libflame revision_anchor
Functions
FLA_QR_UT_vars.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_QR_UT_unb_var1 (FLA_Obj A, FLA_Obj t)
 
FLA_Error FLA_QR_UT_blk_var1 (FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl)
 
FLA_Error FLA_QR_UT_opt_var1 (FLA_Obj A, FLA_Obj t)
 
FLA_Error FLA_QR_UT_ops_var1 (int m_A, int n_A, float *A, int rs_A, int cs_A, float *t, int inc_t)
 
FLA_Error FLA_QR_UT_opd_var1 (int m_A, int n_A, double *A, int rs_A, int cs_A, double *t, int inc_t)
 
FLA_Error FLA_QR_UT_opc_var1 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A, scomplex *t, int inc_t)
 
FLA_Error FLA_QR_UT_opz_var1 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A, dcomplex *t, int inc_t)
 
FLA_Error FLA_QR_UT_unb_var2 (FLA_Obj A, FLA_Obj T)
 
FLA_Error FLA_QR_UT_blk_var2 (FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl)
 
FLA_Error FLA_QR_UT_opt_var2 (FLA_Obj A, FLA_Obj T)
 
FLA_Error FLA_QR_UT_ops_var2 (int m_A, int n_A, float *A, int rs_A, int cs_A, float *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR_UT_opd_var2 (int m_A, int n_A, double *A, int rs_A, int cs_A, double *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR_UT_opc_var2 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A, scomplex *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR_UT_opz_var2 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A, dcomplex *T, int rs_T, int cs_T)
 
FLA_Error FLA_QR_UT_blk_var3 (FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl)
 

Function Documentation

◆ FLA_QR_UT_blk_var1()

FLA_Error FLA_QR_UT_blk_var1 ( FLA_Obj  A,
FLA_Obj  T,
fla_qrut_t *  cntl 
)
14{
15 FLA_Obj ATL, ATR, A00, A01, A02,
16 ABL, ABR, A10, A11, A12,
17 A20, A21, A22;
18
19 FLA_Obj TL, TR, T0, T1, W12;
20
21 FLA_Obj T1T, T2B;
22
23 FLA_Obj AB1, AB2;
24
25 dim_t b_alg, b;
26
27 // Query the algorithmic blocksize by inspecting the length of T.
28 b_alg = FLA_Obj_length( T );
29
30 FLA_Part_2x2( A, &ATL, &ATR,
31 &ABL, &ABR, 0, 0, FLA_TL );
32
33 FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT );
34
35 while ( FLA_Obj_min_dim( ABR ) > 0 ){
36
37 b = min( b_alg, FLA_Obj_min_dim( ABR ) );
38
39 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
40 /* ************* */ /* ******************** */
41 &A10, /**/ &A11, &A12,
42 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
43 b, b, FLA_BR );
44
45 FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &W12,
46 b, FLA_RIGHT );
47
48 /*------------------------------------------------------------*/
49
50 FLA_Part_2x1( T1, &T1T,
51 &T2B, b, FLA_TOP );
52
53 FLA_Merge_2x1( A11,
54 A21, &AB1 );
55
56 // Perform a QR factorization via the UT transform on AB1:
57 //
58 // / A11 \ -> QB1 R11
59 // \ A21 /
60 //
61 // where:
62 // - QB1 is formed from UB1 (which is stored column-wise below the
63 // diagonal of AB1) and T11 (which is stored to the upper triangle
64 // of T11).
65 // - R11 is stored to the upper triangle of AB1.
66
67 FLA_QR_UT_internal( AB1, T1T,
68 FLA_Cntl_sub_qrut( cntl ) );
69
70
71 if ( FLA_Obj_width( A12 ) > 0 )
72 {
73 FLA_Merge_2x1( A12,
74 A22, &AB2 );
75
76 // Apply the Householder transforms associated with UB1 and T11 to
77 // AB2:
78 //
79 // / A12 \ := QB1' / A12 \
80 // \ A22 / \ A22 /
81 //
82 // where QB1 is formed from UB1 and T11.
83
84 FLA_Apply_Q_UT_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
85 AB1, T1T, W12, AB2,
86 FLA_Cntl_sub_apqut( cntl ) );
87 }
88
89 /*------------------------------------------------------------*/
90
91 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
92 A10, A11, /**/ A12,
93 /* ************** */ /* ****************** */
94 &ABL, /**/ &ABR, A20, A21, /**/ A22,
95 FLA_TL );
96
97 FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ W12,
98 FLA_LEFT );
99 }
100
101 return FLA_SUCCESS;
102}
FLA_Apply_Q_UT_internal
FLA_Error FLA_Apply_Q_UT_internal(FLA_Side side, FLA_Trans trans, FLA_Direct direct, FLA_Store storev, FLA_Obj A, FLA_Obj T, FLA_Obj W, FLA_Obj B, fla_apqut_t *cntl)
Definition FLA_Apply_Q_UT_internal.c:17
FLA_QR_UT_internal
FLA_Error FLA_QR_UT_internal(FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl)
Definition FLA_QR_UT_internal.c:17
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_Merge_2x1
FLA_Error FLA_Merge_2x1(FLA_Obj AT, FLA_Obj AB, FLA_Obj *A)
Definition FLA_View.c:541
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_Apply_Q_UT_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Merge_2x1(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_QR_UT_internal(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLA_QR_UT_internal().

