PLASMA  2.4.5
PLASMA - Parallel Linear Algebra for Scalable Multi-core Architectures
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups
core_sblas.h
Go to the documentation of this file.
1 
18 #ifndef _PLASMA_CORE_SBLAS_H_
19 #define _PLASMA_CORE_SBLAS_H_
20 
21 #define REAL
22 
23 #ifdef __cplusplus
24 extern "C" {
25 #endif
26 
31 int CORE_slarfx2(int side, int N,
32  float V,
33  float TAU,
34  float *C1, int LDC1,
35  float *C2, int LDC2);
36 int CORE_slarfx2c(int uplo,
37  float V,
38  float TAU,
39  float *C1,
40  float *C2,
41  float *C3);
42 int CORE_slarfx2ce(int uplo,
43  float *V,
44  float *TAU,
45  float *C1,
46  float *C2,
47  float *C3);
48 int CORE_shbelr(int uplo, int N,
49  PLASMA_desc *A,
50  float *V,
51  float *TAU,
52  int st,
53  int ed,
54  int eltsize);
55 int CORE_shbrce(int uplo, int N,
56  PLASMA_desc *A,
57  float *V,
58  float *TAU,
59  int st,
60  int ed,
61  int eltsize);
62 int CORE_shblrx(int uplo, int N,
63  PLASMA_desc *A,
64  float *V,
65  float *TAU,
66  int st,
67  int ed,
68  int eltsize);
69 int CORE_sgbelr(int uplo, int N,
70  PLASMA_desc *A,
71  float *V,
72  float *TAU,
73  int st,
74  int ed,
75  int eltsize);
76 int CORE_sgbrce(int uplo, int N,
77  PLASMA_desc *A,
78  float *V,
79  float *TAU,
80  int st,
81  int ed,
82  int eltsize);
83 int CORE_sgblrx(int uplo, int N,
84  PLASMA_desc *A,
85  float *V,
86  float *TAU,
87  int st,
88  int ed,
89  int eltsize);
90 void CORE_sasum(int storev, int uplo, int M, int N,
91  float *A, int lda, float *work);
92 void CORE_sgeadd(int M, int N, float alpha,
93  float *A, int LDA,
94  float *B, int LDB);
95 void CORE_sbrdalg(PLASMA_enum uplo, int N, int NB,
96  PLASMA_desc *pA, float *C, float *S,
97  int i, int j, int m, int grsiz);
98 int CORE_sgelqt(int M, int N, int IB,
99  float *A, int LDA,
100  float *T, int LDT,
101  float *TAU, float *WORK);
102 void CORE_sgemm(int transA, int transB,
103  int M, int N, int K,
104  float alpha, float *A, int LDA,
105  float *B, int LDB,
106  float beta, float *C, int LDC);
107 int CORE_sgeqrt(int M, int N, int IB,
108  float *A, int LDA,
109  float *T, int LDT,
110  float *TAU, float *WORK);
111 int CORE_sgessm(int M, int N, int K, int IB,
112  int *IPIV,
113  float *L, int LDL,
114  float *A, int LDA);
115 int CORE_sgetrf(int M, int N,
116  float *A, int LDA,
117  int *IPIV, int *INFO);
118 int CORE_sgetrf_incpiv(int M, int N, int IB,
119  float *A, int LDA,
120  int *IPIV, int *INFO);
121 int CORE_sgetrf_reclap(const int M, const int N,
122  float *A, const int LDA,
123  int *IPIV, int *info);
124 int CORE_sgetrf_rectil(const PLASMA_desc A, int *IPIV, int *info);
125 void CORE_sgetrip(int m, int n, float *A,
126  float *work);
127 #ifdef COMPLEX
128 void CORE_ssygst(int itype, int uplo, int N,
129  float *A, int LDA,
130  float *B, int LDB, int *INFO);
131 void CORE_ssymm(int side, int uplo,
132  int M, int N,
133  float alpha, float *A, int LDA,
134  float *B, int LDB,
135  float beta, float *C, int LDC);
136 void CORE_ssyrk(int uplo, int trans,
137  int N, int K,
138  float alpha, float *A, int LDA,
139  float beta, float *C, int LDC);
140 void CORE_ssyr2k(int uplo, int trans,
141  int N, int K,
142  float alpha, float *A, int LDA,
143  float *B, int LDB,
144  float beta, float *C, int LDC);
145 int CORE_ssyrfb(PLASMA_enum uplo, int N, int K, int IB, int NB,
146  float *A, int LDA,
147  float *T, int LDT,
148  float *C, int LDC,
149  float *WORK, int LDWORK);
150 #endif
151 void CORE_slacpy(PLASMA_enum uplo, int M, int N,
152  float *A, int LDA,
