ppcg.c

   1 /*
   2  * Copyright 2011      INRIA Saclay
   3  *
   4  * Use of this software is governed by the MIT 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 <stdlib.h>
  14 #include <string.h>
  15 #include <isl/ctx.h>
  16 #include <isl/flow.h>
  17 #include <isl/options.h>
  18 #include <isl/schedule.h>
  19 #include <isl/ast_build.h>
  20 #include <isl/schedule.h>
  21 #include <pet.h>
  22 #include "ppcg.h"
  23 #include "ppcg_options.h"
  24 #include "cuda.h"
  25 #include "cpu.h"
  26
  27 struct options {
  28         struct isl_options *isl;
  29         struct pet_options *pet;
  30         struct ppcg_options *ppcg;
  31         char *input;
  32         char *output;
  33 };
  34
  35 const char *ppcg_version(void);
  36 static void print_version(void)
  37 {
  38         printf("%s", ppcg_version());
  39 }
  40
  41 ISL_ARGS_START(struct options, options_args)
  42 ISL_ARG_CHILD(struct options, isl, "isl", &isl_options_args, "isl options")
  43 ISL_ARG_CHILD(struct options, pet, "pet", &pet_options_args, "pet options")
  44 ISL_ARG_CHILD(struct options, ppcg, NULL, &ppcg_options_args, "ppcg options")
  45 ISL_ARG_STR(struct options, output, 'o', NULL,
  46         "filename", NULL, "output filename (c target)")
  47 ISL_ARG_ARG(struct options, input, "input", NULL)
  48 ISL_ARG_VERSION(print_version)
  49 ISL_ARGS_END
  50
  51 ISL_ARG_DEF(options, struct options, options_args)
  52
  53 /* Copy the base name of "input" to "name" and return its length.
  54  * "name" is not NULL terminated.
  55  *
  56  * In particular, remove all leading directory components and
  57  * the final extension, if any.
  58  */
  59 int ppcg_extract_base_name(char *name, const char *input)
  60 {
  61         const char *base;
  62         const char *ext;
  63         int len;
  64
  65         base = strrchr(input, '/');
  66         if (base)
  67                 base++;
  68         else
  69                 base = input;
  70         ext = strrchr(base, '.');
  71         len = ext ? ext - base : strlen(base);
  72
  73         memcpy(name, base, len);
  74
  75         return len;
  76 }
  77
  78 /* Is "stmt" a kill statement?
  79  */
  80 static int is_kill(struct pet_stmt *stmt)
  81 {
  82         if (stmt->body->type != pet_expr_unary)
  83                 return 0;
  84         return stmt->body->op == pet_op_kill;
  85 }
  86
  87 /* Is "stmt" not a kill statement?
  88  */
  89 static int is_not_kill(struct pet_stmt *stmt)
  90 {
  91         return !is_kill(stmt);
  92 }
  93
  94 /* Collect the iteration domains of the statements in "scop" that
  95  * satisfy "pred".
  96  */
  97 static __isl_give isl_union_set *collect_domains(struct pet_scop *scop,
  98         int (*pred)(struct pet_stmt *stmt))
  99 {
 100         int i;
 101         isl_set *domain_i;
 102         isl_union_set *domain;
 103
 104         if (!scop)
 105                 return NULL;
 106
 107         domain = isl_union_set_empty(isl_set_get_space(scop->context));
 108
 109         for (i = 0; i < scop->n_stmt; ++i) {
 110                 struct pet_stmt *stmt = scop->stmts[i];
 111
 112                 if (!pred(stmt))
 113                         continue;
 114
 115                 if (stmt->n_arg > 0)
 116                         isl_die(isl_union_set_get_ctx(domain),
 117                                 isl_error_unsupported,
 118                                 "data dependent conditions not supported",
 119                                 return isl_union_set_free(domain));
 120
 121                 domain_i = isl_set_copy(scop->stmts[i]->domain);
 122                 domain = isl_union_set_add_set(domain, domain_i);
 123         }
 124
 125         return domain;
 126 }
 127
 128 /* Collect the iteration domains of the statements in "scop",
 129  * skipping kill statements.
 130  */
 131 static __isl_give isl_union_set *collect_non_kill_domains(struct pet_scop *scop)
 132 {
 133         return collect_domains(scop, &is_not_kill);
 134 }
 135
 136 /* Does "expr" contain any call expressions?
