libflame revision_anchor
Functions
FLA_QR2_UT_vars.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_QR2_UT_blk_var1 (FLA_Obj U, FLA_Obj D, FLA_Obj T, fla_qr2ut_t *cntl)
 
FLA_Error FLA_QR2_UT_blk_var2 (FLA_Obj U, FLA_Obj D, FLA_Obj T, fla_qr2ut_t *cntl)
 
FLA_Error FLA_QR2_UT_unb_var1 (FLA_Obj U, FLA_Obj D, FLA_Obj T)
 
FLA_Error FLA_QR2_UT_opt_var1 (FLA_Obj U, FLA_Obj D, FLA_Obj T)
 
FLA_Error FLA_QR2_UT_ops_var1 (int m_UT, int m_D, float *U, int rs_U, int cs_U, float *D, int rs_D, int cs_D, float *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR2_UT_opd_var1 (int m_UT, int m_D, double *U, int rs_U, int cs_U, double *D, int rs_D, int cs_D, double *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR2_UT_opc_var1 (int m_UT, int m_D, scomplex *U, int rs_U, int cs_U, scomplex *D, int rs_D, int cs_D, scomplex *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR2_UT_opz_var1 (int m_UT, int m_D, dcomplex *U, int rs_U, int cs_U, dcomplex *D, int rs_D, int cs_D, dcomplex *T, int rs_T, int cs_T)
 

Function Documentation

◆ FLA_QR2_UT_blk_var1()

