gcc/graphite-blocking.c

   1 /* Heuristics and transform for loop blocking and strip mining on
   2    polyhedral representation.
   3
   4    Copyright (C) 2009-2013 Free Software Foundation, Inc.
   5    Contributed by Sebastian Pop <sebastian.pop@amd.com> and
   6    Pranav Garg  <pranav.garg2107@gmail.com>.
   7
   8 This file is part of GCC.
   9
  10 GCC is free software; you can redistribute it and/or modify
  11 it under the terms of the GNU General Public License as published by
  12 the Free Software Foundation; either version 3, or (at your option)
  13 any later version.
  14
  15 GCC is distributed in the hope that it will be useful,
  16 but WITHOUT ANY WARRANTY; without even the implied warranty of
  17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18 GNU General Public License for more details.
  19
  20 You should have received a copy of the GNU General Public License
  21 along with GCC; see the file COPYING3.  If not see
  22 <http://www.gnu.org/licenses/>.  */
  23
  24 #include "config.h"
  25
  26 #ifdef HAVE_cloog
  27 #include <isl/set.h>
  28 #include <isl/map.h>
  29 #include <isl/union_map.h>
  30 #include <isl/constraint.h>
  31 #include <cloog/cloog.h>
  32 #include <cloog/isl/domain.h>
  33 #endif
  34
  35 #include "system.h"
  36 #include "coretypes.h"
  37 #include "tree.h"
  38 #include "gimple.h"
  39 #include "tree-ssa-loop.h"
  40 #include "dumpfile.h"
  41 #include "cfgloop.h"
  42 #include "tree-chrec.h"
  43 #include "tree-data-ref.h"
  44 #include "sese.h"
  45
  46 #ifdef HAVE_cloog
  47 #include "graphite-poly.h"
  48
  49
  50 /* Strip mines with a factor STRIDE the scattering (time) dimension
  51    around PBB at depth TIME_DEPTH.
  52
  53    The following example comes from the wiki page:
  54    http://gcc.gnu.org/wiki/Graphite/Strip_mine
  55
  56    The strip mine of a loop with a tile of 64 can be obtained with a
  57    scattering function as follows:
  58
  59    $ cat ./albert_strip_mine.cloog
  60    # language: C
  61    c
  62
  63    # parameter {n | n >= 0}
  64    1 3
  65    #  n  1
  66    1  1  0
  67    1
  68    n
  69
  70    1 # Number of statements:
  71
  72    1
  73    # {i | 0 <= i <= n}
  74    2 4
  75    #  i  n   1
  76    1  1  0   0
  77    1 -1  1   0
  78
  79    0  0  0
  80    1
  81    i
  82
  83    1 # Scattering functions
  84
  85    3 6
  86    #  NEW  OLD    i    n    1
  87    1  -64    0    1    0    0
  88    1   64    0   -1    0   63
  89    0    0    1   -1    0    0
  90
  91    1
  92    NEW  OLD
  93
  94    #the output of CLooG is like this:
  95    #$ cloog ./albert_strip_mine.cloog
  96    # for (NEW=0;NEW<=floord(n,64);NEW++) {
  97    #   for (OLD=max(64*NEW,0);OLD<=min(64*NEW+63,n);OLD++) {
  98    #     S1(i = OLD) ;
  99    #   }
 100    # }
 101 */
 102
 103 static void
 104 pbb_strip_mine_time_depth (poly_bb_p pbb, int time_depth, int stride)
 105 {
 106   isl_space *d;
 107   isl_constraint *c;
 108   int iter, strip;
 109   /* STRIP is the dimension that iterates with stride STRIDE.  */
 110   /* ITER is the dimension that enumerates single iterations inside
 111      one strip that has at most STRIDE iterations.  */
 112   strip = time_depth;
 113   iter = strip + 2;
 114
 115   pbb->transformed = isl_map_insert_dims (pbb->transformed, isl_dim_out,
 116                                           strip, 2);
 117
 118   /* Lower bound of the striped loop.  */
 119   d = isl_map_get_space (pbb->transformed);
 120   c = isl_inequality_alloc (isl_local_space_from_space (d));
 121   c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip, -stride);
 122   c = isl_constraint_set_coefficient_si (c, isl_dim_out, iter, 1);
 123   pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
 124
 125   /* Upper bound of the striped loop.  */
 126   d = isl_map_get_space (pbb->transformed);
 127   c = isl_inequality_alloc (isl_local_space_from_space (d));
 128   c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip, stride);
 129   c = isl_constraint_set_coefficient_si (c, isl_dim_out, iter, -1);
 130   c = isl_constraint_set_constant_si (c, stride - 1);
 131   pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
 132
 133   /* Static scheduling for ITER level.