◆ FLA_QR_UT_blk_var2()

FLA_Error FLA_QR_UT_blk_var2 ( FLA_Obj  A,
FLA_Obj  T,
fla_qrut_t *  cntl 
)
14{
15 FLA_Obj ATL, ATR, A00, A01, A02,
16 ABL, ABR, A10, A11, A12,
17 A20, A21, A22;
18
19 FLA_Obj TTL, TTR, T00, T01, T02,
20 TBL, TBR, T10, T11, W12,
21 T20, T21, T22;
22
23 FLA_Obj AB1, AB2;
24
25 dim_t b;
26
27 FLA_Part_2x2( A, &ATL, &ATR,
28 &ABL, &ABR, 0, 0, FLA_TL );
29
30 FLA_Part_2x2( T, &TTL, &TTR,
31 &TBL, &TBR, 0, 0, FLA_TL );
32
33 while ( FLA_Obj_min_dim( ABR ) > 0 ){
34
35 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
36
37 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
38 /* ************* */ /* ******************** */
39 &A10, /**/ &A11, &A12,
40 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
41 b, b, FLA_BR );
42
43 FLA_Repart_2x2_to_3x3( TTL, /**/ TTR, &T00, /**/ &T01, &T02,
44 /* ************* */ /* ******************** */
45 &T10, /**/ &T11, &W12,
46 TBL, /**/ TBR, &T20, /**/ &T21, &T22,
47 b, b, FLA_BR );
48
49 /*------------------------------------------------------------*/
50
51 FLA_Merge_2x1( A11,
52 A21, &AB1 );
53
54 // Perform a QR factorization via the UT transform on AB1:
55 //
56 // / A11 \ -> QB1 R11
57 // \ A21 /
58 //
59 // where:
60 // - QB1 is formed from UB1 (which is stored column-wise below the
61 // diagonal of AB1) and T11 (which is stored to the upper triangle
62 // of T11).
63 // - R11 is stored to the upper triangle of AB1.
64
65 FLA_QR_UT_internal( AB1, T11,
66 FLA_Cntl_sub_qrut( cntl ) );
67
68
69 if ( FLA_Obj_width( A12 ) > 0 )
70 {
71 FLA_Merge_2x1( A12,
72 A22, &AB2 );
73
74 // Apply the Householder transforms associated with UB1 and T11 to
75 // AB2:
76 //
77 // / A12 \ := QB1' / A12 \
78 // \ A22 / \ A22 /
79 //
80 // where QB1 is formed from UB1 and T11.
81
82 FLA_Apply_Q_UT_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
83 AB1, T11, W12, AB2,
84 FLA_Cntl_sub_apqut( cntl ) );
85 }
86
87
88 // Update T
89 //
90 // T01 = A10' * U11 + A20' * U21;
91 //
92 // Recall: U11 = trilu( A11 );
93 // U21 = A21;
94
95 FLA_Copyt_external( FLA_CONJ_TRANSPOSE, A10, T01 );
96 FLA_Trmm_external( FLA_RIGHT, FLA_LOWER_TRIANGULAR,
97 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
98 FLA_ONE, A11, T01 );
99 FLA_Gemm_external( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
100 FLA_ONE, A20, A21, FLA_ONE, T01 );
101
102 /*------------------------------------------------------------*/
103
104 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
105 A10, A11, /**/ A12,
106 /* ************** */ /* ****************** */
107 &ABL, /**/ &ABR, A20, A21, /**/ A22,
108 FLA_TL );
109
110 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, T01, /**/ T02,
111 T10, T11, /**/ W12,
112 /* ************** */ /* ****************** */
113 &TBL, /**/ &TBR, T20, T21, /**/ T22,
114 FLA_TL );
115 }
116
117 return FLA_SUCCESS;
118}
FLA_Copyt_external
FLA_Error FLA_Copyt_external(FLA_Trans trans, FLA_Obj A, FLA_Obj B)
Definition FLA_Copyt_external.c:13
FLA_Gemm_external
FLA_Error FLA_Gemm_external(FLA_Trans transa, FLA_Trans transb, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
Definition FLA_Gemm_external.c:13
FLA_Trmm_external
FLA_Error FLA_Trmm_external(FLA_Side side, FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B)
Definition FLA_Trmm_external.c:13
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
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_Apply_Q_UT_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Copyt_external(), FLA_Determine_blocksize(), FLA_Gemm_external(), FLA_Merge_2x1(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_QR_UT_internal(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_external(), and i.

Referenced by FLA_QR_UT_internal().

