libflame revision_anchor
Functions
FLA_Sylv_nh_blk_var9.c File Reference

(r)

Functions

FLA_Error FLA_Sylv_nh_blk_var9 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl)
 

Function Documentation

◆ FLA_Sylv_nh_blk_var9()

FLA_Error FLA_Sylv_nh_blk_var9 ( 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_BR );
36
37 FLA_Part_2x2( C, &CTL, &CTR,
38 &CBL, &CBR, 0, 0, FLA_BR );
39
40 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
41
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 &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 &C10, &C11, /**/ &C12,
58 /* ************* */ /* ******************** */
59 CBL, /**/ CBR, &C20, &C21, /**/ &C22,
60 b, b, FLA_TL );
61
62 // Loop Invariant:
63 // CTL =
64 // CTR =
65 // CBL =
66 // CBR =
67
68 /*------------------------------------------------------------*/
69
70 // C11 = sylv( A11, B11', C11 - A12 * C21 -/+ C12 * B12' );
73 FLA_Cntl_sub_gemm1( cntl ) );
74
77 FLA_Cntl_sub_gemm2( cntl ) );
78
80 isgn, A11, B11, C11, scale,
81 FLA_Cntl_sub_sylv1( cntl ) );
82
83 // C01 = sylv( A00, B11', C01 - A01 * C11 - A02 * C21 -/+ C02 * B12' );
86 FLA_Cntl_sub_gemm3( cntl ) );
87
90 FLA_Cntl_sub_gemm4( cntl ) );
91
94 FLA_Cntl_sub_gemm5( cntl ) );
95
97 isgn, A00, B11, C01, scale,
98 FLA_Cntl_sub_sylv2( cntl ) );
99
100 // C10 = sylv( A11, B00', C10 - A12 * C20 -/+ C12 * B02' -/+ C11 * B01' );
103 FLA_Cntl_sub_gemm6( cntl ) );
104
107 FLA_Cntl_sub_gemm7( cntl ) );
108
111 FLA_Cntl_sub_gemm8( cntl ) );
112
114 isgn, A11, B00, C10, scale,
115 FLA_Cntl_sub_sylv3( cntl ) );
116
117 /*------------------------------------------------------------*/
118
119 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
120 /* ************** */ /* ****************** */
121 A10, /**/ A11, A12,
122 &ABL, /**/ &ABR, A20, /**/ A21, A22,
123 FLA_BR );
124
125 FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, /**/ B01, B02,
126 /* ************** */ /* ****************** */
127 B10, /**/ B11, B12,
128 &BBL, /**/ &BBR, B20, /**/ B21, B22,
129 FLA_BR );
130
131 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
132 /* ************** */ /* ****************** */
133 C10, /**/ C11, C12,
134 &CBL, /**/ &CBR, C20, /**/ C21, C22,
135 FLA_BR );
136
137 }
138
139 return FLA_SUCCESS;
140}
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
int i
Definition bl1_axmyv2.c:145
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(), FLA_Sylv_internal(), and i.

Referenced by FLA_Sylv_nh().