libflame revision_anchor
Functions
FLA_Bidiag_UT_realify.c File Reference

(r)

Functions

FLA_Error FLA_Bidiag_UT_realify (FLA_Obj A, FLA_Obj d, FLA_Obj e)
 
FLA_Error FLA_Bidiag_UT_l_realify_unb (FLA_Obj A, FLA_Obj d, FLA_Obj e)
 
FLA_Error FLA_Bidiag_UT_l_realify_opt (FLA_Obj A, FLA_Obj d, FLA_Obj e)
 
FLA_Error FLA_Bidiag_UT_u_realify_unb (FLA_Obj A, FLA_Obj d, FLA_Obj e)
 
FLA_Error FLA_Bidiag_UT_u_realify_opt (FLA_Obj A, FLA_Obj d, FLA_Obj e)
 

Function Documentation

◆ FLA_Bidiag_UT_l_realify_opt()

FLA_Error FLA_Bidiag_UT_l_realify_opt ( FLA_Obj A,
FLA_Obj d,
FLA_Obj e )
160{
161 FLA_Datatype datatype;
162 int m_A, n_A;
163 int min_m_n;
164 int rs_A, cs_A;
165 int inc_d;
166 int inc_e;
167 int i;
168
169 datatype = FLA_Obj_datatype( A );
170
171 m_A = FLA_Obj_length( A );
172 n_A = FLA_Obj_width( A );
173 min_m_n = FLA_Obj_min_dim( A );
174
175 rs_A = FLA_Obj_row_stride( A );
176 cs_A = FLA_Obj_col_stride( A );
177
178 inc_d = FLA_Obj_vector_inc( d );
179
180 inc_e = FLA_Obj_vector_inc( e );
181
182
183 switch ( datatype )
184 {
185 case FLA_FLOAT:
186 {
187 float* buff_d = FLA_FLOAT_PTR( d );
188 float* buff_e = FLA_FLOAT_PTR( e );
189 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
190
191 bl1_ssetv( min_m_n,
192 buff_1,
193 buff_d, inc_d );
194
195 bl1_ssetv( min_m_n,
196 buff_1,
197 buff_e, inc_e );
198
199 break;
200 }
201
202 case FLA_DOUBLE:
203 {
204 double* buff_d = FLA_DOUBLE_PTR( d );
205 double* buff_e = FLA_DOUBLE_PTR( e );
206 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
207
208 bl1_dsetv( min_m_n,
209 buff_1,
210 buff_d, inc_d );
211
212 bl1_dsetv( min_m_n,
213 buff_1,
214 buff_e, inc_e );
215
216 break;
217 }
218
219 case FLA_COMPLEX:
220 {
221 scomplex* buff_A = FLA_COMPLEX_PTR( A );
222 scomplex* buff_d = FLA_COMPLEX_PTR( d );
223 scomplex* buff_e = FLA_COMPLEX_PTR( e );
224 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
225 float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
226
227 for ( i = 0; i < min_m_n; ++i )
228 {
229
230 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
231 scomplex* delta1 = buff_d + (i )*inc_d;
232 scomplex* epsilon1 = buff_e + (i )*inc_e;
233 scomplex absv;
234
235 int m_ahead = m_A - i - 1;
236 int m_behind = i;
237
238 if ( m_behind == 0 )
239 {
240 // FLA_Set( FLA_ONE, delta1 );
241 *delta1 = *buff_1;
242 }
243 else
244 {
245 scomplex* a10t_r = buff_A + (i-1)*cs_A + (i )*rs_A;
246 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, a10t_r, delta1 );
247 // FLA_Copyt( FLA_NO_TRANSPOSE, a10t_r, absv );
248 // FLA_Absolute_value( absv );
249 // FLA_Inv_scal( absv, delta1 );
250 bl1_ccopys( BLIS1_CONJUGATE, a10t_r, delta1 );
251 bl1_cabsval2( a10t_r, &absv );
252 bl1_cinvscals( &absv, delta1 );
253
254 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, a10t_r );
255 // FLA_Obj_set_imag_part( FLA_ZERO, a10t_r );
256 bl1_cscals( delta1, a10t_r );
257 a10t_r->imag = *buff_0;
258
259 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, alpha11 );
260 bl1_cscals( delta1, alpha11 );
261 }
262
263 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha11, epsilon1 );
264 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha11, absv );
265 // FLA_Absolute_value( absv );
266 // FLA_Inv_scal( absv, epsilon1 );
267 bl1_ccopys( BLIS1_CONJUGATE, alpha11, epsilon1 );
268 bl1_cabsval2( alpha11, &absv );
269 bl1_cinvscals( &absv, epsilon1 );
270
