libflame  revision_anchor
Functions
FLA_Eig_gest_iu_blk_var3.c File Reference

(r)

Functions

FLA_Error FLA_Eig_gest_iu_blk_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl)
 

Function Documentation

◆ FLA_Eig_gest_iu_blk_var3()

FLA_Error FLA_Eig_gest_iu_blk_var3 ( FLA_Obj  A,
FLA_Obj  Y,
FLA_Obj  B,
fla_eig_gest_t cntl 
)

References FLA_Axpy_internal(), 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_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.

Referenced by FLA_Eig_gest_iu().

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 YTL, YTR, Y00, Y01, Y02,
26  YBL, YBR, Y10, Y11, Y12,
27  Y20, Y21, Y22;
28 
29  dim_t b;
30 
31  FLA_Part_2x2( A, &ATL, &ATR,
32  &ABL, &ABR, 0, 0, FLA_TL );
33 
34  FLA_Part_2x2( B, &BTL, &BTR,
35  &BBL, &BBR, 0, 0, FLA_TL );
36 
37  FLA_Part_2x2( Y, &YTL, &YTR,
38  &YBL, &YBR, 0, 0, FLA_TL );
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_2x2_to_3x3( YTL, /**/ YTR, &Y00, /**/ &Y01, &Y02,
57  /* ************* */ /* ******************** */
58  &Y10, /**/ &Y11, &Y12,
59  YBL, /**/ YBR, &Y20, /**/ &Y21, &Y22,
60  b, b, FLA_BR );
61 
62  /*------------------------------------------------------------*/
63 
64  // A01 = A01 - 1/2 * Y01;
66  FLA_Cntl_sub_axpy1( cntl ) );
67 
68  // A11 = A11 - A01' * B01 - B01' * A01;
69  FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
70  FLA_MINUS_ONE, A01, B01, FLA_ONE, A11,
71  FLA_Cntl_sub_her2k( cntl ) );
72 
73  // A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
74  FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
75  A11, Y11, B11,
76  FLA_Cntl_sub_eig_gest( cntl ) );
77 
78  // A12 = A12 - B01' * A02;
79  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
80  FLA_MINUS_ONE, B01, A02, FLA_ONE, A12,
81  FLA_Cntl_sub_gemm1( cntl ) );
82 
83  // A12 = A12 * inv( triu( B11 )' );
84  FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
85  FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
86  FLA_ONE, B11, A12,
87  FLA_Cntl_sub_trsm1( cntl ) );
88 
89  // A01 = A01 - 1/2 * Y01;
91  FLA_Cntl_sub_axpy2( cntl ) );
92 
93  // A01 = A01 * inv( triu( B11 ) );
94  FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
95  FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
96  FLA_ONE, B11, A01,
97  FLA_Cntl_sub_trsm2( cntl ) );
98 
99  // Y02 = Y02 + A01 * B12;
100  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
101  FLA_ONE, A01, B12, FLA_ONE, Y02,
102  FLA_Cntl_sub_gemm2( cntl ) );
103 
104  // Y12 = A11 * B12;
105  FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
106  FLA_ONE, A11, B12, FLA_ZERO, Y12,
107  FLA_Cntl_sub_hemm( cntl ) );
108 
109  // Y12 = Y12 + A01' * B02;
110  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
111  FLA_ONE, A01, B02, FLA_ONE, Y12,
112  FLA_Cntl_sub_gemm3( cntl ) );
113 
114  /*------------------------------------------------------------*/
115 
116  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
117  A10, A11, /**/ A12,
118  /* ************** */ /* ****************** */
119  &ABL, /**/ &ABR, A20, A21, /**/ A22,
120  FLA_TL );
121 
122  FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
123  B10, B11, /**/ B12,
124  /* ************** */ /* ****************** */
125  &BBL, /**/ &BBR, B20, B21, /**/ B22,
126  FLA_TL );
127 
128  FLA_Cont_with_3x3_to_2x2( &YTL, /**/ &YTR, Y00, Y01, /**/ Y02,
129  Y10, Y11, /**/ Y12,
130  /* ************** */ /* ****************** */
131  &YBL, /**/ &YBR, Y20, Y21, /**/ Y22,
132  FLA_TL );
133  }
134 
135  return FLA_SUCCESS;
136 }
FLA_Obj FLA_MINUS_ONE_HALF
Definition: FLA_Init.c:21
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
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
FLA_Obj FLA_MINUS_ONE
Definition: FLA_Init.c:22
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition: FLA_Blocksize.c:234
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
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_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_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
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