File: | polly/lib/External/isl/isl_scheduler.c |
Location: | line 2417, column 2 |
Description: | Value stored to 'nrow' is never read |
1 | /* |
2 | * Copyright 2011 INRIA Saclay |
3 | * Copyright 2012-2014 Ecole Normale Superieure |
4 | * Copyright 2015 Sven Verdoolaege |
5 | * |
6 | * Use of this software is governed by the MIT license |
7 | * |
8 | * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France, |
9 | * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod, |
10 | * 91893 Orsay, France |
11 | * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France |
12 | */ |
13 | |
14 | #include <isl_ctx_private.h> |
15 | #include <isl_map_private.h> |
16 | #include <isl_space_private.h> |
17 | #include <isl_aff_private.h> |
18 | #include <isl/hash.h> |
19 | #include <isl/constraint.h> |
20 | #include <isl/schedule.h> |
21 | #include <isl/schedule_node.h> |
22 | #include <isl_mat_private.h> |
23 | #include <isl_vec_private.h> |
24 | #include <isl/set.h> |
25 | #include <isl/union_set.h> |
26 | #include <isl_seq.h> |
27 | #include <isl_tab.h> |
28 | #include <isl_dim_map.h> |
29 | #include <isl/map_to_basic_set.h> |
30 | #include <isl_sort.h> |
31 | #include <isl_options_private.h> |
32 | #include <isl_tarjan.h> |
33 | #include <isl_morph.h> |
34 | |
35 | /* |
36 | * The scheduling algorithm implemented in this file was inspired by |
37 | * Bondhugula et al., "Automatic Transformations for Communication-Minimized |
38 | * Parallelization and Locality Optimization in the Polyhedral Model". |
39 | */ |
40 | |
41 | enum isl_edge_type { |
42 | isl_edge_validity = 0, |
43 | isl_edge_first = isl_edge_validity, |
44 | isl_edge_coincidence, |
45 | isl_edge_condition, |
46 | isl_edge_conditional_validity, |
47 | isl_edge_proximity, |
48 | isl_edge_last = isl_edge_proximity |
49 | }; |
50 | |
51 | /* The constraints that need to be satisfied by a schedule on "domain". |
52 | * |
53 | * "context" specifies extra constraints on the parameters. |
54 | * |
55 | * "validity" constraints map domain elements i to domain elements |
56 | * that should be scheduled after i. (Hard constraint) |
57 | * "proximity" constraints map domain elements i to domains elements |
58 | * that should be scheduled as early as possible after i (or before i). |
59 | * (Soft constraint) |
60 | * |
61 | * "condition" and "conditional_validity" constraints map possibly "tagged" |
62 | * domain elements i -> s to "tagged" domain elements j -> t. |
63 | * The elements of the "conditional_validity" constraints, but without the |
64 | * tags (i.e., the elements i -> j) are treated as validity constraints, |
65 | * except that during the construction of a tilable band, |
66 | * the elements of the "conditional_validity" constraints may be violated |
67 | * provided that all adjacent elements of the "condition" constraints |
68 | * are local within the band. |
69 | * A dependence is local within a band if domain and range are mapped |
70 | * to the same schedule point by the band. |
71 | */ |
72 | struct isl_schedule_constraints { |
73 | isl_union_set *domain; |
74 | isl_setisl_map *context; |
75 | |
76 | isl_union_map *constraint[isl_edge_last + 1]; |
77 | }; |
78 | |
79 | __isl_give isl_schedule_constraints *isl_schedule_constraints_copy( |
80 | __isl_keep isl_schedule_constraints *sc) |
81 | { |
82 | isl_ctx *ctx; |
83 | isl_schedule_constraints *sc_copy; |
84 | enum isl_edge_type i; |
85 | |
86 | ctx = isl_union_set_get_ctx(sc->domain); |
87 | sc_copy = isl_calloc_type(ctx, struct isl_schedule_constraints)((struct isl_schedule_constraints *)isl_calloc_or_die(ctx, 1, sizeof(struct isl_schedule_constraints))); |
88 | if (!sc_copy) |
89 | return NULL((void*)0); |
90 | |
91 | sc_copy->domain = isl_union_set_copy(sc->domain); |
92 | sc_copy->context = isl_set_copy(sc->context); |
93 | if (!sc_copy->domain || !sc_copy->context) |
94 | return isl_schedule_constraints_free(sc_copy); |
95 | |
96 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
97 | sc_copy->constraint[i] = isl_union_map_copy(sc->constraint[i]); |
98 | if (!sc_copy->constraint[i]) |
99 | return isl_schedule_constraints_free(sc_copy); |
100 | } |
101 | |
102 | return sc_copy; |
103 | } |
104 | |
105 | |
106 | /* Construct an isl_schedule_constraints object for computing a schedule |
107 | * on "domain". The initial object does not impose any constraints. |
108 | */ |
109 | __isl_give isl_schedule_constraints *isl_schedule_constraints_on_domain( |
110 | __isl_take isl_union_set *domain) |
111 | { |
112 | isl_ctx *ctx; |
113 | isl_space *space; |
114 | isl_schedule_constraints *sc; |
115 | isl_union_map *empty; |
116 | enum isl_edge_type i; |
117 | |
118 | if (!domain) |
119 | return NULL((void*)0); |
120 | |
121 | ctx = isl_union_set_get_ctx(domain); |
122 | sc = isl_calloc_type(ctx, struct isl_schedule_constraints)((struct isl_schedule_constraints *)isl_calloc_or_die(ctx, 1, sizeof(struct isl_schedule_constraints))); |
123 | if (!sc) |
124 | goto error; |
125 | |
126 | space = isl_union_set_get_space(domain); |
127 | sc->domain = domain; |
128 | sc->context = isl_set_universe(isl_space_copy(space)); |
129 | empty = isl_union_map_empty(space); |
130 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
131 | sc->constraint[i] = isl_union_map_copy(empty); |
132 | if (!sc->constraint[i]) |
133 | sc->domain = isl_union_set_free(sc->domain); |
134 | } |
135 | isl_union_map_free(empty); |
136 | |
137 | if (!sc->domain || !sc->context) |
138 | return isl_schedule_constraints_free(sc); |
139 | |
140 | return sc; |
141 | error: |
142 | isl_union_set_free(domain); |
143 | return NULL((void*)0); |
144 | } |
145 | |
146 | /* Replace the context of "sc" by "context". |
147 | */ |
148 | __isl_give isl_schedule_constraints *isl_schedule_constraints_set_context( |
149 | __isl_take isl_schedule_constraints *sc, __isl_take isl_setisl_map *context) |
150 | { |
151 | if (!sc || !context) |
152 | goto error; |
153 | |
154 | isl_set_free(sc->context); |
155 | sc->context = context; |
156 | |
157 | return sc; |
158 | error: |
159 | isl_schedule_constraints_free(sc); |
160 | isl_set_free(context); |
161 | return NULL((void*)0); |
162 | } |
163 | |
164 | /* Replace the validity constraints of "sc" by "validity". |
165 | */ |
166 | __isl_give isl_schedule_constraints *isl_schedule_constraints_set_validity( |
167 | __isl_take isl_schedule_constraints *sc, |
168 | __isl_take isl_union_map *validity) |
169 | { |
170 | if (!sc || !validity) |
171 | goto error; |
172 | |
173 | isl_union_map_free(sc->constraint[isl_edge_validity]); |
174 | sc->constraint[isl_edge_validity] = validity; |
175 | |
176 | return sc; |
177 | error: |
178 | isl_schedule_constraints_free(sc); |
179 | isl_union_map_free(validity); |
180 | return NULL((void*)0); |
181 | } |
182 | |
183 | /* Replace the coincidence constraints of "sc" by "coincidence". |
184 | */ |
185 | __isl_give isl_schedule_constraints *isl_schedule_constraints_set_coincidence( |
186 | __isl_take isl_schedule_constraints *sc, |
187 | __isl_take isl_union_map *coincidence) |
188 | { |
189 | if (!sc || !coincidence) |
190 | goto error; |
191 | |
192 | isl_union_map_free(sc->constraint[isl_edge_coincidence]); |
193 | sc->constraint[isl_edge_coincidence] = coincidence; |
194 | |
195 | return sc; |
196 | error: |
197 | isl_schedule_constraints_free(sc); |
198 | isl_union_map_free(coincidence); |
199 | return NULL((void*)0); |
200 | } |
201 | |
202 | /* Replace the proximity constraints of "sc" by "proximity". |
203 | */ |
204 | __isl_give isl_schedule_constraints *isl_schedule_constraints_set_proximity( |
205 | __isl_take isl_schedule_constraints *sc, |
206 | __isl_take isl_union_map *proximity) |
207 | { |
208 | if (!sc || !proximity) |
209 | goto error; |
210 | |
211 | isl_union_map_free(sc->constraint[isl_edge_proximity]); |
212 | sc->constraint[isl_edge_proximity] = proximity; |
213 | |
214 | return sc; |
215 | error: |
216 | isl_schedule_constraints_free(sc); |
217 | isl_union_map_free(proximity); |
218 | return NULL((void*)0); |
219 | } |
220 | |
221 | /* Replace the conditional validity constraints of "sc" by "condition" |
222 | * and "validity". |
223 | */ |
224 | __isl_give isl_schedule_constraints * |
225 | isl_schedule_constraints_set_conditional_validity( |
226 | __isl_take isl_schedule_constraints *sc, |
227 | __isl_take isl_union_map *condition, |
228 | __isl_take isl_union_map *validity) |
229 | { |
230 | if (!sc || !condition || !validity) |
231 | goto error; |
232 | |
233 | isl_union_map_free(sc->constraint[isl_edge_condition]); |
234 | sc->constraint[isl_edge_condition] = condition; |
235 | isl_union_map_free(sc->constraint[isl_edge_conditional_validity]); |
236 | sc->constraint[isl_edge_conditional_validity] = validity; |
237 | |
238 | return sc; |
239 | error: |
240 | isl_schedule_constraints_free(sc); |
241 | isl_union_map_free(condition); |
242 | isl_union_map_free(validity); |
243 | return NULL((void*)0); |
244 | } |
245 | |
246 | __isl_null isl_schedule_constraints *isl_schedule_constraints_free( |
247 | __isl_take isl_schedule_constraints *sc) |
248 | { |
249 | enum isl_edge_type i; |
250 | |
251 | if (!sc) |
252 | return NULL((void*)0); |
253 | |
254 | isl_union_set_free(sc->domain); |
255 | isl_set_free(sc->context); |
256 | for (i = isl_edge_first; i <= isl_edge_last; ++i) |
257 | isl_union_map_free(sc->constraint[i]); |
258 | |
259 | free(sc); |
260 | |
261 | return NULL((void*)0); |
262 | } |
263 | |
264 | isl_ctx *isl_schedule_constraints_get_ctx( |
265 | __isl_keep isl_schedule_constraints *sc) |
266 | { |
267 | return sc ? isl_union_set_get_ctx(sc->domain) : NULL((void*)0); |
268 | } |
269 | |
270 | /* Return the validity constraints of "sc". |
271 | */ |
272 | __isl_give isl_union_map *isl_schedule_constraints_get_validity( |
273 | __isl_keep isl_schedule_constraints *sc) |
274 | { |
275 | if (!sc) |
276 | return NULL((void*)0); |
277 | |
278 | return isl_union_map_copy(sc->constraint[isl_edge_validity]); |
279 | } |
280 | |
281 | /* Return the coincidence constraints of "sc". |
282 | */ |
283 | __isl_give isl_union_map *isl_schedule_constraints_get_coincidence( |
284 | __isl_keep isl_schedule_constraints *sc) |
285 | { |
286 | if (!sc) |
287 | return NULL((void*)0); |
288 | |
289 | return isl_union_map_copy(sc->constraint[isl_edge_coincidence]); |
290 | } |
291 | |
292 | /* Return the conditional validity constraints of "sc". |
293 | */ |
294 | __isl_give isl_union_map *isl_schedule_constraints_get_conditional_validity( |
295 | __isl_keep isl_schedule_constraints *sc) |
296 | { |
297 | if (!sc) |
298 | return NULL((void*)0); |
299 | |
300 | return |
301 | isl_union_map_copy(sc->constraint[isl_edge_conditional_validity]); |
302 | } |
303 | |
304 | /* Return the conditions for the conditional validity constraints of "sc". |
305 | */ |
306 | __isl_give isl_union_map * |
307 | isl_schedule_constraints_get_conditional_validity_condition( |
308 | __isl_keep isl_schedule_constraints *sc) |
309 | { |
310 | if (!sc) |
311 | return NULL((void*)0); |
312 | |
313 | return isl_union_map_copy(sc->constraint[isl_edge_condition]); |
314 | } |
315 | |
316 | void isl_schedule_constraints_dump(__isl_keep isl_schedule_constraints *sc) |
317 | { |
318 | if (!sc) |
319 | return; |
320 | |
321 | fprintf(stderr, "domain: ")__fprintf_chk (stderr, 2 - 1, "domain: "); |
322 | isl_union_set_dump(sc->domain); |
323 | fprintf(stderr, "context: ")__fprintf_chk (stderr, 2 - 1, "context: "); |
324 | isl_set_dump(sc->context); |
325 | fprintf(stderr, "validity: ")__fprintf_chk (stderr, 2 - 1, "validity: "); |
326 | isl_union_map_dump(sc->constraint[isl_edge_validity]); |
327 | fprintf(stderr, "proximity: ")__fprintf_chk (stderr, 2 - 1, "proximity: "); |
328 | isl_union_map_dump(sc->constraint[isl_edge_proximity]); |
329 | fprintf(stderr, "coincidence: ")__fprintf_chk (stderr, 2 - 1, "coincidence: "); |
330 | isl_union_map_dump(sc->constraint[isl_edge_coincidence]); |
331 | fprintf(stderr, "condition: ")__fprintf_chk (stderr, 2 - 1, "condition: "); |
332 | isl_union_map_dump(sc->constraint[isl_edge_condition]); |
333 | fprintf(stderr, "conditional_validity: ")__fprintf_chk (stderr, 2 - 1, "conditional_validity: "); |
334 | isl_union_map_dump(sc->constraint[isl_edge_conditional_validity]); |
335 | } |
336 | |
337 | /* Align the parameters of the fields of "sc". |
338 | */ |
339 | static __isl_give isl_schedule_constraints * |
340 | isl_schedule_constraints_align_params(__isl_take isl_schedule_constraints *sc) |
341 | { |
342 | isl_space *space; |
343 | enum isl_edge_type i; |
344 | |
345 | if (!sc) |
346 | return NULL((void*)0); |
347 | |
348 | space = isl_union_set_get_space(sc->domain); |
349 | space = isl_space_align_params(space, isl_set_get_space(sc->context)); |
350 | for (i = isl_edge_first; i <= isl_edge_last; ++i) |
351 | space = isl_space_align_params(space, |
352 | isl_union_map_get_space(sc->constraint[i])); |
353 | |
354 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
355 | sc->constraint[i] = isl_union_map_align_params( |
356 | sc->constraint[i], isl_space_copy(space)); |
357 | if (!sc->constraint[i]) |
358 | space = isl_space_free(space); |
359 | } |
360 | sc->context = isl_set_align_params(sc->context, isl_space_copy(space)); |
361 | sc->domain = isl_union_set_align_params(sc->domain, space); |
362 | if (!sc->context || !sc->domain) |
363 | return isl_schedule_constraints_free(sc); |
364 | |
365 | return sc; |
366 | } |
367 | |
368 | /* Return the total number of isl_maps in the constraints of "sc". |
369 | */ |
370 | static __isl_give int isl_schedule_constraints_n_map( |
371 | __isl_keep isl_schedule_constraints *sc) |
372 | { |
373 | enum isl_edge_type i; |
374 | int n = 0; |
375 | |
376 | for (i = isl_edge_first; i <= isl_edge_last; ++i) |
377 | n += isl_union_map_n_map(sc->constraint[i]); |
378 | |
379 | return n; |
380 | } |
381 | |
382 | /* Internal information about a node that is used during the construction |
383 | * of a schedule. |
384 | * space represents the space in which the domain lives |
385 | * sched is a matrix representation of the schedule being constructed |
386 | * for this node; if compressed is set, then this schedule is |
387 | * defined over the compressed domain space |
388 | * sched_map is an isl_map representation of the same (partial) schedule |
389 | * sched_map may be NULL; if compressed is set, then this map |