271 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, alpha11 );
272 // FLA_Obj_set_imag_part( FLA_ZERO, alpha11 );
273 bl1_cscals( epsilon1, alpha11 );
274 alpha11->imag = *buff_0;
275
276 if ( m_ahead > 0 )
277 {
278 scomplex* a21_t = buff_A + (i )*cs_A + (i+1)*rs_A;
279 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, a21_t );
280 bl1_cscals( epsilon1, a21_t );
281 }
282 }
283
284 break;
285 }
286
287 case FLA_DOUBLE_COMPLEX:
288 {
289 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
290 dcomplex* buff_d = FLA_DOUBLE_COMPLEX_PTR( d );
291 dcomplex* buff_e = FLA_DOUBLE_COMPLEX_PTR( e );
292 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
293 double* buff_0 = FLA_DOUBLE_PTR( FLA_ZERO );
294
295 for ( i = 0; i < min_m_n; ++i )
296 {
297
298 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
299 dcomplex* delta1 = buff_d + (i )*inc_d;
300 dcomplex* epsilon1 = buff_e + (i )*inc_e;
301 dcomplex absv;
302
303 int m_ahead = m_A - i - 1;
304 int m_behind = i;
305
306 if ( m_behind == 0 )
307 {
308 // FLA_Set( FLA_ONE, delta1 );
309 *delta1 = *buff_1;
310 }
311 else
312 {
313 dcomplex* a10t_r = buff_A + (i-1)*cs_A + (i )*rs_A;
314 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, a10t_r, delta1 );
315 // FLA_Copyt( FLA_NO_TRANSPOSE, a10t_r, absv );
316 // FLA_Absolute_value( absv );
317 // FLA_Inv_scal( absv, delta1 );
318 bl1_zcopys( BLIS1_CONJUGATE, a10t_r, delta1 );
319 bl1_zabsval2( a10t_r, &absv );
320 bl1_zinvscals( &absv, delta1 );
321
322 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, a10t_r );
323 // FLA_Obj_set_imag_part( FLA_ZERO, a10t_r );
324 bl1_zscals( delta1, a10t_r );
325 a10t_r->imag = *buff_0;
326
327 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, alpha11 );
328 bl1_zscals( delta1, alpha11 );
329 }
330
331 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha11, epsilon1 );
332 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha11, absv );
333 // FLA_Absolute_value( absv );
334 // FLA_Inv_scal( absv, epsilon1 );
335 bl1_zcopys( BLIS1_CONJUGATE, alpha11, epsilon1 );
336 bl1_zabsval2( alpha11, &absv );
337 bl1_zinvscals( &absv, epsilon1 );
338
339 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, alpha11 );
340 // FLA_Obj_set_imag_part( FLA_ZERO, alpha11 );
341 bl1_zscals( epsilon1, alpha11 );
342 alpha11->imag = *buff_0;
343
344 if ( m_ahead > 0 )
345 {
346 dcomplex* a21_t = buff_A + (i )*cs_A + (i+1)*rs_A;
347 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, a21_t );
348 bl1_zscals( epsilon1, a21_t );
349 }
350 }
351
352 break;
353 }
354 }
355
356 return FLA_SUCCESS;
357}
FLA_ZERO
FLA_Obj FLA_ZERO
Definition FLA_Init.c:20
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
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_min_dim
dim_t FLA_Obj_min_dim(FLA_Obj obj)
Definition FLA_Query.c:153
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
i
int i
Definition bl1_axmyv2.c:145
bl1_zscals
bl1_zscals(beta, rho_yz)
bl1_dsetv
void bl1_dsetv(int m, double *sigma, double *x, int incx)
Definition bl1_setv.c:39
bl1_ssetv
void bl1_ssetv(int m, float *sigma, float *x, int incx)
Definition bl1_setv.c:26
BLIS1_CONJUGATE
@ BLIS1_CONJUGATE
Definition blis_type_defs.h:82
dcomplex
Definition blis_type_defs.h:138
dcomplex::imag
double imag
Definition blis_type_defs.h:139
scomplex
Definition blis_type_defs.h:133
scomplex::imag
float imag
Definition blis_type_defs.h:134

