol, krow, mycol, myrow, pc, pr;
int_t mybufmax[NBUFFERS];
NRformat_loc *Astore;
double *a;
int_t *asub, *xa;
int_t *xa_begin, *xa_end;
int_t *xsup = Glu_persist->xsup; /* supernode and column mapping */
int_t *supno = Glu_persist->supno;
int_t *lsub, *xlsub, *usub, *usub1, *xusub;
int_t nsupers;
int_t next_lind; /* next available position in index[*] */
int_t next_lval; /* next available position in nzval[*] */
int_t *index; /* indices consist of headers and row subscripts */
int_t *index_srt; /* indices consist of headers and row subscripts */
int *index1; /* temporary pointer to array of int */
double *lusup, *lusup_srt, *uval; /* nonzero values in L and U */
double **Lnzval_bc_ptr; /* size ceil(NSUPERS/Pc) */
int_t **Lrowind_bc_ptr; /* size ceil(NSUPERS/Pc) */
int_t **Lindval_loc_bc_ptr; /* size ceil(NSUPERS/Pc) */
int_t *Unnz; /* size ceil(NSUPERS/Pc) */
double **Unzval_br_ptr; /* size ceil(NSUPERS/Pr) */
int_t **Ufstnz_br_ptr; /* size ceil(NSUPERS/Pr) */
BcTree *LBtree_ptr; /* size ceil(NSUPERS/Pc) */
RdTree *LRtree_ptr; /* size ceil(NSUPERS/Pr) */
BcTree *UBtree_ptr; /* size ceil(NSUPERS/Pc) */
RdTree *URtree_ptr; /* size ceil(NSUPERS/Pr) */
int msgsize;
int_t *Urbs,*Urbs1; /* Number of row blocks in each block column of U. */
Ucb_indptr_t **Ucb_indptr;/* Vertical linked list pointing to Uindex[] */
int_t **Ucb_valptr; /* Vertical linked list pointing to Unzval[] */
/*-- Counts to be used in factorization. --*/
int *ToRecv, *ToSendD, **ToSendR;
/*-- Counts to be used in lower triangular solve. --*/
int_t *fmod; /* Modification count for L-solve. */
int_t **fsendx_plist; /* Column process list to send down Xk. */
int_t nfrecvx = 0; /* Number of Xk I will receive. */
int_t nfsendx = 0; /* Number of Xk I will send */
int_t kseen;
/*-- Counts to be used in upper triangular solve. --*/
int_t *bmod; /* Modification count for U-solve. */
int_t **bsendx_plist; /* Column process list to send down Xk. */
int_t nbrecvx = 0; /* Number of Xk I will receive. */
int_t nbsendx = 0; /* Number of Xk I will send */
int_t *ilsum; /* starting position of each supernode in
the full array (local) */
/*-- Auxiliary arrays; freed on return --*/
int_t *rb_marker; /* block hit marker; size ceil(NSUPERS/Pr) */
int_t *Urb_length; /* U block length; size ceil(NSUPERS/Pr) */
int_t *Urb_indptr; /* pointers to U index[]; size ceil(NSUPERS/Pr) */
int_t *Urb_fstnz; /* # of fstnz in a block row; size ceil(NSUPERS/Pr) */
int_t *Ucbs; /* number of column blocks in a block row */
int_t *Lrb_length; /* L block length; size ceil(NSUPERS/Pr) */
int_t *Lrb_number; /* global block number; size ceil(NSUPERS/Pr) */
int_t *Lrb_indptr; /* pointers to L index[]; size ceil(NSUPERS/Pr) */
int_t *Lrb_valptr; /* pointers to L nzval[]; size ceil(NSUPERS/Pr) */
