libflame  revision_anchor
Functions
FLA_Her2k_un_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_Her2k_un_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 

Function Documentation

◆ FLA_Her2k_un_blk_var2()

FLA_Error FLA_Her2k_un_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_internal().

Referenced by FLA_Her2k_un().

16 {
17  FLA_Obj AT, A0,
18  AB, A1,
19  A2;
20 
21  FLA_Obj BT, B0,
22  BB, B1,
23  B2;
24 
25  FLA_Obj CTL, CTR, C00, C01, C02,
26  CBL, CBR, C10, C11, C12,
27  C20, C21, C22;
28 
29  dim_t b;
30 
31  FLA_Scalr_internal( FLA_UPPER_TRIANGULAR, beta, C,
32  FLA_Cntl_sub_scalr( cntl ) );
33 
34  FLA_Part_2x1( A, &AT,
35  &AB, 0, FLA_TOP );
36 
37  FLA_Part_2x1( B, &BT,
38  &BB, 0, FLA_TOP );
39 
40  FLA_Part_2x2( C, &CTL, &CTR,
41  &CBL, &CBR, 0, 0, FLA_TL );
42 
43  while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
44 
45  b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
46 
47  FLA_Repart_2x1_to_3x1( AT, &A0,
48  /* ** */ /* ** */
49  &A1,
50  AB, &A2, b, FLA_BOTTOM );
51 
52  FLA_Repart_2x1_to_3x1( BT, &B0,
53  /* ** */ /* ** */
54  &B1,
55  BB, &B2, b, FLA_BOTTOM );
56 
57  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
58  /* ************* */ /* ******************** */
59  &C10, /**/ &C11, &C12,
60  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
61  b, b, FLA_BR );
62 
63  /*------------------------------------------------------------*/
64 
65  /* C01 = C01 + B0 * A1' */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
67  alpha, B0, A1, FLA_ONE, C01,
68  FLA_Cntl_sub_gemm1( cntl ) );
69 
70  /* C12 = C12 + A1 * B2' */
71  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
72  alpha, A1, B2, FLA_ONE, C12,
73  FLA_Cntl_sub_gemm2( cntl ) );
74 
75  /* C11 = C11 + A1 * B1' + B1 * A1' */
76  FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
77  alpha, A1, B1, FLA_ONE, C11,
78  FLA_Cntl_sub_her2k( cntl ) );
79 
80  /*------------------------------------------------------------*/
81 
82  FLA_Cont_with_3x1_to_2x1( &AT, A0,
83  A1,
84  /* ** */ /* ** */
85  &AB, A2, FLA_TOP );
86 
87  FLA_Cont_with_3x1_to_2x1( &BT, B0,
88  B1,
89  /* ** */ /* ** */
90  &BB, B2, FLA_TOP );
91 
92  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
93  C10, C11, /**/ C12,
94  /* ************** */ /* ****************** */
95  &CBL, /**/ &CBR, C20, C21, /**/ C22,
96  FLA_TL );
97 
98  }
99 
100  return FLA_SUCCESS;
101 }
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_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
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_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_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_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_Scalr_internal(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, fla_scalr_t *cntl)
Definition: FLA_Scalr_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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116