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 /* Given a union of "tagged" access relations of the form
171 * [S_i[...] -> R_j[]] -> A_k[...]
173 * project out the "tags" (R_j[]).
174 * That is, return a union of relations of the form
176 * S_i[...] -> A_k[...]
178 static __isl_give isl_union_map
*project_out_tags(
179 __isl_take isl_union_map
*umap
)
183 proj
= isl_union_map_universe(isl_union_map_copy(umap
));
184 proj
= isl_union_set_unwrap(isl_union_map_domain(proj
));
185 proj
= isl_union_map_domain_map(proj
);
187 umap
= isl_union_map_apply_domain(umap
, proj
);
192 /* Construct a relation from the iteration domains to tagged iteration
193 * domains with as range the reference tags that appear
194 * in any of the reads, writes or kills.
195 * Store the result in ps->tagger.
197 * For example, if the statement with iteration space S[i,j]
198 * contains two array references R_1[] and R_2[], then ps->tagger will contain
200 * { S[i,j] -> [S[i,j] -> R_1[]]; S[i,j] -> [S[i,j] -> R_2[]] }
202 static void compute_tagger(struct ppcg_scop
*ps
)
204 isl_union_map
*tagged
, *tagger
;
206 tagged
= isl_union_map_copy(ps
->tagged_reads
);
207 tagged
= isl_union_map_union(tagged
,
208 isl_union_map_copy(ps
->tagged_may_writes
));
209 tagged
= isl_union_map_union(tagged
,
210 isl_union_map_copy(ps
->tagged_must_kills
));
212 tagger
= isl_union_map_universe(tagged
);
213 tagger
= isl_union_set_unwrap(isl_union_map_domain(tagger
));
214 tagger
= isl_union_map_reverse(isl_union_map_domain_map(tagger
));
219 /* Compute the live out accesses, i.e., the writes that are
220 * potentially not killed by any kills or any other writes, and
221 * store them in ps->live_out.
223 * We compute the "dependence" of any "kill" (an explicit kill
224 * or a must write) on any may write.
225 * The may writes with a "depending" kill are definitely killed.
226 * The remaining may writes can potentially be live out.
228 static void compute_live_out(struct ppcg_scop
*ps
)
230 isl_union_map
*tagger
;
231 isl_union_map
*schedule
;
232 isl_union_map
*empty
;
233 isl_union_map
*kills
;
234 isl_union_map
*exposed
;
235 isl_union_map
*covering
;
237 tagger
= isl_union_map_copy(ps
->tagger
);
238 schedule
= isl_union_map_copy(ps
->schedule
);
239 schedule
= isl_union_map_apply_domain(schedule
,
240 isl_union_map_copy(tagger
));
241 empty
= isl_union_map_empty(isl_union_set_get_space(ps
->domain
));
242 kills
= isl_union_map_union(isl_union_map_copy(ps
->tagged_must_writes
),
243 isl_union_map_copy(ps
->tagged_must_kills
));
244 isl_union_map_compute_flow(kills
, empty
,
245 isl_union_map_copy(ps
->tagged_may_writes
),
246 schedule
, NULL
, &covering
, NULL
, NULL
);
247 exposed
= isl_union_map_copy(ps
->tagged_may_writes
);
248 exposed
= isl_union_map_subtract_domain(exposed
,
249 isl_union_map_domain(covering
));
250 exposed
= isl_union_map_apply_range(tagger
, exposed
);
251 ps
->live_out
= exposed
;
254 /* Compute the flow dependences and the live_in accesses and store
255 * the results in ps->dep_flow and ps->live_in.
256 * A copy of the flow dependences, tagged with the reference tags
257 * is stored in ps->tagged_dep_flow.
259 * We first compute ps->tagged_dep_flow, i.e., the tagged flow dependences
260 * and then project out the tags.
