libflame  revision_anchor
Functions
FLA_Sylv_hh_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_Sylv_hh_blk_var2 (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_var2()

FLA_Error FLA_Sylv_hh_blk_var2 ( FLA_Obj  isgn,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  C,
FLA_Obj  scale,
fla_sylv_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 BTL, BTR, B00, B01, B02,
22  BBL, BBR, B10, B11, B12,
23  B20, B21, B22;
24 
25  FLA_Obj CTL, CTR, C00, C01, C02,
26  CBL, CBR, C10, C11, C12,
27  C20, C21, C22;
28 
29  dim_t b;
30 
31  FLA_Part_2x2( A, &ATL, &ATR,
32  &ABL, &ABR, 0, 0, FLA_TL );
33 
34  FLA_Part_2x2( B, &BTL, &BTR,
35  &BBL, &BBR, 0, 0, FLA_BR );
36 
37  FLA_Part_2x2( C, &CTL, &CTR,
38  &CBL, &CBR, 0, 0, FLA_TR );
39 
40  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
41 
42  b = FLA_Determine_blocksize( CBL, FLA_BL, FLA_Cntl_blocksize( cntl ) );
43 
44  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
45  /* ************* */ /* ******************** */
46  &A10, /**/ &A11, &A12,
47  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
48  b, b, FLA_BR );
49 
50  FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, &B01, /**/ &B02,
51  &B10, &B11, /**/ &B12,
52  /* ************* */ /* ******************** */
53  BBL, /**/ BBR, &B20, &B21, /**/ &B22,
54  b, b, FLA_TL );
55 
56  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
57  /* ************* */ /* ******************** */
58  &C10, &C11, /**/ &C12,
59  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
60  b, b, FLA_BL );
61 
62  // Loop Invariant:
63  // CTL =
64  // CTR =
65  // CBL =
66  // CBR =
67 
68  /*------------------------------------------------------------*/
69 
70  // C12 = sylv( A11', B22', C12 );
71  FLA_Sylv_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
72  isgn, A11, B22, C12, scale,
73  FLA_Cntl_sub_sylv1( cntl ) );
74 
75  // C22 = C22 - A12' * C12;
76  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
77  FLA_MINUS_ONE, A12, C12, FLA_ONE, C22,
78  FLA_Cntl_sub_gemm1( cntl ) );
79 
80  // C01 = sylv( A00', B11', C01 -/+ C02 * B12' );
81  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
82  FLA_NEGATE( isgn ), C02, B12, FLA_ONE, C01,
83  FLA_Cntl_sub_gemm2( cntl ) );
84 
85  FLA_Sylv_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
86  isgn, A00, B11, C01, scale,
87  FLA_Cntl_sub_sylv2( cntl ) );
88 
89  // C11 = sylv( A11', B11', C11 - A01' * C01 -/+ C12 * B12' );
90  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
91  FLA_NEGATE( isgn ), C12, B12, FLA_ONE, C11,
92  FLA_Cntl_sub_gemm3( cntl ) );
93 
94  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
95  FLA_MINUS_ONE, A01, C01, FLA_ONE, C11,
96  FLA_Cntl_sub_gemm4( cntl ) );
97 
98  FLA_Sylv_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
99  isgn, A11, B11, C11, scale,
100  FLA_Cntl_sub_sylv3( cntl ) );
101 
102  // C21 = C21 - A12' * C11 - A02' * C01;
103  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
104  FLA_MINUS_ONE, A02, C01, FLA_ONE, C21,
105  FLA_Cntl_sub_gemm5( cntl ) );
106 
107  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
108  FLA_MINUS_ONE, A12, C11, FLA_ONE, C21,
109  FLA_Cntl_sub_gemm6( cntl ) );
110 
111  /*------------------------------------------------------------*/
112 
113  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
114  A10, A11, /**/ A12,
115  /* ************** */ /* ****************** */
116  &ABL, /**/ &ABR, A20, A21, /**/ A22,
117  FLA_TL );
118 
119  FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, /**/ B01, B02,
120  /* ************** */ /* ****************** */
121  B10, /**/ B11, B12,
122  &BBL, /**/ &BBR, B20, /**/ B21, B22,
123  FLA_BR );
124 
125  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
126  C10, /**/ C11, C12,
127  /* ************** */ /* ****************** */
128  &CBL, /**/ &CBR, C20, /**/ C21, C22,
129  FLA_TR );
130 
131  }
132 
133  return FLA_SUCCESS;
134 }
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
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_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().

Referenced by FLA_Sylv_hh().