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_Trsm_luh_blk_var1(FLA_Diag diagA,FLA_Obj alpha,FLA_Obj A,FLA_Obj B,fla_trsm_t * cntl)15 FLA_Error FLA_Trsm_luh_blk_var1( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_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 BT,              B0,
22           BB,              B1,
23                            B2;
24 
25   dim_t b;
26 
27   FLA_Scal_internal( alpha, B,
28                      FLA_Cntl_sub_scal( cntl ) );
29 
30   FLA_Part_2x2( A,    &ATL, &ATR,
31                       &ABL, &ABR,     0, 0, FLA_TL );
32 
33   FLA_Part_2x1( B,    &BT,
34                       &BB,            0, FLA_TOP );
35 
36   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
37 
38     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
39 
40     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
41                         /* ************* */   /* ******************** */
42                                                 &A10, /**/ &A11, &A12,
43                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
44                            b, b, FLA_BR );
45 
46     FLA_Repart_2x1_to_3x1( BT,                &B0,
47                         /* ** */            /* ** */
48                                               &B1,
49                            BB,                &B2,        b, FLA_BOTTOM );
50 
51     /*------------------------------------------------------------*/
52 
53     /* B1 = B1 - A01' * B0 */
54     FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
55                        FLA_MINUS_ONE, A01, B0, FLA_ONE, B1,
56                        FLA_Cntl_sub_gemm( cntl ) );
57 
58     /* B1 = triu( A11' ) \ B1 */
59     FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA,
60                        FLA_ONE, A11, B1,
61                        FLA_Cntl_sub_trsm( cntl ) );
62 
63     /*------------------------------------------------------------*/
64 
65     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
66                                                      A10, A11, /**/ A12,
67                             /* ************** */  /* ****************** */
68                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
69                               FLA_TL );
70 
71     FLA_Cont_with_3x1_to_2x1( &BT,                B0,
72                                                   B1,
73                             /* ** */           /* ** */
74                               &BB,                B2,     FLA_TOP );
75 
76   }
77 
78   return FLA_SUCCESS;
79 }
80 
81 #endif
82