262 static void compute_tagged_flow_dep(struct ppcg_scop
*ps
)
264 isl_union_map
*tagger
;
265 isl_union_map
*schedule
;
266 isl_union_map
*may_flow
;
267 isl_union_map
*live_in
, *may_live_in
;
269 tagger
= isl_union_map_copy(ps
->tagger
);
270 schedule
= isl_union_map_copy(ps
->schedule
);
271 schedule
= isl_union_map_apply_domain(schedule
, tagger
);
272 isl_union_map_compute_flow(isl_union_map_copy(ps
->tagged_reads
),
273 isl_union_map_copy(ps
->tagged_must_writes
),
274 isl_union_map_copy(ps
->tagged_may_writes
),
275 schedule
, &ps
->tagged_dep_flow
, &may_flow
,
276 &live_in
, &may_live_in
);
277 ps
->tagged_dep_flow
= isl_union_map_union(ps
->tagged_dep_flow
,
279 ps
->dep_flow
= isl_union_map_copy(ps
->tagged_dep_flow
);
280 ps
->dep_flow
= isl_union_map_zip(ps
->dep_flow
);
281 ps
->dep_flow
= isl_union_set_unwrap(isl_union_map_domain(ps
->dep_flow
));
282 live_in
= isl_union_map_union(live_in
, may_live_in
);
283 ps
->live_in
= project_out_tags(live_in
);
286 /* Compute the potential flow dependences and the potential live in
289 static void compute_flow_dep(struct ppcg_scop
*ps
)
291 isl_union_map
*may_flow
;
292 isl_union_map
*may_live_in
;
294 isl_union_map_compute_flow(isl_union_map_copy(ps
->reads
),
295 isl_union_map_copy(ps
->must_writes
),
296 isl_union_map_copy(ps
->may_writes
),
297 isl_union_map_copy(ps
->schedule
),
298 &ps
->dep_flow
, &may_flow
,
299 &ps
->live_in
, &may_live_in
);
301 ps
->dep_flow
= isl_union_map_union(ps
->dep_flow
, may_flow
);
302 ps
->live_in
= isl_union_map_union(ps
->live_in
, may_live_in
);
305 /* Compute the dependences of the program represented by "scop".
306 * Store the computed potential flow dependences
307 * in scop->dep_flow and the reads with potentially no corresponding writes in
309 * Store the potential live out accesses in scop->live_out.
310 * Store the potential false (anti and output) dependences in scop->dep_false.
312 static void compute_dependences(struct ppcg_scop
*scop
)
314 isl_union_map
*dep1
, *dep2
;
315 isl_union_map
*may_source
;
320 compute_live_out(scop
);
322 if (scop
->options
->target
!= PPCG_TARGET_C
)
323 compute_tagged_flow_dep(scop
);
325 compute_flow_dep(scop
);
327 may_source
= isl_union_map_union(isl_union_map_copy(scop
->may_writes
),
328 isl_union_map_copy(scop
->reads
));
329 isl_union_map_compute_flow(isl_union_map_copy(scop
->may_writes
),
330 isl_union_map_copy(scop
->must_writes
),
331 may_source
, isl_union_map_copy(scop
->schedule
),
332 &dep1
, &dep2
, NULL
, NULL
);
334 scop
->dep_false
= isl_union_map_union(dep1
, dep2
);
335 scop
->dep_false
= isl_union_map_coalesce(scop
->dep_false
);
338 /* Eliminate dead code from ps->domain.
340 * In particular, intersect ps->domain with the (parts of) iteration
341 * domains that are needed to produce the output or for statement
342 * iterations that call functions.
344 * We start with the iteration domains that call functions
345 * and the set of iterations that last write to an array
346 * (except those that are later killed).
348 * Then we add those statement iterations that produce
349 * something needed by the "live" statements iterations.
350 * We keep doing this until no more statement iterations can be added.
351 * To ensure that the procedure terminates, we compute the affine
352 * hull of the live iterations (bounded to the original iteration
353 * domains) each time we have added extra iterations.
