2 * Copyright 2011 INRIA Saclay
3 * Copyright 2013 Ecole Normale Superieure
5 * Use of this software is governed by the MIT license
7 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
8 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
10 * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France
19 #include <isl/options.h>
20 #include <isl/schedule.h>
21 #include <isl/ast_build.h>
22 #include <isl/schedule.h>
25 #include "ppcg_options.h"
30 struct isl_options
*isl
;
31 struct pet_options
*pet
;
32 struct ppcg_options
*ppcg
;
37 const char *ppcg_version(void);
38 static void print_version(void)
40 printf("%s", ppcg_version());
43 ISL_ARGS_START(struct options
, options_args
)
44 ISL_ARG_CHILD(struct options
, isl
, "isl", &isl_options_args
, "isl options")
45 ISL_ARG_CHILD(struct options
, pet
, "pet", &pet_options_args
, "pet options")
46 ISL_ARG_CHILD(struct options
, ppcg
, NULL
, &ppcg_options_args
, "ppcg options")
47 ISL_ARG_STR(struct options
, output
, 'o', NULL
,
48 "filename", NULL
, "output filename (c target)")
49 ISL_ARG_ARG(struct options
, input
, "input", NULL
)
50 ISL_ARG_VERSION(print_version
)
53 ISL_ARG_DEF(options
, struct options
, options_args
)
55 /* Copy the base name of "input" to "name" and return its length.
56 * "name" is not NULL terminated.
58 * In particular, remove all leading directory components and
59 * the final extension, if any.
61 int ppcg_extract_base_name(char *name
, const char *input
)
67 base
= strrchr(input
, '/');
72 ext
= strrchr(base
, '.');
73 len
= ext
? ext
- base
: strlen(base
);
75 memcpy(name
, base
, len
);
80 /* Is "stmt" a kill statement?
82 static int is_kill(struct pet_stmt
*stmt
)
84 if (stmt
->body
->type
!= pet_expr_unary
)
86 return stmt
->body
->op
== pet_op_kill
;
89 /* Is "stmt" not a kill statement?
91 static int is_not_kill(struct pet_stmt
*stmt
)
93 return !is_kill(stmt
);
96 /* Collect the iteration domains of the statements in "scop" that
99 static __isl_give isl_union_set
*collect_domains(struct pet_scop
*scop
,
100 int (*pred
)(struct pet_stmt
*stmt
))
104 isl_union_set
*domain
;
109 domain
= isl_union_set_empty(isl_set_get_space(scop
->context
));
111 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
112 struct pet_stmt
*stmt
= scop
->stmts
[i
];
118 isl_die(isl_union_set_get_ctx(domain
),
119 isl_error_unsupported
,
120 "data dependent conditions not supported",
121 return isl_union_set_free(domain
));
123 domain_i
= isl_set_copy(scop
->stmts
[i
]->domain
);
124 domain
= isl_union_set_add_set(domain
, domain_i
);
130 /* Collect the iteration domains of the statements in "scop",
131 * skipping kill statements.
133 static __isl_give isl_union_set
*collect_non_kill_domains(struct pet_scop
*scop
)
135 return collect_domains(scop
, &is_not_kill
);
138 /* Does "expr" contain any call expressions?
140 static int expr_has_call(struct pet_expr
*expr
)
144 if (expr
->type
== pet_expr_call
)
147 for (i
= 0; i
< expr
->n_arg
; ++i
)
148 if (expr_has_call(expr
->args
[i
]))
154 /* Does "stmt" contain any call expressions?
156 static int has_call(struct pet_stmt
*stmt
)
158 return expr_has_call(stmt
->body
);
161 /* Collect the iteration domains of the statements in "scop"
162 * that contain a call expression.
164 static __isl_give isl_union_set
*collect_call_domains(struct pet_scop
*scop
)
166 return collect_domains(scop
, &has_call
);
169 /* Construct a relation from the iteration domains to tagged iteration
170 * domains with as range the reference tags that appear
171 * in any of the reads, writes or kills.
172 * Store the result in ps->tagger.
