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 "gimple-iterator.h"
  40 #include "tree-ssa-loop.h"
  41 #include "dumpfile.h"
  42 #include "cfgloop.h"
  43 #include "tree-chrec.h"
  44 #include "tree-data-ref.h"
  45 #include "sese.h"
  46
  47 #ifdef HAVE_cloog
  48 #include "graphite-poly.h"
  49
  50
  51 /* Strip mines with a factor STRIDE the scattering (time) dimension
  52    around PBB at depth TIME_DEPTH.
  53
  54    The following example comes from the wiki page:
  55    http://gcc.gnu.org/wiki/Graphite/Strip_mine
  56
  57    The strip mine of a loop with a tile of 64 can be obtained with a
  58    scattering function as follows:
  59
  60    $ cat ./albert_strip_mine.cloog
  61    # language: C
  62    c
  63
  64    # parameter {n | n >= 0}
  65    1 3
  66    #  n  1
  67    1  1  0
  68    1
  69    n
  70
  71    1 # Number of statements:
  72
  73    1
  74    # {i | 0 <= i <= n}
  75    2 4
  76    #  i  n   1
  77    1  1  0   0
  78    1 -1  1   0
  79
  80    0  0  0
  81    1
  82    i
  83
  84    1 # Scattering functions
  85
  86    3 6
  87    #  NEW  OLD    i    n    1
  88    1  -64    0    1    0    0
  89    1   64    0   -1    0   63
  90    0    0    1   -1    0    0
  91
  92    1
  93    NEW  OLD
  94
  95    #the output of CLooG is like this:
  96    #$ cloog ./albert_strip_mine.cloog
  97    # for (NEW=0;NEW<=floord(n,64);NEW++) {
  98    #   for (OLD=max(64*NEW,0);OLD<=min(64*NEW+63,n);OLD++) {
  99    #     S1(i = OLD) ;
 100    #   }
 101    # }
 102 */
 103
 104 static void
 105 pbb_strip_mine_time_depth (poly_bb_p pbb, int time_depth, int stride)
 106 {
 107   isl_space *d;
 108   isl_constraint *c;
 109   int iter, strip;
 110   /* STRIP is the dimension that iterates with stride STRIDE.  */
 111   /* ITER is the dimension that enumerates single iterations inside
 112      one strip that has at most STRIDE iterations.  */
 113   strip = time_depth;
 114   iter = strip + 2;
 115
 116   pbb->transformed = isl_map_insert_dims (pbb->transformed, isl_dim_out,
 117                                           strip, 2);
 118
 119   /* Lower bound of the striped loop.  */
 120   d = isl_map_get_space (pbb->transformed);
 121   c = isl_inequality_alloc (isl_local_space_from_space (d));
 122   c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip, -stride);
 123   c = isl_constraint_set_coefficient_si (c, isl_dim_out, iter, 1);
 124   pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
 125
 126   /* Upper bound of the striped loop.  */
 127   d = isl_map_get_space (pbb->transformed);
 128   c = isl_inequality_alloc (isl_local_space_from_space (d));
 129   c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip, stride);
 130   c = isl_constraint_set_coefficient_si (c, isl_dim_out, iter, -1);
 131   c = isl_constraint_set_constant_si (c, stride - 1);
 132   pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
 133
 134   /* Static scheduling for ITER level.