153  float *B, int LDB);
154 void CORE_slange(int norm, int M, int N,
155  float *A, int LDA,
156  float *work, float *normA);
157 #ifdef COMPLEX
158 void CORE_slansy(int norm, int uplo, int N,
159  float *A, int LDA,
160  float *work, float *normA);
161 #endif
162 void CORE_slansy(int norm, int uplo, int N,
163  float *A, int LDA,
164  float *work, float *normA);
165 void CORE_slaset(PLASMA_enum uplo, int n1, int n2, float alpha,
166  float beta, float *tileA, int ldtilea);
167 void CORE_slaset2(PLASMA_enum uplo, int n1, int n2, float alpha,
168  float *tileA, int ldtilea);
169 void CORE_slaswp(int N, float *A, int LDA,
170  int I1, int I2, int *IPIV, int INC);
171 int CORE_slaswp_ontile(PLASMA_desc descA, int i1, int i2, int *ipiv, int inc);
172 int CORE_slaswpc_ontile(PLASMA_desc descA, int i1, int i2, int *ipiv, int inc);
173 void CORE_slauum(int uplo, int N, float *A, int LDA);
174 int CORE_spamm(int op, int side, int storev,
175  int M, int N, int K, int L,
176  float *A1, int LDA1,
177  float *A2, int LDA2,
178  float *V, int LDV,
179  float *W, int LDW);
180 int CORE_sparfb(int side, int trans, int direct, int storev,
181  int M1, int N1, int M2, int N2, int K, int L,
182  float *A1, int LDA1,
183  float *A2, int LDA2,
184  float *V, int LDV,
185  float *T, int LDT,
186  float *WORK, int LDWORK);
187 int CORE_spemv(int trans, int storev,
188  int M, int N, int L,
189  float ALPHA,
190  float *A, int LDA,
191  float *X, int INCX,
192  float BETA,
193  float *Y, int INCY,
194  float *WORK);
195 void CORE_splgsy(float bump, int m, int n, float *A, int lda,
196  int bigM, int m0, int n0, unsigned long long int seed );
197 void CORE_splgsy(float bump, int m, int n, float *A, int lda,
198  int bigM, int m0, int n0, unsigned long long int seed );
199 void CORE_splrnt(int m, int n, float *A, int lda,
200  int bigM, int m0, int n0, unsigned long long int seed );
201 void CORE_spotrf(int uplo, int N, float *A, int LDA, int *INFO);
202 void CORE_sshift(int s, int m, int n, int L,
203  float *A);
204 void CORE_sshiftw(int s, int cl, int m, int n, int L,
205  float *A, float *W);
206 int CORE_sssssm(int M1, int N1, int M2, int N2, int K, int IB,
207  float *A1, int LDA1,
208  float *A2, int LDA2,
209  float *L1, int LDL1,
210  float *L2, int LDL2,
211  int *IPIV);
212 void CORE_ssymm(int side, int uplo,
213  int M, int N,
214  float alpha, float *A, int LDA,
215  float *B, int LDB,
216  float beta, float *C, int LDC);
217 void CORE_ssyrk(int uplo, int trans,
218  int N, int K,
219  float alpha, float *A, int LDA,
220  float beta, float *C, int LDC);
221 void CORE_ssyr2k(int uplo, int trans,
222  int N, int K,
223  float alpha, float *A, int LDA,
224  float *B, int LDB,
225  float beta, float *C, int LDC);
226 void CORE_sswpab(int i, int n1, int n2,
227  float *A, float *work);
228 int CORE_sswptr_ontile(PLASMA_desc descA, int i1, int i2, int *ipiv, int inc,
229  float *Akk, int ldak);
230 void CORE_strdalg(PLASMA_enum uplo, int N, int NB,
231  PLASMA_desc *pA, float *C, float *S,
232  int i, int j, int m, int grsiz);
233 void CORE_strmm(int side, int uplo,
234  int transA, int diag,
235  int M, int N,
236  float alpha, float *A, int LDA,
237  float *B, int LDB);
238 void CORE_strsm(int side, int uplo,
239  int transA, int diag,
240  int M, int N,
241  float alpha, float *A, int LDA,
242  float *B, int LDB);
243 void CORE_strtri(int uplo, int diag, int N, float *A, int LDA, int *info);
244 int CORE_stslqt(int M, int N, int IB,
245  float *A1, int LDA1,
246  float *A2, int LDA2,
247  float *T, int LDT,