174 * For example, if the statement with iteration space S[i,j]
175 * contains two array references R_1[] and R_2[], then ps->tagger will contain
177 * { S[i,j] -> [S[i,j] -> R_1[]]; S[i,j] -> [S[i,j] -> R_2[]] }
179 static void compute_tagger(struct ppcg_scop
*ps
)
181 isl_union_map
*tagged
, *tagger
;
183 tagged
= isl_union_map_copy(ps
->tagged_reads
);
184 tagged
= isl_union_map_union(tagged
,
185 isl_union_map_copy(ps
->tagged_may_writes
));
186 tagged
= isl_union_map_union(tagged
,
187 isl_union_map_copy(ps
->tagged_must_kills
));
189 tagger
= isl_union_map_universe(tagged
);
190 tagger
= isl_union_set_unwrap(isl_union_map_domain(tagger
));
191 tagger
= isl_union_map_reverse(isl_union_map_domain_map(tagger
));
196 /* Compute the live out accesses, i.e., the writes that are
197 * potentially not killed by any kills or any other writes, and
198 * store them in ps->live_out.
200 * We compute the "dependence" of any "kill" (an explicit kill
201 * or a must write) on any may write.
202 * The may writes with a "depending" kill are definitely killed.
203 * The remaining may writes can potentially be live out.
205 static void compute_live_out(struct ppcg_scop
*ps
)
207 isl_union_map
*tagger
;
208 isl_union_map
*schedule
;
209 isl_union_map
*empty
;
210 isl_union_map
*kills
;
211 isl_union_map
*exposed
;
212 isl_union_map
*covering
;
214 tagger
= isl_union_map_copy(ps
->tagger
);
215 schedule
= isl_union_map_copy(ps
->schedule
);
216 schedule
= isl_union_map_apply_domain(schedule
,
217 isl_union_map_copy(tagger
));
218 empty
= isl_union_map_empty(isl_union_set_get_space(ps
->domain
));
219 kills
= isl_union_map_union(isl_union_map_copy(ps
->tagged_must_writes
),
220 isl_union_map_copy(ps
->tagged_must_kills
));
221 isl_union_map_compute_flow(kills
, empty
,
222 isl_union_map_copy(ps
->tagged_may_writes
),
223 schedule
, NULL
, &covering
, NULL
, NULL
);
224 exposed
= isl_union_map_copy(ps
->tagged_may_writes
);
225 exposed
= isl_union_map_subtract_domain(exposed
,
226 isl_union_map_domain(covering
));
227 exposed
= isl_union_map_apply_range(tagger
, exposed
);
228 ps
->live_out
= exposed
;
231 /* Compute the potential flow dependences and the potential live in
234 static void compute_flow_dep(struct ppcg_scop
*ps
)
236 isl_union_map
*may_flow
;
237 isl_union_map
*may_live_in
;
239 isl_union_map_compute_flow(isl_union_map_copy(ps
->reads
),
240 isl_union_map_copy(ps
->must_writes
),
241 isl_union_map_copy(ps
->may_writes
),
242 isl_union_map_copy(ps
->schedule
),
243 &ps
->dep_flow
, &may_flow
,
244 &ps
->live_in
, &may_live_in
);
246 ps
->dep_flow
= isl_union_map_union(ps
->dep_flow
, may_flow
);
247 ps
->live_in
= isl_union_map_union(ps
->live_in
, may_live_in
);
250 /* Compute the dependences of the program represented by "scop".
251 * Store the computed potential flow dependences
252 * in scop->dep_flow and the reads with potentially no corresponding writes in
254 * Store the potential live out accesses in scop->live_out.
255 * Store the potential false (anti and output) dependences in scop->dep_false.
257 static void compute_dependences(struct ppcg_scop
*scop
)
259 isl_union_map
*dep1
, *dep2
;
260 isl_union_map
*may_source
;
265 compute_live_out(scop
);
267 compute_flow_dep(scop
);
269 may_source
= isl_union_map_union(isl_union_map_copy(scop
->may_writes
),
270 isl_union_map_copy(scop
->reads
));
271 isl_union_map_compute_flow(isl_union_map_copy(scop
->may_writes
),
272 isl_union_map_copy(scop
->must_writes
),
273 may_source
, isl_union_map_copy(scop
->schedule
),
274 &dep1
, &dep2
, NULL
, NULL
);
276 scop
->dep_false
= isl_union_map_union(dep1
, dep2
);
277 scop
->dep_false
= isl_union_map_coalesce(scop
->dep_false
);
280 /* Eliminate dead code from ps->domain.
