libflame  revision_anchor
Functions
FLA_Bidiag_UT_u_unb_var4.c File Reference

(r)

Functions

FLA_Error FLA_Bidiag_UT_u_unb_var4 (FLA_Obj A, FLA_Obj TU, FLA_Obj TV)
 
FLA_Error FLA_Bidiag_UT_u_step_unb_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj Z, FLA_Obj T, FLA_Obj S)
 

Function Documentation

◆ FLA_Bidiag_UT_u_step_unb_var4()

FLA_Error FLA_Bidiag_UT_u_step_unb_var4 ( FLA_Obj  A,
FLA_Obj  Y,
FLA_Obj  Z,
FLA_Obj  T,
FLA_Obj  S 
)
36 {
37  FLA_Obj ATL, ATR, A00, a01, A02,
38  ABL, ABR, a10t, alpha11, a12t,
39  A20, a21, A22;
40  FLA_Obj YTL, YTR, Y00, y01, Y02,
41  YBL, YBR, y10t, psi11, y12t,
42  Y20, y21, Y22;
43  FLA_Obj ZTL, ZTR, Z00, z01, Z02,
44  ZBL, ZBR, z10t, zeta11, z12t,
45  Z20, z21, Z22;
46  FLA_Obj TTL, TTR, T00, t01, T02,
47  TBL, TBR, t10t, tau11, t12t,
48  T20, t21, T22;
49  FLA_Obj STL, STR, S00, s01, S02,
50  SBL, SBR, s10t, sigma11, s12t,
51  S20, s21, S22;
52  FLA_Obj wT, w01,
53  wB, omega11,
54  w21;
55  FLA_Obj alT, a01l,
56  alB, alpha11l,
57  a22l;
58  FLA_Obj apT, a01p,
59  apB, alpha11p,
60  a12p;
61  FLA_Obj uT, u01,
62  uB, upsilon11,
63  u21;
64  FLA_Obj uTp, u01p,
65  uBp, upsilon11p,
66  u21p;
67  FLA_Obj vT, v01,
68  vB, nu11,
69  v21;
70  FLA_Obj dT, d0,
71  dB, delta1,
72  d2;
73  FLA_Obj eT, e0,
74  eB, epsilon1,
75  e2;
76  FLA_Obj fT, f0,
77  fB, phi1,
78  f2;
79  FLA_Obj gT, g0,
80  gB, ghi1,
81  g2;
82  FLA_Obj w, al, ap, u, up, v;
83  FLA_Obj d, e, f, g;
84 
85  FLA_Obj minus_inv_tau11;
86  FLA_Obj last_elem;
87  FLA_Obj beta;
88  FLA_Obj alpha12;
89  FLA_Obj minus_alpha12;
90  FLA_Obj minus_conj_alpha12;
91  FLA_Obj psi11_minus_alpha12;
92  FLA_Obj minus_upsilon11;
93  FLA_Obj minus_conj_nu11;
94  FLA_Obj minus_conj_psi11;
95  FLA_Obj minus_zeta11;
96 
97  FLA_Obj a01_t,
98  a01_b;
99  FLA_Obj A02_l, A02_r;
100  FLA_Obj a12t_l, a12t_r;
101  FLA_Obj a12p_t,
102  a12p_b;
103  FLA_Obj A22_l, A22_r;
104  FLA_Obj v21_t,
105  v21_b;
106  FLA_Obj Y20_t,
107  Y20_b;
108  FLA_Obj a2;
109 
110  FLA_Datatype datatype_A;
111  dim_t m_A, n_A;
112  dim_t b_alg;
113 
114 
115  b_alg = FLA_Obj_length( T );
116 
117  datatype_A = FLA_Obj_datatype( A );
118  m_A = FLA_Obj_length( A );
119  n_A = FLA_Obj_width( A );
120 
121  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_inv_tau11 );
122  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &last_elem );
123  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &beta );
124  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &alpha12 );
125  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_alpha12 );
126  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_conj_alpha12 );
127  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &psi11_minus_alpha12 );
128  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_upsilon11 );
129  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_conj_nu11 );
130  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_conj_psi11 );
131  FLA_Obj_create( datatype_A, 1, 1, 0, 0, &minus_zeta11 );
132  FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &w );
