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_sbrdalg(PLASMA_enum uplo,
int N,
int NB,
34 int i,
int j,
int m,
int grsiz);
35 int CORE_sgbelr(PLASMA_enum uplo,
int N,
37 int st,
int ed,
int eltsize);
38 int CORE_sgbrce(PLASMA_enum uplo,
int N,
40 int st,
int ed,
int eltsize);
41 int CORE_sgblrx(PLASMA_enum uplo,
int N,
43 int st,
int ed,
int eltsize);
44 int CORE_sgeadd(
int M,
int N,
float alpha,
45 const float *A,
int LDA,
47 int CORE_sgelqt(
int M,
int N,
int IB,
52 void CORE_sgemm(PLASMA_enum transA, PLASMA_enum transB,
54 float alpha,
const float *A,
int LDA,
55 const float *B,
int LDB,
56 float beta,
float *C,
int LDC);
57 int CORE_sgeqrt(
int M,
int N,
int IB,
60 float *TAU,
float *WORK);
61 int CORE_sgessm(
int M,
int N,
int K,
int IB,
63 const float *L,
int LDL,
65 int CORE_sgessq(
int M,
int N,
66 const float *A,
int LDA,
67 float *scale,
float *sumsq);
68 int CORE_sgetrf(
int M,
int N,
70 int *IPIV,
int *INFO);
71 int CORE_sgetrf_incpiv(
int M,
int N,
int IB,
73 int *IPIV,
int *INFO);
74 int CORE_sgetrf_reclap(
int M,
int N,
76 int *IPIV,
int *info);
77 int CORE_sgetrf_rectil(
const PLASMA_desc A,
int *IPIV,
int *info);
78 void CORE_sgetrip(
int m,
int n,
float *A,
80 int CORE_shbelr(PLASMA_enum uplo,
int N,
82 int st,
int ed,
int eltsize);
83 int CORE_shblrx(PLASMA_enum uplo,
int N,
85 int st,
int ed,
int eltsize);
86 int CORE_shbrce(PLASMA_enum uplo,
int N,
88 int st,
int ed,
int eltsize);
89 void CORE_ssbtype1cb(
int N,
int NB,
92 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
94 void CORE_ssbtype2cb(
int N,
int NB,
97 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
99 void CORE_ssbtype3cb(
int N,
int NB,
101 const float *V,
const float *TAU,
102 int st,
int ed,
int sweep,
int Vblksiz,
int WANTZ,
104 void CORE_ssygst(
int itype, PLASMA_enum uplo,
int N,
106 float *B,
int LDB,
int *INFO);
108 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
110 float alpha,
const float *A,
int LDA,
111 const float *B,
int LDB,
112 float beta,
float *C,
int LDC);
113 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
115 float alpha,
const float *A,
int LDA,
116 float beta,
float *C,
int LDC);
117 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
119 float alpha,
const float *A,
int LDA,
120 const float *B,
int LDB,
121 float beta,
float *C,
int LDC);
122 int CORE_shessq(PLASMA_enum uplo,
int N,
123 const float *A,
int LDA,
124 float *scale,
float *sumsq);
126 int CORE_ssyrfb(PLASMA_enum uplo,
int N,
int K,
int IB,
int NB,
127 const float *A,
int LDA,
128 const float *T,
int LDT,
130 float *WORK,
int LDWORK);
131 void CORE_slacpy(PLASMA_enum uplo,
int M,
int N,
132 const float *A,
int LDA,
134 void CORE_slange(
int norm,
int M,
int N,
135 const float *A,
int LDA,
136 float *work,
float *normA);
138 void CORE_slansy(
int norm, PLASMA_enum uplo,
int N,
139 const float *A,
int LDA,
140 float *work,
float *normA);
142 void CORE_slansy(
int norm, PLASMA_enum uplo,
int N,
143 const float *A,
int LDA,
144 float *work,
float *normA);
145 int CORE_slarfb_gemm(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
147 const float *V,
int LDV,
148 const float *T,
int LDT,
150 float *WORK,
int LDWORK);
151 int CORE_slarfx2(PLASMA_enum side,
int N,
155 float *C2,
int LDC2);
