libflame revision_anchor
Functions
FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1.c File Reference

(r)

Functions

FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1 (FLA_Obj alpha, FLA_Obj tau, FLA_Obj u, FLA_Obj y, FLA_Obj z, FLA_Obj v, FLA_Obj A, FLA_Obj up, FLA_Obj a, FLA_Obj w)
 
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1 (int m_A, int n_A, float *buff_tau, float *buff_alpha, float *buff_u, int inc_u, float *buff_y, int inc_y, float *buff_z, int inc_z, float *buff_v, int inc_v, float *buff_A, int rs_A, int cs_A, float *buff_up, int inc_up, float *buff_a, int inc_a, float *buff_w, int inc_w)
 
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1 (int m_A, int n_A, double *buff_tau, double *buff_alpha, double *buff_u, int inc_u, double *buff_y, int inc_y, double *buff_z, int inc_z, double *buff_v, int inc_v, double *buff_A, int rs_A, int cs_A, double *buff_up, int inc_up, double *buff_a, int inc_a, double *buff_w, int inc_w)
 
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1 (int m_A, int n_A, scomplex *buff_tau, scomplex *buff_alpha, scomplex *buff_u, int inc_u, scomplex *buff_y, int inc_y, scomplex *buff_z, int inc_z, scomplex *buff_v, int inc_v, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_up, int inc_up, scomplex *buff_a, int inc_a, scomplex *buff_w, int inc_w)
 
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1 (int m_A, int n_A, dcomplex *buff_tau, dcomplex *buff_alpha, dcomplex *buff_u, int inc_u, dcomplex *buff_y, int inc_y, dcomplex *buff_z, int inc_z, dcomplex *buff_v, int inc_v, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_up, int inc_up, dcomplex *buff_a, int inc_a, dcomplex *buff_w, int inc_w)
 

Function Documentation

◆ FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1()

FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1 ( int m_A,
int n_A,
scomplex * buff_tau,
scomplex * buff_alpha,
scomplex * buff_u,
int inc_u,
scomplex * buff_y,
int inc_y,
scomplex * buff_z,
int inc_z,
scomplex * buff_v,
int inc_v,
scomplex * buff_A,
int rs_A,
int cs_A,
scomplex * buff_up,
int inc_up,
scomplex * buff_a,
int inc_a,
scomplex * buff_w,
int inc_w )
436{
437 scomplex* buff_0 = FLA_COMPLEX_PTR( FLA_ZERO );
438 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
439 scomplex minus_inv_tau;
440 scomplex conj_psi1;
441 scomplex conj_nu1;
442 scomplex conj_alpha1;
443 int i;
444
445 bl1_csetv( m_A,
446 buff_0,
447 buff_w, inc_w );
448
449 bl1_cdiv3( buff_m1, buff_tau, &minus_inv_tau );
450
451 for ( i = 0; i < n_A; ++i )
452 {
453 scomplex* a1 = buff_A + (i )*cs_A + (0 )*rs_A;
454 scomplex* u = buff_u;
455 scomplex* psi1 = buff_y + (i )*inc_y;
456 scomplex* nu1 = buff_v + (i )*inc_v;
457 scomplex* z = buff_z;
458 scomplex* up = buff_up;
459 scomplex* alpha1 = buff_a + (i )*inc_a;
460 scomplex* w = buff_w;
461 scomplex* alpha = buff_alpha;
462 scomplex temp1;
463 scomplex temp2;
464
465 /*------------------------------------------------------------*/
466
467 bl1_ccopyconj( psi1, &conj_psi1 );
468 bl1_cmult3( alpha, &conj_psi1, &temp1 );
469
470 bl1_ccopyconj( nu1, &conj_nu1 );
471 bl1_cmult3( alpha, &conj_nu1, &temp2 );
472
473 bl1_caxpyv( BLIS1_NO_CONJUGATE,
474 m_A,
475 &temp1,
476 u, inc_u,
477 a1, rs_A );
478 //F77_caxpy( &m_A,
479 // &temp1,
480 // u, &inc_u,
481 // a1, &rs_A );
482
483 bl1_caxpyv( BLIS1_NO_CONJUGATE,
484 m_A,
485 &temp2,
486 z, inc_z,
487 a1, rs_A );
488 //F77_caxpy( &m_A,
489 // &temp2,
490 // z, &inc_z,
491 // a1, &rs_A );
492
493 bl1_cdot( BLIS1_CONJUGATE,
494 m_A,
495 a1, rs_A,
496 up, inc_up,
497 psi1 );
498
499 bl1_ccopyconj( psi1, &conj_psi1 );
500 bl1_cmult4( &minus_inv_tau, &conj_psi1, alpha1, alpha1 );
501
502 bl1_ccopyconj( alpha1, &conj_alpha1 );
503
504 bl1_caxpyv( BLIS1_NO_CONJUGATE,
505 m_A,
506 &conj_alpha1,
507 a1, rs_A,
508 w, inc_w );
509 //F77_caxpy( &m_A,
510 // &conj_alpha1,
511 // a1, &rs_A,
512 // w, &inc_w );
513
514 /*------------------------------------------------------------*/
515
516 }
517
518 return FLA_SUCCESS;
519}
FLA_MINUS_ONE
FLA_Obj FLA_MINUS_ONE
Definition FLA_Init.c:22
FLA_ZERO
FLA_Obj FLA_ZERO
Definition FLA_Init.c:20
psi1
double *restrict psi1
Definition bl1_axmyv2.c:141
i
int i
Definition bl1_axmyv2.c:145
temp2
double temp2
Definition bl1_axpyv2b.c:147
temp1
double temp1
Definition bl1_axpyv2b.c:146
alpha1
double *restrict alpha1
Definition bl1_axpyv2bdotaxpy.c:198
bl1_caxpyv
void bl1_caxpyv(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy)
Definition bl1_axpyv.c:29
bl1_cdot
void bl1_cdot(conj1_t conj, int n, scomplex *x, int incx, scomplex *y, int incy, scomplex *rho)
Definition bl1_dot.c:39
bl1_csetv
void bl1_csetv(int m, scomplex *sigma, scomplex *x, int incx)
Definition bl1_setv.c:52
BLIS1_CONJUGATE
@ BLIS1_CONJUGATE
Definition blis_type_defs.h:82
BLIS1_NO_CONJUGATE
@ BLIS1_NO_CONJUGATE
Definition blis_type_defs.h:81
scomplex
Definition blis_type_defs.h:133

