libflame revision_anchor
Functions
FLA_LQ_UT_vars.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_LQ_UT_unb_var1 (FLA_Obj A, FLA_Obj t)
 
FLA_Error FLA_LQ_UT_blk_var1 (FLA_Obj A, FLA_Obj T, fla_lqut_t *cntl)
 
FLA_Error FLA_LQ_UT_opt_var1 (FLA_Obj A, FLA_Obj t)
 
FLA_Error FLA_LQ_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_LQ_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_LQ_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_LQ_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_LQ_UT_unb_var2 (FLA_Obj A, FLA_Obj T)
 
FLA_Error FLA_LQ_UT_blk_var2 (FLA_Obj A, FLA_Obj T, fla_lqut_t *cntl)
 
FLA_Error FLA_LQ_UT_opt_var2 (FLA_Obj A, FLA_Obj T)
 
FLA_Error FLA_LQ_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_LQ_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_LQ_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_LQ_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_LQ_UT_blk_var3 (FLA_Obj A, FLA_Obj T, fla_lqut_t *cntl)
 

Function Documentation

◆ FLA_LQ_UT_blk_var1()

FLA_Error FLA_LQ_UT_blk_var1 ( FLA_Obj  A,
FLA_Obj  T,
fla_lqut_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 AR1, AR2;
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_1x2( A11, A12, &AR1 );
54
55 // Perform an LQ factorization via the UT transform on AR1:
56 //
57 // ( A11 A12 ) -> L11 QR1
58 //
59 // where:
60 // - QR1 is formed from UR1 (which is stored row-wise above the
61 // diagonal of AR1) and T11 (which is stored to the upper triangle
62 // of T11).
63 // - L11 is stored to the lower triangle of AR1.
64
65 FLA_LQ_UT_internal( AR1, T1T,
66 FLA_Cntl_sub_lqut( cntl ));
67
68
69 if ( FLA_Obj_length( A21 ) > 0 )
70 {
71 FLA_Merge_1x2( A21, A22, &AR2 );
72
73 // Apply the Householder transforms associated with UR1 and T11 to
74 // AR2:
75 //
76 // ( A21 A22 ) := ( A21 A22 ) Q1
77 //
78 // where QR1 is formed from UR1 and T11.
79
80 FLA_Apply_Q_UT_internal( FLA_RIGHT, FLA_NO_TRANSPOSE, FLA_FORWARD, FLA_ROWWISE,
81 AR1, T1T, W12, AR2,
82 FLA_Cntl_sub_apqut( cntl ) );
83 }
84
85 /*------------------------------------------------------------*/
86
87 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
88 A10, A11, /**/ A12,
89 /* ************** */ /* ****************** */
90 &ABL, /**/ &ABR, A20, A21, /**/ A22,
91 FLA_TL );
92
93 FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ W12,
94 FLA_LEFT );
95 }
96
97 return FLA_SUCCESS;
98}
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_LQ_UT_internal
FLA_Error FLA_LQ_UT_internal(FLA_Obj A, FLA_Obj T, fla_lqut_t *cntl)
Definition FLA_LQ_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_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_Merge_1x2
FLA_Error FLA_Merge_1x2(FLA_Obj AL, FLA_Obj AR, FLA_Obj *A)
Definition FLA_View.c:562
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_Apply_Q_UT_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_LQ_UT_internal(), FLA_Merge_1x2(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLA_LQ_UT_internal().