355 static void eliminate_dead_code(struct ppcg_scop
*ps
)
360 live
= isl_union_map_domain(isl_union_map_copy(ps
->live_out
));
361 if (!isl_union_set_is_empty(ps
->call
)) {
362 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
363 live
= isl_union_set_coalesce(live
);
366 dep
= isl_union_map_copy(ps
->dep_flow
);
367 dep
= isl_union_map_reverse(dep
);
370 isl_union_set
*extra
;
372 extra
= isl_union_set_apply(isl_union_set_copy(live
),
373 isl_union_map_copy(dep
));
374 if (isl_union_set_is_subset(extra
, live
)) {
375 isl_union_set_free(extra
);
379 live
= isl_union_set_union(live
, extra
);
380 live
= isl_union_set_affine_hull(live
);
381 live
= isl_union_set_intersect(live
,
382 isl_union_set_copy(ps
->domain
));
385 isl_union_map_free(dep
);
387 ps
->domain
= isl_union_set_intersect(ps
->domain
, live
);
390 /* Intersect "set" with the set described by "str", taking the NULL
391 * string to represent the universal set.
393 static __isl_give isl_set
*set_intersect_str(__isl_take isl_set
*set
,
402 ctx
= isl_set_get_ctx(set
);
403 set2
= isl_set_read_from_str(ctx
, str
);
404 set
= isl_set_intersect(set
, set2
);
409 /* Does "expr" involve any data dependent accesses?
411 static int expr_has_data_dependent_accesses(struct pet_expr
*expr
)
415 for (i
= 0; i
< expr
->n_arg
; ++i
)
416 if (expr_has_data_dependent_accesses(expr
->args
[i
]))
419 if (expr
->type
== pet_expr_access
&& expr
->n_arg
> 0)
425 /* Does "stmt" contain any data dependent accesses?
427 static int stmt_has_data_dependent_accesses(struct pet_stmt
*stmt
)
429 return expr_has_data_dependent_accesses(stmt
->body
);
432 /* Does "scop" contain any data dependent accesses?
434 static int scop_has_data_dependent_accesses(struct pet_scop
*scop
)
440 for (i
= 0; i
< scop
->n_stmt
; ++i
)
441 if (stmt_has_data_dependent_accesses(scop
->stmts
[i
]))
447 static void *ppcg_scop_free(struct ppcg_scop
*ps
)
452 isl_set_free(ps
->context
);
453 isl_union_set_free(ps
->domain
);
454 isl_union_set_free(ps
->call
);
455 isl_union_map_free(ps
->tagged_reads
);
456 isl_union_map_free(ps
->reads
);
457 isl_union_map_free(ps
->live_in
);
458 isl_union_map_free(ps
->tagged_may_writes
);
459 isl_union_map_free(ps
->tagged_must_writes
);
460 isl_union_map_free(ps
->may_writes
);
461 isl_union_map_free(ps
->must_writes
);
462 isl_union_map_free(ps
->live_out
);
463 isl_union_map_free(ps
->tagged_must_kills
);
464 isl_union_map_free(ps
->tagged_dep_flow
);
465 isl_union_map_free(ps
->dep_flow
);
466 isl_union_map_free(ps
->dep_false
);
467 isl_union_map_free(ps
->schedule
);
468 isl_union_map_free(ps
->tagger
);
475 /* Extract a ppcg_scop from a pet_scop.
477 * The constructed ppcg_scop refers to elements from the pet_scop
478 * so the pet_scop should not be freed before the ppcg_scop.