156 int CORE_slarfx2c(PLASMA_enum uplo,
162 int CORE_slarfx2ce(PLASMA_enum uplo,
168 void CORE_slarfy(
int N,
173 void CORE_slaset(PLASMA_enum uplo,
int n1,
int n2,
174 float alpha,
float beta,
175 float *tileA,
int ldtilea);
176 void CORE_slaset2(PLASMA_enum uplo,
int n1,
int n2,
float alpha,
177 float *tileA,
int ldtilea);
178 void CORE_slaswp(
int N,
float *A,
int LDA,
179 int I1,
int I2,
const int *IPIV,
int INC);
180 int CORE_slaswp_ontile(
PLASMA_desc descA,
int i1,
int i2,
const int *ipiv,
int inc);
181 int CORE_slaswpc_ontile(
PLASMA_desc descA,
int i1,
int i2,
const int *ipiv,
int inc);
182 int CORE_slatro(PLASMA_enum uplo, PLASMA_enum trans,
184 const float *A,
int LDA,
186 void CORE_slauum(PLASMA_enum uplo,
int N,
float *A,
int LDA);
187 int CORE_spamm(
int op, PLASMA_enum side, PLASMA_enum storev,
188 int M,
int N,
int K,
int L,
189 const float *A1,
int LDA1,
191 const float *V,
int LDV,
193 int CORE_sparfb(PLASMA_enum side, PLASMA_enum trans, PLASMA_enum direct, PLASMA_enum storev,
194 int M1,
int N1,
int M2,
int N2,
int K,
int L,
197 const float *V,
int LDV,
198 const float *T,
int LDT,
199 float *WORK,
int LDWORK);
200 int CORE_spemv(PLASMA_enum trans, PLASMA_enum storev,
203 const float *A,
int LDA,
204 const float *X,
int INCX,
208 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
209 int bigM,
int m0,
int n0,
unsigned long long int seed );
210 void CORE_splgsy(
float bump,
int m,
int n,
float *A,
int lda,
211 int bigM,
int m0,
int n0,
unsigned long long int seed );
212 void CORE_splrnt(
int m,
int n,
float *A,
int lda,
213 int bigM,
int m0,
int n0,
unsigned long long int seed );
214 void CORE_spotrf(PLASMA_enum uplo,
int N,
float *A,
int LDA,
int *INFO);
215 void CORE_sshift(
int s,
int m,
int n,
int L,
217 void CORE_sshiftw(
int s,
int cl,
int m,
int n,
int L,
219 int CORE_sssssm(
int M1,
int N1,
int M2,
int N2,
int K,
int IB,
222 const float *L1,
int LDL1,
223 const float *L2,
int LDL2,
225 void CORE_ssymm(PLASMA_enum side, PLASMA_enum uplo,
227 float alpha,
const float *A,
int LDA,
228 const float *B,
int LDB,
229 float beta,
float *C,
int LDC);
230 void CORE_ssyrk(PLASMA_enum uplo, PLASMA_enum trans,
232 float alpha,
const float *A,
int LDA,
233 float beta,
float *C,
int LDC);
234 void CORE_ssyr2k(PLASMA_enum uplo, PLASMA_enum trans,
236 float alpha,
const float *A,
int LDA,
237 const float *B,
int LDB,
238 float beta,
float *C,
int LDC);
239 int CORE_ssyssq(PLASMA_enum uplo,
int N,
240 const float *A,
int LDA,
241 float *scale,
float *sumsq);
242 void CORE_sswpab(
int i,
int n1,
int n2,
243 float *A,
float *work);
244 int CORE_sswptr_ontile(
PLASMA_desc descA,
int i1,
int i2,
const int *ipiv,
int inc,
245 const float *Akk,
int ldak);
246 void CORE_strdalg(PLASMA_enum uplo,
int N,
int NB,
248 int i,
int j,
int m,
int grsiz);
249 void CORE_strmm(PLASMA_enum side, PLASMA_enum uplo,
250 PLASMA_enum transA, PLASMA_enum diag,
252 float alpha,
const float *A,
int LDA,
254 void CORE_strsm(PLASMA_enum side, PLASMA_enum uplo,
255 PLASMA_enum transA, PLASMA_enum diag,
257 float alpha,
const float *A,
int LDA,
259 void CORE_strtri(PLASMA_enum uplo, PLASMA_enum diag,
int N,
260 float *A,
int LDA,
int *info);
261 int CORE_stslqt(
int M,
int N,
int IB,
265 float *TAU,
float *WORK);
