18 #ifndef _PLASMA_CORE_SBLAS_H_
19 #define _PLASMA_CORE_SBLAS_H_
30 void CORE_sasum(
int storev, PLASMA_enum uplo,
int M,
int N,
31 const float *A,
int lda,
float *work);
41 int Vblksiz,
int wantz,
42 int i,
int sweepid,
int m,
int grsiz,
46 int st,
int ed,
int eltsize);
49 int st,
int ed,
int eltsize);
52 int st,
int ed,
int eltsize);
54 const float *A,
int LDA,
61 void CORE_sgemm(PLASMA_enum transA, PLASMA_enum transB,
63 float alpha,
const float *A,
int LDA,
64 const float *B,
int LDB,
65 float beta,
float *C,
int LDC);
66 void CORE_sgemv(PLASMA_enum trans,
int M,
int N,
67 float alpha,
const float *A,
int LDA,
68 const float *x,
int incx,
69 float beta,
float *y,
int incy);
72 float *tau,
float *beta );
75 int jj,
int k,
int *jpvt,
76 float *norms1,
float *norms2,
int *info );
77 int CORE_sgeqp3_tntpiv(
int m,
int n,
79 int *IPIV,
float *tau,
83 const float *Fk,
int ldf,
84 int joff,
int k,
int koff,
int nb,
85 float *norms1,
float *norms2,
90 float *TAU,
float *WORK);
93 const float *L,
int LDL,
95 int CORE_sgessq(
int M,
int N,
96 const float *A,
int LDA,
97 float *scale,
float *sumsq);
102 int *IPIV,
int *INFO);
105 int *IPIV,
int *INFO);
110 int *IPIV,
int *info);
111 void CORE_sgetrf_reclap_init(
void);
113 void CORE_sgetrf_rectil_init(
void);
118 int st,
int ed,
int eltsize);
121 int st,
int ed,
int eltsize);
124 int st,
int ed,
int eltsize);
127 float *V,
float *TAU,
128 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
132 float *V,
float *TAU,
133 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
137 const float *V,
const float *TAU,
138 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
142 float *VQ,
float *TAUQ,
143 float *VP,
float *TAUP,
144 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
148 float *VQ,
float *TAUQ,
149 float *VP,
float *TAUP,
150 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
154 float *VQ,
float *TAUQ,
155 float *VP,
float *TAUP,
156 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
158 void CORE_ssygst(
int itype, PLASMA_enum uplo,
int N,
160 float *B,
int LDB,
int *INFO);
162 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
164 float alpha,
const float *A,
int LDA,
165 const float *B,
int LDB,
166 float beta,
float *C,
int LDC);
167 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
169 float alpha,
const float *A,
int LDA,
170 float beta,
float *C,
int LDC);
171 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
173 float alpha,
const float *A,
int LDA,
174 const float *B,
int LDB,
175 float beta,
float *C,
int LDC);
176 int CORE_shessq(PLASMA_enum uplo,
int N,
177 const float *A,
int LDA,
178 float *scale,
float *sumsq);
180 int CORE_ssyrfb(PLASMA_enum uplo,
int N,
int K,
int IB,
int NB,
181 const float *A,
int LDA,
182 const float *T,
int LDT,
184 float *WORK,
int LDWORK);
186 const float *A,
int LDA,
190 int k1,
int k2,
const int *ipiv,
191 int *rankin,
int *rankout,
195 const float *A,
int LDA,
196 float *work,
float *normA);
198 void CORE_slansy(
int norm, PLASMA_enum uplo,
int N,
199 const float *A,
int LDA,
200 float *work,
float *normA);
202 void CORE_slansy(
int norm, PLASMA_enum uplo,
int N,
203 const float *A,
int LDA,
204 float *work,
float *normA);
205 void CORE_slantr(PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag,
207 const float *A,
int LDA,
208 float *work,
float *normA);
209 int CORE_slarfb_gemm(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
211 const float *V,
int LDV,
212 const float *T,
int LDT,
214 float *WORK,
int LDWORK);
219 float *C2,
int LDC2);
238 float alpha,
float beta,
239 float *tileA,
int ldtilea);
240 void CORE_slaset2(PLASMA_enum uplo,
int n1,
int n2,
float alpha,
241 float *tileA,
int ldtilea);
243 int I1,
int I2,
const int *IPIV,
int INC);
246 int CORE_slatro(PLASMA_enum uplo, PLASMA_enum trans,
248 const float *A,
int LDA,
250 void CORE_slauum(PLASMA_enum uplo,
int N,
float *A,
int LDA);
