libflame revision_anchor
Functions
FLA_Apply_HUD_UT_l.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Apply_HUD_UT_l_unb_var1 (FLA_Obj tau, FLA_Obj w12t, FLA_Obj r12t, FLA_Obj u1, FLA_Obj C2, FLA_Obj v1, FLA_Obj D2)
 
FLA_Error FLA_Apply_HUD_UT_l_opt_var1 (FLA_Obj tau, FLA_Obj w12t, FLA_Obj r12t, FLA_Obj u1, FLA_Obj C2, FLA_Obj v1, FLA_Obj D2)
 
FLA_Error FLA_Apply_HUD_UT_l_ops_var1 (int m_u1_C2, int m_v1_D2, int n_r12t, float *tau, float *w12t, int inc_w12t, float *r12t, int inc_r12t, float *u1, int inc_u1, float *C2, int rs_C2, int cs_C2, float *v1, int inc_v1, float *D2, int rs_D2, int cs_D2)
 
FLA_Error FLA_Apply_HUD_UT_l_opd_var1 (int m_u1_C2, int m_v1_D2, int n_r12t, double *tau, double *w12t, int inc_w12t, double *r12t, int inc_r12t, double *u1, int inc_u1, double *C2, int rs_C2, int cs_C2, double *v1, int inc_v1, double *D2, int rs_D2, int cs_D2)
 
FLA_Error FLA_Apply_HUD_UT_l_opc_var1 (int m_u1_C2, int m_v1_D2, int n_r12t, scomplex *tau, scomplex *w12t, int inc_w12t, scomplex *r12t, int inc_r12t, scomplex *u1, int inc_u1, scomplex *C2, int rs_C2, int cs_C2, scomplex *v1, int inc_v1, scomplex *D2, int rs_D2, int cs_D2)
 
FLA_Error FLA_Apply_HUD_UT_l_opz_var1 (int m_u1_C2, int m_v1_D2, int n_r12t, dcomplex *tau, dcomplex *w12t, int inc_w12t, dcomplex *r12t, int inc_r12t, dcomplex *u1, int inc_u1, dcomplex *C2, int rs_C2, int cs_C2, dcomplex *v1, int inc_v1, dcomplex *D2, int rs_D2, int cs_D2)
 

Function Documentation

◆ FLA_Apply_HUD_UT_l_opc_var1()

