18 #ifndef _PLASMA_CORE_SBLAS_H_
19 #define _PLASMA_CORE_SBLAS_H_
32 void CORE_sasum(
int storev, PLASMA_enum uplo,
int M,
int N,
33 const float *A,
int lda,
float *work);
43 int Vblksiz,
int wantz,
44 int i,
int sweepid,
int m,
int grsiz,
48 int st,
int ed,
int eltsize);
51 int st,
int ed,
int eltsize);
54 int st,
int ed,
int eltsize);
57 const float *A,
int LDA,
65 void CORE_sgemm(PLASMA_enum transA, PLASMA_enum transB,
67 float alpha,
const float *A,
int LDA,
68 const float *B,
int LDB,
69 float beta,
float *C,
int LDC);
70 void CORE_sgemv(PLASMA_enum trans,
int M,
int N,
71 float alpha,
const float *A,
int LDA,
72 const float *x,
int incx,
73 float beta,
float *y,
int incy);
76 float *tau,
float *beta );
79 int jj,
int k,
int *jpvt,
80 float *norms1,
float *norms2,
int *info );
81 int CORE_sgeqp3_tntpiv(
int m,
int n,
83 int *IPIV,
float *tau,
87 const float *Fk,
int ldf,
88 int joff,
int k,
int koff,
int nb,
89 float *norms1,
float *norms2,
94 float *TAU,
float *WORK);
97 const float *L,
int LDL,
99 int CORE_sgessq(
int M,
int N,
100 const float *A,
int LDA,
101 float *scale,
float *sumsq);
106 int *IPIV,
int *INFO);
109 int *IPIV,
int *INFO);
114 int *IPIV,
int *info);
122 int st,
int ed,
int eltsize);
125 int st,
int ed,
int eltsize);
128 int st,
int ed,
int eltsize);
131 float *V,
float *TAU,
132 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
136 float *V,
float *TAU,
137 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
141 const float *V,
const float *TAU,
142 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
146 float *VQ,
float *TAUQ,
147 float *VP,
float *TAUP,
148 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
152 float *VQ,
float *TAUQ,
153 float *VP,
float *TAUP,
154 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
158 float *VQ,
float *TAUQ,
159 float *VP,
float *TAUP,
160 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
162 void CORE_ssygst(
int itype, PLASMA_enum uplo,
int N,
164 float *B,
int LDB,
int *INFO);
166 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
168 float alpha,
const float *A,
int LDA,
169 const float *B,
int LDB,
170 float beta,
float *C,
int LDC);
171 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
173 float alpha,
const float *A,
int LDA,
174 float beta,
float *C,
int LDC);
175 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
177 float alpha,
const float *A,
int LDA,
178 const float *B,
int LDB,
179 float beta,
float *C,
int LDC);
180 int CORE_shessq(PLASMA_enum uplo,
int N,
181 const float *A,
int LDA,
182 float *scale,
float *sumsq);
184 int CORE_ssyrfb(PLASMA_enum uplo,
int N,
int K,
int IB,
int NB,
185 const float *A,
int LDA,
186 const float *T,
int LDT,
188 float *WORK,
int LDWORK);
190 const float *A,
int LDA,
194 int k1,
int k2,
const int *ipiv,
195 int *rankin,
int *rankout,
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);
206 void CORE_slansy(
int norm, PLASMA_enum uplo,
int N,
207 const float *A,
int LDA,
208 float *work,
float *normA);
209 void CORE_slantr(PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag,
211 const float *A,
int LDA,
212 float *work,
float *normA);
213 int CORE_slarfb_gemm(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
215 const float *V,
int LDV,
216 const float *T,
int LDT,
218 float *WORK,
int LDWORK);
223 float *C2,
int LDC2);
242 float alpha,
float *A,
int lda);
244 float alpha,
float beta,
245 float *tileA,
int ldtilea);
246 void CORE_slaset2(PLASMA_enum uplo,
int n1,
int n2,
float alpha,
247 float *tileA,
int ldtilea);
249 int I1,
int I2,
const int *IPIV,
int INC);
252 int CORE_slatro(PLASMA_enum uplo, PLASMA_enum trans,
254 const float *A,
int LDA,
256 void CORE_slauum(PLASMA_enum uplo,
int N,
float *A,
int LDA);
257 int CORE_spamm(
int op, PLASMA_enum side, PLASMA_enum storev,
258 int M,
int N,
int K,
int L,
259 const float *A1,
int LDA1,
261 const float *V,
int LDV,
263 int CORE_sparfb(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
264 int M1,
int N1,
int M2,
int N2,
int K,
int L,
267 const float *V,
int LDV,
268 const float *T,
int LDT,
269 float *WORK,
int LDWORK);
270 int CORE_spemv(PLASMA_enum trans, PLASMA_enum storev,
273 const float *A,
int LDA,
274 const float *X,
int INCX,
278 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
279 int bigM,
int m0,
int n0,
unsigned long long int seed );
280 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
281 int bigM,
int m0,
int n0,
unsigned long long int seed );
283 int bigM,
int m0,
int n0,
unsigned long long int seed );
284 int CORE_spltmg(PLASMA_enum mtxtype,
int m,
int n,
float *A,
int lda,
285 int gM,
int gN,
int m0,
int n0,
unsigned long long int seed );
287 int gN,
int m0,
int n0,
290 int gM,
int m0,
int n0,
294 const float *X,
int incX,
295 const float *Y,
int incY,
298 int m0,
int n0,
int nb,
302 unsigned long long int seed );
306 void CORE_spotrf(PLASMA_enum uplo,
int N,
float *A,
int LDA,
int *INFO);
308 void CORE_sshift(
int s,
int m,
int n,
int L,
312 int CORE_sssssm(
int M1,
int N1,
int M2,
int N2,
int K,
int IB,
315 const float *L1,
int LDL1,
316 const float *L2,
int LDL2,
321 float *WORK,
int LWORK,
323 float *RWORK,
int LRWORK,
325 int *IWORK,
int LIWORK);
330 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
332 float alpha,
const float *A,
int LDA,
333 const float *B,
int LDB,
334 float beta,
float *C,
int LDC);
335 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
337 float alpha,
const float *A,
int LDA,
338 float beta,
float *C,
int LDC);
339 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
341 float alpha,
const float *A,
int LDA,
342 const float *B,
int LDB,
343 float beta,
float *C,
int LDC);
344 int CORE_ssyssq(PLASMA_enum uplo,
int N,
345 const float *A,
int LDA,
346 float *scale,
float *sumsq);
348 float *A,
float *work);
350 const float *Akk,
int ldak);
351 int CORE_stradd(PLASMA_enum uplo, PLASMA_enum trans,
int M,
int N,
353 const float *A,
int LDA,
356 void CORE_strasm(PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag,
357 int M,
int N,
const float *A,
int lda,
float *work);
364 int Vblksiz,
int wantz,
365 int i,
int sweepid,
int m,
int grsiz,
367 void CORE_strmm(PLASMA_enum side, PLASMA_enum uplo,
368 PLASMA_enum transA, PLASMA_enum diag,
370 float alpha,
const float *A,
int LDA,
372 void CORE_strsm(PLASMA_enum side, PLASMA_enum uplo,
373 PLASMA_enum transA, PLASMA_enum diag,
375 float alpha,
const float *A,
int LDA,
377 int CORE_strssq(PLASMA_enum uplo, PLASMA_enum diag,
int M,
int N,
378 const float *A,
int LDA,
379 float *scale,
float *sumsq);
380 void CORE_strtri(PLASMA_enum uplo, PLASMA_enum diag,
int N,
381 float *A,
int LDA,
int *info);
386 float *TAU,
float *WORK);
387 int CORE_stsmlq(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);
410 int CORE_stsmqr(PLASMA_enum side, PLASMA_enum trans,
411 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
414 const float *V,