266 int CORE_stsmlq(PLASMA_enum side, PLASMA_enum trans,
267 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
270 const float *V,
int LDV,
271 const float *T,
int LDT,
272 float *WORK,
int LDWORK);
273 int CORE_stsmlq_corner(
int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
274 int k,
int ib,
int nb,
278 const float *V,
int ldv,
279 const float *T,
int ldt,
280 float *WORK,
int ldwork);
281 int CORE_stsmlq_sytra1( PLASMA_enum side, PLASMA_enum trans,
282 int m1,
int n1,
int m2,
int n2,
286 const float *V,
int ldv,
287 const float *T,
int ldt,
288 float *WORK,
int ldwork);
289 int CORE_stsmqr(PLASMA_enum side, PLASMA_enum trans,
290 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
293 const float *V,
int LDV,
294 const float *T,
int LDT,
295 float *WORK,
int LDWORK);
296 int CORE_stsmqr_corner(
int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
297 int k,
int ib,
int nb,
301 const float *V,
int ldv,
302 const float *T,
int ldt,
303 float *WORK,
int ldwork);
304 int CORE_stsmqr_sytra1( PLASMA_enum side, PLASMA_enum trans,
305 int m1,
int n1,
int m2,
int n2,
309 const float *V,
int ldv,
310 const float *T,
int ldt,
311 float *WORK,
int ldwork);
312 int CORE_stsqrt(
int M,
int N,
int IB,
316 float *TAU,
float *WORK);
317 int CORE_ststrf(
int M,
int N,
int IB,
int NB,
321 int *IPIV,
float *WORK,
322 int LDWORK,
int *INFO);
323 int CORE_sttmqr(PLASMA_enum side, PLASMA_enum trans,
324 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
327 const float *V,
int LDV,
328 const float *T,
int LDT,
329 float *WORK,
int LDWORK);
330 int CORE_sttqrt(
int M,
int N,
int IB,
336 int CORE_sttmlq(PLASMA_enum side, PLASMA_enum trans,
337 int M1,
int N1,
int M2,
int N2,
int K,
int IB,
340 const float *V,
int LDV,
341 const float *T,
int LDT,
342 float *WORK,
int LDWORK);
343 int CORE_sttlqt(
int M,
int N,
int IB,
349 int CORE_sormlq(PLASMA_enum side, PLASMA_enum trans,
350 int M,
int N,
int IB,
int K,
351 const float *V,
int LDV,
352 const float *T,
int LDT,
354 float *WORK,
int LDWORK);
355 int CORE_sormqr(PLASMA_enum side, PLASMA_enum trans,
356 int M,
int N,
int K,
int IB,
357 const float *V,
int LDV,
358 const float *T,
int LDT,
360 float *WORK,
int LDWORK);
365 void QUARK_CORE_sasum(Quark *quark, Quark_Task_Flags *task_flags,
366 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
367 const float *A,
int lda,
int szeA,
368 float *work,
int szeW);
369 void QUARK_CORE_sasum_f1(Quark *quark, Quark_Task_Flags *task_flags,
370 PLASMA_enum storev, PLASMA_enum uplo,
int m,
int n,
371 const float *A,
int lda,
int szeA,
372 float *work,
int szeW,
373 float *fake,
int szeF);
374 void QUARK_CORE_sgeadd(Quark *quark, Quark_Task_Flags *task_flags,
375 int m,
int n,
int nb,
float alpha,
376 const float *A,
int lda,
378 void QUARK_CORE_sbrdalg(Quark *quark, Quark_Task_Flags *task_flags,
384 int i,
int j,
int m,
int grsiz,
int BAND,
385 const int *PCOL,
const int *ACOL,
int *MCOL);
386 void QUARK_CORE_sgelqt(Quark *quark, Quark_Task_Flags *task_flags,
387 int m,
int n,
int ib,
int nb,
390 void QUARK_CORE_sgemm(Quark *quark, Quark_Task_Flags *task_flags,
391 PLASMA_enum transA, PLASMA_enum transB,
392 int m,
int n,
int k,
int nb,
393 float alpha,
const float *A,
int lda,
394 const float *B,
int ldb,
395 float beta,
float *C,
int ldc);
