ppcg.c

   1 /*
   2  * Copyright 2011      INRIA Saclay
   3  *
   4  * Use of this software is governed by the GNU LGPLv2.1 license
   5  *
   6  * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
   7  * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
   8  * 91893 Orsay, France
   9  */
  10
  11 #include <assert.h>
  12 #include <stdio.h>
  13 #include <isl/ctx.h>
  14 #include <isl/flow.h>
  15 #include <isl/options.h>
  16 #include <isl/schedule.h>
  17 #include <pet.h>
  18 #include "ppcg.h"
  19 #include "ppcg_options.h"
  20 #include "cuda.h"
  21 #include "cpu.h"
  22
  23 struct options {
  24         struct isl_options *isl;
  25         struct pet_options *pet;
  26         struct ppcg_options *ppcg;
  27         char *input;
  28 };
  29
  30 const char *ppcg_version(void);
  31 static void print_version(void)
  32 {
  33         printf("%s", ppcg_version());
  34 }
  35
  36 ISL_ARGS_START(struct options, options_args)
  37 ISL_ARG_CHILD(struct options, isl, "isl", &isl_options_args, "isl options")
  38 ISL_ARG_CHILD(struct options, pet, "pet", &pet_options_args, "pet options")
  39 ISL_ARG_CHILD(struct options, ppcg, NULL, &ppcg_options_args, "ppcg options")
  40 ISL_ARG_ARG(struct options, input, "input", NULL)
  41 ISL_ARG_VERSION(print_version)
  42 ISL_ARGS_END
  43
  44 ISL_ARG_DEF(options, struct options, options_args)
  45
  46 /* Is "stmt" a kill statement?
  47  */
  48 static int is_kill(struct pet_stmt *stmt)
  49 {
  50         if (stmt->body->type != pet_expr_unary)
  51                 return 0;
  52         return stmt->body->op == pet_op_kill;
  53 }
  54
  55 /* Is "stmt" not a kill statement?
  56  */
  57 static int is_not_kill(struct pet_stmt *stmt)
  58 {
  59         return !is_kill(stmt);
  60 }
  61
  62 /* Collect the iteration domains of the statements in "scop" that
  63  * satisfy "pred".
  64  */
  65 static __isl_give isl_union_set *collect_domains(struct pet_scop *scop,
  66         int (*pred)(struct pet_stmt *stmt))
  67 {
  68         int i;
  69         isl_set *domain_i;
  70         isl_union_set *domain;
  71
  72         if (!scop)
  73                 return NULL;
  74
  75         domain = isl_union_set_empty(isl_set_get_space(scop->context));
  76
  77         for (i = 0; i < scop->n_stmt; ++i) {
  78                 struct pet_stmt *stmt = scop->stmts[i];
  79
  80                 if (!pred(stmt))
  81                         continue;
  82                 domain_i = isl_set_copy(scop->stmts[i]->domain);
  83                 domain = isl_union_set_add_set(domain, domain_i);
  84         }
  85
  86         return domain;
  87 }
  88
  89 /* Collect the iteration domains of the statements in "scop",
  90  * skipping kill statements.
  91  */
  92 static __isl_give isl_union_set *collect_non_kill_domains(struct pet_scop *scop)
  93 {
  94         return collect_domains(scop, &is_not_kill);
  95 }
  96
  97 /* Does "expr" contain any call expressions?
  98  */
  99 static int expr_has_call(struct pet_expr *expr)
 100 {
 101         int i;
 102
 103         if (expr->type == pet_expr_call)
 104                 return 1;
 105
 106         for (i = 0; i < expr->n_arg; ++i)
 107                 if (expr_has_call(expr->args[i]))
 108                         return 1;
 109
 110         return 0;
 111 }
 112
 113 /* Does "stmt" contain any call expressions?
 114  */
 115 static int has_call(struct pet_stmt *stmt)
 116 {
 117         return expr_has_call(stmt->body);
 118 }
 119
 120 /* Collect the iteration domains of the statements in "scop"
 121  * that contain a call expression.
 122  */
 123 static __isl_give isl_union_set *collect_call_domains(struct pet_scop *scop)
 124 {
 125         return collect_domains(scop, &has_call);
 126 }
 127
 128 /* Collect all kill accesses in "scop".
 129  */
 130 static __isl_give isl_union_map *collect_kills(struct pet_scop *scop)
 131 {
 132         int i;
 133         isl_union_map *kills;
 134
 135         if (!scop)
 136                 return NULL;
 137
 138         kills = isl_union_map_empty(isl_set_get_space(scop->context));
 139
 140         for (i = 0; i < scop->n_stmt; ++i) {
 141                 struct pet_stmt *stmt = scop->stmts[i];
 142                 isl_map *kill_i;
 143
 144                 if (!is_kill(stmt))
 145                         continue;
 146                 kill_i = isl_map_copy(stmt->body->args[0]->acc.access);
 147                 kill_i = isl_map_intersect_domain(kill_i,
 148                                                     isl_set_copy(stmt->domain));
 149                 kills = isl_union_map_add_map(kills, kill_i);
 150         }
 151
 152         return kills;
 153 }
 154
 155 /* Compute (flow) dependences and store the resulting flow dependences
 156  * in scop->dep_flow and the reads with no corresponding writes in
 157  * scop->live_in.
 158  */
 159 static void compute_dependences(struct ppcg_scop *scop)
 160 {
 161         isl_union_map *empty;
 162
 163         if (!scop)
 164                 return;
 165
 166         empty = isl_union_map_empty(isl_union_set_get_space(scop->domain));
 167         isl_union_map_compute_flow(isl_union_map_copy(scop->reads),
 168                                 isl_union_map_copy(scop->writes), empty,
 169                                 isl_union_map_copy(scop->schedule),
 170                                 &scop->dep_flow, NULL, &scop->live_in, NULL);
 171 }
 172
 173 /* Eliminate dead code from ps->domain.