 137  */
 138 static int expr_has_call(struct pet_expr *expr)
 139 {
 140         int i;
 141
 142         if (expr->type == pet_expr_call)
 143                 return 1;
 144
 145         for (i = 0; i < expr->n_arg; ++i)
 146                 if (expr_has_call(expr->args[i]))
 147                         return 1;
 148
 149         return 0;
 150 }
 151
 152 /* Does "stmt" contain any call expressions?
 153  */
 154 static int has_call(struct pet_stmt *stmt)
 155 {
 156         return expr_has_call(stmt->body);
 157 }
 158
 159 /* Collect the iteration domains of the statements in "scop"
 160  * that contain a call expression.
 161  */
 162 static __isl_give isl_union_set *collect_call_domains(struct pet_scop *scop)
 163 {
 164         return collect_domains(scop, &has_call);
 165 }
 166
 167 /* Compute the flow dependences and the live in accesses.
 168  */
 169 static void compute_flow_dep(struct ppcg_scop *ps)
 170 {
 171         isl_union_map *empty;
 172
 173         empty = isl_union_map_empty(isl_union_set_get_space(ps->domain));
 174         isl_union_map_compute_flow(isl_union_map_copy(ps->reads),
 175                                 isl_union_map_copy(ps->writes), empty,
 176                                 isl_union_map_copy(ps->schedule),
 177                                 &ps->dep_flow, NULL, &ps->live_in, NULL);
 178 }
 179
 180 /* Compute the dependences of the program represented by "scop".
 181  * Store the computed flow dependences
 182  * in scop->dep_flow and the reads with no corresponding writes in
 183  * scop->live_in.
 184  * Store the false (anti and output) dependences in scop->dep_false.
 185  */
 186 static void compute_dependences(struct ppcg_scop *scop)
 187 {
 188         isl_union_map *dep1, *dep2;
 189
 190         if (!scop)
 191                 return;
 192
 193         compute_flow_dep(scop);
 194
 195         isl_union_map_compute_flow(isl_union_map_copy(scop->writes),
 196                                 isl_union_map_copy(scop->writes),
 197                                 isl_union_map_copy(scop->reads),
 198                                 isl_union_map_copy(scop->schedule),
 199                                 &dep1, &dep2, NULL, NULL);
 200
 201         scop->dep_false = isl_union_map_union(dep1, dep2);
 202         scop->dep_false = isl_union_map_coalesce(scop->dep_false);
 203 }
 204
 205 /* Eliminate dead code from ps->domain.
 206  *
 207  * In particular, intersect ps->domain with the (parts of) iteration
 208  * domains that are needed to produce the output or for statement
 209  * iterations that call functions.
 210  *
 211  * We start with the iteration domains that call functions
 212  * and the set of iterations that last write to an array
 213  * (except those that are later killed).
 214  *
 215  * Then we add those statement iterations that produce
 216  * something needed by the "live" statements iterations.
 217  * We keep doing this until no more statement iterations can be added.
 218  * To ensure that the procedure terminates, we compute the affine
 219  * hull of the live iterations (bounded to the original iteration
 220  * domains) each time we have added extra iterations.
 221  */
 222 static void eliminate_dead_code(struct ppcg_scop *ps)
 223 {
 224         isl_union_map *exposed;
 225         isl_union_set *live;
 226         isl_union_map *dep;
 227
 228         exposed = isl_union_map_union(isl_union_map_copy(ps->writes),
 229                                             isl_union_map_copy(ps->kills));
 230         exposed = isl_union_map_reverse(exposed);
 231         exposed = isl_union_map_apply_range(exposed,
 232                                             isl_union_map_copy(ps->schedule));
 233         exposed = isl_union_map_lexmax(exposed);
 234         exposed = isl_union_map_apply_range(exposed,
 235                     isl_union_map_reverse(isl_union_map_copy(ps->schedule)));
 236
 237         live = isl_union_map_range(exposed);
 238         if (!isl_union_set_is_empty(ps->call)) {
 239                 live = isl_union_set_union(live, isl_union_set_copy(ps->call));
 240                 live = isl_union_set_coalesce(live);
 241         }
 242
 243         dep = isl_union_map_copy(ps->dep_flow);
 244         dep = isl_union_map_reverse(dep);
 245
 246         for (;;) {
 247                 isl_union_set *extra;
 248
 249                 extra = isl_union_set_apply(isl_union_set_copy(live),
 250                                             isl_union_map_copy(dep));
 251                 if (isl_union_set_is_subset(extra, live)) {
 252                         isl_union_set_free(extra);
 253                         break;
 254                 }
 255
 256                 live = isl_union_set_union(live, extra);
 257                 live = isl_union_set_affine_hull(live);
 258                 live = isl_union_set_intersect(live,
 259                                             isl_union_set_copy(ps->domain));
 260         }
 261
 262         isl_union_map_free(dep);
 263
 264         ps->domain = isl_union_set_intersect(ps->domain, live);
 265 }
 266
 267 /* Intersect "set" with the set described by "str", taking the NULL
 268  * string to represent the universal set.