133  FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &al );
134  FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &ap );
135  FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &u );
136  FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &up );
137  FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &v );
138  FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &d );
139  FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &e );
140  FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &f );
141  FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &g );
142 
143  FLA_Set( FLA_ZERO, Y );
144  FLA_Set( FLA_ZERO, Z );
145 
146  FLA_Part_2x2( A, &ATL, &ATR,
147  &ABL, &ABR, 0, 0, FLA_TL );
148  FLA_Part_2x2( Y, &YTL, &YTR,
149  &YBL, &YBR, 0, 0, FLA_TL );
150  FLA_Part_2x2( Z, &ZTL, &ZTR,
151  &ZBL, &ZBR, 0, 0, FLA_TL );
152  FLA_Part_2x2( T, &TTL, &TTR,
153  &TBL, &TBR, 0, 0, FLA_TL );
154  FLA_Part_2x2( S, &STL, &STR,
155  &SBL, &SBR, 0, 0, FLA_TL );
156  FLA_Part_2x1( w, &wT,
157  &wB, 0, FLA_TOP );
158  FLA_Part_2x1( al, &alT,
159  &alB, 0, FLA_TOP );
160  FLA_Part_2x1( ap, &apT,
161  &apB, 0, FLA_TOP );
162  FLA_Part_2x1( u, &uT,
163  &uB, 0, FLA_TOP );
164  FLA_Part_2x1( up, &uTp,
165  &uBp, 0, FLA_TOP );
166  FLA_Part_2x1( v, &vT,
167  &vB, 0, FLA_TOP );
168  FLA_Part_2x1( d, &dT,
169  &dB, 0, FLA_TOP );
170  FLA_Part_2x1( e, &eT,
171  &eB, 0, FLA_TOP );
172  FLA_Part_2x1( f, &fT,
173  &fB, 0, FLA_TOP );
174  FLA_Part_2x1( g, &gT,
175  &gB, 0, FLA_TOP );
176 
177  while ( FLA_Obj_length( ATL ) < b_alg )
178  {
179  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
180  /* ************* */ /* ************************** */
181  &a10t, /**/ &alpha11, &a12t,
182  ABL, /**/ ABR, &A20, /**/ &a21, &A22,
183  1, 1, FLA_BR );
184  FLA_Repart_2x2_to_3x3( YTL, /**/ YTR, &Y00, /**/ &y01, &Y02,
185  /* ************* */ /* ************************ */
186  &y10t, /**/ &psi11, &y12t,
187  YBL, /**/ YBR, &Y20, /**/ &y21, &Y22,
188  1, 1, FLA_BR );
189  FLA_Repart_2x2_to_3x3( ZTL, /**/ ZTR, &Z00, /**/ &z01, &Z02,
190  /* ************* */ /* ************************* */
191  &z10t, /**/ &zeta11, &z12t,
192  ZBL, /**/ ZBR, &Z20, /**/ &z21, &Z22,
193  1, 1, FLA_BR );
194  FLA_Repart_2x2_to_3x3( TTL, /**/ TTR, &T00, /**/ &t01, &T02,
195  /* ************* */ /* ************************** */
196  &t10t, /**/ &tau11, &t12t,
197  TBL, /**/ TBR, &T20, /**/ &t21, &T22,
198  1, 1, FLA_BR );
199  FLA_Repart_2x2_to_3x3( STL, /**/ STR, &S00, /**/ &s01, &S02,
200  /* ************* */ /* ************************** */
201  &s10t, /**/ &sigma11, &s12t,
202  SBL, /**/ SBR, &S20, /**/ &s21, &S22,
203  1, 1, FLA_BR );
204  FLA_Repart_2x1_to_3x1( wT, &w01,
205  /* ** */ /* ***** */
206  &omega11,
207  wB, &w21, 1, FLA_BOTTOM );
208  FLA_Repart_2x1_to_3x1( alT, &a01l,
209  /* ** */ /* ***** */
210  &alpha11l,
211  alB, &a22l, 1, FLA_BOTTOM );
212  FLA_Repart_2x1_to_3x1( apT, &a01p,
213  /* ** */ /* ***** */
214  &alpha11p,
215  apB, &a12p, 1, FLA_BOTTOM );
216  FLA_Repart_2x1_to_3x1( uT, &u01,
217  /* ** */ /* ***** */