480 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
,
481 struct ppcg_options
*options
)
484 struct ppcg_scop
*ps
;
489 ctx
= isl_set_get_ctx(scop
->context
);
491 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
495 ps
->options
= options
;
496 ps
->start
= scop
->start
;
498 ps
->context
= isl_set_copy(scop
->context
);
499 ps
->context
= set_intersect_str(ps
->context
, options
->ctx
);
500 ps
->domain
= collect_non_kill_domains(scop
);
501 ps
->call
= collect_call_domains(scop
);
502 ps
->tagged_reads
= pet_scop_collect_tagged_may_reads(scop
);
503 ps
->reads
= pet_scop_collect_may_reads(scop
);
504 ps
->tagged_may_writes
= pet_scop_collect_tagged_may_writes(scop
);
505 ps
->may_writes
= pet_scop_collect_may_writes(scop
);
506 ps
->tagged_must_writes
= pet_scop_collect_tagged_must_writes(scop
);
507 ps
->must_writes
= pet_scop_collect_must_writes(scop
);
508 ps
->tagged_must_kills
= pet_scop_collect_tagged_must_kills(scop
);
509 ps
->schedule
= pet_scop_collect_schedule(scop
);
510 ps
->n_type
= scop
->n_type
;
511 ps
->types
= scop
->types
;
512 ps
->n_array
= scop
->n_array
;
513 ps
->arrays
= scop
->arrays
;
514 ps
->n_stmt
= scop
->n_stmt
;
515 ps
->stmts
= scop
->stmts
;
518 compute_dependences(ps
);
519 eliminate_dead_code(ps
);
521 if (!ps
->context
|| !ps
->domain
|| !ps
->call
|| !ps
->reads
||
522 !ps
->may_writes
|| !ps
->must_writes
|| !ps
->tagged_must_kills
||
524 return ppcg_scop_free(ps
);
529 /* Internal data structure for ppcg_transform.
531 struct ppcg_transform_data
{
532 struct ppcg_options
*options
;
533 __isl_give isl_printer
*(*transform
)(__isl_take isl_printer
*p
,
534 struct ppcg_scop
*scop
, void *user
);
538 /* Callback for pet_transform_C_source that transforms
539 * the given pet_scop to a ppcg_scop before calling the
540 * ppcg_transform callback.
542 static __isl_give isl_printer
*transform(__isl_take isl_printer
*p
,
543 struct pet_scop
*scop
, void *user
)
545 struct ppcg_transform_data
*data
= user
;
546 struct ppcg_scop
*ps
;
548 if (pet_scop_has_data_dependent_conditions(scop
)) {
549 p
= pet_scop_print_original(scop
, p
);
554 scop
= pet_scop_align_params(scop
);
555 ps
= ppcg_scop_from_pet_scop(scop
, data
->options
);
557 p
= data
->transform(p
, ps
, data
->user
);
565 /* Transform the C source file "input" by rewriting each scop
566 * through a call to "transform".
567 * The transformed C code is written to "out".
569 * This is a wrapper around pet_transform_C_source that transforms
570 * the pet_scop to a ppcg_scop before calling "fn".
572 int ppcg_transform(isl_ctx
*ctx
, const char *input
, FILE *out
,
573 struct ppcg_options
*options
,
574 __isl_give isl_printer
*(*fn
)(__isl_take isl_printer
*p
,
575 struct ppcg_scop
*scop
, void *user
), void *user
)
577 struct ppcg_transform_data data
= { options
, fn
, user
};
578 return pet_transform_C_source(ctx
, input
, out
, &transform
, &data
);
581 /* Check consistency of options.
583 * Return -1 on error.
585 static int check_options(isl_ctx
*ctx
)
587 struct options
*options
;
589 options
= isl_ctx_peek_options(ctx
, &options_args
);
591 isl_die(ctx
, isl_error_internal
,
592 "unable to find options", return -1);
594 if (options
->ppcg
->openmp
&&
595 !isl_options_get_ast_build_atomic_upper_bound(ctx
))
596 isl_die(ctx
, isl_error_invalid
,
597 "OpenMP requires atomic bounds", return -1);
602 int main(int argc
, char **argv
)
606 struct options
*options
;
608 options
= options_new_with_defaults();
611 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
612 isl_options_set_schedule_outer_coincidence(ctx
, 1);
613 isl_options_set_schedule_maximize_band_depth(ctx
, 1);
614 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
616 if (check_options(ctx
) < 0)
618 else if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
619 r
= generate_cuda(ctx
, options
->ppcg
, options
->input
);
621 r
= generate_cpu(ctx
, options
->ppcg
, options
->input
,