int LDV,
415 const float *T,
int LDT,
416 float *WORK,
int LDWORK);
418 int k,
int ib,
int nb,
422 const float *V,
int ldv,
423 const float *T,
int ldt,
424 float *WORK,
int ldwork);
426 int m1,
int n1,
int m2,
int n2,
430 const float *V,
int ldv,
431 const float *T,
int ldt,
432 float *WORK,
int ldwork);
437 float *TAU,
float *WORK);
442 int *IPIV,
float *WORK,
443 int LDWORK,
int *INFO);
444 int CORE_sttmqr(PLASMA_enum side, PLASMA_enum trans,
445 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
448 const float *V,
int LDV,
449 const float *T,
int LDT,
450 float *WORK,
int LDWORK);
457 int CORE_sttmlq(PLASMA_enum side, PLASMA_enum trans,
458 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
461 const float *V,
int LDV,
462 const float *T,
int LDT,
463 float *WORK,
int LDWORK);
470 int CORE_sormlq(PLASMA_enum side, PLASMA_enum trans,
471 int M,
int N,
int IB,
int K,
472 const float *V,
int LDV,
473 const float *T,
int LDT,
475 float *WORK,
int LDWORK);
476 int CORE_sormqr(PLASMA_enum side, PLASMA_enum trans,
477 int M,
int N,
int K,
int IB,
478 const float *V,
int LDV,
479 const float *T,
int LDT,
481 float *WORK,
int LDWORK);
485 float *beta,
float *D,
float *Q,
int LDQ,
486 float *Z,
float *DLAMBDA,
float *W,
487 int *INDX,
int *INDXC,
int *INDXP,
int *INDXQ,
490 const float *Q,
int LDQ,
float *Q2,
493 float *Q,
int LDQ,
const float *Q2,
496 float *D,
float beta,
498 const float *D0,
const float *Z,
500 int start,
int end );
502 const float *Q,
int LDQ,
503 const float *DLAMBDA,
float *W,
507 const float *Q,
int LDQ,
508 const float *Wred,
float *W);
510 float *Q,
int LDQ,
float *W,
float *S,
515 int il_nondef,
int iu_nondef,
516 float *Q,
int ldq,
float *Q2,
517 const int *ctot,
float *WORK,
int start,
int end);
519 void CORE_sswap(
int m,
int n,
float *Q,
int ldq,
520 const float *work,
const int *perm,
522 int CORE_slascl(PLASMA_enum type,
int kl,
int ku,
float cfrom,
float cto,
523 int m,
int n,
float *A,
int lda);
525 int CORE_slag2c(
int m,
int n,
const float *Q,
int LDQ,
530 void CORE_slaed3_freebigwork(
int oper,
float **WORK);
532 float *D,
const float *E);
533 int CORE_slapst(PLASMA_enum type,
int n,
534 const float *D,
int *INDX);
541 void QUARK_CORE_sasum(Quark *quark, Quark_Task_Flags *task_flags,
542 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
543 const float *A,
int lda,
int szeA,
544 float *work,
int szeW);
545 void QUARK_CORE_sasum_f1(Quark *quark, Quark_Task_Flags *task_flags,
546 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
547 const float *A,
int lda,
int szeA,
548 float *work,
int szeW,
549 float *fake,
int szeF);
550 void QUARK_CORE_sgeadd(Quark *quark, Quark_Task_Flags *task_flags,
551 PLASMA_enum trans,
int m,
int n,
int nb,
553 const float *A,
int lda,
556 void QUARK_CORE_sbrdalg1(Quark *quark, Quark_Task_Flags *task_flags,
565 int Vblksiz,
int wantz,
566 int i,
int sweepid,
int m,
int grsiz,
567 int *PCOL,
int *ACOL,
int *MCOL);
568 void QUARK_CORE_sgelqt(Quark *quark, Quark_Task_Flags *task_flags,
569 int m,
int n,
int ib,
int nb,
572 void QUARK_CORE_sgemm(Quark *quark, Quark_Task_Flags *task_flags,
573 PLASMA_enum transA, PLASMA_enum transB,
574 int m,
int n,
int k,
int nb,
575 float alpha,
const float *A,
int lda,
576 const float *B,
int ldb,
577 float beta,
float *C,
int ldc);
578 void QUARK_CORE_sgemm2( Quark *quark, Quark_Task_Flags *task_flags,
579 PLASMA_enum transA, PLASMA_enum transB,
580 int m,
int n,
int k,
int nb,
581 float alpha,
const float *A,
int lda,
582 const float *B,
int ldb,
583 float beta,
float *C,
int ldc);
584 void QUARK_CORE_sgemm_f2(Quark *quark, Quark_Task_Flags *task_flags,
585 PLASMA_enum transA, PLASMA_enum transB,
586 int m,
int n,
int k,
int nb,
587 float alpha,
const float *A,
int lda,
588 const float *B,
int ldb,
589 float beta,
float *C,
int ldc,
590 float *fake1,
int szefake1,
int flag1,
591 float *fake2,
int szefake2,
int flag2);
592 void QUARK_CORE_sgemm_p2(Quark *quark, Quark_Task_Flags *task_flags,
593 PLASMA_enum transA, PLASMA_enum transB,
594 int m,
int n,
int k,
int nb,
595 float alpha,
const float *A,
int lda,
596 const float **B,
int ldb,
597 float beta,
float *C,
int ldc);
598 void QUARK_CORE_sgemm_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
599 PLASMA_enum transA, PLASMA_enum transB,
600 int m,
int n,
int k,
int nb,
601 float alpha,
const float *A,
int lda,
602 const float **B,
int ldb,
603 float beta,
float *C,
int ldc,
604 float *fake1,
int szefake1,
int flag1);
605 void QUARK_CORE_sgemm_p3(Quark *quark, Quark_Task_Flags *task_flags,
606 PLASMA_enum transA, PLASMA_enum transB,
607 int m,
int n,
int k,
int nb,
608 float alpha,
const float *A,
int lda,
609 const float *B,
int ldb,
610 float beta,
float **C,
int ldc);
611 void QUARK_CORE_sgemm_tile(Quark *quark, Quark_Task_Flags *task_flags,
612 PLASMA_enum transA, PLASMA_enum transB,
613 int m,
int n,
int k,
int nb,
614 const float *alpha,
const float *A,
int lda,
615 const float *B,
int ldb,
616 const float *beta,
float *C,
int ldc,
620 void QUARK_CORE_sgemv(Quark *quark, Quark_Task_Flags *task_flags,
621 PLASMA_enum trans,
int m,
int n,
622 float alpha,
const float *A,
int lda,
623 const float *x,
int incx,
624 float beta,
float *y,
int incy);
625 void QUARK_CORE_sgemv_tile(Quark *quark, Quark_Task_Flags *task_flags,
628 const float *alpha,
const float *A,
int lda,
629 const float *x,
int incx,
630 const float *beta,
float *y,
int incy,
634 void QUARK_CORE_sgeqp3_init( Quark *quark, Quark_Task_Flags *task_flags,
636 void QUARK_CORE_sgeqp3_larfg(Quark *quark, Quark_Task_Flags *task_flags,
638 float *tau,
float *beta );
639 void QUARK_CORE_sgeqp3_norms( Quark *quark, Quark_Task_Flags *task_flags,
640 PLASMA_desc A,
int ioff,
int joff,
float *norms1,
float *norms2 );
641 void QUARK_CORE_sgeqp3_pivot( Quark *quark, Quark_Task_Flags *task_flags,
644 int jj,
int k,
int *jpvt,
645 float *norms1,
float *norms2,
int *info );
646 void QUARK_CORE_sgeqp3_tntpiv(Quark *quark, Quark_Task_Flags *task_flags,
647 int m,
int n,
int nb,
651 PLASMA_bool check_info,
int iinfo);
652 void QUARK_CORE_sgeqp3_update( Quark *quark, Quark_Task_Flags *task_flags,
653 float *Ajj,
int lda1,
654 float *Ajk,
int lda2,
656 int joff,
int k,
int koff,
int nb,
657 float *norms1,
float *norms2,
int *info );
658 void QUARK_CORE_sgeqrt(Quark *quark, Quark_Task_Flags *task_flags,
659 int m,
int n,
int ib,
int nb,
662 void QUARK_CORE_sgessm(Quark *quark, Quark_Task_Flags *task_flags,
663 int m,
int n,
int k,
int ib,
int nb,
665 const float *L,
int ldl,
667 void QUARK_CORE_sgessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
668 int m,
int n,
const float *A,
int lda,
669 float *scale,
float *sumsq,
670 float *fake,
int szeF,
int paramF );
671 void QUARK_CORE_sgetrf(Quark *quark, Quark_Task_Flags *task_flags,
672 int m,
int n,
int nb,
676 PLASMA_bool check_info,
int iinfo);