218  &upsilon11,
219  uB, &u21, 1, FLA_BOTTOM );
220  FLA_Repart_2x1_to_3x1( uTp, &u01p,
221  /* ** */ /* ***** */
222  &upsilon11p,
223  uBp, &u21p, 1, FLA_BOTTOM );
224  FLA_Repart_2x1_to_3x1( vT, &v01,
225  /* ** */ /* ***** */
226  &nu11,
227  vB, &v21, 1, FLA_BOTTOM );
228  FLA_Repart_2x1_to_3x1( dT, &d0,
229  /* ** */ /* ****** */
230  &delta1,
231  dB, &d2, 1, FLA_BOTTOM );
232  FLA_Repart_2x1_to_3x1( eT, &e0,
233  /* ** */ /* ******** */
234  &epsilon1,
235  eB, &e2, 1, FLA_BOTTOM );
236  FLA_Repart_2x1_to_3x1( fT, &f0,
237  /* ** */ /* **** */
238  &phi1,
239  fB, &f2, 1, FLA_BOTTOM );
240  FLA_Repart_2x1_to_3x1( gT, &g0,
241  /* ** */ /* **** */
242  &ghi1,
243  gB, &g2, 1, FLA_BOTTOM );
244 
245  /*------------------------------------------------------------*/
246 
247  // Save last element of a01 and set it to one so we can use a01 as
248  // v10t^T in subsequent computations. We will restore a01_b later on.
249  // Also note: V20^T is stored in A02.
250  if ( FLA_Obj_length( ATL ) > 0 )
251  {
252  FLA_Part_2x1( a01, &a01_t,
253  &a01_b, 1, FLA_BOTTOM );
254  FLA_Copy( a01_b, last_elem );
255  FLA_Set( FLA_ONE, a01_b );
256  }
257 
258  FLA_Merge_2x1( alpha11,
259  a21, &a2 );
260 
261  // alpha11 = alpha11 - u10t * y10t' - z10t * v10t';
262  // a21 = a21 - U20 * y10t' - Z20 * v10t';
263  FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, ABL, y10t, FLA_ONE, a2 );
264  FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, ZBL, a01, FLA_ONE, a2 );
265 
266  // a12t = a12t - u10t * Y20' - z10t * V20';
267  FLA_Gemv( FLA_CONJ_NO_TRANSPOSE, FLA_MINUS_ONE, Y20, a10t, FLA_ONE, a12t );
268  FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_MINUS_ONE, A02, z10t, FLA_ONE, a12t );
269 
270  // Restore last element of a01.
271  if ( FLA_Obj_length( ATL ) > 0 )
272  {
273  FLA_Copy( last_elem, a01_b );
274  }
275 
276  // [ alpha11, u21p, tau11 ] = House2( alpha11, a21 );
277  FLA_Househ2_UT( FLA_LEFT,
278  alpha11,
279  a21, tau11 );
280  FLA_Copy( a21, u21p );
281 
282  if ( FLA_Obj_width( A22 ) > 0 )
283  {
284  // minus_inv_tau11 = - 1 / tau11;
285  FLA_Copy( FLA_MINUS_ONE, minus_inv_tau11 );
286  FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, minus_inv_tau11 );
287 
288  // a12p = ( tau11 - 1 ) * a12t^T / tau11;
289  // = a12t^T - ( 1 / tau11 ) * a12t^T;
290  FLA_Copyt( FLA_TRANSPOSE, a12t, a12p );
291  FLA_Axpyt( FLA_TRANSPOSE, minus_inv_tau11, a12t, a12p );
292 
293  // y21 = - Y20 * ( U20' * u21p ) - V20 * ( Z20' * u21p );
294  FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, u21p, FLA_ZERO, d0 );
295  FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, Z20, u21p, FLA_ZERO, e0 );
296 
297  FLA_Set( FLA_ZERO, y21 );
298  FLA_Gemv( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, Y20, d0, FLA_ONE, y21 );
299  FLA_Gemv( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, e0, FLA_ONE, y21 );
300 
301  // t01 = a10t' + U20' * u21;
302  FLA_Copyt( FLA_CONJ_TRANSPOSE, a10t, t01 );
303  FLA_Axpy( FLA_ONE, d0, t01 );
304 
305  // y21 = y21 + A22' * u21p;