FLA_Error FLA_Apply_HUD_UT_l_opc_var1 ( int  m_u1_C2,
int  m_v1_D2,
int  n_r12t,
scomplex *  tau,
scomplex *  w12t,
int  inc_w12t,
scomplex *  r12t,
int  inc_r12t,
scomplex *  u1,
int  inc_u1,
scomplex *  C2,
int  rs_C2,
int  cs_C2,
scomplex *  v1,
int  inc_v1,
scomplex *  D2,
int  rs_D2,
int  cs_D2 
)
339{
340 scomplex* one_p = FLA_COMPLEX_PTR( FLA_ONE );
341 scomplex* minus_one_p = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
342
343 // if ( FLA_Obj_has_zero_dim( r12t ) ) return FLA_SUCCESS;
344 if ( n_r12t == 0 ) return FLA_SUCCESS;
345
346 // // w12t = r12t;
347 // FLA_Copy_external( r12t, w12t );
348 bl1_ccopyv( BLIS1_NO_CONJUGATE,
349 n_r12t,
350 r12t, inc_r12t,
351 w12t, inc_w12t );
352
353 // // w12t = w12t + u1' * C2;
354 // // = w12t + C2^T * conj(u1);
355 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, C2, u1, FLA_ONE, w12t );
356 bl1_cgemv( BLIS1_TRANSPOSE,
357 BLIS1_CONJUGATE,
358 m_u1_C2,
359 n_r12t,
360 one_p,
361 C2, rs_C2, cs_C2,
362 u1, inc_u1,
363 one_p,
364 w12t, inc_w12t );
365
366 // // w12t = w12t + v1' * D2;
367 // // = w12t + D2^T * conj(v1);
368 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, D2, v1, FLA_ONE, w12t );
369 bl1_cgemv( BLIS1_TRANSPOSE,
370 BLIS1_CONJUGATE,
371 m_v1_D2,
372 n_r12t,
373 one_p,
374 D2, rs_D2, cs_D2,
375 v1, inc_v1,
376 one_p,
377 w12t, inc_w12t );
378
379 // // w12t = w12t / tau;
380 // FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w12t );
381 bl1_cinvscalv( BLIS1_NO_CONJUGATE,
382 n_r12t,
383 tau,
384 w12t, inc_w12t );
385
386 // // r12t = - w12t + r12t;
387 // FLA_Axpy_external( FLA_MINUS_ONE, w12t, r12t );
388 bl1_caxpyv( BLIS1_NO_CONJUGATE,
389 n_r12t,
390 minus_one_p,
391 w12t, inc_w12t,
392 r12t, inc_r12t );
393
394 // // C2 = - u1 * w12t + C2;
395 // FLA_Ger_external( FLA_MINUS_ONE, u1, w12t, C2 );
396 bl1_cger( BLIS1_NO_CONJUGATE,
397 BLIS1_NO_CONJUGATE,
398 m_u1_C2,
399 n_r12t,
400 minus_one_p,
401 u1, inc_u1,
402 w12t, inc_w12t,
403 C2, rs_C2, cs_C2 );
404
405 // // D2 = v1 * w12t + D2;
406 // FLA_Ger_external( FLA_ONE, v1, w12t, D2 );
407 bl1_cger( BLIS1_NO_CONJUGATE,
408 BLIS1_NO_CONJUGATE,
409 m_v1_D2,
410 n_r12t,
411 one_p,
412 v1, inc_v1,
413 w12t, inc_w12t,
414 D2, rs_D2, cs_D2 );
415
416 return FLA_SUCCESS;
417}
FLA_MINUS_ONE
FLA_Obj FLA_MINUS_ONE
Definition FLA_Init.c:22
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
i
int i
Definition bl1_axmyv2.c:145
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_ccopyv
void bl1_ccopyv(conj1_t conj, int m, scomplex *x, int incx, scomplex *y, int incy)
Definition bl1_copyv.c:49
bl1_cgemv
void bl1_cgemv(trans1_t transa, conj1_t conjx, int m, int n, scomplex *alpha, scomplex *a, int a_rs, int a_cs, scomplex *x, int incx, scomplex *beta, scomplex *y, int incy)
Definition bl1_gemv.c:125
bl1_cger
void bl1_cger(conj1_t conjx, conj1_t conjy, int m, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy, scomplex *a, int a_rs, int a_cs)
Definition bl1_ger.c:111
bl1_cinvscalv
void bl1_cinvscalv(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx)
Definition bl1_invscalv.c:52
BLIS1_TRANSPOSE
@ BLIS1_TRANSPOSE
Definition blis_type_defs.h:55
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 bl1_caxpyv(), bl1_ccopyv(), bl1_cgemv(), bl1_cger(), bl1_cinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_Apply_HUD_UT_l_opt_var1(), and FLA_UDdate_UT_opc_var1().

◆ FLA_Apply_HUD_UT_l_opd_var1()