251 int CORE_spamm(
int op, PLASMA_enum side, PLASMA_enum storev,
252 int M,
int N,
int K,
int L,
253 const float *A1,
int LDA1,
255 const float *V,
int LDV,
257 int CORE_sparfb(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
258 int M1,
int N1,
int M2,
int N2,
int K,
int L,
261 const float *V,
int LDV,
262 const float *T,
int LDT,
263 float *WORK,
int LDWORK);
264 int CORE_spemv(PLASMA_enum trans, PLASMA_enum storev,
267 const float *A,
int LDA,
268 const float *X,
int INCX,
272 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
273 int bigM,
int m0,
int n0,
unsigned long long int seed );
274 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
275 int bigM,
int m0,
int n0,
unsigned long long int seed );
277 int bigM,
int m0,
int n0,
unsigned long long int seed );
278 int CORE_spltmg(PLASMA_enum mtxtype,
int m,
int n,
float *A,
int lda,
279 int gM,
int gN,
int m0,
int n0,
unsigned long long int seed );
281 int gN,
int m0,
int n0,
284 int gM,
int m0,
int n0,
288 const float *X,
int incX,
289 const float *Y,
int incY,
292 int m0,
int n0,
int nb,
296 unsigned long long int seed );
300 void CORE_spotrf(PLASMA_enum uplo,
int N,
float *A,
int LDA,
int *INFO);
302 void CORE_sshift(
int s,
int m,
int n,
int L,
306 int CORE_sssssm(
int M1,
int N1,
int M2,
int N2,
int K,
int IB,
309 const float *L1,
int LDL1,
310 const float *L2,
int LDL2,
312 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
314 float alpha,
const float *A,
int LDA,
315 const float *B,
int LDB,
316 float beta,
float *C,
int LDC);
317 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
319 float alpha,
const float *A,
int LDA,
320 float beta,
float *C,
int LDC);
321 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
323 float alpha,
const float *A,
int LDA,
324 const float *B,
int LDB,
325 float beta,
float *C,
int LDC);
326 int CORE_ssyssq(PLASMA_enum uplo,
int N,
327 const float *A,
int LDA,
328 float *scale,
float *sumsq);
330 float *A,
float *work);
332 const float *Akk,
int ldak);
333 void CORE_strasm(PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag,
334 int M,
int N,
const float *A,
int lda,
float *work);
341 int Vblksiz,
int wantz,
342 int i,
int sweepid,
int m,
int grsiz,
344 void CORE_strmm(PLASMA_enum side, PLASMA_enum uplo,
345 PLASMA_enum transA, PLASMA_enum diag,
347 float alpha,
const float *A,
int LDA,
349 void CORE_strsm(PLASMA_enum side, PLASMA_enum uplo,
350 PLASMA_enum transA, PLASMA_enum diag,
352 float alpha,
const float *A,
int LDA,
354 int CORE_strssq(PLASMA_enum uplo, PLASMA_enum diag,
int M,
int N,
355 const float *A,
int LDA,
356 float *scale,
float *sumsq);
357 void CORE_strtri(PLASMA_enum uplo, PLASMA_enum diag,
int N,
358 float *A,
int LDA,
int *info);
363 float *TAU,
float *WORK);
364 int CORE_stsmlq(PLASMA_enum side, PLASMA_enum trans,
365 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
368 const float *V,
int LDV,
369 const float *T,
int LDT,
370 float *WORK,
int LDWORK);
372 int k,
int ib,
int nb,
376 const float *V,
int ldv,
377 const float *T,
int ldt,
378 float *WORK,
int ldwork);
380 int m1,
int n1,
int m2,
int n2,
384 const float *V,
int ldv,
385 const float *T,
int ldt,
386 float *WORK,
int ldwork);
387 int CORE_stsmqr(PLASMA_enum side, PLASMA_enum trans,
388 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
391 const float *V,
int LDV,
392 const float *T,
int LDT,
393 float *WORK,
int LDWORK);
395 int k,
int ib,
int nb,
399 const float *V,
int ldv,
400 const float *T,
int ldt,
401 float *WORK,
int ldwork);
403 int m1,
int n1,
int m2,
int n2,
407 const float *V,
int ldv,
408 const float *T,
int ldt,
409 float *WORK,
int ldwork);
414 float *TAU,
float *WORK);
419 int *IPIV,
float *WORK,
420 int LDWORK,
int *INFO);
421 int CORE_sttmqr(PLASMA_enum side, PLASMA_enum trans,
422 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
425 const float *V,
int LDV,
426 const float *T,
int LDT,