int_t *ActiveFlag;
int_t *ActiveFlagAll;
int_t Iactive;
int *ranks;
int_t *idxs;
int_t **nzrows;
double rseed;
int rank_cnt,rank_cnt_ref,Root;
double *dense, *dense_col; /* SPA */
double zero = 0.0;
int_t ldaspa; /* LDA of SPA */
int_t iword, dword;
float mem_use = 0.0;
int_t *mod_bit;
int_t *frecv, *brecv, *lloc;
double **Linv_bc_ptr; /* size ceil(NSUPERS/Pc) */
double **Uinv_bc_ptr; /* size ceil(NSUPERS/Pc) */
double *SeedSTD_BC,*SeedSTD_RD;
int_t idx_indx,idx_lusup;
int_t nbrow;
int_t ik, il, lk, rel, knsupc, idx_r;
int_t lptr1_tmp, idx_i, idx_v,m, uu, aln_i;
int_t nub;
int tag;
#if ( PRNTlevel>=1 )
int_t nLblocks = 0, nUblocks = 0;
#endif
#if ( PROFlevel>=1 )
double t, t_u, t_l;
int_t u_blks;
#endif
/* Initialization. */
iam = grid->iam;
myrow = MYROW( iam, grid );
mycol = MYCOL( iam, grid );
for (i = 0; i < NBUFFERS; ++i) mybufmax[i] = 0;
nsupers = supno[n-1] + 1;
Astore = (NRformat_loc *) A->Store;
//#if ( PRNTlevel>=1 )
iword = sizeof(int_t);
dword = sizeof(double);
aln_i = ceil(CACHELINE/(double)iword);
//#endif
#ifdef oneside
int Pr, Pc;
Pc = grid->npcol;
Pr = grid->nprow;
if ( !(BufSize = (int*)SUPERLU_MALLOC( Pr * Pc * sizeof(int))) )
ABORT("Malloc fails for BufSize[]");
memset(BufSize, 0, Pr * Pc * sizeof(int));
if ( !(BufSize_rd = (int*)SUPERLU_MALLOC( Pc * Pr*sizeof(int))) )
ABORT("Malloc fails for BufSiz_rd[]");
memset(BufSize_rd, 0, Pc * Pr * sizeof(int));
#endif
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter pddistribute()");
#endif
#if ( PROFlevel>=1 )
t = SuperLU_timer_();
#endif
dReDistribute_A(A, ScalePermstruct, Glu_freeable, xsup, supno,
grid, &xa, &asub, &a);
#if ( PROFlevel>=1 )
t = SuperLU_timer_() - t;
if ( !iam ) printf("--------\n"
".. Phase 1 - ReDistribute_A time: %.2f\t\n", t);
#endif
if ( fact == SamePattern_SameRowPerm ) {
#if ( PROFlevel>=1 )
t_l = t_u = 0; u_blks = 0;
#endif
/* We can propagate the new values of A into the existing
L and U data structures. */
ilsum = Llu->ilsum;
ldaspa = Llu->ldalsum;
if ( !(dense = doubleCalloc_dist(ldaspa * sp_ienv_dist(3))) )
ABORT("Calloc fails for SPA dense[].");
nrbu = CEILING( nsupers, grid->nprow ); /* No. of local block rows */
if ( !(Urb_length = intCalloc_dist(nrbu)) )
ABORT("Calloc fails for Urb_length[].");
if ( !(Urb_indptr = intMalloc_dist(nrbu)) )
ABORT("Malloc fails for Urb_indptr[].");
Lrowind_bc_ptr = Llu->Lrowind_bc_ptr;
Lindval_loc_bc_ptr = Llu->Lindval_loc_bc_ptr;
Lnzval_bc_ptr = Llu->Lnzval_bc_ptr;
Ufstnz_br_ptr = Llu->Ufstnz_br_ptr;
Unzval_br_ptr = Llu->Unzval_br_ptr;
Unnz = Llu->Unnz;
#if ( PRNTlevel>=1 )
mem_use += 2.0*nrbu*iword + ldaspa*sp_ienv_dist(3)*dword;
#endif
#if ( PROFlevel>=1 )
t = SuperLU_timer_();
#endif
/* Initialize Uval to zero. */
for (lb = 0; lb < nrbu; ++lb) {