FLA_Error FLA_Apply_HUD_UT_l_opd_var1 ( int  m_u1_C2,
int  m_v1_D2,
int  n_r12t,
double *  tau,
double *  w12t,
int  inc_w12t,
double *  r12t,
int  inc_r12t,
double *  u1,
int  inc_u1,
double *  C2,
int  rs_C2,
int  cs_C2,
double *  v1,
int  inc_v1,
double *  D2,
int  rs_D2,
int  cs_D2 
)
247{
248 double* one_p = FLA_DOUBLE_PTR( FLA_ONE );
249 double* minus_one_p = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
250
251 // if ( FLA_Obj_has_zero_dim( r12t ) ) return FLA_SUCCESS;
252 if ( n_r12t == 0 ) return FLA_SUCCESS;
253
254 // // w12t = r12t;
255 // FLA_Copy_external( r12t, w12t );
256 bl1_dcopyv( BLIS1_NO_CONJUGATE,
257 n_r12t,
258 r12t, inc_r12t,
259 w12t, inc_w12t );
260
261 // // w12t = w12t + u1' * C2;
262 // // = w12t + C2^T * conj(u1);
263 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, C2, u1, FLA_ONE, w12t );
264 bl1_dgemv( BLIS1_TRANSPOSE,
265 BLIS1_CONJUGATE,
266 m_u1_C2,
267 n_r12t,
268 one_p,
269 C2, rs_C2, cs_C2,
270 u1, inc_u1,
271 one_p,
272 w12t, inc_w12t );
273
274 // // w12t = w12t + v1' * D2;
275 // // = w12t + D2^T * conj(v1);
276 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, D2, v1, FLA_ONE, w12t );
277 bl1_dgemv( BLIS1_TRANSPOSE,
278 BLIS1_CONJUGATE,
279 m_v1_D2,
280 n_r12t,
281 one_p,
282 D2, rs_D2, cs_D2,
283 v1, inc_v1,
284 one_p,
285 w12t, inc_w12t );
286
287 // // w12t = w12t / tau;
288 // FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w12t );
289 bl1_dinvscalv( BLIS1_NO_CONJUGATE,
290 n_r12t,
291 tau,
292 w12t, inc_w12t );
293
294 // // r12t = - w12t + r12t;
295 // FLA_Axpy_external( FLA_MINUS_ONE, w12t, r12t );
296 bl1_daxpyv( BLIS1_NO_CONJUGATE,
297 n_r12t,
298 minus_one_p,
299 w12t, inc_w12t,
300 r12t, inc_r12t );
301
302 // // C2 = - u1 * w12t + C2;
303 // FLA_Ger_external( FLA_MINUS_ONE, u1, w12t, C2 );
304 bl1_dger( BLIS1_NO_CONJUGATE,
305 BLIS1_NO_CONJUGATE,
306 m_u1_C2,
307 n_r12t,
308 minus_one_p,
309 u1, inc_u1,
310 w12t, inc_w12t,
311 C2, rs_C2, cs_C2 );
312
313 // // D2 = v1 * w12t + D2;
314 // FLA_Ger_external( FLA_ONE, v1, w12t, D2 );
315 bl1_dger( BLIS1_NO_CONJUGATE,
316 BLIS1_NO_CONJUGATE,
317 m_v1_D2,
318 n_r12t,
319 one_p,
320 v1, inc_v1,
321 w12t, inc_w12t,
322 D2, rs_D2, cs_D2 );
323
324 return FLA_SUCCESS;
325}
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_dcopyv
void bl1_dcopyv(conj1_t conj, int m, double *x, int incx, double *y, int incy)
Definition bl1_copyv.c:42
bl1_dgemv
void bl1_dgemv(trans1_t transa, conj1_t conjx, int m, int n, double *alpha, double *a, int a_rs, int a_cs, double *x, int incx, double *beta, double *y, int incy)
Definition bl1_gemv.c:69
bl1_dger
void bl1_dger(conj1_t conjx, conj1_t conjy, int m, int n, double *alpha, double *x, int incx, double *y, int incy, double *a, int a_rs, int a_cs)
Definition bl1_ger.c:62
bl1_dinvscalv
void bl1_dinvscalv(conj1_t conj, int n, double *alpha, double *x, int incx)
Definition bl1_invscalv.c:26

References bl1_daxpyv(), bl1_dcopyv(), bl1_dgemv(), bl1_dger(), bl1_dinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_Apply_HUD_UT_l_opt_var1(), and FLA_UDdate_UT_opd_var1().

◆ FLA_Apply_HUD_UT_l_ops_var1()

