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);
32 void CORE_sbrdalg1( PLASMA_enum uplo,
41 int Vblksiz,
int wantz,
42 int i,
int sweepid,
int m,
int grsiz,
44 int CORE_sgbelr(PLASMA_enum uplo,
int N,
46 int st,
int ed,
int eltsize);
47 int CORE_sgbrce(PLASMA_enum uplo,
int N,
49 int st,
int ed,
int eltsize);
50 int CORE_sgblrx(PLASMA_enum uplo,
int N,
52 int st,
int ed,
int eltsize);
53 int CORE_sgeadd(
int M,
int N,
float alpha,
54 const float *A,
int LDA,
56 int CORE_sgelqt(
int M,
int N,
int IB,
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);
70 void CORE_sgeqp3_init(
int n,
int *jpvt );
71 void CORE_sgeqp3_larfg(
PLASMA_desc A,
int ii,
int jj,
int i,
int j,
72 float *tau,
float *beta );
73 void CORE_sgeqp3_norms(
PLASMA_desc A,
int ioff,
int joff,
float *norms1,
float *norms2 );
74 void CORE_sgeqp3_pivot(
PLASMA_desc A,
float *F,
int ldf,
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,
81 void CORE_sgeqp3_update(
const float *Ajj,
int lda1,
83 const float *Fk,
int ldf,
84 int joff,
int k,
int koff,
int nb,
85 float *norms1,
float *norms2,
87 int CORE_sgeqrt(
int M,
int N,
int IB,
90 float *TAU,
float *WORK);
91 int CORE_sgessm(
int M,
int N,
int K,
int IB,
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);
98 int CORE_sgetf2_nopiv(
int m,
int n,
100 int CORE_sgetrf(
int M,
int N,
102 int *IPIV,
int *INFO);
103 int CORE_sgetrf_incpiv(
int M,
int N,
int IB,
105 int *IPIV,
int *INFO);
106 int CORE_sgetrf_nopiv(
int m,
int n,
int ib,
108 int CORE_sgetrf_reclap(
int M,
int N,
110 int *IPIV,
int *info);
111 void CORE_sgetrf_reclap_init(
void);
112 int CORE_sgetrf_rectil(
const PLASMA_desc A,
int *IPIV,
int *info);
113 void CORE_sgetrf_rectil_init(
void);
114 void CORE_sgetrip(
int m,
int n,
float *A,
116 int CORE_shbelr(PLASMA_enum uplo,
int N,
118 int st,
int ed,
int eltsize);
119 int CORE_shblrx(PLASMA_enum uplo,
int N,
121 int st,
int ed,
int eltsize);
122 int CORE_shbrce(PLASMA_enum uplo,
int N,
124 int st,
int ed,
int eltsize);
125 void CORE_ssbtype1cb(
int N,
int NB,
127 float *V,
float *TAU,
128 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
130 void CORE_ssbtype2cb(
int N,
int NB,
132 float *V,
float *TAU,
133 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
135 void CORE_ssbtype3cb(
int N,
int NB,
137 const float *V,
const float *TAU,
138 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
140 void CORE_sgbtype1cb(PLASMA_enum uplo,
int N,
int NB,
142 float *VQ,
float *TAUQ,
143 float *VP,
float *TAUP,
144 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
146 void CORE_sgbtype2cb(PLASMA_enum uplo,
int N,
int NB,
148 float *VQ,
float *TAUQ,
149 float *VP,
float *TAUP,
150 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
152 void CORE_sgbtype3cb(PLASMA_enum uplo,
int N,
int NB,
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);
185 void CORE_slacpy(PLASMA_enum uplo,
int M,
int N,
186 const float *A,
int LDA,
190 int k1,
int k2,
const int *ipiv,
191 int *rankin,
int *rankout,
194 void CORE_slange(
int norm,
int M,
int N,
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 int CORE_slarfb_gemm(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
207 const float *V,
int LDV,
208 const float *T,
int LDT,
210 float *WORK,
int LDWORK);
211 int CORE_slarfx2(PLASMA_enum side,
int N,
215 float *C2,
int LDC2);
216 int CORE_slarfx2c(PLASMA_enum uplo,
222 int CORE_slarfx2ce(PLASMA_enum uplo,
228 void CORE_slarfy(
int N,
233 void CORE_slaset(PLASMA_enum uplo,
int n1,
int n2,
234 float alpha,
float beta,
235 float *tileA,
int ldtilea);
236 void CORE_slaset2(PLASMA_enum uplo,
int n1,
int n2,
float alpha,
237 float *tileA,
int ldtilea);
238 void CORE_slaswp(
int N,
float *A,
int LDA,
239 int I1,
int I2,
const int *IPIV,
int INC);
240 int CORE_slaswp_ontile(
PLASMA_desc descA,
int i1,
int i2,
const int *ipiv,
int inc);
241 int CORE_slaswpc_ontile(
PLASMA_desc descA,
int i1,