References bl1_dsetv(), bl1_ssetv(), bl1_zscals(), BLIS1_CONJUGATE, FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), FLA_Obj_width(), FLA_ONE, FLA_ZERO, i, dcomplex::imag, and scomplex::imag.

Referenced by FLA_Bidiag_UT_realify().

◆ FLA_Bidiag_UT_l_realify_unb()

FLA_Error FLA_Bidiag_UT_l_realify_unb ( FLA_Obj A,
FLA_Obj d,
FLA_Obj e )
39{
40 FLA_Obj ATL, ATR, A00, a01, A02,
41 ABL, ABR, a10t, alpha11, a12t,
42 A20, a21, A22;
43
44 FLA_Obj dT, d0,
45 dB, delta1,
46 d2;
47
48 FLA_Obj eT, e0,
49 eB, epsilon1,
50 e2;
51
52 FLA_Obj a10t_l, a10t_r;
53
54 FLA_Obj a21_t,
55 a21_b;
56
57 FLA_Obj absv;
58
59
60 FLA_Obj_create( FLA_Obj_datatype( A ), 1, 1, 0, 0, &absv );
61
62 FLA_Part_2x2( A, &ATL, &ATR,
63 &ABL, &ABR, 0, 0, FLA_TL );
64
65 FLA_Part_2x1( d, &dT,
66 &dB, 0, FLA_TOP );
67
68 FLA_Part_2x1( e, &eT,
69 &eB, 0, FLA_TOP );
70
71 while ( FLA_Obj_min_dim( ABR ) > 0 )
72 {
73 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
74 /* ************* */ /* ************************** */
75 &a10t, /**/ &alpha11, &a12t,
76 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
77 1, 1, FLA_BR );
78
79 FLA_Repart_2x1_to_3x1( dT, &d0,
80 /* ** */ /* ****** */
81 &delta1,
82 dB, &d2, 1, FLA_BOTTOM );
83
84 FLA_Repart_2x1_to_3x1( eT, &e0,
85 /* ** */ /* ******** */
86 &epsilon1,
87 eB, &e2, 1, FLA_BOTTOM );
88
89 /*------------------------------------------------------------*/
90
91 if ( FLA_Obj_width( a10t ) == 0 )
92 {
93 // delta1 = 1;
94 FLA_Set( FLA_ONE, delta1 );
95 }
96 else
97 {
98 FLA_Part_1x2( a10t, &a10t_l, &a10t_r, 1, FLA_RIGHT );
99
100 // delta1 = conj(a10t_r) / abs(a10t_r);
101 FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, a10t_r, delta1 );
102 FLA_Copyt( FLA_NO_TRANSPOSE, a10t_r, absv );
103 FLA_Absolute_value( absv );
104 FLA_Inv_scal( absv, delta1 );
105
106 // a10t_r = delta1 * a10t_r;
107 // a10t_r.imag = 0;
108 FLA_Scalc( FLA_NO_CONJUGATE, delta1, a10t_r );
109 FLA_Obj_set_imag_part( FLA_ZERO, a10t_r );
110
111 // alpha11 = delta1 * alpha11;
112 FLA_Scalc( FLA_NO_CONJUGATE, delta1, alpha11 );
113 }
114
115 // epsilon1 = conj(alpha11) / abs(alpha11);
116 FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha11, epsilon1 );
117 FLA_Copyt( FLA_NO_TRANSPOSE, alpha11, absv );
118 FLA_Absolute_value( absv );
119 FLA_Inv_scal( absv, epsilon1 );
120
121 // alpha11 = epsilon1 * alpha11;
122 // alpha11.imag = 0;
123 FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, alpha11 );
124 FLA_Obj_set_imag_part( FLA_ZERO, alpha11 );
125
126 if ( FLA_Obj_length( a21 ) > 0 )
127 {
128 FLA_Part_2x1( a21, &a21_t,
129 &a21_b, 1, FLA_TOP );
130
131 // a21_t = epsilon1 * a21_t;