Urb_indptr[lb] = BR_HEADER; /* Skip header in U index[]. */
index = Ufstnz_br_ptr[lb];
if ( index ) {
uval = Unzval_br_ptr[lb];
len = index[1];
for (i = 0; i < len; ++i) uval[i] = zero;
} /* if index != NULL */
} /* for lb ... */
for (jb = 0; jb < nsupers; ++jb) { /* Loop through each block column */
pc = PCOL( jb, grid );
if ( mycol == pc ) { /* Block column jb in my process column */
fsupc = FstBlockC( jb );
nsupc = SuperSize( jb );
/* Scatter A into SPA (for L), or into U directly. */
for (j = fsupc, dense_col = dense; j < FstBlockC(jb+1); ++j) {
for (i = xa[j]; i < xa[j+1]; ++i) {
irow = asub[i];
gb = BlockNum( irow );
if ( myrow == PROW( gb, grid ) ) {
lb = LBi( gb, grid );
if ( gb < jb ) { /* in U */
index = Ufstnz_br_ptr[lb];
uval = Unzval_br_ptr[lb];
while ( (k = index[Urb_indptr[lb]]) < jb ) {
/* Skip nonzero values in this block */
Urb_length[lb] += index[Urb_indptr[lb]+1];
/* Move pointer to the next block */
Urb_indptr[lb] += UB_DESCRIPTOR
+ SuperSize( k );
}
/*assert(k == jb);*/
/* start fstnz */
istart = Urb_indptr[lb] + UB_DESCRIPTOR;
len = Urb_length[lb];
fsupc1 = FstBlockC( gb+1 );
k = j - fsupc;
/* Sum the lengths of the leading columns */
for (jj = 0; jj < k; ++jj)
len += fsupc1 - index[istart++];
/*assert(irow>=index[istart]);*/
uval[len + irow - index[istart]] = a[i];
} else { /* in L; put in SPA first */
irow = ilsum[lb] + irow - FstBlockC( gb );
dense_col[irow] = a[i];
}
}
} /* for i ... */
dense_col += ldaspa;
} /* for j ... */
#if ( PROFlevel>=1 )
t_u += SuperLU_timer_() - t;
t = SuperLU_timer_();
#endif
/* Gather the values of A from SPA into Lnzval[]. */
ljb = LBj( jb, grid ); /* Local block number */
index = Lrowind_bc_ptr[ljb];
if ( index ) {
nrbl = index[0]; /* Number of row blocks. */
len = index[1]; /* LDA of lusup[]. */
lusup = Lnzval_bc_ptr[ljb];
next_lind = BC_HEADER;
next_lval = 0;
for (jj = 0; jj < nrbl; ++jj) {
gb = index[next_lind++];
len1 = index[next_lind++]; /* Rows in the block. */
lb = LBi( gb, grid );
for (bnnz = 0; bnnz < len1; ++bnnz) {
irow = index[next_lind++]; /* Global index. */
irow = ilsum[lb] + irow - FstBlockC( gb );
k = next_lval++;
for (j = 0, dense_col = dense; j < nsupc; ++j) {
lusup[k] = dense_col[irow];
dense_col[irow] = zero;
k += len;
dense_col += ldaspa;
}
} /* for bnnz ... */
} /* for jj ... */
} /* if index ... */
#if ( PROFlevel>=1 )
t_l += SuperLU_timer_() - t;
#endif
} /* if mycol == pc */
} /* for jb ... */
SUPERLU_FREE(dense);
SUPERLU_FREE(Urb_length);
SUPERLU_FREE(Urb_indptr);
#if ( PROFlevel>=1 )
if ( !iam ) printf(".. 2nd distribute time: L %.2f\tU %.2f\tu_blks %d\tnrbu %d\n",
t_l, t_u, u_blks, nrbu);
#endif
} else {
/* ------------------------------------------------------------
FIRST TIME CREATING THE L AND U DATA STRUCTURES.
