libflame revision_anchor
Functions
FLA_QR_UT_inc_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_QR_UT_inc_blk_var1 (FLA_Obj A, FLA_Obj TW, fla_qrutinc_t *cntl)
 

Function Documentation

◆ FLA_QR_UT_inc_blk_var1()

FLA_Error FLA_QR_UT_inc_blk_var1 ( FLA_Obj  A,
FLA_Obj  TW,
fla_qrutinc_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, WTR, T00, W01, W02,
20 TBL, TBR, T10, T11, W12,
21 T20, T21, T22;
22
23 dim_t b;
24
25 FLA_Part_2x2( A, &ATL, &ATR,
26 &ABL, &ABR, 0, 0, FLA_TL );
27
29 &TBL, &TBR, 0, 0, FLA_TL );
30
31 while ( FLA_Obj_min_dim( ABR ) > 0 ){
32
34
35 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
36 /* ************* */ /* ******************** */
37 &A10, /**/ &A11, &A12,
38 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
39 b, b, FLA_BR );
40
41 FLA_Repart_2x2_to_3x3( TTL, /**/ WTR, &T00, /**/ &W01, &W02,
42 /* ************* */ /* ******************** */
43 &T10, /**/ &T11, &W12,
44 TBL, /**/ TBR, &T20, /**/ &T21, &T22,
45 b, b, FLA_BR );
46
47 /*------------------------------------------------------------*/
48
49 /*
50 Perform a QR factorization (via UT transform) on A11:
51
52 [ A11, T11 ] = QR_UT( A11, T11 );
53
54 where T11 refers to a single storage block that refers to an
55 b_alg-by-b row-panel of upper triangular block Householder
56 transforms. Here, b is the storage blocksize while b_alg is
57 the algorithmic blocksize used by the QR factorization.
58 Typically b_alg << b.
59
60 */
61
63 FLA_Cntl_sub_qrut( cntl ) );
64
65
66 if ( FLA_Obj_width( A12 ) > 0 )
67 {
68 /*
69 Apply Q^H to A12 from the left:
70
71 A12 = Q^H * A12
72
73 where Q is formed from A11 and T11. Note that W12 refers
74 to a row-panel of blocks where each block refers to an
75 b_alg-by-b row-panel of workspace.
76 */
77
79 A11, T11, W12, A12,
80 FLA_Cntl_sub_apqut( cntl ) );
81 }
82
83
84 /*
85 Update QR factorization of A11 with each block of A21, storing
86 block Householder transforms into corresponding blocks of T21.
87
88 [ A11, ...
89 A21, T21 ] = QR2_UT( A11, ...
90 A21, T21 );
91 */
92
94 A21, T21,
95 FLA_Cntl_sub_qr2ut( cntl ) );
96
97
98 if ( FLA_Obj_width( A12 ) > 0 )
99 {
100 /*
101 Apply Q^H to A12 and A22 from the left:
102
103 / A12 \ = Q^H * / A12 \
104 \ A22 / \ A22 /
105
106 where Q is formed from A21 and T21.
107 */
108
110 A21, T21, W12, A12,
111 A22,
112 FLA_Cntl_sub_apq2ut( cntl ) );
113 }
114
115 /*------------------------------------------------------------*/
116
117 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
118 A10, A11, /**/ A12,
119 /* ************** */ /* ****************** */
120 &ABL, /**/ &ABR, A20, A21, /**/ A22,
121 FLA_TL );
122
123 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &WTR, T00, W01, /**/ W02,
124 T10, T11, /**/ W12,
125 /* ************** */ /* ****************** */
126 &TBL, /**/ &TBR, T20, T21, /**/ T22,
127 FLA_TL );
128
129 }
130
131 return FLA_SUCCESS;
132}
FLA_Error FLA_Apply_Q2_UT_internal(FLA_Side side, FLA_Trans trans, FLA_Direct direct, FLA_Store storev, FLA_Obj D, FLA_Obj T, FLA_Obj W, FLA_Obj C, FLA_Obj E, fla_apq2ut_t *cntl)
Definition FLA_Apply_Q2_UT_internal.c:17
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_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_Error FLA_QR_UT_internal(FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl)
Definition FLA_QR_UT_internal.c:17
FLA_Error FLA_Cont_with_3x3_to_2x2(FLA_Obj *ATL, FLA_Obj *ATR, FLA_Obj A00, FLA_Obj A01, FLA_Obj A02, FLA_Obj A10, FLA_Obj A11, FLA_Obj A12, FLA_Obj *ABL, FLA_Obj *ABR, FLA_Obj A20, FLA_Obj A21, FLA_Obj A22, FLA_Quadrant quadrant)
Definition FLA_View.c:304
FLA_Error FLA_Part_2x2(FLA_Obj A, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition FLA_View.c:17
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
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
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition FLA_Blocksize.c:234
dim_t FLA_Obj_min_dim(FLA_Obj obj)
Definition FLA_Query.c:153
unsigned long dim_t
Definition FLA_type_defs.h:71
int i
Definition bl1_axmyv2.c:145
Definition FLA_type_defs.h:159

References FLA_Apply_Q2_UT_internal(), FLA_Apply_Q_UT_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_Part_2x2(), FLA_QR2_UT_internal(), FLA_QR_UT_internal(), FLA_Repart_2x2_to_3x3(), and i.

Referenced by FLASH_QR_UT_inc_noopt().