427 float *WORK,
int LDWORK);
434 int CORE_sttmlq(PLASMA_enum side, PLASMA_enum trans,
435 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
438 const float *V,
int LDV,
439 const float *T,
int LDT,
440 float *WORK,
int LDWORK);
447 int CORE_sormlq(PLASMA_enum side, PLASMA_enum trans,
448 int M,
int N,
int IB,
int K,
449 const float *V,
int LDV,
450 const float *T,
int LDT,
452 float *WORK,
int LDWORK);
453 int CORE_sormqr(PLASMA_enum side, PLASMA_enum trans,
454 int M,
int N,
int K,
int IB,
455 const float *V,
int LDV,
456 const float *T,
int LDT,
458 float *WORK,
int LDWORK);
464 void QUARK_CORE_sasum(Quark *quark, Quark_Task_Flags *task_flags,
465 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
466 const float *A,
int lda,
int szeA,
467 float *work,
int szeW);
468 void QUARK_CORE_sasum_f1(Quark *quark, Quark_Task_Flags *task_flags,
469 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
470 const float *A,
int lda,
int szeA,
471 float *work,
int szeW,
472 float *fake,
int szeF);
473 void QUARK_CORE_sgeadd(Quark *quark, Quark_Task_Flags *task_flags,
474 int m,
int n,
int nb,
float alpha,
475 const float *A,
int lda,
477 void QUARK_CORE_sbrdalg1(Quark *quark, Quark_Task_Flags *task_flags,
486 int Vblksiz,
int wantz,
487 int i,
int sweepid,
int m,
int grsiz,
488 int *PCOL,
int *ACOL,
int *MCOL);
489 void QUARK_CORE_sgelqt(Quark *quark, Quark_Task_Flags *task_flags,
490 int m,
int n,
int ib,
int nb,
493 void QUARK_CORE_sgemm(Quark *quark, Quark_Task_Flags *task_flags,
494 PLASMA_enum transA, PLASMA_enum transB,
495 int m,
int n,
int k,
int nb,
496 float alpha,
const float *A,
int lda,
497 const float *B,
int ldb,
498 float beta,
float *C,
int ldc);
499 void QUARK_CORE_sgemm2( Quark *quark, Quark_Task_Flags *task_flags,
500 PLASMA_enum transA, PLASMA_enum transB,
501 int m,
int n,
int k,
int nb,
502 float alpha,
const float *A,
int lda,
503 const float *B,
int ldb,
504 float beta,
float *C,
int ldc);
505 void QUARK_CORE_sgemm_f2(Quark *quark, Quark_Task_Flags *task_flags,
506 PLASMA_enum transA, PLASMA_enum transB,
507 int m,
int n,
int k,
int nb,
508 float alpha,
const float *A,
int lda,
509 const float *B,
int ldb,
510 float beta,
float *C,
int ldc,
511 float *fake1,
int szefake1,
int flag1,
512 float *fake2,
int szefake2,
int flag2);
513 void QUARK_CORE_sgemm_p2(Quark *quark, Quark_Task_Flags *task_flags,
514 PLASMA_enum transA, PLASMA_enum transB,
515 int m,
int n,
int k,
int nb,
516 float alpha,
const float *A,
int lda,
517 const float **B,
int ldb,
518 float beta,
float *C,
int ldc);
519 void QUARK_CORE_sgemm_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
520 PLASMA_enum transA, PLASMA_enum transB,
521 int m,
int n,
int k,
int nb,
522 float alpha,
const float *A,
int lda,
523 const float **B,
int ldb,
524 float beta,
float *C,
int ldc,
525 float *fake1,
int szefake1,
int flag1);
526 void QUARK_CORE_sgemm_p3(Quark *quark, Quark_Task_Flags *task_flags,
527 PLASMA_enum transA, PLASMA_enum transB,
528 int m,
int n,
int k,
int nb,
529 float alpha,
const float *A,
int lda,
530 const float *B,
int ldb,
531 float beta,
float **C,
int ldc);
532 void QUARK_CORE_sgemm_tile(Quark *quark, Quark_Task_Flags *task_flags,
533 PLASMA_enum transA, PLASMA_enum transB,
534 int m,
int n,
int k,
int nb,
535 const float *alpha,
const float *A,
int lda,
536 const float *B,
int ldb,
537 const float *beta,
float *C,
int ldc,
541 void QUARK_CORE_sgemv(Quark *quark, Quark_Task_Flags *task_flags,
542 PLASMA_enum trans,
int m,
int n,
543 float alpha,
const float *A,
int lda,
544 const float *x,
int incx,
545 float beta,
float *y,
int incy);
546 void QUARK_CORE_sgemv_tile(Quark *quark, Quark_Task_Flags *task_flags,
549 const float *alpha,
const float *A,
int lda,
550 const float *x,
int incx,
551 const float *beta,
float *y,
int incy,
555 void QUARK_CORE_sgeqp3_init( Quark *quark, Quark_Task_Flags *task_flags,
