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
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@author Hartwig Anzt
@generated from sparse/src/zparilut.cpp, normal z -> s, Wed Jan 22 14:42:45 2025
*/
#include "magmasparse_internal.h"
#ifdef _OPENMP
#include <omp.h>
#endif
#include "../blas/magma_trisolve.h"
#define PRECISION_s
/* For hipSPARSE, they use a separate real type than for hipBLAS */
#ifdef MAGMA_HAVE_HIP
#define float float
#endif
/***************************************************************************//**
Purpose
-------
Prepares the iterative threshold Incomplete LU preconditioner. The strategy
is interleaving a parallel fixed-point iteration that approximates an
incomplete factorization for a given nonzero pattern with a procedure that
adaptively changes the pattern. Much of this new algorithm has fine-grained
parallelism, and we show that it can efficiently exploit the compute power
of shared memory architectures.
This is the routine used in the publication by Anzt, Chow, Dongarra:
''ParILUT - A new parallel threshold ILU factorization''
submitted to SIAM SISC in 2017.
This function requires OpenMP, and is only available if OpenMP is activated.
The parameter list is:
precond.sweeps : number of ParILUT steps
precond.atol : absolute fill ratio (1.0 keeps nnz constant)
precond.rtol : how many candidates are added to the sparsity pattern
* 1.0 one per row
* < 1.0 a fraction of those
* > 1.0 all candidates
Arguments
---------
@param[in]
A magma_s_matrix
input matrix A
@param[in]
b magma_s_matrix
input RHS b
@param[in,out]
precond magma_s_preconditioner*
preconditioner parameters
@param[in]
queue magma_queue_t
Queue to execute in.
@ingroup magmasparse_sgepr
*******************************************************************************/
extern "C"
magma_int_t
magma_sparilut(
magma_s_matrix A,
magma_s_matrix b,
magma_s_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
#ifdef _OPENMP
real_Double_t start, end;
real_Double_t t_rm=0.0, t_add=0.0, t_res=0.0, t_sweep1=0.0, t_sweep2=0.0, t_cand=0.0,
t_transpose1=0.0, t_transpose2=0.0, t_selectrm=0.0,
t_selectadd=0.0, t_nrm=0.0, t_total = 0.0, accum=0.0;
float sum, sumL, sumU;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_s_matrix hA={Magma_CSR}, A0={Magma_CSR}, hAT={Magma_CSR}, hL={Magma_CSR}, hU={Magma_CSR},
oneL={Magma_CSR}, oneU={Magma_CSR},
L={Magma_CSR}, U={Magma_CSR}, L_new={Magma_CSR}, U_new={Magma_CSR}, UT={Magma_CSR};
magma_s_matrix L0={Magma_CSR}, U0={Magma_CSR};
magma_int_t num_rmL, num_rmU;
float thrsL = 0.0;
float thrsU = 0.0;
magma_int_t num_threads, timing = 1; // print timing
magma_int_t L0nnz, U0nnz;
#pragma omp parallel
{
num_threads = omp_get_max_threads();
}
CHECK( magma_smtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smtransfer( A, &A0, A.memory_location, Magma_CPU, queue ));
// in case using fill-in
if( precond->levels > 0 ){
CHECK( magma_ssymbilu( &hA, precond->levels, &hL, &hU , queue ));
}
magma_smfree(&hU, queue );
magma_smfree(&hL, queue );
L.diagorder_type = Magma_VALUE;
magma_smatrix_tril( hA, &L, queue );
magma_smtranspose(hA, &hAT, queue );
U.diagorder_type = Magma_UNITY;
magma_smatrix_tril( hAT, &U, queue );