 135      This is mandatory to keep the 2d + 1 canonical scheduling format.  */
 136   d = isl_map_get_space (pbb->transformed);
 137   c = isl_equality_alloc (isl_local_space_from_space (d));
 138   c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip + 1, 1);
 139   pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
 140 }
 141
 142 /* Returns true when strip mining with STRIDE of the loop LST is
 143    profitable.  */
 144
 145 static bool
 146 lst_strip_mine_profitable_p (lst_p lst, int stride)
 147 {
 148   mpz_t niter, strip_stride;
 149   bool res;
 150
 151   gcc_assert (LST_LOOP_P (lst));
 152   mpz_init (strip_stride);
 153   mpz_init (niter);
 154
 155   mpz_set_si (strip_stride, stride);
 156   lst_niter_for_loop (lst, niter);
 157   res = (mpz_cmp (niter, strip_stride) > 0);
 158
 159   mpz_clear (strip_stride);
 160   mpz_clear (niter);
 161   return res;
 162 }
 163
 164 /* Strip-mines all the loops of LST with STRIDE.  Return the number of
 165    loops strip-mined.  */
 166
 167 static int
 168 lst_do_strip_mine_loop (lst_p lst, int depth, int stride)
 169 {
 170   int i;
 171   lst_p l;
 172   poly_bb_p pbb;
 173
 174   if (!lst)
 175     return 0;
 176
 177   if (LST_LOOP_P (lst))
 178     {
 179       int res = 0;
 180
 181       FOR_EACH_VEC_ELT (LST_SEQ (lst), i, l)
 182         res += lst_do_strip_mine_loop (l, depth, stride);
 183
 184       return res;
 185     }
 186
 187   pbb = LST_PBB (lst);
 188   pbb_strip_mine_time_depth (pbb, psct_dynamic_dim (pbb, depth), stride);
 189   return 1;
 190 }
 191
 192 /* Strip-mines all the loops of LST with STRIDE.  When STRIDE is zero,
 193    read the stride from the PARAM_LOOP_BLOCK_TILE_SIZE.  Return the
 194    number of strip-mined loops.
 195
 196    Strip mining transforms a loop
 197
 198    | for (i = 0; i < N; i++)
 199    |   S (i);
 200
 201    into the following loop nest:
 202
 203    | for (k = 0; k < N; k += STRIDE)
 204    |   for (j = 0; j < STRIDE; j++)
 205    |     S (i = k + j);
 206 */
 207
 208 static int
 209 lst_do_strip_mine (lst_p lst, int stride)
 210 {
 211   int i;
 212   lst_p l;
 213   int res = 0;
 214   int depth;
 215
 216   if (!stride)
 217     stride = PARAM_VALUE (PARAM_LOOP_BLOCK_TILE_SIZE);
 218
 219   if (!lst
 220       || !LST_LOOP_P (lst))
 221     return false;
 222
 223   FOR_EACH_VEC_ELT (LST_SEQ (lst), i, l)
 224     res += lst_do_strip_mine (l, stride);
 225
 226   depth = lst_depth (lst);
 227   if (depth >= 0
 228       && lst_strip_mine_profitable_p (lst, stride))
 229     {
 230       res += lst_do_strip_mine_loop (lst, lst_depth (lst), stride);
 231       lst_add_loop_under_loop (lst);
 232     }
 233
 234   return res;
 235 }
 236
 237 /* Strip mines all the loops in SCOP.  Returns the number of
 238    strip-mined loops.  */
 239
 240 int
 241 scop_do_strip_mine (scop_p scop, int stride)
 242 {
 243   return lst_do_strip_mine (SCOP_TRANSFORMED_SCHEDULE (scop), stride);
 244 }
 245
 246 /* Loop blocks all the loops in SCOP.  Returns true when we manage to
 247    block some loops.  */
 248
 249 bool
 250 scop_do_block (scop_p scop)
 251 {
 252   store_scattering (scop);
 253
 254   /* If we don't strip mine at least two loops, or not interchange
 255      loops, the strip mine alone will not be profitable, and the
 256      transform is not a loop blocking: so revert the transform.  */
 257   if (lst_do_strip_mine (SCOP_TRANSFORMED_SCHEDULE (scop), 0) < 2
 258       || scop_do_interchange (scop) == 0)
 259     {
 260       restore_scattering (scop);
 261       return false;
 262     }
 263
 264   if (dump_file && (dump_flags & TDF_DETAILS))
 265     fprintf (dump_file, "SCoP will be loop blocked.\n");
 266
 267   return true;
 268 }
 269
 270 #endif