396 void QUARK_CORE_sgemm2( Quark *quark, Quark_Task_Flags *task_flags,
397 PLASMA_enum transA, PLASMA_enum transB,
398 int m,
int n,
int k,
int nb,
399 float alpha,
const float *A,
int lda,
400 const float *B,
int ldb,
401 float beta,
float *C,
int ldc);
402 void QUARK_CORE_sgemm_f2(Quark *quark, Quark_Task_Flags *task_flags,
403 PLASMA_enum transA, PLASMA_enum transB,
404 int m,
int n,
int k,
int nb,
405 float alpha,
const float *A,
int lda,
406 const float *B,
int ldb,
407 float beta,
float *C,
int ldc,
408 float *fake1,
int szefake1,
int flag1,
409 float *fake2,
int szefake2,
int flag2);
410 void QUARK_CORE_sgemm_p2(Quark *quark, Quark_Task_Flags *task_flags,
411 PLASMA_enum transA, PLASMA_enum transB,
412 int m,
int n,
int k,
int nb,
413 float alpha,
const float *A,
int lda,
414 const float **B,
int ldb,
415 float beta,
float *C,
int ldc);
416 void QUARK_CORE_sgemm_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
417 PLASMA_enum transA, PLASMA_enum transB,
418 int m,
int n,
int k,
int nb,
419 float alpha,
const float *A,
int lda,
420 const float **B,
int ldb,
421 float beta,
float *C,
int ldc,
422 float *fake1,
int szefake1,
int flag1);
423 void QUARK_CORE_sgemm_p3(Quark *quark, Quark_Task_Flags *task_flags,
424 PLASMA_enum transA, PLASMA_enum transB,
425 int m,
int n,
int k,
int nb,
426 float alpha,
const float *A,
int lda,
427 const float *B,
int ldb,
428 float beta,
float **C,
int ldc);
429 void QUARK_CORE_sgeqrt(Quark *quark, Quark_Task_Flags *task_flags,
430 int m,
int n,
int ib,
int nb,
433 void QUARK_CORE_sgessm(Quark *quark, Quark_Task_Flags *task_flags,
434 int m,
int n,
int k,
int ib,
int nb,
436 const float *L,
int ldl,
438 void QUARK_CORE_sgessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
439 int m,
int n,
const float *A,
int lda,
440 float *scale,
float *sumsq,
441 float *fake,
int szeF,
int paramF );
442 void QUARK_CORE_sgetrf(Quark *quark, Quark_Task_Flags *task_flags,
443 int m,
int n,
int nb,
447 PLASMA_bool check_info,
int iinfo);
448 void QUARK_CORE_sgetrf_incpiv(Quark *quark, Quark_Task_Flags *task_flags,
449 int m,
int n,
int ib,
int nb,
453 PLASMA_bool check_info,
int iinfo);
454 void QUARK_CORE_sgetrf_reclap(Quark *quark, Quark_Task_Flags *task_flags,
455 int m,
int n,
int nb,
459 PLASMA_bool check_info,
int iinfo,
461 void QUARK_CORE_sgetrf_rectil(Quark *quark, Quark_Task_Flags *task_flags,
465 PLASMA_bool check_info,
int iinfo,
467 void QUARK_CORE_sgetrip(Quark *quark, Quark_Task_Flags *task_flags,
468 int m,
int n,
float *A,
int szeA);
469 void QUARK_CORE_sgetrip_f1(Quark *quark, Quark_Task_Flags *task_flags,
470 int m,
int n,
float *A,
int szeA,
471 float *fake,
int szeF,
int paramF);
472 void QUARK_CORE_sgetrip_f2(Quark *quark, Quark_Task_Flags *task_flags,
473 int m,
int n,
float *A,
int szeA,
474 float *fake1,
int szeF1,
int paramF1,
475 float *fake2,
int szeF2,
int paramF2);
476 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
477 PLASMA_enum side, PLASMA_enum uplo,
478 int m,
int n,
int nb,
479 float alpha,
const float *A,
int lda,
480 const float *B,
int ldb,
481 float beta,
float *C,
int ldc);
482 void QUARK_CORE_ssygst(Quark *quark, Quark_Task_Flags *task_flags,
483 int itype, PLASMA_enum uplo,
int N,
488 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