◆ FLA_QR_UT_blk_var3()

FLA_Error FLA_QR_UT_blk_var3 ( FLA_Obj  A,
FLA_Obj  T,
fla_qrut_t *  cntl 
)
14{
15 FLA_Obj ATL, ATR, A00, A01, A02,
16 ABL, ABR, A10, A11, A12,
17 A20, A21, A22;
18
19 FLA_Obj TWTL, TWTR, TW00, TW01, TW02,
20 TWBL, TWBR, TW10, T11, W12,
21 TW20, TW21, TW22;
22
23 FLA_Obj AB1, AB2;
24
25 dim_t b;
26
27 FLA_Part_2x2( A, &ATL, &ATR,
28 &ABL, &ABR, 0, 0, FLA_TL );
29
30 FLA_Part_2x2( TW, &TWTL, &TWTR,
31 &TWBL, &TWBR, 0, 0, FLA_TL );
32
33 while ( FLA_Obj_min_dim( ABR ) > 0 ){
34
35 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
36
37 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
38 /* ************* */ /* ******************** */
39 &A10, /**/ &A11, &A12,
40 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
41 b, b, FLA_BR );
42
43 FLA_Repart_2x2_to_3x3( TWTL, /**/ TWTR, &TW00, /**/ &TW01, &TW02,
44 /* ************* */ /* ******************** */
45 &TW10, /**/ &T11, &W12,
46 TWBL, /**/ TWBR, &TW20, /**/ &TW21, &TW22,
47 b, b, FLA_BR );
48
49 /*------------------------------------------------------------*/
50
51 FLA_Merge_2x1( A11,
52 A21, &AB1 );
53
54 // Perform a QR factorization via the UT transform on AB1:
55 //
56 // / A11 \ -> QB1 R11
57 // \ A21 /
58 //
59 // where:
60 // - QB1 is formed from UB1 (which is stored column-wise below the
61 // diagonal of AB1) and T11 (which is stored to the upper triangle
62 // of T11).
63 // - R11 is stored to the upper triangle of AB1.
64
65 FLA_QR_UT_internal( AB1, T11,
66 FLA_Cntl_sub_qrut( cntl ) );
67
68
69 if ( FLA_Obj_width( A12 ) > 0 )
70 {
71 FLA_Merge_2x1( A12,
72 A22, &AB2 );
73
74 // Apply the Householder transforms associated with UB1 and T11 to
75 // AB2:
76 //
77 // / A12 \ := QB1' / A12 \
78 // \ A22 / \ A22 /
79 //
80 // where QB1 is formed from UB1 and T11.
81
82 FLA_Apply_Q_UT_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
83 AB1, T11, W12, AB2,
84 FLA_Cntl_sub_apqut( cntl ) );
85 }
86
87 /*------------------------------------------------------------*/
88
89 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
90 A10, A11, /**/ A12,
91 /* ************** */ /* ****************** */
92 &ABL, /**/ &ABR, A20, A21, /**/ A22,
93 FLA_TL );
94
95 FLA_Cont_with_3x3_to_2x2( &TWTL, /**/ &TWTR, TW00, TW01, /**/ TW02,
96 TW10, T11, /**/ W12,
97 /* ************** */ /* ****************** */
98 &TWBL, /**/ &TWBR, TW20, TW21, /**/ TW22,
99 FLA_TL );
100 }
101
102 return FLA_SUCCESS;
103}

References FLA_Apply_Q_UT_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Merge_2x1(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_Part_2x2(), FLA_QR_UT_internal(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLA_QR_UT_internal().

◆ FLA_QR_UT_opc_var1()

FLA_Error FLA_QR_UT_opc_var1 ( int  m_A,
int  n_A,
scomplex *  A,
int  rs_A,
int  cs_A,
scomplex *  t,
int  inc_t 
)
190{
191 int min_m_n = min( m_A, n_A );
192 int i;
193
194 for ( i = 0; i < min_m_n; ++i )
195 {
196 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
197 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
198 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
199 scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
200
201 scomplex* tau1 = buff_t + (i )*inc_t;
202
203 int m_ahead = m_A - i - 1;
204 int n_ahead = n_A - i - 1;
205
206 /*------------------------------------------------------------*/
207
208 // FLA_Househ2_UT( FLA_LEFT,
209 // alpha11,
210 // a21, tau1 );
211 FLA_Househ2_UT_l_opc( m_ahead,
212 alpha11,
213 a21, rs_A,
214 tau1 );
215
216 // FLA_Apply_H2_UT( FLA_LEFT, tau1, a21, a12t,
217 // A22 );
218 FLA_Apply_H2_UT_l_opc_var1( m_ahead,
219 n_ahead,
220 tau1,
221 a21, rs_A,
222 a12t, cs_A,
223 A22, rs_A, cs_A );
224
225 /*------------------------------------------------------------*/
226
227 }
228
229 return FLA_SUCCESS;
230}
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_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
scomplex
Definition blis_type_defs.h:133

References FLA_Apply_H2_UT_l_opc_var1(), FLA_Househ2_UT_l_opc(), and i.

Referenced by FLA_QR_UT_opt_var1().