------------------------------------------------------------*/
#if ( PROFlevel>=1 )
t_l = t_u = 0; u_blks = 0;
#endif
/* We first need to set up the L and U data structures and then
* propagate the values of A into them.
*/
lsub = Glu_freeable->lsub; /* compressed L subscripts */
xlsub = Glu_freeable->xlsub;
usub = Glu_freeable->usub; /* compressed U subscripts */
xusub = Glu_freeable->xusub;
if ( !(ToRecv = (int *) SUPERLU_MALLOC(nsupers * sizeof(int))) )
ABORT("Malloc fails for ToRecv[].");
for (i = 0; i < nsupers; ++i) ToRecv[i] = 0;
k = CEILING( nsupers, grid->npcol );/* Number of local column blocks */
if ( !(ToSendR = (int **) SUPERLU_MALLOC(k*sizeof(int*))) )
ABORT("Malloc fails for ToSendR[].");
j = k * grid->npcol;
if ( !(index1 = SUPERLU_MALLOC(j * sizeof(int))) )
ABORT("Malloc fails for index[].");
#if ( PRNTlevel>=1 )
mem_use += (float) k*sizeof(int_t*) + (j + nsupers)*iword;
#endif
for (i = 0; i < j; ++i) index1[i] = EMPTY;
for (i = 0,j = 0; i < k; ++i, j += grid->npcol) ToSendR[i] = &index1[j];
k = CEILING( nsupers, grid->nprow ); /* Number of local block rows */
/* Pointers to the beginning of each block row of U. */
if ( !(Unzval_br_ptr =
(double**)SUPERLU_MALLOC(k * sizeof(double*))) )
ABORT("Malloc fails for Unzval_br_ptr[].");
if ( !(Ufstnz_br_ptr = (int_t**)SUPERLU_MALLOC(k * sizeof(int_t*))) )
ABORT("Malloc fails for Ufstnz_br_ptr[].");
if ( !(ToSendD = SUPERLU_MALLOC(k * sizeof(int))) )
ABORT("Malloc fails for ToSendD[].");
for (i = 0; i < k; ++i) ToSendD[i] = NO;
if ( !(ilsum = intMalloc_dist(k+1)) )
ABORT("Malloc fails for ilsum[].");
/* Auxiliary arrays used to set up U block data structures.
They are freed on return. */
if ( !(rb_marker = intCalloc_dist(k)) )
ABORT("Calloc fails for rb_marker[].");
if ( !(Urb_length = intCalloc_dist(k)) )
ABORT("Calloc fails for Urb_length[].");
if ( !(Urb_indptr = intMalloc_dist(k)) )
ABORT("Malloc fails for Urb_indptr[].");
if ( !(Urb_fstnz = intCalloc_dist(k)) )
ABORT("Calloc fails for Urb_fstnz[].");
if ( !(Ucbs = intCalloc_dist(k)) )
ABORT("Calloc fails for Ucbs[].");
#if ( PRNTlevel>=1 )
mem_use += 2.0*k*sizeof(int_t*) + (7*k+1)*iword;
#endif
/* Compute ldaspa and ilsum[]. */
ldaspa = 0;
ilsum[0] = 0;
for (gb = 0; gb < nsupers; ++gb) {
if ( myrow == PROW( gb, grid ) ) {
i = SuperSize( gb );
ldaspa += i;
lb = LBi( gb, grid );
ilsum[lb + 1] = ilsum[lb] + i;
}
}
#if ( PROFlevel>=1 )
t = SuperLU_timer_();
#endif
/* ------------------------------------------------------------
COUNT NUMBER OF ROW BLOCKS AND THE LENGTH OF EACH BLOCK IN U.
THIS ACCOUNTS FOR ONE-PASS PROCESSING OF G(U).
