PLASMA
Parallel Linear Algebra Software for Multicore Architectures
plasma_core_blas_d.h
1
12#ifndef PLASMA_CORE_BLAS_D_H
13#define PLASMA_CORE_BLAS_D_H
14
15#include "plasma_async.h"
16#include "plasma_barrier.h"
17#include "plasma_descriptor.h"
18#include "plasma_types.h"
19#include "plasma_workspace.h"
20#include "plasma_descriptor.h"
21
22#ifdef __cplusplus
23extern "C" {
24#endif
25
26#define REAL
27
28/******************************************************************************/
29#ifdef COMPLEX
30double fabs(double alpha);
31#endif
32
33int plasma_core_dgeadd(plasma_enum_t transa,
34 int m, int n,
35 double alpha, const double *A, int lda,
36 double beta, double *B, int ldb);
37
38int plasma_core_dgelqt(int m, int n, int ib,
39 double *A, int lda,
40 double *T, int ldt,
41 double *tau,
42 double *work);
43
44void plasma_core_dgemm(plasma_enum_t transa, plasma_enum_t transb,
45 int m, int n, int k,
46 double alpha, const double *A, int lda,
47 const double *B, int ldb,
48 double beta, double *C, int ldc);
49
50int plasma_core_dgeqrt(int m, int n, int ib,
51 double *A, int lda,
52 double *T, int ldt,
53 double *tau,
54 double *work);
55
56void plasma_core_dgessq(int m, int n,
57 const double *A, int lda,
58 double *scale, double *sumsq);
59
60void plasma_core_dgetrf(plasma_desc_t A, int *ipiv, int ib, int rank, int size,
61 volatile int *max_idx, volatile double *max_val,
62 volatile int *info, plasma_barrier_t *barrier);
63
64int plasma_core_dsygst(int itype, plasma_enum_t uplo,
65 int n,
66 double *A, int lda,
67 double *B, int ldb);
68
69void plasma_core_dsymm(plasma_enum_t side, plasma_enum_t uplo,
70 int m, int n,
71 double alpha, const double *A, int lda,
72 const double *B, int ldb,
73 double beta, double *C, int ldc);
74
75void plasma_core_dsyr2k(plasma_enum_t uplo, plasma_enum_t trans,
76 int n, int k,
77 double alpha, const double *A, int lda,
78 const double *B, int ldb,
79 double beta, double *C, int ldc);
80
81void plasma_core_dsyrk(plasma_enum_t uplo, plasma_enum_t trans,
82 int n, int k,
83 double alpha, const double *A, int lda,
84 double beta, double *C, int ldc);
85
86void plasma_core_dsyssq(plasma_enum_t uplo,
87 int n,
88 const double *A, int lda,
89 double *scale, double *sumsq);
90
91void plasma_core_dsyssq(plasma_enum_t uplo,
92 int n,
93 const double *A, int lda,
94 double *scale, double *sumsq);
95
96void plasma_core_dlacpy(plasma_enum_t uplo, plasma_enum_t transa,
97 int m, int n,
98 const double *A, int lda,
99 double *B, int ldb);
100
101void plasma_core_dlacpy_lapack2tile_band(plasma_enum_t uplo,
102 int it, int jt,
103 int m, int n, int nb, int kl, int ku,
104 const double *A, int lda,
105 double *B, int ldb);
106
107void plasma_core_dlacpy_tile2lapack_band(plasma_enum_t uplo,
108 int it, int jt,
109 int m, int n, int nb, int kl, int ku,
110 const double *B, int ldb,
111 double *A, int lda);
112
113void plasma_core_dlange(plasma_enum_t norm,
114 int m, int n,
115 const double *A, int lda,
116 double *work, double *result);
117
118void plasma_core_dlansy(plasma_enum_t norm, plasma_enum_t uplo,
119 int n,
120 const double *A, int lda,
121 double *work, double *value);
122
123void plasma_core_dlansy(plasma_enum_t norm, plasma_enum_t uplo,