 134      This is mandatory to keep the 2d + 1 canonical scheduling format.  */
 135   d = isl_map_get_space (pbb->transformed);
 136   c = isl_equality_alloc (isl_local_space_from_space (d));
 137   c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip + 1, 1);
 138   pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
 139 }
 140
 141 /* Returns true when strip mining with STRIDE of the loop LST is
 142    profitable.  */
 143
 144 static bool
 145 lst_strip_mine_profitable_p (lst_p lst, int stride)
 146 {
 147   mpz_t niter, strip_stride;
 148   bool res;
 149
 150   gcc_assert (LST_LOOP_P (lst));
 151   mpz_init (strip_stride);
 152   mpz_init (niter);
 153
 154   mpz_set_si (strip_stride, stride);
 155   lst_niter_for_loop (lst, niter);
 156   res = (mpz_cmp (niter, strip_stride) > 0);
 157
 158   mpz_clear (strip_stride);
 159   mpz_clear (niter);
 160   return res;
 161 }
 162
 163 /* Strip-mines all the loops of LST with STRIDE.  Return the number of
 164    loops strip-mined.  */
 165
 166 static int
 167 lst_do_strip_mine_loop (lst_p lst, int depth, int stride)
 168 {
 169   int i;
 170   lst_p l;
 171   poly_bb_p pbb;
 172
 173   if (!lst)
 174     return 0;
 175
 176   if (LST_LOOP_P (lst))
 177     {
 178       int res = 0;
 179
 180       FOR_EACH_VEC_ELT (LST_SEQ (lst), i, l)
 181         res += lst_do_strip_mine_loop (l, depth, stride);
 182
 183       return res;
 184     }
 185
 186   pbb = LST_PBB (lst);
 187   pbb_strip_mine_time_depth (pbb, psct_dynamic_dim (pbb, depth), stride);
 188   return 1;
 189 }
 190
 191 /* Strip-mines all the loops of LST with STRIDE.  When STRIDE is zero,
 192    read the stride from the PARAM_LOOP_BLOCK_TILE_SIZE.  Return the
 193    number of strip-mined loops.
 194
 195    Strip mining transforms a loop
 196
 197    | for (i = 0; i < N; i++)
 198    |   S (i);
 199
 200    into the following loop nest:
 201
 202    | for (k = 0; k < N; k += STRIDE)
 203    |   for (j = 0; j < STRIDE; j++)
 204    |     S (i = k + j);
 205 */
 206
 207 static int
 208 lst_do_strip_mine (lst_p lst, int stride)
 209 {
 210   int i;
 211   lst_p l;
 212   int res = 0;
 213   int depth;
 214
 215   if (!stride)
 216     stride = PARAM_VALUE (PARAM_LOOP_BLOCK_TILE_SIZE);
 217
 218   if (!lst
 219       || !LST_LOOP_P (lst))
 220     return false;
 221
 222   FOR_EACH_VEC_ELT (LST_SEQ (lst), i, l)
 223     res += lst_do_strip_mine (l, stride);
 224
 225   depth = lst_depth (lst);
 226   if (depth >= 0
 227       && lst_strip_mine_profitable_p (lst, stride))
 228     {
 229       res += lst_do_strip_mine_loop (lst, lst_depth (lst), stride);
 230       lst_add_loop_under_loop (lst);
 231     }
 232
 233   return res;
 234 }
 235
 236 /* Strip mines all the loops in SCOP.  Returns the number of
 237    strip-mined loops.  */
 238
 239 int
 240 scop_do_strip_mine (scop_p scop, int stride)
 241 {
 242   return lst_do_strip_mine (SCOP_TRANSFORMED_SCHEDULE (scop), stride);
 243 }
 244
 245 /* Loop blocks all the loops in SCOP.  Returns true when we manage to
 246    block some loops.  */
 247
 248 bool
 249 scop_do_block (scop_p scop)
 250 {
 251   store_scattering (scop);
 252
 253   /* If we don't strip mine at least two loops, or not interchange
 254      loops, the strip mine alone will not be profitable, and the
 255      transform is not a loop blocking: so revert the transform.  */
 256   if (lst_do_strip_mine (SCOP_TRANSFORMED_SCHEDULE (scop), 0) < 2
 257       || scop_do_interchange (scop) == 0)
 258     {
 259       restore_scattering (scop);
 260       return false;
 261     }
 262
 263   if (dump_file && (dump_flags & TDF_DETAILS))
 264     fprintf (dump_file, "SCoP will be loop blocked.\n");
 265
 266   return true;
 267 }
 268
 269 #endif