FLA_Error FLA_QR2_UT_blk_var1 ( FLA_Obj  U,
FLA_Obj  D,
FLA_Obj  T,
fla_qr2ut_t *  cntl 
)
15{
16 FLA_Obj UTL, UTR, U00, U01, U02,
17 UBL, UBR, U10, U11, U12,
18 U20, U21, U22;
19
20 FLA_Obj DL, DR, D0, D1, D2;
21
22 FLA_Obj TL, TR, T0, T1, W12;
23
24 FLA_Obj W12T, W12B;
25
26 FLA_Obj T1T, T2B;
27
28 dim_t b_alg, b;
29
30 // Query the algorithmic blocksize by inspecting the length of T.
31 b_alg = FLA_Obj_length( T );
32
33 FLA_Part_2x2( U, &UTL, &UTR,
34 &UBL, &UBR, 0, 0, FLA_TL );
35
36 FLA_Part_1x2( D, &DL, &DR, 0, FLA_LEFT );
37
38 FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT );
39
40 while ( FLA_Obj_min_dim( UBR ) > 0 ){
41
42 b = min( b_alg, FLA_Obj_min_dim( UBR ) );
43
44 FLA_Repart_2x2_to_3x3( UTL, /**/ UTR, &U00, /**/ &U01, &U02,
45 /* ************* */ /* ******************** */
46 &U10, /**/ &U11, &U12,
47 UBL, /**/ UBR, &U20, /**/ &U21, &U22,
48 b, b, FLA_BR );
49
50 FLA_Repart_1x2_to_1x3( DL, /**/ DR, &D0, /**/ &D1, &D2,
51 b, FLA_RIGHT );
52
53 FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &W12,
54 b, FLA_RIGHT );
55
56 /*------------------------------------------------------------*/
57
58 // T1T = FLA_Top_part( T1, b );
59
60 FLA_Part_2x1( T1, &T1T,
61 &T2B, b, FLA_TOP );
62
63 // [ U11, ...
64 // D1, T1 ] = FLA_QR2_UT( U11
65 // D1, T1T );
66
67 FLA_QR2_UT_internal( U11,
68 D1, T1T,
69 FLA_Cntl_sub_qr2ut( cntl ) );
70
71
72 if ( FLA_Obj_width( U12 ) > 0 )
73 {
74 // W12T = FLA_Top_part( W12, b );
75
76 FLA_Part_2x1( W12, &W12T,
77 &W12B, b, FLA_TOP );
78
79 // W12T = inv( triu( T1T ) )' * ( U12 + D1' * D2 );
80
81 FLA_Copy_internal( U12, W12T,
82 FLA_Cntl_sub_copy( cntl ) );
83
84 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
85 FLA_ONE, D1, D2, FLA_ONE, W12T,
86 FLA_Cntl_sub_gemm1( cntl ) );
87
88 FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
89 FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
90 FLA_ONE, T1T, W12T,
91 FLA_Cntl_sub_trsm( cntl ) );
92
93 // U12 = U12 - W12T;
94 // D2 = D2 - D1 * W12T;
95
96 FLA_Axpy_internal( FLA_MINUS_ONE, W12T, U12,
97 FLA_Cntl_sub_axpy( cntl ) );
98
99 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
100 FLA_MINUS_ONE, D1, W12T, FLA_ONE, D2,
101 FLA_Cntl_sub_gemm2( cntl ) );
102 }
103
104 /*------------------------------------------------------------*/
105
106 FLA_Cont_with_3x3_to_2x2( &UTL, /**/ &UTR, U00, U01, /**/ U02,
107 U10, U11, /**/ U12,
108 /* ************** */ /* ****************** */
109 &UBL, /**/ &UBR, U20, U21, /**/ U22,
110 FLA_TL );
111
112 FLA_Cont_with_1x3_to_1x2( &DL, /**/ &DR, D0, D1, /**/ D2,
113 FLA_LEFT );
114
115 FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ W12,
116 FLA_LEFT );
117
118 }
119
120 return FLA_SUCCESS;
121}
FLA_Axpy_internal
FLA_Error FLA_Axpy_internal(FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)
Definition FLA_Axpy_internal.c:16
FLA_Copy_internal
FLA_Error FLA_Copy_internal(FLA_Obj A, FLA_Obj B, fla_copy_t *cntl)
Definition FLA_Copy_internal.c:16
FLA_Gemm_internal
FLA_Error FLA_Gemm_internal(FLA_Trans transa, FLA_Trans transb, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
Definition FLA_Gemm_internal.c:16
FLA_QR2_UT_internal
FLA_Error FLA_QR2_UT_internal(FLA_Obj U, FLA_Obj D, FLA_Obj T, fla_qr2ut_t *cntl)
Definition FLA_QR2_UT_internal.c:16
FLA_Trsm_internal
FLA_Error FLA_Trsm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl)
Definition FLA_Trsm_internal.c:16
FLA_MINUS_ONE
FLA_Obj FLA_MINUS_ONE
Definition FLA_Init.c:22
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
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_1x3_to_1x2
FLA_Error FLA_Cont_with_1x3_to_1x2(FLA_Obj *AL, FLA_Obj *AR, FLA_Obj A0, FLA_Obj A1, FLA_Obj A2, FLA_Side side)
Definition FLA_View.c:475
FLA_Obj_width
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
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_length
dim_t FLA_Obj_length(FLA_Obj obj)
Definition FLA_Query.c:116
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_Repart_1x2_to_1x3
FLA_Error FLA_Repart_1x2_to_1x3(FLA_Obj AL, FLA_Obj AR, FLA_Obj *A0, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition FLA_View.c:267
FLA_Obj_min_dim
dim_t FLA_Obj_min_dim(FLA_Obj obj)
Definition FLA_Query.c:153
dim_t
unsigned long dim_t
Definition FLA_type_defs.h:71
i
int i
Definition bl1_axmyv2.c:145
FLA_Obj_view
Definition FLA_type_defs.h:159

References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_internal(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_QR2_UT_internal(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and i.

Referenced by FLA_QR2_UT_internal().