124 int n,
125 const double *A, int lda,
126 double *work, double *value);
127
128void plasma_core_dlantr(plasma_enum_t norm, plasma_enum_t uplo, plasma_enum_t diag,
129 int m, int n,
130 const double *A, int lda,
131 double *work, double *value);
132
133void plasma_core_dlascl(plasma_enum_t uplo,
134 double cfrom, double cto,
135 int m, int n,
136 double *A, int lda);
137
138void plasma_core_dlaset(plasma_enum_t uplo,
139 int m, int n,
140 double alpha, double beta,
141 double *A, int lda);
142
143void plasma_core_dgeswp(plasma_enum_t colrow,
144 plasma_desc_t A, int k1, int k2, const int *ipiv, int incx);
145
146void plasma_core_dsyswp(int rank, int num_threads,
147 int uplo, plasma_desc_t A, int k1, int k2, const int *ipiv,
148 int incx, plasma_barrier_t *barrier);
149
150int plasma_core_dlauum(plasma_enum_t uplo,
151 int n,
152 double *A, int lda);
153
154int plasma_core_dpamm(plasma_enum_t op, plasma_enum_t side, plasma_enum_t storev,
155 int m, int n, int k, int l,
156 const double *A1, int lda1,
157 double *A2, int lda2,
158 const double *V, int ldv,
159 double *W, int ldw);
160
161int plasma_core_dparfb(plasma_enum_t side, plasma_enum_t trans, plasma_enum_t direct,
162 plasma_enum_t storev,
163 int m1, int n1, int m2, int n2, int k, int l,
164 double *A1, int lda1,
165 double *A2, int lda2,
166 const double *V, int ldv,
167 const double *T, int ldt,
168 double *work, int ldwork);
169
170int plasma_core_dpemv(plasma_enum_t trans, int storev,
171 int m, int n, int l,
172 double alpha,
173 const double *A, int lda,
174 const double *X, int incx,
175 double beta,
176 double *Y, int incy,
177 double *work);
178
179int plasma_core_dpotrf(plasma_enum_t uplo,
180 int n,
181 double *A, int lda);
182
183void plasma_core_dsymm(plasma_enum_t side, plasma_enum_t uplo,
184 int m, int n,
185 double alpha, const double *A, int lda,
186 const double *B, int ldb,
187 double beta, double *C, int ldc);
188
189void plasma_core_dsyr2k(
190 plasma_enum_t uplo, plasma_enum_t trans,
191 int n, int k,
192 double alpha, const double *A, int lda,
193 const double *B, int ldb,
194 double beta, double *C, int ldc);
195
196void plasma_core_dsyrk(plasma_enum_t uplo, plasma_enum_t trans,
197 int n, int k,
198 double alpha, const double *A, int lda,
199 double beta, double *C, int ldc);
200
201int plasma_core_dtradd(plasma_enum_t uplo, plasma_enum_t transa,
202 int m, int n,
203 double alpha, const double *A, int lda,
204 double beta, double *B, int ldb);
205
206void plasma_core_dtrmm(plasma_enum_t side, plasma_enum_t uplo,
207 plasma_enum_t transa, plasma_enum_t diag,
208 int m, int n,
209 double alpha, const double *A, int lda,
210 double *B, int ldb);
211
212void plasma_core_dtrsm(plasma_enum_t side, plasma_enum_t uplo,
213 plasma_enum_t transa, plasma_enum_t diag,
214 int m, int n,
215 double alpha, const double *A, int lda,
216 double *B, int ldb);
217
218void plasma_core_dtrssq(plasma_enum_t uplo, plasma_enum_t diag,
219 int m, int n,
220 const double *A, int lda,
221 double *scale, double *sumsq);
222
223int plasma_core_dtrtri(plasma_enum_t uplo, plasma_enum_t diag,
224 int n,
225 double *A, int lda);
226
227int plasma_core_dtslqt(int m, int n, int ib,
228 double *A1, int lda1,
