]> Gitweb @ Texas Instruments - Open Source Git Repositories - git.TI.com/gitweb - dense-linear-algebra-libraries/linalg.git/blob - blis/frame/2/hemv/bli_hemv_cntl.c
Consolidate all git repos of linalg into one.
[dense-linear-algebra-libraries/linalg.git] / blis / frame / 2 / hemv / bli_hemv_cntl.c
1 /*
3    BLIS    
4    An object-based framework for developing high-performance BLAS-like
5    libraries.
7    Copyright (C) 2014, The University of Texas at Austin
9    Redistribution and use in source and binary forms, with or without
10    modification, are permitted provided that the following conditions are
11    met:
12     - Redistributions of source code must retain the above copyright
13       notice, this list of conditions and the following disclaimer.
14     - Redistributions in binary form must reproduce the above copyright
15       notice, this list of conditions and the following disclaimer in the
16       documentation and/or other materials provided with the distribution.
17     - Neither the name of The University of Texas at Austin nor the names
18       of its contributors may be used to endorse or promote products
19       derived from this software without specific prior written permission.
21    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25    HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 */
35 #include "blis.h"
37 extern scalv_t*   scalv_cntl;
38 extern packm_t*   packm_cntl;
39 extern packv_t*   packv_cntl;
40 extern unpackv_t* unpackv_cntl;
42 extern gemv_t*    gemv_cntl_rp_bs_dot;
43 extern gemv_t*    gemv_cntl_rp_bs_axpy;
44 extern gemv_t*    gemv_cntl_cp_bs_dot;
45 extern gemv_t*    gemv_cntl_cp_bs_axpy;
47 extern blksz_t*   gemv_mc;
49 hemv_t*           hemv_cntl_bs_ke_lrow_ucol;
50 hemv_t*           hemv_cntl_bs_ke_lcol_urow;
51 hemv_t*           hemv_cntl_ge_lrow_ucol;
52 hemv_t*           hemv_cntl_ge_lcol_urow;
55 void bli_hemv_cntl_init()
56 {
57         // Create control trees for the lowest-level kernels. These trees induce
58         // operations on (presumably) relatively small block-subvector problems.
59         hemv_cntl_bs_ke_lrow_ucol
60         =
61         bli_hemv_cntl_obj_create( BLIS_UNB_FUSED,
62                                   BLIS_VARIANT1,
63                                   NULL, NULL, NULL, NULL,
64                                   NULL, NULL, NULL, NULL,
65                                   NULL, NULL, NULL );
66         hemv_cntl_bs_ke_lcol_urow
67         =
68         bli_hemv_cntl_obj_create( BLIS_UNB_FUSED,
69                                   BLIS_VARIANT3,
70                                   NULL, NULL, NULL, NULL,
71                                   NULL, NULL, NULL, NULL,
72                                   NULL, NULL, NULL );
75         // Create control trees for generally large problems. Here, we choose a
76         // variant that prioritizes keeping a subvector of y in cache.
