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_Trinv_lu_blk_var4(FLA_Obj A,fla_trinv_t * cntl)15 FLA_Error FLA_Trinv_lu_blk_var4( FLA_Obj A, fla_trinv_t* cntl )
16 {
17   FLA_Obj ATL,   ATR,      A00, A01, A02,
18           ABL,   ABR,      A10, A11, A12,
19                            A20, A21, A22;
20 
21   dim_t b;
22 
23   FLA_Part_2x2( A,    &ATL, &ATR,
24                       &ABL, &ABR,     0, 0, FLA_TL );
25 
26   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
27 
28     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
29 
30     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
31                         /* ************* */   /* ******************** */
32                                                 &A10, /**/ &A11, &A12,
33                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
34                            b, b, FLA_BR );
35 
36     /*------------------------------------------------------------*/
37 
38     // A21 = -trilu( A22 ) \ A21;
39     FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
40                        FLA_MINUS_ONE, A22, A21,
41                        FLA_Cntl_sub_trsm1( cntl ) );
42 
43     // A20 = -A21 * A10 + A20;
44     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
45                        FLA_MINUS_ONE, A21, A10, FLA_ONE, A20,
46                        FLA_Cntl_sub_gemm( cntl ) );
47 
48     // A10 = A10 * A00;
49     FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
50                        FLA_ONE, A00, A10,
51                        FLA_Cntl_sub_trmm( cntl ) );
52 
53     // A11 = inv( A11 );
54     FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, FLA_UNIT_DIAG, A11,
55                         FLA_Cntl_sub_trinv( cntl ) );
56 
57     /*------------------------------------------------------------*/
58 
59     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
60                                                      A10, A11, /**/ A12,
61                             /* ************** */  /* ****************** */
62                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
63                               FLA_TL );
64 
65   }
66 
67   return FLA_SUCCESS;
68 }
69 
70 #endif
71