------------------------------------------------------------*/
/* Loop through each supernode column. */
for (jb = 0; jb < nsupers; ++jb) {
pc = PCOL( jb, grid );
fsupc = FstBlockC( jb );
nsupc = SuperSize( jb );
/* Loop through each column in the block. */
for (j = fsupc; j < fsupc + nsupc; ++j) {
/* usub[*] contains only "first nonzero" in each segment. */
for (i = xusub[j]; i < xusub[j+1]; ++i) {
irow = usub[i]; /* First nonzero of the segment. */
gb = BlockNum( irow );
kcol = PCOL( gb, grid );
ljb = LBj( gb, grid );
if ( mycol == kcol && mycol != pc ) ToSendR[ljb][pc] = YES;
pr = PROW( gb, grid );
lb = LBi( gb, grid );
if ( mycol == pc ) {
if ( myrow == pr ) {
ToSendD[lb] = YES;
/* Count nonzeros in entire block row. */
Urb_length[lb] += FstBlockC( gb+1 ) - irow;
if (rb_marker[lb] <= jb) {/* First see the block */
rb_marker[lb] = jb + 1;
Urb_fstnz[lb] += nsupc;
++Ucbs[lb]; /* Number of column blocks
in block row lb. */
#if ( PRNTlevel>=1 )
++nUblocks;
#endif
}
ToRecv[gb] = 1;
} else ToRecv[gb] = 2; /* Do I need 0, 1, 2 ? */
}
} /* for i ... */
} /* for j ... */
} /* for jb ... */
/* Set up the initial pointers for each block row in U. */
nrbu = CEILING( nsupers, grid->nprow );/* Number of local block rows */
for (lb = 0; lb < nrbu; ++lb) {
len = Urb_length[lb];
rb_marker[lb] = 0; /* Reset block marker. */
if ( len ) {
/* Add room for descriptors */
len1 = Urb_fstnz[lb] + BR_HEADER + Ucbs[lb] * UB_DESCRIPTOR;
if ( !(index = intMalloc_dist(len1+1)) )
ABORT("Malloc fails for Uindex[].");
Ufstnz_br_ptr[lb] = index;
if ( !(Unzval_br_ptr[lb] = doubleMalloc_dist(len)) )
ABORT("Malloc fails for Unzval_br_ptr[*][].");
mybufmax[2] = SUPERLU_MAX( mybufmax[2], len1 );
mybufmax[3] = SUPERLU_MAX( mybufmax[3], len );
index[0] = Ucbs[lb]; /* Number of column blocks */
index[1] = len; /* Total length of nzval[] */
index[2] = len1; /* Total length of index[] */
index[len1] = -1; /* End marker */
} else {
Ufstnz_br_ptr[lb] = NULL;
Unzval_br_ptr[lb] = NULL;
}
Urb_length[lb] = 0; /* Reset block length. */
Urb_indptr[lb] = BR_HEADER; /* Skip header in U index[]. */
Urb_fstnz[lb] = BR_HEADER;
} /* for lb ... */
SUPERLU_FREE(Ucbs);
#if ( PROFlevel>=1 )
t = SuperLU_timer_() - t;
if ( !iam) printf(".. Phase 2 - setup U strut time: %.2f\t\n", t);
#endif
#if ( PRNTlevel>=1 )
mem_use -= 2.0*k * iword;
#endif
/* Auxiliary arrays used to set up L block data structures.