◆ FLA_QR_UT_opc_var2()

FLA_Error FLA_QR_UT_opc_var2 ( int  m_A,
int  n_A,
scomplex *  A,
int  rs_A,
int  cs_A,
scomplex *  T,
int  rs_T,
int  cs_T 
)
235{
236 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
237 int min_m_n = min( m_A, n_A );
238 int i;
239
240 for ( i = 0; i < min_m_n; ++i )
241 {
242 scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
243 scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
244 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
245 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
246 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
247 scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
248
249 scomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
250 scomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
251
252 int m_ahead = m_A - i - 1;
253 int n_ahead = n_A - i - 1;
254 int n_behind = i;
255
256 /*------------------------------------------------------------*/
257
258 // FLA_Househ2_UT( FLA_LEFT,
259 // alpha11,
260 // a21, tau11 );
261 FLA_Househ2_UT_l_opc( m_ahead,
262 alpha11,
263 a21, rs_A,
264 tau11 );
265
266 // FLA_Apply_H2_UT( FLA_LEFT, tau11, a21, a12t,
267 // A22 );
268 FLA_Apply_H2_UT_l_opc_var1( m_ahead,
269 n_ahead,
270 tau11,
271 a21, rs_A,
272 a12t, cs_A,
273 A22, rs_A, cs_A );
274
275 // FLA_Copyt_external( FLA_CONJ_TRANSPOSE, a10t, t01 );
276 bl1_ccopyv( BLIS1_CONJUGATE,
277 n_behind,
278 a10t, cs_A,
279 t01, rs_T );
280
281 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, t01 );
282 bl1_cgemv( BLIS1_CONJ_TRANSPOSE,
283 BLIS1_NO_CONJUGATE,
284 m_ahead,
285 n_behind,
286 buff_1,
287 A20, rs_A, cs_A,
288 a21, rs_A,
289 buff_1,
290 t01, rs_T );
291
292 /*------------------------------------------------------------*/
293
294 }
295
296 return FLA_SUCCESS;
297}
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
BLIS1_CONJ_TRANSPOSE
@ BLIS1_CONJ_TRANSPOSE
Definition blis_type_defs.h:57
BLIS1_CONJUGATE
@ BLIS1_CONJUGATE
Definition blis_type_defs.h:82
BLIS1_NO_CONJUGATE
@ BLIS1_NO_CONJUGATE
Definition blis_type_defs.h:81

References bl1_ccopyv(), bl1_cgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_opc_var1(), FLA_Househ2_UT_l_opc(), FLA_ONE, and i.

Referenced by FLA_QR_UT_opt_var2().

◆ FLA_QR_UT_opd_var1()

FLA_Error FLA_QR_UT_opd_var1 ( int  m_A,
int  n_A,
double *  A,
int  rs_A,
int  cs_A,
double *  t,
int  inc_t 
)
142{
143 int min_m_n = min( m_A, n_A );
144 int i;
145
146 for ( i = 0; i < min_m_n; ++i )
147 {
148 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
149 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
150 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
151 double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
152
153 double* tau1 = buff_t + (i )*inc_t;
154
155 int m_ahead = m_A - i - 1;
156 int n_ahead = n_A - i - 1;
157
158 /*------------------------------------------------------------*/
159
160 // FLA_Househ2_UT( FLA_LEFT,
161 // alpha11,
162 // a21, tau1 );
163 FLA_Househ2_UT_l_opd( m_ahead,
164 alpha11,
165 a21, rs_A,
166 tau1 );
167
168 // FLA_Apply_H2_UT( FLA_LEFT, tau1, a21, a12t,
169 // A22 );
170 FLA_Apply_H2_UT_l_opd_var1( m_ahead,
171 n_ahead,
172 tau1,
173 a21, rs_A,
174 a12t, cs_A,
175 A22, rs_A, cs_A );
176
177 /*------------------------------------------------------------*/
178
179 }
180
181 return FLA_SUCCESS;
182}
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

References FLA_Apply_H2_UT_l_opd_var1(), FLA_Househ2_UT_l_opd(), and i.

Referenced by FLA_QR_UT_opt_var1().

◆ FLA_QR_UT_opd_var2()

FLA_Error FLA_QR_UT_opd_var2 ( int  m_A,
int  n_A,
double *  A,
int  rs_A,
int  cs_A,
double *  T,
int  rs_T,
int  cs_T 
)
165{
166 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
167 int min_m_n = min( m_A, n_A );
168 int i;
169
170 for ( i = 0; i < min_m_n; ++i )
171 {
172 double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
173 double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
174 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
175 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
176 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
177 double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
178
179 double* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
180 double* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
181
182 int m_ahead = m_A - i - 1;
183 int n_ahead = n_A - i - 1;
184 int n_behind = i;
185
186 /*------------------------------------------------------------*/
187
188 // FLA_Househ2_UT( FLA_LEFT,
189 // alpha11,
190 // a21, tau11 );
191 FLA_Househ2_UT_l_opd( m_ahead,
192 alpha11,
193 a21, rs_A,
194 tau11 );
195
196 // FLA_Apply_H2_UT( FLA_LEFT, tau11, a21, a12t,
197 // A22 );
198 FLA_Apply_H2_UT_l_opd_var1( m_ahead,
199 n_ahead,
200 tau11,
201 a21, rs_A,
202 a12t, cs_A,
203 A22, rs_A, cs_A );
204
205 // FLA_Copyt_external( FLA_CONJ_TRANSPOSE, a10t, t01 );
206 bl1_dcopyv( BLIS1_CONJUGATE,
207 n_behind,
208 a10t, cs_A,
209 t01, rs_T );
210
211 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, t01 );
212 bl1_dgemv( BLIS1_CONJ_TRANSPOSE,
213 BLIS1_NO_CONJUGATE,
214 m_ahead,
215 n_behind,
216 buff_1,
217 A20, rs_A, cs_A,
218 a21, rs_A,
219 buff_1,
220 t01, rs_T );
221
222 /*------------------------------------------------------------*/
223
224 }
225
226 return FLA_SUCCESS;
227}
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

References bl1_dcopyv(), bl1_dgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_opd_var1(), FLA_Househ2_UT_l_opd(), FLA_ONE, and i.