306  FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A22, u21p, FLA_ONE, y21 );
307 
308  // a12p = a12p - conj(y21) / tau11;
309  FLA_Axpyt( FLA_CONJ_NO_TRANSPOSE, minus_inv_tau11, y21, a12p );
310 
311  // w21 = A22 * conj(a12p);
312  FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A22, a12p, FLA_ZERO, w21 );
313 
314  // w21 = w21 - U20 * ( Y20' * conj(a12p) ) - Z20 * ( V20' * conj(a12p) );
315  FLA_Gemvc( FLA_CONJ_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, Y20, a12p, FLA_ZERO, f0 );
316  FLA_Gemvc( FLA_CONJ_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A02, a12p, FLA_ZERO, g0 );
317 
318  FLA_Gemv( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, f0, FLA_ONE, w21 );
319  FLA_Gemv( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, Z20, g0, FLA_ONE, w21 );
320 
321  FLA_Part_1x2( A22, &A22_l, &A22_r, 1, FLA_LEFT );
322  FLA_Part_2x1( Y20, &Y20_t,
323  &Y20_b, 1, FLA_TOP );
324  FLA_Part_1x2( A02, &A02_l, &A02_r, 1, FLA_LEFT );
325 
326  // a22l = A22 * e0 - U20 * ( Y20' * e0 ) - Z20 * ( V20' * e0 );
327  FLA_Copy( A22_l, a22l );
328  FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, Y20_t, FLA_ONE, a22l );
329  FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, Z20, A02_l, FLA_ONE, a22l );
330 
331  // y21 = y21 + conj(a12t)^T;
332  FLA_Axpyt( FLA_CONJ_TRANSPOSE, FLA_ONE, a12t, y21 );
333 
334  FLA_Part_1x2( a12t, &a12t_l, &a12t_r, 1, FLA_LEFT );
335  FLA_Part_2x1( v21, &v21_t,
336  &v21_b, 1, FLA_TOP );
337  FLA_Part_2x1( a12p, &a12p_t,
338  &a12p_b, 1, FLA_TOP );
339 
340  // [ alpha12, psi11_minus_alpha12, sigma11 ] = House2s( a12p_t, a12p_b );
341  FLA_Househ2s_UT( FLA_RIGHT,
342  a12p_t,
343  a12p_b,
344  alpha12, psi11_minus_alpha12, sigma11 );
345 
346  // v21 = conj( ( a12p - alpha12 * e0 ) / ( psi11 - alpha12 ) );
347  FLA_Copy( a12p, v21 );
348  FLA_Mult_add( FLA_MINUS_ONE, alpha12, v21_t );
349  FLA_Inv_scalc( FLA_NO_CONJUGATE, psi11_minus_alpha12, v21 );
350  FLA_Conjugate( v21_b );
351 
352  // minus_conj_alpha12 = - conj(alpha12);
353  FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha12, minus_conj_alpha12 );
354  FLA_Scal( FLA_MINUS_ONE, minus_conj_alpha12 );
355 
356  // s01 = V20' * v21;
357  // = conj(V02) * v21;
358  // = conj(V02) * conj( ( a12p - alpha12 * e0 ) / ( psi11 - alpha12 ) );
359  // = conj(V02) * ( conj(a12p) - conj(alpha12) * e0 ) / conj( psi11 - alpha12 ) );
360  // = ( conj(V02) * conj(a12p) - conj(V02) * conj(alpha12) * e0 ) / conj( psi11 - alpha12 );
361  // = ( g0 - conj(V02) * conj(alpha12) * e0 ) / conj( psi11 - alpha12 );
362  FLA_Copy( g0, s01 );
363  FLA_Axpyt( FLA_CONJ_NO_TRANSPOSE, minus_conj_alpha12, A02_l, s01 );
364  FLA_Inv_scalc( FLA_CONJUGATE, psi11_minus_alpha12, s01 );
365 
366  // a12t_l = alpha12;
367  // a12t_r = v21_b^T;
368  FLA_Copyt( FLA_NO_TRANSPOSE, alpha12, a12t_l );
369  FLA_Copyt( FLA_TRANSPOSE, v21_b, a12t_r );
370  }
371 
372  // u21 = u21p;
373  FLA_Copy( u21p, u21 );
374 
375  if ( FLA_Obj_width( A22 ) > 0 )
376  {
377  // beta = - y21' * v21 / tau11;
378  FLA_Dotc( FLA_CONJUGATE, y21, v21, beta );
