libflame revision_anchor
FLA_Eig_gest_iu_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_Eig_gest_iu_blk_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl)
 

Function Documentation

◆ FLA_Eig_gest_iu_blk_var2()

FLA_Error FLA_Eig_gest_iu_blk_var2 ( FLA_Obj A,
FLA_Obj Y,
FLA_Obj B,
fla_eig_gest_t * cntl )
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BTL, BTR, B00, B01, B02,
22 BBL, BBR, B10, B11, B12,
23 B20, B21, B22;
24
25 FLA_Obj YT, Y01,
26 YB, Y11,
27 Y21;
28
29 FLA_Obj Y01_l, Y01_r;
30
31 dim_t b;
32
33 FLA_Part_2x2( A, &ATL, &ATR,
34 &ABL, &ABR, 0, 0, FLA_TL );
35
36 FLA_Part_2x2( B, &BTL, &BTR,
37 &BBL, &BBR, 0, 0, FLA_TL );
38
39 FLA_Part_2x1( Y, &YT,
40 &YB, 0, FLA_TOP );
41
42 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
43
44 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
45
46 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
47 /* ************* */ /* ******************** */
48 &A10, /**/ &A11, &A12,
49 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
50 b, b, FLA_BR );
51
52 FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
53 /* ************* */ /* ******************** */
54 &B10, /**/ &B11, &B12,
55 BBL, /**/ BBR, &B20, /**/ &B21, &B22,
56 b, b, FLA_BR );
57
58 FLA_Repart_2x1_to_3x1( YT, &Y01,
59 /* ** */ /* *** */
60 &Y11,
61 YB, &Y21, b, FLA_BOTTOM );
62
63 /*------------------------------------------------------------*/
64
65 FLA_Part_1x2( Y01, &Y01_l, &Y01_r, b, FLA_LEFT );
66
67 // Y01 = 1/2 * A00 * B01;
68 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
69 FLA_ONE_HALF, A00, B01, FLA_ZERO, Y01_l,
70 FLA_Cntl_sub_hemm( cntl ) );
71
72 // A01 = A01 - Y01;
74 FLA_Cntl_sub_axpy1( cntl ) );
75
76 // A11 = A11 - A01' * B01 - B01' * A01;
77 FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
78 FLA_MINUS_ONE, A01, B01, FLA_ONE, A11,
79 FLA_Cntl_sub_her2k( cntl ) );
80
81 // A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
82 FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
83 A11, Y11, B11,
84 FLA_Cntl_sub_eig_gest( cntl ) );
85
86 // A12 = A12 - B01' * A02;
87 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
88 FLA_MINUS_ONE, B01, A02, FLA_ONE, A12,
89 FLA_Cntl_sub_gemm1( cntl ) );
90
91 // A12 = inv( triu( B11 )' ) * A12;
92 FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
93 FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
94 FLA_ONE, B11, A12,
95 FLA_Cntl_sub_trsm1( cntl ) );
96
97 // A01 = A01 - Y01;
99 FLA_Cntl_sub_axpy2( cntl ) );
100
101 // A01 = A01 * inv( triu( B11 ) );
102 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
103 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
104 FLA_ONE, B11, A01,
105 FLA_Cntl_sub_trsm2( cntl ) );
106
107 /*------------------------------------------------------------*/
108
109 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
110 A10, A11, /**/ A12,
111 /* ************** */ /* ****************** */
112 &ABL, /**/ &ABR, A20, A21, /**/ A22,
113 FLA_TL );
114
115 FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
116 B10, B11, /**/ B12,
117 /* ************** */ /* ****************** */
118 &BBL, /**/ &BBR, B20, B21, /**/ B22,
119 FLA_TL );
120
121 FLA_Cont_with_3x1_to_2x1( &YT, Y01,
122 Y11,
123 /* ** */ /* *** */
124 &YB, Y21, FLA_TOP );
125 }
126
127 return FLA_SUCCESS;
128}
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_Error FLA_Eig_gest_internal(FLA_Inv inv, FLA_Uplo uplo, FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl)
Definition FLA_Eig_gest_internal.c:17
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_Error FLA_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition FLA_Hemm_internal.c:16
FLA_Error FLA_Her2k_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
Definition FLA_Her2k_internal.c:16
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_Obj FLA_MINUS_ONE
Definition FLA_Init.c:22
FLA_Obj FLA_ONE_HALF
Definition FLA_Init.c:19
FLA_Obj FLA_ZERO
Definition FLA_Init.c:20
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
FLA_Error FLA_Cont_with_3x3_to_2x2(FLA_Obj *ATL, FLA_Obj *ATR, FLA_Obj A00, FLA_Obj A01, FLA_Obj A02, FLA_Obj A10, FLA_Obj A11, FLA_Obj A12, FLA_Obj *ABL, FLA_Obj *ABR, FLA_Obj A20, FLA_Obj A21, FLA_Obj A22, FLA_Quadrant quadrant)
Definition FLA_View.c:304
FLA_Error FLA_Part_2x2(FLA_Obj A, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition FLA_View.c:17
FLA_Error FLA_Cont_with_3x1_to_2x1(FLA_Obj *AT, FLA_Obj A0, FLA_Obj A1, FLA_Obj *AB, FLA_Obj A2, FLA_Side side)
Definition FLA_View.c:428
FLA_Error FLA_Repart_2x1_to_3x1(FLA_Obj AT, FLA_Obj *A0, FLA_Obj *A1, FLA_Obj AB, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition FLA_View.c:226
FLA_Error FLA_Part_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition FLA_View.c:110
FLA_Error FLA_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition FLA_View.c:76
dim_t FLA_Obj_length(FLA_Obj obj)
Definition FLA_Query.c:116
FLA_Error FLA_Repart_2x2_to_3x3(FLA_Obj ATL, FLA_Obj ATR, FLA_Obj *A00, FLA_Obj *A01, FLA_Obj *A02, FLA_Obj *A10, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj ABL, FLA_Obj ABR, FLA_Obj *A20, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition FLA_View.c:142
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition FLA_Blocksize.c:234
unsigned long dim_t
Definition FLA_type_defs.h:71
Definition FLA_type_defs.h:159

References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_ONE_HALF, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.

Referenced by FLA_Eig_gest_iu().