for ( magma_int_t z=0; z<U.num_rows; z++ ){
U.val[U.row[z+1]-1] = MAGMA_S_ONE;
}
CHECK( magma_smtranspose( U, &UT, queue) );
L.rowidx = NULL;
UT.rowidx = NULL;
magma_smatrix_addrowindex( &L, queue );
magma_smatrix_addrowindex( &U, queue );
//CHECK( magma_sparilut_sweep( &A0, &L, &UT, queue ) );
//CHECK( magma_sparilut_sweep( &A0, &L, &UT, queue ) );
//CHECK( magma_sparilut_sweep( &A0, &L, &UT, queue ) );
//CHECK( magma_sparilut_sweep( &A0, &L, &UT, queue ) );
//CHECK( magma_sparilut_sweep( &A0, &L, &UT, queue ) );
L0nnz=L.nnz;
U0nnz=U.nnz;
// need only lower triangular
magma_smfree(&U, queue );
CHECK( magma_smtranspose( UT, &U, queue) );
CHECK( magma_smtransfer( L, &L0, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smtransfer( L, &oneL, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smtransfer( UT, &U0, A.memory_location, Magma_CPU, queue ));
magma_smatrix_addrowindex( &U, queue );
magma_smfree(&UT, queue );
//magma_free_cpu( UT.row ); UT.row = NULL;
//magma_free_cpu( UT.list ); UT.list = NULL;
//CHECK( magma_sparilut_create_collinkedlist( U, &UT, queue) );
if (timing == 1) {
printf("ilut_fill_ratio = %.6f;\n\n", precond->atol );
printf("performance_%d = [\n%%iter L.nnz U.nnz ILU-Norm candidat resid ILU-norm selectad add transp1 sweep1 selectrm remove sweep2 transp2 total accum\n", (int) num_threads);
}
//##########################################################################
for( magma_int_t iters =0; iters<precond->sweeps; iters++ ) {
t_rm=0.0; t_add=0.0; t_res=0.0; t_sweep1=0.0; t_sweep2=0.0; t_cand=0.0;
t_transpose1=0.0; t_transpose2=0.0; t_selectrm=0.0;
t_selectadd=0.0; t_nrm=0.0; t_total = 0.0;
num_rmL = max( (L_new.nnz-L0nnz*(1+precond->atol*(iters+1)/precond->sweeps)), 0 );
num_rmU = max( (U_new.nnz-U0nnz*(1+precond->atol*(iters+1)/precond->sweeps)), 0 );
// magma_free_cpu( UT.row ); UT.row = NULL;
// magma_free_cpu( UT.list ); UT.list = NULL;
// CHECK( magma_sparilut_create_collinkedlist( U, &UT, queue) );
start = magma_sync_wtime( queue );
magma_smfree(&UT, queue );
//magma_smtransposestruct_cpu( U, &UT, queue );
magma_scsrcoo_transpose( U, &UT, queue );
end = magma_sync_wtime( queue ); t_transpose1+=end-start;
start = magma_sync_wtime( queue );
magma_sparilut_candidates( L0, U0, L, UT, &hL, &hU, queue );
end = magma_sync_wtime( queue ); t_cand=+end-start;
if( precond->rtol == 1.0 ){
start = magma_sync_wtime( queue );
magma_sparilut_residuals( hA, L, U, &hL, queue );
magma_sparilut_residuals( hA, L, U, &hU, queue );
end = magma_sync_wtime( queue ); t_res=+end-start;
start = magma_sync_wtime( queue );
magma_smatrix_abssum( hL, &sumL, queue );
magma_smatrix_abssum( hU, &sumU, queue );
sum = sumL + sumU;
end = magma_sync_wtime( queue ); t_nrm+=end-start;
start = magma_sync_wtime( queue );
magma_sparilut_transpose_select_one( hU, &oneU, queue );
magma_sparilut_selectoneperrow( 1, &hL, &oneL, queue );
magma_smfree(&hL, queue );
magma_smfree(&hU, queue );
end = magma_sync_wtime( queue ); t_selectadd+=end-start;
} else if( precond->rtol > 1.0 ) {
start = magma_sync_wtime( queue );
magma_sparilut_residuals( hA, L, U, &hL, queue );
magma_sparilut_residuals( hA, L, U, &hU, queue );
