libflame revision_anchor
Functions
FLA_Syr2k_ut_blk_var3.c File Reference

(r)

Functions

FLA_Error FLA_Syr2k_ut_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 

Function Documentation

◆ FLA_Syr2k_ut_blk_var3()

FLA_Error FLA_Syr2k_ut_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
14{
15 FLA_Obj AL, AR, A0, A1, A2;
16
17 FLA_Obj BL, BR, B0, B1, B2;
18
19 FLA_Obj CTL, CTR, C00, C01, C02,
20 CBL, CBR, C10, C11, C12,
21 C20, C21, C22;
22
23 dim_t b;
24
26 FLA_Cntl_sub_scalr( cntl ) );
27
28 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
29
30 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31
32 FLA_Part_2x2( C, &CTL, &CTR,
33 &CBL, &CBR, 0, 0, FLA_TL );
34
35 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
36
38
39 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
40 b, FLA_RIGHT );
41
42 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
43 b, FLA_RIGHT );
44
45 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
46 /* ************* */ /* ******************** */
47 &C10, /**/ &C11, &C12,
48 CBL, /**/ CBR, &C20, /**/ &C21, &C22,
49 b, b, FLA_BR );
50
51 /*------------------------------------------------------------*/
52
53 /* C01 = C01 + A0' * B1 */
55 alpha, A0, B1, FLA_ONE, C01,
56 FLA_Cntl_sub_gemm1( cntl ) );
57
58 /* C12 = C12 + B1' * A2 */
60 alpha, B1, A2, FLA_ONE, C12,
61 FLA_Cntl_sub_gemm2( cntl ) );
62
63 /* C11 = C11 + A1' * B1 + B1' * A1 */
65 alpha, A1, B1, FLA_ONE, C11,
66 FLA_Cntl_sub_syr2k( cntl ) );
67
68 /*------------------------------------------------------------*/
69
70 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
71 FLA_LEFT );
72
73 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
74 FLA_LEFT );
75
76 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
77 C10, C11, /**/ C12,
78 /* ************** */ /* ****************** */
79 &CBL, /**/ &CBR, C20, C21, /**/ C22,
80 FLA_TL );
81
82 }
83
84 return FLA_SUCCESS;
85}
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_Scalr_internal(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, fla_scalr_t *cntl)
Definition FLA_Scalr_internal.c:16
FLA_Error FLA_Syr2k_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
Definition FLA_Syr2k_internal.c:16
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_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
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
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_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_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
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
int i
Definition bl1_axmyv2.c:145
Definition FLA_type_defs.h:159

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), FLA_Syr2k_internal(), and i.

Referenced by FLA_Syr2k_ut().