379  FLA_Scal( FLA_MINUS_ONE, beta );
380  FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, beta );
381 
382  // z21 = ( w21 - conj(alpha12) * a22l ) / conj(psi11 - alpha12) + beta * u21;
383  FLA_Copy( w21, z21 );
384  FLA_Axpy( minus_conj_alpha12, a22l, z21 );
385  FLA_Inv_scalc( FLA_CONJUGATE, psi11_minus_alpha12, z21 );
386  FLA_Axpy( beta, u21, z21 );
387 
388  // y21 = y21 / tau11;
389  // z21 = z21 / sigma11;
390  FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, y21 );
391  FLA_Inv_scalc( FLA_NO_CONJUGATE, sigma11, z21 );
392  }
393  else // if ( FLA_Obj_width( A22 ) == 0 )
394  {
395  // t01 = a10t' + U20' * u21;
396  FLA_Copyt( FLA_CONJ_TRANSPOSE, a10t, t01 );
397  FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, u21, FLA_ONE, t01 );
398  }
399 
400  /*------------------------------------------------------------*/
401 
402  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
403  a10t, alpha11, /**/ a12t,
404  /* ************** */ /* ************************ */
405  &ABL, /**/ &ABR, A20, a21, /**/ A22,
406  FLA_TL );
407  FLA_Cont_with_3x3_to_2x2( &YTL, /**/ &YTR, Y00, y01, /**/ Y02,
408  y10t, psi11, /**/ y12t,
409  /* ************** */ /* ********************** */
410  &YBL, /**/ &YBR, Y20, y21, /**/ Y22,
411  FLA_TL );
412  FLA_Cont_with_3x3_to_2x2( &ZTL, /**/ &ZTR, Z00, z01, /**/ Z02,
413  z10t, zeta11, /**/ z12t,
414  /* ************** */ /* *********************** */
415  &ZBL, /**/ &ZBR, Z20, z21, /**/ Z22,
416  FLA_TL );
417  FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, t01, /**/ T02,
418  t10t, tau11, /**/ t12t,
419  /* ************** */ /* ************************ */
420  &TBL, /**/ &TBR, T20, t21, /**/ T22,
421  FLA_TL );
422  FLA_Cont_with_3x3_to_2x2( &STL, /**/ &STR, S00, s01, /**/ S02,
423  s10t, sigma11, /**/ s12t,
424  /* ************** */ /* ************************ */
425  &SBL, /**/ &SBR, S20, s21, /**/ S22,
426  FLA_TL );
427  FLA_Cont_with_3x1_to_2x1( &wT, w01,
428  omega11,
429  /* ** */ /* ***** */
430  &wB, w21, FLA_TOP );
431  FLA_Cont_with_3x1_to_2x1( &alT, a01l,
432  alpha11l,
433  /* ** */ /* ***** */
434  &alB, a22l, FLA_TOP );
435  FLA_Cont_with_3x1_to_2x1( &apT, a01p,
436  alpha11p,
437  /* ** */ /* ***** */
438  &apB, a12p, FLA_TOP );
439  FLA_Cont_with_3x1_to_2x1( &uT, u01,
440  upsilon11,
441  /* ** */ /* ***** */
442  &uB, u21, FLA_TOP );
443  FLA_Cont_with_3x1_to_2x1( &uTp, u01p,
444  upsilon11p,
445  /* ** */ /* ***** */
446  &uBp, u21p, FLA_TOP );
447  FLA_Cont_with_3x1_to_2x1( &vT, v01,
448  nu11,
449  /* ** */ /* ***** */
450  &vB, v21, FLA_TOP );
451  FLA_Cont_with_3x1_to_2x1( &dT, d0,
452  delta1,
453  /* ** */ /* ****** */
454  &dB, d2, FLA_TOP );
455  FLA_Cont_with_3x1_to_2x1( &eT, e0,
456  epsilon1,
457  /* ** */ /* ******** */
458  &eB, e2, FLA_TOP );
459  FLA_Cont_with_3x1_to_2x1( &fT, f0,
460  phi1,
461  /* ** */ /* **** */
462  &fB, f2, FLA_TOP );
463  FLA_Cont_with_3x1_to_2x1( &gT, g0,
464  ghi1,
465  /* ** */ /* **** */
466  &gB, g2, FLA_TOP );
467  }
468 
469  FLA_Obj_free( &minus_inv_tau11 );