489 PLASMA_enum uplo, PLASMA_enum trans,
490 int n,
int k,
int nb,
491 float alpha,
const float *A,
int lda,
492 float beta,
float *C,
int ldc);
493 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
494 PLASMA_enum uplo, PLASMA_enum trans,
495 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_ssyrfb(Quark *quark, Quark_Task_Flags *task_flags,
501 int n,
int k,
int ib,
int nb,
502 const float *A,
int lda,
503 const float *T,
int ldt,
505 void QUARK_CORE_shessq_f1( Quark *quark, Quark_Task_Flags *task_flags,
506 PLASMA_enum uplo,
int n,
const float *A,
int lda,
507 float *scale,
float *sumsq,
508 float *fake,
int szeF,
int paramF );
509 void QUARK_CORE_slacpy(Quark *quark, Quark_Task_Flags *task_flags,
510 PLASMA_enum uplo,
int m,
int n,
int mb,
511 const float *A,
int lda,
513 void QUARK_CORE_slange(Quark *quark, Quark_Task_Flags *task_flags,
514 int norm,
int M,
int N,
515 const float *A,
int LDA,
int szeA,
516 int szeW,
float *result);
517 void QUARK_CORE_slange_f1(Quark *quark, Quark_Task_Flags *task_flags,
518 int norm,
int M,
int N,
519 const float *A,
int LDA,
int szeA,
520 int szeW,
float *result,
521 float *fake,
int szeF);
523 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
524 int norm, PLASMA_enum uplo,
int N,
525 const float *A,
int LDA,
int szeA,
526 int szeW,
float *result);
527 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
528 int norm, PLASMA_enum uplo,
int N,
529 const float *A,
int LDA,
int szeA,
530 int szeW,
float *result,
531 float *fake,
int szeF);
533 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
534 int norm, PLASMA_enum uplo,
int N,
535 const float *A,
int LDA,
int szeA,
536 int szeW,
float *result);
537 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
538 int norm, PLASMA_enum uplo,
int N,
539 const float *A,
int LDA,
int szeA,
540 int szeW,
float *result,
541 float *fake,
int szeF);
542 void QUARK_CORE_slaset(Quark *quark, Quark_Task_Flags *task_flags,
543 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
544 float beta,
float *tileA,
int ldtilea);
545 void QUARK_CORE_slaset2(Quark *quark, Quark_Task_Flags *task_flags,
546 PLASMA_enum uplo,
int n1,
int n2,
float alpha,
547 float *tileA,
int ldtilea);
548 void QUARK_CORE_slaswp(Quark *quark, Quark_Task_Flags *task_flags,
549 int n,
float *A,
int lda,
550 int i1,
int i2,
const int *ipiv,
int inc);
551 void QUARK_CORE_slaswp_f2(Quark *quark, Quark_Task_Flags *task_flags,
552 int n,
float *A,
int lda,
553 int i1,
int i2,
const int *ipiv,
int inc,
554 float *fake1,
int szefake1,
int flag1,
555 float *fake2,
int szefake2,
int flag2);
556 void QUARK_CORE_slaswp_ontile(Quark *quark, Quark_Task_Flags *task_flags,
558 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
559 void QUARK_CORE_slaswp_ontile_f2(Quark *quark, Quark_Task_Flags *task_flags,
561 int i1,
int i2,
const int *ipiv,
int inc,
562 float *fake1,
int szefake1,
int flag1,
563 float *fake2,
int szefake2,
int flag2);
564 void QUARK_CORE_slaswpc_ontile(Quark *quark, Quark_Task_Flags *task_flags,
566 int i1,
int i2,
const int *ipiv,
int inc,
float *fakepanel);
567 void QUARK_CORE_slatro(Quark *quark, Quark_Task_Flags *task_flags,
568 PLASMA_enum uplo, PLASMA_enum trans,
int m,
int n,
int mb,
569 const float *A,
int lda,