◆ FLA_LQ_UT_blk_var2()

FLA_Error FLA_LQ_UT_blk_var2 ( FLA_Obj  A,
FLA_Obj  T,
fla_lqut_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 AR1,
24 AR2;
25
26 dim_t b;
27
28 FLA_Part_2x2( A, &ATL, &ATR,
29 &ABL, &ABR, 0, 0, FLA_TL );
30
31 FLA_Part_2x2( T, &TTL, &TTR,
32 &TBL, &TBR, 0, 0, FLA_TL );
33
34 while ( FLA_Obj_min_dim( ABR ) > 0 ){
35
36 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
37
38 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
39 /* ************* */ /* ******************** */
40 &A10, /**/ &A11, &A12,
41 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
42 b, b, FLA_BR );
43
44 FLA_Repart_2x2_to_3x3( TTL, /**/ TTR, &T00, /**/ &T01, &T02,
45 /* ************* */ /* ******************** */
46 &T10, /**/ &T11, &W12,
47 TBL, /**/ TBR, &T20, /**/ &T21, &T22,
48 b, b, FLA_BR );
49
50 /*------------------------------------------------------------*/
51
52 FLA_Merge_1x2( A11, A12, &AR1 );
53
54 // Perform an LQ factorization via the UT transform on AR1:
55 //
56 // ( A11 A12 ) -> L11 QR1
57 //
58 // where:
59 // - QR1 is formed from UR1 (which is stored row-wise above the
60 // diagonal of AR1) and T11 (which is stored to the upper triangle
61 // of T11).
62 // - L11 is stored to the lower triangle of AR1.
63
64 FLA_LQ_UT_internal( AR1, T11,
65 FLA_Cntl_sub_lqut( cntl ) );
66
67
68 if ( FLA_Obj_length( A21 ) > 0 )
69 {
70 FLA_Merge_1x2( A21, A22, &AR2 );
71
72 // Apply the Householder transforms associated with UR1 and T11 to
73 // AR2:
74 //
75 // ( A21 A22 ) := ( A21 A22 ) Q1
76 //
77 // where QR1 is formed from UR1 and T11.
78
79 FLA_Apply_Q_UT_internal( FLA_RIGHT, FLA_NO_TRANSPOSE, FLA_FORWARD, FLA_ROWWISE,
80 AR1, T11, W12, AR2,
81 FLA_Cntl_sub_apqut( cntl ) );
82 }
83
84
85 // Update T
86 //
87 // T01 = conj(A01) * U11^T + conj(A02) * U12^T;
88 //
89 // Recall: U11 = triuu( A11 );
90 // U12 = A12;
91
92 FLA_Copyt_external( FLA_CONJ_NO_TRANSPOSE, A01, T01 );
93 FLA_Trmm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
94 FLA_TRANSPOSE, FLA_UNIT_DIAG,
95 FLA_ONE, A11, T01 );
96 FLA_Gemm_external( FLA_CONJ_NO_TRANSPOSE, FLA_TRANSPOSE,
97 FLA_ONE, A02, A12, FLA_ONE, T01 );
98
99 /*------------------------------------------------------------*/
100
101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
102 A10, A11, /**/ A12,
103 /* ************** */ /* ****************** */
104 &ABL, /**/ &ABR, A20, A21, /**/ A22,
105 FLA_TL );
106
107 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, T01, /**/ T02,
108 T10, T11, /**/ W12,
109 /* ************** */ /* ****************** */
110 &TBL, /**/ &TBR, T20, T21, /**/ T22,
111 FLA_TL );
112 }
113
114 return FLA_SUCCESS;
115}
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_LQ_UT_internal(), FLA_Merge_1x2(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_external(), and i.

Referenced by FLA_LQ_UT_internal().

◆ FLA_LQ_UT_blk_var3()