390 | * is defined over the uncompressed domain space |
391 | * rank is the number of linearly independent rows in the linear part |
392 | * of sched |
393 | * the columns of cmap represent a change of basis for the schedule |
394 | * coefficients; the first rank columns span the linear part of |
395 | * the schedule rows |
396 | * cinv is the inverse of cmap. |
397 | * start is the first variable in the LP problem in the sequences that |
398 | * represents the schedule coefficients of this node |
399 | * nvar is the dimension of the domain |
400 | * nparam is the number of parameters or 0 if we are not constructing |
401 | * a parametric schedule |
402 | * |
403 | * If compressed is set, then hull represents the constraints |
404 | * that were used to derive the compression, while compress and |
405 | * decompress map the original space to the compressed space and |
406 | * vice versa. |
407 | * |
408 | * scc is the index of SCC (or WCC) this node belongs to |
409 | * |
410 | * coincident contains a boolean for each of the rows of the schedule, |
411 | * indicating whether the corresponding scheduling dimension satisfies |
412 | * the coincidence constraints in the sense that the corresponding |
413 | * dependence distances are zero. |
414 | */ |
415 | struct isl_sched_node { |
416 | isl_space *space; |
417 | int compressed; |
418 | isl_setisl_map *hull; |
419 | isl_multi_aff *compress; |
420 | isl_multi_aff *decompress; |
421 | isl_mat *sched; |
422 | isl_map *sched_map; |
423 | int rank; |
424 | isl_mat *cmap; |
425 | isl_mat *cinv; |
426 | int start; |
427 | int nvar; |
428 | int nparam; |
429 | |
430 | int scc; |
431 | |
432 | int *coincident; |
433 | }; |
434 | |
435 | static int node_has_space(const void *entry, const void *val) |
436 | { |
437 | struct isl_sched_node *node = (struct isl_sched_node *)entry; |
438 | isl_space *dim = (isl_space *)val; |
439 | |
440 | return isl_space_is_equal(node->space, dim); |
441 | } |
442 | |
443 | static int node_scc_exactly(struct isl_sched_node *node, int scc) |
444 | { |
445 | return node->scc == scc; |
446 | } |
447 | |
448 | static int node_scc_at_most(struct isl_sched_node *node, int scc) |
449 | { |
450 | return node->scc <= scc; |
451 | } |
452 | |
453 | static int node_scc_at_least(struct isl_sched_node *node, int scc) |
454 | { |
455 | return node->scc >= scc; |
456 | } |
457 | |
458 | /* An edge in the dependence graph. An edge may be used to |
459 | * ensure validity of the generated schedule, to minimize the dependence |
460 | * distance or both |
461 | * |
462 | * map is the dependence relation, with i -> j in the map if j depends on i |
463 | * tagged_condition and tagged_validity contain the union of all tagged |
464 | * condition or conditional validity dependence relations that |
465 | * specialize the dependence relation "map"; that is, |
466 | * if (i -> a) -> (j -> b) is an element of "tagged_condition" |
467 | * or "tagged_validity", then i -> j is an element of "map". |
468 | * If these fields are NULL, then they represent the empty relation. |
469 | * src is the source node |
470 | * dst is the sink node |
471 | * validity is set if the edge is used to ensure correctness |
472 | * coincidence is used to enforce zero dependence distances |
473 | * proximity is set if the edge is used to minimize dependence distances |
474 | * condition is set if the edge represents a condition |
475 | * for a conditional validity schedule constraint |
476 | * local can only be set for condition edges and indicates that |
477 | * the dependence distance over the edge should be zero |
478 | * conditional_validity is set if the edge is used to conditionally |
479 | * ensure correctness |
480 | * |
481 | * For validity edges, start and end mark the sequence of inequality |
482 | * constraints in the LP problem that encode the validity constraint |
483 | * corresponding to this edge. |
484 | */ |
485 | struct isl_sched_edge { |
486 | isl_map *map; |
487 | isl_union_map *tagged_condition; |
488 | isl_union_map *tagged_validity; |
489 | |
490 | struct isl_sched_node *src; |
491 | struct isl_sched_node *dst; |
492 | |
493 | unsigned validity : 1; |
494 | unsigned coincidence : 1; |
495 | unsigned proximity : 1; |
496 | unsigned local : 1; |
497 | unsigned condition : 1; |
498 | unsigned conditional_validity : 1; |
499 | |
500 | int start; |
501 | int end; |
502 | }; |
503 | |
504 | /* Internal information about the dependence graph used during |
505 | * the construction of the schedule. |
506 | * |
507 | * intra_hmap is a cache, mapping dependence relations to their dual, |
508 | * for dependences from a node to itself |
509 | * inter_hmap is a cache, mapping dependence relations to their dual, |
510 | * for dependences between distinct nodes |
511 | * if compression is involved then the key for these maps |
512 | * it the original, uncompressed dependence relation, while |
513 | * the value is the dual of the compressed dependence relation. |
514 | * |
515 | * n is the number of nodes |
516 | * node is the list of nodes |
517 | * maxvar is the maximal number of variables over all nodes |
518 | * max_row is the allocated number of rows in the schedule |
519 | * n_row is the current (maximal) number of linearly independent |
520 | * rows in the node schedules |
521 | * n_total_row is the current number of rows in the node schedules |
522 | * band_start is the starting row in the node schedules of the current band |
523 | * root is set if this graph is the original dependence graph, |
524 | * without any splitting |
525 | * |
526 | * sorted contains a list of node indices sorted according to the |
527 | * SCC to which a node belongs |
528 | * |
529 | * n_edge is the number of edges |
530 | * edge is the list of edges |
531 | * max_edge contains the maximal number of edges of each type; |
532 | * in particular, it contains the number of edges in the inital graph. |
533 | * edge_table contains pointers into the edge array, hashed on the source |
534 | * and sink spaces; there is one such table for each type; |
535 | * a given edge may be referenced from more than one table |
536 | * if the corresponding relation appears in more than one of the |
537 | * sets of dependences |
538 | * |
539 | * node_table contains pointers into the node array, hashed on the space |
540 | * |
541 | * region contains a list of variable sequences that should be non-trivial |
542 | * |
543 | * lp contains the (I)LP problem used to obtain new schedule rows |
544 | * |
545 | * src_scc and dst_scc are the source and sink SCCs of an edge with |
546 | * conflicting constraints |
547 | * |
548 | * scc represents the number of components |
549 | * weak is set if the components are weakly connected |
550 | */ |
551 | struct isl_sched_graph { |
552 | isl_map_to_basic_set *intra_hmap; |
553 | isl_map_to_basic_set *inter_hmap; |
554 | |
555 | struct isl_sched_node *node; |
556 | int n; |
557 | int maxvar; |
558 | int max_row; |
559 | int n_row; |
560 | |
561 | int *sorted; |
562 | |
563 | int n_total_row; |
564 | int band_start; |
565 | |
566 | int root; |
567 | |
568 | struct isl_sched_edge *edge; |
569 | int n_edge; |
570 | int max_edge[isl_edge_last + 1]; |
571 | struct isl_hash_table *edge_table[isl_edge_last + 1]; |
572 | |
573 | struct isl_hash_table *node_table; |
574 | struct isl_region *region; |
575 | |
576 | isl_basic_setisl_basic_map *lp; |
577 | |
578 | int src_scc; |
579 | int dst_scc; |
580 | |
581 | int scc; |
582 | int weak; |
583 | }; |
584 | |
585 | /* Initialize node_table based on the list of nodes. |
586 | */ |
587 | static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph) |
588 | { |
589 | int i; |
590 | |
591 | graph->node_table = isl_hash_table_alloc(ctx, graph->n); |
592 | if (!graph->node_table) |
593 | return -1; |
594 | |
595 | for (i = 0; i < graph->n; ++i) { |
596 | struct isl_hash_table_entry *entry; |
597 | uint32_t hash; |
598 | |
599 | hash = isl_space_get_hash(graph->node[i].space); |
600 | entry = isl_hash_table_find(ctx, graph->node_table, hash, |
601 | &node_has_space, |
602 | graph->node[i].space, 1); |
603 | if (!entry) |
604 | return -1; |
605 | entry->data = &graph->node[i]; |
606 | } |
607 | |
608 | return 0; |
609 | } |
610 | |
611 | /* Return a pointer to the node that lives within the given space, |
612 | * or NULL if there is no such node. |
613 | */ |
614 | static struct isl_sched_node *graph_find_node(isl_ctx *ctx, |
615 | struct isl_sched_graph *graph, __isl_keep isl_space *dim) |
616 | { |
617 | struct isl_hash_table_entry *entry; |
618 | uint32_t hash; |
619 | |
620 | hash = isl_space_get_hash(dim); |
621 | entry = isl_hash_table_find(ctx, graph->node_table, hash, |
622 | &node_has_space, dim, 0); |
623 | |
624 | return entry ? entry->data : NULL((void*)0); |
625 | } |
626 | |
627 | static int edge_has_src_and_dst(const void *entry, const void *val) |
628 | { |
629 | const struct isl_sched_edge *edge = entry; |
630 | const struct isl_sched_edge *temp = val; |
631 | |
632 | return edge->src == temp->src && edge->dst == temp->dst; |
633 | } |
634 | |
635 | /* Add the given edge to graph->edge_table[type]. |
636 | */ |
637 | static isl_stat graph_edge_table_add(isl_ctx *ctx, |
638 | struct isl_sched_graph *graph, enum isl_edge_type type, |
639 | struct isl_sched_edge *edge) |
640 | { |
641 | struct isl_hash_table_entry *entry; |
642 | uint32_t hash; |
643 | |
644 | hash = isl_hash_init()(2166136261u); |
645 | hash = isl_hash_builtin(hash, edge->src)isl_hash_mem(hash, &edge->src, sizeof(edge->src)); |
646 | hash = isl_hash_builtin(hash, edge->dst)isl_hash_mem(hash, &edge->dst, sizeof(edge->dst)); |
647 | entry = isl_hash_table_find(ctx, graph->edge_table[type], hash, |
648 | &edge_has_src_and_dst, edge, 1); |
649 | if (!entry) |
650 | return isl_stat_error; |
651 | entry->data = edge; |
652 | |
653 | return isl_stat_ok; |
654 | } |
655 | |
656 | /* Allocate the edge_tables based on the maximal number of edges of |
657 | * each type. |
658 | */ |
659 | static int graph_init_edge_tables(isl_ctx *ctx, struct isl_sched_graph *graph) |
660 | { |
661 | int i; |
662 | |
663 | for (i = 0; i <= isl_edge_last; ++i) { |
664 | graph->edge_table[i] = isl_hash_table_alloc(ctx, |
665 | graph->max_edge[i]); |
666 | if (!graph->edge_table[i]) |
667 | return -1; |
668 | } |
669 | |
670 | return 0; |
671 | } |
672 | |
673 | /* If graph->edge_table[type] contains an edge from the given source |
674 | * to the given destination, then return the hash table entry of this edge. |
675 | * Otherwise, return NULL. |
676 | */ |
677 | static struct isl_hash_table_entry *graph_find_edge_entry( |
678 | struct isl_sched_graph *graph, |
679 | enum isl_edge_type type, |
680 | struct isl_sched_node *src, struct isl_sched_node *dst) |
681 | { |
682 | isl_ctx *ctx = isl_space_get_ctx(src->space); |
683 | uint32_t hash; |
684 | struct isl_sched_edge temp = { .src = src, .dst = dst }; |
685 | |
686 | hash = isl_hash_init()(2166136261u); |
687 | hash = isl_hash_builtin(hash, temp.src)isl_hash_mem(hash, &temp.src, sizeof(temp.src)); |
688 | hash = isl_hash_builtin(hash, temp.dst)isl_hash_mem(hash, &temp.dst, sizeof(temp.dst)); |
689 | return isl_hash_table_find(ctx, graph->edge_table[type], hash, |
690 | &edge_has_src_and_dst, &temp, 0); |
691 | } |
692 | |
693 | |
694 | /* If graph->edge_table[type] contains an edge from the given source |
695 | * to the given destination, then return this edge. |
696 | * Otherwise, return NULL. |
697 | */ |
698 | static struct isl_sched_edge *graph_find_edge(struct isl_sched_graph *graph, |
699 | enum isl_edge_type type, |
700 | struct isl_sched_node *src, struct isl_sched_node *dst) |
701 | { |
702 | struct isl_hash_table_entry *entry; |
703 | |
704 | entry = graph_find_edge_entry(graph, type, src, dst); |
705 | if (!entry) |
706 | return NULL((void*)0); |
707 | |
708 | return entry->data; |
709 | } |
710 | |
711 | /* Check whether the dependence graph has an edge of the given type |
712 | * between the given two nodes. |
713 | */ |
714 | static isl_bool graph_has_edge(struct isl_sched_graph *graph, |
715 | enum isl_edge_type type, |
716 | struct isl_sched_node *src, struct isl_sched_node *dst) |
717 | { |
718 | struct isl_sched_edge *edge; |
719 | isl_bool empty; |
720 | |
721 | edge = graph_find_edge(graph, type, src, dst); |
722 | if (!edge) |
723 | return 0; |
724 | |
725 | empty = isl_map_plain_is_empty(edge->map); |
726 | if (empty < 0) |
727 | return isl_bool_error; |
728 | |
729 | return !empty; |
730 | } |
731 | |
732 | /* Look for any edge with the same src, dst and map fields as "model". |
733 | * |
734 | * Return the matching edge if one can be found. |
735 | * Return "model" if no matching edge is found. |
736 | * Return NULL on error. |
737 | */ |
738 | static struct isl_sched_edge *graph_find_matching_edge( |
739 | struct isl_sched_graph *graph, struct isl_sched_edge *model) |
740 | { |
741 | enum isl_edge_type i; |
742 | struct isl_sched_edge *edge; |
743 | |
744 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
745 | int is_equal; |
746 | |
747 | edge = graph_find_edge(graph, i, model->src, model->dst); |
748 | if (!edge) |
749 | continue; |
750 | is_equal = isl_map_plain_is_equal(model->map, edge->map); |
751 | if (is_equal < 0) |
752 | return NULL((void*)0); |
753 | if (is_equal) |
754 | return edge; |
755 | } |
756 | |
757 | return model; |
758 | } |
759 | |
760 | /* Remove the given edge from all the edge_tables that refer to it. |
761 | */ |
762 | static void graph_remove_edge(struct isl_sched_graph *graph, |
763 | struct isl_sched_edge *edge) |
764 | { |
765 | isl_ctx *ctx = isl_map_get_ctx(edge->map); |
766 | enum isl_edge_type i; |
767 | |
768 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
769 | struct isl_hash_table_entry *entry; |
770 | |
771 | entry = graph_find_edge_entry(graph, i, edge->src, edge->dst); |
772 | if (!entry) |
773 | continue; |
774 | if (entry->data != edge) |
775 | continue; |
776 | isl_hash_table_remove(ctx, graph->edge_table[i], entry); |
777 | } |
778 | } |
779 | |
780 | /* Check whether the dependence graph has any edge |
781 | * between the given two nodes. |
782 | */ |
783 | static isl_bool graph_has_any_edge(struct isl_sched_graph *graph, |
784 | struct isl_sched_node *src, struct isl_sched_node *dst) |
785 | { |