571 void QUARK_CORE_slauum(Quark *quark, Quark_Task_Flags *task_flags,
572 PLASMA_enum uplo,
int n,
int nb,
574 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
575 float bump,
int m,
int n,
float *A,
int lda,
576 int bigM,
int m0,
int n0,
unsigned long long int seed );
577 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
578 float bump,
int m,
int n,
float *A,
int lda,
579 int bigM,
int m0,
int n0,
unsigned long long int seed );
580 void QUARK_CORE_splrnt(Quark *quark, Quark_Task_Flags *task_flags,
581 int m,
int n,
float *A,
int lda,
582 int bigM,
int m0,
int n0,
unsigned long long int seed );
583 void QUARK_CORE_spotrf(Quark *quark, Quark_Task_Flags *task_flags,
584 PLASMA_enum uplo,
int n,
int nb,
588 void QUARK_CORE_sshift( Quark *quark, Quark_Task_Flags *task_flags,
589 int s,
int m,
int n,
int L,
591 void QUARK_CORE_sshiftw(Quark *quark, Quark_Task_Flags *task_flags,
592 int s,
int cl,
int m,
int n,
int L,
594 void QUARK_CORE_sssssm(Quark *quark, Quark_Task_Flags *task_flags,
595 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
598 const float *L1,
int ldl1,
599 const float *L2,
int ldl2,
601 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
602 PLASMA_enum side, PLASMA_enum uplo,
603 int m,
int n,
int nb,
604 float alpha,
const float *A,
int lda,
605 const float *B,
int ldb,
606 float beta,
float *C,
int ldc);
607 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
608 PLASMA_enum uplo, PLASMA_enum trans,
609 int n,
int k,
int nb,
610 float alpha,
const float *A,
int lda,
611 float beta,
float *C,
int ldc);
612 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
613 PLASMA_enum uplo, PLASMA_enum trans,
614 int n,
int k,
int nb,
615 float alpha,
const float *A,
int lda,
616 const float *B,
int LDB,
617 float beta,
float *C,
int ldc);
618 void QUARK_CORE_ssyssq_f1( Quark *quark, Quark_Task_Flags *task_flags,
619 PLASMA_enum uplo,
int n,
const float *A,
int lda,
620 float *scale,
float *sumsq,
621 float *fake,
int szeF,
int paramF );
622 void QUARK_CORE_sswpab(Quark *quark, Quark_Task_Flags *task_flags,
623 int i,
int n1,
int n2,
625 void QUARK_CORE_sswptr_ontile(Quark *quark, Quark_Task_Flags *task_flags,
627 int i1,
int i2,
const int *ipiv,
int inc,
628 const float *Akk,
int ldak);
629 void QUARK_CORE_strdalg(Quark *quark, Quark_Task_Flags *task_flags,
635 int i,
int j,
int m,
int grsiz,
int BAND,
636 const int *PCOL,
const int *ACOL,
int *MCOL);
637 void QUARK_CORE_strmm(Quark *quark, Quark_Task_Flags *task_flags,
638 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
639 int m,
int n,
int nb,
640 float alpha,
const float *A,
int lda,
642 void QUARK_CORE_strmm_p2(Quark *quark, Quark_Task_Flags *task_flags,
643 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
644 int m,
int n,
int nb,
645 float alpha,
const float *A,
int lda,
647 void QUARK_CORE_strsm(Quark *quark, Quark_Task_Flags *task_flags,
648 PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum transA, PLASMA_enum diag,
649 int m,
int n,
int nb,
650 float alpha,
const float *A,
int lda,
652 void QUARK_CORE_strtri(Quark *quark, Quark_Task_Flags *task_flags,
653 PLASMA_enum uplo, PLASMA_enum diag,
int n,
int nb,
657 void QUARK_CORE_stslqt(Quark *quark, Quark_Task_Flags *task_flags,
658 int m,
int n,
int ib,
int nb,