229 double *A2, int lda2,
230 double *T, int ldt,
231 double *tau,
232 double *work);
233
234int plasma_core_dtsmlq(plasma_enum_t side, plasma_enum_t trans,
235 int m1, int n1, int m2, int n2, int k, int ib,
236 double *A1, int lda1,
237 double *A2, int lda2,
238 const double *V, int ldv,
239 const double *T, int ldt,
240 double *work, int ldwork);
241
242int plasma_core_dtsmqr(plasma_enum_t side, plasma_enum_t trans,
243 int m1, int n1, int m2, int n2, int k, int ib,
244 double *A1, int lda1,
245 double *A2, int lda2,
246 const double *V, int ldv,
247 const double *T, int ldt,
248 double *work, int ldwork);
249
250int plasma_core_dtsqrt(int m, int n, int ib,
251 double *A1, int lda1,
252 double *A2, int lda2,
253 double *T, int ldt,
254 double *tau,
255 double *work);
256
257int plasma_core_dttlqt(int m, int n, int ib,
258 double *A1, int lda1,
259 double *A2, int lda2,
260 double *T, int ldt,
261 double *tau,
262 double *work);
263
264int plasma_core_dttmlq(plasma_enum_t side, plasma_enum_t trans,
265 int m1, int n1, int m2, int n2, int k, int ib,
266 double *A1, int lda1,
267 double *A2, int lda2,
268 const double *V, int ldv,
269 const double *T, int ldt,
270 double *work, int ldwork);
271
272int plasma_core_dttmqr(plasma_enum_t side, plasma_enum_t trans,
273 int m1, int n1, int m2, int n2, int k, int ib,
274 double *A1, int lda1,
275 double *A2, int lda2,
276 const double *V, int ldv,
277 const double *T, int ldt,
278 double *work, int ldwork);
279
280int plasma_core_dttqrt(int m, int n, int ib,
281 double *A1, int lda1,
282 double *A2, int lda2,
283 double *T, int ldt,
284 double *tau,
285 double *work);
286
287int plasma_core_dormlq(plasma_enum_t side, plasma_enum_t trans,
288 int m, int n, int k, int ib,
289 const double *A, int lda,
290 const double *T, int ldt,
291 double *C, int ldc,
292 double *work, int ldwork);
293
294int plasma_core_dormqr(plasma_enum_t side, plasma_enum_t trans,
295 int m, int n, int k, int ib,
296 const double *A, int lda,
297 const double *T, int ldt,
298 double *C, int ldc,
299 double *work, int ldwork);
300
301/******************************************************************************/
302void plasma_core_omp_damax(int colrow, int m, int n,
303 const double *A, int lda,
304 double *values,
305 plasma_sequence_t *sequence, plasma_request_t *request);
306
307void plasma_core_omp_dgeadd(
308 plasma_enum_t transa, int m, int n,
309 double alpha, const double *A, int lda,
310 double beta, double *B, int ldb,
311 plasma_sequence_t *sequence, plasma_request_t *request);
312
313void plasma_core_omp_dgelqt(int m, int n, int ib,
314 double *A, int lda,
315 double *T, int ldt,
316 plasma_workspace_t work,
317 plasma_sequence_t *sequence, plasma_request_t *request);
318
319void plasma_core_omp_dgemm(
320 plasma_enum_t transa, plasma_enum_t transb,
321 int m, int n, int k,
322 double alpha, const double *A, int lda,
323 const double *B, int ldb,
324 double beta, double *C, int ldc,
325 plasma_sequence_t *sequence, plasma_request_t *request);
326
327void plasma_core_omp_dgeqrt(int m, int n, int ib,
328 double *A, int lda,
329 double *T, int ldt,
330 plasma_workspace_t work,
331 plasma_sequence_t *sequence, plasma_request_t *request);
332