557 void QUARK_CORE_sgeqp3_larfg(Quark *quark, Quark_Task_Flags *task_flags,
559 float *tau,
float *beta );
560 void QUARK_CORE_sgeqp3_norms( Quark *quark, Quark_Task_Flags *task_flags,
561 PLASMA_desc A,
int ioff,
int joff,
float *norms1,
float *norms2 );
562 void QUARK_CORE_sgeqp3_pivot( Quark *quark, Quark_Task_Flags *task_flags,
565 int jj,
int k,
int *jpvt,
566 float *norms1,
float *norms2,
int *info );
567 void QUARK_CORE_sgeqp3_tntpiv(Quark *quark, Quark_Task_Flags *task_flags,
568 int m,
int n,
int nb,
572 PLASMA_bool check_info,
int iinfo);
573 void QUARK_CORE_sgeqp3_update( Quark *quark, Quark_Task_Flags *task_flags,
574 float *Ajj,
int lda1,
575 float *Ajk,
int lda2,
577 int joff,
int k,
int koff,
int nb,
578 float *norms1,
float *norms2,
int *info );
579 void QUARK_CORE_sgeqrt(Quark *quark, Quark_Task_Flags *task_flags,
580 int m,
int n,
int ib,
int nb,
583 void QUARK_CORE_sgessm(Quark *quark, Quark_Task_Flags *task_flags,
584 int m,
int n,
int k,
int ib,
int nb,
586 const float *L,
int ldl,
588 void QUARK_CORE_sgessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
589 int m,
int n,
const float *A,
int lda,
590 float *scale,
float *sumsq,
591 float *fake,
int szeF,
int paramF );
592 void QUARK_CORE_sgetrf(Quark *quark, Quark_Task_Flags *task_flags,
593 int m,
int n,
int nb,
597 PLASMA_bool check_info,
int iinfo);
598 void QUARK_CORE_sgetrf_incpiv(Quark *quark, Quark_Task_Flags *task_flags,
599 int m,
int n,
int ib,
int nb,
603 PLASMA_bool check_info,
int iinfo);
604 void QUARK_CORE_sgetrf_nopiv(Quark *quark, Quark_Task_Flags *task_flags,
605 int m,
int n,
int ib,
int nb,
609 void QUARK_CORE_sgetrf_reclap(Quark *quark, Quark_Task_Flags *task_flags,
610 int m,
int n,
int nb,
614 PLASMA_bool check_info,
int iinfo,
616 void QUARK_CORE_sgetrf_rectil(Quark *quark, Quark_Task_Flags *task_flags,
620 PLASMA_bool check_info,
int iinfo,
622 void QUARK_CORE_sgetrip(Quark *quark, Quark_Task_Flags *task_flags,
623 int m,
int n,
float *A,
int szeA);
624 void QUARK_CORE_sgetrip_f1(Quark *quark, Quark_Task_Flags *task_flags,
625 int m,
int n,
float *A,
int szeA,
626 float *fake,
int szeF,
int paramF);
627 void QUARK_CORE_sgetrip_f2(Quark *quark, Quark_Task_Flags *task_flags,
628 int m,
int n,
float *A,
int szeA,
629 float *fake1,
int szeF1,
int paramF1,
630 float *fake2,
int szeF2,
int paramF2);
631 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
632 PLASMA_enum side, PLASMA_enum uplo,
633 int m,
int n,
int nb,
634 float alpha,
const float *A,
int lda,
635 const float *B,
int ldb,
636 float beta,
float *C,
int ldc);
637 void QUARK_CORE_ssygst(Quark *quark, Quark_Task_Flags *task_flags,
638 int itype, PLASMA_enum uplo,
int N,
643 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
644 PLASMA_enum uplo, PLASMA_enum trans,
645 int n,
int k,
int nb,
646 float alpha,
const float *A,
int lda,
647 float beta,
float *C,
int ldc);
648 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
649 PLASMA_enum uplo, PLASMA_enum trans,
650 int n,
int k,
int nb,
651 float alpha,
const float *A,
int lda,
652 const float *B,
int LDB,
653 float beta,
float *C,
int ldc);
654 void QUARK_CORE_ssyrfb(Quark *quark, Quark_Task_Flags *task_flags,
656 int n,
int k,
int ib,
int nb,
657 const float *A,
int lda,
658 const float *T,
int ldt,
660 void QUARK_CORE_shessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
661 PLASMA_enum uplo,
int n,
const float *A,
int lda,
662 float *scale,
float *sumsq,
663 float *fake,
int szeF,
int paramF );
664 void QUARK_CORE_slacpy(Quark *quark, Quark_Task_Flags *task_flags,
665 PLASMA_enum uplo,
int m,
int n,
int mb,
666 const float *A,
int lda,
668 void QUARK_CORE_slacpy_f1(Quark *quark, Quark_Task_Flags *task_flags,
669 PLASMA_enum uplo,
int m,
int n,
int nb,
670 const float *A,
int lda,
672 float *fake1,
int szefake1,
int flag1);
673 void QUARK_CORE_slacpy_pivot(Quark *quark, Quark_Task_Flags *task_flags,
676 int k1,
int k2,