FLA_Error FLA_Apply_HUD_UT_l_ops_var1 ( int  m_u1_C2,
int  m_v1_D2,
int  n_r12t,
float *  tau,
float *  w12t,
int  inc_w12t,
float *  r12t,
int  inc_r12t,
float *  u1,
int  inc_u1,
float *  C2,
int  rs_C2,
int  cs_C2,
float *  v1,
int  inc_v1,
float *  D2,
int  rs_D2,
int  cs_D2 
)
155{
156 float* one_p = FLA_FLOAT_PTR( FLA_ONE );
157 float* minus_one_p = FLA_FLOAT_PTR( FLA_MINUS_ONE );
158
159 // if ( FLA_Obj_has_zero_dim( r12t ) ) return FLA_SUCCESS;
160 if ( n_r12t == 0 ) return FLA_SUCCESS;
161
162 // // w12t = r12t;
163 // FLA_Copy_external( r12t, w12t );
164 bl1_scopyv( BLIS1_NO_CONJUGATE,
165 n_r12t,
166 r12t, inc_r12t,
167 w12t, inc_w12t );
168
169 // // w12t = w12t + u1' * C2;
170 // // = w12t + C2^T * conj(u1);
171 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, C2, u1, FLA_ONE, w12t );
172 bl1_sgemv( BLIS1_TRANSPOSE,
173 BLIS1_CONJUGATE,
174 m_u1_C2,
175 n_r12t,
176 one_p,
177 C2, rs_C2, cs_C2,
178 u1, inc_u1,
179 one_p,
180 w12t, inc_w12t );
181
182 // // w12t = w12t + v1' * D2;
183 // // = w12t + D2^T * conj(v1);
184 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, D2, v1, FLA_ONE, w12t );
185 bl1_sgemv( BLIS1_TRANSPOSE,
186 BLIS1_CONJUGATE,
187 m_v1_D2,
188 n_r12t,
189 one_p,
190 D2, rs_D2, cs_D2,
191 v1, inc_v1,
192 one_p,
193 w12t, inc_w12t );
194
195 // // w12t = w12t / tau;
196 // FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w12t );
197 bl1_sinvscalv( BLIS1_NO_CONJUGATE,
198 n_r12t,
199 tau,
200 w12t, inc_w12t );
201
202 // // r12t = - w12t + r12t;
203 // FLA_Axpy_external( FLA_MINUS_ONE, w12t, r12t );
204 bl1_saxpyv( BLIS1_NO_CONJUGATE,
205 n_r12t,
206 minus_one_p,
207 w12t, inc_w12t,
208 r12t, inc_r12t );
209
210 // // C2 = - u1 * w12t + C2;
211 // FLA_Ger_external( FLA_MINUS_ONE, u1, w12t, C2 );
212 bl1_sger( BLIS1_NO_CONJUGATE,
213 BLIS1_NO_CONJUGATE,
214 m_u1_C2,
215 n_r12t,
216 minus_one_p,
217 u1, inc_u1,
218 w12t, inc_w12t,
219 C2, rs_C2, cs_C2 );
220
221 // // D2 = v1 * w12t + D2;
222 // FLA_Ger_external( FLA_ONE, v1, w12t, D2 );
223 bl1_sger( BLIS1_NO_CONJUGATE,
224 BLIS1_NO_CONJUGATE,
225 m_v1_D2,
226 n_r12t,
227 one_p,
228 v1, inc_v1,
229 w12t, inc_w12t,
230 D2, rs_D2, cs_D2 );
231
232 return FLA_SUCCESS;
233}
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_scopyv
void bl1_scopyv(conj1_t conj, int m, float *x, int incx, float *y, int incy)
Definition bl1_copyv.c:35
bl1_sgemv
void bl1_sgemv(trans1_t transa, conj1_t conjx, int m, int n, float *alpha, float *a, int a_rs, int a_cs, float *x, int incx, float *beta, float *y, int incy)
Definition bl1_gemv.c:13
bl1_sger
void bl1_sger(conj1_t conjx, conj1_t conjy, int m, int n, float *alpha, float *x, int incx, float *y, int incy, float *a, int a_rs, int a_cs)
Definition bl1_ger.c:13
bl1_sinvscalv
void bl1_sinvscalv(conj1_t conj, int n, float *alpha, float *x, int incx)
Definition bl1_invscalv.c:13

References bl1_saxpyv(), bl1_scopyv(), bl1_sgemv(), bl1_sger(), bl1_sinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_Apply_HUD_UT_l_opt_var1(), and FLA_UDdate_UT_ops_var1().

◆ FLA_Apply_HUD_UT_l_opt_var1()