They are freed on return.
k is the number of local row blocks. */
if ( !(Lrb_length = intCalloc_dist(k)) )
ABORT("Calloc fails for Lrb_length[].");
if ( !(Lrb_number = intMalloc_dist(k)) )
ABORT("Malloc fails for Lrb_number[].");
if ( !(Lrb_indptr = intMalloc_dist(k)) )
ABORT("Malloc fails for Lrb_indptr[].");
if ( !(Lrb_valptr = intMalloc_dist(k)) )
ABORT("Malloc fails for Lrb_valptr[].");
if ( !(dense = doubleCalloc_dist(ldaspa * sp_ienv_dist(3))) )
ABORT("Calloc fails for SPA dense[].");
/* These counts will be used for triangular solves. */
if ( !(fmod = intCalloc_dist(k)) )
ABORT("Calloc fails for fmod[].");
if ( !(bmod = intCalloc_dist(k)) )
ABORT("Calloc fails for bmod[].");
/* ------------------------------------------------ */
#if ( PRNTlevel>=1 )
mem_use += 6.0*k*iword + ldaspa*sp_ienv_dist(3)*dword;
#endif
k = CEILING( nsupers, grid->npcol );/* Number of local block columns */
/* Pointers to the beginning of each block column of L. */
if ( !(Lnzval_bc_ptr =
(double**)SUPERLU_MALLOC(k * sizeof(double*))) )
ABORT("Malloc fails for Lnzval_bc_ptr[].");
if ( !(Lrowind_bc_ptr = (int_t**)SUPERLU_MALLOC(k * sizeof(int_t*))) )
ABORT("Malloc fails for Lrowind_bc_ptr[].");
Lrowind_bc_ptr[k-1] = NULL;
if ( !(Lindval_loc_bc_ptr =
(int_t**)SUPERLU_MALLOC(k * sizeof(int_t*))) )
ABORT("Malloc fails for Lindval_loc_bc_ptr[].");
Lindval_loc_bc_ptr[k-1] = NULL;
if ( !(Linv_bc_ptr =
(double**)SUPERLU_MALLOC(k * sizeof(double*))) ) {
fprintf(stderr, "Malloc fails for Linv_bc_ptr[].");
}
if ( !(Uinv_bc_ptr =
(double**)SUPERLU_MALLOC(k * sizeof(double*))) ) {
fprintf(stderr, "Malloc fails for Uinv_bc_ptr[].");
}
Linv_bc_ptr[k-1] = NULL;
Uinv_bc_ptr[k-1] = NULL;
if ( !(Unnz =
(int_t*)SUPERLU_MALLOC(k * sizeof(int_t))) )
ABORT("Malloc fails for Unnz[].");
/* These lists of processes will be used for triangular solves. */
if ( !(fsendx_plist = (int_t **) SUPERLU_MALLOC(k*sizeof(int_t*))) )
ABORT("Malloc fails for fsendx_plist[].");
len = k * grid->nprow;
if ( !(index = intMalloc_dist(len)) )
ABORT("Malloc fails for fsendx_plist[0]");
for (i = 0; i < len; ++i) index[i] = EMPTY;
for (i = 0, j = 0; i < k; ++i, j += grid->nprow)
fsendx_plist[i] = &index[j];
if ( !(bsendx_plist = (int_t **) SUPERLU_MALLOC(k*sizeof(int_t*))) )
ABORT("Malloc fails for bsendx_plist[].");
if ( !(index = intMalloc_dist(len)) )
ABORT("Malloc fails for bsendx_plist[0]");
for (i = 0; i < len; ++i) index[i] = EMPTY;
for (i = 0, j = 0; i < k; ++i, j += grid->nprow)
bsendx_plist[i] = &index[j];
/* -------------------------------------------------------------- */
#if ( PRNTlevel>=1 )
mem_use += 4.0*k*sizeof(int_t*) + 2.0*len*iword;
#endif
/*------------------------------------------------------------
PROPAGATE ROW SUBSCRIPTS AND VALUES OF A INTO L AND U BLOCKS.
THIS ACCOUNTS FOR ONE-PASS PROCESSING OF A, L AND U.
