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_Trmm_rlc_blk_var2(FLA_Diag diagA,FLA_Obj alpha,FLA_Obj A,FLA_Obj B,fla_trmm_t * cntl)13 FLA_Error FLA_Trmm_rlc_blk_var2( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_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 BL,    BR,       B0,  B1,  B2;
20 
21   dim_t b;
22 
23   FLA_Scal_internal( alpha, B,
24                      FLA_Cntl_sub_scal( cntl ) );
25 
26   FLA_Part_2x2( A,    &ATL, &ATR,
27                       &ABL, &ABR,     0, 0, FLA_TL );
28 
29   FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_LEFT );
30 
31   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
32 
33     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
34 
35     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
36                         /* ************* */   /* ******************** */
37                                                 &A10, /**/ &A11, &A12,
38                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
39                            b, b, FLA_BR );
40 
41     FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, /**/ &B1, &B2,
42                            b, FLA_RIGHT );
43 
44     /*------------------------------------------------------------*/
45 
46     /* B0 = B0 + B1 * A10; */
47     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE,
48                        FLA_ONE, B1, A10, FLA_ONE, B0,
49                        FLA_Cntl_sub_gemm( cntl ) );
50 
51     /* B1 = B1 * tril( A11 ); */
52     FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, diagA,
53                        FLA_ONE, A11, B1,
54                        FLA_Cntl_sub_trmm( cntl ) );
55 
56     /*------------------------------------------------------------*/
57 
58     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
59                                                      A10, A11, /**/ A12,
60                             /* ************** */  /* ****************** */
61                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
62                               FLA_TL );
63 
64     FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, B1, /**/ B2,
65                               FLA_LEFT );
66 
67   }
68 
69   return FLA_SUCCESS;
70 }
71 
72