blis/frame/3/trsm/bli_trsm_blk_var2f.c

   1 /*
   2
   3    BLIS
   4    An object-based framework for developing high-performance BLAS-like
   5    libraries.
   6
   7    Copyright (C) 2014, The University of Texas at Austin
   8
   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.
  20
  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.
  32
  33 */
  34
  35 #include "blis.h"
  36
  37 void bli_trsm_blk_var2f( obj_t*  a,
  38                          obj_t*  b,
  39                          obj_t*  c,
  40                          trsm_t* cntl,
  41                          trsm_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;
  50
  51         dim_t i;
  52         dim_t b_alg;
  53         dim_t n_trans;
  54
  55         // Initialize pack objects for A that are passed into packm_init().
  56     if( thread_am_ochief( thread ) ) {
  57             bli_obj_init_pack( &a_pack_s );
  58
  59         // Initialize object for packing A.
  60         bli_packm_init( a, &a_pack_s,
  61                         cntl_sub_packm_a( cntl ) );
  62
  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 );
  69
  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 );
  77
  78         // Pack A (if instructed).
  79         bli_packm_int( a, a_pack,
  80                        cntl_sub_packm_a( cntl ),
  81                    trmm_thread_sub_opackm( thread ) );
  82
  83         // Query dimension in partitioning direction.
  84         n_trans = bli_obj_width_after_trans( *b );
  85     dim_t start, end;
  86     num_t datatype = bli_obj_execution_datatype( *a );
  87     bli_get_range( thread, 0, n_trans,
  88                    bli_lcm( bli_info_get_default_nr( datatype ), bli_info_get_default_mr( datatype ) ),
  89                    &start, &end );
  90
  91         // Partition along the n dimension.
  92         for ( i = start; i < end; i += b_alg )
  93         {
  94                 // Determine the current algorithmic blocksize.
  95                 b_alg = bli_determine_blocksize_f( i, end, b,
  96                                                    cntl_blocksize( cntl ) );
  97
  98                 // Acquire partitions for B1 and C1.
  99                 bli_acquire_mpart_l2r( BLIS_SUBPART1,
 100                                        i, b_alg, b, &b1 );
 101                 bli_acquire_mpart_l2r( BLIS_SUBPART1,
 102                                        i, b_alg, c, &c1 );
 103
 104                 // Initialize objects for packing A1 and B1.
 105         if( thread_am_ichief( thread ) ) {
 106             bli_packm_init( &b1, b1_pack,
 107                             cntl_sub_packm_b( cntl ) );
 108             bli_packm_init( &c1, c1_pack,
 109                             cntl_sub_packm_c( cntl ) );
 110         }
 111         thread_ibarrier( thread );
 112
 113                 // Pack B1 (if instructed).
 114                 bli_packm_int( &b1, b1_pack,
 115                                cntl_sub_packm_b( cntl ),
 116                        trsm_thread_sub_ipackm( thread ) );
 117
 118                 // Pack C1 (if instructed).
 119                 bli_packm_int( &c1, c1_pack,
 120                                cntl_sub_packm_c( cntl ),
 121                        trsm_thread_sub_ipackm( thread ) );
 122
 123                 // Perform trsm subproblem.
 124                 bli_trsm_int( &BLIS_ONE,
 125                               a_pack,
 126                               b1_pack,
 127                               &BLIS_ONE,
 128                               c1_pack,
 129                               cntl_sub_trsm( cntl ),
 130                       trsm_thread_sub_trsm( thread ) );
 131
 132                 // Unpack C1 (if C1 was packed).
 133         bli_unpackm_int( c1_pack, &c1,
 134                          cntl_sub_unpackm_c( cntl ),
 135                          trsm_thread_sub_ipackm( thread ) );
 136         }
 137
 138         // If any packing buffers were acquired within packm, release them back
 139         // to the memory manager.
 140     thread_obarrier( thread );
 141     if( thread_am_ochief( thread ) )
 142         bli_obj_release_pack( a_pack, cntl_sub_packm_a(cntl) );
 143     if( thread_am_ichief( thread ) ) {
 144         bli_obj_release_pack( b1_pack, cntl_sub_packm_b(cntl) );
 145         bli_obj_release_pack( c1_pack, cntl_sub_packm_c(cntl));
 146     }
 147 }
 148