677 void QUARK_CORE_sgetrf_incpiv(Quark *quark, Quark_Task_Flags *task_flags,
678 int m,
int n,
int ib,
int nb,
682 PLASMA_bool check_info,
int iinfo);
683 void QUARK_CORE_sgetrf_nopiv(Quark *quark, Quark_Task_Flags *task_flags,
684 int m,
int n,
int ib,
int nb,
688 void QUARK_CORE_sgetrf_reclap(Quark *quark, Quark_Task_Flags *task_flags,
693 PLASMA_bool check_info,
int iinfo,
695 void QUARK_CORE_sgetrf_rectil(Quark *quark, Quark_Task_Flags *task_flags,
700 PLASMA_bool check_info,
int iinfo,
702 void QUARK_CORE_sgetrip(Quark *quark, Quark_Task_Flags *task_flags,
703 int m,
int n,
float *A,
int szeA);
704 void QUARK_CORE_sgetrip_f1(Quark *quark, Quark_Task_Flags *task_flags,
705 int m,
int n,
float *A,
int szeA,
706 float *fake,
int szeF,
int paramF);
707 void QUARK_CORE_sgetrip_f2(Quark *quark, Quark_Task_Flags *task_flags,
708 int m,
int n,
float *A,
int szeA,
709 float *fake1,
int szeF1,
int paramF1,
710 float *fake2,
int szeF2,
int paramF2);
711 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
712 PLASMA_enum side, PLASMA_enum uplo,
713 int m,
int n,
int nb,
714 float alpha,
const float *A,
int lda,
715 const float *B,
int ldb,
716 float beta,
float *C,
int ldc);
717 void QUARK_CORE_ssygst(Quark *quark, Quark_Task_Flags *task_flags,
718 int itype, PLASMA_enum uplo,
int N,
723 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
724 PLASMA_enum uplo, PLASMA_enum trans,
725 int n,
int k,
int nb,
726 float alpha,
const float *A,
int lda,
727 float beta,
float *C,
int ldc);
728 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
729 PLASMA_enum uplo, PLASMA_enum trans,
730 int n,
int k,
int nb,
731 float alpha,
const float *A,
int lda,
732 const float *B,
int LDB,
733 float beta,
float *C,
int ldc);
734 void QUARK_CORE_ssyrfb(Quark *quark, Quark_Task_Flags *task_flags,
736 int n,
int k,
int ib,
int nb,
737 const float *A,
int lda,
738 const float *T,
int ldt,
740 void QUARK_CORE_shessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
741 PLASMA_enum uplo,
int n,
const float *A,
int lda,
742 float *scale,
float *sumsq,
743 float *fake,
int szeF,
int paramF );
744 void QUARK_CORE_slacpy(Quark *quark, Quark_Task_Flags *task_flags,
745 PLASMA_enum uplo,
int m,
int n,
int mb,
746 const float *A,
int lda,
748 void QUARK_CORE_slacpy_f1(Quark *quark, Quark_Task_Flags *task_flags,
749 PLASMA_enum uplo,
int m,
int n,
int nb,
750 const float *A,
int lda,
752 float *fake1,
int szefake1,
int flag1);
753 void QUARK_CORE_slacpy_pivot(Quark *quark, Quark_Task_Flags *task_flags,
756 int k1,
int k2,
const int *ipiv,
757 int *rankin,
int *rankout,
760 void QUARK_CORE_slange(Quark *quark, Quark_Task_Flags *task_flags,
761 int norm,
int M,
int N,
762 const float *A,
int LDA,
int szeA,
763 int szeW,
float *result);
764 void QUARK_CORE_slange_f1(Quark *quark, Quark_Task_Flags *task_flags,
765 int norm,
int M,
int N,
766 const float *A,
int LDA,
int szeA,
767 int szeW,
float *result,
768 float *fake,
int szeF);
770 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
771 int norm, PLASMA_enum uplo,
int N,
772 const float *A,
int LDA,
int szeA,
773 int szeW,
float *result);
774 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
775 int norm, PLASMA_enum uplo,
int N,
776 const float *A,
int LDA,
int szeA,
777 int szeW,
float *result,
778 float *fake,
int szeF);
780 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
781 int norm, PLASMA_enum uplo,
int N,
782 const float *A,
int LDA,
int szeA,
783 int szeW,
float *result);
784 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
785 int norm, PLASMA_enum uplo,
int N,
786 const float *A,
int LDA,
int szeA,
787 int szeW,
float *result,
788 float *fake,
int szeF);
789 void QUARK_CORE_slantr(Quark *quark, Quark_Task_Flags *task_flags,
790 PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag,
int M,
int N,
791 const float *A,
int LDA,
int szeA,
792 int szeW,
float *result);
793 void QUARK_CORE_slantr_f1(Quark *quark, Quark_Task_Flags *task_flags,
794 PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag,
int M,
int N,
795 const float *A,
int LDA,
int szeA,
796 int szeW,
float *result,
797 float *fake,
int szeF);
798 void QUARK_CORE_slascal(Quark *quark, Quark_Task_Flags *task_flags,
799 PLASMA_enum uplo,
int m,
int n,
int nb,
800 float alpha,
float *A,
int lda);
801 void QUARK_CORE_slaset(Quark *quark, Quark_Task_Flags *task_flags,
802 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
803 float beta,
float *tileA,
int ldtilea);
804 void QUARK_CORE_slaset2(Quark *quark, Quark_Task_Flags *task_flags,
805 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
806 float *tileA,
int ldtilea);
807 void QUARK_CORE_slaswp(Quark *quark, Quark_Task_Flags *task_flags,
808 int n,
float *A,
int lda,
809 int i1,
int i2,
const int *ipiv,
int inc);
810 void QUARK_CORE_slaswp_f2(Quark *quark, Quark_Task_Flags *task_flags,
811 int n,
float *A,
int lda,
812 int i1,
int i2,
const int *ipiv,
int inc,
813 float *fake1,
int szefake1,
int flag1,
814 float *fake2,
int szefake2,
int flag2);
815 void QUARK_CORE_slaswp_ontile(Quark *quark, Quark_Task_Flags *task_flags,
817 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
818 void QUARK_CORE_slaswp_ontile_f2(Quark *quark, Quark_Task_Flags *task_flags,
820 int i1,
int i2,
const int *ipiv,
int inc,
821 float *fake1,
int szefake1,
int flag1,
822 float *fake2,
int szefake2,
int flag2);
823 void QUARK_CORE_slaswpc_ontile(Quark *quark, Quark_Task_Flags *task_flags,
825 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
826 void QUARK_CORE_slatro(Quark *quark, Quark_Task_Flags *task_flags,
827 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
828 const float *A,
int lda,
830 void QUARK_CORE_slatro_f1(Quark *quark, Quark_Task_Flags *task_flags,
831 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
832 const float *A,
int lda,
834 float *fake1,
int szefake1,
int flag1);
835 void QUARK_CORE_slauum(Quark *quark, Quark_Task_Flags *task_flags,
836 PLASMA_enum uplo,
int n,
int nb,
838 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
839 float bump,
int m,
int n,
float *A,
int lda,
840 int bigM,
int m0,
int n0,
unsigned long long int seed );
841 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
842 float bump,
int m,
int n,
float *A,
int lda,
843 int bigM,
int m0,
int n0,
unsigned long long int seed );
844 void QUARK_CORE_splrnt(Quark *quark, Quark_Task_Flags *task_flags,
845 int m,
int n,
float *A,
int lda,
846 int bigM,
int m0,
int n0,
unsigned long long int seed );
847 void QUARK_CORE_spltmg(Quark *quark, Quark_Task_Flags *task_flags,
848 PLASMA_enum mtxtype,
int m,
int n,
float *A,
int lda,
849 int gM,
int gN,
int m0,
int n0,
unsigned long long int seed );
850 void QUARK_CORE_spltmg_chebvand( Quark *quark, Quark_Task_Flags *task_flags,
851 int M,
int N,
float *A,
int LDA,
852 int gN,
int m0,
int n0,
854 void QUARK_CORE_spltmg_circul( Quark *quark, Quark_Task_Flags *task_flags,