int i2,
const int *ipiv,
int inc);
242 int CORE_slatro(PLASMA_enum uplo, PLASMA_enum trans,
244 const float *A,
int LDA,
246 void CORE_slauum(PLASMA_enum uplo,
int N,
float *A,
int LDA);
247 int CORE_spamm(
int op, PLASMA_enum side, PLASMA_enum storev,
248 int M,
int N,
int K,
int L,
249 const float *A1,
int LDA1,
251 const float *V,
int LDV,
253 int CORE_sparfb(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
254 int M1,
int N1,
int M2,
int N2,
int K,
int L,
257 const float *V,
int LDV,
258 const float *T,
int LDT,
259 float *WORK,
int LDWORK);
260 int CORE_spemv(PLASMA_enum trans, PLASMA_enum storev,
263 const float *A,
int LDA,
264 const float *X,
int INCX,
268 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
269 int bigM,
int m0,
int n0,
unsigned long long int seed );
270 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
271 int bigM,
int m0,
int n0,
unsigned long long int seed );
272 void CORE_splrnt(
int m,
int n,
float *A,
int lda,
273 int bigM,
int m0,
int n0,
unsigned long long int seed );
274 int CORE_spltmg(PLASMA_enum mtxtype,
int m,
int n,
float *A,
int lda,
275 int gM,
int gN,
int m0,
int n0,
unsigned long long int seed );
276 int CORE_spltmg_chebvand(
int M,
int N,
float *A,
int LDA,
277 int gN,
int m0,
int n0,
279 int CORE_spltmg_circul(
int M,
int N,
float *A,
int LDA,
280 int gM,
int m0,
int n0,
282 void CORE_spltmg_condexq(
int M,
int N,
float *Q,
int LDQ );
283 void CORE_spltmg_fiedler(
int m,
int n,
284 const float *X,
int incX,
285 const float *Y,
int incY,
287 int CORE_spltmg_hankel( PLASMA_enum uplo,
int M,
int N,
float *A,
int LDA,
288 int m0,
int n0,
int nb,
291 void CORE_spltmg_toeppd1(
int gM,
int m0,
int M,
float *W,
292 unsigned long long int seed );
293 void CORE_spltmg_toeppd2(
int M,
int N,
int K,
int m0,
int n0,
296 void CORE_spotrf(PLASMA_enum uplo,
int N,
float *A,
int LDA,
int *INFO);
297 void CORE_ssetvar(
const float *alpha,
float *x);
298 void CORE_sshift(
int s,
int m,
int n,
int L,
300 void CORE_sshiftw(
int s,
int cl,
int m,
int n,
int L,
302 int CORE_sssssm(
int M1,
int N1,
int M2,
int N2,
int K,
int IB,
305 const float *L1,
int LDL1,
306 const float *L2,
int LDL2,
308 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
310 float alpha,
const float *A,
int LDA,
311 const float *B,
int LDB,
312 float beta,
float *C,
int LDC);
313 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
315 float alpha,
const float *A,
int LDA,
316 float beta,
float *C,
int LDC);
317 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
319 float alpha,
const float *A,
int LDA,
320 const float *B,
int LDB,
321 float beta,
float *C,
int LDC);
322 int CORE_ssyssq(PLASMA_enum uplo,
int N,
323 const float *A,
int LDA,
324 float *scale,
float *sumsq);
325 void CORE_sswpab(
int i,
int n1,
int n2,
326 float *A,
float *work);
327 int CORE_sswptr_ontile(
PLASMA_desc descA,
int i1,
int i2,
const int *ipiv,
int inc,
328 const float *Akk,
int ldak);
329 void CORE_strdalg1(
int n,
335 int Vblksiz,
int wantz,
336 int i,
int sweepid,
int m,
int grsiz,
338 void CORE_strmm(PLASMA_enum side, PLASMA_enum uplo,
339 PLASMA_enum transA, PLASMA_enum diag,
341 float alpha,
const float *A,
int LDA,
343 void CORE_strsm(PLASMA_enum side, PLASMA_enum uplo,
344 PLASMA_enum transA, PLASMA_enum diag,
346 float alpha,
const float *A,
int LDA,
348 void CORE_strtri(PLASMA_enum uplo, PLASMA_enum diag,
int N,
349 float *A,
int LDA,
int *info);
350 int CORE_stslqt(
int M,
int N,
int IB,
354 float *TAU,
float *WORK);
355 int CORE_stsmlq(PLASMA_enum side, PLASMA_enum trans,
356 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
359 const float *V,
int LDV,
360 const float *T,
int LDT,
361 float *WORK,
int LDWORK);
362 int CORE_stsmlq_corner(
int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
363 int k,
int ib,
int nb,
367 const float *V,
int ldv,
368 const float *T,
int ldt,
369 float *WORK,
int ldwork);