References alpha1, bl1_caxpyv(), bl1_cdot(), bl1_csetv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_MINUS_ONE, FLA_ZERO, i, psi1, temp1, and temp2.

Referenced by FLA_Bidiag_UT_u_step_ofc_var3(), and FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1().

◆ FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1()

FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1 ( int m_A,
int n_A,
double * buff_tau,
double * buff_alpha,
double * buff_u,
int inc_u,
double * buff_y,
int inc_y,
double * buff_z,
int inc_z,
double * buff_v,
int inc_v,
double * buff_A,
int rs_A,
int cs_A,
double * buff_up,
int inc_up,
double * buff_a,
int inc_a,
double * buff_w,
int inc_w )
279{
280 double zero = bl1_d0();
281 double minus_one = bl1_dm1();
282 double* restrict u = buff_u;
283 double* restrict up = buff_up;
284 double* restrict w = buff_w;
285 double* restrict z = buff_z;
286 double* restrict alpha = buff_alpha;
287 double* restrict a1;
288 double* restrict a2;
289 double* restrict psi1;
290 double* restrict psi2;
291 double* restrict alpha1;
292 double* restrict alpha2;
293 double* restrict nu1;
294 double* restrict nu2;
295
296 double minus_inv_tau;
297 double alpha_conj_psi1;
298 double alpha_conj_psi2;
299 double alpha_conj_nu1;
300 double alpha_conj_nu2;
301 int i;
302 int n_run = n_A / 2;
303 int n_left = n_A % 2;
304 int twocs_A = 2*cs_A;
305 int twoinc_y = 2*inc_y;
306 int twoinc_a = 2*inc_a;
307 int twoinc_v = 2*inc_v;
308
309
310 bl1_dsetv( m_A,
311 &zero,
312 buff_w, inc_w );
313
314 bl1_ddiv3( &minus_one, buff_tau, &minus_inv_tau );
315
316 a1 = buff_A;
317 a2 = buff_A + cs_A;
318 psi1 = buff_y;
319 psi2 = buff_y + inc_y;
320 alpha1 = buff_a;
321 alpha2 = buff_a + inc_a;
322 nu1 = buff_v;
323 nu2 = buff_v + inc_v;
324
325 for ( i = 0; i < n_run; ++i )
326 {
327
328 /*------------------------------------------------------------*/
329
330 bl1_dmult3( alpha, psi1, &alpha_conj_psi1 );
331 bl1_dmult3( alpha, psi2, &alpha_conj_psi2 );
332
333 bl1_dmult3( alpha, nu1, &alpha_conj_nu1 );
334 bl1_dmult3( alpha, nu2, &alpha_conj_nu2 );
335
336/*
337 Effective computation:
338 A = A + alpha * ( u * y' + z * v' );
339 y = A' * up;
340 a = a - conj(y) / tau;
341 w = A * conj(a);
342*/
343 bl1_daxpyv2b( m_A,
344 &alpha_conj_psi1,
345 &alpha_conj_nu1,
346 u, inc_u,
347 z, inc_z,
348 a1, rs_A );
349 bl1_daxpyv2b( m_A,
350 &alpha_conj_psi2,
351 &alpha_conj_nu2,
352 u, inc_u,
353 z, inc_z,
354 a2, rs_A );
355
356
357 bl1_ddotsv2( BLIS1_CONJUGATE,
