libflame  revision_anchor
Functions
FLA_Symm_rl_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_Symm_rl_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_symm_t *cntl)
 

Function Documentation

◆ FLA_Symm_rl_blk_var1()

FLA_Error FLA_Symm_rl_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_symm_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and FLA_Symm_internal().

Referenced by FLA_Symm_rl().

16 {
17  FLA_Obj ATL, ATR, A00, A01, A02,
18  ABL, ABR, A10, A11, A12,
19  A20, A21, A22;
20 
21  FLA_Obj BL, BR, B0, B1, B2;
22 
23  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
28  FLA_Cntl_sub_scal( cntl ) );
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_TL );
32 
33  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
36 
37  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
42  /* ************* */ /* ******************** */
43  &A10, /**/ &A11, &A12,
44  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
45  b, b, FLA_BR );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
48  b, FLA_RIGHT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
51  b, FLA_RIGHT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C0 = C0 + B1 * A10 */
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, B1, A10, FLA_ONE, C0,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C1 = C1 + B0 * A10' */
61  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
62  alpha, B0, A10, FLA_ONE, C1,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C1 = C1 + B1 * A11 */
66  FLA_Symm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR,
67  alpha, A11, B1, FLA_ONE, C1,
68  FLA_Cntl_sub_symm( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
73  A10, A11, /**/ A12,
74  /* ************** */ /* ****************** */
75  &ABL, /**/ &ABR, A20, A21, /**/ A22,
76  FLA_TL );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
79  FLA_LEFT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
82  FLA_LEFT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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
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_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
FLA_Error FLA_Symm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_symm_t *cntl)
Definition: FLA_Symm_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_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:110
FLA_Error FLA_Scal_internal(FLA_Obj alpha, FLA_Obj A, fla_scal_t *cntl)
Definition: FLA_Scal_internal.c:16
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116