132 FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, a21_t );
133 }
134
135 /*------------------------------------------------------------*/
136
137 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
138 a10t, alpha11, /**/ a12t,
139 /* ************** */ /* ************************ */
140 &ABL, /**/ &ABR, A20, a21, /**/ A22,
141 FLA_TL );
142
143 FLA_Cont_with_3x1_to_2x1( &dT, d0,
144 delta1,
145 /* ** */ /* ****** */
146 &dB, d2, FLA_TOP );
147
148 FLA_Cont_with_3x1_to_2x1( &eT, e0,
149 epsilon1,
150 /* ** */ /* ******** */
151 &eB, e2, FLA_TOP );
152 }
153
154 FLA_Obj_free( &absv );
155
156 return FLA_SUCCESS;
157}
FLA_Scalc
FLA_Error FLA_Scalc(FLA_Conj conjalpha, FLA_Obj alpha, FLA_Obj A)
Definition FLA_Scalc.c:13
FLA_Inv_scal
FLA_Error FLA_Inv_scal(FLA_Obj alpha, FLA_Obj A)
Definition FLA_Inv_scal.c:13
FLA_Copyt
FLA_Error FLA_Copyt(FLA_Trans trans, FLA_Obj A, FLA_Obj B)
Definition FLA_Copyt.c:15
FLA_Cont_with_3x3_to_2x2
FLA_Error FLA_Cont_with_3x3_to_2x2(FLA_Obj *ATL, FLA_Obj *ATR, FLA_Obj A00, FLA_Obj A01, FLA_Obj A02, FLA_Obj A10, FLA_Obj A11, FLA_Obj A12, FLA_Obj *ABL, FLA_Obj *ABR, FLA_Obj A20, FLA_Obj A21, FLA_Obj A22, FLA_Quadrant quadrant)
Definition FLA_View.c:304
FLA_Part_2x2
FLA_Error FLA_Part_2x2(FLA_Obj A, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition FLA_View.c:17
FLA_Cont_with_3x1_to_2x1
FLA_Error FLA_Cont_with_3x1_to_2x1(FLA_Obj *AT, FLA_Obj A0, FLA_Obj A1, FLA_Obj *AB, FLA_Obj A2, FLA_Side side)
Definition FLA_View.c:428
FLA_Repart_2x1_to_3x1
FLA_Error FLA_Repart_2x1_to_3x1(FLA_Obj AT, FLA_Obj *A0, FLA_Obj *A1, FLA_Obj AB, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition FLA_View.c:226
FLA_Obj_create
FLA_Error FLA_Obj_create(FLA_Datatype datatype, dim_t m, dim_t n, dim_t rs, dim_t cs, FLA_Obj *obj)
Definition FLA_Obj.c:55
FLA_Part_1x2
FLA_Error FLA_Part_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition FLA_View.c:110
FLA_Part_2x1
FLA_Error FLA_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition FLA_View.c:76
FLA_Obj_set_imag_part
void FLA_Obj_set_imag_part(FLA_Obj alpha, FLA_Obj beta)
Definition FLA_Misc.c:229
FLA_Repart_2x2_to_3x3
FLA_Error FLA_Repart_2x2_to_3x3(FLA_Obj ATL, FLA_Obj ATR, FLA_Obj *A00, FLA_Obj *A01, FLA_Obj *A02, FLA_Obj *A10, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj ABL, FLA_Obj ABR, FLA_Obj *A20, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition FLA_View.c:142
FLA_Obj_free
FLA_Error FLA_Obj_free(FLA_Obj *obj)
Definition FLA_Obj.c:588
FLA_Absolute_value
FLA_Error FLA_Absolute_value(FLA_Obj alpha)
Definition FLA_Absolute_value.c:13
FLA_Set
FLA_Error FLA_Set(FLA_Obj alpha, FLA_Obj A)
Definition FLA_Set.c:13
FLA_Obj_view
Definition FLA_type_defs.h:159