248  float *TAU, float *WORK);
249 int CORE_stsmlq(int side, int trans,
250  int M1, int N1, int M2, int N2, int K, int IB,
251  float *A1, int LDA1,
252  float *A2, int LDA2,
253  float *V, int LDV,
254  float *T, int LDT,
255  float *WORK, int LDWORK);
256 int CORE_stsmlq_corner( int m1, int n1, int m2, int n2, int m3, int n3,
257  int k, int ib, int nb,
258  float *A1, int lda1,
259  float *A2, int lda2,
260  float *A3, int lda3,
261  float *V, int ldv,
262  float *T, int ldt,
263  float *WORK, int ldwork);
264 int CORE_stsmlq_sytra1( int side, int trans,
265  int m1, int n1, int m2, int n2,
266  int k, int ib,
267  float *A1, int lda1,
268  float *A2, int lda2,
269  float *V, int ldv,
270  float *T, int ldt,
271  float *WORK, int ldwork);
272 int CORE_stsmqr(int side, int trans,
273  int M1, int N1, int M2, int N2, int K, int IB,
274  float *A1, int LDA1,
275  float *A2, int LDA2,
276  float *V, int LDV,
277  float *T, int LDT,
278  float *WORK, int LDWORK);
279 int CORE_stsmqr_corner( int m1, int n1, int m2, int n2, int m3, int n3,
280  int k, int ib, int nb,
281  float *A1, int lda1,
282  float *A2, int lda2,
283  float *A3, int lda3,
284  float *V, int ldv,
285  float *T, int ldt,
286  float *WORK, int ldwork);
287 int CORE_stsmqr_sytra1( int side, int trans,
288  int m1, int n1, int m2, int n2,
289  int k, int ib,
290  float *A1, int lda1,
291  float *A2, int lda2,
292  float *V, int ldv,
293  float *T, int ldt,
294  float *WORK, int ldwork);
295 int CORE_stsqrt(int M, int N, int IB,
296  float *A1, int LDA1,
297  float *A2, int LDA2,
298  float *T, int LDT,
299  float *TAU, float *WORK);
300 int CORE_ststrf(int M, int N, int IB, int NB,
301  float *U, int LDU,
302  float *A, int LDA,
303  float *L, int LDL,
304  int *IPIV, float *WORK,
305  int LDWORK, int *INFO);
306 int CORE_sttmqr(int side, int trans,
307  int M1, int N1, int M2, int N2, int K, int IB,
308  float *A1, int LDA1,
309  float *A2, int LDA2,
310  float *V, int LDV,
311  float *T, int LDT,
312  float *WORK, int LDWORK);
313 int CORE_sttqrt(int M, int N, int IB,
314  float *A1, int LDA1,
315  float *A2, int LDA2,
316  float *T, int LDT,
317  float *TAU,
318  float *WORK);
319 int CORE_sttmlq(int side, int trans,
320  int M1, int N1, int M2, int N2, int K, int IB,
321  float *A1, int LDA1,
322  float *A2, int LDA2,
323  float *V, int LDV,
324  float *T, int LDT,
325  float *WORK, int LDWORK);
326 int CORE_sttlqt(int M, int N, int IB,
327  float *A1, int LDA1,
328  float *A2, int LDA2,
329  float *T, int LDT,
330  float *TAU,
331  float *WORK);
332 int CORE_sormlq(int side, int trans,
333  int M, int N, int IB, int K,
334  float *V, int LDV,
335  float *T, int LDT,
336  float *C, int LDC,
337  float *WORK, int LDWORK);
338 int CORE_sormqr(int side, int trans,
339  int M, int N, int K, int IB,
340  float *V, int LDV,
341  float *T, int LDT,
342  float *C, int LDC,
343  float *WORK, int LDWORK);
344 
348 void QUARK_CORE_sasum(Quark *quark, Quark_Task_Flags *task_flags,
349  PLASMA_enum storev, PLASMA_enum uplo, int m, int n,
350  float *A, int lda, int szeA,
351  float *work, int szeW);
352 void QUARK_CORE_sasum_f1(Quark *quark, Quark_Task_Flags *task_flags,
353  PLASMA_enum storev, PLASMA_enum uplo, int m, int n,
354  float *A, int lda, int szeA,
355  float *work, int szeW,
356  float *fake, int szeF);
357 void QUARK_CORE_sgeadd(Quark *quark, Quark_Task_Flags *task_flags,
358  int m, int n, int nb, float alpha,
359  float *A, int lda,
360  float *B, int ldb);
361 void QUARK_CORE_sbrdalg(Quark *quark, Quark_Task_Flags *task_flags,