Referenced by FLA_QR_UT_opt_var2().

◆ FLA_QR_UT_ops_var1()

FLA_Error FLA_QR_UT_ops_var1 ( int  m_A,
int  n_A,
float *  A,
int  rs_A,
int  cs_A,
float *  t,
int  inc_t 
)
94{
95 int min_m_n = min( m_A, n_A );
96 int i;
97
98 for ( i = 0; i < min_m_n; ++i )
99 {
100 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
101 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
102 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
103 float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
104
105 float* tau1 = buff_t + (i )*inc_t;
106
107 int m_ahead = m_A - i - 1;
108 int n_ahead = n_A - i - 1;
109
110 /*------------------------------------------------------------*/
111
112 // FLA_Househ2_UT( FLA_LEFT,
113 // alpha11,
114 // a21, tau1 );
115 FLA_Househ2_UT_l_ops( m_ahead,
116 alpha11,
117 a21, rs_A,
118 tau1 );
119
120 // FLA_Apply_H2_UT( FLA_LEFT, tau1, a21, a12t,
121 // A22 );
122 FLA_Apply_H2_UT_l_ops_var1( m_ahead,
123 n_ahead,
124 tau1,
125 a21, rs_A,
126 a12t, cs_A,
127 A22, rs_A, cs_A );
128
129 /*------------------------------------------------------------*/
130
131 }
132
133 return FLA_SUCCESS;
134}
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

References FLA_Apply_H2_UT_l_ops_var1(), FLA_Househ2_UT_l_ops(), and i.

Referenced by FLA_QR_UT_opt_var1().

◆ FLA_QR_UT_ops_var2()

FLA_Error FLA_QR_UT_ops_var2 ( int  m_A,
int  n_A,
float *  A,
int  rs_A,
int  cs_A,
float *  T,
int  rs_T,
int  cs_T 
)
95{
96 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
97 int min_m_n = min( m_A, n_A );
98 int i;
99
100 for ( i = 0; i < min_m_n; ++i )
101 {
102 float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
103 float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
104 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
105 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
106 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
107 float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
108
109 float* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
110 float* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
111
112 int m_ahead = m_A - i - 1;
113 int n_ahead = n_A - i - 1;
114 int n_behind = i;
115
116 /*------------------------------------------------------------*/
117
118 // FLA_Househ2_UT( FLA_LEFT,
119 // alpha11,
120 // a21, tau11 );
121 FLA_Househ2_UT_l_ops( m_ahead,
122 alpha11,
123 a21, rs_A,
124 tau11 );
125
126 // FLA_Apply_H2_UT( FLA_LEFT, tau11, a21, a12t,
127 // A22 );
128 FLA_Apply_H2_UT_l_ops_var1( m_ahead,
129 n_ahead,
130 tau11,
131 a21, rs_A,
132 a12t, cs_A,
133 A22, rs_A, cs_A );
134
135 // FLA_Copyt_external( FLA_CONJ_TRANSPOSE, a10t, t01 );
136 bl1_scopyv( BLIS1_CONJUGATE,
137 n_behind,
138 a10t, cs_A,
139 t01, rs_T );
140
141 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, t01 );
142 bl1_sgemv( BLIS1_CONJ_TRANSPOSE,
143 BLIS1_NO_CONJUGATE,
144 m_ahead,
145 n_behind,
146 buff_1,
147 A20, rs_A, cs_A,
148 a21, rs_A,
149 buff_1,
150 t01, rs_T );
151
152 /*------------------------------------------------------------*/
153
154 }
155
156 return FLA_SUCCESS;
157}
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

References bl1_scopyv(), bl1_sgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_ops_var1(), FLA_Househ2_UT_l_ops(), FLA_ONE, and i.

Referenced by FLA_QR_UT_opt_var2().

