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 
FLA_Trsm_luh_blk_var2(FLA_Diag diagA,FLA_Obj alpha,FLA_Obj A,FLA_Obj B,fla_trsm_t * cntl)13 FLA_Error FLA_Trsm_luh_blk_var2( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_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 BT,              B0,
20           BB,              B1,
21                            B2;
22 
23   dim_t b;
24 
25   FLA_Scal_internal( alpha, B,
26                      FLA_Cntl_sub_scal( cntl ) );
27 
28   FLA_Part_2x2( A,    &ATL, &ATR,
29                       &ABL, &ABR,     0, 0, FLA_TL );
30 
31   FLA_Part_2x1( B,    &BT,
32                       &BB,            0, FLA_TOP );
33 
34   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
35 
36     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
37 
38     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
39                         /* ************* */   /* ******************** */
40                                                 &A10, /**/ &A11, &A12,
41                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
42                            b, b, FLA_BR );
43 
44     FLA_Repart_2x1_to_3x1( BT,                &B0,
45                         /* ** */            /* ** */
46                                               &B1,
47                            BB,                &B2,        b, FLA_BOTTOM );
48 
49     /*------------------------------------------------------------*/
50 
51     /* B1 = triu( A11' ) \ B1 */
52     FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA,
53                        FLA_ONE, A11, B1,
54                        FLA_Cntl_sub_trsm( cntl ) );
55 
56     /* B2 = B2 - A12' * B1 */
57     FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
58                        FLA_MINUS_ONE, A12, B1, FLA_ONE, B2,
59                        FLA_Cntl_sub_gemm( 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( &BT,                B0,
70                                                   B1,
71                             /* ** */           /* ** */
72                               &BB,                B2,     FLA_TOP );
73 
74   }
75 
76   return FLA_SUCCESS;
77 }
78 
79