362  int uplo,
363  int N, int NB,
364  PLASMA_desc *A,
365  float *C,
366  float *S,
367  int i, int j, int m, int grsiz, int BAND,
368  int *PCOL, int *ACOL, int *MCOL);
369 void QUARK_CORE_sgelqt(Quark *quark, Quark_Task_Flags *task_flags,
370  int m, int n, int ib, int nb,
371  float *A, int lda,
372  float *T, int ldt);
373 void QUARK_CORE_sgemm(Quark *quark, Quark_Task_Flags *task_flags,
374  int transA, int transB,
375  int m, int n, int k, int nb,
376  float alpha, float *A, int lda,
377  float *B, int ldb,
378  float beta, float *C, int ldc);
379 void QUARK_CORE_sgemm2( Quark *quark, Quark_Task_Flags *task_flags,
380  int transA, int transB,
381  int m, int n, int k, int nb,
382  float alpha, float *A, int lda,
383  float *B, int ldb,
384  float beta, float *C, int ldc);
385 void QUARK_CORE_sgemm_f2(Quark *quark, Quark_Task_Flags *task_flags,
386  int transA, int transB,
387  int m, int n, int k, int nb,
388  float alpha, float *A, int lda,
389  float *B, int ldb,
390  float beta, float *C, int ldc,
391  float *fake1, int szefake1, int flag1,
392  float *fake2, int szefake2, int flag2);
393 void QUARK_CORE_sgemm_p2(Quark *quark, Quark_Task_Flags *task_flags,
394  int transA, int transB,
395  int m, int n, int k, int nb,
396  float alpha, float *A, int lda,
397  float **B, int ldb,
398  float beta, float *C, int ldc);
399 void QUARK_CORE_sgemm_p2f1(Quark *quark, Quark_Task_Flags *task_flags,
400  int transA, int transB,
401  int m, int n, int k, int nb,
402  float alpha, float *A, int lda,
403  float **B, int ldb,
404  float beta, float *C, int ldc,
405  float *fake1, int szefake1, int flag1);
406 void QUARK_CORE_sgemm_p3(Quark *quark, Quark_Task_Flags *task_flags,
407  int transA, int transB,
408  int m, int n, int k, int nb,
409  float alpha, float *A, int lda,
410  float *B, int ldb,
411  float beta, float **C, int ldc);
412 void QUARK_CORE_sgeqrt(Quark *quark, Quark_Task_Flags *task_flags,
413  int m, int n, int ib, int nb,
414  float *A, int lda,
415  float *T, int ldt);
416 void QUARK_CORE_sgessm(Quark *quark, Quark_Task_Flags *task_flags,
417  int m, int n, int k, int ib, int nb,
418  int *IPIV,
419  float *L, int ldl,
420  float *A, int lda);
421 void QUARK_CORE_sgetrf(Quark *quark, Quark_Task_Flags *task_flags,
422  int m, int n, int nb,
423  float *A, int lda,
424  int *IPIV,
425  PLASMA_sequence *sequence, PLASMA_request *request,
426  PLASMA_bool check_info, int iinfo);
427 void QUARK_CORE_sgetrf_incpiv(Quark *quark, Quark_Task_Flags *task_flags,
428  int m, int n, int ib, int nb,
429  float *A, int lda,
430  int *IPIV,
431  PLASMA_sequence *sequence, PLASMA_request *request,
432  PLASMA_bool check_info, int iinfo);
433 void QUARK_CORE_sgetrf_reclap(Quark *quark, Quark_Task_Flags *task_flags,
434  int m, int n, int nb,
435  float *A, int lda,
436  int *IPIV,
437  PLASMA_sequence *sequence, PLASMA_request *request,
438  PLASMA_bool check_info, int iinfo,
439  int nbthread);
440 void QUARK_CORE_sgetrf_rectil(Quark *quark, Quark_Task_Flags *task_flags,
441  PLASMA_desc A, float *Amn, int size,
442  int *IPIV,
443  PLASMA_sequence *sequence, PLASMA_request *request,
444  PLASMA_bool check_info, int iinfo,
445  int nbthread);
446 void QUARK_CORE_sgetrip(Quark *quark, Quark_Task_Flags *task_flags,
447  int m, int n, float *A, int szeA);
448 void QUARK_CORE_sgetrip_f1(Quark *quark, Quark_Task_Flags *task_flags,
449  int m, int n, float *A, int szeA,
450  float *fake, int szeF, int paramF);
451 void QUARK_CORE_sgetrip_f2(Quark *quark, Quark_Task_Flags *task_flags,