◆ FLA_QR2_UT_blk_var2()

FLA_Error FLA_QR2_UT_blk_var2 ( FLA_Obj  U,
FLA_Obj  D,
FLA_Obj  T,
fla_qr2ut_t *  cntl 
)
15{
16 FLA_Obj DT, D0,
17 DB, D1,
18 D2;
19
20 FLA_Obj TT, T0,
21 TB, T1,
22 T2;
23
24 dim_t b;
25
26 FLA_Part_2x1( D, &DT,
27 &DB, 0, FLA_TOP );
28
29 FLA_Part_2x1( T, &TT,
30 &TB, 0, FLA_TOP );
31
32 while ( FLA_Obj_length( DT ) < FLA_Obj_length( D ) ){
33
34 b = FLA_Determine_blocksize( DB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
35
36 FLA_Repart_2x1_to_3x1( DT, &D0,
37 /* ** */ /* ****** */
38 &D1,
39 DB, &D2, b, FLA_BOTTOM );
40
41 FLA_Repart_2x1_to_3x1( TT, &T0,
42 /* ** */ /* ****** */
43 &T1,
44 TB, &T2, b, FLA_BOTTOM );
45
46 /*------------------------------------------------------------*/
47
48 // [ U, ...
49 // D1, T ] = FLA_QR2_UT( U
50 // D1, T1 );
51
52 FLA_QR2_UT_internal( U,
53 D1, T1,
54 FLA_Cntl_sub_qr2ut( cntl ) );
55
56 /*------------------------------------------------------------*/
57
58 FLA_Cont_with_3x1_to_2x1( &DT, D0,
59 D1,
60 /* ** */ /* ****** */
61 &DB, D2, FLA_TOP );
62
63 FLA_Cont_with_3x1_to_2x1( &TT, T0,
64 T1,
65 /* ** */ /* ****** */
66 &TB, T2, FLA_TOP );
67 }
68
69 return FLA_SUCCESS;
70}
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_Determine_blocksize
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition FLA_Blocksize.c:234

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), FLA_QR2_UT_internal(), FLA_Repart_2x1_to_3x1(), and i.

Referenced by FLA_QR2_UT_internal().

◆ FLA_QR2_UT_opc_var1()

FLA_Error FLA_QR2_UT_opc_var1 ( int  m_UT,
int  m_D,
scomplex *  U,
int  rs_U,
int  cs_U,
scomplex *  D,
int  rs_D,
int  cs_D,
scomplex *  T,
int  rs_T,
int  cs_T 
)
236{
237 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
238 scomplex* buff_0 = FLA_COMPLEX_PTR( FLA_ZERO );
239 int i;
240
241 for ( i = 0; i < m_UT; ++i )
242 {
243 scomplex* upsilon11 = buff_U + (i )*cs_U + (i )*rs_U;
244 scomplex* u12t = buff_U + (i+1)*cs_U + (i )*rs_U;
245
246 scomplex* D0 = buff_D + (0 )*cs_D + (0 )*rs_D;
247 scomplex* d1 = buff_D + (i )*cs_D + (0 )*rs_D;
248 scomplex* D2 = buff_D + (i+1)*cs_D + (0 )*rs_D;
249
250 scomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
251 scomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
252
253 int mn_ahead = m_UT - i - 1;
254 int mn_behind = i;
255
256 //------------------------------------------------------------//
257
258 // FLA_Househ2_UT( FLA_LEFT,
259 // upsilon11,
260 // d1, tau11 );
261 FLA_Househ2_UT_l_opc( m_D,
262 upsilon11,
263 d1, rs_D,
264 tau11 );
265
266 // FLA_Apply_H2_UT( FLA_LEFT, tau11, d1, u12t,
267 // D2 );
268 FLA_Apply_H2_UT_l_opc_var1( m_D,
269 mn_ahead,
270 tau11,
271 d1, rs_D,
272 u12t, cs_U,
273 D2, rs_D, cs_D );
274
275 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, D0, d1, FLA_ZERO, t01 );
276 bl1_cgemv( BLIS1_CONJ_TRANSPOSE,
277 BLIS1_NO_CONJUGATE,
278 m_D,
279 mn_behind,
280 buff_1,
281 D0, rs_D, cs_D,
282 d1, rs_D,
283 buff_0,
284 t01, rs_T );
285
286 //------------------------------------------------------------//
287
288 }
289
290 return FLA_SUCCESS;
291}
FLA_Apply_H2_UT_l_opc_var1
FLA_Error FLA_Apply_H2_UT_l_opc_var1(int m_u2_A2, int n_a1t, scomplex *tau, scomplex *u2, int inc_u2, scomplex *a1t, int inc_a1t, scomplex *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_l_opt_var1.c:269
FLA_ZERO
FLA_Obj FLA_ZERO
Definition FLA_Init.c:20
FLA_Househ2_UT_l_opc
FLA_Error FLA_Househ2_UT_l_opc(int m_x2, scomplex *chi_1, scomplex *x2, int inc_x2, scomplex *tau)
Definition FLA_Househ2_UT.c:390
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
BLIS1_CONJ_TRANSPOSE
@ BLIS1_CONJ_TRANSPOSE
Definition blis_type_defs.h:57
BLIS1_NO_CONJUGATE
@ BLIS1_NO_CONJUGATE
Definition blis_type_defs.h:81
scomplex
Definition blis_type_defs.h:133