786 | enum isl_edge_type i; |
787 | isl_bool r; |
788 | |
789 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
790 | r = graph_has_edge(graph, i, src, dst); |
791 | if (r < 0 || r) |
792 | return r; |
793 | } |
794 | |
795 | return r; |
796 | } |
797 | |
798 | /* Check whether the dependence graph has a validity edge |
799 | * between the given two nodes. |
800 | * |
801 | * Conditional validity edges are essentially validity edges that |
802 | * can be ignored if the corresponding condition edges are iteration private. |
803 | * Here, we are only checking for the presence of validity |
804 | * edges, so we need to consider the conditional validity edges too. |
805 | * In particular, this function is used during the detection |
806 | * of strongly connected components and we cannot ignore |
807 | * conditional validity edges during this detection. |
808 | */ |
809 | static isl_bool graph_has_validity_edge(struct isl_sched_graph *graph, |
810 | struct isl_sched_node *src, struct isl_sched_node *dst) |
811 | { |
812 | isl_bool r; |
813 | |
814 | r = graph_has_edge(graph, isl_edge_validity, src, dst); |
815 | if (r < 0 || r) |
816 | return r; |
817 | |
818 | return graph_has_edge(graph, isl_edge_conditional_validity, src, dst); |
819 | } |
820 | |
821 | static int graph_alloc(isl_ctx *ctx, struct isl_sched_graph *graph, |
822 | int n_node, int n_edge) |
823 | { |
824 | int i; |
825 | |
826 | graph->n = n_node; |
827 | graph->n_edge = n_edge; |
828 | graph->node = isl_calloc_array(ctx, struct isl_sched_node, graph->n)((struct isl_sched_node *)isl_calloc_or_die(ctx, graph->n, sizeof(struct isl_sched_node))); |
829 | graph->sorted = isl_calloc_array(ctx, int, graph->n)((int *)isl_calloc_or_die(ctx, graph->n, sizeof(int))); |
830 | graph->region = isl_alloc_array(ctx, struct isl_region, graph->n)((struct isl_region *)isl_malloc_or_die(ctx, (graph->n)*sizeof (struct isl_region))); |
831 | graph->edge = isl_calloc_array(ctx,((struct isl_sched_edge *)isl_calloc_or_die(ctx, graph->n_edge , sizeof(struct isl_sched_edge))) |
832 | struct isl_sched_edge, graph->n_edge)((struct isl_sched_edge *)isl_calloc_or_die(ctx, graph->n_edge , sizeof(struct isl_sched_edge))); |
833 | |
834 | graph->intra_hmap = isl_map_to_basic_set_alloc(ctx, 2 * n_edge); |
835 | graph->inter_hmap = isl_map_to_basic_set_alloc(ctx, 2 * n_edge); |
836 | |
837 | if (!graph->node || !graph->region || (graph->n_edge && !graph->edge) || |
838 | !graph->sorted) |
839 | return -1; |
840 | |
841 | for(i = 0; i < graph->n; ++i) |
842 | graph->sorted[i] = i; |
843 | |
844 | return 0; |
845 | } |
846 | |
847 | static void graph_free(isl_ctx *ctx, struct isl_sched_graph *graph) |
848 | { |
849 | int i; |
850 | |
851 | isl_map_to_basic_set_free(graph->intra_hmap); |
852 | isl_map_to_basic_set_free(graph->inter_hmap); |
853 | |
854 | if (graph->node) |
855 | for (i = 0; i < graph->n; ++i) { |
856 | isl_space_free(graph->node[i].space); |
857 | isl_set_free(graph->node[i].hull); |
858 | isl_multi_aff_free(graph->node[i].compress); |
859 | isl_multi_aff_free(graph->node[i].decompress); |
860 | isl_mat_free(graph->node[i].sched); |
861 | isl_map_free(graph->node[i].sched_map); |
862 | isl_mat_free(graph->node[i].cmap); |
863 | isl_mat_free(graph->node[i].cinv); |
864 | if (graph->root) |
865 | free(graph->node[i].coincident); |
866 | } |
867 | free(graph->node); |
868 | free(graph->sorted); |
869 | if (graph->edge) |
870 | for (i = 0; i < graph->n_edge; ++i) { |
871 | isl_map_free(graph->edge[i].map); |
872 | isl_union_map_free(graph->edge[i].tagged_condition); |
873 | isl_union_map_free(graph->edge[i].tagged_validity); |
874 | } |
875 | free(graph->edge); |
876 | free(graph->region); |
877 | for (i = 0; i <= isl_edge_last; ++i) |
878 | isl_hash_table_free(ctx, graph->edge_table[i]); |
879 | isl_hash_table_free(ctx, graph->node_table); |
880 | isl_basic_set_free(graph->lp); |
881 | } |
882 | |
883 | /* For each "set" on which this function is called, increment |
884 | * graph->n by one and update graph->maxvar. |
885 | */ |
886 | static isl_stat init_n_maxvar(__isl_take isl_setisl_map *set, void *user) |
887 | { |
888 | struct isl_sched_graph *graph = user; |
889 | int nvar = isl_set_dim(set, isl_dim_set); |
890 | |
891 | graph->n++; |
892 | if (nvar > graph->maxvar) |
893 | graph->maxvar = nvar; |
894 | |
895 | isl_set_free(set); |
896 | |
897 | return isl_stat_ok; |
898 | } |
899 | |
900 | /* Add the number of basic maps in "map" to *n. |
901 | */ |
902 | static isl_stat add_n_basic_map(__isl_take isl_map *map, void *user) |
903 | { |
904 | int *n = user; |
905 | |
906 | *n += isl_map_n_basic_map(map); |
907 | isl_map_free(map); |
908 | |
909 | return isl_stat_ok; |
910 | } |
911 | |
912 | /* Compute the number of rows that should be allocated for the schedule. |
913 | * In particular, we need one row for each variable or one row |
914 | * for each basic map in the dependences. |
915 | * Note that it is practically impossible to exhaust both |
916 | * the number of dependences and the number of variables. |
917 | */ |
918 | static int compute_max_row(struct isl_sched_graph *graph, |
919 | __isl_keep isl_schedule_constraints *sc) |
920 | { |
921 | enum isl_edge_type i; |
922 | int n_edge; |
923 | |
924 | graph->n = 0; |
925 | graph->maxvar = 0; |
926 | if (isl_union_set_foreach_set(sc->domain, &init_n_maxvar, graph) < 0) |
927 | return -1; |
928 | n_edge = 0; |
929 | for (i = isl_edge_first; i <= isl_edge_last; ++i) |
930 | if (isl_union_map_foreach_map(sc->constraint[i], |
931 | &add_n_basic_map, &n_edge) < 0) |
932 | return -1; |
933 | graph->max_row = n_edge + graph->maxvar; |
934 | |
935 | return 0; |
936 | } |
937 | |
938 | /* Does "bset" have any defining equalities for its set variables? |
939 | */ |
940 | static int has_any_defining_equality(__isl_keep isl_basic_setisl_basic_map *bset) |
941 | { |
942 | int i, n; |
943 | |
944 | if (!bset) |
945 | return -1; |
946 | |
947 | n = isl_basic_set_dim(bset, isl_dim_set); |
948 | for (i = 0; i < n; ++i) { |
949 | int has; |
950 | |
951 | has = isl_basic_set_has_defining_equality(bset, isl_dim_set, i, |
952 | NULL((void*)0)); |
953 | if (has < 0 || has) |
954 | return has; |
955 | } |
956 | |
957 | return 0; |
958 | } |
959 | |
960 | /* Add a new node to the graph representing the given space. |
961 | * "nvar" is the (possibly compressed) number of variables and |
962 | * may be smaller than then number of set variables in "space" |
963 | * if "compressed" is set. |
964 | * If "compressed" is set, then "hull" represents the constraints |
965 | * that were used to derive the compression, while "compress" and |
966 | * "decompress" map the original space to the compressed space and |
967 | * vice versa. |
968 | * If "compressed" is not set, then "hull", "compress" and "decompress" |
969 | * should be NULL. |
970 | */ |
971 | static isl_stat add_node(struct isl_sched_graph *graph, |
972 | __isl_take isl_space *space, int nvar, int compressed, |
973 | __isl_take isl_setisl_map *hull, __isl_take isl_multi_aff *compress, |
974 | __isl_take isl_multi_aff *decompress) |
975 | { |
976 | int nparam; |
977 | isl_ctx *ctx; |
978 | isl_mat *sched; |
979 | int *coincident; |
980 | |
981 | if (!space) |
982 | return isl_stat_error; |
983 | |
984 | ctx = isl_space_get_ctx(space); |
985 | nparam = isl_space_dim(space, isl_dim_param); |
986 | if (!ctx->opt->schedule_parametric) |
987 | nparam = 0; |
988 | sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar); |
989 | graph->node[graph->n].space = space; |
990 | graph->node[graph->n].nvar = nvar; |
991 | graph->node[graph->n].nparam = nparam; |
992 | graph->node[graph->n].sched = sched; |
993 | graph->node[graph->n].sched_map = NULL((void*)0); |
994 | coincident = isl_calloc_array(ctx, int, graph->max_row)((int *)isl_calloc_or_die(ctx, graph->max_row, sizeof(int) )); |
995 | graph->node[graph->n].coincident = coincident; |
996 | graph->node[graph->n].compressed = compressed; |
997 | graph->node[graph->n].hull = hull; |
998 | graph->node[graph->n].compress = compress; |
999 | graph->node[graph->n].decompress = decompress; |
1000 | graph->n++; |
1001 | |
1002 | if (!space || !sched || (graph->max_row && !coincident)) |
1003 | return isl_stat_error; |
1004 | if (compressed && (!hull || !compress || !decompress)) |
1005 | return isl_stat_error; |
1006 | |
1007 | return isl_stat_ok; |
1008 | } |
1009 | |
1010 | /* Add a new node to the graph representing the given set. |
1011 | * |
1012 | * If any of the set variables is defined by an equality, then |
1013 | * we perform variable compression such that we can perform |
1014 | * the scheduling on the compressed domain. |
1015 | */ |
1016 | static isl_stat extract_node(__isl_take isl_setisl_map *set, void *user) |
1017 | { |
1018 | int nvar; |
1019 | int has_equality; |
1020 | isl_space *space; |
1021 | isl_basic_setisl_basic_map *hull; |
1022 | isl_setisl_map *hull_set; |
1023 | isl_morph *morph; |
1024 | isl_multi_aff *compress, *decompress; |
1025 | struct isl_sched_graph *graph = user; |
1026 | |
1027 | space = isl_set_get_space(set); |
1028 | hull = isl_set_affine_hull(set); |
1029 | hull = isl_basic_set_remove_divs(hull); |
1030 | nvar = isl_space_dim(space, isl_dim_set); |
1031 | has_equality = has_any_defining_equality(hull); |
1032 | |
1033 | if (has_equality < 0) |
1034 | goto error; |
1035 | if (!has_equality) { |
1036 | isl_basic_set_free(hull); |
1037 | return add_node(graph, space, nvar, 0, NULL((void*)0), NULL((void*)0), NULL((void*)0)); |
1038 | } |
1039 | |
1040 | morph = isl_basic_set_variable_compression(hull, isl_dim_set); |
1041 | nvar = isl_morph_ran_dim(morph, isl_dim_set); |
1042 | compress = isl_morph_get_var_multi_aff(morph); |
1043 | morph = isl_morph_inverse(morph); |
1044 | decompress = isl_morph_get_var_multi_aff(morph); |
1045 | isl_morph_free(morph); |
1046 | |
1047 | hull_set = isl_set_from_basic_set(hull); |
1048 | return add_node(graph, space, nvar, 1, hull_set, compress, decompress); |
1049 | error: |
1050 | isl_basic_set_free(hull); |
1051 | isl_space_free(space); |
1052 | return isl_stat_error; |
1053 | } |
1054 | |
1055 | struct isl_extract_edge_data { |
1056 | enum isl_edge_type type; |
1057 | struct isl_sched_graph *graph; |
1058 | }; |
1059 | |
1060 | /* Merge edge2 into edge1, freeing the contents of edge2. |
1061 | * "type" is the type of the schedule constraint from which edge2 was |
1062 | * extracted. |
1063 | * Return 0 on success and -1 on failure. |
1064 | * |
1065 | * edge1 and edge2 are assumed to have the same value for the map field. |
1066 | */ |
1067 | static int merge_edge(enum isl_edge_type type, struct isl_sched_edge *edge1, |
1068 | struct isl_sched_edge *edge2) |
1069 | { |
1070 | edge1->validity |= edge2->validity; |
1071 | edge1->coincidence |= edge2->coincidence; |
1072 | edge1->proximity |= edge2->proximity; |
1073 | edge1->condition |= edge2->condition; |
1074 | edge1->conditional_validity |= edge2->conditional_validity; |
1075 | isl_map_free(edge2->map); |
1076 | |
1077 | if (type == isl_edge_condition) { |
1078 | if (!edge1->tagged_condition) |
1079 | edge1->tagged_condition = edge2->tagged_condition; |
1080 | else |
1081 | edge1->tagged_condition = |
1082 | isl_union_map_union(edge1->tagged_condition, |
1083 | edge2->tagged_condition); |
1084 | } |
1085 | |
1086 | if (type == isl_edge_conditional_validity) { |
1087 | if (!edge1->tagged_validity) |
1088 | edge1->tagged_validity = edge2->tagged_validity; |
1089 | else |
1090 | edge1->tagged_validity = |
1091 | isl_union_map_union(edge1->tagged_validity, |
1092 | edge2->tagged_validity); |
1093 | } |
1094 | |
1095 | if (type == isl_edge_condition && !edge1->tagged_condition) |
1096 | return -1; |
1097 | if (type == isl_edge_conditional_validity && !edge1->tagged_validity) |
1098 | return -1; |
1099 | |
1100 | return 0; |
1101 | } |
1102 | |
1103 | /* Insert dummy tags in domain and range of "map". |
1104 | * |
1105 | * In particular, if "map" is of the form |
1106 | * |
1107 | * A -> B |
1108 | * |
1109 | * then return |
1110 | * |
1111 | * [A -> dummy_tag] -> [B -> dummy_tag] |
1112 | * |
1113 | * where the dummy_tags are identical and equal to any dummy tags |
1114 | * introduced by any other call to this function. |
1115 | */ |
1116 | static __isl_give isl_map *insert_dummy_tags(__isl_take isl_map *map) |
1117 | { |
1118 | static char dummy; |
1119 | isl_ctx *ctx; |
1120 | isl_id *id; |
1121 | isl_space *space; |
1122 | isl_setisl_map *domain, *range; |
1123 | |
1124 | ctx = isl_map_get_ctx(map); |
1125 | |
1126 | id = isl_id_alloc(ctx, NULL((void*)0), &dummy); |
1127 | space = isl_space_params(isl_map_get_space(map)); |
1128 | space = isl_space_set_from_params(space); |
1129 | space = isl_space_set_tuple_id(space, isl_dim_set, id); |
1130 | space = isl_space_map_from_set(space); |
1131 | |
1132 | domain = isl_map_wrap(map); |
1133 | range = isl_map_wrap(isl_map_universe(space)); |
1134 | map = isl_map_from_domain_and_range(domain, range); |
1135 | map = isl_map_zip(map); |
1136 | |
1137 | return map; |
1138 | } |
1139 | |
1140 | /* Given that at least one of "src" or "dst" is compressed, return |
1141 | * a map between the spaces of these nodes restricted to the affine |
1142 | * hull that was used in the compression. |
1143 | */ |
1144 | static __isl_give isl_map *extract_hull(struct isl_sched_node *src, |
1145 | struct isl_sched_node *dst) |
1146 | { |
1147 | isl_setisl_map *dom, *ran; |
1148 | |
1149 | if (src->compressed) |
1150 | dom = isl_set_copy(src->hull); |
1151 | else |
1152 | dom = isl_set_universe(isl_space_copy(src->space)); |
1153 | if (dst->compressed) |
1154 | ran = isl_set_copy(dst->hull); |
1155 | else |
1156 | ran = isl_set_universe(isl_space_copy(dst->space)); |
1157 | |
1158 | return isl_map_from_domain_and_range(dom, ran); |
1159 | } |
1160 | |
1161 | /* Intersect the domains of the nested relations in domain and range |
1162 | * of "tagged" with "map". |
1163 | */ |
1164 | static __isl_give isl_map *map_intersect_domains(__isl_take isl_map *tagged, |
1165 | __isl_keep isl_map *map) |
1166 | { |
1167 | isl_setisl_map *set; |
1168 | |
1169 | tagged = isl_map_zip(tagged); |
1170 | set = isl_map_wrap(isl_map_copy(map)); |
1171 | tagged = isl_map_intersect_domain(tagged, set); |
1172 | tagged = isl_map_zip(tagged); |
1173 | return tagged; |
1174 | } |
1175 | |