452  int m, int n, float *A, int szeA,
453  float *fake1, int szeF1, int paramF1,
454  float *fake2, int szeF2, int paramF2);
455 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
456  int side, int uplo,
457  int m, int n, int nb,
458  float alpha, float *A, int lda,
459  float *B, int ldb,
460  float beta, float *C, int ldc);
461 void QUARK_CORE_ssygst(Quark *quark, Quark_Task_Flags *task_flags,
462  int itype, int uplo, int N,
463  float *A, int LDA,
464  float *B, int LDB,
465  PLASMA_sequence *sequence, PLASMA_request *request,
466  int iinfo);
467 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
468  int uplo, int trans,
469  int n, int k, int nb,
470  float alpha, float *A, int lda,
471  float beta, float *C, int ldc);
472 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
473  int uplo, int trans,
474  int n, int k, int nb,
475  float alpha, float *A, int lda,
476  float *B, int LDB,
477  float beta, float *C, int ldc);
478 void QUARK_CORE_ssyrfb(Quark *quark, Quark_Task_Flags *task_flags,
479  int uplo,
480  int n, int k, int ib, int nb,
481  float *A, int lda,
482  float *T, int ldt,
483  float *C, int ldc);
484 void QUARK_CORE_slacpy(Quark *quark, Quark_Task_Flags *task_flags,
485  PLASMA_enum uplo, int m, int n, int mb,
486  float *A, int lda,
487  float *B, int ldb);
488 void QUARK_CORE_slange(Quark *quark, Quark_Task_Flags *task_flags,
489  int norm, int M, int N,
490  float *A, int LDA, int szeA,
491  int szeW, float *result);
492 void QUARK_CORE_slange_f1(Quark *quark, Quark_Task_Flags *task_flags,
493  int norm, int M, int N,
494  float *A, int LDA, int szeA,
495  int szeW, float *result,
496  float *fake, int szeF);
497 #ifdef COMPLEX
498 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
499  int norm, int uplo, int N,
500  float *A, int LDA, int szeA,
501  int szeW, float *result);
502 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
503  int norm, int uplo, int N,
504  float *A, int LDA, int szeA,
505  int szeW, float *result,
506  float *fake, int szeF);
507 #endif
508 void QUARK_CORE_slansy(Quark *quark, Quark_Task_Flags *task_flags,
509  int norm, int uplo, int N,
510  float *A, int LDA, int szeA,
511  int szeW, float *result);
512 void QUARK_CORE_slansy_f1(Quark *quark, Quark_Task_Flags *task_flags,
513  int norm, int uplo, int N,
514  float *A, int LDA, int szeA,
515  int szeW, float *result,
516  float *fake, int szeF);
517 void QUARK_CORE_slaset(Quark *quark, Quark_Task_Flags *task_flags,
518  PLASMA_enum uplo, int n1, int n2, float alpha,
519  float beta, float *tileA, int ldtilea);
520 void QUARK_CORE_slaset2(Quark *quark, Quark_Task_Flags *task_flags,
521  PLASMA_enum uplo, int n1, int n2, float alpha,
522  float *tileA, int ldtilea);
523 void QUARK_CORE_slaswp(Quark *quark, Quark_Task_Flags *task_flags,
524  int n, float *A, int lda,
525  int i1, int i2, int *ipiv, int inc);
526 void QUARK_CORE_slaswp_f2(Quark *quark, Quark_Task_Flags *task_flags,
527  int n, float *A, int lda,
528  int i1, int i2, int *ipiv, int inc,
529  float *fake1, int szefake1, int flag1,
530  float *fake2, int szefake2, int flag2);
531 void QUARK_CORE_slaswp_ontile(Quark *quark, Quark_Task_Flags *task_flags,
532  PLASMA_desc descA, float *A,
533  int i1, int i2, int *ipiv, int inc, float *fakepanel);
534 void QUARK_CORE_slaswp_ontile_f2(Quark *quark, Quark_Task_Flags *task_flags,
535  PLASMA_desc descA, float *A,
536  int i1, int i2, int *ipiv, int inc,
537  float *fake1, int szefake1, int flag1,
538  float *fake2, int szefake2, int flag2);