◆ FLA_QR_UT_opt_var1()

FLA_Error FLA_QR_UT_opt_var1 ( FLA_Obj  A,
FLA_Obj  t 
)
14{
15 FLA_Datatype datatype;
16 int m_A, n_A;
17 int rs_A, cs_A;
18 int inc_t;
19
20 datatype = FLA_Obj_datatype( A );
21
22 m_A = FLA_Obj_length( A );
23 n_A = FLA_Obj_width( A );
24 rs_A = FLA_Obj_row_stride( A );
25 cs_A = FLA_Obj_col_stride( A );
26
27 inc_t = FLA_Obj_vector_inc( t );
28
29
30 switch ( datatype )
31 {
32 case FLA_FLOAT:
33 {
34 float* buff_A = FLA_FLOAT_PTR( A );
35 float* buff_t = FLA_FLOAT_PTR( t );
36
37 FLA_QR_UT_ops_var1( m_A,
38 n_A,
39 buff_A, rs_A, cs_A,
40 buff_t, inc_t );
41
42 break;
43 }
44
45 case FLA_DOUBLE:
46 {
47 double* buff_A = FLA_DOUBLE_PTR( A );
48 double* buff_t = FLA_DOUBLE_PTR( t );
49
50 FLA_QR_UT_opd_var1( m_A,
51 n_A,
52 buff_A, rs_A, cs_A,
53 buff_t, inc_t );
54
55 break;
56 }
57
58 case FLA_COMPLEX:
59 {
60 scomplex* buff_A = FLA_COMPLEX_PTR( A );
61 scomplex* buff_t = FLA_COMPLEX_PTR( t );
62
63 FLA_QR_UT_opc_var1( m_A,
64 n_A,
65 buff_A, rs_A, cs_A,
66 buff_t, inc_t );
67
68 break;
69 }
70
71 case FLA_DOUBLE_COMPLEX:
72 {
73 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
74 dcomplex* buff_t = FLA_DOUBLE_COMPLEX_PTR( t );
75
76 FLA_QR_UT_opz_var1( m_A,
77 n_A,
78 buff_A, rs_A, cs_A,
79 buff_t, inc_t );
80
81 break;
82 }
83 }
84
85 return FLA_SUCCESS;
86}
FLA_QR_UT_opz_var1
FLA_Error FLA_QR_UT_opz_var1(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_t, int inc_t)
Definition FLA_QR_UT_opt_var1.c:234
FLA_QR_UT_opd_var1
FLA_Error FLA_QR_UT_opd_var1(int m_A, int n_A, double *buff_A, int rs_A, int cs_A, double *buff_t, int inc_t)
Definition FLA_QR_UT_opt_var1.c:138
FLA_QR_UT_opc_var1
FLA_Error FLA_QR_UT_opc_var1(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_t, int inc_t)
Definition FLA_QR_UT_opt_var1.c:186
FLA_QR_UT_ops_var1
FLA_Error FLA_QR_UT_ops_var1(int m_A, int n_A, float *buff_A, int rs_A, int cs_A, float *buff_t, int inc_t)
Definition FLA_QR_UT_opt_var1.c:90
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_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_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), FLA_Obj_width(), FLA_QR_UT_opc_var1(), FLA_QR_UT_opd_var1(), FLA_QR_UT_ops_var1(), FLA_QR_UT_opz_var1(), and i.

Referenced by FLA_QR_UT_internal().

◆ FLA_QR_UT_opt_var2()

FLA_Error FLA_QR_UT_opt_var2 ( FLA_Obj  A,
FLA_Obj  T 
)
14{
15 FLA_Datatype datatype;
16 int m_A, n_A;
17 int rs_A, cs_A;
18 int rs_T, cs_T;
19
20 datatype = FLA_Obj_datatype( A );
21
22 m_A = FLA_Obj_length( A );
23 n_A = FLA_Obj_width( A );
24 rs_A = FLA_Obj_row_stride( A );
25 cs_A = FLA_Obj_col_stride( A );
26
27 rs_T = FLA_Obj_row_stride( T );
28 cs_T = FLA_Obj_col_stride( T );
29
30
31 switch ( datatype )
32 {
33 case FLA_FLOAT:
34 {
35 float* buff_A = FLA_FLOAT_PTR( A );
36 float* buff_T = FLA_FLOAT_PTR( T );
37
38 FLA_QR_UT_ops_var2( m_A,
39 n_A,
40 buff_A, rs_A, cs_A,
41 buff_T, rs_T, cs_T );
42
43 break;
44 }
45
46 case FLA_DOUBLE:
47 {
48 double* buff_A = FLA_DOUBLE_PTR( A );
49 double* buff_T = FLA_DOUBLE_PTR( T );
50
51 FLA_QR_UT_opd_var2( m_A,
52 n_A,
53 buff_A, rs_A, cs_A,
54 buff_T, rs_T, cs_T );
55
56 break;
57 }
58
59 case FLA_COMPLEX:
60 {
61 scomplex* buff_A = FLA_COMPLEX_PTR( A );
62 scomplex* buff_T = FLA_COMPLEX_PTR( T );
63
64 FLA_QR_UT_opc_var2( m_A,
65 n_A,
66 buff_A, rs_A, cs_A,
67 buff_T, rs_T, cs_T );
68
69 break;
70 }
71
72 case FLA_DOUBLE_COMPLEX:
73 {
74 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
75 dcomplex* buff_T = FLA_DOUBLE_COMPLEX_PTR( T );
76
77 FLA_QR_UT_opz_var2( m_A,
78 n_A,
79 buff_A, rs_A, cs_A,
80 buff_T, rs_T, cs_T );
81
82 break;
83 }
84 }
85
86 return FLA_SUCCESS;
87}
FLA_QR_UT_ops_var2
FLA_Error FLA_QR_UT_ops_var2(int m_A, int n_A, float *buff_A, int rs_A, int cs_A, float *buff_T, int rs_T, int cs_T)
Definition FLA_QR_UT_opt_var2.c:91
FLA_QR_UT_opc_var2
FLA_Error FLA_QR_UT_opc_var2(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_T, int rs_T, int cs_T)
Definition FLA_QR_UT_opt_var2.c:231
FLA_QR_UT_opz_var2
FLA_Error FLA_QR_UT_opz_var2(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_T, int rs_T, int cs_T)
Definition FLA_QR_UT_opt_var2.c:301
FLA_QR_UT_opd_var2
FLA_Error FLA_QR_UT_opd_var2(int m_A, int n_A, double *buff_A, int rs_A, int cs_A, double *buff_T, int rs_T, int cs_T)
Definition FLA_QR_UT_opt_var2.c:161