855 int M,
int N,
float *A,
int LDA,
856 int gM,
int m0,
int n0,
858 void QUARK_CORE_spltmg_fiedler(Quark *quark, Quark_Task_Flags *task_flags,
860 const float *X,
int incX,
861 const float *Y,
int incY,
863 void QUARK_CORE_spltmg_hankel( Quark *quark, Quark_Task_Flags *task_flags,
864 PLASMA_enum uplo,
int M,
int N,
float *A,
int LDA,
865 int m0,
int n0,
int nb,
868 void QUARK_CORE_spltmg_toeppd1(Quark *quark, Quark_Task_Flags *task_flags,
869 int gM,
int m0,
int M,
871 unsigned long long int seed);
872 void QUARK_CORE_spltmg_toeppd2(Quark *quark, Quark_Task_Flags *task_flags,
873 int M,
int N,
int K,
int m0,
int n0,
876 void QUARK_CORE_spotrf(Quark *quark, Quark_Task_Flags *task_flags,
877 PLASMA_enum uplo,
int n,
int nb,
881 void QUARK_CORE_ssetvar(Quark *quark, Quark_Task_Flags *task_flags,
882 const float *alpha,
float *x,
884 void QUARK_CORE_sshift( Quark *quark, Quark_Task_Flags *task_flags,
885 int s,
int m,
int n,
int L,
887 void QUARK_CORE_sshiftw(Quark *quark, Quark_Task_Flags *task_flags,
888 int s,
int cl,
int m,
int n,
int L,
890 void QUARK_CORE_sssssm(Quark *quark, Quark_Task_Flags *task_flags,
891 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
894 const float *L1,
int ldl1,
895 const float *L2,
int ldl2,
897 void QUARK_CORE_sstedc(Quark *quark, Quark_Task_Flags *task_flags,
898 PLASMA_enum compz,
int n,
901 void QUARK_CORE_sstedc_f2(Quark *quark, Quark_Task_Flags *task_flags,
902 PLASMA_enum compz,
int n,
905 void *fake1,
int szefake1,
int flag1,
906 void *fake2,
int szefake2,
int flag2);
907 void QUARK_CORE_ssteqr(Quark *quark, Quark_Task_Flags *task_flags,
908 PLASMA_enum compz,
int n,
911 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
912 PLASMA_enum side, PLASMA_enum uplo,
913 int m,
int n,
int nb,
914 float alpha,
const float *A,
int lda,
915 const float *B,
int ldb,
916 float beta,
float *C,
int ldc);
917 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
918 PLASMA_enum uplo, PLASMA_enum trans,
919 int n,
int k,
int nb,
920 float alpha,
const float *A,
int lda,
921 float beta,
float *C,
int ldc);
922 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
923 PLASMA_enum uplo, PLASMA_enum trans,
924 int n,
int k,
int nb,
925 float alpha,
const float *A,
int lda,
926 const float *B,
int LDB,
927 float beta,
float *C,
int ldc);
928 void QUARK_CORE_ssyssq_f1( Quark *quark, Quark_Task_Flags *task_flags,
929 PLASMA_enum uplo,
int n,
const float *A,
int lda,
930 float *scale,
float *sumsq,
931 float *fake,
int szeF,
int paramF );
932 void QUARK_CORE_sswpab(Quark *quark, Quark_Task_Flags *task_flags,
933 int i,
int n1,
int n2,
935 void QUARK_CORE_sswptr_ontile(Quark *quark, Quark_Task_Flags *task_flags,
937 int i1,
int i2,
const int *ipiv,
int inc,
938 const float *Akk,
int ldak);
939 void QUARK_CORE_stradd(Quark *quark, Quark_Task_Flags *task_flags,
940 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int nb,
942 const float *A,
int lda,
945 void QUARK_CORE_strasm(Quark *quark, Quark_Task_Flags *task_flags,
946 PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag,
int m,
int n,
947 const float *A,
int lda,
int szeA,
948 float *work,
int szeW);
949 void QUARK_CORE_strasm_f1(Quark *quark, Quark_Task_Flags *task_flags,
950 PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag,
int m,
int n,
951 const float *A,
int lda,
int szeA,
952 float *work,
int szeW,
953 float *fake,
int szeF);
954 void QUARK_CORE_strdalg1(Quark *quark, Quark_Task_Flags *task_flags,
961 int Vblksiz,
int wantz,
962 int i,
int sweepid,
int m,
int grsiz,
963 int *PCOL,
int *ACOL,
int *MCOL);
964 void QUARK_CORE_strmm(Quark *quark, Quark_Task_Flags *task_flags,
965 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
966 int m,
int n,
int nb,
967 float alpha,
const float *A,
int lda,
969 void QUARK_CORE_strmm_p2(Quark *quark, Quark_Task_Flags *task_flags,
970 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
971 int m,
int n,
int nb,
972 float alpha,
const float *A,
int lda,
974 void QUARK_CORE_strsm(Quark *quark, Quark_Task_Flags *task_flags,
975 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
976 int m,
int n,
int nb,
977 float alpha,
const float *A,
int lda,
979 void QUARK_CORE_strssq_f1( Quark *quark, Quark_Task_Flags *task_flags,
980 PLASMA_enum uplo, PLASMA_enum diag,
981 int m,
int n,
const float *A,
int lda,
982 float *scale,
float *sumsq,
983 float *fake,
int szeF,
int paramF );
984 void QUARK_CORE_strtri(Quark *quark, Quark_Task_Flags *task_flags,
985 PLASMA_enum uplo, PLASMA_enum diag,
int n,
int nb,
989 void QUARK_CORE_stslqt(Quark *quark, Quark_Task_Flags *task_flags,
990 int m,
int n,
int ib,
int nb,
994 void QUARK_CORE_stsmlq(Quark *quark, Quark_Task_Flags *task_flags,
995 PLASMA_enum side, PLASMA_enum trans,
996 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
999 const float *V,
int ldv,
1000 const float *T,
int ldt);
1001 void QUARK_CORE_stsmlq_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
1002 PLASMA_enum side, PLASMA_enum trans,
1003 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
1004 float *A1,
int lda1,
1005 float *A2,
int lda2,
1006 const float *V,
int ldv,
1007 const float *T,
int ldt);
1008 void QUARK_CORE_stsmlq_corner(Quark *quark, Quark_Task_Flags *task_flags,
1009 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
1010 float *A1,
int lda1,
1011 float *A2,
int lda2,
1012 float *A3,
int lda3,
1013 const float *V,
int ldv,
1014 const float *T,
int ldt);
1015 void QUARK_CORE_stsmqr(Quark *quark, Quark_Task_Flags *task_flags,
1016 PLASMA_enum side, PLASMA_enum trans,
1017 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
1018 float *A1,
int lda1,
1019 float *A2,
int lda2,
1020 const float *V,
int ldv,
1021 const float *T,
int ldt);
1022 void QUARK_CORE_stsmqr_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
1023 PLASMA_enum side, PLASMA_enum trans,
1024 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
1025 float *A1,
int lda1,
1026 float *A2,
int lda2,
1027 const float *V,
int ldv,
1028 const float *T,
int ldt);
1029 void QUARK_CORE_stsmqr_corner(Quark *quark, Quark_Task_Flags *task_flags,
1030 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
1031 float *A1,
int lda1,
1032 float *A2,
int lda2,
1033 float *A3,
int lda3,
1034 const float *V,
int ldv,
1035 const float *T,
int ldt);
1036 void QUARK_CORE_stsqrt(Quark *quark, Quark_Task_Flags *task_flags,
1037 int m,
int n,
int ib,
int nb,
1038 float *A1,
int lda1,
1039 float *A2,
int lda2,
1041 void QUARK_CORE_ststrf(Quark *quark, Quark_Task_Flags *task_flags,
1042 int m,
int n,
int ib,
int nb,
1048 PLASMA_bool check_info,
int iinfo);
1049 void QUARK_CORE_sttmqr(Quark *quark, Quark_Task_Flags *task_flags,
1050 PLASMA_enum side, PLASMA_enum trans,
1051 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
1052 float *A1,
int lda1,
1053 float *A2,
int lda2,
1054 const float *V,
int ldv,
1055 const float *T,