370 int CORE_stsmlq_sytra1( PLASMA_enum side, PLASMA_enum trans,
371 int m1,
int n1,
int m2,
int n2,
375 const float *V,
int ldv,
376 const float *T,
int ldt,
377 float *WORK,
int ldwork);
378 int CORE_stsmqr(PLASMA_enum side, PLASMA_enum trans,
379 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
382 const float *V,
int LDV,
383 const float *T,
int LDT,
384 float *WORK,
int LDWORK);
385 int CORE_stsmqr_corner(
int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
386 int k,
int ib,
int nb,
390 const float *V,
int ldv,
391 const float *T,
int ldt,
392 float *WORK,
int ldwork);
393 int CORE_stsmqr_sytra1( PLASMA_enum side, PLASMA_enum trans,
394 int m1,
int n1,
int m2,
int n2,
398 const float *V,
int ldv,
399 const float *T,
int ldt,
400 float *WORK,
int ldwork);
401 int CORE_stsqrt(
int M,
int N,
int IB,
405 float *TAU,
float *WORK);
406 int CORE_ststrf(
int M,
int N,
int IB,
int NB,
410 int *IPIV,
float *WORK,
411 int LDWORK,
int *INFO);
412 int CORE_sttmqr(PLASMA_enum side, PLASMA_enum trans,
413 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
416 const float *V,
int LDV,
417 const float *T,
int LDT,
418 float *WORK,
int LDWORK);
419 int CORE_sttqrt(
int M,
int N,
int IB,
425 int CORE_sttmlq(PLASMA_enum side, PLASMA_enum trans,
426 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
429 const float *V,
int LDV,
430 const float *T,
int LDT,
431 float *WORK,
int LDWORK);
432 int CORE_sttlqt(
int M,
int N,
int IB,
438 int CORE_sormlq(PLASMA_enum side, PLASMA_enum trans,
439 int M,
int N,
int IB,
int K,
440 const float *V,
int LDV,
441 const float *T,
int LDT,
443 float *WORK,
int LDWORK);
444 int CORE_sormqr(PLASMA_enum side, PLASMA_enum trans,
445 int M,
int N,
int K,
int IB,
446 const float *V,
int LDV,
447 const float *T,
int LDT,
449 float *WORK,
int LDWORK);
455 void QUARK_CORE_sasum(Quark *quark, Quark_Task_Flags *task_flags,
456 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
457 const float *A,
int lda,
int szeA,
458 float *work,
int szeW);
459 void QUARK_CORE_sasum_f1(Quark *quark, Quark_Task_Flags *task_flags,
460 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
461 const float *A,
int lda,
int szeA,
462 float *work,
int szeW,
463 float *fake,
int szeF);
464 void QUARK_CORE_sgeadd(Quark *quark, Quark_Task_Flags *task_flags,
465 int m,
int n,
int nb,
float alpha,
466 const float *A,
int lda,
468 void QUARK_CORE_sbrdalg1(Quark *quark, Quark_Task_Flags *task_flags,
477 int Vblksiz,
int wantz,
478 int i,
int sweepid,
int m,
int grsiz,
479 int *PCOL,
int *ACOL,
int *MCOL);
480 void QUARK_CORE_sgelqt(Quark *quark, Quark_Task_Flags *task_flags,
481 int m,
int n,
int ib,
int nb,
484 void QUARK_CORE_sgemm(Quark *quark, Quark_Task_Flags *task_flags,
485 PLASMA_enum transA, PLASMA_enum transB,
486 int m,
int n,
int k,
int nb,
487 float alpha,
const float *A,
int lda,
488 const float *B,
int ldb,
489 float beta,
float *C,
int ldc);
490 void QUARK_CORE_sgemm2( Quark *quark, Quark_Task_Flags *task_flags,
491 PLASMA_enum transA, PLASMA_enum transB,
492 int m,
int n,
int k,
int nb,
493 float alpha,
const float *A,
int lda,
494 const float *B,
int ldb,
495 float beta,
float *C,
int ldc);
496 void QUARK_CORE_sgemm_f2(Quark *quark, Quark_Task_Flags *task_flags,
497 PLASMA_enum transA, PLASMA_enum transB,
498 int m,
int n,
int k,
int nb,
499 float alpha,
const float *A,
int lda,
500 const float *B,
int ldb,
501 float beta,
float *C,
int ldc,
502 float *fake1,
int szefake1,
int flag1,
503 float *fake2,
int szefake2,
int flag2);
504 void QUARK_CORE_sgemm_p2(Quark *quark, Quark_Task_Flags *task_flags,
505 PLASMA_enum transA, PLASMA_enum transB,
506 int m,
int n,
int k,
int nb,
507 float alpha,
const float *A,
int lda,
508 const float **B,
int ldb,
509 float beta,
float *C,
int ldc);
510 void QUARK_CORE_sgemm_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
511 PLASMA_enum transA, PLASMA_enum transB,
512 int m,
int n,
