libflame  revision_anchor
Functions
FLA_Hemm_ll_blk_var3.c File Reference

(r)

Functions

FLA_Error FLA_Hemm_ll_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 

Function Documentation

◆ FLA_Hemm_ll_blk_var3()

FLA_Error FLA_Hemm_ll_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16 {
17  FLA_Obj ATL, ATR, A00, A01, A02,
18  ABL, ABR, A10, A11, A12,
19  A20, A21, A22;
20 
21  FLA_Obj BT, B0,
22  BB, B1,
23  B2;
24 
25  FLA_Obj CT, C0,
26  CB, C1,
27  C2;
28 
29  dim_t b;
30 
31  FLA_Scal_internal( beta, C,
32  FLA_Cntl_sub_scal( cntl ) );
33 
34  FLA_Part_2x2( A, &ATL, &ATR,
35  &ABL, &ABR, 0, 0, FLA_TL );
36 
37  FLA_Part_2x1( B, &BT,
38  &BB, 0, FLA_TOP );
39 
40  FLA_Part_2x1( C, &CT,
41  &CB, 0, FLA_TOP );
42 
43  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
44 
45  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
46 
47  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
48  /* ************* */ /* ******************** */
49  &A10, /**/ &A11, &A12,
50  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
51  b, b, FLA_BR );
52 
53  FLA_Repart_2x1_to_3x1( BT, &B0,
54  /* ** */ /* ** */
55  &B1,
56  BB, &B2, b, FLA_BOTTOM );
57 
58  FLA_Repart_2x1_to_3x1( CT, &C0,
59  /* ** */ /* ** */
60  &C1,
61  CB, &C2, b, FLA_BOTTOM );
62 
63  /*------------------------------------------------------------*/
64 
65  /* C1 = C1 + A11 * B1 */
66  FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
67  alpha, A11, B1, FLA_ONE, C1,
68  FLA_Cntl_sub_hemm( cntl ) );
69 
70  /* C1 = C1 + A21' * B2 */
71  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
72  alpha, A21, B2, FLA_ONE, C1,
73  FLA_Cntl_sub_gemm1( cntl ) );
74 
75  /* C2 = C2 + A21 * B1 */
76  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
77  alpha, A21, B1, FLA_ONE, C2,
78  FLA_Cntl_sub_gemm2( cntl ) );
79 
80  /*------------------------------------------------------------*/
81 
82  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
83  A10, A11, /**/ A12,
84  /* ************** */ /* ****************** */
85  &ABL, /**/ &ABR, A20, A21, /**/ A22,
86  FLA_TL );
87 
88  FLA_Cont_with_3x1_to_2x1( &BT, B0,
89  B1,
90  /* ** */ /* ** */
91  &BB, B2, FLA_TOP );
92 
93  FLA_Cont_with_3x1_to_2x1( &CT, C0,
94  C1,
95  /* ** */ /* ** */
96  &CB, C2, FLA_TOP );
97 
98  }
99 
100  return FLA_SUCCESS;
101 }
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
FLA_Error FLA_Scal_internal(FLA_Obj alpha, FLA_Obj A, fla_scal_t *cntl)
Definition: FLA_Scal_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_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_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_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition: FLA_View.c:76
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
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
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_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().

Referenced by FLA_Hemm_ll().