FLA_Error FLA_LQ_UT_blk_var3 ( FLA_Obj  A,
FLA_Obj  T,
fla_lqut_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 AR1,
24 AR2;
25
26 dim_t b;
27
28 FLA_Part_2x2( A, &ATL, &ATR,
29 &ABL, &ABR, 0, 0, FLA_TL );
30
31 FLA_Part_2x2( TW, &TWTL, &TWTR,
32 &TWBL, &TWBR, 0, 0, FLA_TL );
33
34 while ( FLA_Obj_min_dim( ABR ) > 0 ){
35
36 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
37
38 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
39 /* ************* */ /* ******************** */
40 &A10, /**/ &A11, &A12,
41 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
42 b, b, FLA_BR );
43
44 FLA_Repart_2x2_to_3x3( TWTL, /**/ TWTR, &TW00, /**/ &TW01, &TW02,
45 /* ************* */ /* ******************** */
46 &TW10, /**/ &T11, &W12,
47 TWBL, /**/ TWBR, &TW20, /**/ &TW21, &TW22,
48 b, b, FLA_BR );
49
50 /*------------------------------------------------------------*/
51
52 FLA_Merge_1x2( A11, A12, &AR1 );
53
54 // Perform an LQ factorization via the UT transform on AR1:
55 //
56 // ( A11 A12 ) -> L11 QR1
57 //
58 // where:
59 // - QR1 is formed from UR1 (which is stored row-wise above the
60 // diagonal of AR1) and T11 (which is stored to the upper triangle
61 // of T11).
62 // - L11 is stored to the lower triangle of AR1.
63
64 FLA_LQ_UT_internal( AR1, T11,
65 FLA_Cntl_sub_lqut( cntl ) );
66
67
68 if ( FLA_Obj_length( A21 ) > 0 )
69 {
70 FLA_Merge_1x2( A21, A22, &AR2 );
71
72 // Apply the Householder transforms associated with UR1 and T11 to
73 // AR2:
74 //
75 // ( A21 A22 ) := ( A21 A22 ) Q1
76 //
77 // where QR1 is formed from UR1 and T11.
78
79 FLA_Apply_Q_UT_internal( FLA_RIGHT, FLA_NO_TRANSPOSE, FLA_FORWARD, FLA_ROWWISE,
80 AR1, T11, W12, AR2,
81 FLA_Cntl_sub_apqut( cntl ) );
82 }
83
84 /*------------------------------------------------------------*/
85
86 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
87 A10, A11, /**/ A12,
88 /* ************** */ /* ****************** */
89 &ABL, /**/ &ABR, A20, A21, /**/ A22,
90 FLA_TL );
91
92 FLA_Cont_with_3x3_to_2x2( &TWTL, /**/ &TWTR, TW00, TW01, /**/ TW02,
93 TW10, T11, /**/ W12,
94 /* ************** */ /* ****************** */
95 &TWBL, /**/ &TWBR, TW20, TW21, /**/ TW22,
96 FLA_TL );
97 }
98
99 return FLA_SUCCESS;
100}

References FLA_Apply_Q_UT_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_LQ_UT_internal(), FLA_Merge_1x2(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLA_LQ_UT_internal().

◆ FLA_LQ_UT_opc_var1()

FLA_Error FLA_LQ_UT_opc_var1 ( int  m_A,
int  n_A,
scomplex *  A,
int  rs_A,
int  cs_A,
scomplex *  t,
int  inc_t 
)
186{
187 int min_m_n = min( m_A, n_A );
188 int i;
189
190 for ( i = 0; i < min_m_n; ++i )
191 {
192 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
193 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
194 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
195 scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
196
197 scomplex* tau1 = buff_t + (i )*inc_t;
198
199 int m_ahead = m_A - i - 1;
200 int n_ahead = n_A - i - 1;
201
202 /*------------------------------------------------------------*/
203
204 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
205 // tau1 );
206 FLA_Househ2_UT_r_opc( n_ahead,
207 alpha11,
208 a12t, cs_A,
209 tau1 );
210
211 // FLA_Apply_H2_UT( FLA_RIGHT, tau1, a12t, a21, A22 );
212 FLA_Apply_H2_UT_r_opc_var1( m_ahead,
213 n_ahead,
214 tau1,
215 a12t, cs_A,
216 a21, rs_A,
217 A22, rs_A, cs_A );
218
219 /*------------------------------------------------------------*/
220
221 }
222
223 return FLA_SUCCESS;
224}
FLA_Apply_H2_UT_r_opc_var1
FLA_Error FLA_Apply_H2_UT_r_opc_var1(int n_u2h_A2, int m_a1, scomplex *tau, scomplex *u2h, int inc_u2h, scomplex *a1, int inc_a1, scomplex *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_r_opt_var1.c:254
FLA_Househ2_UT_r_opc
FLA_Error FLA_Househ2_UT_r_opc(int m_x2, scomplex *chi_1, scomplex *x2, int inc_x2, scomplex *tau)
Definition FLA_Househ2_UT.c:677
scomplex
Definition blis_type_defs.h:133

References FLA_Apply_H2_UT_r_opc_var1(), FLA_Househ2_UT_r_opc(), and i.