int k,
int nb,
513 float alpha,
const float *A,
int lda,
514 const float **B,
int ldb,
515 float beta,
float *C,
int ldc,
516 float *fake1,
int szefake1,
int flag1);
517 void QUARK_CORE_sgemm_p3(Quark *quark, Quark_Task_Flags *task_flags,
518 PLASMA_enum transA, PLASMA_enum transB,
519 int m,
int n,
int k,
int nb,
520 float alpha,
const float *A,
int lda,
521 const float *B,
int ldb,
522 float beta,
float **C,
int ldc);
523 void QUARK_CORE_sgemm_tile(Quark *quark, Quark_Task_Flags *task_flags,
524 PLASMA_enum transA, PLASMA_enum transB,
525 int m,
int n,
int k,
int nb,
526 const float *alpha,
const float *A,
int lda,
527 const float *B,
int ldb,
528 const float *beta,
float *C,
int ldc,
532 void QUARK_CORE_sgemv(Quark *quark, Quark_Task_Flags *task_flags,
533 PLASMA_enum trans,
int m,
int n,
534 float alpha,
const float *A,
int lda,
535 const float *x,
int incx,
536 float beta,
float *y,
int incy);
537 void QUARK_CORE_sgemv_tile(Quark *quark, Quark_Task_Flags *task_flags,
540 const float *alpha,
const float *A,
int lda,
541 const float *x,
int incx,
542 const float *beta,
float *y,
int incy,
546 void QUARK_CORE_sgeqp3_init( Quark *quark, Quark_Task_Flags *task_flags,
548 void QUARK_CORE_sgeqp3_larfg(Quark *quark, Quark_Task_Flags *task_flags,
550 float *tau,
float *beta );
551 void QUARK_CORE_sgeqp3_norms( Quark *quark, Quark_Task_Flags *task_flags,
552 PLASMA_desc A,
int ioff,
int joff,
float *norms1,
float *norms2 );
553 void QUARK_CORE_sgeqp3_pivot( Quark *quark, Quark_Task_Flags *task_flags,
556 int jj,
int k,
int *jpvt,
557 float *norms1,
float *norms2,
int *info );
558 void QUARK_CORE_sgeqp3_tntpiv(Quark *quark, Quark_Task_Flags *task_flags,
559 int m,
int n,
int nb,
563 PLASMA_bool check_info,
int iinfo);
564 void QUARK_CORE_sgeqp3_update( Quark *quark, Quark_Task_Flags *task_flags,
565 float *Ajj,
int lda1,
566 float *Ajk,
int lda2,
568 int joff,
int k,
int koff,
int nb,
569 float *norms1,
float *norms2,
int *info );
570 void QUARK_CORE_sgeqrt(Quark *quark, Quark_Task_Flags *task_flags,
571 int m,
int n,
int ib,
int nb,
574 void QUARK_CORE_sgessm(Quark *quark, Quark_Task_Flags *task_flags,
575 int m,
int n,
int k,
int ib,
int nb,
577 const float *L,
int ldl,
579 void QUARK_CORE_sgessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
580 int m,
int n,
const float *A,
int lda,
581 float *scale,
float *sumsq,
582 float *fake,
int szeF,
int paramF );
583 void QUARK_CORE_sgetrf(Quark *quark, Quark_Task_Flags *task_flags,
584 int m,
int n,
int nb,
588 PLASMA_bool check_info,
int iinfo);
589 void QUARK_CORE_sgetrf_incpiv(Quark *quark, Quark_Task_Flags *task_flags,
590 int m,
int n,
int ib,
int nb,
594 PLASMA_bool check_info,
int iinfo);
595 void QUARK_CORE_sgetrf_nopiv(Quark *quark, Quark_Task_Flags *task_flags,
596 int m,
int n,
int ib,
int nb,
600 void QUARK_CORE_sgetrf_reclap(Quark *quark, Quark_Task_Flags *task_flags,
601 int m,
int n,
int nb,
605 PLASMA_bool check_info,
int iinfo,
607 void QUARK_CORE_sgetrf_rectil(Quark *quark, Quark_Task_Flags *task_flags,
611 PLASMA_bool check_info,
int iinfo,
613 void QUARK_CORE_sgetrip(Quark *quark, Quark_Task_Flags *task_flags,
614 int m,
int n,
float *A,
int szeA);
615 void QUARK_CORE_sgetrip_f1(Quark *quark, Quark_Task_Flags *task_flags,
616 int m,
int n,
float *A,
int szeA,
617 float *fake,
int szeF,
int paramF);
618 void QUARK_CORE_sgetrip_f2(Quark *quark, Quark_Task_Flags *task_flags,
619 int m,
int n,
float *A,
int szeA,
620 float *fake1,
int szeF1,
int paramF1,
621 float *fake2,
int szeF2,
int paramF2);
622 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
623 PLASMA_enum side, PLASMA_enum uplo,
624 int m,
int n,
int nb,
625 float alpha,
const float *A,
int lda,
626 const float *B,
int ldb,
627 float beta,
float *C,
int ldc);
628 void QUARK_CORE_ssygst(Quark *quark, Quark_Task_Flags *task_flags,
629 int itype, PLASMA_enum uplo,
int N,
634 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