const int *ipiv,
677 int *rankin,
int *rankout,
680 void QUARK_CORE_slange(Quark *quark, Quark_Task_Flags *task_flags,
681 int norm,
int M,
int N,
682 const float *A,
int LDA,
int szeA,
683 int szeW,
float *result);
684 void QUARK_CORE_slange_f1(Quark *quark, Quark_Task_Flags *task_flags,
685 int norm,
int M,
int N,
686 const float *A,
int LDA,
int szeA,
687 int szeW,
float *result,
688 float *fake,
int szeF);
690 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
691 int norm, PLASMA_enum uplo,
int N,
692 const float *A,
int LDA,
int szeA,
693 int szeW,
float *result);
694 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
695 int norm, PLASMA_enum uplo,
int N,
696 const float *A,
int LDA,
int szeA,
697 int szeW,
float *result,
698 float *fake,
int szeF);
700 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
701 int norm, PLASMA_enum uplo,
int N,
702 const float *A,
int LDA,
int szeA,
703 int szeW,
float *result);
704 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
705 int norm, PLASMA_enum uplo,
int N,
706 const float *A,
int LDA,
int szeA,
707 int szeW,
float *result,
708 float *fake,
int szeF);
709 void QUARK_CORE_slantr(Quark *quark, Quark_Task_Flags *task_flags,
710 PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag,
int M,
int N,
711 const float *A,
int LDA,
int szeA,
712 int szeW,
float *result);
713 void QUARK_CORE_slantr_f1(Quark *quark, Quark_Task_Flags *task_flags,
714 PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag,
int M,
int N,
715 const float *A,
int LDA,
int szeA,
716 int szeW,
float *result,
717 float *fake,
int szeF);
718 void QUARK_CORE_slaset(Quark *quark, Quark_Task_Flags *task_flags,
719 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
720 float beta,
float *tileA,
int ldtilea);
721 void QUARK_CORE_slaset2(Quark *quark, Quark_Task_Flags *task_flags,
722 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
723 float *tileA,
int ldtilea);
724 void QUARK_CORE_slaswp(Quark *quark, Quark_Task_Flags *task_flags,
725 int n,
float *A,
int lda,
726 int i1,
int i2,
const int *ipiv,
int inc);
727 void QUARK_CORE_slaswp_f2(Quark *quark, Quark_Task_Flags *task_flags,
728 int n,
float *A,
int lda,
729 int i1,
int i2,
const int *ipiv,
int inc,
730 float *fake1,
int szefake1,
int flag1,
731 float *fake2,
int szefake2,
int flag2);
732 void QUARK_CORE_slaswp_ontile(Quark *quark, Quark_Task_Flags *task_flags,
734 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
735 void QUARK_CORE_slaswp_ontile_f2(Quark *quark, Quark_Task_Flags *task_flags,
737 int i1,
int i2,
const int *ipiv,
int inc,
738 float *fake1,
int szefake1,
int flag1,
739 float *fake2,
int szefake2,
int flag2);
740 void QUARK_CORE_slaswpc_ontile(Quark *quark, Quark_Task_Flags *task_flags,
742 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
743 void QUARK_CORE_slatro(Quark *quark, Quark_Task_Flags *task_flags,
744 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
745 const float *A,
int lda,
747 void QUARK_CORE_slatro_f1(Quark *quark, Quark_Task_Flags *task_flags,
748 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
749 const float *A,
int lda,
751 float *fake1,
int szefake1,
int flag1);
752 void QUARK_CORE_slauum(Quark *quark, Quark_Task_Flags *task_flags,
753 PLASMA_enum uplo,
int n,
int nb,
755 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
756 float bump,
int m,
int n,
float *A,
int lda,
757 int bigM,
int m0,
int n0,
unsigned long long int seed );
758 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
759 float bump,
int m,
int n,
float *A,
int lda,
760 int bigM,
int m0,
int n0,
unsigned long long int seed );
761 void QUARK_CORE_splrnt(Quark *quark, Quark_Task_Flags *task_flags,
762 int m,
int n,
float *A,
int lda,
763 int bigM,
int m0,
int n0,
unsigned long long int seed );
764 void QUARK_CORE_spltmg(Quark *quark, Quark_Task_Flags *task_flags,
765 PLASMA_enum mtxtype,
int m,
int n,