References FLA_Absolute_value(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Copyt(), FLA_Inv_scal(), FLA_Obj_create(), FLA_Obj_datatype(), FLA_Obj_free(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_set_imag_part(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalc(), FLA_Set(), and FLA_ZERO.

◆ FLA_Bidiag_UT_realify()

FLA_Error FLA_Bidiag_UT_realify ( FLA_Obj A,
FLA_Obj d,
FLA_Obj e )
14{
15 FLA_Error r_val = FLA_SUCCESS;
16
17 if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
18 FLA_Bidiag_UT_realify_check( A, d, e );
19
20 if ( FLA_Obj_is_real( A ) )
21 {
22 FLA_Set( FLA_ONE, d );
23 FLA_Set( FLA_ONE, e );
24 return FLA_SUCCESS;
25 }
26
27 if ( FLA_Obj_length( A ) < FLA_Obj_width( A ) )
28 //r_val = FLA_Bidiag_UT_l_realify_unb( A, d, e );
29 r_val = FLA_Bidiag_UT_l_realify_opt( A, d, e );
30 else
31 //r_val = FLA_Bidiag_UT_u_realify_unb( A, d, e );
32 r_val = FLA_Bidiag_UT_u_realify_opt( A, d, e );
33
34 return r_val;
35}
FLA_Bidiag_UT_u_realify_opt
FLA_Error FLA_Bidiag_UT_u_realify_opt(FLA_Obj A, FLA_Obj d, FLA_Obj e)
Definition FLA_Bidiag_UT_realify.c:482
FLA_Bidiag_UT_l_realify_opt
FLA_Error FLA_Bidiag_UT_l_realify_opt(FLA_Obj A, FLA_Obj d, FLA_Obj e)
Definition FLA_Bidiag_UT_realify.c:159
FLA_Bidiag_UT_realify_check
FLA_Error FLA_Bidiag_UT_realify_check(FLA_Obj A, FLA_Obj d, FLA_Obj e)
Definition FLA_Bidiag_UT_realify_check.c:13
FLA_Obj_is_real
FLA_Bool FLA_Obj_is_real(FLA_Obj A)
Definition FLA_Query.c:307
FLA_Check_error_level
unsigned int FLA_Check_error_level(void)
Definition FLA_Check.c:18
FLA_Error
int FLA_Error
Definition FLA_type_defs.h:47

References FLA_Bidiag_UT_l_realify_opt(), FLA_Bidiag_UT_realify_check(), FLA_Bidiag_UT_u_realify_opt(), FLA_Check_error_level(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, and FLA_Set().

Referenced by FLA_Svd_ext_u_unb_var1(), FLA_Svd_uv_unb_var1(), and FLA_Svd_uv_unb_var2().