FLA_Error FLA_Apply_HUD_UT_l_opt_var1 ( FLA_Obj  tau,
FLA_Obj  w12t,
FLA_Obj  r12t,
FLA_Obj  u1,
FLA_Obj  C2,
FLA_Obj  v1,
FLA_Obj  D2 
)
17{
18 FLA_Datatype datatype;
19 int m_u1_C2;
20 int m_v1_D2;
21 int n_r12t;
22 int inc_u1;
23 int inc_v1;
24 int inc_w12t;
25 int inc_r12t;
26 int rs_C2, cs_C2;
27 int rs_D2, cs_D2;
28
29 if ( FLA_Obj_has_zero_dim( r12t ) ) return FLA_SUCCESS;
30
31 datatype = FLA_Obj_datatype( C2 );
32
33 m_u1_C2 = FLA_Obj_length( u1 );
34 m_v1_D2 = FLA_Obj_length( v1 );
35 n_r12t = FLA_Obj_width( r12t );
36 inc_w12t = FLA_Obj_vector_inc( w12t );
37 inc_r12t = FLA_Obj_vector_inc( r12t );
38 inc_u1 = FLA_Obj_vector_inc( u1 );
39 rs_C2 = FLA_Obj_row_stride( C2 );
40 cs_C2 = FLA_Obj_col_stride( C2 );
41 inc_v1 = FLA_Obj_vector_inc( v1 );
42 rs_D2 = FLA_Obj_row_stride( D2 );
43 cs_D2 = FLA_Obj_col_stride( D2 );
44
45 switch ( datatype )
46 {
47 case FLA_FLOAT:
48 {
49 float* tau_p = ( float* ) FLA_FLOAT_PTR( tau );
50 float* w12t_p = ( float* ) FLA_FLOAT_PTR( w12t );
51 float* r12t_p = ( float* ) FLA_FLOAT_PTR( r12t );
52 float* u1_p = ( float* ) FLA_FLOAT_PTR( u1 );
53 float* C2_p = ( float* ) FLA_FLOAT_PTR( C2 );
54 float* v1_p = ( float* ) FLA_FLOAT_PTR( v1 );
55 float* D2_p = ( float* ) FLA_FLOAT_PTR( D2 );
56
57 FLA_Apply_HUD_UT_l_ops_var1( m_u1_C2,
58 m_v1_D2,
59 n_r12t,
60 tau_p,
61 w12t_p, inc_w12t,
62 r12t_p, inc_r12t,
63 u1_p, inc_u1,
64 C2_p, rs_C2, cs_C2,
65 v1_p, inc_v1,
66 D2_p, rs_D2, cs_D2 );
67 break;
68 }
69
70 case FLA_DOUBLE:
71 {
72 double* tau_p = ( double* ) FLA_DOUBLE_PTR( tau );
73 double* w12t_p = ( double* ) FLA_DOUBLE_PTR( w12t );
74 double* r12t_p = ( double* ) FLA_DOUBLE_PTR( r12t );
75 double* u1_p = ( double* ) FLA_DOUBLE_PTR( u1 );
76 double* C2_p = ( double* ) FLA_DOUBLE_PTR( C2 );
77 double* v1_p = ( double* ) FLA_DOUBLE_PTR( v1 );
78 double* D2_p = ( double* ) FLA_DOUBLE_PTR( D2 );
79
80 FLA_Apply_HUD_UT_l_opd_var1( m_u1_C2,
81 m_v1_D2,
82 n_r12t,
83 tau_p,
84 w12t_p, inc_w12t,
85 r12t_p, inc_r12t,
86 u1_p, inc_u1,
87 C2_p, rs_C2, cs_C2,
88 v1_p, inc_v1,
89 D2_p, rs_D2, cs_D2 );
90 break;
91 }
92
93 case FLA_COMPLEX:
94 {
95 scomplex* tau_p = ( scomplex* ) FLA_COMPLEX_PTR( tau );
96 scomplex* w12t_p = ( scomplex* ) FLA_COMPLEX_PTR( w12t );
97 scomplex* r12t_p = ( scomplex* ) FLA_COMPLEX_PTR( r12t );
98 scomplex* u1_p = ( scomplex* ) FLA_COMPLEX_PTR( u1 );
99 scomplex* C2_p = ( scomplex* ) FLA_COMPLEX_PTR( C2 );
100 scomplex* v1_p = ( scomplex* ) FLA_COMPLEX_PTR( v1 );
101 scomplex* D2_p = ( scomplex* ) FLA_COMPLEX_PTR( D2 );
102
103 FLA_Apply_HUD_UT_l_opc_var1( m_u1_C2,
104 m_v1_D2,
105 n_r12t,
106 tau_p,
107 w12t_p, inc_w12t,
108 r12t_p, inc_r12t,
109 u1_p, inc_u1,
110 C2_p, rs_C2, cs_C2,
111 v1_p, inc_v1,