References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_width(), FLA_QR_UT_opc_var2(), FLA_QR_UT_opd_var2(), FLA_QR_UT_ops_var2(), FLA_QR_UT_opz_var2(), and i.

Referenced by FLA_QR_UT_internal().

◆ FLA_QR_UT_opz_var1()

FLA_Error FLA_QR_UT_opz_var1 ( int  m_A,
int  n_A,
dcomplex *  A,
int  rs_A,
int  cs_A,
dcomplex *  t,
int  inc_t 
)
238{
239 int min_m_n = min( m_A, n_A );
240 int i;
241
242 for ( i = 0; i < min_m_n; ++i )
243 {
244 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
245 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
246 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
247 dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
248
249 dcomplex* tau1 = buff_t + (i )*inc_t;
250
251 int m_ahead = m_A - i - 1;
252 int n_ahead = n_A - i - 1;
253
254 /*------------------------------------------------------------*/
255
256 // FLA_Househ2_UT( FLA_LEFT,
257 // alpha11,
258 // a21, tau1 );
259 FLA_Househ2_UT_l_opz( m_ahead,
260 alpha11,
261 a21, rs_A,
262 tau1 );
263
264 // FLA_Apply_H2_UT( FLA_LEFT, tau1, a21, a12t,
265 // A22 );
266 FLA_Apply_H2_UT_l_opz_var1( m_ahead,
267 n_ahead,
268 tau1,
269 a21, rs_A,
270 a12t, cs_A,
271 A22, rs_A, cs_A );
272
273 /*------------------------------------------------------------*/
274
275 }
276
277 return FLA_SUCCESS;
278}
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

References FLA_Apply_H2_UT_l_opz_var1(), FLA_Househ2_UT_l_opz(), and i.

Referenced by FLA_QR_UT_opt_var1().

◆ FLA_QR_UT_opz_var2()

FLA_Error FLA_QR_UT_opz_var2 ( int  m_A,
int  n_A,
dcomplex *  A,
int  rs_A,
int  cs_A,
dcomplex *  T,
int  rs_T,
int  cs_T 
)
305{
306 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
307 int min_m_n = min( m_A, n_A );
308 int i;
309
310 for ( i = 0; i < min_m_n; ++i )
311 {
312 dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
313 dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
314 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
315 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
316 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
317 dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
318
319 dcomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
320 dcomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
321
322 int m_ahead = m_A - i - 1;
323 int n_ahead = n_A - i - 1;
324 int n_behind = i;
325
326 /*------------------------------------------------------------*/
327
328 // FLA_Househ2_UT( FLA_LEFT,
329 // alpha11,
330 // a21, tau11 );
331 FLA_Househ2_UT_l_opz( m_ahead,
332 alpha11,
333 a21, rs_A,
334 tau11 );
335
336 // FLA_Apply_H2_UT( FLA_LEFT, tau11, a21, a12t,
337 // A22 );
338 FLA_Apply_H2_UT_l_opz_var1( m_ahead,
339 n_ahead,
340 tau11,
341 a21, rs_A,
342 a12t, cs_A,
343 A22, rs_A, cs_A );
344
345 // FLA_Copyt_external( FLA_CONJ_TRANSPOSE, a10t, t01 );
346 bl1_zcopyv( BLIS1_CONJUGATE,
347 n_behind,
348 a10t, cs_A,
349 t01, rs_T );
350
351 // FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, t01 );
352 bl1_zgemv( BLIS1_CONJ_TRANSPOSE,
353 BLIS1_NO_CONJUGATE,
354 m_ahead,
355 n_behind,
356 buff_1,
357 A20, rs_A, cs_A,
358 a21, rs_A,
359 buff_1,
360 t01, rs_T );
361
362 /*------------------------------------------------------------*/
363
364 }
365
366 return FLA_SUCCESS;
367}
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

References bl1_zcopyv(), bl1_zgemv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_l_opz_var1(), FLA_Househ2_UT_l_opz(), FLA_ONE, and i.

Referenced by FLA_QR_UT_opt_var2().