358 m_A,
359 a1, rs_A,
360 a2, rs_A,
361 up, inc_up,
362 &zero,
363 psi1,
364 psi2 );
365
366 bl1_dmult4( &minus_inv_tau, psi1, alpha1, alpha1 );
367 bl1_dmult4( &minus_inv_tau, psi2, alpha2, alpha2 );
368
369 bl1_daxpyv2b( m_A,
370 alpha1,
371 alpha2,
372 a1, rs_A,
373 a2, rs_A,
374 w, inc_w );
375
376 /*------------------------------------------------------------*/
377
378 a1 += twocs_A;
379 a2 += twocs_A;
380 psi1 += twoinc_y;
381 psi2 += twoinc_y;
382 alpha1 += twoinc_a;
383 alpha2 += twoinc_a;
384 nu1 += twoinc_v;
385 nu2 += twoinc_v;
386 }
387
388 if ( n_left == 1 )
389 {
390 double rho1;
391
392 bl1_dmult3( alpha, psi1, &alpha_conj_psi1 );
393 bl1_dmult3( alpha, nu1, &alpha_conj_nu1 );
394
395 bl1_daxpyv2b( m_A,
396 &alpha_conj_psi1,
397 &alpha_conj_nu1,
398 u, inc_u,
399 z, inc_z,
400 a1, rs_A );
401
402 bl1_ddot( BLIS1_CONJUGATE,
403 m_A,
404 a1, rs_A,
405 up, inc_up,
406 &rho1 );
407 bl1_dscals( &zero, psi1 );
408 bl1_dadd3( psi1, &rho1, psi1 );
409
410 bl1_dmult4( &minus_inv_tau, psi1, alpha1, alpha1 );
411
412 bl1_daxpyv( BLIS1_NO_CONJUGATE,
413 m_A,
414 alpha1,
415 a1, rs_A,
416 w, inc_w );
417 }
418
419 return FLA_SUCCESS;
420}
n_left
int n_left
Definition bl1_axmyv2.c:149
n_run
int n_run
Definition bl1_axmyv2.c:148
twoinc_y
int twoinc_y
Definition bl1_axpyv2b.c:154
bl1_daxpyv2b
void bl1_daxpyv2b(int n, double *alpha1, double *alpha2, double *x1, int inc_x1, double *x2, int inc_x2, double *y, int inc_y)
Definition bl1_axpyv2b.c:31
bl1_daxpyv
void bl1_daxpyv(conj1_t conj, int n, double *alpha, double *x, int incx, double *y, int incy)
Definition bl1_axpyv.c:21
bl1_ddot
void bl1_ddot(conj1_t conj, int n, double *x, int incx, double *y, int incy, double *rho)
Definition bl1_dot.c:26
rho1
double rho1
Definition bl1_dotsv2.c:149
bl1_ddotsv2
void bl1_ddotsv2(conj1_t conjxy, int n, double *x, int inc_x, double *y, int inc_y, double *z, int inc_z, double *beta, double *rho_xz, double *rho_yz)
Definition bl1_dotsv2.c:35
alpha2
double *restrict alpha2
Definition bl1_dotv2axpyv2b.c:188
bl1_dm1
double bl1_dm1(void)
Definition bl1_constants.c:182
bl1_dsetv
void bl1_dsetv(int m, double *sigma, double *x, int incx)
Definition bl1_setv.c:39
bl1_d0
double bl1_d0(void)
Definition bl1_constants.c:118

References alpha1, alpha2, bl1_d0(), bl1_daxpyv(), bl1_daxpyv2b(), bl1_ddot(), bl1_ddotsv2(), bl1_dm1(), bl1_dsetv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, i, n_left, n_run, psi1, rho1, and twoinc_y.

