libflame revision_anchor
FLA_Sylv_nn_blk_var11.c File Reference

(r)

Functions

FLA_Error FLA_Sylv_nn_blk_var11 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl)
 

Function Documentation

◆ FLA_Sylv_nn_blk_var11()

FLA_Error FLA_Sylv_nn_blk_var11 ( 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_BR );
33
34 FLA_Part_2x2( B, &BTL, &BTR,
35 &BBL, &BBR, 0, 0, FLA_TL );
36
37 FLA_Part_2x2( C, &CTL, &CTR,
38 &CBL, &CBR, 0, 0, FLA_BL );
39
40 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
41
42 b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) );
43
44 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
45 &A10, &A11, /**/ &A12,
46 /* ************* */ /* ******************** */
47 ABL, /**/ ABR, &A20, &A21, /**/ &A22,
48 b, b, FLA_TL );
49
50 FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
51 /* ************* */ /* ******************** */
52 &B10, /**/ &B11, &B12,
53 BBL, /**/ BBR, &B20, /**/ &B21, &B22,
54 b, b, FLA_BR );
55
56 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
57 &C10, /**/ &C11, &C12,
58 /* ************* */ /* ******************** */
59 CBL, /**/ CBR, &C20, /**/ &C21, &C22,
60 b, b, FLA_TR );
61
62 // Loop Invariant:
63 // CTL = CTL - ATR * sylv( ABR, BTL, CBL )
64 // CTR = CTR - ATR * sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR )
65 // CBL = sylv( ABR, BTL, CBL )
66 // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR )
67
68 /*------------------------------------------------------------*/
69
70 // C10 = sylv( A11, B00, C10 );
71 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
72 isgn, A11, B00, C10, scale,
73 FLA_Cntl_sub_sylv1( cntl ) );
74
75 // C00 = C00 - A01 * C10;
76 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
77 FLA_MINUS_ONE, A01, C10, FLA_ONE, C00,
78 FLA_Cntl_sub_gemm1( cntl ) );
79
80 // C11 = sylv( A11, B11, C11 -/+ C10 * B01 );
81 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
82 FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11,
83 FLA_Cntl_sub_gemm2( cntl ) );
84
85 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
86 isgn, A11, B11, C11, scale,
87 FLA_Cntl_sub_sylv2( cntl ) );
88
89 // C01 = C01 - A01 * C11;
90 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
91 FLA_MINUS_ONE, A01, C11, FLA_ONE, C01,
92 FLA_Cntl_sub_gemm3( cntl ) );
93
94 // C12 = sylv( A11, B22, C12 -/+ C10 * B02 -/+ C11 * B12 );
95 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
96 FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12,
97 FLA_Cntl_sub_gemm4( cntl ) );
98
99 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
100 FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12,
101 FLA_Cntl_sub_gemm5( cntl ) );
102
103 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
104 isgn, A11, B22, C12, scale,
105 FLA_Cntl_sub_sylv3( cntl ) );
106
107 // C02 = C02 - A01 * C12;
108 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
109 FLA_MINUS_ONE, A01, C12, FLA_ONE, C02,
110 FLA_Cntl_sub_gemm6( cntl ) );
111
112 /*------------------------------------------------------------*/
113
114 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
115 /* ************** */ /* ****************** */
116 A10, /**/ A11, A12,
117 &ABL, /**/ &ABR, A20, /**/ A21, A22,
118 FLA_BR );
119
120 FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
121 B10, B11, /**/ B12,
122 /* ************** */ /* ****************** */
123 &BBL, /**/ &BBR, B20, B21, /**/ B22,
124 FLA_TL );
125
126 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
127 /* ************** */ /* ****************** */
128 C10, C11, /**/ C12,
129 &CBL, /**/ &CBR, C20, C21, /**/ C22,
130 FLA_BL );
131
132 }
133
134 return FLA_SUCCESS;
135}
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_nn().