1176 | /* Add a new edge to the graph based on the given map |
1177 | * and add it to data->graph->edge_table[data->type]. |
1178 | * If a dependence relation of a given type happens to be identical |
1179 | * to one of the dependence relations of a type that was added before, |
1180 | * then we don't create a new edge, but instead mark the original edge |
1181 | * as also representing a dependence of the current type. |
1182 | * |
1183 | * Edges of type isl_edge_condition or isl_edge_conditional_validity |
1184 | * may be specified as "tagged" dependence relations. That is, "map" |
1185 | * may contain elements (i -> a) -> (j -> b), where i -> j denotes |
1186 | * the dependence on iterations and a and b are tags. |
1187 | * edge->map is set to the relation containing the elements i -> j, |
1188 | * while edge->tagged_condition and edge->tagged_validity contain |
1189 | * the union of all the "map" relations |
1190 | * for which extract_edge is called that result in the same edge->map. |
1191 | * |
1192 | * If the source or the destination node is compressed, then |
1193 | * intersect both "map" and "tagged" with the constraints that |
1194 | * were used to construct the compression. |
1195 | * This ensures that there are no schedule constraints defined |
1196 | * outside of these domains, while the scheduler no longer has |
1197 | * any control over those outside parts. |
1198 | */ |
1199 | static isl_stat extract_edge(__isl_take isl_map *map, void *user) |
1200 | { |
1201 | isl_ctx *ctx = isl_map_get_ctx(map); |
1202 | struct isl_extract_edge_data *data = user; |
1203 | struct isl_sched_graph *graph = data->graph; |
1204 | struct isl_sched_node *src, *dst; |
1205 | isl_space *dim; |
1206 | struct isl_sched_edge *edge; |
1207 | isl_map *tagged = NULL((void*)0); |
1208 | |
1209 | if (data->type == isl_edge_condition || |
1210 | data->type == isl_edge_conditional_validity) { |
1211 | if (isl_map_can_zip(map)) { |
1212 | tagged = isl_map_copy(map); |
1213 | map = isl_set_unwrap(isl_map_domain(isl_map_zip(map))); |
1214 | } else { |
1215 | tagged = insert_dummy_tags(isl_map_copy(map)); |
1216 | } |
1217 | } |
1218 | |
1219 | dim = isl_space_domain(isl_map_get_space(map)); |
1220 | src = graph_find_node(ctx, graph, dim); |
1221 | isl_space_free(dim); |
1222 | dim = isl_space_range(isl_map_get_space(map)); |
1223 | dst = graph_find_node(ctx, graph, dim); |
1224 | isl_space_free(dim); |
1225 | |
1226 | if (!src || !dst) { |
1227 | isl_map_free(map); |
1228 | isl_map_free(tagged); |
1229 | return isl_stat_ok; |
1230 | } |
1231 | |
1232 | if (src->compressed || dst->compressed) { |
1233 | isl_map *hull; |
1234 | hull = extract_hull(src, dst); |
1235 | if (tagged) |
1236 | tagged = map_intersect_domains(tagged, hull); |
1237 | map = isl_map_intersect(map, hull); |
1238 | } |
1239 | |
1240 | graph->edge[graph->n_edge].src = src; |
1241 | graph->edge[graph->n_edge].dst = dst; |
1242 | graph->edge[graph->n_edge].map = map; |
1243 | graph->edge[graph->n_edge].validity = 0; |
1244 | graph->edge[graph->n_edge].coincidence = 0; |
1245 | graph->edge[graph->n_edge].proximity = 0; |
1246 | graph->edge[graph->n_edge].condition = 0; |
1247 | graph->edge[graph->n_edge].local = 0; |
1248 | graph->edge[graph->n_edge].conditional_validity = 0; |
1249 | graph->edge[graph->n_edge].tagged_condition = NULL((void*)0); |
1250 | graph->edge[graph->n_edge].tagged_validity = NULL((void*)0); |
1251 | if (data->type == isl_edge_validity) |
1252 | graph->edge[graph->n_edge].validity = 1; |
1253 | if (data->type == isl_edge_coincidence) |
1254 | graph->edge[graph->n_edge].coincidence = 1; |
1255 | if (data->type == isl_edge_proximity) |
1256 | graph->edge[graph->n_edge].proximity = 1; |
1257 | if (data->type == isl_edge_condition) { |
1258 | graph->edge[graph->n_edge].condition = 1; |
1259 | graph->edge[graph->n_edge].tagged_condition = |
1260 | isl_union_map_from_map(tagged); |
1261 | } |
1262 | if (data->type == isl_edge_conditional_validity) { |
1263 | graph->edge[graph->n_edge].conditional_validity = 1; |
1264 | graph->edge[graph->n_edge].tagged_validity = |
1265 | isl_union_map_from_map(tagged); |
1266 | } |
1267 | |
1268 | edge = graph_find_matching_edge(graph, &graph->edge[graph->n_edge]); |
1269 | if (!edge) { |
1270 | graph->n_edge++; |
1271 | return isl_stat_error; |
1272 | } |
1273 | if (edge == &graph->edge[graph->n_edge]) |
1274 | return graph_edge_table_add(ctx, graph, data->type, |
1275 | &graph->edge[graph->n_edge++]); |
1276 | |
1277 | if (merge_edge(data->type, edge, &graph->edge[graph->n_edge]) < 0) |
1278 | return -1; |
1279 | |
1280 | return graph_edge_table_add(ctx, graph, data->type, edge); |
1281 | } |
1282 | |
1283 | /* Check whether there is any dependence from node[j] to node[i] |
1284 | * or from node[i] to node[j]. |
1285 | */ |
1286 | static isl_bool node_follows_weak(int i, int j, void *user) |
1287 | { |
1288 | isl_bool f; |
1289 | struct isl_sched_graph *graph = user; |
1290 | |
1291 | f = graph_has_any_edge(graph, &graph->node[j], &graph->node[i]); |
1292 | if (f < 0 || f) |
1293 | return f; |
1294 | return graph_has_any_edge(graph, &graph->node[i], &graph->node[j]); |
1295 | } |
1296 | |
1297 | /* Check whether there is a (conditional) validity dependence from node[j] |
1298 | * to node[i], forcing node[i] to follow node[j]. |
1299 | */ |
1300 | static isl_bool node_follows_strong(int i, int j, void *user) |
1301 | { |
1302 | struct isl_sched_graph *graph = user; |
1303 | |
1304 | return graph_has_validity_edge(graph, &graph->node[j], &graph->node[i]); |
1305 | } |
1306 | |
1307 | /* Use Tarjan's algorithm for computing the strongly connected components |
1308 | * in the dependence graph (only validity edges). |
1309 | * If weak is set, we consider the graph to be undirected and |
1310 | * we effectively compute the (weakly) connected components. |
1311 | * Additionally, we also consider other edges when weak is set. |
1312 | */ |
1313 | static int detect_ccs(isl_ctx *ctx, struct isl_sched_graph *graph, int weak) |
1314 | { |
1315 | int i, n; |
1316 | struct isl_tarjan_graph *g = NULL((void*)0); |
1317 | |
1318 | g = isl_tarjan_graph_init(ctx, graph->n, |
1319 | weak ? &node_follows_weak : &node_follows_strong, graph); |
1320 | if (!g) |
1321 | return -1; |
1322 | |
1323 | graph->weak = weak; |
1324 | graph->scc = 0; |
1325 | i = 0; |
1326 | n = graph->n; |
1327 | while (n) { |
1328 | while (g->order[i] != -1) { |
1329 | graph->node[g->order[i]].scc = graph->scc; |
1330 | --n; |
1331 | ++i; |
1332 | } |
1333 | ++i; |
1334 | graph->scc++; |
1335 | } |
1336 | |
1337 | isl_tarjan_graph_free(g); |
1338 | |
1339 | return 0; |
1340 | } |
1341 | |
1342 | /* Apply Tarjan's algorithm to detect the strongly connected components |
1343 | * in the dependence graph. |
1344 | */ |
1345 | static int detect_sccs(isl_ctx *ctx, struct isl_sched_graph *graph) |
1346 | { |
1347 | return detect_ccs(ctx, graph, 0); |
1348 | } |
1349 | |
1350 | /* Apply Tarjan's algorithm to detect the (weakly) connected components |
1351 | * in the dependence graph. |
1352 | */ |
1353 | static int detect_wccs(isl_ctx *ctx, struct isl_sched_graph *graph) |
1354 | { |
1355 | return detect_ccs(ctx, graph, 1); |
1356 | } |
1357 | |
1358 | static int cmp_scc(const void *a, const void *b, void *data) |
1359 | { |
1360 | struct isl_sched_graph *graph = data; |
1361 | const int *i1 = a; |
1362 | const int *i2 = b; |
1363 | |
1364 | return graph->node[*i1].scc - graph->node[*i2].scc; |
1365 | } |
1366 | |
1367 | /* Sort the elements of graph->sorted according to the corresponding SCCs. |
1368 | */ |
1369 | static int sort_sccs(struct isl_sched_graph *graph) |
1370 | { |
1371 | return isl_sort(graph->sorted, graph->n, sizeof(int), &cmp_scc, graph); |
1372 | } |
1373 | |
1374 | /* Given a dependence relation R from "node" to itself, |
1375 | * construct the set of coefficients of valid constraints for elements |
1376 | * in that dependence relation. |
1377 | * In particular, the result contains tuples of coefficients |
1378 | * c_0, c_n, c_x such that |
1379 | * |
1380 | * c_0 + c_n n + c_x y - c_x x >= 0 for each (x,y) in R |
1381 | * |
1382 | * or, equivalently, |
1383 | * |
1384 | * c_0 + c_n n + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R } |
1385 | * |
1386 | * We choose here to compute the dual of delta R. |
1387 | * Alternatively, we could have computed the dual of R, resulting |
1388 | * in a set of tuples c_0, c_n, c_x, c_y, and then |
1389 | * plugged in (c_0, c_n, c_x, -c_x). |
1390 | * |
1391 | * If "node" has been compressed, then the dependence relation |
1392 | * is also compressed before the set of coefficients is computed. |
1393 | */ |
1394 | static __isl_give isl_basic_setisl_basic_map *intra_coefficients( |
1395 | struct isl_sched_graph *graph, struct isl_sched_node *node, |
1396 | __isl_take isl_map *map) |
1397 | { |
1398 | isl_setisl_map *delta; |
1399 | isl_map *key; |
1400 | isl_basic_setisl_basic_map *coef; |
1401 | |
1402 | if (isl_map_to_basic_set_has(graph->intra_hmap, map)) |
1403 | return isl_map_to_basic_set_get(graph->intra_hmap, map); |
1404 | |
1405 | key = isl_map_copy(map); |
1406 | if (node->compressed) { |
1407 | map = isl_map_preimage_domain_multi_aff(map, |
1408 | isl_multi_aff_copy(node->decompress)); |
1409 | map = isl_map_preimage_range_multi_aff(map, |
1410 | isl_multi_aff_copy(node->decompress)); |
1411 | } |
1412 | delta = isl_set_remove_divs(isl_map_deltas(map)); |
1413 | coef = isl_set_coefficients(delta); |
1414 | graph->intra_hmap = isl_map_to_basic_set_set(graph->intra_hmap, key, |
1415 | isl_basic_set_copy(coef)); |
1416 | |
1417 | return coef; |
1418 | } |
1419 | |
1420 | /* Given a dependence relation R, construct the set of coefficients |
1421 | * of valid constraints for elements in that dependence relation. |
1422 | * In particular, the result contains tuples of coefficients |
1423 | * c_0, c_n, c_x, c_y such that |
1424 | * |
1425 | * c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R |
1426 | * |
1427 | * If the source or destination nodes of "edge" have been compressed, |
1428 | * then the dependence relation is also compressed before |
1429 | * the set of coefficients is computed. |
1430 | */ |
1431 | static __isl_give isl_basic_setisl_basic_map *inter_coefficients( |
1432 | struct isl_sched_graph *graph, struct isl_sched_edge *edge, |
1433 | __isl_take isl_map *map) |
1434 | { |
1435 | isl_setisl_map *set; |
1436 | isl_map *key; |
1437 | isl_basic_setisl_basic_map *coef; |
1438 | |
1439 | if (isl_map_to_basic_set_has(graph->inter_hmap, map)) |
1440 | return isl_map_to_basic_set_get(graph->inter_hmap, map); |
1441 | |
1442 | key = isl_map_copy(map); |
1443 | if (edge->src->compressed) |
1444 | map = isl_map_preimage_domain_multi_aff(map, |
1445 | isl_multi_aff_copy(edge->src->decompress)); |
1446 | if (edge->dst->compressed) |
1447 | map = isl_map_preimage_range_multi_aff(map, |
1448 | isl_multi_aff_copy(edge->dst->decompress)); |
1449 | set = isl_map_wrap(isl_map_remove_divs(map)); |
1450 | coef = isl_set_coefficients(set); |
1451 | graph->inter_hmap = isl_map_to_basic_set_set(graph->inter_hmap, key, |
1452 | isl_basic_set_copy(coef)); |
1453 | |
1454 | return coef; |
1455 | } |
1456 | |
1457 | /* Add constraints to graph->lp that force validity for the given |
1458 | * dependence from a node i to itself. |
1459 | * That is, add constraints that enforce |
1460 | * |
1461 | * (c_i_0 + c_i_n n + c_i_x y) - (c_i_0 + c_i_n n + c_i_x x) |
1462 | * = c_i_x (y - x) >= 0 |
1463 | * |
1464 | * for each (x,y) in R. |
1465 | * We obtain general constraints on coefficients (c_0, c_n, c_x) |
1466 | * of valid constraints for (y - x) and then plug in (0, 0, c_i_x^+ - c_i_x^-), |
1467 | * where c_i_x = c_i_x^+ - c_i_x^-, with c_i_x^+ and c_i_x^- non-negative. |
1468 | * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart. |
1469 | * |
1470 | * Actually, we do not construct constraints for the c_i_x themselves, |
1471 | * but for the coefficients of c_i_x written as a linear combination |
1472 | * of the columns in node->cmap. |
1473 | */ |
1474 | static int add_intra_validity_constraints(struct isl_sched_graph *graph, |
1475 | struct isl_sched_edge *edge) |
1476 | { |
1477 | unsigned total; |
1478 | isl_map *map = isl_map_copy(edge->map); |
1479 | isl_ctx *ctx = isl_map_get_ctx(map); |
1480 | isl_space *dim; |
1481 | isl_dim_map *dim_map; |
1482 | isl_basic_setisl_basic_map *coef; |
1483 | struct isl_sched_node *node = edge->src; |
1484 | |
1485 | coef = intra_coefficients(graph, node, map); |
1486 | |
1487 | dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); |
1488 | |
1489 | coef = isl_basic_set_transform_dims(coef, isl_dim_set, |
1490 | isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap)); |
1491 | if (!coef) |
1492 | goto error; |
1493 | |
1494 | total = isl_basic_set_total_dim(graph->lp); |
1495 | dim_map = isl_dim_map_alloc(ctx, total); |
1496 | isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2, |
1497 | isl_space_dim(dim, isl_dim_set), 1, |
1498 | node->nvar, -1); |
1499 | isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2, |
1500 | isl_space_dim(dim, isl_dim_set), 1, |
1501 | node->nvar, 1); |
1502 | graph->lp = isl_basic_set_extend_constraints(graph->lp, |
1503 | coef->n_eq, coef->n_ineq); |
1504 | graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, |
1505 | coef, dim_map); |
1506 | isl_space_free(dim); |
1507 | |
1508 | return 0; |
1509 | error: |
1510 | isl_space_free(dim); |
1511 | return -1; |
1512 | } |
1513 | |
1514 | /* Add constraints to graph->lp that force validity for the given |
1515 | * dependence from node i to node j. |
1516 | * That is, add constraints that enforce |
1517 | * |
1518 | * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) >= 0 |
1519 | * |
1520 | * for each (x,y) in R. |
1521 | * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y) |
1522 | * of valid constraints for R and then plug in |
1523 | * (c_j_0 - c_i_0, c_j_n^+ - c_j_n^- - (c_i_n^+ - c_i_n^-), |
1524 | * c_j_x^+ - c_j_x^- - (c_i_x^+ - c_i_x^-)), |
1525 | * where c_* = c_*^+ - c_*^-, with c_*^+ and c_*^- non-negative. |
1526 | * In graph->lp, the c_*^- appear before their c_*^+ counterpart. |
1527 | * |
1528 | * Actually, we do not construct constraints for the c_*_x themselves, |
1529 | * but for the coefficients of c_*_x written as a linear combination |
