libflame  revision_anchor
Functions
FLA_Syrk_ln_blk_var4.c File Reference

(r)

Functions

FLA_Error FLA_Syrk_ln_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 

Function Documentation

◆ FLA_Syrk_ln_blk_var4()

FLA_Error FLA_Syrk_ln_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

16 {
17  FLA_Obj AT, A0,
18  AB, A1,
19  A2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Part_2x1( A, &AT,
28  &AB, 0, FLA_BOTTOM );
29 
30  FLA_Part_2x2( C, &CTL, &CTR,
31  &CBL, &CBR, 0, 0, FLA_BR );
32 
33  while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
34 
35  b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
36 
37  FLA_Repart_2x1_to_3x1( AT, &A0,
38  &A1,
39  /* ** */ /* ** */
40  AB, &A2, b, FLA_TOP );
41 
42  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
43  &C10, &C11, /**/ &C12,
44  /* ************* */ /* ******************** */
45  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
46  b, b, FLA_TL );
47 
48  /*------------------------------------------------------------*/
49 
50  /* C10 = C10 + A1 * A0' */
51  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
52  alpha, A1, A0, beta, C10,
53  FLA_Cntl_sub_gemm( cntl ) );
54 
55  /* C11 = C11 + A1 * A1' */
56  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE,
57  alpha, A1, beta, C11,
58  FLA_Cntl_sub_syrk( cntl ) );
59 
60  /*------------------------------------------------------------*/
61 
62  FLA_Cont_with_3x1_to_2x1( &AT, A0,
63  /* ** */ /* ** */
64  A1,
65  &AB, A2, FLA_BOTTOM );
66 
67  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
68  /* ************** */ /* ****************** */
69  C10, /**/ C11, C12,
70  &CBL, /**/ &CBR, C20, /**/ C21, C22,
71  FLA_BR );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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
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
Definition: FLA_type_defs.h:158
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_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_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