◆ FLA_Bidiag_UT_u_realify_opt()

FLA_Error FLA_Bidiag_UT_u_realify_opt ( FLA_Obj A,
FLA_Obj d,
FLA_Obj e )
483{
484 FLA_Datatype datatype;
485 int m_A, n_A;
486 int min_m_n;
487 int rs_A, cs_A;
488 int inc_d;
489 int inc_e;
490 int i;
491
492 datatype = FLA_Obj_datatype( A );
493
494 m_A = FLA_Obj_length( A );
495 n_A = FLA_Obj_width( A );
496 min_m_n = FLA_Obj_min_dim( A );
497
498 rs_A = FLA_Obj_row_stride( A );
499 cs_A = FLA_Obj_col_stride( A );
500
501 inc_d = FLA_Obj_vector_inc( d );
502
503 inc_e = FLA_Obj_vector_inc( e );
504
505
506 switch ( datatype )
507 {
508 case FLA_FLOAT:
509 {
510 float* buff_d = FLA_FLOAT_PTR( d );
511 float* buff_e = FLA_FLOAT_PTR( e );
512 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
513
514 bl1_ssetv( min_m_n,
515 buff_1,
516 buff_d, inc_d );
517
518 bl1_ssetv( min_m_n,
519 buff_1,
520 buff_e, inc_e );
521
522 break;
523 }
524
525 case FLA_DOUBLE:
526 {
527 double* buff_d = FLA_DOUBLE_PTR( d );
528 double* buff_e = FLA_DOUBLE_PTR( e );
529 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
530
531 bl1_dsetv( min_m_n,
532 buff_1,
533 buff_d, inc_d );
534
535 bl1_dsetv( min_m_n,
536 buff_1,
537 buff_e, inc_e );
538
539 break;
540 }
541
542 case FLA_COMPLEX:
543 {
544 scomplex* buff_A = FLA_COMPLEX_PTR( A );
545 scomplex* buff_d = FLA_COMPLEX_PTR( d );
546 scomplex* buff_e = FLA_COMPLEX_PTR( e );
547 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
548 float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
549
550 for ( i = 0; i < min_m_n; ++i )
551 {
552 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
553 scomplex* delta1 = buff_d + (i )*inc_d;
554 scomplex* epsilon1 = buff_e + (i )*inc_e;
555 scomplex absv;
556
557 int n_ahead = n_A - i - 1;
558 int n_behind = i;
559
560 if ( n_behind == 0 )
561 {
562 // FLA_Set( FLA_ONE, epsilon1 );
563 *epsilon1 = *buff_1;
564 }
565 else
566 {
567 scomplex* a01_b = buff_A + (i )*cs_A + (i-1)*rs_A;
568 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, a01_b, epsilon1 );
569 // FLA_Copyt( FLA_NO_TRANSPOSE, a01_b, absv );
570 // FLA_Absolute_value( absv );
571 // FLA_Inv_scal( absv, epsilon1 );
572 bl1_ccopys( BLIS1_CONJUGATE, a01_b, epsilon1 );
573 bl1_cabsval2( a01_b, &absv );
574 bl1_cinvscals( &absv, epsilon1 );
575
576 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, a01_b );
577 // FLA_Obj_set_imag_part( FLA_ZERO, a01_b );
578 bl1_cscals( epsilon1, a01_b );
579 a01_b->imag = *buff_0;
580
581 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, alpha11 );
582 bl1_cscals( epsilon1, alpha11 );
583 }
584
585 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha11, delta1 );
586 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha11, absv );
587 // FLA_Absolute_value( absv );