Referenced by FLA_LQ_UT_opt_var1().

◆ FLA_LQ_UT_opc_var2()

FLA_Error FLA_LQ_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 
)
233{
234 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
235 int min_m_n = min( m_A, n_A );
236 int i;
237
238 for ( i = 0; i < min_m_n; ++i )
239 {
240 scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
241 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
242 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
243 scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
244 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
245 scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
246
247 scomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
248 scomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
249
250 int m_ahead = m_A - i - 1;
251 int n_ahead = n_A - i - 1;
252 int m_behind = i;
253
254 /*------------------------------------------------------------*/
255
256 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
257 // tau11 );
258 FLA_Househ2_UT_r_opc( n_ahead,
259 alpha11,
260 a12t, cs_A,
261 tau11 );
262
263 // FLA_Apply_H2_UT( FLA_RIGHT, tau11, a12t, a21, A22 );
264 FLA_Apply_H2_UT_r_opc_var1( m_ahead,
265 n_ahead,
266 tau11,
267 a12t, cs_A,
268 a21, rs_A,
269 A22, rs_A, cs_A );
270
271 // FLA_Copyt_external( FLA_CONJ_NO_TRANSPOSE, a01, t01 );
272 bl1_ccopyv( BLIS1_CONJUGATE,
273 m_behind,
274 a01, rs_A,
275 t01, rs_T );
276
277 // FLA_Gemvc_external( FLA_CONJ_NO_TRANSPOSE, FLA_NO_CONJUGATE,
278 // FLA_ONE, A02, a12t, FLA_ONE, t01 );
279 bl1_cgemv( BLIS1_CONJ_NO_TRANSPOSE,
280 BLIS1_NO_CONJUGATE,
281 m_behind,
282 n_ahead,
283 buff_1,
284 A02, rs_A, cs_A,
285 a12t, cs_A,
286 buff_1,
287 t01, rs_T );
288
289 /*------------------------------------------------------------*/
290
291 }
292
293 return FLA_SUCCESS;
294}
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_NO_TRANSPOSE
@ BLIS1_CONJ_NO_TRANSPOSE
Definition blis_type_defs.h:56
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_NO_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_r_opc_var1(), FLA_Househ2_UT_r_opc(), FLA_ONE, and i.

Referenced by FLA_LQ_UT_opt_var2().

◆ FLA_LQ_UT_opd_var1()

FLA_Error FLA_LQ_UT_opd_var1 ( int  m_A,
int  n_A,
double *  A,
int  rs_A,
int  cs_A,
double *  t,
int  inc_t 
)
140{
141 int min_m_n = min( m_A, n_A );
142 int i;
143
144 for ( i = 0; i < min_m_n; ++i )
145 {
146 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
147 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
148 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
149 double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
150
151 double* tau1 = buff_t + (i )*inc_t;
152
153 int m_ahead = m_A - i - 1;
154 int n_ahead = n_A - i - 1;
155
156 /*------------------------------------------------------------*/
157
158 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
159 // tau1 );
160 FLA_Househ2_UT_r_opd( n_ahead,
161 alpha11,
162 a12t, cs_A,
163 tau1 );
164
165 // FLA_Apply_H2_UT( FLA_RIGHT, tau1, a12t, a21, A22 );
166 FLA_Apply_H2_UT_r_opd_var1( m_ahead,
167 n_ahead,
168 tau1,
169 a12t, cs_A,
170 a21, rs_A,
171 A22, rs_A, cs_A );
172
173 /*------------------------------------------------------------*/
174
175 }
176
177 return FLA_SUCCESS;
178}
FLA_Apply_H2_UT_r_opd_var1
FLA_Error FLA_Apply_H2_UT_r_opd_var1(int n_u2h_A2, int m_a1, double *tau, double *u2h, int inc_u2h, double *a1, int inc_a1, double *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_r_opt_var1.c:181
FLA_Househ2_UT_r_opd
FLA_Error FLA_Househ2_UT_r_opd(int m_x2, double *chi_1, double *x2, int inc_x2, double *tau)
Definition FLA_Househ2_UT.c:664