662 void QUARK_CORE_stsmlq(Quark *quark, Quark_Task_Flags *task_flags,
663 PLASMA_enum side, PLASMA_enum trans,
664 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
667 const float *V,
int ldv,
668 const float *T,
int ldt);
669 void QUARK_CORE_stsmlq_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
670 PLASMA_enum side, PLASMA_enum trans,
671 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
674 const float *V,
int ldv,
675 const float *T,
int ldt);
676 void QUARK_CORE_stsmlq_corner(Quark *quark, Quark_Task_Flags *task_flags,
677 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
681 const float *V,
int ldv,
682 const float *T,
int ldt);
683 void QUARK_CORE_stsmqr(Quark *quark, Quark_Task_Flags *task_flags,
684 PLASMA_enum side, PLASMA_enum trans,
685 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
688 const float *V,
int ldv,
689 const float *T,
int ldt);
690 void QUARK_CORE_stsmqr_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
691 PLASMA_enum side, PLASMA_enum trans,
692 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
695 const float *V,
int ldv,
696 const float *T,
int ldt);
697 void QUARK_CORE_stsmqr_corner(Quark *quark, Quark_Task_Flags *task_flags,
698 int m1,
int n1,
int m2,
int n2,
int m3,
int n3,
int k,
int ib,
int nb,
702 const float *V,
int ldv,
703 const float *T,
int ldt);
704 void QUARK_CORE_stsqrt(Quark *quark, Quark_Task_Flags *task_flags,
705 int m,
int n,
int ib,
int nb,
709 void QUARK_CORE_ststrf(Quark *quark, Quark_Task_Flags *task_flags,
710 int m,
int n,
int ib,
int nb,
716 PLASMA_bool check_info,
int iinfo);
717 void QUARK_CORE_sttmqr(Quark *quark, Quark_Task_Flags *task_flags,
718 PLASMA_enum side, PLASMA_enum trans,
719 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
722 const float *V,
int ldv,
723 const float *T,
int ldt);
724 void QUARK_CORE_sttqrt(Quark *quark, Quark_Task_Flags *task_flags,
725 int m,
int n,
int ib,
int nb,
729 void QUARK_CORE_sttmlq(Quark *quark, Quark_Task_Flags *task_flags,
730 PLASMA_enum side, PLASMA_enum trans,
731 int m1,
int n1,
int m2,
int n2,
int k,
int ib,
int nb,
734 const float *V,
int ldv,
735 const float *T,
int ldt);
736 void QUARK_CORE_sttlqt(Quark *quark, Quark_Task_Flags *task_flags,
737 int m,
int n,
int ib,
int nb,
741 void QUARK_CORE_spamm(Quark *quark, Quark_Task_Flags *task_flags,
742 int op, PLASMA_enum side, PLASMA_enum storev,
743 int m,
int n,
int k,
int l,
744 const float *A1,
int lda1,
746 const float *V,
int ldv,
748 void QUARK_CORE_splssq( Quark *quark, Quark_Task_Flags *task_flags,
749 int m,
const float *A,
float *result );
750 void QUARK_CORE_sormlq(Quark *quark, Quark_Task_Flags *task_flags,
751 PLASMA_enum side, PLASMA_enum trans,
752 int m,
int n,
int ib,
int nb,
int k,
753 const float *A,
int lda,
754 const float *T,
int ldt,
756 void QUARK_CORE_sormqr(Quark *quark, Quark_Task_Flags *task_flags,
757 PLASMA_enum side, PLASMA_enum trans,
758 int m,
int n,
int k,
int ib,
int nb,
759 const float *A,
int lda,
760 const float *T,
int ldt,
766 void CORE_sasum_quark(Quark *quark);
767 void CORE_sasum_f1_quark(Quark *quark);
768 void CORE_sgeadd_quark(Quark *quark);
769 void CORE_sbrdalg_quark(Quark *quark);
770 void CORE_sgelqt_quark(Quark *quark);
771 void CORE_sgemm_quark(Quark *quark);