588 // FLA_Inv_scal( absv, delta1 );
589 bl1_ccopys( BLIS1_CONJUGATE, alpha11, delta1 );
590 bl1_cabsval2( alpha11, &absv );
591 bl1_cinvscals( &absv, delta1 );
592
593 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, alpha11 );
594 // FLA_Obj_set_imag_part( FLA_ZERO, alpha11 );
595 bl1_cscals( delta1, alpha11 );
596 alpha11->imag = *buff_0;
597
598 if ( n_ahead > 0 )
599 {
600 scomplex* a12t_l = buff_A + (i+1)*cs_A + (i )*rs_A;
601 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, a12t_l );
602 bl1_cscals( delta1, a12t_l );
603 }
604 }
605
606 break;
607 }
608
609 case FLA_DOUBLE_COMPLEX:
610 {
611 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
612 dcomplex* buff_d = FLA_DOUBLE_COMPLEX_PTR( d );
613 dcomplex* buff_e = FLA_DOUBLE_COMPLEX_PTR( e );
614 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
615 double* buff_0 = FLA_DOUBLE_PTR( FLA_ZERO );
616
617 for ( i = 0; i < min_m_n; ++i )
618 {
619 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
620 dcomplex* delta1 = buff_d + (i )*inc_d;
621 dcomplex* epsilon1 = buff_e + (i )*inc_e;
622 dcomplex absv;
623
624 int n_ahead = n_A - i - 1;
625 int n_behind = i;
626
627 if ( n_behind == 0 )
628 {
629 // FLA_Set( FLA_ONE, epsilon1 );
630 *epsilon1 = *buff_1;
631 }
632 else
633 {
634 dcomplex* a01_b = buff_A + (i )*cs_A + (i-1)*rs_A;
635 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, a01_b, epsilon1 );
636 // FLA_Copyt( FLA_NO_TRANSPOSE, a01_b, absv );
637 // FLA_Absolute_value( absv );
638 // FLA_Inv_scal( absv, epsilon1 );
639 bl1_zcopys( BLIS1_CONJUGATE, a01_b, epsilon1 );
640 bl1_zabsval2( a01_b, &absv );
641 bl1_zinvscals( &absv, epsilon1 );
642
643 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, a01_b );
644 // FLA_Obj_set_imag_part( FLA_ZERO, a01_b );
645 bl1_zscals( epsilon1, a01_b );
646 a01_b->imag = *buff_0;
647
648 // FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, alpha11 );
649 bl1_zscals( epsilon1, alpha11 );
650 }
651
652 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha11, delta1 );
653 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha11, absv );
654 // FLA_Absolute_value( absv );
655 // FLA_Inv_scal( absv, delta1 );
656 bl1_zcopys( BLIS1_CONJUGATE, alpha11, delta1 );
657 bl1_zabsval2( alpha11, &absv );
658 bl1_zinvscals( &absv, delta1 );
659
660 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, alpha11 );
661 // FLA_Obj_set_imag_part( FLA_ZERO, alpha11 );
662 bl1_zscals( delta1, alpha11 );
663 alpha11->imag = *buff_0;
664
665 if ( n_ahead > 0 )
666 {
667 dcomplex* a12t_l = buff_A + (i+1)*cs_A + (i )*rs_A;
668 // FLA_Scalc( FLA_NO_CONJUGATE, delta1, a12t_l );
669 bl1_zscals( delta1, a12t_l );
670 }
671 }
672
673 break;
674 }
675 }
676
677 return FLA_SUCCESS;
678}