end = magma_sync_wtime( queue ); t_res=+end-start;
start = magma_sync_wtime( queue );
magma_smatrix_abssum( hL, &sumL, queue );
magma_smatrix_abssum( hU, &sumU, queue );
sum = sumL + sumU;
end = magma_sync_wtime( queue ); t_nrm+=end-start;
CHECK( magma_smatrix_swap( &hL, &oneL, queue) );
magma_smfree(&hL, queue );
start = magma_sync_wtime( queue );
if( precond->pattern == 2 ){
// align residuals in U
magma_sparilut_align_residuals( L, U, &hL, &hU, queue );
} else if(precond->pattern == 0){
for(magma_int_t z=0; z<hL.nnz; z++)
hL.val[z] = MAGMA_S_ZERO;
for(magma_int_t z=0; z<hU.nnz; z++)
hU.val[z] = MAGMA_S_ZERO;
}
magma_scsrcoo_transpose( hU, &oneU, queue );
end = magma_sync_wtime( queue ); t_transpose2+=end-start;
magma_smfree(&hU, queue );
magma_smfree(&UT, queue );
} else {
start = magma_sync_wtime( queue );
magma_sparilut_residuals( hA, L, U, &hL, queue );
magma_sparilut_residuals( hA, L, U, &hU, queue );
end = magma_sync_wtime( queue ); t_res=+end-start;
start = magma_sync_wtime( queue );
magma_smatrix_abssum( hL, &sumL, queue );
magma_smatrix_abssum( hU, &sumU, queue );
sum = sumL + sumU;
end = magma_sync_wtime( queue ); t_nrm+=end-start;
start = magma_sync_wtime( queue );
magma_sparilut_transpose_select_one( hU, &oneU, queue );
magma_smfree(&hU, queue );
magma_smfree(&UT, queue );
// magma_smatrix_addrowindex( &hU, queue );
// CHECK( magma_smatrix_swap( &oneU, &hU, queue) );
// magma_smfree(&oneU, queue );
// magma_smatrix_addrowindex( &hU, queue );
end = magma_sync_wtime( queue ); t_transpose2+=end-start;
magma_sparilut_selectoneperrow( 1, &hL, &oneL, queue );
//CHECK( magma_smatrix_swap( &oneL, &hL, queue) );
//CHECK( magma_smatrix_swap( &oneU, &hU, queue) );
// use only a subset of the candidates
//magma_smfree( &oneL, queue );
//magma_smfree( &oneU, queue );
num_rmL = max(oneL.nnz * ( precond->rtol ),0);
num_rmU = max(oneU.nnz * ( precond->rtol ),0);
// num_rmL = max(hL.nnz * ( precond->rtol-0.15*iters ),0);
// num_rmU = max(hU.nnz * ( precond->rtol-0.15*iters ),0);
//printf("hL:%d hU:%d\n", num_rmL, num_rmU);
//#pragma omp parallel
{
// magma_int_t id = omp_get_thread_num();
//if( id == 0 ){
if( num_rmL>0 ){
magma_sparilut_set_thrs_randomselect( num_rmL, &hL, 1, &thrsL, queue );
} else {
thrsL = 1e6;
}
//}
//if( id == num_threads-1 ){
if( num_rmU>0 ){
magma_sparilut_set_thrs_randomselect( num_rmU, &hU, 1, &thrsU, queue );
} else {
thrsU = 1e6;
}
//}
}
magma_sparilut_thrsrm( 1, &oneL, &thrsL, queue );
magma_sparilut_thrsrm( 1, &oneU, &thrsU, queue );
}
end = magma_sync_wtime( queue ); t_selectadd+=end-start;
start = magma_sync_wtime( queue );
#pragma omp parallel
for(int row=0; row<hL.num_rows; row++){
magma_sindexsort( &hL.col[hL.row[row]], 0, hL.row[row+1]-hL.row[row]-1, queue );
}
#pragma omp parallel
for(int row=0; row<hL.num_rows; row++){
magma_sindexsort( &hU.col[hU.row[row]], 0, hU.row[row+1]-hU.row[row]-1, queue );
}
CHECK( magma_smatrix_cup( L, oneL, &L_new, queue ) );
CHECK( magma_smatrix_cup( U, oneU, &U_new, queue ) );
//magma_smatrix_addrowindex( &U, queue );
end = magma_sync_wtime( queue ); t_add=+end-start;
magma_smfree( &oneL, queue );
magma_smfree( &oneU, queue );
// using linked list