Referenced by FLA_Bidiag_UT_u_step_ofd_var3(), and FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1().

◆ FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1()

FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1 ( int m_A,
int n_A,
float * buff_tau,
float * buff_alpha,
float * buff_u,
int inc_u,
float * buff_y,
int inc_y,
float * buff_z,
int inc_z,
float * buff_v,
int inc_v,
float * buff_A,
int rs_A,
int cs_A,
float * buff_up,
int inc_up,
float * buff_a,
int inc_a,
float * buff_w,
int inc_w )
182{
183 float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
184 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
185 float minus_inv_tau;
186 int i;
187
188 bl1_ssetv( m_A,
189 buff_0,
190 buff_w, inc_w );
191
192 minus_inv_tau = *buff_m1 / *buff_tau;
193
194 for ( i = 0; i < n_A; ++i )
195 {
196 float* a1 = buff_A + (i )*cs_A + (0 )*rs_A;
197 float* u = buff_u;
198 float* psi1 = buff_y + (i )*inc_y;
199 float* nu1 = buff_v + (i )*inc_v;
200 float* z = buff_z;
201 float* up = buff_up;
202 float* alpha1 = buff_a + (i )*inc_a;
203 float* w = buff_w;
204 float* alpha = buff_alpha;
205 float temp1;
206 float temp2;
207
208 /*------------------------------------------------------------*/
209
210 // bl1_smult3( alpha, psi1, &temp1 );
211 temp1 = *alpha * *psi1;
212
213 // bl1_smult3( alpha, nu1, &temp2 );
214 temp2 = *alpha * *nu1;
215
216 bl1_saxpyv( BLIS1_NO_CONJUGATE,
217 m_A,
218 &temp1,
219 u, inc_u,
220 a1, rs_A );
221 //F77_saxpy( &m_A,
222 // &temp1,
223 // u, &inc_u,
224 // a1, &rs_A );
225
226 bl1_saxpyv( BLIS1_NO_CONJUGATE,
227 m_A,
228 &temp2,
229 z, inc_z,
230 a1, rs_A );
231 //F77_saxpy( &m_A,
232 // &temp2,
233 // z, &inc_z,
234 // a1, &rs_A );
235
236 bl1_sdot( BLIS1_CONJUGATE,
237 m_A,
238 a1, rs_A,
239 up, inc_up,
240 psi1 );
241 //*psi1 = F77_sdot( &m_A,
242 // a1, &rs_A,
243 // up, &inc_up );
244
245 // bl1_smult4( &minus_inv_tau, psi1, alpha1, alpha1 );
246 *alpha1 = *alpha1 + minus_inv_tau * *psi1;
247
248 bl1_saxpyv( BLIS1_NO_CONJUGATE,
249 m_A,
250 alpha1,
251 a1, rs_A,
252 w, inc_w );
253 //F77_saxpy( &m_A,
254 // alpha1,
255 // a1, &rs_A,
256 // w, &inc_w );
257
258 /*------------------------------------------------------------*/
259
260 }
261
262 return FLA_SUCCESS;
263}
bl1_saxpyv
void bl1_saxpyv(conj1_t conj, int n, float *alpha, float *x, int incx, float *y, int incy)
Definition bl1_axpyv.c:13
bl1_sdot
void bl1_sdot(conj1_t conj, int n, float *x, int incx, float *y, int incy, float *rho)
Definition bl1_dot.c:13
bl1_ssetv
void bl1_ssetv(int m, float *sigma, float *x, int incx)
Definition bl1_setv.c:26

References alpha1, bl1_saxpyv(), bl1_sdot(), bl1_ssetv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_MINUS_ONE, FLA_ZERO, i, psi1, temp1, and temp2.