References bl1_cgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_opc_var1(), FLA_Househ2_UT_l_opc(), FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_QR2_UT_opt_var1().

◆ FLA_QR2_UT_opd_var1()

FLA_Error FLA_QR2_UT_opd_var1 ( int  m_UT,
int  m_D,
double *  U,
int  rs_U,
int  cs_U,
double *  D,
int  rs_D,
int  cs_D,
double *  T,
int  rs_T,
int  cs_T 
)
172{
173 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
174 double* buff_0 = FLA_DOUBLE_PTR( FLA_ZERO );
175 int i;
176
177 for ( i = 0; i < m_UT; ++i )
178 {
179 double* upsilon11 = buff_U + (i )*cs_U + (i )*rs_U;
180 double* u12t = buff_U + (i+1)*cs_U + (i )*rs_U;
181
182 double* D0 = buff_D + (0 )*cs_D + (0 )*rs_D;
183 double* d1 = buff_D + (i )*cs_D + (0 )*rs_D;
184 double* D2 = buff_D + (i+1)*cs_D + (0 )*rs_D;
185
186 double* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
187 double* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
188
189 int mn_ahead = m_UT - i - 1;
190 int mn_behind = i;
191
192 //------------------------------------------------------------//
193
194 // FLA_Househ2_UT( FLA_LEFT,
195 // upsilon11,
196 // d1, tau11 );
197 FLA_Househ2_UT_l_opd( m_D,
198 upsilon11,
199 d1, rs_D,
200 tau11 );
201
202 // FLA_Apply_H2_UT( FLA_LEFT, tau11, d1, u12t,
203 // D2 );
204 FLA_Apply_H2_UT_l_opd_var1( m_D,
205 mn_ahead,
206 tau11,
207 d1, rs_D,
208 u12t, cs_U,
209 D2, rs_D, cs_D );
210
211 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, D0, d1, FLA_ZERO, t01 );
212 bl1_dgemv( BLIS1_CONJ_TRANSPOSE,
213 BLIS1_NO_CONJUGATE,
214 m_D,
215 mn_behind,
216 buff_1,
217 D0, rs_D, cs_D,
218 d1, rs_D,
219 buff_0,
220 t01, rs_T );
221
222 //------------------------------------------------------------//
223
224 }
225
226 return FLA_SUCCESS;
227}
FLA_Apply_H2_UT_l_opd_var1
FLA_Error FLA_Apply_H2_UT_l_opd_var1(int m_u2_A2, int n_a1t, double *tau, double *u2, int inc_u2, double *a1t, int inc_a1t, double *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_l_opt_var1.c:195
FLA_Househ2_UT_l_opd
FLA_Error FLA_Househ2_UT_l_opd(int m_x2, double *chi_1, double *x2, int inc_x2, double *tau)
Definition FLA_Househ2_UT.c:274
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

References bl1_dgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_opd_var1(), FLA_Househ2_UT_l_opd(), FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_QR2_UT_opt_var1().