float *A,
int lda,
766 int gM,
int gN,
int m0,
int n0,
unsigned long long int seed );
767 void QUARK_CORE_spltmg_chebvand( Quark *quark, Quark_Task_Flags *task_flags,
768 int M,
int N,
float *A,
int LDA,
769 int gN,
int m0,
int n0,
771 void QUARK_CORE_spltmg_circul( Quark *quark, Quark_Task_Flags *task_flags,
772 int M,
int N,
float *A,
int LDA,
773 int gM,
int m0,
int n0,
775 void QUARK_CORE_spltmg_fiedler(Quark *quark, Quark_Task_Flags *task_flags,
777 const float *X,
int incX,
778 const float *Y,
int incY,
780 void QUARK_CORE_spltmg_hankel( Quark *quark, Quark_Task_Flags *task_flags,
781 PLASMA_enum uplo,
int M,
int N,
float *A,
int LDA,
782 int m0,
int n0,
int nb,
785 void QUARK_CORE_spltmg_toeppd1(Quark *quark, Quark_Task_Flags *task_flags,
786 int gM,
int m0,
int M,
788 unsigned long long int seed);
789 void QUARK_CORE_spltmg_toeppd2(Quark *quark, Quark_Task_Flags *task_flags,
790 int M,
int N,
int K,
int m0,
int n0,
793 void QUARK_CORE_spotrf(Quark *quark, Quark_Task_Flags *task_flags,
794 PLASMA_enum uplo,
int n,
int nb,
798 void QUARK_CORE_ssetvar(Quark *quark, Quark_Task_Flags *task_flags,
799 const float *alpha,
float *x,
801 void QUARK_CORE_sshift( Quark *quark, Quark_Task_Flags *task_flags,
802 int s,
int m,
int n,
int L,
804 void QUARK_CORE_sshiftw(Quark *quark, Quark_Task_Flags *task_flags,
805 int s,
int cl,
int m,
int n,
int L,
807 void QUARK_CORE_sssssm(Quark *quark, Quark_Task_Flags *task_flags,
808 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
811 const float *L1,
int ldl1,
812 const float *L2,
int ldl2,
814 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
815 PLASMA_enum side, PLASMA_enum uplo,
816 int m,
int n,
int nb,
817 float alpha,
const float *A,
int lda,
818 const float *B,
int ldb,
819 float beta,
float *C,
int ldc);
820 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
821 PLASMA_enum uplo, PLASMA_enum trans,
822 int n,
int k,
int nb,
823 float alpha,
const float *A,
int lda,
824 float beta,
float *C,
int ldc);
825 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
826 PLASMA_enum uplo, PLASMA_enum trans,
827 int n,
int k,
int nb,
828 float alpha,
const float *A,
int lda,
829 const float *B,
int LDB,
830 float beta,
float *C,
int ldc);
831 void QUARK_CORE_ssyssq_f1( Quark *quark, Quark_Task_Flags *task_flags,
832 PLASMA_enum uplo,
int n,
const float *A,
int lda,
833 float *scale,
float *sumsq,
834 float *fake,
int szeF,
int paramF );
835 void QUARK_CORE_sswpab(Quark *quark, Quark_Task_Flags *task_flags,
836 int i,
int n1,
int n2,
838 void QUARK_CORE_sswptr_ontile(Quark *quark, Quark_Task_Flags *task_flags,
840 int i1,
int i2,
const int *ipiv,
int inc,
841 const float *Akk,
int ldak);
842 void QUARK_CORE_strasm(Quark *quark, Quark_Task_Flags *task_flags,
843 PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag,
int m,
int n,
844 const float *A,
int lda,
int szeA,
845 float *work,
int szeW);
846 void QUARK_CORE_strasm_f1(Quark *quark, Quark_Task_Flags *task_flags,
847 PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag,
int m,
int n,
848 const float *A,
int lda,
int szeA,
849 float *work,
int szeW,
850 float *fake,
int szeF);
851 void QUARK_CORE_strdalg1(Quark *quark, Quark_Task_Flags *task_flags,
858 int Vblksiz,
int wantz,
859 int i,
int sweepid,
int m,
int grsiz,
860 int *PCOL,
int *ACOL,
int *MCOL);
861 void QUARK_CORE_strmm(Quark *quark, Quark_Task_Flags *task_flags,
862 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
863 int m,
int n,
int nb,
864 float alpha,
const float *A,
int lda,
866 void QUARK_CORE_strmm_p2(Quark *quark, Quark_Task_Flags *task_flags,
867 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
868 int m,
int n,
int nb,
869 float alpha,
const float *A,
int lda,
871 void QUARK_CORE_strsm(Quark *quark, Quark_Task_Flags *task_flags,
872 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
873 int m,
int n,
int nb,
874 float alpha,