635 PLASMA_enum uplo, PLASMA_enum trans,
636 int n,
int k,
int nb,
637 float alpha,
const float *A,
int lda,
638 float beta,
float *C,
int ldc);
639 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
640 PLASMA_enum uplo, PLASMA_enum trans,
641 int n,
int k,
int nb,
642 float alpha,
const float *A,
int lda,
643 const float *B,
int LDB,
644 float beta,
float *C,
int ldc);
645 void QUARK_CORE_ssyrfb(Quark *quark, Quark_Task_Flags *task_flags,
647 int n,
int k,
int ib,
int nb,
648 const float *A,
int lda,
649 const float *T,
int ldt,
651 void QUARK_CORE_shessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
652 PLASMA_enum uplo,
int n,
const float *A,
int lda,
653 float *scale,
float *sumsq,
654 float *fake,
int szeF,
int paramF );
655 void QUARK_CORE_slacpy(Quark *quark, Quark_Task_Flags *task_flags,
656 PLASMA_enum uplo,
int m,
int n,
int mb,
657 const float *A,
int lda,
659 void QUARK_CORE_slacpy_f1(Quark *quark, Quark_Task_Flags *task_flags,
660 PLASMA_enum uplo,
int m,
int n,
int nb,
661 const float *A,
int lda,
663 float *fake1,
int szefake1,
int flag1);
664 void QUARK_CORE_slacpy_pivot(Quark *quark, Quark_Task_Flags *task_flags,
667 int k1,
int k2,
const int *ipiv,
668 int *rankin,
int *rankout,
671 void QUARK_CORE_slange(Quark *quark, Quark_Task_Flags *task_flags,
672 int norm,
int M,
int N,
673 const float *A,
int LDA,
int szeA,
674 int szeW,
float *result);
675 void QUARK_CORE_slange_f1(Quark *quark, Quark_Task_Flags *task_flags,
676 int norm,
int M,
int N,
677 const float *A,
int LDA,
int szeA,
678 int szeW,
float *result,
679 float *fake,
int szeF);
681 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
682 int norm, PLASMA_enum uplo,
int N,
683 const float *A,
int LDA,
int szeA,
684 int szeW,
float *result);
685 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
686 int norm, PLASMA_enum uplo,
int N,
687 const float *A,
int LDA,
int szeA,
688 int szeW,
float *result,
689 float *fake,
int szeF);
691 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
692 int norm, PLASMA_enum uplo,
int N,
693 const float *A,
int LDA,
int szeA,
694 int szeW,
float *result);
695 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
696 int norm, PLASMA_enum uplo,
int N,
697 const float *A,
int LDA,
int szeA,
698 int szeW,
float *result,
699 float *fake,
int szeF);
700 void QUARK_CORE_slaset(Quark *quark, Quark_Task_Flags *task_flags,
701 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
702 float beta,
float *tileA,
int ldtilea);
703 void QUARK_CORE_slaset2(Quark *quark, Quark_Task_Flags *task_flags,
704 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
705 float *tileA,
int ldtilea);
706 void QUARK_CORE_slaswp(Quark *quark, Quark_Task_Flags *task_flags,
707 int n,
float *A,
int lda,
708 int i1,
int i2,
const int *ipiv,
int inc);
709 void QUARK_CORE_slaswp_f2(Quark *quark, Quark_Task_Flags *task_flags,
710 int n,
float *A,
int lda,
711 int i1,
int i2,
const int *ipiv,
int inc,
712 float *fake1,
int szefake1,
int flag1,
713 float *fake2,
int szefake2,
int flag2);
714 void QUARK_CORE_slaswp_ontile(Quark *quark, Quark_Task_Flags *task_flags,
716 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
717 void QUARK_CORE_slaswp_ontile_f2(Quark *quark, Quark_Task_Flags *task_flags,
719 int i1,
int i2,
const int *ipiv,
int inc,
720 float *fake1,
int szefake1,
int flag1,
721 float *fake2,
int szefake2,
int flag2);
722 void QUARK_CORE_slaswpc_ontile(Quark *quark, Quark_Task_Flags *task_flags,
724 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
725 void QUARK_CORE_slatro(Quark *quark, Quark_Task_Flags *task_flags,
726 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
727 const float *A,
int lda,
729 void QUARK_CORE_slatro_f1(Quark *quark, Quark_Task_Flags *task_flags,
730 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
731 const float *A,
int lda,
733 float *fake1,
int szefake1,