77         hemv_cntl_ge_lrow_ucol
78         =
79         bli_hemv_cntl_obj_create( BLIS_BLOCKED,
80                                   BLIS_VARIANT2,
81                                   gemv_mc,
82                                   scalv_cntl,           // scale y up-front
83                                   packm_cntl,           // pack A11 (if needed)
84                                   packv_cntl,           // pack x1 (if needed)
85                                   packv_cntl,           // pack y1 (if needed)
86                                   gemv_cntl_rp_bs_dot,  // gemv_n_rp needed by var2
87                                   NULL,                 // gemv_n_cp not used by var2
88                                   NULL,                 // gemv_t_rp not used by var2
89                                   gemv_cntl_rp_bs_axpy, // gemv_t_cp needed by var2
90                                   hemv_cntl_bs_ke_lrow_ucol,
91                                   unpackv_cntl );       // unpack y1 (if packed)
92         hemv_cntl_ge_lcol_urow
93         =
94         bli_hemv_cntl_obj_create( BLIS_BLOCKED,
95                                   BLIS_VARIANT2,
96                                   gemv_mc,
97                                   scalv_cntl,           // scale y up-front
98                                   packm_cntl,           // pack A11 (if needed)
99                                   packv_cntl,           // pack x1 (if needed)
100                                   packv_cntl,           // pack y1 (if needed)
101                                   gemv_cntl_rp_bs_axpy, // gemv_n_rp needed by var2
102                                   NULL,                 // gemv_n_cp not used by var2
103                                   NULL,                 // gemv_t_rp not used by var2
104                                   gemv_cntl_rp_bs_dot,  // gemv_t_cp needed by var2
105                                   hemv_cntl_bs_ke_lcol_urow,
106                                   unpackv_cntl );       // unpack y1 (if packed)
109 void bli_hemv_cntl_finalize()
111         bli_cntl_obj_free( hemv_cntl_bs_ke_lrow_ucol );
112         bli_cntl_obj_free( hemv_cntl_bs_ke_lcol_urow );
113         bli_cntl_obj_free( hemv_cntl_ge_lrow_ucol );
114         bli_cntl_obj_free( hemv_cntl_ge_lcol_urow );
118 hemv_t* bli_hemv_cntl_obj_create( impl_t     impl_type,
119                                   varnum_t   var_num,
120                                   blksz_t*   b,
121                                   scalv_t*   sub_scalv,
122                                   packm_t*   sub_packm_a11,
123                                   packv_t*   sub_packv_x1,
124                                   packv_t*   sub_packv_y1,
125                                   gemv_t*    sub_gemv_n_rp,
126                                   gemv_t*    sub_gemv_n_cp,
127                                   gemv_t*    sub_gemv_t_rp,
128                                   gemv_t*    sub_gemv_t_cp,
129                                   hemv_t*    sub_hemv,
130                                   unpackv_t* sub_unpackv_y1 )
132         hemv_t* cntl;
134         cntl = ( hemv_t* ) bli_malloc( sizeof(hemv_t) );        
136         cntl->impl_type      = impl_type;
137         cntl->var_num        = var_num;
138         cntl->b              = b;
139         cntl->sub_scalv      = sub_scalv;
140         cntl->sub_packm_a11  = sub_packm_a11;
141         cntl->sub_packv_x1   = sub_packv_x1;
142         cntl->sub_packv_y1   = sub_packv_y1;
143         cntl->sub_gemv_n_rp  = sub_gemv_n_rp;
144         cntl->sub_gemv_n_cp  = sub_gemv_n_cp;
145         cntl->sub_gemv_t_rp  = sub_gemv_t_rp;
146         cntl->sub_gemv_t_cp  = sub_gemv_t_cp;
147         cntl->sub_hemv       = sub_hemv;
148         cntl->sub_unpackv_y1 = sub_unpackv_y1;
150         return cntl;
153 void bli_hemv_cntl_obj_init( hemv_t*    cntl,
154                              impl_t     impl_type,
155                              varnum_t   var_num,
156                              blksz_t*   b,
157                              scalv_t*   sub_scalv,
158                              packm_t*   sub_packm_a11,
159                              packv_t*   sub_packv_x1,
160                              packv_t*   sub_packv_y1,
161                              gemv_t*    sub_gemv_n_rp,
162                              gemv_t*    sub_gemv_n_cp,
163                              gemv_t*    sub_gemv_t_rp,
164                              gemv_t*    sub_gemv_t_cp,
165                              hemv_t*    sub_hemv,
166                              unpackv_t* sub_unpackv_y1 )
168         cntl->impl_type      = impl_type;
169         cntl->var_num        = var_num;
170         cntl->b              = b;
171         cntl->sub_scalv      = sub_scalv;
172         cntl->sub_packm_a11  = sub_packm_a11;
173         cntl->sub_packv_x1   = sub_packv_x1;
174         cntl->sub_packv_y1   = sub_packv_y1;
175         cntl->sub_gemv_n_rp  = sub_gemv_n_rp;
176         cntl->sub_gemv_n_cp  = sub_gemv_n_cp;
177         cntl->sub_gemv_t_rp  = sub_gemv_t_rp;
178         cntl->sub_gemv_t_cp  = sub_gemv_t_cp;
179         cntl->sub_hemv       = sub_hemv;
180         cntl->sub_unpackv_y1 = sub_unpackv_y1;