◆ FLA_QR2_UT_ops_var1()

FLA_Error FLA_QR2_UT_ops_var1 ( int  m_UT,
int  m_D,
float *  U,
int  rs_U,
int  cs_U,
float *  D,
int  rs_D,
int  cs_D,
float *  T,
int  rs_T,
int  cs_T 
)
108{
109 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
110 float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
111 int i;
112
113 for ( i = 0; i < m_UT; ++i )
114 {
115 float* upsilon11 = buff_U + (i )*cs_U + (i )*rs_U;
116 float* u12t = buff_U + (i+1)*cs_U + (i )*rs_U;
117
118 float* D0 = buff_D + (0 )*cs_D + (0 )*rs_D;
119 float* d1 = buff_D + (i )*cs_D + (0 )*rs_D;
120 float* D2 = buff_D + (i+1)*cs_D + (0 )*rs_D;
121
122 float* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
123 float* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
124
125 int mn_ahead = m_UT - i - 1;
126 int mn_behind = i;
127
128 //------------------------------------------------------------//
129
130 // FLA_Househ2_UT( FLA_LEFT,
131 // upsilon11,
132 // d1, tau11 );
133 FLA_Househ2_UT_l_ops( m_D,
134 upsilon11,
135 d1, rs_D,
136 tau11 );
137
138 // FLA_Apply_H2_UT( FLA_LEFT, tau11, d1, u12t,
139 // D2 );
140 FLA_Apply_H2_UT_l_ops_var1( m_D,
141 mn_ahead,
142 tau11,
143 d1, rs_D,
144 u12t, cs_U,
145 D2, rs_D, cs_D );
146
147 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, D0, d1, FLA_ZERO, t01 );
148 bl1_sgemv( BLIS1_CONJ_TRANSPOSE,
149 BLIS1_NO_CONJUGATE,
150 m_D,
151 mn_behind,
152 buff_1,
153 D0, rs_D, cs_D,
154 d1, rs_D,
155 buff_0,
156 t01, rs_T );
157
158 //------------------------------------------------------------//
159
160 }
161
162 return FLA_SUCCESS;
163}
FLA_Apply_H2_UT_l_ops_var1
FLA_Error FLA_Apply_H2_UT_l_ops_var1(int m_u2_A2, int n_a1t, float *tau, float *u2, int inc_u2, float *a1t, int inc_a1t, float *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_l_opt_var1.c:121
FLA_Househ2_UT_l_ops
FLA_Error FLA_Househ2_UT_l_ops(int m_x2, float *chi_1, float *x2, int inc_x2, float *tau)
Definition FLA_Househ2_UT.c:160
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

References bl1_sgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_ops_var1(), FLA_Househ2_UT_l_ops(), FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_QR2_UT_opt_var1().

◆ FLA_QR2_UT_opt_var1()