int flag1);
734 void QUARK_CORE_slauum(Quark *quark, Quark_Task_Flags *task_flags,
735 PLASMA_enum uplo,
int n,
int nb,
737 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
738 float bump,
int m,
int n,
float *A,
int lda,
739 int bigM,
int m0,
int n0,
unsigned long long int seed );
740 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
741 float bump,
int m,
int n,
float *A,
int lda,
742 int bigM,
int m0,
int n0,
unsigned long long int seed );
743 void QUARK_CORE_splrnt(Quark *quark, Quark_Task_Flags *task_flags,
744 int m,
int n,
float *A,
int lda,
745 int bigM,
int m0,
int n0,
unsigned long long int seed );
746 void QUARK_CORE_spltmg(Quark *quark, Quark_Task_Flags *task_flags,
747 PLASMA_enum mtxtype,
int m,
int n,
float *A,
int lda,
748 int gM,
int gN,
int m0,
int n0,
unsigned long long int seed );
749 void QUARK_CORE_spltmg_chebvand( Quark *quark, Quark_Task_Flags *task_flags,
750 int M,
int N,
float *A,
int LDA,
751 int gN,
int m0,
int n0,
753 void QUARK_CORE_spltmg_circul( Quark *quark, Quark_Task_Flags *task_flags,
754 int M,
int N,
float *A,
int LDA,
755 int gM,
int m0,
int n0,
757 void QUARK_CORE_spltmg_fiedler(Quark *quark, Quark_Task_Flags *task_flags,
759 const float *X,
int incX,
760 const float *Y,
int incY,
762 void QUARK_CORE_spltmg_hankel( Quark *quark, Quark_Task_Flags *task_flags,
763 PLASMA_enum uplo,
int M,
int N,
float *A,
int LDA,
764 int m0,
int n0,
int nb,
767 void QUARK_CORE_spltmg_toeppd1(Quark *quark, Quark_Task_Flags *task_flags,
768 int gM,
int m0,
int M,
770 unsigned long long int seed);
771 void QUARK_CORE_spltmg_toeppd2(Quark *quark, Quark_Task_Flags *task_flags,
772 int M,
int N,
int K,
int m0,
int n0,
775 void QUARK_CORE_spotrf(Quark *quark, Quark_Task_Flags *task_flags,
776 PLASMA_enum uplo,
int n,
int nb,
780 void QUARK_CORE_ssetvar(Quark *quark, Quark_Task_Flags *task_flags,
781 const float *alpha,
float *x,
783 void QUARK_CORE_sshift( Quark *quark, Quark_Task_Flags *task_flags,
784 int s,
int m,
int n,
int L,
786 void QUARK_CORE_sshiftw(Quark *quark, Quark_Task_Flags *task_flags,
787 int s,
int cl,
int m,
int n,
int L,
789 void QUARK_CORE_sssssm(Quark *quark, Quark_Task_Flags *task_flags,
790 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
793 const float *L1,
int ldl1,
794 const float *L2,
int ldl2,
796 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
797 PLASMA_enum side, PLASMA_enum uplo,
798 int m,
int n,
int nb,
799 float alpha,
const float *A,
int lda,
800 const float *B,
int ldb,
801 float beta,
float *C,
int ldc);
802 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
803 PLASMA_enum uplo, PLASMA_enum trans,
804 int n,
int k,
int nb,
805 float alpha,
const float *A,
int lda,
806 float beta,
float *C,
int ldc);
807 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
808 PLASMA_enum uplo, PLASMA_enum trans,
809 int n,
int k,
int nb,
810 float alpha,
const float *A,
int lda,
811 const float *B,
int LDB,
812 float beta,
float *C,
int ldc);
813 void QUARK_CORE_ssyssq_f1( Quark *quark, Quark_Task_Flags *task_flags,
814 PLASMA_enum uplo,
int n,
const float *A,
int lda,
815 float *scale,
float *sumsq,
816 float *fake,
int szeF,
int paramF );
817 void QUARK_CORE_sswpab(Quark *quark, Quark_Task_Flags *task_flags,
818 int i,
int n1,
int n2,
820 void QUARK_CORE_sswptr_ontile(Quark *quark, Quark_Task_Flags *task_flags,
822 int i1,
int i2,
const int *ipiv,
int inc,
823 const float *Akk,
int ldak);
824 void QUARK_CORE_strdalg1(Quark *quark, Quark_Task_Flags *task_flags,
831 int Vblksiz,
int wantz,
832 int i,
int sweepid,
int m,
int grsiz,
833 int *PCOL,
int *ACOL,
int *MCOL);
834 void QUARK_CORE_strmm(Quark *quark, Quark_Task_Flags *task_flags,
835 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
836 int m,
int n,
int nb,
837 float alpha,
const float *A,
int lda,
839 void QUARK_CORE_strmm_p2(Quark *quark, Quark_Task_Flags *task_flags,
840 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