282 * In particular, intersect ps->domain with the (parts of) iteration
283 * domains that are needed to produce the output or for statement
284 * iterations that call functions.
286 * We start with the iteration domains that call functions
287 * and the set of iterations that last write to an array
288 * (except those that are later killed).
290 * Then we add those statement iterations that produce
291 * something needed by the "live" statements iterations.
292 * We keep doing this until no more statement iterations can be added.
293 * To ensure that the procedure terminates, we compute the affine
294 * hull of the live iterations (bounded to the original iteration
295 * domains) each time we have added extra iterations.
297 static void eliminate_dead_code(struct ppcg_scop
*ps
)
302 live
= isl_union_map_domain(isl_union_map_copy(ps
->live_out
));
303 if (!isl_union_set_is_empty(ps
->call
)) {
304 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
305 live
= isl_union_set_coalesce(live
);
308 dep
= isl_union_map_copy(ps
->dep_flow
);
309 dep
= isl_union_map_reverse(dep
);
312 isl_union_set
*extra
;
314 extra
= isl_union_set_apply(isl_union_set_copy(live
),
315 isl_union_map_copy(dep
));
316 if (isl_union_set_is_subset(extra
, live
)) {
317 isl_union_set_free(extra
);
321 live
= isl_union_set_union(live
, extra
);
322 live
= isl_union_set_affine_hull(live
);
323 live
= isl_union_set_intersect(live
,
324 isl_union_set_copy(ps
->domain
));
327 isl_union_map_free(dep
);
329 ps
->domain
= isl_union_set_intersect(ps
->domain
, live
);
332 /* Intersect "set" with the set described by "str", taking the NULL
333 * string to represent the universal set.
335 static __isl_give isl_set
*set_intersect_str(__isl_take isl_set
*set
,
344 ctx
= isl_set_get_ctx(set
);
345 set2
= isl_set_read_from_str(ctx
, str
);
346 set
= isl_set_intersect(set
, set2
);
351 /* Does "expr" involve any data dependent accesses?
353 static int expr_has_data_dependent_accesses(struct pet_expr
*expr
)
357 for (i
= 0; i
< expr
->n_arg
; ++i
)
358 if (expr_has_data_dependent_accesses(expr
->args
[i
]))
361 if (expr
->type
== pet_expr_access
&& expr
->n_arg
> 0)
367 /* Does "stmt" contain any data dependent accesses?
369 static int stmt_has_data_dependent_accesses(struct pet_stmt
*stmt
)
371 return expr_has_data_dependent_accesses(stmt
->body
);
374 /* Does "scop" contain any data dependent accesses?
376 static int scop_has_data_dependent_accesses(struct pet_scop
*scop
)
382 for (i
= 0; i
< scop
->n_stmt
; ++i
)
383 if (stmt_has_data_dependent_accesses(scop
->stmts
[i
]))
389 static void *ppcg_scop_free(struct ppcg_scop
*ps
)
394 isl_set_free(ps
->context
);
395 isl_union_set_free(ps
->domain
);
396 isl_union_set_free(ps
->call
);
397 isl_union_map_free(ps
->tagged_reads
);
398 isl_union_map_free(ps
->reads
);
399 isl_union_map_free(ps
->live_in
);
400 isl_union_map_free(ps
->tagged_may_writes
);
401 isl_union_map_free(ps
->tagged_must_writes
);
402 isl_union_map_free(ps
->may_writes
);
403 isl_union_map_free(ps
->must_writes
);
404 isl_union_map_free(ps
->live_out
);
405 isl_union_map_free(ps
->tagged_must_kills
);
406 isl_union_map_free(ps
->dep_flow
);
407 isl_union_map_free(ps
->dep_false
);
408 isl_union_map_free(ps
->schedule
);
409 isl_union_map_free(ps
->tagger
);
416 /* Extract a ppcg_scop from a pet_scop.
418 * The constructed ppcg_scop refers to elements from the pet_scop
419 * so the pet_scop should not be freed before the ppcg_scop.