1530 | * of the columns in node->cmap. |
1531 | */ |
1532 | static int add_inter_validity_constraints(struct isl_sched_graph *graph, |
1533 | struct isl_sched_edge *edge) |
1534 | { |
1535 | unsigned total; |
1536 | isl_map *map = isl_map_copy(edge->map); |
1537 | isl_ctx *ctx = isl_map_get_ctx(map); |
1538 | isl_space *dim; |
1539 | isl_dim_map *dim_map; |
1540 | isl_basic_setisl_basic_map *coef; |
1541 | struct isl_sched_node *src = edge->src; |
1542 | struct isl_sched_node *dst = edge->dst; |
1543 | |
1544 | coef = inter_coefficients(graph, edge, map); |
1545 | |
1546 | dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); |
1547 | |
1548 | coef = isl_basic_set_transform_dims(coef, isl_dim_set, |
1549 | isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap)); |
1550 | coef = isl_basic_set_transform_dims(coef, isl_dim_set, |
1551 | isl_space_dim(dim, isl_dim_set) + src->nvar, |
1552 | isl_mat_copy(dst->cmap)); |
1553 | if (!coef) |
1554 | goto error; |
1555 | |
1556 | total = isl_basic_set_total_dim(graph->lp); |
1557 | dim_map = isl_dim_map_alloc(ctx, total); |
1558 | |
1559 | isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1); |
1560 | isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1); |
1561 | isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1); |
1562 | isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2, |
1563 | isl_space_dim(dim, isl_dim_set) + src->nvar, 1, |
1564 | dst->nvar, -1); |
1565 | isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2, |
1566 | isl_space_dim(dim, isl_dim_set) + src->nvar, 1, |
1567 | dst->nvar, 1); |
1568 | |
1569 | isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1); |
1570 | isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1); |
1571 | isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1); |
1572 | isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2, |
1573 | isl_space_dim(dim, isl_dim_set), 1, |
1574 | src->nvar, 1); |
1575 | isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2, |
1576 | isl_space_dim(dim, isl_dim_set), 1, |
1577 | src->nvar, -1); |
1578 | |
1579 | edge->start = graph->lp->n_ineq; |
1580 | graph->lp = isl_basic_set_extend_constraints(graph->lp, |
1581 | coef->n_eq, coef->n_ineq); |
1582 | graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, |
1583 | coef, dim_map); |
1584 | if (!graph->lp) |
1585 | goto error; |
1586 | isl_space_free(dim); |
1587 | edge->end = graph->lp->n_ineq; |
1588 | |
1589 | return 0; |
1590 | error: |
1591 | isl_space_free(dim); |
1592 | return -1; |
1593 | } |
1594 | |
1595 | /* Add constraints to graph->lp that bound the dependence distance for the given |
1596 | * dependence from a node i to itself. |
1597 | * If s = 1, we add the constraint |
1598 | * |
1599 | * c_i_x (y - x) <= m_0 + m_n n |
1600 | * |
1601 | * or |
1602 | * |
1603 | * -c_i_x (y - x) + m_0 + m_n n >= 0 |
1604 | * |
1605 | * for each (x,y) in R. |
1606 | * If s = -1, we add the constraint |
1607 | * |
1608 | * -c_i_x (y - x) <= m_0 + m_n n |
1609 | * |
1610 | * or |
1611 | * |
1612 | * c_i_x (y - x) + m_0 + m_n n >= 0 |
1613 | * |
1614 | * for each (x,y) in R. |
1615 | * We obtain general constraints on coefficients (c_0, c_n, c_x) |
1616 | * of valid constraints for (y - x) and then plug in (m_0, m_n, -s * c_i_x), |
1617 | * with each coefficient (except m_0) represented as a pair of non-negative |
1618 | * coefficients. |
1619 | * |
1620 | * Actually, we do not construct constraints for the c_i_x themselves, |
1621 | * but for the coefficients of c_i_x written as a linear combination |
1622 | * of the columns in node->cmap. |
1623 | * |
1624 | * |
1625 | * If "local" is set, then we add constraints |
1626 | * |
1627 | * c_i_x (y - x) <= 0 |
1628 | * |
1629 | * or |
1630 | * |
1631 | * -c_i_x (y - x) <= 0 |
1632 | * |
1633 | * instead, forcing the dependence distance to be (less than or) equal to 0. |
1634 | * That is, we plug in (0, 0, -s * c_i_x), |
1635 | * Note that dependences marked local are treated as validity constraints |
1636 | * by add_all_validity_constraints and therefore also have |
1637 | * their distances bounded by 0 from below. |
1638 | */ |
1639 | static int add_intra_proximity_constraints(struct isl_sched_graph *graph, |
1640 | struct isl_sched_edge *edge, int s, int local) |
1641 | { |
1642 | unsigned total; |
1643 | unsigned nparam; |
1644 | isl_map *map = isl_map_copy(edge->map); |
1645 | isl_ctx *ctx = isl_map_get_ctx(map); |
1646 | isl_space *dim; |
1647 | isl_dim_map *dim_map; |
1648 | isl_basic_setisl_basic_map *coef; |
1649 | struct isl_sched_node *node = edge->src; |
1650 | |
1651 | coef = intra_coefficients(graph, node, map); |
1652 | |
1653 | dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); |
1654 | |
1655 | coef = isl_basic_set_transform_dims(coef, isl_dim_set, |
1656 | isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap)); |
1657 | if (!coef) |
1658 | goto error; |
1659 | |
1660 | nparam = isl_space_dim(node->space, isl_dim_param); |
1661 | total = isl_basic_set_total_dim(graph->lp); |
1662 | dim_map = isl_dim_map_alloc(ctx, total); |
1663 | |
1664 | if (!local) { |
1665 | isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1); |
1666 | isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1); |
1667 | isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1); |
1668 | } |
1669 | isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2, |
1670 | isl_space_dim(dim, isl_dim_set), 1, |
1671 | node->nvar, s); |
1672 | isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2, |
1673 | isl_space_dim(dim, isl_dim_set), 1, |
1674 | node->nvar, -s); |
1675 | graph->lp = isl_basic_set_extend_constraints(graph->lp, |
1676 | coef->n_eq, coef->n_ineq); |
1677 | graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, |
1678 | coef, dim_map); |
1679 | isl_space_free(dim); |
1680 | |
1681 | return 0; |
1682 | error: |
1683 | isl_space_free(dim); |
1684 | return -1; |
1685 | } |
1686 | |
1687 | /* Add constraints to graph->lp that bound the dependence distance for the given |
1688 | * dependence from node i to node j. |
1689 | * If s = 1, we add the constraint |
1690 | * |
1691 | * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) |
1692 | * <= m_0 + m_n n |
1693 | * |
1694 | * or |
1695 | * |
1696 | * -(c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x) + |
1697 | * m_0 + m_n n >= 0 |
1698 | * |
1699 | * for each (x,y) in R. |
1700 | * If s = -1, we add the constraint |
1701 | * |
1702 | * -((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)) |
1703 | * <= m_0 + m_n n |
1704 | * |
1705 | * or |
1706 | * |
1707 | * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) + |
1708 | * m_0 + m_n n >= 0 |
1709 | * |
1710 | * for each (x,y) in R. |
1711 | * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y) |
1712 | * of valid constraints for R and then plug in |
1713 | * (m_0 - s*c_j_0 + s*c_i_0, m_n - s*c_j_n + s*c_i_n, |
1714 | * -s*c_j_x+s*c_i_x) |
1715 | * with each coefficient (except m_0, c_j_0 and c_i_0) |
1716 | * represented as a pair of non-negative coefficients. |
1717 | * |
1718 | * Actually, we do not construct constraints for the c_*_x themselves, |
1719 | * but for the coefficients of c_*_x written as a linear combination |
1720 | * of the columns in node->cmap. |
1721 | * |
1722 | * |
1723 | * If "local" is set, then we add constraints |
1724 | * |
1725 | * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) <= 0 |
1726 | * |
1727 | * or |
1728 | * |
1729 | * -((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)) <= 0 |
1730 | * |
1731 | * instead, forcing the dependence distance to be (less than or) equal to 0. |
1732 | * That is, we plug in |
1733 | * (-s*c_j_0 + s*c_i_0, -s*c_j_n + s*c_i_n, -s*c_j_x+s*c_i_x). |
1734 | * Note that dependences marked local are treated as validity constraints |
1735 | * by add_all_validity_constraints and therefore also have |
1736 | * their distances bounded by 0 from below. |
1737 | */ |
1738 | static int add_inter_proximity_constraints(struct isl_sched_graph *graph, |
1739 | struct isl_sched_edge *edge, int s, int local) |
1740 | { |
1741 | unsigned total; |
1742 | unsigned nparam; |
1743 | isl_map *map = isl_map_copy(edge->map); |
1744 | isl_ctx *ctx = isl_map_get_ctx(map); |
1745 | isl_space *dim; |
1746 | isl_dim_map *dim_map; |
1747 | isl_basic_setisl_basic_map *coef; |
1748 | struct isl_sched_node *src = edge->src; |
1749 | struct isl_sched_node *dst = edge->dst; |
1750 | |
1751 | coef = inter_coefficients(graph, edge, map); |
1752 | |
1753 | dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); |
1754 | |
1755 | coef = isl_basic_set_transform_dims(coef, isl_dim_set, |
1756 | isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap)); |
1757 | coef = isl_basic_set_transform_dims(coef, isl_dim_set, |
1758 | isl_space_dim(dim, isl_dim_set) + src->nvar, |
1759 | isl_mat_copy(dst->cmap)); |
1760 | if (!coef) |
1761 | goto error; |
1762 | |
1763 | nparam = isl_space_dim(src->space, isl_dim_param); |
1764 | total = isl_basic_set_total_dim(graph->lp); |
1765 | dim_map = isl_dim_map_alloc(ctx, total); |
1766 | |
1767 | if (!local) { |
1768 | isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1); |
1769 | isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1); |
1770 | isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1); |
1771 | } |
1772 | |
1773 | isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, -s); |
1774 | isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, s); |
1775 | isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, -s); |
1776 | isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2, |
1777 | isl_space_dim(dim, isl_dim_set) + src->nvar, 1, |
1778 | dst->nvar, s); |
1779 | isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2, |
1780 | isl_space_dim(dim, isl_dim_set) + src->nvar, 1, |
1781 | dst->nvar, -s); |
1782 | |
1783 | isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, s); |
1784 | isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, -s); |
1785 | isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, s); |
1786 | isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2, |
1787 | isl_space_dim(dim, isl_dim_set), 1, |
1788 | src->nvar, -s); |
1789 | isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2, |
1790 | isl_space_dim(dim, isl_dim_set), 1, |
1791 | src->nvar, s); |
1792 | |
1793 | graph->lp = isl_basic_set_extend_constraints(graph->lp, |
1794 | coef->n_eq, coef->n_ineq); |
1795 | graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, |
1796 | coef, dim_map); |
1797 | isl_space_free(dim); |
1798 | |
1799 | return 0; |
1800 | error: |
1801 | isl_space_free(dim); |
1802 | return -1; |
1803 | } |
1804 | |
1805 | /* Add all validity constraints to graph->lp. |
1806 | * |
1807 | * An edge that is forced to be local needs to have its dependence |
1808 | * distances equal to zero. We take care of bounding them by 0 from below |
1809 | * here. add_all_proximity_constraints takes care of bounding them by 0 |
1810 | * from above. |
1811 | * |
1812 | * If "use_coincidence" is set, then we treat coincidence edges as local edges. |
1813 | * Otherwise, we ignore them. |
1814 | */ |
1815 | static int add_all_validity_constraints(struct isl_sched_graph *graph, |
1816 | int use_coincidence) |
1817 | { |
1818 | int i; |
1819 | |
1820 | for (i = 0; i < graph->n_edge; ++i) { |
1821 | struct isl_sched_edge *edge= &graph->edge[i]; |
1822 | int local; |
1823 | |
1824 | local = edge->local || (edge->coincidence && use_coincidence); |
1825 | if (!edge->validity && !local) |
1826 | continue; |
1827 | if (edge->src != edge->dst) |
1828 | continue; |
1829 | if (add_intra_validity_constraints(graph, edge) < 0) |
1830 | return -1; |
1831 | } |
1832 | |
1833 | for (i = 0; i < graph->n_edge; ++i) { |
1834 | struct isl_sched_edge *edge = &graph->edge[i]; |
1835 | int local; |
1836 | |
1837 | local = edge->local || (edge->coincidence && use_coincidence); |
1838 | if (!edge->validity && !local) |
1839 | continue; |
1840 | if (edge->src == edge->dst) |
1841 | continue; |
1842 | if (add_inter_validity_constraints(graph, edge) < 0) |
1843 | return -1; |
1844 | } |
1845 | |
1846 | return 0; |
1847 | } |
1848 | |
1849 | /* Add constraints to graph->lp that bound the dependence distance |
1850 | * for all dependence relations. |
1851 | * If a given proximity dependence is identical to a validity |
1852 | * dependence, then the dependence distance is already bounded |
1853 | * from below (by zero), so we only need to bound the distance |
1854 | * from above. (This includes the case of "local" dependences |
1855 | * which are treated as validity dependence by add_all_validity_constraints.) |
1856 | * Otherwise, we need to bound the distance both from above and from below. |
1857 | * |
1858 | * If "use_coincidence" is set, then we treat coincidence edges as local edges. |
1859 | * Otherwise, we ignore them. |
1860 | */ |
1861 | static int add_all_proximity_constraints(struct isl_sched_graph *graph, |
1862 | int use_coincidence) |
1863 | { |
1864 | int i; |
1865 | |
1866 | for (i = 0; i < graph->n_edge; ++i) { |
1867 | struct isl_sched_edge *edge= &graph->edge[i]; |
1868 | int local; |
1869 | |
1870 | local = edge->local || (edge->coincidence && use_coincidence); |
1871 | if (!edge->proximity && !local) |
1872 | continue; |
1873 | if (edge->src == edge->dst && |
1874 | add_intra_proximity_constraints(graph, edge, 1, local) < 0) |
1875 | return -1; |
1876 | if (edge->src != edge->dst && |
1877 | add_inter_proximity_constraints(graph, edge, 1, local) < 0) |
1878 | return -1; |
1879 | if (edge->validity || local) |
1880 | continue; |
1881 | if (edge->src == edge->dst && |
1882 | add_intra_proximity_constraints(graph, edge, -1, 0) < 0) |
1883 | return -1; |
1884 | if (edge->src != edge->dst && |
1885 | add_inter_proximity_constraints(graph, edge, -1, 0) < 0) |
1886 | return -1; |
1887 | } |
1888 | |
1889 | return 0; |
1890 | } |
1891 | |
1892 | /* Compute a basis for the rows in the linear part of the schedule |
1893 | * and extend this basis to a full basis. The remaining rows |
1894 | * can then be used to force linear independence from the rows |
1895 | * in the schedule. |
1896 | * |
1897 | * In particular, given the schedule rows S, we compute |
1898 | * |
1899 | * S = H Q |
1900 | * S U = H |
