1 /*
2
3 Copyright (C) 2014, The University of Texas at Austin
4
5 This file is part of libflame and is available under the 3-Clause
6 BSD license, which can be found in the LICENSE file at the top-level
7 directory, or at http://opensource.org/licenses/BSD-3-Clause
8
9 */
10
11 #include "FLAME.h"
12
13 #ifdef FLA_ENABLE_NON_CRITICAL_CODE
14
FLA_Hemm_rl_unb_var2(FLA_Obj alpha,FLA_Obj A,FLA_Obj B,FLA_Obj beta,FLA_Obj C)15 FLA_Error FLA_Hemm_rl_unb_var2( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
16 {
17 FLA_Obj ATL, ATR, A00, a01, A02,
18 ABL, ABR, a10t, alpha11, a12t,
19 A20, a21, A22;
20
21 FLA_Obj BL, BR, B0, b1t, B2;
22
23 FLA_Obj CL, CR, C0, c1t, C2;
24
25 FLA_Scal_external( beta, C );
26
27 FLA_Part_2x2( A, &ATL, &ATR,
28 &ABL, &ABR, 0, 0, FLA_TL );
29
30 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31
32 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
33
34 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
35
36 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
37 /* ************* */ /* ************************** */
38 &a10t, /**/ &alpha11, &a12t,
39 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
40 1, 1, FLA_BR );
41
42 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
43 1, FLA_RIGHT );
44
45 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
46 1, FLA_RIGHT );
47
48 /*------------------------------------------------------------*/
49
50 /* c1t = c1t + B0 * a10t' */
51 FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, B0, a10t, FLA_ONE, c1t );
52
53 /* c1t = c1t + b1t * alpha11 */
54 FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
55
56 /* c1t = c1t + B2 * a21 */
57 FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B2, a21, FLA_ONE, c1t );
58
59 /*------------------------------------------------------------*/
60
61 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
62 a10t, alpha11, /**/ a12t,
63 /* ************** */ /* ************************ */
64 &ABL, /**/ &ABR, A20, a21, /**/ A22,
65 FLA_TL );
66
67 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
68 FLA_LEFT );
69
70 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
71 FLA_LEFT );
72
73 }
74
75 return FLA_SUCCESS;
76 }
77
78 #endif
79