// start = magma_sync_wtime( queue );
// magma_free_cpu( UT.row ); UT.row = NULL;
// magma_free_cpu( UT.list ); UT.list = NULL;
// CHECK( magma_sparilut_create_collinkedlist( U_new, &UT, queue) );
// end = magma_sync_wtime( queue ); t_transpose2+=end-start;
start = magma_sync_wtime( queue );
// CHECK( magma_sparilut_sweep( &A0, &L_new, &U_new, queue ) );
CHECK( magma_sparilut_sweep_sync( &A0, &L_new, &U_new, queue ) );
end = magma_sync_wtime( queue ); t_sweep1+=end-start;
num_rmL = max( (L_new.nnz-L0nnz*(1+(precond->atol-1.)*(iters+1)/precond->sweeps)), 0 );
num_rmU = max( (U_new.nnz-U0nnz*(1+(precond->atol-1.)*(iters+1)/precond->sweeps)), 0 );
start = magma_sync_wtime( queue );
// pre-select: ignore the diagonal entries
magma_sparilut_preselect( 0, &L_new, &oneL, queue );
magma_sparilut_preselect( 0, &U_new, &oneU, queue );
//#pragma omp parallel
{
// magma_int_t id = omp_get_thread_num();
//if( id == 0 ){
if( num_rmL>0 ){
magma_sparilut_set_thrs_randomselect( num_rmL, &oneL, 0, &thrsL, queue );
} else {
thrsL = 0.0;
}
//}
//if( id == num_threads-1 ){
if( num_rmU>0 ){
magma_sparilut_set_thrs_randomselect( num_rmU, &oneU, 0, &thrsU, queue );
} else {
thrsU = 0.0;
}
//}
}
// if(thrsL > thrsL_old){
// thrsL=thrsL*2;
// } else {
// thrsL_old=thrsL;
// }
// if(thrsU > thrsU_old){
// thrsU=thrsU*2;
// } else {
// thrsU_old=thrsU;
// }
// magma_sparilut_set_thrs_randomselect( num_rmL, &L_new, 0, &thrsL, queue );
// magma_sparilut_set_thrs_randomselect( num_rmU, &UT, 0, &thrsU, queue );
end = magma_sync_wtime( queue ); t_selectrm=end-start;
magma_smfree( &oneL, queue );
magma_smfree( &oneU, queue );
start = magma_sync_wtime( queue );
magma_sparilut_thrsrm( 1, &L_new, &thrsL, queue );//printf("done...");fflush(stdout);
magma_sparilut_thrsrm( 1, &U_new, &thrsU, queue );//printf("done...");fflush(stdout);
// magma_sparilut_thrsrm_U( 1, L_new, &U_new, &thrsU, queue );
// for(int z=0; z<L_new.nnz; z++){
// if(MAGMA_S_ABS(L_new.val[z])<thrsL){
// printf("invalid element here:%.4e < %.4e <%.4e> \n",MAGMA_S_ABS(L_new.val[z]),thrsL, MAGMA_S_REAL(L_new.val[z]));
// }
// }
//
// for(int z=0; z<U_new.nnz; z++){
// if(MAGMA_S_ABS(U_new.val[z])<thrsU){
// printf("invalid element here:%.4e < %.4e <%.4e> \n",MAGMA_S_ABS(U_new.val[z]),thrsU, MAGMA_S_REAL(U_new.val[z]));
// }
// }
// magma_sparilut_thrsrm_semilinked( &U_new, &UT, &thrsU, queue );//printf("done.\n");fflush(stdout);
CHECK( magma_smatrix_swap( &L_new, &L, queue) );
CHECK( magma_smatrix_swap( &U_new, &U, queue) );
magma_smfree( &L_new, queue );
magma_smfree( &U_new, queue );
end = magma_sync_wtime( queue ); t_rm=end-start;
start = magma_sync_wtime( queue );
// magma_free_cpu( UT.row ); UT.row = NULL;
// magma_free_cpu( UT.list ); UT.list = NULL;
// CHECK( magma_sparilut_create_collinkedlist( U, &UT, queue) );
// end = magma_sync_wtime( queue ); t_transpose1+=end-start;
start = magma_sync_wtime( queue );
//magma_free_cpu( U.rowidx ); U.rowidx = NULL;
// CHECK( magma_sparilut_sweep( &A0, &L, &U, queue ) );
CHECK( magma_sparilut_sweep_sync( &A0, &L, &U, queue ) );
end = magma_sync_wtime( queue ); t_sweep2+=end-start;
start = magma_sync_wtime( queue );
end = magma_sync_wtime( queue ); t_rm+=end-start;
