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 extern fla_herk_t* fla_herk_cntl_blas;
14 extern fla_trsm_t* fla_trsm_cntl_blas;
15
16 fla_chol_t* fla_chol_cntl = NULL;
17 fla_chol_t* fla_chol_cntl2 = NULL;
18
19 fla_chol_t* fla_chol_cntl_in = NULL;
20 fla_chol_t* fla_chol_cntl_leaf = NULL;
21 fla_blocksize_t* fla_chol_var3_bsize = NULL;
22 fla_blocksize_t* fla_chol_var3_bsize_in = NULL;
23 double fla_chol_var3_in_to_ou_bsize_ratio = 0.25;
24
FLA_Chol_cntl_init()25 void FLA_Chol_cntl_init()
26 {
27 // Set blocksize with default values for conventional storage.
28 fla_chol_var3_bsize = FLA_Query_blocksizes( FLA_DIMENSION_MIN );
29 fla_chol_var3_bsize_in = FLA_Blocksize_create_copy( fla_chol_var3_bsize );
30 FLA_Blocksize_scale( fla_chol_var3_bsize_in, fla_chol_var3_in_to_ou_bsize_ratio );
31
32 // Create a control tree to invoke LAPACK.
33 fla_chol_cntl_leaf = FLA_Cntl_chol_obj_create( FLA_FLAT,
34 #ifdef FLA_ENABLE_EXTERNAL_LAPACK_FOR_SUBPROBLEMS
35 FLA_BLOCKED_EXTERN,
36 #else
37 FLA_UNB_OPT_VARIANT2,
38 #endif
39 NULL,
40 NULL,
41 NULL,
42 NULL,
43 NULL );
44
45 // Create a control tree for small subproblems.
46 fla_chol_cntl_in = FLA_Cntl_chol_obj_create( FLA_FLAT,
47 FLA_BLOCKED_VARIANT3,
48 fla_chol_var3_bsize_in,
49 fla_chol_cntl_leaf,
50 fla_herk_cntl_blas,
51 fla_trsm_cntl_blas,
52 NULL );
53
54 // Create a control tree for larger problems with one level of recursion.
55 fla_chol_cntl2 = FLA_Cntl_chol_obj_create( FLA_FLAT,
56 FLA_BLOCKED_VARIANT3,
57 fla_chol_var3_bsize,
58 fla_chol_cntl_in,
59 fla_herk_cntl_blas,
60 fla_trsm_cntl_blas,
61 NULL );
62
63 // Create a control tree for large problems with no extra recursion.
64 fla_chol_cntl = FLA_Cntl_chol_obj_create( FLA_FLAT,
65 FLA_BLOCKED_VARIANT3,
66 fla_chol_var3_bsize,
67 fla_chol_cntl_leaf,
68 fla_herk_cntl_blas,
69 fla_trsm_cntl_blas,
70 NULL );
71 }
72
FLA_Chol_cntl_finalize()73 void FLA_Chol_cntl_finalize()
74 {
75 FLA_Cntl_obj_free( fla_chol_cntl );
76 FLA_Cntl_obj_free( fla_chol_cntl2 );
77 FLA_Cntl_obj_free( fla_chol_cntl_leaf );
78 FLA_Cntl_obj_free( fla_chol_cntl_in );
79
80 FLA_Blocksize_free( fla_chol_var3_bsize );
81 FLA_Blocksize_free( fla_chol_var3_bsize_in );
82 }
83
84