Referenced by FLA_Bidiag_UT_u_step_ofs_var3(), and FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1().

◆ FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1()

FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1 ( FLA_Obj alpha,
FLA_Obj tau,
FLA_Obj u,
FLA_Obj y,
FLA_Obj z,
FLA_Obj v,
FLA_Obj A,
FLA_Obj up,
FLA_Obj a,
FLA_Obj w )
14{
15/*
16 Effective computation:
17 A = A + alpha * ( u * y' + z * v' );
18 y = A' * up;
19 a = a - conj(y) / tau;
20 w = A * conj(a);
21*/
22 FLA_Datatype datatype;
23 int m_A, n_A;
24 int rs_A, cs_A;
25 int inc_u, inc_y, inc_z, inc_v;
26 int inc_up, inc_a, inc_w;
27
28 datatype = FLA_Obj_datatype( A );
29
30 m_A = FLA_Obj_length( A );
31 n_A = FLA_Obj_width( A );
32
33 rs_A = FLA_Obj_row_stride( A );
34 cs_A = FLA_Obj_col_stride( A );
35
36 inc_u = FLA_Obj_vector_inc( u );
37 inc_y = FLA_Obj_vector_inc( y );
38 inc_z = FLA_Obj_vector_inc( z );
39 inc_v = FLA_Obj_vector_inc( v );
40
41 inc_up = FLA_Obj_vector_inc( up );
42 inc_a = FLA_Obj_vector_inc( a );
43 inc_w = FLA_Obj_vector_inc( w );
44
45
46 switch ( datatype )
47 {
48 case FLA_FLOAT:
49 {
50 float* buff_A = FLA_FLOAT_PTR( A );
51 float* buff_u = FLA_FLOAT_PTR( u );
52 float* buff_y = FLA_FLOAT_PTR( y );
53 float* buff_z = FLA_FLOAT_PTR( z );
54 float* buff_v = FLA_FLOAT_PTR( v );
55 float* buff_up = FLA_FLOAT_PTR( up );
56 float* buff_a = FLA_FLOAT_PTR( a );
57 float* buff_w = FLA_FLOAT_PTR( w );
58 float* buff_tau = FLA_FLOAT_PTR( tau );
59 float* buff_alpha = FLA_FLOAT_PTR( alpha );
60
61 FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1( m_A,
62 n_A,
63 buff_tau,
64 buff_alpha,
65 buff_u, inc_u,
66 buff_y, inc_y,
67 buff_z, inc_z,
68 buff_v, inc_v,
69 buff_A, rs_A, cs_A,
70 buff_up, inc_up,
71 buff_a, inc_a,
72 buff_w, inc_w );
73
74 break;
75 }
76
77 case FLA_DOUBLE:
78 {
79 double* buff_A = FLA_DOUBLE_PTR( A );
80 double* buff_u = FLA_DOUBLE_PTR( u );
81 double* buff_y = FLA_DOUBLE_PTR( y );
82 double* buff_z = FLA_DOUBLE_PTR( z );
83 double* buff_v = FLA_DOUBLE_PTR( v );
84 double* buff_up = FLA_DOUBLE_PTR( up );
85 double* buff_a = FLA_DOUBLE_PTR( a );
86 double* buff_w = FLA_DOUBLE_PTR( w );
87 double* buff_tau = FLA_DOUBLE_PTR( tau );
88 double* buff_alpha = FLA_DOUBLE_PTR( alpha );
89
90 FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1( m_A,
91 n_A,
92 buff_tau,
93 buff_alpha,
94 buff_u, inc_u,
95 buff_y, inc_y,
96 buff_z, inc_z,
97 buff_v, inc_v,
98 buff_A, rs_A, cs_A,
99 buff_up, inc_up,
100 buff_a, inc_a,
101 buff_w, inc_w );
102
103 break;
104 }
105
106 case FLA_COMPLEX:
107 {
108 scomplex* buff_A = FLA_COMPLEX_PTR( A );
109 scomplex* buff_u = FLA_COMPLEX_PTR( u );
110 scomplex* buff_y = FLA_COMPLEX_PTR( y );
111 scomplex* buff_z = FLA_COMPLEX_PTR( z );
112 scomplex* buff_v = FLA_COMPLEX_PTR( v );
113 scomplex* buff_up = FLA_COMPLEX_PTR( up );
114 scomplex* buff_a = FLA_COMPLEX_PTR( a );
115 scomplex* buff_w = FLA_COMPLEX_PTR( w );
116 scomplex* buff_tau = FLA_COMPLEX_PTR( tau );
117 scomplex* buff_alpha = FLA_COMPLEX_PTR( alpha );
118
119 FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1( m_A,
120 n_A,
121 buff_tau,
122 buff_alpha,
123 buff_u, inc_u,
124 buff_y, inc_y,
125 buff_z, inc_z,
126 buff_v, inc_v,
127 buff_A, rs_A, cs_A,
128 buff_up, inc_up,
129 buff_a, inc_a,
130 buff_w, inc_w );
131
132 break;
133 }
134
135 case FLA_DOUBLE_COMPLEX:
136 {