 269  */
 270 static __isl_give isl_set *set_intersect_str(__isl_take isl_set *set,
 271         const char *str)
 272 {
 273         isl_ctx *ctx;
 274         isl_set *set2;
 275
 276         if (!str)
 277                 return set;
 278
 279         ctx = isl_set_get_ctx(set);
 280         set2 = isl_set_read_from_str(ctx, str);
 281         set = isl_set_intersect(set, set2);
 282
 283         return set;
 284 }
 285
 286 /* Does "expr" involve any data dependent accesses?
 287  */
 288 static int expr_has_data_dependent_accesses(struct pet_expr *expr)
 289 {
 290         int i;
 291
 292         for (i = 0; i < expr->n_arg; ++i)
 293                 if (expr_has_data_dependent_accesses(expr->args[i]))
 294                         return 1;
 295
 296         if (expr->type == pet_expr_access && expr->n_arg > 0)
 297                 return 1;
 298
 299         return 0;
 300 }
 301
 302 /* Does "stmt" contain any data dependent accesses?
 303  */
 304 static int stmt_has_data_dependent_accesses(struct pet_stmt *stmt)
 305 {
 306         return expr_has_data_dependent_accesses(stmt->body);
 307 }
 308
 309 /* Does "scop" contain any data dependent accesses?
 310  */
 311 static int scop_has_data_dependent_accesses(struct pet_scop *scop)
 312 {
 313         int i;
 314
 315         if (!scop)
 316                 return -1;
 317         for (i = 0; i < scop->n_stmt; ++i)
 318                 if (stmt_has_data_dependent_accesses(scop->stmts[i]))
 319                         return 1;
 320
 321         return 0;
 322 }
 323
 324 static void *ppcg_scop_free(struct ppcg_scop *ps)
 325 {
 326         if (!ps)
 327                 return NULL;
 328
 329         isl_set_free(ps->context);
 330         isl_union_set_free(ps->domain);
 331         isl_union_set_free(ps->call);
 332         isl_union_map_free(ps->reads);
 333         isl_union_map_free(ps->live_in);
 334         isl_union_map_free(ps->writes);
 335         isl_union_map_free(ps->kills);
 336         isl_union_map_free(ps->dep_flow);
 337         isl_union_map_free(ps->dep_false);
 338         isl_union_map_free(ps->schedule);
 339
 340         free(ps);
 341
 342         return NULL;
 343 }
 344
 345 /* Extract a ppcg_scop from a pet_scop.
 346  *
 347  * The constructed ppcg_scop refers to elements from the pet_scop
 348  * so the pet_scop should not be freed before the ppcg_scop.
 349  */
 350 static struct ppcg_scop *ppcg_scop_from_pet_scop(struct pet_scop *scop,
 351         struct ppcg_options *options)
 352 {
 353         isl_ctx *ctx;
 354         struct ppcg_scop *ps;
 355
 356         if (!scop)
 357                 return NULL;
 358
 359         ctx = isl_set_get_ctx(scop->context);
 360
 361         if (scop_has_data_dependent_accesses(scop))
 362                 isl_die(ctx, isl_error_unsupported,
 363                         "data dependent accesses not supported",
 364                         return NULL);
 365
 366         ps = isl_calloc_type(ctx, struct ppcg_scop);
 367         if (!ps)
 368                 return NULL;
 369
 370         ps->options = options;
 371         ps->start = scop->start;
 372         ps->end = scop->end;
 373         ps->context = isl_set_copy(scop->context);
 374         ps->context = set_intersect_str(ps->context, options->ctx);
 375         ps->domain = collect_non_kill_domains(scop);
 376         ps->call = collect_call_domains(scop);
 377         ps->reads = pet_scop_collect_may_reads(scop);
 378         ps->writes = pet_scop_collect_may_writes(scop);
 379         ps->kills = pet_scop_collect_must_kills(scop);
 380         ps->schedule = pet_scop_collect_schedule(scop);
 381         ps->n_type = scop->n_type;
 382         ps->types = scop->types;
 383         ps->n_array = scop->n_array;
 384         ps->arrays = scop->arrays;
 385         ps->n_stmt = scop->n_stmt;
 386         ps->stmts = scop->stmts;
 387
 388         compute_dependences(ps);
 389         eliminate_dead_code(ps);
 390
 391         if (!ps->context || !ps->domain || !ps->call || !ps->reads ||
 392             !ps->writes || !ps->kills || !ps->schedule)
 393                 return ppcg_scop_free(ps);
 394
 395         return ps;
 396 }
 397
 398 /* Internal data structure for ppcg_transform.
