libflame  revision_anchor
Functions
FLA_Eig_gest_nl_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_Eig_gest_nl_blk_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl)
 

Function Documentation

◆ FLA_Eig_gest_nl_blk_var2()

FLA_Error FLA_Eig_gest_nl_blk_var2 ( FLA_Obj  A,
FLA_Obj  Y,
FLA_Obj  B,
fla_eig_gest_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 BTL, BTR, B00, B01, B02,
20  BBL, BBR, B10, B11, B12,
21  B20, B21, B22;
22 
23  FLA_Obj YT, Y01,
24  YB, Y11,
25  Y21;
26 
27  FLA_Obj Y21_l, Y21_r;
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_2x1( Y, &YT,
38  &YB, 0, FLA_TOP );
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_2x1_to_3x1( YT, &Y01,
57  /* ** */ /* *** */
58  &Y11,
59  YB, &Y21, b, FLA_BOTTOM );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Part_1x2( Y21, &Y21_l, &Y21_r, b, FLA_LEFT );
64 
65  // A10 = tril( B11 )' * A10;
66  FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
67  FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
68  FLA_ONE, B11, A10,
69  FLA_Cntl_sub_trmm1( cntl ) );
70 
71  // A10 = A10 + B21' * A20;
72  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
73  FLA_ONE, B21, A20, FLA_ONE, A10,
74  FLA_Cntl_sub_gemm1( cntl ) );
75 
76  // Y21 = A22 * B21;
77  FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
78  FLA_ONE, A22, B21, FLA_ZERO, Y21_l,
79  FLA_Cntl_sub_hemm( cntl ) );
80 
81  // A21 = A21 * tril( B11 );
82  FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR,
83  FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
84  FLA_ONE, B11, A21,
85  FLA_Cntl_sub_trmm2( cntl ) );
86 
87  // A21 = A21 + 1/2 * Y21;
88  FLA_Axpy_internal( FLA_ONE_HALF, Y21_l, A21,
89  FLA_Cntl_sub_axpy1( cntl ) );
90 
91  // A11 = tril( B11 )' * A11 * tril( B11 );
92  FLA_Eig_gest_internal( FLA_NO_INVERSE, FLA_LOWER_TRIANGULAR,
93  A11, Y11, B11,
94  FLA_Cntl_sub_eig_gest( cntl ) );
95 
96  // A11 = A11 + A21' * B21 + B21' * A21;
97  FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
98  FLA_ONE, A21, B21, FLA_ONE, A11,
99  FLA_Cntl_sub_her2k( cntl ) );
100 
101  // A21 = A21 + 1/2 * Y21;
102  FLA_Axpy_internal( FLA_ONE_HALF, Y21_l, A21,
103  FLA_Cntl_sub_axpy2( cntl ) );
104 
105  /*------------------------------------------------------------*/
106 
107  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
108  A10, A11, /**/ A12,
109  /* ************** */ /* ****************** */
110  &ABL, /**/ &ABR, A20, A21, /**/ A22,
111  FLA_TL );
112 
113  FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
114  B10, B11, /**/ B12,
115  /* ************** */ /* ****************** */
116  &BBL, /**/ &BBR, B20, B21, /**/ B22,
117  FLA_TL );
118 
119  FLA_Cont_with_3x1_to_2x1( &YT, Y01,
120  Y11,
121  /* ** */ /* *** */
122  &YB, Y21, FLA_TOP );
123  }
124 
125  return FLA_SUCCESS;
126 }
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_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
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_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_Trmm_internal(), and FLA_ZERO.

Referenced by FLA_Eig_gest_nl().