1901 | * |
1902 | * with H the Hermite normal form of S. That is, all but the |
1903 | * first rank columns of H are zero and so each row in S is |
1904 | * a linear combination of the first rank rows of Q. |
1905 | * The matrix Q is then transposed because we will write the |
1906 | * coefficients of the next schedule row as a column vector s |
1907 | * and express this s as a linear combination s = Q c of the |
1908 | * computed basis. |
1909 | * Similarly, the matrix U is transposed such that we can |
1910 | * compute the coefficients c = U s from a schedule row s. |
1911 | */ |
1912 | static int node_update_cmap(struct isl_sched_node *node) |
1913 | { |
1914 | isl_mat *H, *U, *Q; |
1915 | int n_row = isl_mat_rows(node->sched); |
1916 | |
1917 | H = isl_mat_sub_alloc(node->sched, 0, n_row, |
1918 | 1 + node->nparam, node->nvar); |
1919 | |
1920 | H = isl_mat_left_hermite(H, 0, &U, &Q); |
1921 | isl_mat_free(node->cmap); |
1922 | isl_mat_free(node->cinv); |
1923 | node->cmap = isl_mat_transpose(Q); |
1924 | node->cinv = isl_mat_transpose(U); |
1925 | node->rank = isl_mat_initial_non_zero_cols(H); |
1926 | isl_mat_free(H); |
1927 | |
1928 | if (!node->cmap || !node->cinv || node->rank < 0) |
1929 | return -1; |
1930 | return 0; |
1931 | } |
1932 | |
1933 | /* How many times should we count the constraints in "edge"? |
1934 | * |
1935 | * If carry is set, then we are counting the number of |
1936 | * (validity or conditional validity) constraints that will be added |
1937 | * in setup_carry_lp and we count each edge exactly once. |
1938 | * |
1939 | * Otherwise, we count as follows |
1940 | * validity -> 1 (>= 0) |
1941 | * validity+proximity -> 2 (>= 0 and upper bound) |
1942 | * proximity -> 2 (lower and upper bound) |
1943 | * local(+any) -> 2 (>= 0 and <= 0) |
1944 | * |
1945 | * If an edge is only marked conditional_validity then it counts |
1946 | * as zero since it is only checked afterwards. |
1947 | * |
1948 | * If "use_coincidence" is set, then we treat coincidence edges as local edges. |
1949 | * Otherwise, we ignore them. |
1950 | */ |
1951 | static int edge_multiplicity(struct isl_sched_edge *edge, int carry, |
1952 | int use_coincidence) |
1953 | { |
1954 | if (carry && !edge->validity && !edge->conditional_validity) |
1955 | return 0; |
1956 | if (carry) |
1957 | return 1; |
1958 | if (edge->proximity || edge->local) |
1959 | return 2; |
1960 | if (use_coincidence && edge->coincidence) |
1961 | return 2; |
1962 | if (edge->validity) |
1963 | return 1; |
1964 | return 0; |
1965 | } |
1966 | |
1967 | /* Count the number of equality and inequality constraints |
1968 | * that will be added for the given map. |
1969 | * |
1970 | * "use_coincidence" is set if we should take into account coincidence edges. |
1971 | */ |
1972 | static int count_map_constraints(struct isl_sched_graph *graph, |
1973 | struct isl_sched_edge *edge, __isl_take isl_map *map, |
1974 | int *n_eq, int *n_ineq, int carry, int use_coincidence) |
1975 | { |
1976 | isl_basic_setisl_basic_map *coef; |
1977 | int f = edge_multiplicity(edge, carry, use_coincidence); |
1978 | |
1979 | if (f == 0) { |
1980 | isl_map_free(map); |
1981 | return 0; |
1982 | } |
1983 | |
1984 | if (edge->src == edge->dst) |
1985 | coef = intra_coefficients(graph, edge->src, map); |
1986 | else |
1987 | coef = inter_coefficients(graph, edge, map); |
1988 | if (!coef) |
1989 | return -1; |
1990 | *n_eq += f * coef->n_eq; |
1991 | *n_ineq += f * coef->n_ineq; |
1992 | isl_basic_set_free(coef); |
1993 | |
1994 | return 0; |
1995 | } |
1996 | |
1997 | /* Count the number of equality and inequality constraints |
1998 | * that will be added to the main lp problem. |
1999 | * We count as follows |
2000 | * validity -> 1 (>= 0) |
2001 | * validity+proximity -> 2 (>= 0 and upper bound) |
2002 | * proximity -> 2 (lower and upper bound) |
2003 | * local(+any) -> 2 (>= 0 and <= 0) |
2004 | * |
2005 | * If "use_coincidence" is set, then we treat coincidence edges as local edges. |
2006 | * Otherwise, we ignore them. |
2007 | */ |
2008 | static int count_constraints(struct isl_sched_graph *graph, |
2009 | int *n_eq, int *n_ineq, int use_coincidence) |
2010 | { |
2011 | int i; |
2012 | |
2013 | *n_eq = *n_ineq = 0; |
2014 | for (i = 0; i < graph->n_edge; ++i) { |
2015 | struct isl_sched_edge *edge= &graph->edge[i]; |
2016 | isl_map *map = isl_map_copy(edge->map); |
2017 | |
2018 | if (count_map_constraints(graph, edge, map, n_eq, n_ineq, |
2019 | 0, use_coincidence) < 0) |
2020 | return -1; |
2021 | } |
2022 | |
2023 | return 0; |
2024 | } |
2025 | |
2026 | /* Count the number of constraints that will be added by |
2027 | * add_bound_coefficient_constraints and increment *n_eq and *n_ineq |
2028 | * accordingly. |
2029 | * |
2030 | * In practice, add_bound_coefficient_constraints only adds inequalities. |
2031 | */ |
2032 | static int count_bound_coefficient_constraints(isl_ctx *ctx, |
2033 | struct isl_sched_graph *graph, int *n_eq, int *n_ineq) |
2034 | { |
2035 | int i; |
2036 | |
2037 | if (ctx->opt->schedule_max_coefficient == -1) |
2038 | return 0; |
2039 | |
2040 | for (i = 0; i < graph->n; ++i) |
2041 | *n_ineq += 2 * graph->node[i].nparam + 2 * graph->node[i].nvar; |
2042 | |
2043 | return 0; |
2044 | } |
2045 | |
2046 | /* Add constraints that bound the values of the variable and parameter |
2047 | * coefficients of the schedule. |
2048 | * |
2049 | * The maximal value of the coefficients is defined by the option |
2050 | * 'schedule_max_coefficient'. |
2051 | */ |
2052 | static int add_bound_coefficient_constraints(isl_ctx *ctx, |
2053 | struct isl_sched_graph *graph) |
2054 | { |
2055 | int i, j, k; |
2056 | int max_coefficient; |
2057 | int total; |
2058 | |
2059 | max_coefficient = ctx->opt->schedule_max_coefficient; |
2060 | |
2061 | if (max_coefficient == -1) |
2062 | return 0; |
2063 | |
2064 | total = isl_basic_set_total_dim(graph->lp); |
2065 | |
2066 | for (i = 0; i < graph->n; ++i) { |
2067 | struct isl_sched_node *node = &graph->node[i]; |
2068 | for (j = 0; j < 2 * node->nparam + 2 * node->nvar; ++j) { |
2069 | int dim; |
2070 | k = isl_basic_set_alloc_inequality(graph->lp); |
2071 | if (k < 0) |
2072 | return -1; |
2073 | dim = 1 + node->start + 1 + j; |
2074 | isl_seq_clr(graph->lp->ineq[k], 1 + total); |
2075 | isl_int_set_si(graph->lp->ineq[k][dim], -1)isl_sioimath_set_si((graph->lp->ineq[k][dim]), -1); |
2076 | isl_int_set_si(graph->lp->ineq[k][0], max_coefficient)isl_sioimath_set_si((graph->lp->ineq[k][0]), max_coefficient ); |
2077 | } |
2078 | } |
2079 | |
2080 | return 0; |
2081 | } |
2082 | |
2083 | /* Construct an ILP problem for finding schedule coefficients |
2084 | * that result in non-negative, but small dependence distances |
2085 | * over all dependences. |
2086 | * In particular, the dependence distances over proximity edges |
2087 | * are bounded by m_0 + m_n n and we compute schedule coefficients |
2088 | * with small values (preferably zero) of m_n and m_0. |
2089 | * |
2090 | * All variables of the ILP are non-negative. The actual coefficients |
2091 | * may be negative, so each coefficient is represented as the difference |
2092 | * of two non-negative variables. The negative part always appears |
2093 | * immediately before the positive part. |
2094 | * Other than that, the variables have the following order |
2095 | * |
2096 | * - sum of positive and negative parts of m_n coefficients |
2097 | * - m_0 |
2098 | * - sum of positive and negative parts of all c_n coefficients |
2099 | * (unconstrained when computing non-parametric schedules) |
2100 | * - sum of positive and negative parts of all c_x coefficients |
2101 | * - positive and negative parts of m_n coefficients |
2102 | * - for each node |
2103 | * - c_i_0 |
2104 | * - positive and negative parts of c_i_n (if parametric) |
2105 | * - positive and negative parts of c_i_x |
2106 | * |
2107 | * The c_i_x are not represented directly, but through the columns of |
2108 | * node->cmap. That is, the computed values are for variable t_i_x |
2109 | * such that c_i_x = Q t_i_x with Q equal to node->cmap. |
2110 | * |
2111 | * The constraints are those from the edges plus two or three equalities |
2112 | * to express the sums. |
2113 | * |
2114 | * If "use_coincidence" is set, then we treat coincidence edges as local edges. |
2115 | * Otherwise, we ignore them. |
2116 | */ |
2117 | static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph, |
2118 | int use_coincidence) |
2119 | { |
2120 | int i, j; |
2121 | int k; |
2122 | unsigned nparam; |
2123 | unsigned total; |
2124 | isl_space *dim; |
2125 | int parametric; |
2126 | int param_pos; |
2127 | int n_eq, n_ineq; |
2128 | int max_constant_term; |
2129 | |
2130 | max_constant_term = ctx->opt->schedule_max_constant_term; |
2131 | |
2132 | parametric = ctx->opt->schedule_parametric; |
2133 | nparam = isl_space_dim(graph->node[0].space, isl_dim_param); |
2134 | param_pos = 4; |
2135 | total = param_pos + 2 * nparam; |
2136 | for (i = 0; i < graph->n; ++i) { |
2137 | struct isl_sched_node *node = &graph->node[graph->sorted[i]]; |
2138 | if (node_update_cmap(node) < 0) |
2139 | return -1; |
2140 | node->start = total; |
2141 | total += 1 + 2 * (node->nparam + node->nvar); |
2142 | } |
2143 | |
2144 | if (count_constraints(graph, &n_eq, &n_ineq, use_coincidence) < 0) |
2145 | return -1; |
2146 | if (count_bound_coefficient_constraints(ctx, graph, &n_eq, &n_ineq) < 0) |
2147 | return -1; |
2148 | |
2149 | dim = isl_space_set_alloc(ctx, 0, total); |
2150 | isl_basic_set_free(graph->lp); |
2151 | n_eq += 2 + parametric; |
2152 | if (max_constant_term != -1) |
2153 | n_ineq += graph->n; |
2154 | |
2155 | graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq); |
2156 | |
2157 | k = isl_basic_set_alloc_equality(graph->lp); |
2158 | if (k < 0) |
2159 | return -1; |
2160 | isl_seq_clr(graph->lp->eq[k], 1 + total); |
2161 | isl_int_set_si(graph->lp->eq[k][1], -1)isl_sioimath_set_si((graph->lp->eq[k][1]), -1); |
2162 | for (i = 0; i < 2 * nparam; ++i) |
2163 | isl_int_set_si(graph->lp->eq[k][1 + param_pos + i], 1)isl_sioimath_set_si((graph->lp->eq[k][1 + param_pos + i ]), 1); |
2164 | |
2165 | if (parametric) { |
2166 | k = isl_basic_set_alloc_equality(graph->lp); |
2167 | if (k < 0) |
2168 | return -1; |
2169 | isl_seq_clr(graph->lp->eq[k], 1 + total); |
2170 | isl_int_set_si(graph->lp->eq[k][3], -1)isl_sioimath_set_si((graph->lp->eq[k][3]), -1); |
2171 | for (i = 0; i < graph->n; ++i) { |
2172 | int pos = 1 + graph->node[i].start + 1; |
2173 | |
2174 | for (j = 0; j < 2 * graph->node[i].nparam; ++j) |
2175 | isl_int_set_si(graph->lp->eq[k][pos + j], 1)isl_sioimath_set_si((graph->lp->eq[k][pos + j]), 1); |
2176 | } |
2177 | } |
2178 | |
2179 | k = isl_basic_set_alloc_equality(graph->lp); |
2180 | if (k < 0) |
2181 | return -1; |
2182 | isl_seq_clr(graph->lp->eq[k], 1 + total); |
2183 | isl_int_set_si(graph->lp->eq[k][4], -1)isl_sioimath_set_si((graph->lp->eq[k][4]), -1); |
2184 | for (i = 0; i < graph->n; ++i) { |
2185 | struct isl_sched_node *node = &graph->node[i]; |
2186 | int pos = 1 + node->start + 1 + 2 * node->nparam; |
2187 | |
2188 | for (j = 0; j < 2 * node->nvar; ++j) |
2189 | isl_int_set_si(graph->lp->eq[k][pos + j], 1)isl_sioimath_set_si((graph->lp->eq[k][pos + j]), 1); |
2190 | } |
2191 | |
2192 | if (max_constant_term != -1) |
2193 | for (i = 0; i < graph->n; ++i) { |
2194 | struct isl_sched_node *node = &graph->node[i]; |
2195 | k = isl_basic_set_alloc_inequality(graph->lp); |
2196 | if (k < 0) |
2197 | return -1; |
2198 | isl_seq_clr(graph->lp->ineq[k], 1 + total); |
2199 | isl_int_set_si(graph->lp->ineq[k][1 + node->start], -1)isl_sioimath_set_si((graph->lp->ineq[k][1 + node->start ]), -1); |
2200 | isl_int_set_si(graph->lp->ineq[k][0], max_constant_term)isl_sioimath_set_si((graph->lp->ineq[k][0]), max_constant_term ); |
2201 | } |
2202 | |
2203 | if (add_bound_coefficient_constraints(ctx, graph) < 0) |
2204 | return -1; |
2205 | if (add_all_validity_constraints(graph, use_coincidence) < 0) |
2206 | return -1; |
2207 | if (add_all_proximity_constraints(graph, use_coincidence) < 0) |
2208 | return -1; |
2209 | |
2210 | return 0; |
2211 | } |
2212 | |
2213 | /* Analyze the conflicting constraint found by |
2214 | * isl_tab_basic_set_non_trivial_lexmin. If it corresponds to the validity |
2215 | * constraint of one of the edges between distinct nodes, living, moreover |
2216 | * in distinct SCCs, then record the source and sink SCC as this may |
2217 | * be a good place to cut between SCCs. |
2218 | */ |
2219 | static int check_conflict(int con, void *user) |
2220 | { |
2221 | int i; |
2222 | struct isl_sched_graph *graph = user; |
2223 | |
2224 | if (graph->src_scc >= 0) |
2225 | return 0; |
2226 | |
2227 | con -= graph->lp->n_eq; |
2228 | |
2229 | if (con >= graph->lp->n_ineq) |
2230 | return 0; |
2231 | |
2232 | for (i = 0; i < graph->n_edge; ++i) { |
2233 | if (!graph->edge[i].validity) |
2234 | continue; |
2235 | if (graph->edge[i].src == graph->edge[i].dst) |
2236 | continue; |
2237 | if (graph->edge[i].src->scc == graph->edge[i].dst->scc) |
2238 | continue; |
2239 | if (graph->edge[i].start > con) |
2240 | continue; |
2241 | if (graph->edge[i].end <= con) |
2242 | continue; |
2243 | graph->src_scc = graph->edge[i].src->scc; |
2244 | graph->dst_scc = graph->edge[i].dst->scc; |
2245 | } |
2246 | |
2247 | return 0; |
2248 | } |
2249 | |
2250 | /* Check whether the next schedule row of the given node needs to be |
2251 | * non-trivial. Lower-dimensional domains may have some trivial rows, |
2252 | * but as soon as the number of remaining required non-trivial rows |
2253 | * is as large as the number or remaining rows to be computed, |
2254 | * all remaining rows need to be non-trivial. |
2255 | */ |
2256 | static int needs_row(struct isl_sched_graph *graph, struct isl_sched_node *node) |
2257 | { |
2258 | return node->nvar - node->rank >= graph->maxvar - graph->n_row; |
2259 | } |
2260 | |
2261 | /* Solve the ILP problem constructed in setup_lp. |
2262 | * For each node such that all the remaining rows of its schedule |
2263 | * need to be non-trivial, we construct a non-triviality region. |
2264 | * This region imposes that the next row is independent of previous rows. |
2265 | * In particular the coefficients c_i_x are represented by t_i_x |