FLA_Error FLA_QR2_UT_opt_var1 ( FLA_Obj  U,
FLA_Obj  D,
FLA_Obj  T 
)
15{
16 FLA_Datatype datatype;
17 int m_UT, m_D;
18 int rs_U, cs_U;
19 int rs_D, cs_D;
20 int rs_T, cs_T;
21
22 datatype = FLA_Obj_datatype( U );
23
24 m_UT = FLA_Obj_length( U );
25 m_D = FLA_Obj_length( D );
26
27 rs_U = FLA_Obj_row_stride( U );
28 cs_U = FLA_Obj_col_stride( U );
29 rs_D = FLA_Obj_row_stride( D );
30 cs_D = FLA_Obj_col_stride( D );
31 rs_T = FLA_Obj_row_stride( T );
32 cs_T = FLA_Obj_col_stride( T );
33
34
35 switch ( datatype )
36 {
37 case FLA_FLOAT:
38 {
39 float* buff_U = FLA_FLOAT_PTR( U );
40 float* buff_D = FLA_FLOAT_PTR( D );
41 float* buff_T = FLA_FLOAT_PTR( T );
42
43 FLA_QR2_UT_ops_var1( m_UT,
44 m_D,
45 buff_U, rs_U, cs_U,
46 buff_D, rs_D, cs_D,
47 buff_T, rs_T, cs_T );
48
49 break;
50 }
51
52 case FLA_DOUBLE:
53 {
54 double* buff_U = FLA_DOUBLE_PTR( U );
55 double* buff_D = FLA_DOUBLE_PTR( D );
56 double* buff_T = FLA_DOUBLE_PTR( T );
57
58 FLA_QR2_UT_opd_var1( m_UT,
59 m_D,
60 buff_U, rs_U, cs_U,
61 buff_D, rs_D, cs_D,
62 buff_T, rs_T, cs_T );
63
64 break;
65 }
66
67 case FLA_COMPLEX:
68 {
69 scomplex* buff_U = FLA_COMPLEX_PTR( U );
70 scomplex* buff_D = FLA_COMPLEX_PTR( D );
71 scomplex* buff_T = FLA_COMPLEX_PTR( T );
72
73 FLA_QR2_UT_opc_var1( m_UT,
74 m_D,
75 buff_U, rs_U, cs_U,
76 buff_D, rs_D, cs_D,
77 buff_T, rs_T, cs_T );
78
79 break;
80 }
81
82 case FLA_DOUBLE_COMPLEX:
83 {
84 dcomplex* buff_U = FLA_DOUBLE_COMPLEX_PTR( U );
85 dcomplex* buff_D = FLA_DOUBLE_COMPLEX_PTR( D );
86 dcomplex* buff_T = FLA_DOUBLE_COMPLEX_PTR( T );
87
88 FLA_QR2_UT_opz_var1( m_UT,
89 m_D,
90 buff_U, rs_U, cs_U,
91 buff_D, rs_D, cs_D,
92 buff_T, rs_T, cs_T );
93
94 break;
95 }
96 }
97
98 return FLA_SUCCESS;
99}
FLA_QR2_UT_opz_var1
FLA_Error FLA_QR2_UT_opz_var1(int m_UT, int m_D, dcomplex *buff_U, int rs_U, int cs_U, dcomplex *buff_D, int rs_D, int cs_D, dcomplex *buff_T, int rs_T, int cs_T)
Definition FLA_QR2_UT_opt_var1.c:295
FLA_QR2_UT_opd_var1
FLA_Error FLA_QR2_UT_opd_var1(int m_UT, int m_D, double *buff_U, int rs_U, int cs_U, double *buff_D, int rs_D, int cs_D, double *buff_T, int rs_T, int cs_T)
Definition FLA_QR2_UT_opt_var1.c:167
FLA_QR2_UT_opc_var1
FLA_Error FLA_QR2_UT_opc_var1(int m_UT, int m_D, scomplex *buff_U, int rs_U, int cs_U, scomplex *buff_D, int rs_D, int cs_D, scomplex *buff_T, int rs_T, int cs_T)
Definition FLA_QR2_UT_opt_var1.c:231
FLA_QR2_UT_ops_var1
FLA_Error FLA_QR2_UT_ops_var1(int m_UT, int m_D, float *buff_U, int rs_U, int cs_U, float *buff_D, int rs_D, int cs_D, float *buff_T, int rs_T, int cs_T)
Definition FLA_QR2_UT_opt_var1.c:103
FLA_Obj_row_stride
dim_t FLA_Obj_row_stride(FLA_Obj obj)
Definition FLA_Query.c:167
FLA_Obj_col_stride
dim_t FLA_Obj_col_stride(FLA_Obj obj)
Definition FLA_Query.c:174
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_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_QR2_UT_opc_var1(), FLA_QR2_UT_opd_var1(), FLA_QR2_UT_ops_var1(), FLA_QR2_UT_opz_var1(), and i.