int ldt);
1056 void QUARK_CORE_sttqrt(Quark *quark, Quark_Task_Flags *task_flags,
1057 int m,
int n,
int ib,
int nb,
1058 float *A1,
int lda1,
1059 float *A2,
int lda2,
1061 void QUARK_CORE_sttmlq(Quark *quark, Quark_Task_Flags *task_flags,
1062 PLASMA_enum side, PLASMA_enum trans,
1063 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
1064 float *A1,
int lda1,
1065 float *A2,
int lda2,
1066 const float *V,
int ldv,
1067 const float *T,
int ldt);
1068 void QUARK_CORE_sttlqt(Quark *quark, Quark_Task_Flags *task_flags,
1069 int m,
int n,
int ib,
int nb,
1070 float *A1,
int lda1,
1071 float *A2,
int lda2,
1073 void QUARK_CORE_spamm(Quark *quark, Quark_Task_Flags *task_flags,
1074 int op, PLASMA_enum side, PLASMA_enum storev,
1075 int m,
int n,
int k,
int l,
1076 const float *A1,
int lda1,
1077 float *A2,
int lda2,
1078 const float *V,
int ldv,
1080 void QUARK_CORE_splssq( Quark *quark, Quark_Task_Flags *task_flags,
1081 int m,
const float *A,
float *result );
1082 void QUARK_CORE_sormlq(Quark *quark, Quark_Task_Flags *task_flags,
1083 PLASMA_enum side, PLASMA_enum trans,
1084 int m,
int n,
int ib,
int nb,
int k,
1085 const float *A,
int lda,
1086 const float *T,
int ldt,
1088 void QUARK_CORE_sormqr(Quark *quark, Quark_Task_Flags *task_flags,
1089 PLASMA_enum side, PLASMA_enum trans,
1090 int m,
int n,
int k,
int ib,
int nb,
1091 const float *A,
int lda,
1092 const float *T,
int ldt,
1096 void QUARK_CORE_slascl(Quark *quark, Quark_Task_Flags *task_flags,
1097 PLASMA_enum type,
int kl,
int ku,
float cfrom,
float cto,
1098 int m,
int n,
float *A,
int lda);
1099 void QUARK_CORE_slascl_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
1100 PLASMA_enum type,
int kl,
int ku,
float *cfrom,
float *cto,
1101 int m,
int n,
float *A,
int lda,
1102 void *fake,
int szefake,
int flag);
1103 void QUARK_CORE_slaed0_lascl( Quark *quark, Quark_Task_Flags *task_flags,
1104 int n,
float *scale,
float *D,
float *E);
1105 void QUARK_CORE_slaed0_betaapprox(Quark *quark, Quark_Task_Flags *task_flags,
1106 int subpbs,
const int *subpbs_info,
1107 float *D,
const float *E);
1110 void QUARK_CORE_slaed2_computeK(Quark *quark, Quark_Task_Flags *task_flags,
1111 int *K1,
int n,
int n1,
1112 float *beta,
float *D,
float *Q,
int LDQ,
1113 float *Z,
float *DLAMBDA,
float *W,
1114 int *INDX,
int *INDXC,
int *INDXP,
int *INDXQ,
1116 float **Qmerge,
int wsmode,
1119 void QUARK_CORE_slaed1_pipelined(Quark *quark, Quark_Task_Flags *task_flags,
1120 int n,
int n1,
const int *K,
1121 const int *INDX,
const int *ctot,
1122 float *D,
const float *beta,
1123 float *Q,
int LDQ,
float *Q2,
1124 const float *DLAMBDA,
const float *W,
float *Wred,
1125 int start,
int end);
1126 void QUARK_CORE_slaed2_compressq(Quark *quark, Quark_Task_Flags *task_flags,
1127 int n,
int n1,
int start,
int end,
1128 const int *INDX,
const int *ctot,
1129 const float *Q,
int LDQ,
1131 void QUARK_CORE_slaed4_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
1132 int n,
const int *K,
1133 float *D,
const float *beta,
1134 float **Q,
const int *LDQ,
1135 const float *DLAMBDA,
const float *W,
const int *INDX,
1138 void *fakeQ,
int flagfQ);
1139 void QUARK_CORE_slaed3_compW_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
1140 int n,
const int *K,
1141 float **Q,
const int *LDQ,
1142 const float *DLAMBDA,
float *W,
1145 void *fakeQ,
int flagfQ,
1146 void *fakeW,
int flagfW);
1148 void QUARK_CORE_slaed3_reduceW(Quark *quark, Quark_Task_Flags *task_flags,
1149 int n,
int n1,
const int *K,
int l,
1150 const float *Q,
int LDQ,
1151 const float *Wred,
float *W);
1152 void QUARK_CORE_slaed3_reduceW_p2(Quark *quark, Quark_Task_Flags *task_flags,
1153 int n,
int n1,
const int *K,
int l,
1154 float **Q,
const int *LDQ,
1155 const float *Wred,
float *W);
1157 void QUARK_CORE_slaed2_copydef(Quark *quark, Quark_Task_Flags *task_flags,
1158 int n,
int n1,
const int *K,
const int *ctot,
1159 float *Q,
int LDQ,
const float *Q2,
1160 int start,
int end);
1161 void QUARK_CORE_slaed3_computevectors(Quark *quark, Quark_Task_Flags *task_flags,
1162 int wsmode,
int n,
const int *K,
1163 const int *il_nondef,
const int *iu_nondef,
1164 float *Q,
int LDQ,
float *W,
const int *INDXC,
1165 float **WSglobal,
float **WSlocal,
1166 int start,
int end );
1167 void QUARK_CORE_slaed3_wscopy( Quark *quark, Quark_Task_Flags *task_flags,
1168 const int *K,
const int *il_nondef,
const int *iu_nondef,
1169 const float *Q,
int LDQ,
float **WORK,
1170 int start,
int end );
1171 void QUARK_CORE_slaed3_updatevectors(Quark *quark, Quark_Task_Flags *task_flags,
1172 int oper,
int wsmode,
int n,
int n1,
int *K,
1173 int *il_nondef,
int *iu_nondef,
1174 float *D,
float *Q,
int LDQ,
float *Q2,
1175 int *INDXQ,
int *COLTYP,
float **WORK,
1176 int start,
int end,
float **WORKDEP);
1177 void QUARK_CORE_slaed3_pipelined(Quark *quark, Quark_Task_Flags *task_flags,
1178 int n,
int n1,
int *K,
int *il_nondef,
int *iu_nondef,
1179 float *D,
float *Q,
int LDQ,
float *Q2,
1180 int *INDXC,
int *INDXQ,
int *COLTYP,
float *W,
1181 int start,
int end2);
1183 void QUARK_CORE_sDC_fakedep(Quark *quark, Quark_Task_Flags *task_flags,
1184 int nb_tasks,
int nb,
float *Q,
int LDQ,
float *W);
1187 void QUARK_CORE_sswap(Quark *quark, Quark_Task_Flags *task_flags,
1188 int m,
int n,
float *Q,
1189 int LDQ,
float *work,
1190 int *perm,
int begin,
int end);
1192 void QUARK_CORE_slag2c(Quark *quark, Quark_Task_Flags *task_flags,
1194 const float *Q,
int LDQ,
1197 void QUARK_CORE_slaed3_freebigwork(Quark *quark, Quark_Task_Flags *task_flags,
1198 int *K_bis,
int largework,
float **WORK);
1199 void QUARK_CORE_slaset_identity(Quark *quark, Quark_Task_Flags *task_flags,
1200 int n,
int start,
int size,
1206 void CORE_sasum_quark(Quark *quark);
1207 void CORE_sasum_f1_quark(Quark *quark);
1208 void CORE_sgeadd_quark(Quark *quark);
1209 void CORE_sbrdalg1_quark(Quark *quark);
1210 void CORE_sgelqt_quark(Quark *quark);
1211 void CORE_sgemm_quark(Quark *quark);
1212 void CORE_sgemm_tile_quark(Quark *quark);
1213 void CORE_sgemv_quark(Quark *quark);
1214 void CORE_sgemv_tile_quark(Quark *quark);
1215 void CORE_sgeqp3_init_quark(Quark *quark);
1216 void CORE_sgeqp3_larfg_quark(Quark *quark);
1217 void CORE_sgeqp3_norms_quark(Quark *quark);
1218 void CORE_sgeqp3_pivot_quark(Quark *quark);
1219 void CORE_sgeqp3_tntpiv_quark(Quark *quark);
1220 void CORE_sgeqp3_update_quark(Quark *quark);
1221 void CORE_sgeqrt_quark(Quark *quark);
1222 void CORE_sgessm_quark(Quark *quark);
1223 void CORE_sgessq_quark(Quark *quark);
1224 void CORE_sgessq_f1_quark(Quark *quark);
1225 void CORE_sgetrf_quark(Quark *quark);
1226 void CORE_sgetrf_incpiv_quark(Quark *quark);
1227 void CORE_sgetrf_nopiv_quark(Quark* quark);
1228 void CORE_sgetrf_reclap_quark(Quark *quark);
1229 void CORE_sgetrf_rectil_quark(Quark* quark);
1230 void CORE_sgetrip_quark(Quark *quark);
1231 void CORE_sgetrip_f1_quark(Quark *quark);