2266 | * variables with c_i_x = Q t_i_x and Q a unimodular matrix such that |
2267 | * its first columns span the rows of the previously computed part |
2268 | * of the schedule. The non-triviality region enforces that at least |
2269 | * one of the remaining components of t_i_x is non-zero, i.e., |
2270 | * that the new schedule row depends on at least one of the remaining |
2271 | * columns of Q. |
2272 | */ |
2273 | static __isl_give isl_vec *solve_lp(struct isl_sched_graph *graph) |
2274 | { |
2275 | int i; |
2276 | isl_vec *sol; |
2277 | isl_basic_setisl_basic_map *lp; |
2278 | |
2279 | for (i = 0; i < graph->n; ++i) { |
2280 | struct isl_sched_node *node = &graph->node[i]; |
2281 | int skip = node->rank; |
2282 | graph->region[i].pos = node->start + 1 + 2*(node->nparam+skip); |
2283 | if (needs_row(graph, node)) |
2284 | graph->region[i].len = 2 * (node->nvar - skip); |
2285 | else |
2286 | graph->region[i].len = 0; |
2287 | } |
2288 | lp = isl_basic_set_copy(graph->lp); |
2289 | sol = isl_tab_basic_set_non_trivial_lexmin(lp, 2, graph->n, |
2290 | graph->region, &check_conflict, graph); |
2291 | return sol; |
2292 | } |
2293 | |
2294 | /* Update the schedules of all nodes based on the given solution |
2295 | * of the LP problem. |
2296 | * The new row is added to the current band. |
2297 | * All possibly negative coefficients are encoded as a difference |
2298 | * of two non-negative variables, so we need to perform the subtraction |
2299 | * here. Moreover, if use_cmap is set, then the solution does |
2300 | * not refer to the actual coefficients c_i_x, but instead to variables |
2301 | * t_i_x such that c_i_x = Q t_i_x and Q is equal to node->cmap. |
2302 | * In this case, we then also need to perform this multiplication |
2303 | * to obtain the values of c_i_x. |
2304 | * |
2305 | * If coincident is set, then the caller guarantees that the new |
2306 | * row satisfies the coincidence constraints. |
2307 | */ |
2308 | static int update_schedule(struct isl_sched_graph *graph, |
2309 | __isl_take isl_vec *sol, int use_cmap, int coincident) |
2310 | { |
2311 | int i, j; |
2312 | isl_vec *csol = NULL((void*)0); |
2313 | |
2314 | if (!sol) |
2315 | goto error; |
2316 | if (sol->size == 0) |
2317 | isl_die(sol->ctx, isl_error_internal,do { isl_handle_error(sol->ctx, isl_error_internal, "no solution found" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2318); goto error; } while (0) |
2318 | "no solution found", goto error)do { isl_handle_error(sol->ctx, isl_error_internal, "no solution found" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2318); goto error; } while (0); |
2319 | if (graph->n_total_row >= graph->max_row) |
2320 | isl_die(sol->ctx, isl_error_internal,do { isl_handle_error(sol->ctx, isl_error_internal, "too many schedule rows" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2321); goto error; } while (0) |
2321 | "too many schedule rows", goto error)do { isl_handle_error(sol->ctx, isl_error_internal, "too many schedule rows" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2321); goto error; } while (0); |
2322 | |
2323 | for (i = 0; i < graph->n; ++i) { |
2324 | struct isl_sched_node *node = &graph->node[i]; |
2325 | int pos = node->start; |
2326 | int row = isl_mat_rows(node->sched); |
2327 | |
2328 | isl_vec_free(csol); |
2329 | csol = isl_vec_alloc(sol->ctx, node->nvar); |
2330 | if (!csol) |
2331 | goto error; |
2332 | |
2333 | isl_map_free(node->sched_map); |
2334 | node->sched_map = NULL((void*)0); |
2335 | node->sched = isl_mat_add_rows(node->sched, 1); |
2336 | if (!node->sched) |
2337 | goto error; |
2338 | node->sched = isl_mat_set_element(node->sched, row, 0, |
2339 | sol->el[1 + pos]); |
2340 | for (j = 0; j < node->nparam + node->nvar; ++j) |
2341 | isl_int_sub(sol->el[1 + pos + 1 + 2 * j + 1],isl_sioimath_sub((sol->el[1 + pos + 1 + 2 * j + 1]), *(sol ->el[1 + pos + 1 + 2 * j + 1]), *(sol->el[1 + pos + 1 + 2 * j])) |
2342 | sol->el[1 + pos + 1 + 2 * j + 1],isl_sioimath_sub((sol->el[1 + pos + 1 + 2 * j + 1]), *(sol ->el[1 + pos + 1 + 2 * j + 1]), *(sol->el[1 + pos + 1 + 2 * j])) |
2343 | sol->el[1 + pos + 1 + 2 * j])isl_sioimath_sub((sol->el[1 + pos + 1 + 2 * j + 1]), *(sol ->el[1 + pos + 1 + 2 * j + 1]), *(sol->el[1 + pos + 1 + 2 * j])); |
2344 | for (j = 0; j < node->nparam; ++j) |
2345 | node->sched = isl_mat_set_element(node->sched, |
2346 | row, 1 + j, sol->el[1+pos+1+2*j+1]); |
2347 | for (j = 0; j < node->nvar; ++j) |
2348 | isl_int_set(csol->el[j],isl_sioimath_set((csol->el[j]), *(sol->el[1+pos+1+2*(node ->nparam+j)+1])) |
2349 | sol->el[1+pos+1+2*(node->nparam+j)+1])isl_sioimath_set((csol->el[j]), *(sol->el[1+pos+1+2*(node ->nparam+j)+1])); |
2350 | if (use_cmap) |
2351 | csol = isl_mat_vec_product(isl_mat_copy(node->cmap), |
2352 | csol); |
2353 | if (!csol) |
2354 | goto error; |
2355 | for (j = 0; j < node->nvar; ++j) |
2356 | node->sched = isl_mat_set_element(node->sched, |
2357 | row, 1 + node->nparam + j, csol->el[j]); |
2358 | node->coincident[graph->n_total_row] = coincident; |
2359 | } |
2360 | isl_vec_free(sol); |
2361 | isl_vec_free(csol); |
2362 | |
2363 | graph->n_row++; |
2364 | graph->n_total_row++; |
2365 | |
2366 | return 0; |
2367 | error: |
2368 | isl_vec_free(sol); |
2369 | isl_vec_free(csol); |
2370 | return -1; |
2371 | } |
2372 | |
2373 | /* Convert row "row" of node->sched into an isl_aff living in "ls" |
2374 | * and return this isl_aff. |
2375 | */ |
2376 | static __isl_give isl_aff *extract_schedule_row(__isl_take isl_local_space *ls, |
2377 | struct isl_sched_node *node, int row) |
2378 | { |
2379 | int j; |
2380 | isl_int v; |
2381 | isl_aff *aff; |
2382 | |
2383 | isl_int_init(v)isl_sioimath_init((v)); |
2384 | |
2385 | aff = isl_aff_zero_on_domain(ls); |
2386 | isl_mat_get_element(node->sched, row, 0, &v); |
2387 | aff = isl_aff_set_constant(aff, v); |
2388 | for (j = 0; j < node->nparam; ++j) { |
2389 | isl_mat_get_element(node->sched, row, 1 + j, &v); |
2390 | aff = isl_aff_set_coefficient(aff, isl_dim_param, j, v); |
2391 | } |
2392 | for (j = 0; j < node->nvar; ++j) { |
2393 | isl_mat_get_element(node->sched, row, 1 + node->nparam + j, &v); |
2394 | aff = isl_aff_set_coefficient(aff, isl_dim_in, j, v); |
2395 | } |
2396 | |
2397 | isl_int_clear(v)isl_sioimath_clear((v)); |
2398 | |
2399 | return aff; |
2400 | } |
2401 | |
2402 | /* Convert the "n" rows starting at "first" of node->sched into a multi_aff |
2403 | * and return this multi_aff. |
2404 | * |
2405 | * The result is defined over the uncompressed node domain. |
2406 | */ |
2407 | static __isl_give isl_multi_aff *node_extract_partial_schedule_multi_aff( |
2408 | struct isl_sched_node *node, int first, int n) |
2409 | { |
2410 | int i; |
2411 | isl_space *space; |
2412 | isl_local_space *ls; |
2413 | isl_aff *aff; |
2414 | isl_multi_aff *ma; |
2415 | int nrow; |
2416 | |
2417 | nrow = isl_mat_rows(node->sched); |
Value stored to 'nrow' is never read | |
2418 | if (node->compressed) |
2419 | space = isl_multi_aff_get_domain_space(node->decompress); |
2420 | else |
2421 | space = isl_space_copy(node->space); |
2422 | ls = isl_local_space_from_space(isl_space_copy(space)); |
2423 | space = isl_space_from_domain(space); |
2424 | space = isl_space_add_dims(space, isl_dim_out, n); |
2425 | ma = isl_multi_aff_zero(space); |
2426 | |
2427 | for (i = first; i < first + n; ++i) { |
2428 | aff = extract_schedule_row(isl_local_space_copy(ls), node, i); |
2429 | ma = isl_multi_aff_set_aff(ma, i - first, aff); |
2430 | } |
2431 | |
2432 | isl_local_space_free(ls); |
2433 | |
2434 | if (node->compressed) |
2435 | ma = isl_multi_aff_pullback_multi_aff(ma, |
2436 | isl_multi_aff_copy(node->compress)); |
2437 | |
2438 | return ma; |
2439 | } |
2440 | |
2441 | /* Convert node->sched into a multi_aff and return this multi_aff. |
2442 | * |
2443 | * The result is defined over the uncompressed node domain. |
2444 | */ |
2445 | static __isl_give isl_multi_aff *node_extract_schedule_multi_aff( |
2446 | struct isl_sched_node *node) |
2447 | { |
2448 | int nrow; |
2449 | |
2450 | nrow = isl_mat_rows(node->sched); |
2451 | return node_extract_partial_schedule_multi_aff(node, 0, nrow); |
2452 | } |
2453 | |
2454 | /* Convert node->sched into a map and return this map. |
2455 | * |
2456 | * The result is cached in node->sched_map, which needs to be released |
2457 | * whenever node->sched is updated. |
2458 | * It is defined over the uncompressed node domain. |
2459 | */ |
2460 | static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node) |
2461 | { |
2462 | if (!node->sched_map) { |
2463 | isl_multi_aff *ma; |
2464 | |
2465 | ma = node_extract_schedule_multi_aff(node); |
2466 | node->sched_map = isl_map_from_multi_aff(ma); |
2467 | } |
2468 | |
2469 | return isl_map_copy(node->sched_map); |
2470 | } |
2471 | |
2472 | /* Construct a map that can be used to update a dependence relation |
2473 | * based on the current schedule. |
2474 | * That is, construct a map expressing that source and sink |
2475 | * are executed within the same iteration of the current schedule. |
2476 | * This map can then be intersected with the dependence relation. |
2477 | * This is not the most efficient way, but this shouldn't be a critical |
2478 | * operation. |
2479 | */ |
2480 | static __isl_give isl_map *specializer(struct isl_sched_node *src, |
2481 | struct isl_sched_node *dst) |
2482 | { |
2483 | isl_map *src_sched, *dst_sched; |
2484 | |
2485 | src_sched = node_extract_schedule(src); |
2486 | dst_sched = node_extract_schedule(dst); |
2487 | return isl_map_apply_range(src_sched, isl_map_reverse(dst_sched)); |
2488 | } |
2489 | |
2490 | /* Intersect the domains of the nested relations in domain and range |
2491 | * of "umap" with "map". |
2492 | */ |
2493 | static __isl_give isl_union_map *intersect_domains( |
2494 | __isl_take isl_union_map *umap, __isl_keep isl_map *map) |
2495 | { |
2496 | isl_union_set *uset; |
2497 | |
2498 | umap = isl_union_map_zip(umap); |
2499 | uset = isl_union_set_from_set(isl_map_wrap(isl_map_copy(map))); |
2500 | umap = isl_union_map_intersect_domain(umap, uset); |
2501 | umap = isl_union_map_zip(umap); |
2502 | return umap; |
2503 | } |
2504 | |
2505 | /* Update the dependence relation of the given edge based |
2506 | * on the current schedule. |
2507 | * If the dependence is carried completely by the current schedule, then |
2508 | * it is removed from the edge_tables. It is kept in the list of edges |
2509 | * as otherwise all edge_tables would have to be recomputed. |
2510 | */ |
2511 | static int update_edge(struct isl_sched_graph *graph, |
2512 | struct isl_sched_edge *edge) |
2513 | { |
2514 | int empty; |
2515 | isl_map *id; |
2516 | |
2517 | id = specializer(edge->src, edge->dst); |
2518 | edge->map = isl_map_intersect(edge->map, isl_map_copy(id)); |
2519 | if (!edge->map) |
2520 | goto error; |
2521 | |
2522 | if (edge->tagged_condition) { |
2523 | edge->tagged_condition = |
2524 | intersect_domains(edge->tagged_condition, id); |
2525 | if (!edge->tagged_condition) |
2526 | goto error; |
2527 | } |
2528 | if (edge->tagged_validity) { |
2529 | edge->tagged_validity = |
2530 | intersect_domains(edge->tagged_validity, id); |
2531 | if (!edge->tagged_validity) |
2532 | goto error; |
2533 | } |
2534 | |
2535 | empty = isl_map_plain_is_empty(edge->map); |
2536 | if (empty < 0) |
2537 | goto error; |
2538 | if (empty) |
2539 | graph_remove_edge(graph, edge); |
2540 | |
2541 | isl_map_free(id); |
2542 | return 0; |
2543 | error: |
2544 | isl_map_free(id); |
2545 | return -1; |
2546 | } |
2547 | |
2548 | /* Does the domain of "umap" intersect "uset"? |
2549 | */ |
2550 | static int domain_intersects(__isl_keep isl_union_map *umap, |
2551 | __isl_keep isl_union_set *uset) |
2552 | { |
2553 | int empty; |
2554 | |
2555 | umap = isl_union_map_copy(umap); |
2556 | umap = isl_union_map_intersect_domain(umap, isl_union_set_copy(uset)); |
2557 | empty = isl_union_map_is_empty(umap); |
2558 | isl_union_map_free(umap); |
2559 | |
2560 | return empty < 0 ? -1 : !empty; |
2561 | } |
2562 | |
2563 | /* Does the range of "umap" intersect "uset"? |
2564 | */ |
2565 | static int range_intersects(__isl_keep isl_union_map *umap, |
2566 | __isl_keep isl_union_set *uset) |
2567 | { |
2568 | int empty; |
2569 | |
2570 | umap = isl_union_map_copy(umap); |
2571 | umap = isl_union_map_intersect_range(umap, isl_union_set_copy(uset)); |
2572 | empty = isl_union_map_is_empty(umap); |
2573 | isl_union_map_free(umap); |
2574 | |
2575 | return empty < 0 ? -1 : !empty; |
2576 | } |
2577 | |
2578 | /* Are the condition dependences of "edge" local with respect to |
2579 | * the current schedule? |
2580 | * |
2581 | * That is, are domain and range of the condition dependences mapped |
2582 | * to the same point? |
2583 | * |
2584 | * In other words, is the condition false? |
2585 | */ |
2586 | static int is_condition_false(struct isl_sched_edge *edge) |
2587 | { |
2588 | isl_union_map *umap; |
2589 | isl_map *map, *sched, *test; |
2590 | int empty, local; |
2591 | |
2592 | empty = isl_union_map_is_empty(edge->tagged_condition); |
2593 | if (empty < 0 || empty) |
2594 | return empty; |
2595 | |
2596 | umap = isl_union_map_copy(edge->tagged_condition); |
2597 | umap = isl_union_map_zip(umap); |
2598 | umap = isl_union_set_unwrap(isl_union_map_domain(umap)); |
2599 | map = isl_map_from_union_map(umap); |
2600 | |
2601 | sched = node_extract_schedule(edge->src); |
2602 | map = isl_map_apply_domain(map, sched); |
2603 | sched = node_extract_schedule(edge->dst); |
2604 | map = isl_map_apply_range(map, sched); |
2605 | |
2606 | test = isl_map_identity(isl_map_get_space(map)); |
2607 | local = isl_map_is_subset(map, test); |
2608 | isl_map_free(map); |
2609 | isl_map_free(test); |
2610 | |
2611 | return local; |
2612 | } |
2613 | |
2614 | /* For each conditional validity constraint that is adjacent |
2615 | * to a condition with domain in condition_source or range in condition_sink, |
2616 | * turn it into an unconditional validity constraint. |
2617 | */ |
2618 | static int unconditionalize_adjacent_validity(struct isl_sched_graph *graph, |
2619 | __isl_take isl_union_set *condition_source, |
2620 | __isl_take isl_union_set *condition_sink) |
2621 | { |
2622 | int i; |
2623 | |
2624 | condition_source = isl_union_set_coalesce(condition_source); |
2625 | condition_sink = isl_union_set_coalesce(condition_sink); |