Referenced by FLA_QR2_UT_internal().

◆ FLA_QR2_UT_opz_var1()

FLA_Error FLA_QR2_UT_opz_var1 ( int  m_UT,
int  m_D,
dcomplex *  U,
int  rs_U,
int  cs_U,
dcomplex *  D,
int  rs_D,
int  cs_D,
dcomplex *  T,
int  rs_T,
int  cs_T 
)
300{
301 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
302 dcomplex* buff_0 = FLA_DOUBLE_COMPLEX_PTR( FLA_ZERO );
303 int i;
304
305 for ( i = 0; i < m_UT; ++i )
306 {
307 dcomplex* upsilon11 = buff_U + (i )*cs_U + (i )*rs_U;
308 dcomplex* u12t = buff_U + (i+1)*cs_U + (i )*rs_U;
309
310 dcomplex* D0 = buff_D + (0 )*cs_D + (0 )*rs_D;
311 dcomplex* d1 = buff_D + (i )*cs_D + (0 )*rs_D;
312 dcomplex* D2 = buff_D + (i+1)*cs_D + (0 )*rs_D;
313
314 dcomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
315 dcomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
316
317 int mn_ahead = m_UT - i - 1;
318 int mn_behind = i;
319
320 //------------------------------------------------------------//
321
322 // FLA_Househ2_UT( FLA_LEFT,
323 // upsilon11,
324 // d1, tau11 );
325 FLA_Househ2_UT_l_opz( m_D,
326 upsilon11,
327 d1, rs_D,
328 tau11 );
329
330 // FLA_Apply_H2_UT( FLA_LEFT, tau11, d1, u12t,
331 // D2 );
332 FLA_Apply_H2_UT_l_opz_var1( m_D,
333 mn_ahead,
334 tau11,
335 d1, rs_D,
336 u12t, cs_U,
337 D2, rs_D, cs_D );
338
339 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, D0, d1, FLA_ZERO, t01 );
340 bl1_zgemv( BLIS1_CONJ_TRANSPOSE,
341 BLIS1_NO_CONJUGATE,
342 m_D,
343 mn_behind,
344 buff_1,
345 D0, rs_D, cs_D,
346 d1, rs_D,
347 buff_0,
348 t01, rs_T );
349
350 //------------------------------------------------------------//
351
352 }
353
354 return FLA_SUCCESS;
355}
FLA_Apply_H2_UT_l_opz_var1
FLA_Error FLA_Apply_H2_UT_l_opz_var1(int m_u2_A2, int n_a1t, dcomplex *tau, dcomplex *u2, int inc_u2, dcomplex *a1t, int inc_a1t, dcomplex *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_l_opt_var1.c:343
FLA_Househ2_UT_l_opz
FLA_Error FLA_Househ2_UT_l_opz(int m_x2, dcomplex *chi_1, dcomplex *x2, int inc_x2, dcomplex *tau)
Definition FLA_Househ2_UT.c:521
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

References bl1_zgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_opz_var1(), FLA_Househ2_UT_l_opz(), FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_QR2_UT_opt_var1().