841 int m,
int n,
int nb,
842 float alpha,
const float *A,
int lda,
844 void QUARK_CORE_strsm(Quark *quark, Quark_Task_Flags *task_flags,
845 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
846 int m,
int n,
int nb,
847 float alpha,
const float *A,
int lda,
849 void QUARK_CORE_strtri(Quark *quark, Quark_Task_Flags *task_flags,
850 PLASMA_enum uplo, PLASMA_enum diag,
int n,
int nb,
854 void QUARK_CORE_stslqt(Quark *quark, Quark_Task_Flags *task_flags,
855 int m,
int n,
int ib,
int nb,
859 void QUARK_CORE_stsmlq(Quark *quark, Quark_Task_Flags *task_flags,
860 PLASMA_enum side, PLASMA_enum trans,
861 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
864 const float *V,
int ldv,
865 const float *T,
int ldt);
866 void QUARK_CORE_stsmlq_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
867 PLASMA_enum side, PLASMA_enum trans,
868 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
871 const float *V,
int ldv,
872 const float *T,
int ldt);
873 void QUARK_CORE_stsmlq_corner(Quark *quark, Quark_Task_Flags *task_flags,
874 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
878 const float *V,
int ldv,
879 const float *T,
int ldt);
880 void QUARK_CORE_stsmqr(Quark *quark, Quark_Task_Flags *task_flags,
881 PLASMA_enum side, PLASMA_enum trans,
882 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
885 const float *V,
int ldv,
886 const float *T,
int ldt);
887 void QUARK_CORE_stsmqr_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
888 PLASMA_enum side, PLASMA_enum trans,
889 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
892 const float *V,
int ldv,
893 const float *T,
int ldt);
894 void QUARK_CORE_stsmqr_corner(Quark *quark, Quark_Task_Flags *task_flags,
895 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
899 const float *V,
int ldv,
900 const float *T,
int ldt);
901 void QUARK_CORE_stsqrt(Quark *quark, Quark_Task_Flags *task_flags,
902 int m,
int n,
int ib,
int nb,
906 void QUARK_CORE_ststrf(Quark *quark, Quark_Task_Flags *task_flags,
907 int m,
int n,
int ib,
int nb,
913 PLASMA_bool check_info,
int iinfo);
914 void QUARK_CORE_sttmqr(Quark *quark, Quark_Task_Flags *task_flags,
915 PLASMA_enum side, PLASMA_enum trans,
916 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
919 const float *V,
int ldv,
920 const float *T,
int ldt);
921 void QUARK_CORE_sttqrt(Quark *quark, Quark_Task_Flags *task_flags,
922 int m,
int n,
int ib,
int nb,
926 void QUARK_CORE_sttmlq(Quark *quark, Quark_Task_Flags *task_flags,
927 PLASMA_enum side, PLASMA_enum trans,
928 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
931 const float *V,
int ldv,
932 const float *T,
int ldt);
933 void QUARK_CORE_sttlqt(Quark *quark, Quark_Task_Flags *task_flags,
934 int m,
int n,
int ib,
int nb,
938 void QUARK_CORE_spamm(Quark *quark, Quark_Task_Flags *task_flags,
939 int op, PLASMA_enum side, PLASMA_enum storev,
940 int m,
int n,
int k,
int l,
941 const float *A1,
int lda1,
943 const float *V,
int ldv,
945 void QUARK_CORE_splssq( Quark *quark, Quark_Task_Flags *task_flags,
946 int m,
const float *A,
float *result );
947 void QUARK_CORE_sormlq(Quark *quark, Quark_Task_Flags *task_flags,
948 PLASMA_enum side, PLASMA_enum trans,
949 int m,
int n,
int ib,
int nb,
int k,
950 const float *A,
int lda,
951 const float *T,
int ldt,
953 void QUARK_CORE_sormqr(Quark *quark, Quark_Task_Flags *task_flags,
954 PLASMA_enum side, PLASMA_enum trans,
955 int m,
int n,
int k,
int ib,
int nb,
956 const float *A,
int lda,
957 const float *T,
int ldt,
963 void CORE_sasum_quark(Quark *quark);
964 void CORE_sasum_f1_quark(Quark *quark);
965 void CORE_sgeadd_quark(Quark *quark);
966 void CORE_sbrdalg1_quark(Quark *quark);
967 void CORE_sgelqt_quark(Quark *quark);
968 void CORE_sgemm_quark(Quark *quark);
969 void CORE_sgemm_tile_quark(Quark *quark);
970 void CORE_sgemv_quark(Quark *quark);
971 void CORE_sgemv_tile_quark(Quark *quark);