1232 void CORE_sgetrip_f2_quark(Quark *quark);
1234 void CORE_ssymm_quark(Quark *quark);
1235 void CORE_ssyrk_quark(Quark *quark);
1236 void CORE_ssyr2k_quark(Quark *quark);
1238 void CORE_ssygst_quark(Quark *quark);
1239 void CORE_ssyrfb_quark(Quark *quark);
1240 void CORE_shessq_quark(Quark *quark);
1241 void CORE_shessq_f1_quark(Quark *quark);
1242 void CORE_slacpy_quark(Quark *quark);
1243 void CORE_slacpy_f1_quark(Quark *quark);
1244 void CORE_slacpy_pivot_quark(Quark *quark);
1245 void CORE_slatro_quark(Quark *quark);
1246 void CORE_slatro_f1_quark(Quark *quark);
1247 void CORE_slange_quark(Quark *quark);
1248 void CORE_slange_f1_quark(Quark *quark);
1250 void CORE_slansy_quark(Quark *quark);
1251 void CORE_slansy_f1_quark(Quark *quark);
1253 void CORE_slansy_quark(Quark *quark);
1254 void CORE_slansy_f1_quark(Quark *quark);
1255 void CORE_slaset_quark(Quark *quark);
1256 void CORE_slaset2_quark(Quark *quark);
1257 void CORE_slatro_quark(Quark *quark);
1258 void CORE_slauum_quark(Quark *quark);
1259 void CORE_spamm_quark(Quark *quark);
1260 void CORE_splgsy_quark(Quark *quark);
1261 void CORE_splgsy_quark(Quark *quark);
1262 void CORE_splrnt_quark(Quark *quark);
1263 void CORE_spltmg_quark(Quark *quark);
1264 void CORE_splssq_quark(Quark *quark);
1265 void CORE_spotrf_quark(Quark *quark);
1266 void CORE_ssetvar_quark(Quark *quark);
1267 void CORE_sshift_quark(Quark *quark);
1268 void CORE_sshiftw_quark(Quark *quark);
1269 void CORE_sssssm_quark(Quark *quark);
1270 void CORE_ssymm_quark(Quark *quark);
1271 void CORE_ssyrk_quark(Quark *quark);
1272 void CORE_ssyr2k_quark(Quark *quark);
1273 void CORE_ssyssq_quark(Quark *quark);
1274 void CORE_ssyssq_f1_quark(Quark *quark);
1275 void CORE_sswpab_quark(Quark *quark);
1276 void CORE_sswptr_ontile_quark(Quark *quark);
1277 void CORE_strdalg1_quark(Quark *quark);
1278 void CORE_strmm_quark(Quark *quark);
1279 void CORE_strsm_quark(Quark *quark);
1280 void CORE_strtri_quark(Quark *quark);
1281 void CORE_stslqt_quark(Quark *quark);
1282 void CORE_stsmlq_quark(Quark *quark);
1283 void CORE_stsmlq_sytra1_quark(Quark *quark);
1284 void CORE_stsmlq_corner_quark(Quark *quark);
1285 void CORE_stsmqr_quark(Quark *quark);
1286 void CORE_stsmqr_sytra1_quark(Quark *quark);
1287 void CORE_stsmqr_corner_quark(Quark *quark);
1288 void CORE_stsqrt_quark(Quark *quark);
1289 void CORE_ststrf_quark(Quark *quark);
1290 void CORE_sttmqr_quark(Quark *quark);
1291 void CORE_sttqrt_quark(Quark *quark);
1292 void CORE_sttmlq_quark(Quark *quark);
1293 void CORE_sttlqt_quark(Quark *quark);
1294 void CORE_sormlq_quark(Quark *quark);
1295 void CORE_sormqr_quark(Quark *quark);
1296 void CORE_slaswp_quark(Quark* quark);
1297 void CORE_slaswp_f2_quark(Quark* quark);
1298 void CORE_slaswp_ontile_quark(Quark *quark);
1299 void CORE_slaswp_ontile_f2_quark(Quark *quark);
1300 void CORE_slaswpc_ontile_quark(Quark *quark);
1301 void CORE_strmm_p2_quark(Quark* quark);
1302 void CORE_sgemm_f2_quark(Quark* quark);
1303 void CORE_sgemm_p2_quark(Quark* quark);
1304 void CORE_sgemm_p2f1_quark(Quark* quark);
1305 void CORE_sgemm_p3_quark(Quark* quark);
void CORE_slaed0_betaapprox(int subpbs, const int *subpbs_info, float *D, const float *E)
Definition: core_slaed0_betaapprox.c:49
int CORE_sttqrt(int M, int N, int IB, float *A1, int LDA1, float *A2, int LDA2, float *T, int LDT, float *TAU, float *WORK)
Definition: core_sttqrt.c:116
int CORE_stsmlq(PLASMA_enum side, PLASMA_enum trans, int M1, int N1, int M2, int N2, int K, int IB, float *A1, int LDA1, float *A2, int LDA2, const float *V, int LDV, const float *T, int LDT, float *WORK, int LDWORK)
Definition: core_stsmlq.c:124
int CORE_sgetf2_nopiv(int m, int n, float *A, int lda)
Definition: core_sgetf2_nopiv.c:64
void CORE_spotrf(PLASMA_enum uplo, int N, float *A, int LDA, int *INFO)
Definition: core_spotrf.c:68
int CORE_slarfx2c(PLASMA_enum uplo, float V, float TAU, float *C1, float *C2, float *C3)
Definition: core_slarfx_tbrd.c:177
void CORE_slaed3_computeW(int n, int K, const float *Q, int LDQ, const float *DLAMBDA, float *W, const int *INDX, int start, int end)
Definition: core_slaed3_computeW.c:66
int CORE_sswptr_ontile(PLASMA_desc descA, int i1, int i2, const int *ipiv, int inc, const float *Akk, int ldak)
Definition: core_slaswp.c:224
void CORE_sgbtype1cb(PLASMA_enum uplo, int N, int NB, float *A, int LDA, float *VQ, float *TAUQ, float *VP, float *TAUP, int st, int ed, int sweep, int Vblksiz, int WANTZ, float *WORK)
Definition: core_sgbtype1cb.c:121
void CORE_strsm(PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag, int M, int N, float alpha, const float *A, int LDA, float *B, int LDB)
Definition: core_strsm.c:80
int CORE_sttmlq(PLASMA_enum side, PLASMA_enum trans, int M1, int N1, int M2, int N2, int K, int IB, float *A1, int LDA1, float *A2, int LDA2, const float *V, int LDV, const float *T, int LDT, float *WORK, int LDWORK)
Definition: core_sttmlq.c:116
int CORE_spltmg_chebvand(int M, int N, float *A, int LDA, int gN, int m0, int n0, float *W)
Definition: core_spltmg_chebvand.c:83
int CORE_spamm(int op, PLASMA_enum side, PLASMA_enum storev, int M, int N, int K, int L, const float *A1, int LDA1, float *A2, int LDA2, const float *V, int LDV, float *W, int LDW)
Definition: core_spamm.c:175
void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans, int N, int K, float alpha, const float *A, int LDA, float beta, float *C, int LDC)
Definition: core_ssyrk.c:80
void CORE_ssbtype2cb(int N, int NB, float *A, int LDA, float *V, float *TAU, int st, int ed, int sweep, int Vblksiz, int WANTZ, float *WORK)
Definition: core_ssbtype2cb.c:101
void CORE_sgeqp3_larfg(PLASMA_desc A, int ii, int jj, int i, int j, float *tau, float *beta)
Definition: core_sgeqp3_larfg.c:78
int CORE_shbrce(PLASMA_enum uplo, int N, PLASMA_desc *A, float *V, float *TAU, int st, int ed, int eltsize)
Definition: core_shbrce.c:78
void CORE_sswap(int m, int n, float *Q, int ldq, const float *work, const int *perm, int start, int end)
Definition: core_sswap.c:67
void CORE_slaed2_copydef(int n, int n1, int K, const int *ctot, float *Q, int LDQ, const float *Q2, int start, int end)
Definition: core_slaed2_compressq.c:180
void CORE_slaed3_merge(int n, int K, float *D, int *INDXQ)
Definition: core_slaed3_merge.c:46
int CORE_spltmg(PLASMA_enum mtxtype, int m, int n, float *A, int lda, int gM, int gN, int m0, int n0, unsigned long long int seed)
Definition: core_spltmg.c:90
int CORE_slaswp_ontile(PLASMA_desc descA, int i1, int i2, const int *ipiv, int inc)
Definition: core_slaswp.c:107
void CORE_slaset(PLASMA_enum uplo, int n1, int n2, float alpha, float beta, float *tileA, int ldtilea)
Definition: core_slaset.c:57
int CORE_ssyrfb(PLASMA_enum uplo, int N, int K, int IB, int NB, const float *A, int LDA, const float *T, int LDT, float *C, int LDC, float *WORK, int LDWORK)
Definition: core_ssyrfb.c:110