137 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
138 dcomplex* buff_u = FLA_DOUBLE_COMPLEX_PTR( u );
139 dcomplex* buff_y = FLA_DOUBLE_COMPLEX_PTR( y );
140 dcomplex* buff_z = FLA_DOUBLE_COMPLEX_PTR( z );
141 dcomplex* buff_v = FLA_DOUBLE_COMPLEX_PTR( v );
142 dcomplex* buff_up = FLA_DOUBLE_COMPLEX_PTR( up );
143 dcomplex* buff_a = FLA_DOUBLE_COMPLEX_PTR( a );
144 dcomplex* buff_w = FLA_DOUBLE_COMPLEX_PTR( w );
145 dcomplex* buff_tau = FLA_DOUBLE_COMPLEX_PTR( tau );
146 dcomplex* buff_alpha = FLA_DOUBLE_COMPLEX_PTR( alpha );
147
148 FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1( m_A,
149 n_A,
150 buff_tau,
151 buff_alpha,
152 buff_u, inc_u,
153 buff_y, inc_y,
154 buff_z, inc_z,
155 buff_v, inc_v,
156 buff_A, rs_A, cs_A,
157 buff_up, inc_up,
158 buff_a, inc_a,
159 buff_w, inc_w );
160
161 break;
162 }
163 }
164
165 return FLA_SUCCESS;
166}
FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1(int m_A, int n_A, float *buff_tau, float *buff_alpha, float *buff_u, int inc_u, float *buff_y, int inc_y, float *buff_z, int inc_z, float *buff_v, int inc_v, float *buff_A, int rs_A, int cs_A, float *buff_up, int inc_up, float *buff_a, int inc_a, float *buff_w, int inc_w)
Definition FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1.c:170
FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1(int m_A, int n_A, dcomplex *buff_tau, dcomplex *buff_alpha, dcomplex *buff_u, int inc_u, dcomplex *buff_y, int inc_y, dcomplex *buff_z, int inc_z, dcomplex *buff_v, int inc_v, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_up, int inc_up, dcomplex *buff_a, int inc_a, dcomplex *buff_w, int inc_w)
Definition FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1.c:523
FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1(int m_A, int n_A, scomplex *buff_tau, scomplex *buff_alpha, scomplex *buff_u, int inc_u, scomplex *buff_y, int inc_y, scomplex *buff_z, int inc_z, scomplex *buff_v, int inc_v, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_up, int inc_up, scomplex *buff_a, int inc_a, scomplex *buff_w, int inc_w)
Definition FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1.c:424
FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1
FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1(int m_A, int n_A, double *buff_tau, double *buff_alpha, double *buff_u, int inc_u, double *buff_y, int inc_y, double *buff_z, int inc_z, double *buff_v, int inc_v, double *buff_A, int rs_A, int cs_A, double *buff_up, int inc_up, double *buff_a, int inc_a, double *buff_w, int inc_w)
Definition FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1.c:267
FLA_Obj_width
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
FLA_Obj_row_stride
dim_t FLA_Obj_row_stride(FLA_Obj obj)
Definition FLA_Query.c:167
FLA_Obj_length
dim_t FLA_Obj_length(FLA_Obj obj)
Definition FLA_Query.c:116
FLA_Obj_col_stride
dim_t FLA_Obj_col_stride(FLA_Obj obj)
Definition FLA_Query.c:174
FLA_Obj_vector_inc
dim_t FLA_Obj_vector_inc(FLA_Obj obj)
Definition FLA_Query.c:145
FLA_Obj_datatype
FLA_Datatype FLA_Obj_datatype(FLA_Obj obj)
Definition FLA_Query.c:13
FLA_Datatype
int FLA_Datatype
Definition FLA_type_defs.h:49
dcomplex
Definition blis_type_defs.h:138

References FLA_Fused_Gerc2_Ahx_Axpy_Ax_opc_var1(), FLA_Fused_Gerc2_Ahx_Axpy_Ax_opd_var1(), FLA_Fused_Gerc2_Ahx_Axpy_Ax_ops_var1(), FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Obj_width().