333void plasma_core_omp_dgessq(int m, int n,
334 const double *A, int lda,
335 double *scale, double *sumsq,
336 plasma_sequence_t *sequence, plasma_request_t *request);
337
338void plasma_core_omp_dgessq_aux(int n,
339 const double *scale, const double *sumsq,
340 double *value,
341 plasma_sequence_t *sequence,
342 plasma_request_t *request);
343
344void plasma_core_omp_dsygst(int itype, plasma_enum_t uplo,
345 int n,
346 double *A, int lda,
347 double *B, int ldb,
348 plasma_sequence_t *sequence, plasma_request_t *request);
349
350void plasma_core_omp_dsymm(
351 plasma_enum_t side, plasma_enum_t uplo,
352 int m, int n,
353 double alpha, const double *A, int lda,
354 const double *B, int ldb,
355 double beta, double *C, int ldc,
356 plasma_sequence_t *sequence, plasma_request_t *request);
357
358void plasma_core_omp_dsyr2k(
359 plasma_enum_t uplo, plasma_enum_t trans,
360 int n, int k,
361 double alpha, const double *A, int lda,
362 const double *B, int ldb,
363 double beta, double *C, int ldc,
364 plasma_sequence_t *sequence, plasma_request_t *request);
365
366void plasma_core_omp_dsyrk(plasma_enum_t uplo, plasma_enum_t trans,
367 int n, int k,
368 double alpha, const double *A, int lda,
369 double beta, double *C, int ldc,
370 plasma_sequence_t *sequence, plasma_request_t *request);
371
372void plasma_core_omp_dsyssq(plasma_enum_t uplo,
373 int n,
374 const double *A, int lda,
375 double *scale, double *sumsq,
376 plasma_sequence_t *sequence, plasma_request_t *request);
377
378void plasma_core_omp_dsyssq(plasma_enum_t uplo,
379 int n,
380 const double *A, int lda,
381 double *scale, double *sumsq,
382 plasma_sequence_t *sequence, plasma_request_t *request);
383
384void plasma_core_omp_dsyssq_aux(int m, int n,
385 const double *scale, const double *sumsq,
386 double *value,
387 plasma_sequence_t *sequence,
388 plasma_request_t *request);
389
390void plasma_core_omp_dlacpy(plasma_enum_t uplo, plasma_enum_t transa,
391 int m, int n,
392 const double *A, int lda,
393 double *B, int ldb,
394 plasma_sequence_t *sequence, plasma_request_t *request);
395
396void plasma_core_omp_dlacpy_lapack2tile_band(plasma_enum_t uplo,
397 int it, int jt,
398 int m, int n, int nb, int kl, int ku,
399 const double *A, int lda,
400 double *B, int ldb);
401
402void plasma_core_omp_dlacpy_tile2lapack_band(plasma_enum_t uplo,
403 int it, int jt,
404 int m, int n, int nb, int kl, int ku,
405 const double *B, int ldb,
406 double *A, int lda);
407
408void plasma_core_omp_dlange(plasma_enum_t norm,
409 int m, int n,
410 const double *A, int lda,
411 double *work, double *result,
412 plasma_sequence_t *sequence, plasma_request_t *request);
413
414void plasma_core_omp_dlange_aux(plasma_enum_t norm,
415 int m, int n,
416 const double *A, int lda,
417 double *value,
418 plasma_sequence_t *sequence,
419 plasma_request_t *request);
420
421void plasma_core_omp_dlansy(plasma_enum_t norm, plasma_enum_t uplo,
422 int n,
423 const double *A, int lda,
424 double *work, double *value,
425 plasma_sequence_t *sequence, plasma_request_t *request);
426
427void plasma_core_omp_dlansy_aux(plasma_enum_t norm, plasma_enum_t uplo,
428 int n,
429 const double *A, int lda,
430 double *value,
431 plasma_sequence_t *sequence,