References bl1_dsetv(), bl1_ssetv(), bl1_zscals(), BLIS1_CONJUGATE, FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), FLA_Obj_width(), FLA_ONE, FLA_ZERO, i, dcomplex::imag, and scomplex::imag.

Referenced by FLA_Bidiag_UT_realify().

◆ FLA_Bidiag_UT_u_realify_unb()

FLA_Error FLA_Bidiag_UT_u_realify_unb ( FLA_Obj A,
FLA_Obj d,
FLA_Obj e )
362{
363 FLA_Obj ATL, ATR, A00, a01, A02,
364 ABL, ABR, a10t, alpha11, a12t,
365 A20, a21, A22;
366
367 FLA_Obj dT, d0,
368 dB, delta1,
369 d2;
370
371 FLA_Obj eT, e0,
372 eB, epsilon1,
373 e2;
374
375 FLA_Obj a01_t,
376 a01_b;
377
378 FLA_Obj a12t_l, a12t_r;
379
380 FLA_Obj absv;
381
382
383 FLA_Obj_create( FLA_Obj_datatype( A ), 1, 1, 0, 0, &absv );
384
385 FLA_Part_2x2( A, &ATL, &ATR,
386 &ABL, &ABR, 0, 0, FLA_TL );
387
388 FLA_Part_2x1( d, &dT,
389 &dB, 0, FLA_TOP );
390
391 FLA_Part_2x1( e, &eT,
392 &eB, 0, FLA_TOP );
393
394 while ( FLA_Obj_min_dim( ABR ) > 0 )
395 {
396 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
397 /* ************* */ /* ************************** */
398 &a10t, /**/ &alpha11, &a12t,
399 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
400 1, 1, FLA_BR );
401
402 FLA_Repart_2x1_to_3x1( dT, &d0,
403 /* ** */ /* ****** */
404 &delta1,
405 dB, &d2, 1, FLA_BOTTOM );
406
407 FLA_Repart_2x1_to_3x1( eT, &e0,
408 /* ** */ /* ******** */
409 &epsilon1,
410 eB, &e2, 1, FLA_BOTTOM );
411
412 /*------------------------------------------------------------*/
413
414 if ( FLA_Obj_length( a01 ) == 0 )
415 {
416 // epsilon1 = 1;
417 FLA_Set( FLA_ONE, epsilon1 );
418 }
419 else
420 {
421 FLA_Part_2x1( a01, &a01_t,
422 &a01_b, 1, FLA_BOTTOM );
423
424 // epsilon1 = conj(a01_b) / abs(a01_b);
425 FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, a01_b, epsilon1 );
426 FLA_Copyt( FLA_NO_TRANSPOSE, a01_b, absv );
427 FLA_Absolute_value( absv );
428 FLA_Inv_scal( absv, epsilon1 );
429
430 // a01_b = epsilon1 * a01_b;
431 // a01_b.imag = 0;
432 FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, a01_b );
433 FLA_Obj_set_imag_part( FLA_ZERO, a01_b );
434
435 // alpha11 = epsilon1 * alpha11;
436 FLA_Scalc( FLA_NO_CONJUGATE, epsilon1, alpha11 );
437 }
438
439 // delta1 = conj(alpha11) / abs(alpha11);
440 FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha11, delta1 );
441 FLA_Copyt( FLA_NO_TRANSPOSE, alpha11, absv );
442 FLA_Absolute_value( absv );
443 FLA_Inv_scal( absv, delta1 );
444
445 // alpha11 = delta1 * alpha11;
446 // alpha11.imag = 0;
447 FLA_Scalc( FLA_NO_CONJUGATE, delta1, alpha11 );
448 FLA_Obj_set_imag_part( FLA_ZERO, alpha11 );
449
450 if ( FLA_Obj_width( a12t ) > 0 )
451 {
452 FLA_Part_1x2( a12t, &a12t_l, &a12t_r, 1, FLA_LEFT );
453
454 // a12t_l = delta1 * a12t_l;
455 FLA_Scalc( FLA_NO_CONJUGATE, delta1, a12t_l );
456 }
457
458 /*------------------------------------------------------------*/
459
460 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
461 a10t, alpha11, /**/ a12t,
462 /* ************** */ /* ************************ */
463 &ABL, /**/ &ABR, A20, a21, /**/ A22,
464 FLA_TL );
465
466 FLA_Cont_with_3x1_to_2x1( &dT, d0,
467 delta1,
468 /* ** */ /* ****** */
469 &dB, d2, FLA_TOP );
470
471 FLA_Cont_with_3x1_to_2x1( &eT, e0,
472 epsilon1,
473 /* ** */ /* ******** */
474 &eB, e2, FLA_TOP );
475 }
476
477 FLA_Obj_free( &absv );
478
479 return FLA_SUCCESS;
480}

References FLA_Absolute_value(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Copyt(), FLA_Inv_scal(), FLA_Obj_create(), FLA_Obj_datatype(), FLA_Obj_free(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_set_imag_part(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalc(), FLA_Set(), and FLA_ZERO.

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