◆ FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1()

FLA_Error FLA_Fused_Gerc2_Ahx_Axpy_Ax_opz_var1 ( int m_A,
int n_A,
dcomplex * buff_tau,
dcomplex * buff_alpha,
dcomplex * buff_u,
int inc_u,
dcomplex * buff_y,
int inc_y,
dcomplex * buff_z,
int inc_z,
dcomplex * buff_v,
int inc_v,
dcomplex * buff_A,
int rs_A,
int cs_A,
dcomplex * buff_up,
int inc_up,
dcomplex * buff_a,
int inc_a,
dcomplex * buff_w,
int inc_w )
535{
536 dcomplex zero = bl1_z0();
537 dcomplex minus_one = bl1_zm1();
538 dcomplex* restrict u = buff_u;
539 dcomplex* restrict up = buff_up;
540 dcomplex* restrict w = buff_w;
541 dcomplex* restrict z = buff_z;
542 dcomplex* restrict alpha = buff_alpha;
543 dcomplex* restrict a1;
544 dcomplex* restrict a2;
545 dcomplex* restrict psi1;
546 dcomplex* restrict psi2;
547 dcomplex* restrict alpha1;
548 dcomplex* restrict alpha2;
549 dcomplex* restrict nu1;
550 dcomplex* restrict nu2;
551
552 dcomplex minus_inv_tau;
553 dcomplex conj_psi1;
554 dcomplex conj_psi2;
555 dcomplex conj_nu1;
556 dcomplex conj_nu2;
557 dcomplex conj_alpha1;
558 dcomplex conj_alpha2;
559 dcomplex alpha_conj_psi1;
560 dcomplex alpha_conj_psi2;
561 dcomplex alpha_conj_nu1;
562 dcomplex alpha_conj_nu2;
563 int i;
564 int n_run = n_A / 2;
565 int n_left = n_A % 2;
566 int twocs_A = 2*cs_A;
567 int twoinc_y = 2*inc_y;
568 int twoinc_a = 2*inc_a;
569 int twoinc_v = 2*inc_v;
570
571
572 bl1_zsetv( m_A,
573 &zero,
574 buff_w, inc_w );
575
576 bl1_zdiv3( &minus_one, buff_tau, &minus_inv_tau );
577
578 a1 = buff_A;
579 a2 = buff_A + cs_A;
580 psi1 = buff_y;
581 psi2 = buff_y + inc_y;
582 alpha1 = buff_a;
583 alpha2 = buff_a + inc_a;
584 nu1 = buff_v;
585 nu2 = buff_v + inc_v;
586
587 for ( i = 0; i < n_run; ++i )
588 {
589
590 /*------------------------------------------------------------*/
591
592 bl1_zcopyconj( psi1, &conj_psi1 );
593 bl1_zcopyconj( psi2, &conj_psi2 );
594 bl1_zmult3( alpha, &conj_psi1, &alpha_conj_psi1 );
595 bl1_zmult3( alpha, &conj_psi2, &alpha_conj_psi2 );
596
597 bl1_zcopyconj( nu1, &conj_nu1 );
598 bl1_zcopyconj( nu2, &conj_nu2 );
599 bl1_zmult3( alpha, &conj_nu1, &alpha_conj_nu1 );
600 bl1_zmult3( alpha, &conj_nu2, &alpha_conj_nu2 );
601
602 bl1_zaxpyv2b( m_A,
603 &alpha_conj_psi1,
604 &alpha_conj_nu1,
605 u, inc_u,
606 z, inc_z,
607 a1, rs_A );
608 bl1_zaxpyv2b( m_A,
609 &alpha_conj_psi2,
610 &alpha_conj_nu2,
611 u, inc_u,
612 z, inc_z,
613 a2, rs_A );
614
615
616 bl1_zdotsv2( BLIS1_CONJUGATE,
617 m_A,
618 a1, rs_A,
619 a2, rs_A,
620 up, inc_up,
621 &zero,
622 psi1,
623 psi2 );
624
625 bl1_zcopyconj( psi1, &conj_psi1 );
626 bl1_zcopyconj( psi2, &conj_psi2 );