References FLA_Apply_H2_UT_r_opd_var1(), FLA_Househ2_UT_r_opd(), and i.

Referenced by FLA_LQ_UT_opt_var1().

◆ FLA_LQ_UT_opd_var2()

FLA_Error FLA_LQ_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 
)
164{
165 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
166 int min_m_n = min( m_A, n_A );
167 int i;
168
169 for ( i = 0; i < min_m_n; ++i )
170 {
171 double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
172 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
173 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
174 double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
175 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
176 double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
177
178 double* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
179 double* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
180
181 int m_ahead = m_A - i - 1;
182 int n_ahead = n_A - i - 1;
183 int m_behind = i;
184
185 /*------------------------------------------------------------*/
186
187 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
188 // tau11 );
189 FLA_Househ2_UT_r_opd( n_ahead,
190 alpha11,
191 a12t, cs_A,
192 tau11 );
193
194 // FLA_Apply_H2_UT( FLA_RIGHT, tau11, a12t, a21, A22 );
195 FLA_Apply_H2_UT_r_opd_var1( m_ahead,
196 n_ahead,
197 tau11,
198 a12t, cs_A,
199 a21, rs_A,
200 A22, rs_A, cs_A );
201
202 // FLA_Copyt_external( FLA_CONJ_NO_TRANSPOSE, a01, t01 );
203 bl1_dcopyv( BLIS1_CONJUGATE,
204 m_behind,
205 a01, rs_A,
206 t01, rs_T );
207
208 // FLA_Gemvc_external( FLA_CONJ_NO_TRANSPOSE, FLA_NO_CONJUGATE,
209 // FLA_ONE, A02, a12t, FLA_ONE, t01 );
210 bl1_dgemv( BLIS1_CONJ_NO_TRANSPOSE,
211 BLIS1_NO_CONJUGATE,
212 m_behind,
213 n_ahead,
214 buff_1,
215 A02, rs_A, cs_A,
216 a12t, cs_A,
217 buff_1,
218 t01, rs_T );
219
220 /*------------------------------------------------------------*/
221
222 }
223
224 return FLA_SUCCESS;
225}
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_NO_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_r_opd_var1(), FLA_Househ2_UT_r_opd(), FLA_ONE, and i.

Referenced by FLA_LQ_UT_opt_var2().

◆ FLA_LQ_UT_ops_var1()

FLA_Error FLA_LQ_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_RIGHT, alpha11, a12t
113 // tau1 );
114 FLA_Househ2_UT_r_ops( n_ahead,
115 alpha11,
116 a12t, cs_A,
117 tau1 );
118
119 // FLA_Apply_H2_UT( FLA_RIGHT, tau1, a12t, a21, A22 );
120 FLA_Apply_H2_UT_r_ops_var1( m_ahead,
121 n_ahead,
122 tau1,
123 a12t, cs_A,
124 a21, rs_A,
125 A22, rs_A, cs_A );
126
127 /*------------------------------------------------------------*/
128
129 }
130
131 return FLA_SUCCESS;
132}
FLA_Apply_H2_UT_r_ops_var1
FLA_Error FLA_Apply_H2_UT_r_ops_var1(int n_u2h_A2, int m_a1, float *tau, float *u2h, int inc_u2h, float *a1, int inc_a1, float *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_r_opt_var1.c:108
FLA_Househ2_UT_r_ops
FLA_Error FLA_Househ2_UT_r_ops(int m_x2, float *chi_1, float *x2, int inc_x2, float *tau)
Definition FLA_Househ2_UT.c:651

References FLA_Apply_H2_UT_r_ops_var1(), FLA_Househ2_UT_r_ops(), and i.