// end using linked list
if( timing == 1 ){
t_total = t_cand+t_res+t_nrm+t_selectadd+t_add+t_transpose1+t_sweep1+t_selectrm+t_rm+t_sweep2+t_transpose2;
accum = accum + t_total;
printf("%5lld %5lld %5lld %.4e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e\n",
(long long) iters, (long long) L.nnz, (long long) U.nnz, (float) sum,
t_cand, t_res, t_nrm, t_selectadd, t_add, t_transpose1, t_sweep1, t_selectrm, t_rm, t_sweep2, t_transpose2, t_total, accum );
fflush(stdout);
}
}
if (timing == 1) {
printf("]; \n");
}
//##########################################################################
//printf("%% check L:\n"); fflush(stdout);
//magma_sdiagcheck_cpu( hL, queue );
//printf("%% check U:\n"); fflush(stdout);
//magma_sdiagcheck_cpu( hU, queue );
// for CUSPARSE
CHECK( magma_smtransfer( L, &precond->L, Magma_CPU, Magma_DEV , queue ));
magma_scsrcoo_transpose( U, &UT, queue );
//magma_smtranspose(U, &UT, queue );
CHECK( magma_smtransfer( UT, &precond->U, Magma_CPU, Magma_DEV , queue ));
CHECK(magma_strisolve_analysis(precond->L, &precond->cuinfoL, false, false, false, queue));
CHECK(magma_strisolve_analysis(precond->U, &precond->cuinfoU, true, false, false, queue));
if( precond->trisolver != 0 && precond->trisolver != Magma_CUSOLVE ){
//prepare for iterative solves
// extract the diagonal of L into precond->d
CHECK( magma_sjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_svinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_S_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_sjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_svinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_S_ZERO, queue ));
}
if( precond->trisolver == Magma_JACOBI && precond->pattern == 1 ){
// dirty workaround for Jacobi trisolves....
magma_smfree( &hL, queue );
magma_smfree( &hU, queue );
CHECK( magma_smtransfer( precond->U, &hU, Magma_DEV, Magma_CPU , queue ));
CHECK( magma_smtransfer( precond->L, &hL, Magma_DEV, Magma_CPU , queue ));
magma_smfree( &hAT, queue );
hAT.diagorder_type = Magma_VALUE;
CHECK( magma_smconvert( hL, &hAT , Magma_CSR, Magma_CSRU, queue ));
#pragma omp parallel for
for (magma_int_t i=0; i<hAT.nnz; i++) {
hAT.val[i] = MAGMA_S_ONE/hAT.val[i];
}
CHECK( magma_smtransfer( hAT, &(precond->LD), Magma_CPU, Magma_DEV, queue ));
magma_smfree( &hAT, queue );
hAT.diagorder_type = Magma_VALUE;
CHECK( magma_smconvert( hU, &hAT , Magma_CSR, Magma_CSRL, queue ));
#pragma omp parallel for
for (magma_int_t i=0; i<hAT.nnz; i++) {
hAT.val[i] = MAGMA_S_ONE/hAT.val[i];
}
CHECK( magma_smtransfer( hAT, &(precond->UD), Magma_CPU, Magma_DEV, queue ));
}
cleanup:
cusparseDestroy( cusparseHandle );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseHandle=NULL;
descrL=NULL;
descrU=NULL;
magma_smfree( &hA, queue );
magma_smfree( &hAT, queue );
magma_smfree( &A0, queue );
magma_smfree( &L0, queue );
magma_smfree( &U0, queue );
magma_smfree( &hAT, queue );
magma_smfree( &hL, queue );
magma_smfree( &L, queue );
magma_smfree( &L_new, queue );
magma_smfree( &hU, queue );
magma_smfree( &U, queue );
magma_smfree( &UT, queue );
magma_smfree( &U_new, queue );
//magma_smfree( &UT, queue );
#endif
return info;
}
|