const float *A,
int lda,
876 void QUARK_CORE_strssq_f1( Quark *quark, Quark_Task_Flags *task_flags,
877 PLASMA_enum uplo, PLASMA_enum diag,
878 int m,
int n,
const float *A,
int lda,
879 float *scale,
float *sumsq,
880 float *fake,
int szeF,
int paramF );
881 void QUARK_CORE_strtri(Quark *quark, Quark_Task_Flags *task_flags,
882 PLASMA_enum uplo, PLASMA_enum diag,
int n,
int nb,
886 void QUARK_CORE_stslqt(Quark *quark, Quark_Task_Flags *task_flags,
887 int m,
int n,
int ib,
int nb,
891 void QUARK_CORE_stsmlq(Quark *quark, Quark_Task_Flags *task_flags,
892 PLASMA_enum side, PLASMA_enum trans,
893 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
896 const float *V,
int ldv,
897 const float *T,
int ldt);
898 void QUARK_CORE_stsmlq_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
899 PLASMA_enum side, PLASMA_enum trans,
900 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
903 const float *V,
int ldv,
904 const float *T,
int ldt);
905 void QUARK_CORE_stsmlq_corner(Quark *quark, Quark_Task_Flags *task_flags,
906 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
910 const float *V,
int ldv,
911 const float *T,
int ldt);
912 void QUARK_CORE_stsmqr(Quark *quark, Quark_Task_Flags *task_flags,
913 PLASMA_enum side, PLASMA_enum trans,
914 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
917 const float *V,
int ldv,
918 const float *T,
int ldt);
919 void QUARK_CORE_stsmqr_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
920 PLASMA_enum side, PLASMA_enum trans,
921 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
924 const float *V,
int ldv,
925 const float *T,
int ldt);
926 void QUARK_CORE_stsmqr_corner(Quark *quark, Quark_Task_Flags *task_flags,
927 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
931 const float *V,
int ldv,
932 const float *T,
int ldt);
933 void QUARK_CORE_stsqrt(Quark *quark, Quark_Task_Flags *task_flags,
934 int m,
int n,
int ib,
int nb,
938 void QUARK_CORE_ststrf(Quark *quark, Quark_Task_Flags *task_flags,
939 int m,
int n,
int ib,
int nb,
945 PLASMA_bool check_info,
int iinfo);
946 void QUARK_CORE_sttmqr(Quark *quark, Quark_Task_Flags *task_flags,
947 PLASMA_enum side, PLASMA_enum trans,
948 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
951 const float *V,
int ldv,
952 const float *T,
int ldt);
953 void QUARK_CORE_sttqrt(Quark *quark, Quark_Task_Flags *task_flags,
954 int m,
int n,
int ib,
int nb,
958 void QUARK_CORE_sttmlq(Quark *quark, Quark_Task_Flags *task_flags,
959 PLASMA_enum side, PLASMA_enum trans,
960 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
963 const float *V,
int ldv,
964 const float *T,
int ldt);
965 void QUARK_CORE_sttlqt(Quark *quark, Quark_Task_Flags *task_flags,
966 int m,
int n,
int ib,
int nb,
970 void QUARK_CORE_spamm(Quark *quark, Quark_Task_Flags *task_flags,
971 int op, PLASMA_enum side, PLASMA_enum storev,
972 int m,
int n,
int k,
int l,
973 const float *A1,
int lda1,
975 const float *V,
int ldv,
977 void QUARK_CORE_splssq( Quark *quark, Quark_Task_Flags *task_flags,
978 int m,
const float *A,
float *result );
979 void QUARK_CORE_sormlq(Quark *quark, Quark_Task_Flags *task_flags,
980 PLASMA_enum side, PLASMA_enum trans,
981 int m,
int n,
int ib,
int nb,
int k,
982 const float *A,
int lda,
983 const float *T,
int ldt,
985 void QUARK_CORE_sormqr(Quark *quark, Quark_Task_Flags *task_flags,
986 PLASMA_enum side, PLASMA_enum trans,
987 int m,
int n,
int k,
int ib,
int nb,
988 const float *A,
int lda,
989 const float *T,
int ldt,
995 void CORE_sasum_quark(Quark *quark);
996 void CORE_sasum_f1_quark(Quark *quark);
997 void CORE_sgeadd_quark(Quark *quark);
998 void CORE_sbrdalg1_quark(Quark *quark);
999 void CORE_sgelqt_quark(Quark *quark);
1000 void CORE_sgemm_quark(Quark *quark);
1001 void CORE_sgemm_tile_quark(Quark *quark);
1002 void CORE_sgemv_quark(Quark *quark);
1003 void CORE_sgemv_tile_quark(Quark *quark);
1004 void CORE_sgeqp3_init_quark(Quark *quark);