Referenced by FLA_LQ_UT_opt_var1().

◆ FLA_LQ_UT_ops_var2()

FLA_Error FLA_LQ_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* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
103 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
104 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
105 float* A02 = buff_A + (i+1)*cs_A + (0 )*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 m_behind = i;
115
116 /*------------------------------------------------------------*/
117
118 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
119 // tau11 );
120 FLA_Househ2_UT_r_ops( n_ahead,
121 alpha11,
122 a12t, cs_A,
123 tau11 );
124
125 // FLA_Apply_H2_UT( FLA_RIGHT, tau11, a12t, a21, A22 );
126 FLA_Apply_H2_UT_r_ops_var1( m_ahead,
127 n_ahead,
128 tau11,
129 a12t, cs_A,
130 a21, rs_A,
131 A22, rs_A, cs_A );
132
133 // FLA_Copyt_external( FLA_CONJ_NO_TRANSPOSE, a01, t01 );
134 bl1_scopyv( BLIS1_CONJUGATE,
135 m_behind,
136 a01, rs_A,
137 t01, rs_T );
138
139 // FLA_Gemvc_external( FLA_CONJ_NO_TRANSPOSE, FLA_NO_CONJUGATE,
140 // FLA_ONE, A02, a12t, FLA_ONE, t01 );
141 bl1_sgemv( BLIS1_CONJ_NO_TRANSPOSE,
142 BLIS1_NO_CONJUGATE,
143 m_behind,
144 n_ahead,
145 buff_1,
146 A02, rs_A, cs_A,
147 a12t, cs_A,
148 buff_1,
149 t01, rs_T );
150
151 /*------------------------------------------------------------*/
152
153 }
154
155 return FLA_SUCCESS;
156}
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_NO_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_r_ops_var1(), FLA_Househ2_UT_r_ops(), FLA_ONE, and i.

Referenced by FLA_LQ_UT_opt_var2().

◆ FLA_LQ_UT_opt_var1()

FLA_Error FLA_LQ_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_LQ_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_LQ_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_LQ_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_LQ_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_LQ_UT_ops_var1
FLA_Error FLA_LQ_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_LQ_UT_opt_var1.c:90
FLA_LQ_UT_opz_var1
FLA_Error FLA_LQ_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_LQ_UT_opt_var1.c:228
FLA_LQ_UT_opd_var1
FLA_Error FLA_LQ_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_LQ_UT_opt_var1.c:136
FLA_LQ_UT_opc_var1
FLA_Error FLA_LQ_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_LQ_UT_opt_var1.c:182
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_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_LQ_UT_opc_var1(), FLA_LQ_UT_opd_var1(), FLA_LQ_UT_ops_var1(), FLA_LQ_UT_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), FLA_Obj_width(), and i.

Referenced by FLA_LQ_UT_internal().

◆ FLA_LQ_UT_opt_var2()

FLA_Error FLA_LQ_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_LQ_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_LQ_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_LQ_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_LQ_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_LQ_UT_ops_var2
FLA_Error FLA_LQ_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_LQ_UT_opt_var2.c:91
FLA_LQ_UT_opc_var2
FLA_Error FLA_LQ_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_LQ_UT_opt_var2.c:229
FLA_LQ_UT_opd_var2
FLA_Error FLA_LQ_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_LQ_UT_opt_var2.c:160
FLA_LQ_UT_opz_var2
FLA_Error FLA_LQ_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_LQ_UT_opt_var2.c:298

References FLA_LQ_UT_opc_var2(), FLA_LQ_UT_opd_var2(), FLA_LQ_UT_ops_var2(), FLA_LQ_UT_opz_var2(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_width(), and i.

Referenced by FLA_LQ_UT_internal().