539 void QUARK_CORE_slaswpc_ontile(Quark *quark, Quark_Task_Flags *task_flags,
540  PLASMA_desc descA, float *A,
541  int i1, int i2, int *ipiv, int inc, float *fakepanel);
542 void QUARK_CORE_slauum(Quark *quark, Quark_Task_Flags *task_flags,
543  int uplo, int n, int nb,
544  float *A, int lda);
545 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
546  float bump, int m, int n, float *A, int lda,
547  int bigM, int m0, int n0, unsigned long long int seed );
548 void QUARK_CORE_splgsy(Quark *quark, Quark_Task_Flags *task_flags,
549  float bump, int m, int n, float *A, int lda,
550  int bigM, int m0, int n0, unsigned long long int seed );
551 void QUARK_CORE_splrnt(Quark *quark, Quark_Task_Flags *task_flags,
552  int m, int n, float *A, int lda,
553  int bigM, int m0, int n0, unsigned long long int seed );
554 void QUARK_CORE_spotrf(Quark *quark, Quark_Task_Flags *task_flags,
555  int uplo, int n, int nb,
556  float *A, int lda,
557  PLASMA_sequence *sequence, PLASMA_request *request,
558  int iinfo);
559 void QUARK_CORE_sshift( Quark *quark, Quark_Task_Flags *task_flags,
560  int s, int m, int n, int L,
561  float *A);
562 void QUARK_CORE_sshiftw(Quark *quark, Quark_Task_Flags *task_flags,
563  int s, int cl, int m, int n, int L,
564  float *A, float *W);
565 void QUARK_CORE_sssssm(Quark *quark, Quark_Task_Flags *task_flags,
566  int m1, int n1, int m2, int n2, int k, int ib, int nb,
567  float *A1, int lda1,
568  float *A2, int lda2,
569  float *L1, int ldl1,
570  float *L2, int ldl2,
571  int *IPIV);
572 void QUARK_CORE_ssymm(Quark *quark, Quark_Task_Flags *task_flags,
573  int side, int uplo,
574  int m, int n, int nb,
575  float alpha, float *A, int lda,
576  float *B, int ldb,
577  float beta, float *C, int ldc);
578 void QUARK_CORE_ssyrk(Quark *quark, Quark_Task_Flags *task_flags,
579  int uplo, int trans,
580  int n, int k, int nb,
581  float alpha, float *A, int lda,
582  float beta, float *C, int ldc);
583 void QUARK_CORE_ssyr2k(Quark *quark, Quark_Task_Flags *task_flags,
584  int uplo, int trans,
585  int n, int k, int nb,
586  float alpha, float *A, int lda,
587  float *B, int LDB,
588  float beta, float *C, int ldc);
589 void QUARK_CORE_sswpab(Quark *quark, Quark_Task_Flags *task_flags,
590  int i, int n1, int n2,
591  float *A, int szeA);
592 void QUARK_CORE_sswptr_ontile(Quark *quark, Quark_Task_Flags *task_flags,
593  PLASMA_desc descA, float *Aij,
594  int i1, int i2, int *ipiv, int inc,
595  float *Akk, int ldak);
596 void QUARK_CORE_strdalg(Quark *quark, Quark_Task_Flags *task_flags,
597  int uplo,
598  int N, int NB,
599  PLASMA_desc *A,
600  float *C,
601  float *S,
602  int i, int j, int m, int grsiz, int BAND,
603  int *PCOL, int *ACOL, int *MCOL);
604 void QUARK_CORE_strmm(Quark *quark, Quark_Task_Flags *task_flags,
605  int side, int uplo, int transA, int diag,
606  int m, int n, int nb,
607  float alpha, float *A, int lda,
608  float *B, int ldb);
609 void QUARK_CORE_strmm_p2(Quark *quark, Quark_Task_Flags *task_flags,
610  int side, int uplo, int transA, int diag,
611  int m, int n, int nb,
612  float alpha, float *A, int lda,
613  float **B, int ldb);
614 void QUARK_CORE_strsm(Quark *quark, Quark_Task_Flags *task_flags,
615  int side, int uplo, int transA, int diag,
616  int m, int n, int nb,
617  float alpha, float *A, int lda,
618  float *B, int ldb);
619 void QUARK_CORE_strtri(Quark *quark, Quark_Task_Flags *task_flags,
620  int uplo, int diag, int n, int nb,
621  float *A, int lda,
622  PLASMA_sequence *sequence, PLASMA_request *request,
623  int iinfo);