627 bl1_zmult4( &minus_inv_tau, &conj_psi1, alpha1, alpha1 );
628 bl1_zmult4( &minus_inv_tau, &conj_psi2, alpha2, alpha2 );
629 bl1_zcopyconj( alpha1, &conj_alpha1 );
630 bl1_zcopyconj( alpha2, &conj_alpha2 );
631
632 bl1_zaxpyv2b( m_A,
633 &conj_alpha1,
634 &conj_alpha2,
635 a1, rs_A,
636 a2, rs_A,
637 w, inc_w );
638
639 /*------------------------------------------------------------*/
640
641 a1 += twocs_A;
642 a2 += twocs_A;
643 psi1 += twoinc_y;
644 psi2 += twoinc_y;
645 alpha1 += twoinc_a;
646 alpha2 += twoinc_a;
647 nu1 += twoinc_v;
648 nu2 += twoinc_v;
649 }
650
651 if ( n_left == 1 )
652 {
653 dcomplex rho1;
654
655 bl1_zcopyconj( psi1, &conj_psi1 );
656 bl1_zmult3( alpha, &conj_psi1, &alpha_conj_psi1 );
657 bl1_zcopyconj( nu1, &conj_nu1 );
658 bl1_zmult3( alpha, &conj_nu1, &alpha_conj_nu1 );
659
660 bl1_zaxpyv2b( m_A,
661 &alpha_conj_psi1,
662 &alpha_conj_nu1,
663 u, inc_u,
664 z, inc_z,
665 a1, rs_A );
666
667 bl1_zdot( BLIS1_CONJUGATE,
668 m_A,
669 a1, rs_A,
670 up, inc_up,
671 &rho1 );
672 bl1_zscals( &zero, psi1 );
673 bl1_zadd3( psi1, &rho1, psi1 );
674
675 bl1_zcopyconj( psi1, &conj_psi1 );
676 bl1_zmult4( &minus_inv_tau, &conj_psi1, alpha1, alpha1 );
677 bl1_zcopyconj( alpha1, &conj_alpha1 );
678
679 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
680 m_A,
681 &conj_alpha1,
682 a1, rs_A,
683 w, inc_w );
684 }
685
686 return FLA_SUCCESS;
687}
bl1_zaxpyv2b
void bl1_zaxpyv2b(int n, dcomplex *alpha1, dcomplex *alpha2, dcomplex *x1, int inc_x1, dcomplex *x2, int inc_x2, dcomplex *y, int inc_y)
Definition bl1_axpyv2b.c:210
bl1_zaxpyv
void bl1_zaxpyv(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy)
Definition bl1_axpyv.c:60
bl1_zdot
void bl1_zdot(conj1_t conj, int n, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *rho)
Definition bl1_dot.c:65
bl1_zscals
bl1_zscals(beta, rho_yz)
bl1_zdotsv2
void bl1_zdotsv2(conj1_t conjxy, int n, dcomplex *x, int inc_x, dcomplex *y, int inc_y, dcomplex *z, int inc_z, dcomplex *beta, dcomplex *rho_xz, dcomplex *rho_yz)
Definition bl1_dotsv2.c:248
bl1_z0
dcomplex bl1_z0(void)
Definition bl1_constants.c:133
bl1_zsetv
void bl1_zsetv(int m, dcomplex *sigma, dcomplex *x, int incx)
Definition bl1_setv.c:66
bl1_zm1
dcomplex bl1_zm1(void)
Definition bl1_constants.c:197

References alpha1, alpha2, bl1_z0(), bl1_zaxpyv(), bl1_zaxpyv2b(), bl1_zdot(), bl1_zdotsv2(), bl1_zm1(), bl1_zscals(), bl1_zsetv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, i, n_left, n_run, psi1, rho1, and twoinc_y.

Referenced by FLA_Bidiag_UT_u_step_ofz_var3(), and FLA_Fused_Gerc2_Ahx_Axpy_Ax_opt_var1().

Generated on Thu Sep 23 2021 17:46:17 for libflame by doxygen 1.12.0