]> Gitweb @ Texas Instruments - Open Source Git Repositories - git.TI.com/gitweb - dense-linear-algebra-libraries/linalg.git/blob - blis/frame/3/gemm/bli_gemm_blk_var2f.c
LINALG 1.2.0 iteration 1.
[dense-linear-algebra-libraries/linalg.git] / blis / frame / 3 / gemm / bli_gemm_blk_var2f.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 void bli_gemm_blk_var2f( obj_t*  a,
38                          obj_t*  b,
39                          obj_t*  c,
40                          gemm_t* cntl,
41                          gemm_thrinfo_t* thread )
42 {
43         obj_t a_pack_s;
44     obj_t b1_pack_s, c1_pack_s;
45     
46     obj_t b1, c1;
47     obj_t*  a_pack = NULL;
48         obj_t*  b1_pack = NULL;
49         obj_t*  c1_pack = NULL;
51         dim_t i;
52         dim_t b_alg;
53         dim_t n_trans;
55 //      printf("blk var 2\n");
57     if( thread_am_ochief( thread ) ) {
58         // Initialize object for packing A
59             bli_obj_init_pack( &a_pack_s );
60         bli_packm_init( a, &a_pack_s,
61                         cntl_sub_packm_a( cntl ) );
63         // Scale C by beta (if instructed).
64         bli_scalm_int( &BLIS_ONE,
65                        c,
66                        cntl_sub_scalm( cntl ) );
67     }
68     a_pack = thread_obroadcast( thread, &a_pack_s );
70         // Initialize pack objects for B and C that are passed into packm_init().
71     if( thread_am_ichief( thread ) ) {
72         bli_obj_init_pack( &b1_pack_s );
73         bli_obj_init_pack( &c1_pack_s );
74     }
75     b1_pack = thread_ibroadcast( thread, &b1_pack_s );
76     c1_pack = thread_ibroadcast( thread, &c1_pack_s );
78         // Pack A (if instructed).
79         bli_packm_int( a, a_pack,
80                        cntl_sub_packm_a( cntl ),
81                    gemm_thread_sub_opackm( thread ) );
83         // Query dimension in partitioning direction.
84         n_trans = bli_obj_width_after_trans( *b );
85     dim_t start, end;
86 //    bli_get_range( thread, 0, n_trans,
87 //                   bli_determine_reg_blocksize( b, cntl_blocksize( cntl ) ),
88 //                   &start, &end );
90     bli_get_range_l2r( thread, 0, n_trans,
91                 bli_determine_reg_blocksize( b, cntl_blocksize( cntl ) ),
92                        &start, &end );
94         // Partition along the n dimension.
95         for ( i = start; i < end; i += b_alg )
96         {
97                 // Determine the current algorithmic blocksize.
98                 // NOTE: Use of b (for execution datatype) is intentional!
99                 // This causes the right blocksize to be used if c and a are
100                 // complex and b is real.
101                 b_alg = bli_determine_blocksize_f( i, end, b,
102                                                    cntl_blocksize( cntl ) );
104                 // Acquire partitions for B1 and C1.
105                 bli_acquire_mpart_l2r( BLIS_SUBPART1,
106                                        i, b_alg, b, &b1 );
107                 bli_acquire_mpart_l2r( BLIS_SUBPART1,
108                                        i, b_alg, c, &c1 );
110                 // Initialize objects for packing A1 and B1.
111         if( thread_am_ichief( thread ) ) {
112             bli_packm_init( &b1, b1_pack,
113                             cntl_sub_packm_b( cntl ) );
114             bli_packm_init( &c1, c1_pack,
115                             cntl_sub_packm_c( cntl ) );
116         }
117         thread_ibarrier( thread );
119                 // Pack B1 (if instructed).
120                 bli_packm_int( &b1, b1_pack,
121                                cntl_sub_packm_b( cntl ),
122                        gemm_thread_sub_ipackm( thread ) );
124                 // Pack C1 (if instructed).
125                 bli_packm_int( &c1, c1_pack,
126                                cntl_sub_packm_c( cntl ),
127                        gemm_thread_sub_ipackm( thread ) );
129                 // Perform gemm subproblem.
130                 bli_gemm_int( &BLIS_ONE,
131                               a_pack,
132                               b1_pack,
133                               &BLIS_ONE,
134                               c1_pack,
135                               cntl_sub_gemm( cntl ),
136                       gemm_thread_sub_gemm( thread ) );
138         // Unpack C1 (if C1 was packed).
139         // Currently must be done by 1 thread
140         bli_unpackm_int( c1_pack, &c1,
141                          cntl_sub_unpackm_c( cntl ),
142                          gemm_thread_sub_ipackm( thread ) );
143         }
145         // If any packing buffers were acquired within packm, release them back
146         // to the memory manager.
147     thread_obarrier( thread );
148     if( thread_am_ochief( thread ) )
149         bli_obj_release_pack( a_pack, cntl_sub_packm_a( cntl ) );
150     if( thread_am_ichief( thread ) ) {
151         bli_obj_release_pack( b1_pack, cntl_sub_packm_b( cntl ) );
152         bli_obj_release_pack( c1_pack, cntl_sub_packm_c( cntl ) );
153     }
154 //    printf("blk var 2 done\n");