◆ FLA_QR_UT_unb_var1()

FLA_Error FLA_QR_UT_unb_var1 ( FLA_Obj  A,
FLA_Obj  t 
)
14{
15 FLA_Obj ATL, ATR, A00, a01, A02,
16 ABL, ABR, a10t, alpha11, a12t,
17 A20, a21, A22;
18
19 FLA_Obj tLt, tRt, t0t, tau1, t2t;
20
21
22 FLA_Part_2x2( A, &ATL, &ATR,
23 &ABL, &ABR, 0, 0, FLA_TL );
24
25 FLA_Part_1x2( t, &tLt, &tRt, 0, FLA_LEFT );
26
27 while ( FLA_Obj_min_dim( ABR ) > 0 ){
28
29 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
30 /* ************* */ /* ************************** */
31 &a10t, /**/ &alpha11, &a12t,
32 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
33 1, 1, FLA_BR );
34
35 FLA_Repart_1x2_to_1x3( tLt, /**/ tRt, &t0t, /**/ &tau1, &t2t,
36 1, FLA_RIGHT );
37
38 /*------------------------------------------------------------*/
39
40 // Compute tau11 and u21 from alpha11 and a21 such that tau11 and u21
41 // determine a Householder transform H such that applying H from the
42 // left to the column vector consisting of alpha11 and a21 annihilates
43 // the entries in a21 (and updates alpha11).
44 FLA_Househ2_UT( FLA_LEFT,
45 alpha11,
46 a21, tau1 );
47
48 // / a12t \ = H / a12t \
49 // \ A22 / \ A22 /
50 //
51 // where H is formed from tau11 and u21.
52 FLA_Apply_H2_UT( FLA_LEFT, tau1, a21, a12t,
53 A22 );
54
55 /*------------------------------------------------------------*/
56
57 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
58 a10t, alpha11, /**/ a12t,
59 /* ************** */ /* ************************ */
60 &ABL, /**/ &ABR, A20, a21, /**/ A22,
61 FLA_TL );
62
63 FLA_Cont_with_1x3_to_1x2( &tLt, /**/ &tRt, t0t, tau1, /**/ t2t,
64 FLA_LEFT );
65 }
66
67 return FLA_SUCCESS;
68}
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_Househ2_UT(), FLA_Obj_min_dim(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLA_QR_UT_internal().

◆ FLA_QR_UT_unb_var2()

FLA_Error FLA_QR_UT_unb_var2 ( FLA_Obj  A,
FLA_Obj  T 
)
14{
15 FLA_Obj ATL, ATR, A00, a01, A02,
16 ABL, ABR, a10t, alpha11, a12t,
17 A20, a21, A22;
18
19 FLA_Obj TTL, TTR, T00, t01, T02,
20 TBL, TBR, t10t, tau11, t12t,
21 T20, t21, T22;
22
23
24 FLA_Part_2x2( A, &ATL, &ATR,
25 &ABL, &ABR, 0, 0, FLA_TL );
26
27 FLA_Part_2x2( T, &TTL, &TTR,
28 &TBL, &TBR, 0, 0, FLA_TL );
29
30 while ( FLA_Obj_min_dim( ABR ) > 0 ){
31
32 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
33 /* ************* */ /* ************************** */
34 &a10t, /**/ &alpha11, &a12t,
35 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
36 1, 1, FLA_BR );
37
38 FLA_Repart_2x2_to_3x3( TTL, /**/ TTR, &T00, /**/ &t01, &T02,
39 /* ************* */ /* ************************ */
40 &t10t, /**/ &tau11, &t12t,
41 TBL, /**/ TBR, &T20, /**/ &t21, &T22,
42 1, 1, FLA_BR );
43
44 /*------------------------------------------------------------*/
45
46 // Compute tau11 and u21 from alpha11 and a21 such that tau11 and u21
47 // determine a Householder transform H such that applying H from the
48 // left to the column vector consisting of alpha11 and a21 annihilates
49 // the entries in a21 (and updates alpha11).
50 FLA_Househ2_UT( FLA_LEFT,
51 alpha11,
52 a21, tau11 );
53
54 // / a12t \ = H / a12t \
55 // \ A22 / \ A22 /
56 //
57 // where H is formed from tau11 and u21.
58 FLA_Apply_H2_UT( FLA_LEFT, tau11, a21, a12t,
59 A22 );
60
61 // t01 = a10t' + A20' * u21;
62 FLA_Copyt_external( FLA_CONJ_TRANSPOSE, a10t, t01 );
63 FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, t01 );
64
65 /*------------------------------------------------------------*/
66
67 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
68 a10t, alpha11, /**/ a12t,
69 /* ************** */ /* ************************ */
70 &ABL, /**/ &ABR, A20, a21, /**/ A22,
71 FLA_TL );
72
73 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, t01, /**/ T02,
74 t10t, tau11, /**/ t12t,
75 /* ************** */ /* ********************** */
76 &TBL, /**/ &TBR, T20, t21, /**/ T22,
77 FLA_TL );
78
79 }
80
81 return FLA_SUCCESS;
82}
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

References FLA_Apply_H2_UT(), FLA_Cont_with_3x3_to_2x2(), FLA_Copyt_external(), FLA_Gemv_external(), FLA_Househ2_UT(), FLA_Obj_min_dim(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLA_QR_UT_internal().

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