421 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
,
422 struct ppcg_options
*options
)
425 struct ppcg_scop
*ps
;
430 ctx
= isl_set_get_ctx(scop
->context
);
432 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
436 ps
->options
= options
;
437 ps
->start
= scop
->start
;
439 ps
->context
= isl_set_copy(scop
->context
);
440 ps
->context
= set_intersect_str(ps
->context
, options
->ctx
);
441 ps
->domain
= collect_non_kill_domains(scop
);
442 ps
->call
= collect_call_domains(scop
);
443 ps
->tagged_reads
= pet_scop_collect_tagged_may_reads(scop
);
444 ps
->reads
= pet_scop_collect_may_reads(scop
);
445 ps
->tagged_may_writes
= pet_scop_collect_tagged_may_writes(scop
);
446 ps
->may_writes
= pet_scop_collect_may_writes(scop
);
447 ps
->tagged_must_writes
= pet_scop_collect_tagged_must_writes(scop
);
448 ps
->must_writes
= pet_scop_collect_must_writes(scop
);
449 ps
->tagged_must_kills
= pet_scop_collect_tagged_must_kills(scop
);
450 ps
->schedule
= pet_scop_collect_schedule(scop
);
451 ps
->n_type
= scop
->n_type
;
452 ps
->types
= scop
->types
;
453 ps
->n_array
= scop
->n_array
;
454 ps
->arrays
= scop
->arrays
;
455 ps
->n_stmt
= scop
->n_stmt
;
456 ps
->stmts
= scop
->stmts
;
459 compute_dependences(ps
);
460 eliminate_dead_code(ps
);
462 if (!ps
->context
|| !ps
->domain
|| !ps
->call
|| !ps
->reads
||
463 !ps
->may_writes
|| !ps
->must_writes
|| !ps
->tagged_must_kills
||
465 return ppcg_scop_free(ps
);
470 /* Internal data structure for ppcg_transform.
472 struct ppcg_transform_data
{
473 struct ppcg_options
*options
;
474 __isl_give isl_printer
*(*transform
)(__isl_take isl_printer
*p
,
475 struct ppcg_scop
*scop
, void *user
);
479 /* Callback for pet_transform_C_source that transforms
480 * the given pet_scop to a ppcg_scop before calling the
481 * ppcg_transform callback.
483 static __isl_give isl_printer
*transform(__isl_take isl_printer
*p
,
484 struct pet_scop
*scop
, void *user
)
486 struct ppcg_transform_data
*data
= user
;
487 struct ppcg_scop
*ps
;
489 if (pet_scop_has_data_dependent_conditions(scop
)) {
490 p
= pet_scop_print_original(scop
, p
);
495 scop
= pet_scop_align_params(scop
);
496 ps
= ppcg_scop_from_pet_scop(scop
, data
->options
);
498 p
= data
->transform(p
, ps
, data
->user
);
506 /* Transform the C source file "input" by rewriting each scop
507 * through a call to "transform".
508 * The transformed C code is written to "out".
510 * This is a wrapper around pet_transform_C_source that transforms
511 * the pet_scop to a ppcg_scop before calling "fn".
513 int ppcg_transform(isl_ctx
*ctx
, const char *input
, FILE *out
,
514 struct ppcg_options
*options
,
515 __isl_give isl_printer
*(*fn
)(__isl_take isl_printer
*p
,
516 struct ppcg_scop
*scop
, void *user
), void *user
)
518 struct ppcg_transform_data data
= { options
, fn
, user
};
519 return pet_transform_C_source(ctx
, input
, out
, &transform
, &data
);
522 /* Check consistency of options.
524 * Return -1 on error.
526 static int check_options(isl_ctx
*ctx
)
528 struct options
*options
;
530 options
= isl_ctx_peek_options(ctx
, &options_args
);
532 isl_die(ctx
, isl_error_internal
,
533 "unable to find options", return -1);
535 if (options
->ppcg
->openmp
&&
536 !isl_options_get_ast_build_atomic_upper_bound(ctx
))
537 isl_die(ctx
, isl_error_invalid
,
538 "OpenMP requires atomic bounds", return -1);
543 int main(int argc
, char **argv
)
547 struct options
*options
;
549 options
= options_new_with_defaults();
552 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
553 isl_options_set_schedule_outer_coincidence(ctx
, 1);
554 isl_options_set_schedule_maximize_band_depth(ctx
, 1);
555 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
557 if (check_options(ctx
) < 0)
559 else if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
560 r
= generate_cuda(ctx
, options
->ppcg
, options
->input
);
562 r
= generate_cpu(ctx
, options
->ppcg
, options
->input
,