------------------------------------------------------------*/
for (jb = 0; jb < nsupers; ++jb) { /* for each block column ... */
pc = PCOL( jb, grid );
if ( mycol == pc ) { /* Block column jb in my process column */
fsupc = FstBlockC( jb );
nsupc = SuperSize( jb );
ljb = LBj( jb, grid ); /* Local block number */
/* Scatter A into SPA. */
for (j = fsupc, dense_col = dense; j < FstBlockC(jb+1); ++j) {
for (i = xa[j]; i < xa[j+1]; ++i) {
irow = asub[i];
gb = BlockNum( irow );
if ( myrow == PROW( gb, grid ) ) {
lb = LBi( gb, grid );
irow = ilsum[lb] + irow - FstBlockC( gb );
dense_col[irow] = a[i];
}
}
dense_col += ldaspa;
} /* for j ... */
jbrow = PROW( jb, grid );
/*------------------------------------------------
* SET UP U BLOCKS.
*------------------------------------------------*/
#if ( PROFlevel>=1 )
t = SuperLU_timer_();
#endif
kseen = 0;
dense_col = dense;
/* Loop through each column in the block column. */
for (j = fsupc; j < FstBlockC( jb+1 ); ++j) {
istart = xusub[j];
/* NOTE: Only the first nonzero index of the segment
is stored in usub[]. */
for (i = istart; i < xusub[j+1]; ++i) {
irow = usub[i]; /* First nonzero in the segment. */
gb = BlockNum( irow );
pr = PROW( gb, grid );
if ( pr != jbrow &&
myrow == jbrow && /* diag. proc. owning jb */
bsendx_plist[ljb][pr] == EMPTY ) {
bsendx_plist[ljb][pr] = YES;
++nbsendx;
}
if ( myrow == pr ) {
lb = LBi( gb, grid ); /* Local block number */
index = Ufstnz_br_ptr[lb];
uval = Unzval_br_ptr[lb];
fsupc1 = FstBlockC( gb+1 );
if (rb_marker[lb] <= jb) { /* First time see
the block */
rb_marker[lb] = jb + 1;
Urb_indptr[lb] = Urb_fstnz[lb];;
index[Urb_indptr[lb]] = jb; /* Descriptor */
Urb_indptr[lb] += UB_DESCRIPTOR;
/* Record the first location in index[] of the
next block */
Urb_fstnz[lb] = Urb_indptr[lb] + nsupc;
len = Urb_indptr[lb];/* Start fstnz in index */
index[len-1] = 0;
for (k = 0; k < nsupc; ++k)
index[len+k] = fsupc1;
if ( gb != jb )/* Exclude diagonal block. */
++bmod[lb];/* Mod. count for back solve */
if ( kseen == 0 && myrow != jbrow ) {
++nbrecvx;
kseen = 1;
}
} else { /* Already saw the block */
len = Urb_indptr[lb];/* Start fstnz in index */
}
jj = j - fsupc;
index[len+jj] = irow;
/* Load the numerical values */
k = fsupc1 - irow; /* No. of nonzeros in segment */
index[len-1] += k; /* Increment block length in
Descriptor */
irow = ilsum[lb] + irow - FstBlockC( gb );
for (ii = 0; ii < k; ++ii) {
uval[Urb_length[lb]++] = dense_col[irow + ii];
dense_col[irow + ii] = zero;
}
} /* if myrow == pr ... */
} /* for i ... */
dense_col += ldaspa;
} /* for j ... */
#if ( PROFlevel>=1 )
t_u += SuperLU_timer_() - t;
t = SuperLU_timer_();
#endif
/*------------------------------------------------
* SET UP L BLOCKS.
*------------------------------------------------*/
/* Count number of blocks and length of each block. */
nrbl = 0;
len = 0; /* Number of row subscripts I own. */
kseen = 0;
istart = xlsub[fsupc];
for (i = istart; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
gb = BlockNum( irow ); /* Global block number */
pr = PROW( gb, grid ); /* Process row owning this block */
if ( pr != jbrow &&
myrow == jbrow && /* diag. proc. owning jb */
fsendx_plist[ljb][pr] == EMPTY /* first time */ ) {
fsendx_plist[ljb][pr] = YES;
++nfsendx;