1005 void CORE_sgeqp3_larfg_quark(Quark *quark);
1006 void CORE_sgeqp3_norms_quark(Quark *quark);
1007 void CORE_sgeqp3_pivot_quark(Quark *quark);
1008 void CORE_sgeqp3_tntpiv_quark(Quark *quark);
1009 void CORE_sgeqp3_update_quark(Quark *quark);
1010 void CORE_sgeqrt_quark(Quark *quark);
1011 void CORE_sgessm_quark(Quark *quark);
1012 void CORE_sgessq_quark(Quark *quark);
1013 void CORE_sgessq_f1_quark(Quark *quark);
1014 void CORE_sgetrf_quark(Quark *quark);
1015 void CORE_sgetrf_incpiv_quark(Quark *quark);
1016 void CORE_sgetrf_nopiv_quark(Quark* quark);
1017 void CORE_sgetrf_reclap_quark(Quark *quark);
1018 void CORE_sgetrf_rectil_quark(Quark* quark);
1019 void CORE_sgetrip_quark(Quark *quark);
1020 void CORE_sgetrip_f1_quark(Quark *quark);
1021 void CORE_sgetrip_f2_quark(Quark *quark);
1023 void CORE_ssymm_quark(Quark *quark);
1024 void CORE_ssyrk_quark(Quark *quark);
1025 void CORE_ssyr2k_quark(Quark *quark);
1027 void CORE_ssygst_quark(Quark *quark);
1028 void CORE_ssyrfb_quark(Quark *quark);
1029 void CORE_shessq_quark(Quark *quark);
1030 void CORE_shessq_f1_quark(Quark *quark);
1031 void CORE_slacpy_quark(Quark *quark);
1032 void CORE_slacpy_f1_quark(Quark *quark);
1033 void CORE_slacpy_pivot_quark(Quark *quark);
1034 void CORE_slatro_quark(Quark *quark);
1035 void CORE_slatro_f1_quark(Quark *quark);
1036 void CORE_slange_quark(Quark *quark);
1037 void CORE_slange_f1_quark(Quark *quark);
1039 void CORE_slansy_quark(Quark *quark);
1040 void CORE_slansy_f1_quark(Quark *quark);
1042 void CORE_slansy_quark(Quark *quark);
1043 void CORE_slansy_f1_quark(Quark *quark);
1044 void CORE_slaset_quark(Quark *quark);
1045 void CORE_slaset2_quark(Quark *quark);
1046 void CORE_slatro_quark(Quark *quark);
1047 void CORE_slauum_quark(Quark *quark);
1048 void CORE_spamm_quark(Quark *quark);
1049 void CORE_splgsy_quark(Quark *quark);
1050 void CORE_splgsy_quark(Quark *quark);
1051 void CORE_splrnt_quark(Quark *quark);
1052 void CORE_spltmg_quark(Quark *quark);
1053 void CORE_splssq_quark(Quark *quark);
1054 void CORE_spotrf_quark(Quark *quark);
1055 void CORE_ssetvar_quark(Quark *quark);
1056 void CORE_sshift_quark(Quark *quark);
1057 void CORE_sshiftw_quark(Quark *quark);
1058 void CORE_sssssm_quark(Quark *quark);
1059 void CORE_ssymm_quark(Quark *quark);
1060 void CORE_ssyrk_quark(Quark *quark);
1061 void CORE_ssyr2k_quark(Quark *quark);
1062 void CORE_ssyssq_quark(Quark *quark);
1063 void CORE_ssyssq_f1_quark(Quark *quark);
1064 void CORE_sswpab_quark(Quark *quark);
1065 void CORE_sswptr_ontile_quark(Quark *quark);
1066 void CORE_strdalg1_quark(Quark *quark);
1067 void CORE_strmm_quark(Quark *quark);
1068 void CORE_strsm_quark(Quark *quark);
1069 void CORE_strtri_quark(Quark *quark);
1070 void CORE_stslqt_quark(Quark *quark);
1071 void CORE_stsmlq_quark(Quark *quark);
1072 void CORE_stsmlq_sytra1_quark(Quark *quark);
1073 void CORE_stsmlq_corner_quark(Quark *quark);
1074 void CORE_stsmqr_quark(Quark *quark);
1075 void CORE_stsmqr_sytra1_quark(Quark *quark);
1076 void CORE_stsmqr_corner_quark(Quark *quark);
1077 void CORE_stsqrt_quark(Quark *quark);
1078 void CORE_ststrf_quark(Quark *quark);
1079 void CORE_sttmqr_quark(Quark *quark);
1080 void CORE_sttqrt_quark(Quark *quark);
1081 void CORE_sttmlq_quark(Quark *quark);
1082 void CORE_sttlqt_quark(Quark *quark);
1083 void CORE_sormlq_quark(Quark *quark);
1084 void CORE_sormqr_quark(Quark *quark);
1085 void CORE_slaswp_quark(Quark* quark);
1086 void CORE_slaswp_f2_quark(Quark* quark);
1087 void CORE_slaswp_ontile_quark(Quark *quark);
1088 void CORE_slaswp_ontile_f2_quark(Quark *quark);
1089 void CORE_slaswpc_ontile_quark(Quark *quark);
1090 void CORE_strmm_p2_quark(Quark* quark);
1091 void CORE_sgemm_f2_quark(Quark* quark);
1092 void CORE_sgemm_p2_quark(Quark* quark);
1093 void CORE_sgemm_p2f1_quark(Quark* quark);
1094 void CORE_sgemm_p3_quark(Quark* quark);