432 plasma_request_t *request);
433
434void plasma_core_omp_dlansy(plasma_enum_t norm, plasma_enum_t uplo,
435 int n,
436 const double *A, int lda,
437 double *work, double *value,
438 plasma_sequence_t *sequence, plasma_request_t *request);
439
440void plasma_core_omp_dlansy_aux(plasma_enum_t norm, plasma_enum_t uplo,
441 int n,
442 const double *A, int lda,
443 double *value,
444 plasma_sequence_t *sequence,
445 plasma_request_t *request);
446
447void plasma_core_omp_dlantr(plasma_enum_t norm, plasma_enum_t uplo, plasma_enum_t diag,
448 int m, int n,
449 const double *A, int lda,
450 double *work, double *value,
451 plasma_sequence_t *sequence, plasma_request_t *request);
452
453void plasma_core_omp_dlantr_aux(plasma_enum_t norm, plasma_enum_t uplo,
454 plasma_enum_t diag,
455 int m, int n,
456 const double *A, int lda,
457 double *value,
458 plasma_sequence_t *sequence,
459 plasma_request_t *request);
460
461void plasma_core_omp_dlascl(plasma_enum_t uplo,
462 double cfrom, double cto,
463 int m, int n,
464 double *A, int lda,
465 plasma_sequence_t *sequence, plasma_request_t *request);
466
467void plasma_core_omp_dlaset(plasma_enum_t uplo,
468 int mb, int nb,
469 int i, int j,
470 int m, int n,
471 double alpha, double beta,
472 double *A);
473
474void plasma_core_omp_dlauum(plasma_enum_t uplo,
475 int n,
476 double *A, int lda,
477 plasma_sequence_t *sequence, plasma_request_t *request);
478
479void plasma_core_omp_dpotrf(plasma_enum_t uplo,
480 int n,
481 double *A, int lda,
482 int iinfo,
483 plasma_sequence_t *sequence, plasma_request_t *request);
484
485void plasma_core_omp_dsymm(
486 plasma_enum_t side, plasma_enum_t uplo,
487 int m, int n,
488 double alpha, const double *A, int lda,
489 const double *B, int ldb,
490 double beta, double *C, int ldc,
491 plasma_sequence_t *sequence, plasma_request_t *request);
492
493void plasma_core_omp_dsyr2k(
494 plasma_enum_t uplo, plasma_enum_t trans,
495 int n, int k,
496 double alpha, const double *A, int lda,
497 const double *B, int ldb,
498 double beta, double *C, int ldc,
499 plasma_sequence_t *sequence, plasma_request_t *request);
500
501void plasma_core_omp_dsyrk(
502 plasma_enum_t uplo, plasma_enum_t trans,
503 int n, int k,
504 double alpha, const double *A, int lda,
505 double beta, double *C, int ldc,
506 plasma_sequence_t *sequence, plasma_request_t *request);
507
508void plasma_core_omp_dtradd(
509 plasma_enum_t uplo, plasma_enum_t transa,
510 int m, int n,
511 double alpha, const double *A, int lda,
512 double beta, double *B, int ldb,
513 plasma_sequence_t *sequence, plasma_request_t *request);
514
515void plasma_core_omp_dtrmm(
516 plasma_enum_t side, plasma_enum_t uplo,
517 plasma_enum_t transa, plasma_enum_t diag,
518 int m, int n,
519 double alpha, const double *A, int lda,
520 double *B, int ldb,
521 plasma_sequence_t *sequence, plasma_request_t *request);
522
523void plasma_core_omp_dtrsm(
524 plasma_enum_t side, plasma_enum_t uplo,
525 plasma_enum_t transa, plasma_enum_t diag,
526 int m, int n,
527 double alpha, const double *A, int lda,
528 double *B, int ldb,
529 plasma_sequence_t *sequence, plasma_request_t *request);
530
531void plasma_core_omp_dtrssq(plasma_enum_t uplo, plasma_enum_t diag,
532 int m, int n,