112 D2_p, rs_D2, cs_D2 );
113 break;
114 }
115
116 case FLA_DOUBLE_COMPLEX:
117 {
118 dcomplex* tau_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( tau );
119 dcomplex* w12t_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( w12t );
120 dcomplex* r12t_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( r12t );
121 dcomplex* u1_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( u1 );
122 dcomplex* C2_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( C2 );
123 dcomplex* v1_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( v1 );
124 dcomplex* D2_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( D2 );
125
126 FLA_Apply_HUD_UT_l_opz_var1( m_u1_C2,
127 m_v1_D2,
128 n_r12t,
129 tau_p,
130 w12t_p, inc_w12t,
131 r12t_p, inc_r12t,
132 u1_p, inc_u1,
133 C2_p, rs_C2, cs_C2,
134 v1_p, inc_v1,
135 D2_p, rs_D2, cs_D2 );
136 break;
137 }
138 }
139
140 return FLA_SUCCESS;
141}
FLA_Apply_HUD_UT_l_opd_var1
FLA_Error FLA_Apply_HUD_UT_l_opd_var1(int m_u1_C2, int m_v1_D2, int n_r12t, double *tau, double *w12t, int inc_w12t, double *r12t, int inc_r12t, double *u1, int inc_u1, double *C2, int rs_C2, int cs_C2, double *v1, int inc_v1, double *D2, int rs_D2, int cs_D2)
Definition FLA_Apply_HUD_UT_l_opt_var1.c:237
FLA_Apply_HUD_UT_l_ops_var1
FLA_Error FLA_Apply_HUD_UT_l_ops_var1(int m_u1_C2, int m_v1_D2, int n_r12t, float *tau, float *w12t, int inc_w12t, float *r12t, int inc_r12t, float *u1, int inc_u1, float *C2, int rs_C2, int cs_C2, float *v1, int inc_v1, float *D2, int rs_D2, int cs_D2)
Definition FLA_Apply_HUD_UT_l_opt_var1.c:145
FLA_Apply_HUD_UT_l_opc_var1
FLA_Error FLA_Apply_HUD_UT_l_opc_var1(int m_u1_C2, int m_v1_D2, int n_r12t, scomplex *tau, scomplex *w12t, int inc_w12t, scomplex *r12t, int inc_r12t, scomplex *u1, int inc_u1, scomplex *C2, int rs_C2, int cs_C2, scomplex *v1, int inc_v1, scomplex *D2, int rs_D2, int cs_D2)
Definition FLA_Apply_HUD_UT_l_opt_var1.c:329
FLA_Apply_HUD_UT_l_opz_var1
FLA_Error FLA_Apply_HUD_UT_l_opz_var1(int m_u1_C2, int m_v1_D2, int n_r12t, dcomplex *tau, dcomplex *w12t, int inc_w12t, dcomplex *r12t, int inc_r12t, dcomplex *u1, int inc_u1, dcomplex *C2, int rs_C2, int cs_C2, dcomplex *v1, int inc_v1, dcomplex *D2, int rs_D2, int cs_D2)
Definition FLA_Apply_HUD_UT_l_opt_var1.c:421
FLA_Obj_width
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
FLA_Obj_has_zero_dim
FLA_Bool FLA_Obj_has_zero_dim(FLA_Obj A)
Definition FLA_Query.c:400
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_Apply_HUD_UT_l_opc_var1(), FLA_Apply_HUD_UT_l_opd_var1(), FLA_Apply_HUD_UT_l_ops_var1(), FLA_Apply_HUD_UT_l_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_has_zero_dim(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), FLA_Obj_width(), and i.