◆ FLA_LQ_UT_opz_var1()

FLA_Error FLA_LQ_UT_opz_var1 ( int  m_A,
int  n_A,
dcomplex *  A,
int  rs_A,
int  cs_A,
dcomplex *  t,
int  inc_t 
)
232{
233 int min_m_n = min( m_A, n_A );
234 int i;
235
236 for ( i = 0; i < min_m_n; ++i )
237 {
238 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
239 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
240 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
241 dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
242
243 dcomplex* tau1 = buff_t + (i )*inc_t;
244
245 int m_ahead = m_A - i - 1;
246 int n_ahead = n_A - i - 1;
247
248 /*------------------------------------------------------------*/
249
250 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
251 // tau1 );
252 FLA_Househ2_UT_r_opz( n_ahead,
253 alpha11,
254 a12t, cs_A,
255 tau1 );
256
257 // FLA_Apply_H2_UT( FLA_RIGHT, tau1, a12t, a21, A22 );
258 FLA_Apply_H2_UT_r_opz_var1( m_ahead,
259 n_ahead,
260 tau1,
261 a12t, cs_A,
262 a21, rs_A,
263 A22, rs_A, cs_A );
264
265 /*------------------------------------------------------------*/
266
267 }
268
269 return FLA_SUCCESS;
270}
FLA_Apply_H2_UT_r_opz_var1
FLA_Error FLA_Apply_H2_UT_r_opz_var1(int n_u2h_A2, int m_a1, dcomplex *tau, dcomplex *u2h, int inc_u2h, dcomplex *a1, int inc_a1, dcomplex *A2, int rs_A2, int cs_A2)
Definition FLA_Apply_H2_UT_r_opt_var1.c:327
FLA_Househ2_UT_r_opz
FLA_Error FLA_Househ2_UT_r_opz(int m_x2, dcomplex *chi_1, dcomplex *x2, int inc_x2, dcomplex *tau)
Definition FLA_Househ2_UT.c:693

References FLA_Apply_H2_UT_r_opz_var1(), FLA_Househ2_UT_r_opz(), and i.

Referenced by FLA_LQ_UT_opt_var1().

◆ FLA_LQ_UT_opz_var2()

FLA_Error FLA_LQ_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 
)
302{
303 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
304 int min_m_n = min( m_A, n_A );
305 int i;
306
307 for ( i = 0; i < min_m_n; ++i )
308 {
309 dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
310 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
311 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
312 dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
313 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
314 dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
315
316 dcomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
317 dcomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
318
319 int m_ahead = m_A - i - 1;
320 int n_ahead = n_A - i - 1;
321 int m_behind = i;
322
323 /*------------------------------------------------------------*/
324
325 // FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t
326 // tau11 );
327 FLA_Househ2_UT_r_opz( n_ahead,
328 alpha11,
329 a12t, cs_A,
330 tau11 );
331
332 // FLA_Apply_H2_UT( FLA_RIGHT, tau11, a12t, a21, A22 );
333 FLA_Apply_H2_UT_r_opz_var1( m_ahead,
334 n_ahead,
335 tau11,
336 a12t, cs_A,
337 a21, rs_A,
338 A22, rs_A, cs_A );
339
340 // FLA_Copyt_external( FLA_CONJ_NO_TRANSPOSE, a01, t01 );
341 bl1_zcopyv( BLIS1_CONJUGATE,
342 m_behind,
343 a01, rs_A,
344 t01, rs_T );
345
346 // FLA_Gemvc_external( FLA_CONJ_NO_TRANSPOSE, FLA_NO_CONJUGATE,
347 // FLA_ONE, A02, a12t, FLA_ONE, t01 );
348 bl1_zgemv( BLIS1_CONJ_NO_TRANSPOSE,
349 BLIS1_NO_CONJUGATE,
350 m_behind,
351 n_ahead,
352 buff_1,
353 A02, rs_A, cs_A,
354 a12t, cs_A,
355 buff_1,
356 t01, rs_T );
357
358 /*------------------------------------------------------------*/
359
360 }
361
362 return FLA_SUCCESS;
363}
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_NO_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, FLA_Apply_H2_UT_r_opz_var1(), FLA_Househ2_UT_r_opz(), FLA_ONE, and i.