533 const double *A, int lda,
534 double *scale, double *sumsq,
535 plasma_sequence_t *sequence, plasma_request_t *request);
536
537void plasma_core_omp_dtrtri(plasma_enum_t uplo, plasma_enum_t diag,
538 int n,
539 double *A, int lda,
540 int iinfo,
541 plasma_sequence_t *sequence, plasma_request_t *request);
542
543void plasma_core_omp_dtslqt(int m, int n, int ib,
544 double *A1, int lda1,
545 double *A2, int lda2,
546 double *T, int ldt,
547 plasma_workspace_t work,
548 plasma_sequence_t *sequence, plasma_request_t *request);
549
550void plasma_core_omp_dtsmlq(plasma_enum_t side, plasma_enum_t trans,
551 int m1, int n1, int m2, int n2, int k, int ib,
552 double *A1, int lda1,
553 double *A2, int lda2,
554 const double *V, int ldv,
555 const double *T, int ldt,
556 plasma_workspace_t work,
557 plasma_sequence_t *sequence, plasma_request_t *request);
558
559void plasma_core_omp_dtsmqr(plasma_enum_t side, plasma_enum_t trans,
560 int m1, int n1, int m2, int n2, int k, int ib,
561 double *A1, int lda1,
562 double *A2, int lda2,
563 const double *V, int ldv,
564 const double *T, int ldt,
565 plasma_workspace_t work,
566 plasma_sequence_t *sequence, plasma_request_t *request);
567
568void plasma_core_omp_dtsqrt(int m, int n, int ib,
569 double *A1, int lda1,
570 double *A2, int lda2,
571 double *T, int ldt,
572 plasma_workspace_t work,
573 plasma_sequence_t *sequence, plasma_request_t *request);
574
575void plasma_core_omp_dttlqt(int m, int n, int ib,
576 double *A1, int lda1,
577 double *A2, int lda2,
578 double *T, int ldt,
579 plasma_workspace_t work,
580 plasma_sequence_t *sequence, plasma_request_t *request);
581
582void plasma_core_omp_dttmlq(plasma_enum_t side, plasma_enum_t trans,
583 int m1, int n1, int m2, int n2, int k, int ib,
584 double *A1, int lda1,
585 double *A2, int lda2,
586 const double *V, int ldv,
587 const double *T, int ldt,
588 plasma_workspace_t work,
589 plasma_sequence_t *sequence, plasma_request_t *request);
590
591void plasma_core_omp_dttmqr(plasma_enum_t side, plasma_enum_t trans,
592 int m1, int n1, int m2, int n2, int k, int ib,
593 double *A1, int lda1,
594 double *A2, int lda2,
595 const double *V, int ldv,
596 const double *T, int ldt,
597 plasma_workspace_t work,
598 plasma_sequence_t *sequence, plasma_request_t *request);
599
600void plasma_core_omp_dttqrt(int m, int n, int ib,
601 double *A1, int lda1,
602 double *A2, int lda2,
603 double *T, int ldt,
604 plasma_workspace_t work,
605 plasma_sequence_t *sequence, plasma_request_t *request);
606
607void plasma_core_omp_dormlq(plasma_enum_t side, plasma_enum_t trans,
608 int m, int n, int k, int ib,
609 const double *A, int lda,
610 const double *T, int ldt,
611 double *C, int ldc,
612 plasma_workspace_t work,
613 plasma_sequence_t *sequence, plasma_request_t *request);
614
615void plasma_core_omp_dormqr(plasma_enum_t side, plasma_enum_t trans,
616 int m, int n, int k, int ib,
617 const double *A, int lda,
618 const double *T, int ldt,
619 double *C, int ldc,
620 plasma_workspace_t work,
621 plasma_sequence_t *sequence, plasma_request_t *request);
622
623#undef REAL
624
625#ifdef __cplusplus
626} // extern "C"
627#endif
628
629#endif // PLASMA_CORE_BLAS_D_H
Definition: plasma_descriptor.h:40