Referenced by FLA_Apply_HUD_UT_internal().

◆ FLA_Apply_HUD_UT_l_opz_var1()

FLA_Error FLA_Apply_HUD_UT_l_opz_var1 ( int  m_u1_C2,
int  m_v1_D2,
int  n_r12t,
dcomplex *  tau,
dcomplex *  w12t,
int  inc_w12t,
dcomplex *  r12t,
int  inc_r12t,
dcomplex *  u1,
int  inc_u1,
dcomplex *  C2,
int  rs_C2,
int  cs_C2,
dcomplex *  v1,
int  inc_v1,
dcomplex *  D2,
int  rs_D2,
int  cs_D2 
)
431{
432 dcomplex* one_p = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
433 dcomplex* minus_one_p = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
434
435 // if ( FLA_Obj_has_zero_dim( r12t ) ) return FLA_SUCCESS;
436 if ( n_r12t == 0 ) return FLA_SUCCESS;
437
438 // // w12t = r12t;
439 // FLA_Copy_external( r12t, w12t );
440 bl1_zcopyv( BLIS1_NO_CONJUGATE,
441 n_r12t,
442 r12t, inc_r12t,
443 w12t, inc_w12t );
444
445 // // w12t = w12t + u1' * C2;
446 // // = w12t + C2^T * conj(u1);
447 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, C2, u1, FLA_ONE, w12t );
448 bl1_zgemv( BLIS1_TRANSPOSE,
449 BLIS1_CONJUGATE,
450 m_u1_C2,
451 n_r12t,
452 one_p,
453 C2, rs_C2, cs_C2,
454 u1, inc_u1,
455 one_p,
456 w12t, inc_w12t );
457
458 // // w12t = w12t + v1' * D2;
459 // // = w12t + D2^T * conj(v1);
460 // FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, D2, v1, FLA_ONE, w12t );
461 bl1_zgemv( BLIS1_TRANSPOSE,
462 BLIS1_CONJUGATE,
463 m_v1_D2,
464 n_r12t,
465 one_p,
466 D2, rs_D2, cs_D2,
467 v1, inc_v1,
468 one_p,
469 w12t, inc_w12t );
470
471 // // w12t = w12t / tau;
472 // FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w12t );
473 bl1_zinvscalv( BLIS1_NO_CONJUGATE,
474 n_r12t,
475 tau,
476 w12t, inc_w12t );
477
478 // // r12t = - w12t + r12t;
479 // FLA_Axpy_external( FLA_MINUS_ONE, w12t, r12t );
480 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
481 n_r12t,
482 minus_one_p,
483 w12t, inc_w12t,
484 r12t, inc_r12t );
485
486 // // C2 = - u1 * w12t + C2;
487 // FLA_Ger_external( FLA_MINUS_ONE, u1, w12t, C2 );
488 bl1_zger( BLIS1_NO_CONJUGATE,
489 BLIS1_NO_CONJUGATE,
490 m_u1_C2,
491 n_r12t,
492 minus_one_p,
493 u1, inc_u1,
494 w12t, inc_w12t,
495 C2, rs_C2, cs_C2 );
496
497 // // D2 = v1 * w12t + D2;
498 // FLA_Ger_external( FLA_ONE, v1, w12t, D2 );
499 bl1_zger( BLIS1_NO_CONJUGATE,
500 BLIS1_NO_CONJUGATE,
501 m_v1_D2,
502 n_r12t,
503 one_p,
504 v1, inc_v1,
505 w12t, inc_w12t,
506 D2, rs_D2, cs_D2 );
507
508 return FLA_SUCCESS;
509}
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_zcopyv
void bl1_zcopyv(conj1_t conj, int m, dcomplex *x, int incx, dcomplex *y, int incy)
Definition bl1_copyv.c:63
bl1_zgemv
void bl1_zgemv(trans1_t transa, conj1_t conjx, int m, int n, dcomplex *alpha, dcomplex *a, int a_rs, int a_cs, dcomplex *x, int incx, dcomplex *beta, dcomplex *y, int incy)
Definition bl1_gemv.c:255
bl1_zger
void bl1_zger(conj1_t conjx, conj1_t conjy, int m, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *a, int a_rs, int a_cs)
Definition bl1_ger.c:194
bl1_zinvscalv
void bl1_zinvscalv(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx)
Definition bl1_invscalv.c:78

References bl1_zaxpyv(), bl1_zcopyv(), bl1_zgemv(), bl1_zger(), bl1_zinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_Apply_HUD_UT_l_opt_var1(), and FLA_UDdate_UT_opz_var1().

◆ FLA_Apply_HUD_UT_l_unb_var1()

FLA_Error FLA_Apply_HUD_UT_l_unb_var1 ( FLA_Obj  tau,
FLA_Obj  w12t,
FLA_Obj  r12t,
FLA_Obj  u1,
FLA_Obj  C2,
FLA_Obj  v1,
FLA_Obj  D2 
)
Generated on Thu Sep 23 2021 17:46:17 for libflame by doxygen 1.9.8