624 void QUARK_CORE_stslqt(Quark *quark, Quark_Task_Flags *task_flags,
625  int m, int n, int ib, int nb,
626  float *A1, int lda1,
627  float *A2, int lda2,
628  float *T, int ldt);
629 void QUARK_CORE_stsmlq(Quark *quark, Quark_Task_Flags *task_flags,
630  int side, int trans,
631  int m1, int n1, int m2, int n2, int k, int ib, int nb,
632  float *A1, int lda1,
633  float *A2, int lda2,
634  float *V, int ldv,
635  float *T, int ldt);
636 void QUARK_CORE_stsmlq_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
637  int side, int trans,
638  int m1, int n1, int m2, int n2, int k, int ib, int nb,
639  float *A1, int lda1,
640  float *A2, int lda2,
641  float *V, int ldv,
642  float *T, int ldt);
643 void QUARK_CORE_stsmlq_corner(Quark *quark, Quark_Task_Flags *task_flags,
644  int m1, int n1, int m2, int n2, int m3, int n3, int k, int ib, int nb,
645  float *A1, int lda1,
646  float *A2, int lda2,
647  float *A3, int lda3,
648  float *V, int ldv,
649  float *T, int ldt);
650 void QUARK_CORE_stsmqr(Quark *quark, Quark_Task_Flags *task_flags,
651  int side, int trans,
652  int m1, int n1, int m2, int n2, int k, int ib, int nb,
653  float *A1, int lda1,
654  float *A2, int lda2,
655  float *V, int ldv,
656  float *T, int ldt);
657 void QUARK_CORE_stsmqr_sytra1(Quark *quark, Quark_Task_Flags *task_flags,
658  int side, int trans,
659  int m1, int n1, int m2, int n2, int k, int ib, int nb,
660  float *A1, int lda1,
661  float *A2, int lda2,
662  float *V, int ldv,
663  float *T, int ldt);
664 void QUARK_CORE_stsmqr_corner(Quark *quark, Quark_Task_Flags *task_flags,
665  int m1, int n1, int m2, int n2, int m3, int n3, int k, int ib, int nb,
666  float *A1, int lda1,
667  float *A2, int lda2,
668  float *A3, int lda3,
669  float *V, int ldv,
670  float *T, int ldt);
671 void QUARK_CORE_stsqrt(Quark *quark, Quark_Task_Flags *task_flags,
672  int m, int n, int ib, int nb,
673  float *A1, int lda1,
674  float *A2, int lda2,
675  float *T, int ldt);
676 void QUARK_CORE_ststrf(Quark *quark, Quark_Task_Flags *task_flags,
677  int m, int n, int ib, int nb,
678  float *U, int ldu,
679  float *A, int lda,
680  float *L, int ldl,
681  int *IPIV,
682  PLASMA_sequence *sequence, PLASMA_request *request,
683  PLASMA_bool check_info, int iinfo);
684 void QUARK_CORE_sttmqr(Quark *quark, Quark_Task_Flags *task_flags,
685  int side, int trans,
686  int m1, int n1, int m2, int n2, int k, int ib, int nb,
687  float *A1, int lda1,
688  float *A2, int lda2,
689  float *V, int ldv,
690  float *T, int ldt);
691 void QUARK_CORE_sttqrt(Quark *quark, Quark_Task_Flags *task_flags,
692  int m, int n, int ib, int nb,
693  float *A1, int lda1,
694  float *A2, int lda2,
695  float *T, int ldt);
696 void QUARK_CORE_sttmlq(Quark *quark, Quark_Task_Flags *task_flags,
697  int side, int trans,
698  int m1, int n1, int m2, int n2, int k, int ib, int nb,
699  float *A1, int lda1,
700  float *A2, int lda2,
701  float *V, int ldv,
702  float *T, int ldt);
703 void QUARK_CORE_sttlqt(Quark *quark, Quark_Task_Flags *task_flags,
704  int m, int n, int ib, int nb,
705  float *A1, int lda1,
706  float *A2, int lda2,
707  float *T, int ldt);
708 void QUARK_CORE_spamm(Quark *quark, Quark_Task_Flags *task_flags,
709  int op, int side, int storev,
710  int m, int n, int k, int l,
711  float *A1, int lda1,
712  float *A2, int lda2,
713  float *V, int ldv,
714  float *W, int ldw);
715 void QUARK_CORE_sormlq(Quark *quark, Quark_Task_Flags *task_flags,
716  int side, int trans,
717  int m, int n, int ib, int nb, int k,
718  float *A, int lda,