int CORE_sgessm(int M, int N, int K, int IB, const int *IPIV, const float *L, int LDL, float *A, int LDA)
Definition: core_sgessm.c:70
int CORE_shblrx(PLASMA_enum uplo, int N, PLASMA_desc *A, float *V, float *TAU, int st, int ed, int eltsize)
Definition: core_shblrx.c:78
int CORE_slarfx2(PLASMA_enum side, int N, float V, float TAU, float *C1, int LDC1, float *C2, int LDC2)
Definition: core_slarfx_tbrd.c:82
void CORE_sgeqp3_pivot(PLASMA_desc A, float *F, int ldf, int jj, int k, int *jpvt, float *norms1, float *norms2, int *info)
Definition: core_sgeqp3_pivot.c:68
void CORE_sshiftw(int s, int cl, int m, int n, int L, float *A, float *W)
Definition: core_sshift.c:66
void CORE_strasm(PLASMA_enum storev, PLASMA_enum uplo, PLASMA_enum diag, int M, int N, const float *A, int lda, float *work)
Definition: core_strasm.c:66
void CORE_spltmg_condexq(int M, int N, float *Q, int LDQ)
Definition: core_spltmg_condex.c:53
void CORE_sgeqp3_update(const float *Ajj, int lda1, float *Ajk, int lda2, const float *Fk, int ldf, int joff, int k, int koff, int nb, float *norms1, float *norms2, int *info)
Definition: core_sgeqp3_update.c:82
int CORE_slag2c(int m, int n, const float *R, int ldr, PLASMA_Complex32_t *Z, int ldz)
Definition: core_slag2c.c:58
void CORE_slaed2_compressq(int n, int n1, const int *INDX, const int *ctot, const float *Q, int LDQ, float *Q2, int start, int end)
Definition: core_slaed2_compressq.c:85
void CORE_sgeqp3_norms(PLASMA_desc A, int ioff, int joff, float *norms1, float *norms2)
Definition: core_sgeqp3_norms.c:62
void CORE_ssygst(int itype, PLASMA_enum uplo, int N, float *A, int LDA, float *B, int LDB, int *INFO)
Definition: core_ssygst.c:80
int CORE_sgetrf_reclap(CORE_sgetrf_data_t *data, int M, int N, float *A, int LDA, int *IPIV, int *info)
Definition: core_sgetrf_reclap.c:125
void CORE_strmm(PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag, int M, int N, float alpha, const float *A, int LDA, float *B, int LDB)
Definition: core_strmm.c:84
void CORE_slantr(PLASMA_enum norm, PLASMA_enum uplo, PLASMA_enum diag, int M, int N, const float *A, int LDA, float *work, float *normA)
Definition: core_slantr.c:86
int CORE_sgetrf_incpiv(int M, int N, int IB, float *A, int LDA, int *IPIV, int *INFO)
Definition: core_sgetrf_incpiv.c:83
int CORE_stsmqr(PLASMA_enum side, PLASMA_enum trans, int M1, int N1, int M2, int N2, int K, int IB, float *A1, int LDA1, float *A2, int LDA2, const float *V, int LDV, const float *T, int LDT, float *WORK, int LDWORK)
Definition: core_stsmqr.c:124
int CORE_sgetrf_rectil(CORE_sgetrf_data_t *data, const PLASMA_desc A, int *IPIV, int *info)
Definition: core_sgetrf_rectil.c:122
void CORE_sbrdalg1(PLASMA_enum uplo, int n, int nb, float *A, int lda, float *VQ, float *TAUQ, float *VP, float *TAUP, int Vblksiz, int wantz, int i, int sweepid, int m, int grsiz, float *work)
Definition: core_sbrdalg1.c:111
int CORE_sparfb(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev, int M1, int N1, int M2, int N2, int K, int L, float *A1, int LDA1, float *A2, int LDA2, const float *V, int LDV, const float *T, int LDT, float *WORK, int LDWORK)
Definition: core_sparfb.c:132
void CORE_slaed3_updatevectors(int op, int wsmode, int n, int n1, int K, int il_nondef, int iu_nondef, float *Q, int ldq, float *Q2, const int *ctot, float *WORK, int start, int end)
Definition: core_slaed3_updatevectors.c:95
int CORE_stsmqr_corner(int m1, int n1, int m2, int n2, int m3, int n3, int k, int ib, int nb, float *A1, int lda1, float *A2, int lda2, float *A3, int lda3, const float *V, int ldv, const float *T, int ldt, float *WORK, int ldwork)
Definition: core_stsmqr_corner.c:130
int CORE_stradd(PLASMA_enum uplo, PLASMA_enum trans, int M, int N, float alpha, const float *A, int LDA, float beta, float *B, int LDB)
Definition: core_stradd.c:86
int CORE_sgbrce(PLASMA_enum uplo, int N, PLASMA_desc *A, float *V, float *TAU, int st, int ed, int eltsize)
Definition: core_sgbrce.c:78
int CORE_stsmlq_corner(int m1, int n1, int m2, int n2, int m3, int n3, int k, int ib, int nb, float *A1, int lda1, float *A2, int lda2, float *A3, int lda3, const float *V, int ldv, const float *T, int ldt, float *WORK, int ldwork)
Definition: core_stsmlq_corner.c:130
void CORE_ssbtype1cb(int N, int NB, float *A, int LDA, float *V, float *TAU, int st, int ed, int sweep, int Vblksiz, int WANTZ, float *WORK)
Definition: core_ssbtype1cb.c:97
int CORE_sttmqr(PLASMA_enum side, PLASMA_enum trans, int M1, int N1, int M2, int N2, int K, int IB, float *A1, int LDA1, float *A2, int LDA2, const float *V, int LDV, const float *T, int LDT, float *WORK, int LDWORK)
Definition: core_sttmqr.c:116
int CORE_sgeqrt(int M, int N, int IB, float *A, int LDA, float *T, int LDT, float *TAU, float *WORK)
Definition: core_sgeqrt.c:86
void CORE_splrnt(int m, int n, float *A, int lda, int bigM, int m0, int n0, unsigned long long int seed)
Definition: core_splrnt.c:68
void CORE_slauum(PLASMA_enum uplo, int N, float *A, int LDA)
Definition: core_slauum.c:57
void CORE_ssbtype3cb(int N, int NB, float *A, int LDA, const float *V, const float *TAU, int st, int ed, int sweep, int Vblksiz, int WANTZ, float *WORK)
Definition: core_ssbtype3cb.c:98
void CORE_slange(int norm, int M, int N, const float *A, int LDA, float *work, float *normA)
Definition: core_slange.c:72
int CORE_slatro(PLASMA_enum uplo, PLASMA_enum trans, int M, int N, const float *A, int LDA, float *B, int LDB)
Definition: core_slatro.c:80
void CORE_strdalg1(int n, int nb, float *A, int lda, float *V, float *TAU, int Vblksiz, int wantz, int i, int sweepid, int m, int grsiz, float *work)
Definition: core_strdalg1.c:91
Definition: core_sgetrf_reclap.c:24
int CORE_slarfx2ce(PLASMA_enum uplo, float *V, float *TAU, float *C1, float *C2, float *C3)
Definition: core_slarfx_tbrd.c:323
int CORE_sgbelr(PLASMA_enum uplo, int N, PLASMA_desc *A, float *V, float *TAU, int st, int ed, int eltsize)
Definition: core_sgbelr.c:81
int CORE_stsmqr_sytra1(PLASMA_enum side, PLASMA_enum trans, int m1, int n1, int m2, int n2, int k, int ib, float *A1, int lda1, float *A2, int lda2, const float *V, int ldv, const float *T, int ldt, float *WORK, int ldwork)
Definition: core_stsmqr_sytra1.c:127
int CORE_slacpy_pivot(const PLASMA_desc descA, PLASMA_enum direct, int k1, int k2, const int *ipiv, int *rankin, int *rankout, float *A, int lda, int init)
Definition: core_slacpy_pivot.c:84
int CORE_spltmg_hankel(PLASMA_enum uplo, int M, int N, float *A, int LDA, int m0, int n0, int nb, const float *V1, const float *V2)
Definition: core_spltmg_hankel.c:88
void CORE_sgemm(PLASMA_enum transA, PLASMA_enum transB, int M, int N, int K, float alpha, const float *A, int LDA, const float *B, int LDB, float beta, float *C, int LDC)
Definition: core_sgemm.c:90
int CORE_sormqr(PLASMA_enum side, PLASMA_enum trans, int M, int N, int K, int IB, const float *V, int LDV, const float *T, int LDT, float *C, int LDC, float *WORK, int LDWORK)