 174  *
 175  * In particular, intersect ps->domain with the (parts of) iteration
 176  * domains that are needed to produce the output or for statement
 177  * iterations that call functions.
 178  *
 179  * We start with the iteration domains that call functions
 180  * and the set of iterations that last write to an array
 181  * (except those that are later killed).
 182  *
 183  * Then we add those statement iterations that produce
 184  * something needed by the "live" statements iterations.
 185  * We keep doing this until no more statement iterations can be added.
 186  * To ensure that the procedure terminates, we compute the affine
 187  * hull of the live iterations (bounded to the original iteration
 188  * domains) each time we have added extra iterations.
 189  */
 190 static void eliminate_dead_code(struct ppcg_scop *ps)
 191 {
 192         isl_union_map *exposed;
 193         isl_union_set *live;
 194         isl_union_map *dep;
 195
 196         exposed = isl_union_map_union(isl_union_map_copy(ps->writes),
 197                                             isl_union_map_copy(ps->kills));
 198         exposed = isl_union_map_reverse(exposed);
 199         exposed = isl_union_map_apply_range(exposed,
 200                                             isl_union_map_copy(ps->schedule));
 201         exposed = isl_union_map_lexmax(exposed);
 202         exposed = isl_union_map_apply_range(exposed,
 203                     isl_union_map_reverse(isl_union_map_copy(ps->schedule)));
 204
 205         live = isl_union_map_range(exposed);
 206         live = isl_union_set_union(live, isl_union_set_copy(ps->call));
 207
 208         dep = isl_union_map_copy(ps->dep_flow);
 209         dep = isl_union_map_reverse(dep);
 210
 211         for (;;) {
 212                 isl_union_set *extra;
 213
 214                 extra = isl_union_set_apply(isl_union_set_copy(live),
 215                                             isl_union_map_copy(dep));
 216                 if (isl_union_set_is_subset(extra, live)) {
 217                         isl_union_set_free(extra);
 218                         break;
 219                 }
 220
 221                 live = isl_union_set_union(live, extra);
 222                 live = isl_union_set_affine_hull(live);
 223                 live = isl_union_set_intersect(live,
 224                                             isl_union_set_copy(ps->domain));
 225         }
 226
 227         isl_union_map_free(dep);
 228
 229         ps->domain = isl_union_set_intersect(ps->domain, live);
 230 }
 231
 232 /* Extract a ppcg_scop from a pet_scop.
 233  *
 234  * The constructed ppcg_scop refers to elements from the pet_scop
 235  * so the pet_scop should not be freed before the ppcg_scop.
 236  */
 237 static struct ppcg_scop *ppcg_scop_from_pet_scop(struct pet_scop *scop)
 238 {
 239         isl_ctx *ctx;
 240         struct ppcg_scop *ps;
 241
 242         if (!scop)
 243                 return NULL;
 244
 245         ctx = isl_set_get_ctx(scop->context);
 246
 247         ps = isl_calloc_type(ctx, struct ppcg_scop);
 248         if (!ps)
 249                 return NULL;
 250
 251         ps->context = isl_set_copy(scop->context);
 252         ps->domain = collect_non_kill_domains(scop);
 253         ps->call = collect_call_domains(scop);
 254         ps->reads = pet_scop_collect_reads(scop);
 255         ps->writes = pet_scop_collect_writes(scop);
 256         ps->kills = collect_kills(scop);
 257         ps->schedule = pet_scop_collect_schedule(scop);
 258         ps->n_array = scop->n_array;
 259         ps->arrays = scop->arrays;
 260         ps->n_stmt = scop->n_stmt;
 261         ps->stmts = scop->stmts;
 262
 263         compute_dependences(ps);
 264         eliminate_dead_code(ps);
 265
 266         return ps;
 267 }
 268
 269 static void ppcg_scop_free(struct ppcg_scop *ps)
 270 {
 271         if (!ps)
 272                 return;
 273
 274         isl_set_free(ps->context);
 275         isl_union_set_free(ps->domain);
 276         isl_union_set_free(ps->call);
 277         isl_union_map_free(ps->reads);
 278         isl_union_map_free(ps->live_in);
 279         isl_union_map_free(ps->writes);
 280         isl_union_map_free(ps->kills);
 281         isl_union_map_free(ps->dep_flow);
 282         isl_union_map_free(ps->schedule);
 283
 284         free(ps);
 285 }
 286
 287 int main(int argc, char **argv)
 288 {
 289         int r;
 290         isl_ctx *ctx;
 291         struct options *options;
 292         struct pet_scop *scop;
 293         struct ppcg_scop *ps;
 294
 295         options = options_new_with_defaults();
 296         assert(options);
 297
 298         ctx = isl_ctx_alloc_with_options(&options_args, options);
 299         isl_options_set_schedule_outer_zero_distance(ctx, 1);
 300         argc = options_parse(options, argc, argv, ISL_ARG_ALL);
 301
 302         scop = pet_scop_extract_from_C_source(ctx, options->input, NULL);
 303         scop = pet_scop_align_params(scop);
 304         ps = ppcg_scop_from_pet_scop(scop);
 305
 306         if (options->ppcg->target == PPCG_TARGET_CUDA)
 307                 r = generate_cuda(ctx, ps, options->ppcg, options->input);
 308         else
 309                 r = generate_cpu(ctx, ps, options->ppcg, options->input);
 310
 311         ppcg_scop_free(ps);
 312         pet_scop_free(scop);
 313
 314         isl_ctx_free(ctx);
 315
 316         return r;
 317 }