 399  */
 400 struct ppcg_transform_data {
 401         struct ppcg_options *options;
 402         __isl_give isl_printer *(*transform)(__isl_take isl_printer *p,
 403                 struct ppcg_scop *scop, void *user);
 404         void *user;
 405 };
 406
 407 /* Callback for pet_transform_C_source that transforms
 408  * the given pet_scop to a ppcg_scop before calling the
 409  * ppcg_transform callback.
 410  */
 411 static __isl_give isl_printer *transform(__isl_take isl_printer *p,
 412         struct pet_scop *scop, void *user)
 413 {
 414         struct ppcg_transform_data *data = user;
 415         struct ppcg_scop *ps;
 416
 417         if (pet_scop_has_data_dependent_accesses(scop) ||
 418             pet_scop_has_data_dependent_conditions(scop)) {
 419                 p = pet_scop_print_original(scop, p);
 420                 pet_scop_free(scop);
 421                 return p;
 422         }
 423
 424         scop = pet_scop_align_params(scop);
 425         ps = ppcg_scop_from_pet_scop(scop, data->options);
 426
 427         p = data->transform(p, ps, data->user);
 428
 429         ppcg_scop_free(ps);
 430         pet_scop_free(scop);
 431
 432         return p;
 433 }
 434
 435 /* Transform the C source file "input" by rewriting each scop
 436  * through a call to "transform".
 437  * The transformed C code is written to "out".
 438  *
 439  * This is a wrapper around pet_transform_C_source that transforms
 440  * the pet_scop to a ppcg_scop before calling "fn".
 441  */
 442 int ppcg_transform(isl_ctx *ctx, const char *input, FILE *out,
 443         struct ppcg_options *options,
 444         __isl_give isl_printer *(*fn)(__isl_take isl_printer *p,
 445                 struct ppcg_scop *scop, void *user), void *user)
 446 {
 447         struct ppcg_transform_data data = { options, fn, user };
 448         return pet_transform_C_source(ctx, input, out, &transform, &data);
 449 }
 450
 451 /* Check consistency of options.
 452  *
 453  * Return -1 on error.
 454  */
 455 static int check_options(isl_ctx *ctx)
 456 {
 457         struct options *options;
 458
 459         options = isl_ctx_peek_options(ctx, &options_args);
 460         if (!options)
 461                 isl_die(ctx, isl_error_internal,
 462                         "unable to find options", return -1);
 463
 464         if (options->ppcg->openmp &&
 465             !isl_options_get_ast_build_atomic_upper_bound(ctx))
 466                 isl_die(ctx, isl_error_invalid,
 467                         "OpenMP requires atomic bounds", return -1);
 468
 469         return 0;
 470 }
 471
 472 int main(int argc, char **argv)
 473 {
 474         int r;
 475         isl_ctx *ctx;
 476         struct options *options;
 477
 478         options = options_new_with_defaults();
 479         assert(options);
 480
 481         ctx = isl_ctx_alloc_with_options(&options_args, options);
 482         isl_options_set_schedule_outer_coincidence(ctx, 1);
 483         isl_options_set_schedule_maximize_band_depth(ctx, 1);
 484         argc = options_parse(options, argc, argv, ISL_ARG_ALL);
 485
 486         if (check_options(ctx) < 0)
 487                 r = EXIT_FAILURE;
 488         else if (options->ppcg->target == PPCG_TARGET_CUDA)
 489                 r = generate_cuda(ctx, options->ppcg, options->input);
 490         else
 491                 r = generate_cpu(ctx, options->ppcg, options->input,
 492                                 options->output);
 493
 494         isl_ctx_free(ctx);
 495
 496         return r;
 497 }