470  FLA_Obj_free( &last_elem );
471  FLA_Obj_free( &beta );
472  FLA_Obj_free( &alpha12 );
473  FLA_Obj_free( &minus_alpha12 );
474  FLA_Obj_free( &minus_conj_alpha12 );
475  FLA_Obj_free( &psi11_minus_alpha12 );
476  FLA_Obj_free( &minus_upsilon11 );
477  FLA_Obj_free( &minus_conj_nu11 );
478  FLA_Obj_free( &minus_conj_psi11 );
479  FLA_Obj_free( &minus_zeta11 );
480  FLA_Obj_free( &w );
481  FLA_Obj_free( &al );
482  FLA_Obj_free( &ap );
483  FLA_Obj_free( &u );
484  FLA_Obj_free( &up );
485  FLA_Obj_free( &v );
486  FLA_Obj_free( &d );
487  FLA_Obj_free( &e );
488  FLA_Obj_free( &f );
489  FLA_Obj_free( &g );
490 
491  return FLA_SUCCESS;
492 }
FLA_Error FLA_Copy(FLA_Obj A, FLA_Obj B)
Definition: FLA_Copy.c:15
FLA_Error FLA_Scal(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal.c:15
FLA_Error FLA_Dotc(FLA_Conj conj, FLA_Obj x, FLA_Obj y, FLA_Obj rho)
Definition: FLA_Dotc.c:13
FLA_Error FLA_Axpy(FLA_Obj alpha, FLA_Obj A, FLA_Obj B)
Definition: FLA_Axpy.c:15
FLA_Error FLA_Axpyt(FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj B)
Definition: FLA_Axpyt.c:15
FLA_Error FLA_Copyt(FLA_Trans trans, FLA_Obj A, FLA_Obj B)
Definition: FLA_Copyt.c:15
FLA_Error FLA_Inv_scalc(FLA_Conj conjalpha, FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Inv_scalc.c:13
FLA_Error FLA_Gemvc(FLA_Trans transa, FLA_Conj conjx, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemvc.c:13
FLA_Error FLA_Gemv(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemv.c:15
FLA_Obj FLA_MINUS_ONE
Definition: FLA_Init.c:22
FLA_Obj FLA_ZERO
Definition: FLA_Init.c:20
FLA_Obj FLA_ONE
Definition: FLA_Init.c:18
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_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_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_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
dim_t FLA_Obj_width(FLA_Obj obj)
Definition: FLA_Query.c:123
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_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_Error FLA_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition: FLA_View.c:76
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
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_Error FLA_Merge_2x1(FLA_Obj AT, FLA_Obj AB, FLA_Obj *A)
Definition: FLA_View.c:541
FLA_Error FLA_Obj_free(FLA_Obj *obj)
Definition: FLA_Obj.c:588
FLA_Datatype FLA_Obj_datatype(FLA_Obj obj)
Definition: FLA_Query.c:13
int FLA_Datatype
Definition: FLA_type_defs.h:49
unsigned long dim_t
Definition: FLA_type_defs.h:71
FLA_Error FLA_Mult_add(FLA_Obj alpha, FLA_Obj beta, FLA_Obj gamma)
Definition: FLA_Mult_add.c:13
FLA_Error FLA_Set(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Set.c:13
FLA_Error FLA_Conjugate(FLA_Obj A)
Definition: FLA_Conjugate.c:13
FLA_Error FLA_Househ2s_UT(FLA_Side side, FLA_Obj chi_1, FLA_Obj x2, FLA_Obj alpha, FLA_Obj chi_1_minus_alpha, FLA_Obj tau)
Definition: FLA_Househ2s_UT.c:16
FLA_Error FLA_Househ2_UT(FLA_Side side, FLA_Obj chi_1, FLA_Obj x2, FLA_Obj tau)
Definition: FLA_Househ2_UT.c:16
Definition: FLA_type_defs.h:159