772 void CORE_sgeqrt_quark(Quark *quark);
773 void CORE_sgessm_quark(Quark *quark);
774 void CORE_sgessq_quark(Quark *quark);
775 void CORE_sgessq_f1_quark(Quark *quark);
776 void CORE_sgetrf_quark(Quark *quark);
777 void CORE_sgetrf_incpiv_quark(Quark *quark);
778 void CORE_sgetrf_reclap_quark(Quark *quark);
779 void CORE_sgetrf_rectil_quark(Quark* quark);
780 void CORE_sgetrip_quark(Quark *quark);
781 void CORE_sgetrip_f1_quark(Quark *quark);
782 void CORE_sgetrip_f2_quark(Quark *quark);
784 void CORE_ssymm_quark(Quark *quark);
785 void CORE_ssyrk_quark(Quark *quark);
786 void CORE_ssyr2k_quark(Quark *quark);
788 void CORE_ssygst_quark(Quark *quark);
789 void CORE_ssyrfb_quark(Quark *quark);
790 void CORE_shessq_quark(Quark *quark);
791 void CORE_shessq_f1_quark(Quark *quark);
792 void CORE_slacpy_quark(Quark *quark);
793 void CORE_slatro_quark(Quark *quark);
794 void CORE_slange_quark(Quark *quark);
795 void CORE_slange_f1_quark(Quark *quark);
797 void CORE_slansy_quark(Quark *quark);
798 void CORE_slansy_f1_quark(Quark *quark);
800 void CORE_slansy_quark(Quark *quark);
801 void CORE_slansy_f1_quark(Quark *quark);
802 void CORE_slaset_quark(Quark *quark);
803 void CORE_slaset2_quark(Quark *quark);
804 void CORE_slatro_quark(Quark *quark);
805 void CORE_slauum_quark(Quark *quark);
806 void CORE_spamm_quark(Quark *quark);
807 void CORE_splgsy_quark(Quark *quark);
808 void CORE_splgsy_quark(Quark *quark);
809 void CORE_splrnt_quark(Quark *quark);
810 void CORE_splssq_quark(Quark *quark);
811 void CORE_spotrf_quark(Quark *quark);
812 void CORE_sshift_quark(Quark *quark);
813 void CORE_sshiftw_quark(Quark *quark);
814 void CORE_sssssm_quark(Quark *quark);
815 void CORE_ssymm_quark(Quark *quark);
816 void CORE_ssyrk_quark(Quark *quark);
817 void CORE_ssyr2k_quark(Quark *quark);
818 void CORE_ssyssq_quark(Quark *quark);
819 void CORE_ssyssq_f1_quark(Quark *quark);
820 void CORE_sswpab_quark(Quark *quark);
821 void CORE_sswptr_ontile_quark(Quark *quark);
822 void CORE_strdalg_quark(Quark *quark);
823 void CORE_strmm_quark(Quark *quark);
824 void CORE_strsm_quark(Quark *quark);
825 void CORE_strtri_quark(Quark *quark);
826 void CORE_stslqt_quark(Quark *quark);
827 void CORE_stsmlq_quark(Quark *quark);
828 void CORE_stsmlq_sytra1_quark(Quark *quark);
829 void CORE_stsmlq_corner_quark(Quark *quark);
830 void CORE_stsmqr_quark(Quark *quark);
831 void CORE_stsmqr_sytra1_quark(Quark *quark);
832 void CORE_stsmqr_corner_quark(Quark *quark);
833 void CORE_stsqrt_quark(Quark *quark);
834 void CORE_ststrf_quark(Quark *quark);
835 void CORE_sttmqr_quark(Quark *quark);
836 void CORE_sttqrt_quark(Quark *quark);
837 void CORE_sttmlq_quark(Quark *quark);
838 void CORE_sttlqt_quark(Quark *quark);
839 void CORE_sormlq_quark(Quark *quark);
840 void CORE_sormqr_quark(Quark *quark);
842 void CORE_slaswp_quark(Quark* quark);
843 void CORE_slaswp_f2_quark(Quark* quark);
844 void CORE_slaswp_ontile_quark(Quark *quark);
845 void CORE_slaswp_ontile_f2_quark(Quark *quark);
846 void CORE_slaswpc_ontile_quark(Quark *quark);
847 void CORE_strmm_p2_quark(Quark* quark);
848 void CORE_sgemm_f2_quark(Quark* quark);
849 void CORE_sgemm_p2_quark(Quark* quark);
850 void CORE_sgemm_p2f1_quark(Quark* quark);
851 void CORE_sgemm_p3_quark(Quark* quark);