719  float *T, int ldt,
720  float *C, int ldc);
721 void QUARK_CORE_sormqr(Quark *quark, Quark_Task_Flags *task_flags,
722  int side, int trans,
723  int m, int n, int k, int ib, int nb,
724  float *A, int lda,
725  float *T, int ldt,
726  float *C, int ldc);
727 
731 void CORE_sasum_quark(Quark *quark);
732 void CORE_sasum_f1_quark(Quark *quark);
733 void CORE_sgeadd_quark(Quark *quark);
734 void CORE_sbrdalg_quark(Quark *quark);
735 void CORE_sgelqt_quark(Quark *quark);
736 void CORE_sgemm_quark(Quark *quark);
737 void CORE_sgeqrt_quark(Quark *quark);
738 void CORE_sgessm_quark(Quark *quark);
739 void CORE_sgetrf_quark(Quark *quark);
740 void CORE_sgetrf_incpiv_quark(Quark *quark);
741 void CORE_sgetrf_reclap_quark(Quark *quark);
742 void CORE_sgetrf_rectil_quark(Quark* quark);
743 void CORE_sgetrip_quark(Quark *quark);
744 void CORE_sgetrip_f1_quark(Quark *quark);
745 void CORE_sgetrip_f2_quark(Quark *quark);
746 #ifdef COMPLEX
747 void CORE_ssymm_quark(Quark *quark);
748 void CORE_ssyrk_quark(Quark *quark);
749 void CORE_ssyr2k_quark(Quark *quark);
750 #endif
751 void CORE_ssygst_quark(Quark *quark);
752 void CORE_ssyrfb_quark(Quark *quark);
753 void CORE_slacpy_quark(Quark *quark);
754 void CORE_slange_quark(Quark *quark);
755 void CORE_slange_f1_quark(Quark *quark);
756 #ifdef COMPLEX
757 void CORE_slansy_quark(Quark *quark);
758 void CORE_slansy_f1_quark(Quark *quark);
759 #endif
760 void CORE_slansy_quark(Quark *quark);
761 void CORE_slansy_f1_quark(Quark *quark);
762 void CORE_slaset_quark(Quark *quark);
763 void CORE_slaset2_quark(Quark *quark);
764 void CORE_slauum_quark(Quark *quark);
765 void CORE_spamm_quark(Quark *quark);
766 void CORE_splgsy_quark(Quark *quark);
767 void CORE_splgsy_quark(Quark *quark);
768 void CORE_splrnt_quark(Quark *quark);
769 void CORE_spotrf_quark(Quark *quark);
770 void CORE_sshift_quark(Quark *quark);
771 void CORE_sshiftw_quark(Quark *quark);
772 void CORE_sssssm_quark(Quark *quark);
773 void CORE_ssymm_quark(Quark *quark);
774 void CORE_ssyrk_quark(Quark *quark);
775 void CORE_ssyr2k_quark(Quark *quark);
776 void CORE_sswpab_quark(Quark *quark);
777 void CORE_sswptr_ontile_quark(Quark *quark);
778 void CORE_strdalg_quark(Quark *quark);
779 void CORE_strmm_quark(Quark *quark);
780 void CORE_strsm_quark(Quark *quark);
781 void CORE_strtri_quark(Quark *quark);
782 void CORE_stslqt_quark(Quark *quark);
783 void CORE_stsmlq_quark(Quark *quark);
784 void CORE_stsmlq_sytra1_quark(Quark *quark);
785 void CORE_stsmlq_corner_quark(Quark *quark);
786 void CORE_stsmqr_quark(Quark *quark);
787 void CORE_stsmqr_sytra1_quark(Quark *quark);
788 void CORE_stsmqr_corner_quark(Quark *quark);
789 void CORE_stsqrt_quark(Quark *quark);
790 void CORE_ststrf_quark(Quark *quark);
791 void CORE_sttmqr_quark(Quark *quark);
792 void CORE_sttqrt_quark(Quark *quark);
793 void CORE_sttmlq_quark(Quark *quark);
794 void CORE_sttlqt_quark(Quark *quark);
795 void CORE_sormlq_quark(Quark *quark);
796 void CORE_sormqr_quark(Quark *quark);
797 
798 void CORE_slaswp_quark(Quark* quark);
799 void CORE_slaswp_f2_quark(Quark* quark);
800 void CORE_slaswp_ontile_quark(Quark *quark);
801 void CORE_slaswp_ontile_f2_quark(Quark *quark);
802 void CORE_slaswpc_ontile_quark(Quark *quark);
803 void CORE_strmm_p2_quark(Quark* quark);
804 void CORE_sgemm_f2_quark(Quark* quark);
805 void CORE_sgemm_p2_quark(Quark* quark);
806 void CORE_sgemm_p2f1_quark(Quark* quark);
807 void CORE_sgemm_p3_quark(Quark* quark);
808 
809 #ifdef __cplusplus
810 }
811 #endif
812 
813 #undef COMPLEX
814 
815 #endif