References FLA_Axpy(), FLA_Axpyt(), FLA_Conjugate(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy(), FLA_Copyt(), FLA_Dotc(), FLA_Gemv(), FLA_Gemvc(), FLA_Househ2_UT(), FLA_Househ2s_UT(), FLA_Inv_scalc(), FLA_Merge_2x1(), FLA_MINUS_ONE, FLA_Mult_add(), FLA_Obj_create(), FLA_Obj_datatype(), FLA_Obj_free(), FLA_Obj_length(), 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_Scal(), FLA_Set(), and FLA_ZERO.

Referenced by FLA_Bidiag_UT_u_unb_var4().

◆ FLA_Bidiag_UT_u_unb_var4()

FLA_Error FLA_Bidiag_UT_u_unb_var4 ( FLA_Obj  A,
FLA_Obj  TU,
FLA_Obj  TV 
)
14 {
15  FLA_Error r_val;
16  FLA_Obj Y, Z;
17  FLA_Datatype datatype_A;
18  dim_t m_A, n_A;
19 
20  datatype_A = FLA_Obj_datatype( A );
21  m_A = FLA_Obj_length( A );
22  n_A = FLA_Obj_width( A );
23 
24  FLA_Obj_create( datatype_A, n_A, n_A, 0, 0, &Y );
25  FLA_Obj_create( datatype_A, m_A, n_A, 0, 0, &Z );
26 
27  r_val = FLA_Bidiag_UT_u_step_unb_var4( A, Y, Z, TU, TV );
28 
29  FLA_Obj_free( &Y );
30  FLA_Obj_free( &Z );
31 
32  return r_val;
33 }
FLA_Error FLA_Bidiag_UT_u_step_unb_var4(FLA_Obj A, FLA_Obj Y, FLA_Obj Z, FLA_Obj T, FLA_Obj S)
Definition: FLA_Bidiag_UT_u_unb_var4.c:35
int FLA_Error
Definition: FLA_type_defs.h:47

References FLA_Bidiag_UT_u_step_unb_var4(), FLA_Obj_create(), FLA_Obj_datatype(), FLA_Obj_free(), FLA_Obj_length(), and FLA_Obj_width().

Referenced by FLA_Bidiag_UT_u().