Definition: core_sormqr.c:108
void CORE_slaset2(PLASMA_enum uplo, int n1, int n2, float alpha, float *tileA, int ldtilea)
Definition: core_slaset2.c:55
int CORE_stslqt(int M, int N, int IB, float *A1, int LDA1, float *A2, int LDA2, float *T, int LDT, float *TAU, float *WORK)
Definition: core_stslqt.c:107
void CORE_slarfy(int N, float *A, int LDA, const float *V, const float *TAU, float *WORK)
Definition: core_slarfy.c:61
int CORE_sgetrf_nopiv(int m, int n, int ib, float *A, int lda)
Definition: core_sgetrf_nopiv.c:69
void CORE_slaed3_reduceW(int n, int n1, int K, int l, const float *Q, int LDQ, const float *Wred, float *W)
Definition: core_slaed3_reduceW.c:60
void CORE_slaed2_computeK(int *K, int n, int n1, float *beta, float *D, float *Q, int LDQ, float *Z, float *DLAMBDA, float *W, int *INDX, int *INDXC, int *INDXP, int *INDXQ, int *COLTYP)
Definition: core_slaed2_computeK.c:104
void CORE_sgetrip(int m, int n, float *A, float *work)
Definition: core_sgetrip.c:51
int CORE_shbelr(PLASMA_enum uplo, int N, PLASMA_desc *A, float *V, float *TAU, int st, int ed, int eltsize)
Definition: core_shbelr.c:80
void CORE_spltmg_toeppd1(int gM, int m0, int M, float *W, unsigned long long int seed)
Definition: core_spltmg_toeppd.c:70
int CORE_sormlq(PLASMA_enum side, PLASMA_enum trans, int M, int N, int IB, int K, const float *V, int LDV, const float *T, int LDT, float *C, int LDC, float *WORK, int LDWORK)
Definition: core_sormlq.c:108
int CORE_slascl(PLASMA_enum type, int kl, int ku, float cfrom, float cto, int m, int n, float *A, int lda)
Definition: core_slascl.c:77
int CORE_sttlqt(int M, int N, int IB, float *A1, int LDA1, float *A2, int LDA2, float *T, int LDT, float *TAU, float *WORK)
Definition: core_sttlqt.c:116
void CORE_splgsy(float bump, int m, int n, float *A, int lda, int bigM, int m0, int n0, unsigned long long int seed)
Definition: core_splgsy.c:72
void CORE_sgemv(PLASMA_enum trans, int M, int N, float alpha, const float *A, int LDA, const float *x, int incx, float beta, float *y, int incy)
Definition: core_sgemv.c:78
int CORE_slarfb_gemm(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev, int M, int N, int K, const float *V, int LDV, const float *T, int LDT, float *C, int LDC, float *WORK, int LDWORK)
Definition: core_slarfb_gemm.c:113
void CORE_ssetvar(const float *alpha, float *x)
Definition: core_ssetvar.c:37
int CORE_stsmlq_sytra1(PLASMA_enum side, PLASMA_enum trans, int m1, int n1, int m2, int n2, int k, int ib, float *A1, int lda1, float *A2, int lda2, const float *V, int ldv, const float *T, int ldt, float *WORK, int ldwork)
Definition: core_stsmlq_sytra1.c:125
int CORE_slaswpc_ontile(PLASMA_desc descA, int i1, int i2, const int *ipiv, int inc)
Definition: core_slaswp.c:296
int CORE_spemv(PLASMA_enum trans, PLASMA_enum storev, int M, int N, int L, float ALPHA, const float *A, int LDA, const float *X, int INCX, float BETA, float *Y, int INCY, float *WORK)
Definition: core_spemv.c:118
int CORE_sgelqt(int M, int N, int IB, float *A, int LDA, float *T, int LDT, float *TAU, float *WORK)
Definition: core_sgelqt.c:85
void CORE_sgbtype2cb(PLASMA_enum uplo, int N, int NB, float *A, int LDA, float *VQ, float *TAUQ, float *VP, float *TAUP, int st, int ed, int sweep, int Vblksiz, int WANTZ, float *WORK)
Definition: core_sgbtype2cb.c:126
int CORE_slaed4(int n, int K, float *D, float beta, float *Q, int LDQ, const float *D0, const float *Z, const int *INDX, int start, int end)
Definition: core_slaed4.c:75
int CORE_sstedc(PLASMA_enum compz, int n, float *D, float *E, float *Z, int LDZ, float *WORK, int LWORK, int *IWORK, int LIWORK)
Definition: core_sstedc.c:85
void CORE_slaswp(int N, float *A, int LDA, int I1, int I2, const int *IPIV, int INC)
Definition: core_slaswp.c:62
int CORE_stsqrt(int M, int N, int IB, float *A1, int LDA1, float *A2, int LDA2, float *T, int LDT, float *TAU, float *WORK)
Definition: core_stsqrt.c:97
void CORE_strtri(PLASMA_enum uplo, PLASMA_enum diag, int N, float *A, int LDA, int *info)
Definition: core_strtri.c:66
void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo, int M, int N, float alpha, const float *A, int LDA, const float *B, int LDB, float beta, float *C, int LDC)
Definition: core_ssymm.c:88
void CORE_sasum(int storev, PLASMA_enum uplo, int M, int N, const float *A, int lda, float *work)
Definition: core_sasum.c:61
int CORE_sssssm(int M1, int N1, int M2, int N2, int K, int IB, float *A1, int LDA1, float *A2, int LDA2, const float *L1, int LDL1, const float *L2, int LDL2, const int *IPIV)
Definition: core_sssssm.c:94
void CORE_slaed3_computevectors(int K, int il_nondef, int iu_nondef, float *Q, int LDQ, float *W, float *S, const int *INDXC, int start, int end)
Definition: core_slaed3_computevectors.c:74
void CORE_slacpy(PLASMA_enum uplo, int M, int N, const float *A, int LDA, float *B, int LDB)
Definition: core_slacpy.c:59
Definition: descriptor.h:41
void CORE_spltmg_fiedler(int m, int n, const float *X, int incX, const float *Y, int incY, float *A, int lda)
Definition: core_spltmg_fiedler.c:75
void CORE_slansy(int norm, PLASMA_enum uplo, int N, const float *A, int LDA, float *work, float *normA)
Definition: core_slansy.c:74
void CORE_spltmg_toeppd2(int M, int N, int K, int m0, int n0, const float *W, float *A, int LDA)
Definition: core_spltmg_toeppd.c:149
int CORE_spltmg_circul(int M, int N, float *A, int LDA, int gM, int m0, int n0, const float *V)
Definition: core_spltmg_circul.c:76
int CORE_sgblrx(PLASMA_enum uplo, int N, PLASMA_desc *A, float *V, float *TAU, int st, int ed, int eltsize)
Definition: core_sgblrx.c:80
void CORE_sswpab(int i, int n1, int n2, float *A, float *work)
Definition: core_sswpab.c:63
int CORE_slascal(PLASMA_enum uplo, int m, int n, float alpha, float *A, int lda)
Definition: core_slascal.c:52
int CORE_sgeadd(PLASMA_enum trans, int M, int N, float alpha, const float *A, int LDA, float beta, float *B, int LDB)
Definition: core_sgeadd.c:75
int CORE_ststrf(int M, int N, int IB, int NB, float *U, int LDU, float *A, int LDA, float *L, int LDL, int *IPIV, float *WORK, int LDWORK, int *INFO)
Definition: core_ststrf.c:99
void CORE_sgeqp3_init(int n, int *jpvt)
Definition: core_sgeqp3_init.c:38
int CORE_ssteqr(PLASMA_enum compz, int n, float *D, float *E, float *Z, int LDZ, float *WORK)
Definition: core_ssteqr.c:66
void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans, int N, int K, float alpha, const float *A, int LDA, const float *B, int LDB, float beta, float *C, int LDC)
Definition: core_ssyr2k.c:90
void CORE_sgbtype3cb(PLASMA_enum uplo, int N, int NB, float *A, int LDA, float *VQ, float *TAUQ, float *VP, float *TAUP, int st, int ed, int sweep, int Vblksiz, int WANTZ, float *WORK)
Definition: core_sgbtype3cb.c:121
int CORE_sgetrf(int M, int N, float *A, int LDA, int *IPIV, int *INFO)
Definition: core_sgetrf.c:60