◆ FLA_QR2_UT_unb_var1()

FLA_Error FLA_QR2_UT_unb_var1 ( FLA_Obj  U,
FLA_Obj  D,
FLA_Obj  T 
)
15{
16 FLA_Obj UTL, UTR, U00, u01, U02,
17 UBL, UBR, u10t, upsilon11, u12t,
18 U20, u21, U22;
19
20 FLA_Obj DL, DR, D0, d1, D2;
21
22 FLA_Obj TTL, TTR, T00, t01, T02,
23 TBL, TBR, t10t, tau11, t12t,
24 T20, t21, T22;
25
26
27 FLA_Part_2x2( U, &UTL, &UTR,
28 &UBL, &UBR, 0, 0, FLA_TL );
29
30 FLA_Part_1x2( D, &DL, &DR, 0, FLA_LEFT );
31
32 FLA_Part_2x2( T, &TTL, &TTR,
33 &TBL, &TBR, 0, 0, FLA_TL );
34
35 while ( FLA_Obj_min_dim( UBR ) > 0 ){
36
37 FLA_Repart_2x2_to_3x3( UTL, /**/ UTR, &U00, /**/ &u01, &U02,
38 /* ************* */ /* ************************** */
39 &u10t, /**/ &upsilon11, &u12t,
40 UBL, /**/ UBR, &U20, /**/ &u21, &U22,
41 1, 1, FLA_BR );
42
43 FLA_Repart_1x2_to_1x3( DL, /**/ DR, &D0, /**/ &d1, &D2,
44 1, FLA_RIGHT );
45
46 FLA_Repart_2x2_to_3x3( TTL, /**/ TTR, &T00, /**/ &t01, &T02,
47 /* ************* */ /* ************************ */
48 &t10t, /**/ &tau11, &t12t,
49 TBL, /**/ TBR, &T20, /**/ &t21, &T22,
50 1, 1, FLA_BR );
51
52 /*------------------------------------------------------------*/
53
54 // Compute tau11 and u2 from upsilon11 and d1 such that tau11 and u2
55 // determine a Householder transform H such that applying H from the
56 // left to the column vector consisting of upsilon11 and d1 annihilates
57 // the entries in d1 (and updates upsilon11).
58 FLA_Househ2_UT( FLA_LEFT,
59 upsilon11,
60 d1, tau11 );
61
62 // / u12t \ = H / u12t \
63 // \ D2 / \ D2 /
64 //
65 // where H is formed from tau11 and d1.
66 FLA_Apply_H2_UT( FLA_LEFT, tau11, d1, u12t,
67 D2 );
68
69 // t01 = D0' * d1;
70 FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, D0, d1, FLA_ZERO, t01 );
71
72 /*------------------------------------------------------------*/
73
74 FLA_Cont_with_3x3_to_2x2( &UTL, /**/ &UTR, U00, u01, /**/ U02,
75 u10t, upsilon11, /**/ u12t,
76 /* ************** */ /* ************************ */
77 &UBL, /**/ &UBR, U20, u21, /**/ U22,
78 FLA_TL );
79
80 FLA_Cont_with_1x3_to_1x2( &DL, /**/ &DR, D0, d1, /**/ D2,
81 FLA_LEFT );
82
83 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, t01, /**/ T02,
84 t10t, tau11, /**/ t12t,
85 /* ************** */ /* ********************** */
86 &TBL, /**/ &TBR, T20, t21, /**/ T22,
87 FLA_TL );
88 }
89
90 return FLA_SUCCESS;
91}
FLA_Gemv_external
FLA_Error FLA_Gemv_external(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition FLA_Gemv_external.c:13
FLA_Apply_H2_UT
FLA_Error FLA_Apply_H2_UT(FLA_Side side, FLA_Obj tau, FLA_Obj u2, FLA_Obj a1, FLA_Obj A2)
Definition FLA_Apply_H2_UT.c:13
FLA_Househ2_UT
FLA_Error FLA_Househ2_UT(FLA_Side side, FLA_Obj chi_1, FLA_Obj x2, FLA_Obj tau)
Definition FLA_Househ2_UT.c:16

References FLA_Apply_H2_UT(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Househ2_UT(), FLA_Obj_min_dim(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_ZERO, and i.

Referenced by FLA_QR2_UT_internal().

Generated on Thu Sep 23 2021 17:46:17 for libflame by doxygen 1.9.8