Referenced by FLA_LQ_UT_opt_var2().

◆ FLA_LQ_UT_unb_var1()

FLA_Error FLA_LQ_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 FLA_Part_2x2( A, &ATL, &ATR,
22 &ABL, &ABR, 0, 0, FLA_TL );
23
24 FLA_Part_1x2( t, &tLt, &tRt, 0, FLA_LEFT );
25
26 while ( FLA_Obj_min_dim( ABR ) > 0 ){
27
28 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
29 /* ************* */ /* ************************** */
30 &a10t, /**/ &alpha11, &a12t,
31 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
32 1, 1, FLA_BR );
33
34 FLA_Repart_1x2_to_1x3( tLt, /**/ tRt, &t0t, /**/ &tau1, &t2t,
35 1, FLA_RIGHT );
36
37 /*------------------------------------------------------------*/
38
39 // Compute tau11 and u12t from alpha11 and a12t such that tau11 and u12t
40 // determine a Householder transform H such that applying H from the
41 // right to the row vector consisting of alpha11 and a12t annihilates
42 // the entries in a12t (and updates alpha11).
43 FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t,
44 tau1 );
45
46 // ( a21 A22 ) = ( a21 A22 ) H
47 //
48 // where H is formed from tau11 and u12t.
49 FLA_Apply_H2_UT( FLA_RIGHT, tau1, a12t, a21, A22 );
50
51 /*------------------------------------------------------------*/
52
53 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
54 a10t, alpha11, /**/ a12t,
55 /* ************** */ /* ************************ */
56 &ABL, /**/ &ABR, A20, a21, /**/ A22,
57 FLA_TL );
58
59 FLA_Cont_with_1x3_to_1x2( &tLt, /**/ &tRt, t0t, tau1, /**/ t2t,
60 FLA_LEFT );
61 }
62
63 return FLA_SUCCESS;
64}
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_LQ_UT_internal().

◆ FLA_LQ_UT_unb_var2()

FLA_Error FLA_LQ_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 u12t from alpha11 and a12t such that tau11 and u12t
47 // determine a Householder transform H such that applying H from the
48 // right to the row vector consisting of alpha11 and a12t annihilates
49 // the entries in a12t (and updates alpha11).
50 FLA_Househ2_UT( FLA_RIGHT, alpha11, a12t,
51 tau11 );
52
53 // ( a21 A22 ) = ( a21 A22 ) H
54 //
55 // where H is formed from tau11 and u12t.
56 FLA_Apply_H2_UT( FLA_RIGHT, tau11, a12t, a21, A22 );
57
58 // t01 = conj(a01) + conj(A02) * u12t^T;
59 FLA_Copyt_external( FLA_CONJ_NO_TRANSPOSE, a01, t01 );
60 FLA_Gemvc_external( FLA_CONJ_NO_TRANSPOSE, FLA_NO_CONJUGATE, FLA_ONE, A02, a12t, FLA_ONE, t01 );
61
62 /*------------------------------------------------------------*/
63
64 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
65 a10t, alpha11, /**/ a12t,
66 /* ************** */ /* ************************ */
67 &ABL, /**/ &ABR, A20, a21, /**/ A22,
68 FLA_TL );
69
70 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, t01, /**/ T02,
71 t10t, tau11, /**/ t12t,
72 /* ************** */ /* ********************** */
73 &TBL, /**/ &TBR, T20, t21, /**/ T22,
74 FLA_TL );
75
76 }
77
78 return FLA_SUCCESS;
79}
FLA_Gemvc_external
FLA_Error FLA_Gemvc_external(FLA_Trans transa, FLA_Conj conjx, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition FLA_Gemvc_external.c:13

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

Referenced by FLA_LQ_UT_internal().

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