libflame  revision_anchor
Functions
FLA_Sylv_hh_blk_var15.c File Reference

(r)

Functions

FLA_Error FLA_Sylv_hh_blk_var15 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl)
 

Function Documentation

◆ FLA_Sylv_hh_blk_var15()

FLA_Error FLA_Sylv_hh_blk_var15 ( FLA_Obj  isgn,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  C,
FLA_Obj  scale,
fla_sylv_t cntl 
)
14 {
15  FLA_Obj ATL, ATR, A00, A01, A02,
16  ABL, ABR, A10, A11, A12,
17  A20, A21, A22;
18 
19  FLA_Obj CT, C0,
20  CB, C1,
21  C2;
22 
23  dim_t b;
24 
25  FLA_Part_2x2( A, &ATL, &ATR,
26  &ABL, &ABR, 0, 0, FLA_TL );
27 
28  FLA_Part_2x1( C, &CT,
29  &CB, 0, FLA_TOP );
30 
31  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
32 
33  b = FLA_Determine_blocksize( CB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
34 
35  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
36  /* ************* */ /* ******************** */
37  &A10, /**/ &A11, &A12,
38  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
39  b, b, FLA_BR );
40 
41  FLA_Repart_2x1_to_3x1( CT, &C0,
42  /* ** */ /* ** */
43  &C1,
44  CB, &C2, b, FLA_BOTTOM );
45 
46  // Loop Invariant:
47  // CT = sylv( ATL', B', CT )
48  // CB = CB
49 
50  /*------------------------------------------------------------*/
51 
52  // C1 = sylv( A11', B', C1 - A01' * C0 );
53  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
54  FLA_MINUS_ONE, A01, C0, FLA_ONE, C1,
55  FLA_Cntl_sub_gemm1( cntl ) );
56 
57  FLA_Sylv_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
58  isgn, A11, B, C1, scale,
59  FLA_Cntl_sub_sylv1( cntl ) );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
64  A10, A11, /**/ A12,
65  /* ************** */ /* ****************** */
66  &ABL, /**/ &ABR, A20, A21, /**/ A22,
67  FLA_TL );
68 
69  FLA_Cont_with_3x1_to_2x1( &CT, C0,
70  C1,
71  /* ** */ /* ** */
72  &CB, C2, FLA_TOP );
73 
74  }
75 
76  return FLA_SUCCESS;
77 }
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_Sylv_internal(FLA_Trans transa, FLA_Trans transb, FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl)
Definition: FLA_Sylv_internal.c:16
FLA_Obj FLA_MINUS_ONE
Definition: FLA_Init.c:22
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_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().

Referenced by FLA_Sylv_hh().