libflame  revision_anchor
Functions
FLA_Eig_gest_nu_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_Eig_gest_nu_blk_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl)
 

Function Documentation

◆ FLA_Eig_gest_nu_blk_var1()

FLA_Error FLA_Eig_gest_nu_blk_var1 ( FLA_Obj  A,
FLA_Obj  Y,
FLA_Obj  B,
fla_eig_gest_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_ONE_HALF, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_internal(), and FLA_ZERO.

Referenced by FLA_Eig_gest_nu().

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 YL, YR, Y10, Y11, Y12;
26 
27  FLA_Obj Y12_t,
28  Y12_b;
29 
30  dim_t b;
31 
32  FLA_Part_2x2( A, &ATL, &ATR,
33  &ABL, &ABR, 0, 0, FLA_TL );
34 
35  FLA_Part_2x2( B, &BTL, &BTR,
36  &BBL, &BBR, 0, 0, FLA_TL );
37 
38  FLA_Part_1x2( Y, &YL, &YR, 0, FLA_LEFT );
39 
40  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
41 
42  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
43 
44  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
45  /* ************* */ /* ******************** */
46  &A10, /**/ &A11, &A12,
47  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
48  b, b, FLA_BR );
49 
50  FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
51  /* ************* */ /* ******************** */
52  &B10, /**/ &B11, &B12,
53  BBL, /**/ BBR, &B20, /**/ &B21, &B22,
54  b, b, FLA_BR );
55 
56  FLA_Repart_1x2_to_1x3( YL, /**/ YR, &Y10, /**/ &Y11, &Y12,
57  b, FLA_RIGHT );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Part_2x1( Y12, &Y12_t,
62  &Y12_b, b, FLA_TOP );
63 
64  // Y12 = B12 * A22;
65  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
66  FLA_ONE, A22, B12, FLA_ZERO, Y12_t,
67  FLA_Cntl_sub_hemm( cntl ) );
68 
69  // A12 = triu( B11 ) * A12;
70  FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
71  FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
72  FLA_ONE, B11, A12,
73  FLA_Cntl_sub_trmm1( cntl ) );
74 
75  // A12 = A12 + 1/2 * Y12;
76  FLA_Axpy_internal( FLA_ONE_HALF, Y12_t, A12,
77  FLA_Cntl_sub_axpy1( cntl ) );
78 
79  // A11 = triu( B11 ) * A11 * triu( B11 )';
80  FLA_Eig_gest_internal( FLA_NO_INVERSE, FLA_UPPER_TRIANGULAR,
81  A11, Y11, B11,
82  FLA_Cntl_sub_eig_gest( cntl ) );
83 
84  // A11 = A11 + A12 * B12' + B12 * A12';
85  FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
86  FLA_ONE, A12, B12, FLA_ONE, A11,
87  FLA_Cntl_sub_her2k( cntl ) );
88 
89  // A12 = A12 + 1/2 * Y12;
90  FLA_Axpy_internal( FLA_ONE_HALF, Y12_t, A12,
91  FLA_Cntl_sub_axpy2( cntl ) );
92 
93  // A12 = A12 * triu( B22 )';
94  FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
95  FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
96  FLA_ONE, B22, A12,
97  FLA_Cntl_sub_trmm2( cntl ) );
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( &BTL, /**/ &BTR, B00, B01, /**/ B02,
108  B10, B11, /**/ B12,
109  /* ************** */ /* ****************** */
110  &BBL, /**/ &BBR, B20, B21, /**/ B22,
111  FLA_TL );
112 
113  FLA_Cont_with_1x3_to_1x2( &YL, /**/ &YR, Y10, Y11, /**/ Y12,
114  FLA_LEFT );
115  }
116 
117  return FLA_SUCCESS;
118 }
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
unsigned long dim_t
Definition: FLA_type_defs.h:71
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
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition: FLA_Blocksize.c:234
FLA_Obj FLA_ONE_HALF
Definition: FLA_Init.c:19
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_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_Obj FLA_ONE
Definition: FLA_Init.c:18
FLA_Error FLA_Trmm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl)
Definition: FLA_Trmm_internal.c:16
Definition: FLA_type_defs.h:158
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_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_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_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_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:110
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_Axpy_internal(FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)
Definition: FLA_Axpy_internal.c:16
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
FLA_Obj FLA_ZERO
Definition: FLA_Init.c:20