972 void CORE_sgeqp3_init_quark(Quark *quark);
973 void CORE_sgeqp3_larfg_quark(Quark *quark);
974 void CORE_sgeqp3_norms_quark(Quark *quark);
975 void CORE_sgeqp3_pivot_quark(Quark *quark);
976 void CORE_sgeqp3_tntpiv_quark(Quark *quark);
977 void CORE_sgeqp3_update_quark(Quark *quark);
978 void CORE_sgeqrt_quark(Quark *quark);
979 void CORE_sgessm_quark(Quark *quark);
980 void CORE_sgessq_quark(Quark *quark);
981 void CORE_sgessq_f1_quark(Quark *quark);
982 void CORE_sgetrf_quark(Quark *quark);
983 void CORE_sgetrf_incpiv_quark(Quark *quark);
984 void CORE_sgetrf_nopiv_quark(Quark* quark);
985 void CORE_sgetrf_reclap_quark(Quark *quark);
986 void CORE_sgetrf_rectil_quark(Quark* quark);
987 void CORE_sgetrip_quark(Quark *quark);
988 void CORE_sgetrip_f1_quark(Quark *quark);
989 void CORE_sgetrip_f2_quark(Quark *quark);
991 void CORE_ssymm_quark(Quark *quark);
992 void CORE_ssyrk_quark(Quark *quark);
993 void CORE_ssyr2k_quark(Quark *quark);
995 void CORE_ssygst_quark(Quark *quark);
996 void CORE_ssyrfb_quark(Quark *quark);
997 void CORE_shessq_quark(Quark *quark);
998 void CORE_shessq_f1_quark(Quark *quark);
999 void CORE_slacpy_quark(Quark *quark);
1000 void CORE_slacpy_f1_quark(Quark *quark);
1001 void CORE_slacpy_pivot_quark(Quark *quark);
1002 void CORE_slatro_quark(Quark *quark);
1003 void CORE_slatro_f1_quark(Quark *quark);
1004 void CORE_slange_quark(Quark *quark);
1005 void CORE_slange_f1_quark(Quark *quark);
1007 void CORE_slansy_quark(Quark *quark);
1008 void CORE_slansy_f1_quark(Quark *quark);
1010 void CORE_slansy_quark(Quark *quark);
1011 void CORE_slansy_f1_quark(Quark *quark);
1012 void CORE_slaset_quark(Quark *quark);
1013 void CORE_slaset2_quark(Quark *quark);
1014 void CORE_slatro_quark(Quark *quark);
1015 void CORE_slauum_quark(Quark *quark);
1016 void CORE_spamm_quark(Quark *quark);
1017 void CORE_splgsy_quark(Quark *quark);
1018 void CORE_splgsy_quark(Quark *quark);
1019 void CORE_splrnt_quark(Quark *quark);
1020 void CORE_spltmg_quark(Quark *quark);
1021 void CORE_splssq_quark(Quark *quark);
1022 void CORE_spotrf_quark(Quark *quark);
1023 void CORE_ssetvar_quark(Quark *quark);
1024 void CORE_sshift_quark(Quark *quark);
1025 void CORE_sshiftw_quark(Quark *quark);
1026 void CORE_sssssm_quark(Quark *quark);
1027 void CORE_ssymm_quark(Quark *quark);
1028 void CORE_ssyrk_quark(Quark *quark);
1029 void CORE_ssyr2k_quark(Quark *quark);
1030 void CORE_ssyssq_quark(Quark *quark);
1031 void CORE_ssyssq_f1_quark(Quark *quark);
1032 void CORE_sswpab_quark(Quark *quark);
1033 void CORE_sswptr_ontile_quark(Quark *quark);
1034 void CORE_strdalg1_quark(Quark *quark);
1035 void CORE_strmm_quark(Quark *quark);
1036 void CORE_strsm_quark(Quark *quark);
1037 void CORE_strtri_quark(Quark *quark);
1038 void CORE_stslqt_quark(Quark *quark);
1039 void CORE_stsmlq_quark(Quark *quark);
1040 void CORE_stsmlq_sytra1_quark(Quark *quark);
1041 void CORE_stsmlq_corner_quark(Quark *quark);
1042 void CORE_stsmqr_quark(Quark *quark);
1043 void CORE_stsmqr_sytra1_quark(Quark *quark);
1044 void CORE_stsmqr_corner_quark(Quark *quark);
1045 void CORE_stsqrt_quark(Quark *quark);
1046 void CORE_ststrf_quark(Quark *quark);
1047 void CORE_sttmqr_quark(Quark *quark);
1048 void CORE_sttqrt_quark(Quark *quark);
1049 void CORE_sttmlq_quark(Quark *quark);
1050 void CORE_sttlqt_quark(Quark *quark);
1051 void CORE_sormlq_quark(Quark *quark);
1052 void CORE_sormqr_quark(Quark *quark);
1053 void CORE_slaswp_quark(Quark* quark);
1054 void CORE_slaswp_f2_quark(Quark* quark);
1055 void CORE_slaswp_ontile_quark(Quark *quark);
1056 void CORE_slaswp_ontile_f2_quark(Quark *quark);
1057 void CORE_slaswpc_ontile_quark(Quark *quark);
1058 void CORE_strmm_p2_quark(Quark* quark);
1059 void CORE_sgemm_f2_quark(Quark* quark);
1060 void CORE_sgemm_p2_quark(Quark* quark);
1061 void CORE_sgemm_p2f1_quark(Quark* quark);
1062 void CORE_sgemm_p3_quark(Quark* quark);