libflame  revision_anchor
Functions
FLA_Syrk_ut_blk_var4.c File Reference

(r)

Functions

FLA_Error FLA_Syrk_ut_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 

Function Documentation

◆ FLA_Syrk_ut_blk_var4()

FLA_Error FLA_Syrk_ut_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
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 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_BR );
29 
30  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
31 
32  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
33 
34  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
35  b, FLA_LEFT );
36 
37  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
38  &C10, &C11, /**/ &C12,
39  /* ************* */ /* ******************** */
40  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
41  b, b, FLA_TL );
42 
43  /*------------------------------------------------------------*/
44 
45  /* C01 = C01 + A0' * A1 */
46  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
47  alpha, A0, A1, beta, C01,
48  FLA_Cntl_sub_gemm( cntl ) );
49 
50  /* C11 = C11 + A1' * A1 */
51  FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE,
52  alpha, A1, beta, C11,
53  FLA_Cntl_sub_syrk( cntl ) );
54 
55  /*------------------------------------------------------------*/
56 
57  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
58  FLA_RIGHT );
59 
60  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
61  /* ************** */ /* ****************** */
62  C10, /**/ C11, C12,
63  &CBL, /**/ &CBR, C20, /**/ C21, C22,
64  FLA_BR );
65 
66  }
67 
68  return FLA_SUCCESS;
69 }
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_Syrk_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
Definition: FLA_Syrk_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_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
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_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ut().