2626 | |
2627 | for (i = 0; i < graph->n_edge; ++i) { |
2628 | int adjacent; |
2629 | isl_union_map *validity; |
2630 | |
2631 | if (!graph->edge[i].conditional_validity) |
2632 | continue; |
2633 | if (graph->edge[i].validity) |
2634 | continue; |
2635 | |
2636 | validity = graph->edge[i].tagged_validity; |
2637 | adjacent = domain_intersects(validity, condition_sink); |
2638 | if (adjacent >= 0 && !adjacent) |
2639 | adjacent = range_intersects(validity, condition_source); |
2640 | if (adjacent < 0) |
2641 | goto error; |
2642 | if (!adjacent) |
2643 | continue; |
2644 | |
2645 | graph->edge[i].validity = 1; |
2646 | } |
2647 | |
2648 | isl_union_set_free(condition_source); |
2649 | isl_union_set_free(condition_sink); |
2650 | return 0; |
2651 | error: |
2652 | isl_union_set_free(condition_source); |
2653 | isl_union_set_free(condition_sink); |
2654 | return -1; |
2655 | } |
2656 | |
2657 | /* Update the dependence relations of all edges based on the current schedule |
2658 | * and enforce conditional validity constraints that are adjacent |
2659 | * to satisfied condition constraints. |
2660 | * |
2661 | * First check if any of the condition constraints are satisfied |
2662 | * (i.e., not local to the outer schedule) and keep track of |
2663 | * their domain and range. |
2664 | * Then update all dependence relations (which removes the non-local |
2665 | * constraints). |
2666 | * Finally, if any condition constraints turned out to be satisfied, |
2667 | * then turn all adjacent conditional validity constraints into |
2668 | * unconditional validity constraints. |
2669 | */ |
2670 | static int update_edges(isl_ctx *ctx, struct isl_sched_graph *graph) |
2671 | { |
2672 | int i; |
2673 | int any = 0; |
2674 | isl_union_set *source, *sink; |
2675 | |
2676 | source = isl_union_set_empty(isl_space_params_alloc(ctx, 0)); |
2677 | sink = isl_union_set_empty(isl_space_params_alloc(ctx, 0)); |
2678 | for (i = 0; i < graph->n_edge; ++i) { |
2679 | int local; |
2680 | isl_union_set *uset; |
2681 | isl_union_map *umap; |
2682 | |
2683 | if (!graph->edge[i].condition) |
2684 | continue; |
2685 | if (graph->edge[i].local) |
2686 | continue; |
2687 | local = is_condition_false(&graph->edge[i]); |
2688 | if (local < 0) |
2689 | goto error; |
2690 | if (local) |
2691 | continue; |
2692 | |
2693 | any = 1; |
2694 | |
2695 | umap = isl_union_map_copy(graph->edge[i].tagged_condition); |
2696 | uset = isl_union_map_domain(umap); |
2697 | source = isl_union_set_union(source, uset); |
2698 | |
2699 | umap = isl_union_map_copy(graph->edge[i].tagged_condition); |
2700 | uset = isl_union_map_range(umap); |
2701 | sink = isl_union_set_union(sink, uset); |
2702 | } |
2703 | |
2704 | for (i = graph->n_edge - 1; i >= 0; --i) { |
2705 | if (update_edge(graph, &graph->edge[i]) < 0) |
2706 | goto error; |
2707 | } |
2708 | |
2709 | if (any) |
2710 | return unconditionalize_adjacent_validity(graph, source, sink); |
2711 | |
2712 | isl_union_set_free(source); |
2713 | isl_union_set_free(sink); |
2714 | return 0; |
2715 | error: |
2716 | isl_union_set_free(source); |
2717 | isl_union_set_free(sink); |
2718 | return -1; |
2719 | } |
2720 | |
2721 | static void next_band(struct isl_sched_graph *graph) |
2722 | { |
2723 | graph->band_start = graph->n_total_row; |
2724 | } |
2725 | |
2726 | /* Return the union of the universe domains of the nodes in "graph" |
2727 | * that satisfy "pred". |
2728 | */ |
2729 | static __isl_give isl_union_set *isl_sched_graph_domain(isl_ctx *ctx, |
2730 | struct isl_sched_graph *graph, |
2731 | int (*pred)(struct isl_sched_node *node, int data), int data) |
2732 | { |
2733 | int i; |
2734 | isl_setisl_map *set; |
2735 | isl_union_set *dom; |
2736 | |
2737 | for (i = 0; i < graph->n; ++i) |
2738 | if (pred(&graph->node[i], data)) |
2739 | break; |
2740 | |
2741 | if (i >= graph->n) |
2742 | isl_die(ctx, isl_error_internal,do { isl_handle_error(ctx, isl_error_internal, "empty component" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2743); return ((void*)0); } while (0) |
2743 | "empty component", return NULL)do { isl_handle_error(ctx, isl_error_internal, "empty component" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2743); return ((void*)0); } while (0); |
2744 | |
2745 | set = isl_set_universe(isl_space_copy(graph->node[i].space)); |
2746 | dom = isl_union_set_from_set(set); |
2747 | |
2748 | for (i = i + 1; i < graph->n; ++i) { |
2749 | if (!pred(&graph->node[i], data)) |
2750 | continue; |
2751 | set = isl_set_universe(isl_space_copy(graph->node[i].space)); |
2752 | dom = isl_union_set_union(dom, isl_union_set_from_set(set)); |
2753 | } |
2754 | |
2755 | return dom; |
2756 | } |
2757 | |
2758 | /* Return a list of unions of universe domains, where each element |
2759 | * in the list corresponds to an SCC (or WCC) indexed by node->scc. |
2760 | */ |
2761 | static __isl_give isl_union_set_list *extract_sccs(isl_ctx *ctx, |
2762 | struct isl_sched_graph *graph) |
2763 | { |
2764 | int i; |
2765 | isl_union_set_list *filters; |
2766 | |
2767 | filters = isl_union_set_list_alloc(ctx, graph->scc); |
2768 | for (i = 0; i < graph->scc; ++i) { |
2769 | isl_union_set *dom; |
2770 | |
2771 | dom = isl_sched_graph_domain(ctx, graph, &node_scc_exactly, i); |
2772 | filters = isl_union_set_list_add(filters, dom); |
2773 | } |
2774 | |
2775 | return filters; |
2776 | } |
2777 | |
2778 | /* Return a list of two unions of universe domains, one for the SCCs up |
2779 | * to and including graph->src_scc and another for the other SCCS. |
2780 | */ |
2781 | static __isl_give isl_union_set_list *extract_split(isl_ctx *ctx, |
2782 | struct isl_sched_graph *graph) |
2783 | { |
2784 | isl_union_set *dom; |
2785 | isl_union_set_list *filters; |
2786 | |
2787 | filters = isl_union_set_list_alloc(ctx, 2); |
2788 | dom = isl_sched_graph_domain(ctx, graph, |
2789 | &node_scc_at_most, graph->src_scc); |
2790 | filters = isl_union_set_list_add(filters, dom); |
2791 | dom = isl_sched_graph_domain(ctx, graph, |
2792 | &node_scc_at_least, graph->src_scc + 1); |
2793 | filters = isl_union_set_list_add(filters, dom); |
2794 | |
2795 | return filters; |
2796 | } |
2797 | |
2798 | /* Copy nodes that satisfy node_pred from the src dependence graph |
2799 | * to the dst dependence graph. |
2800 | */ |
2801 | static int copy_nodes(struct isl_sched_graph *dst, struct isl_sched_graph *src, |
2802 | int (*node_pred)(struct isl_sched_node *node, int data), int data) |
2803 | { |
2804 | int i; |
2805 | |
2806 | dst->n = 0; |
2807 | for (i = 0; i < src->n; ++i) { |
2808 | int j; |
2809 | |
2810 | if (!node_pred(&src->node[i], data)) |
2811 | continue; |
2812 | |
2813 | j = dst->n; |
2814 | dst->node[j].space = isl_space_copy(src->node[i].space); |
2815 | dst->node[j].compressed = src->node[i].compressed; |
2816 | dst->node[j].hull = isl_set_copy(src->node[i].hull); |
2817 | dst->node[j].compress = |
2818 | isl_multi_aff_copy(src->node[i].compress); |
2819 | dst->node[j].decompress = |
2820 | isl_multi_aff_copy(src->node[i].decompress); |
2821 | dst->node[j].nvar = src->node[i].nvar; |
2822 | dst->node[j].nparam = src->node[i].nparam; |
2823 | dst->node[j].sched = isl_mat_copy(src->node[i].sched); |
2824 | dst->node[j].sched_map = isl_map_copy(src->node[i].sched_map); |
2825 | dst->node[j].coincident = src->node[i].coincident; |
2826 | dst->n++; |
2827 | |
2828 | if (!dst->node[j].space || !dst->node[j].sched) |
2829 | return -1; |
2830 | if (dst->node[j].compressed && |
2831 | (!dst->node[j].hull || !dst->node[j].compress || |
2832 | !dst->node[j].decompress)) |
2833 | return -1; |
2834 | } |
2835 | |
2836 | return 0; |
2837 | } |
2838 | |
2839 | /* Copy non-empty edges that satisfy edge_pred from the src dependence graph |
2840 | * to the dst dependence graph. |
2841 | * If the source or destination node of the edge is not in the destination |
2842 | * graph, then it must be a backward proximity edge and it should simply |
2843 | * be ignored. |
2844 | */ |
2845 | static int copy_edges(isl_ctx *ctx, struct isl_sched_graph *dst, |
2846 | struct isl_sched_graph *src, |
2847 | int (*edge_pred)(struct isl_sched_edge *edge, int data), int data) |
2848 | { |
2849 | int i; |
2850 | enum isl_edge_type t; |
2851 | |
2852 | dst->n_edge = 0; |
2853 | for (i = 0; i < src->n_edge; ++i) { |
2854 | struct isl_sched_edge *edge = &src->edge[i]; |
2855 | isl_map *map; |
2856 | isl_union_map *tagged_condition; |
2857 | isl_union_map *tagged_validity; |
2858 | struct isl_sched_node *dst_src, *dst_dst; |
2859 | |
2860 | if (!edge_pred(edge, data)) |
2861 | continue; |
2862 | |
2863 | if (isl_map_plain_is_empty(edge->map)) |
2864 | continue; |
2865 | |
2866 | dst_src = graph_find_node(ctx, dst, edge->src->space); |
2867 | dst_dst = graph_find_node(ctx, dst, edge->dst->space); |
2868 | if (!dst_src || !dst_dst) { |
2869 | if (edge->validity || edge->conditional_validity) |
2870 | isl_die(ctx, isl_error_internal,do { isl_handle_error(ctx, isl_error_internal, "backward (conditional) validity edge" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2872); return -1; } while (0) |
2871 | "backward (conditional) validity edge",do { isl_handle_error(ctx, isl_error_internal, "backward (conditional) validity edge" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2872); return -1; } while (0) |
2872 | return -1)do { isl_handle_error(ctx, isl_error_internal, "backward (conditional) validity edge" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 2872); return -1; } while (0); |
2873 | continue; |
2874 | } |
2875 | |
2876 | map = isl_map_copy(edge->map); |
2877 | tagged_condition = isl_union_map_copy(edge->tagged_condition); |
2878 | tagged_validity = isl_union_map_copy(edge->tagged_validity); |
2879 | |
2880 | dst->edge[dst->n_edge].src = dst_src; |
2881 | dst->edge[dst->n_edge].dst = dst_dst; |
2882 | dst->edge[dst->n_edge].map = map; |
2883 | dst->edge[dst->n_edge].tagged_condition = tagged_condition; |
2884 | dst->edge[dst->n_edge].tagged_validity = tagged_validity; |
2885 | dst->edge[dst->n_edge].validity = edge->validity; |
2886 | dst->edge[dst->n_edge].proximity = edge->proximity; |
2887 | dst->edge[dst->n_edge].coincidence = edge->coincidence; |
2888 | dst->edge[dst->n_edge].condition = edge->condition; |
2889 | dst->edge[dst->n_edge].conditional_validity = |
2890 | edge->conditional_validity; |
2891 | dst->n_edge++; |
2892 | |
2893 | if (edge->tagged_condition && !tagged_condition) |
2894 | return -1; |
2895 | if (edge->tagged_validity && !tagged_validity) |
2896 | return -1; |
2897 | |
2898 | for (t = isl_edge_first; t <= isl_edge_last; ++t) { |
2899 | if (edge != |
2900 | graph_find_edge(src, t, edge->src, edge->dst)) |
2901 | continue; |
2902 | if (graph_edge_table_add(ctx, dst, t, |
2903 | &dst->edge[dst->n_edge - 1]) < 0) |
2904 | return -1; |
2905 | } |
2906 | } |
2907 | |
2908 | return 0; |
2909 | } |
2910 | |
2911 | /* Compute the maximal number of variables over all nodes. |
2912 | * This is the maximal number of linearly independent schedule |
2913 | * rows that we need to compute. |
2914 | * Just in case we end up in a part of the dependence graph |
2915 | * with only lower-dimensional domains, we make sure we will |
2916 | * compute the required amount of extra linearly independent rows. |
2917 | */ |
2918 | static int compute_maxvar(struct isl_sched_graph *graph) |
2919 | { |
2920 | int i; |
2921 | |
2922 | graph->maxvar = 0; |
2923 | for (i = 0; i < graph->n; ++i) { |
2924 | struct isl_sched_node *node = &graph->node[i]; |
2925 | int nvar; |
2926 | |
2927 | if (node_update_cmap(node) < 0) |
2928 | return -1; |
2929 | nvar = node->nvar + graph->n_row - node->rank; |
2930 | if (nvar > graph->maxvar) |
2931 | graph->maxvar = nvar; |
2932 | } |
2933 | |
2934 | return 0; |
2935 | } |
2936 | |
2937 | static __isl_give isl_schedule_node *compute_schedule(isl_schedule_node *node, |
2938 | struct isl_sched_graph *graph); |
2939 | static __isl_give isl_schedule_node *compute_schedule_wcc( |
2940 | isl_schedule_node *node, struct isl_sched_graph *graph); |
2941 | |
2942 | /* Compute a schedule for a subgraph of "graph". In particular, for |
2943 | * the graph composed of nodes that satisfy node_pred and edges that |
2944 | * that satisfy edge_pred. The caller should precompute the number |
2945 | * of nodes and edges that satisfy these predicates and pass them along |
2946 | * as "n" and "n_edge". |
2947 | * If the subgraph is known to consist of a single component, then wcc should |
2948 | * be set and then we call compute_schedule_wcc on the constructed subgraph. |
2949 | * Otherwise, we call compute_schedule, which will check whether the subgraph |
2950 | * is connected. |
2951 | * |
2952 | * The schedule is inserted at "node" and the updated schedule node |
2953 | * is returned. |
2954 | */ |
2955 | static __isl_give isl_schedule_node *compute_sub_schedule( |
2956 | __isl_take isl_schedule_node *node, isl_ctx *ctx, |
2957 | struct isl_sched_graph *graph, int n, int n_edge, |
2958 | int (*node_pred)(struct isl_sched_node *node, int data), |
2959 | int (*edge_pred)(struct isl_sched_edge *edge, int data), |
2960 | int data, int wcc) |
2961 | { |
2962 | struct isl_sched_graph split = { 0 }; |
2963 | int t; |
2964 | |
2965 | if (graph_alloc(ctx, &split, n, n_edge) < 0) |
2966 | goto error; |
2967 | if (copy_nodes(&split, graph, node_pred, data) < 0) |
2968 | goto error; |
2969 | if (graph_init_table(ctx, &split) < 0) |
2970 | goto error; |
2971 | for (t = 0; t <= isl_edge_last; ++t) |
2972 | split.max_edge[t] = graph->max_edge[t]; |
2973 | if (graph_init_edge_tables(ctx, &split) < 0) |
2974 | goto error; |
2975 | if (copy_edges(ctx, &split, graph, edge_pred, data) < 0) |
2976 | goto error; |
2977 | split.n_row = graph->n_row; |
2978 | split.max_row = graph->max_row; |
2979 | split.n_total_row = graph->n_total_row; |
2980 | split.band_start = graph->band_start; |
2981 | |
2982 | if (wcc) |
2983 | node = compute_schedule_wcc(node, &split); |
2984 | else |
2985 | node = compute_schedule(node, &split); |
2986 | |
2987 | graph_free(ctx, &split); |
2988 | return node; |
2989 | error: |
2990 | graph_free(ctx, &split); |
2991 | return isl_schedule_node_free(node); |
2992 | } |
2993 | |
2994 | static int edge_scc_exactly(struct isl_sched_edge *edge, int scc) |
2995 | { |
2996 | return edge->src->scc == scc && edge->dst->scc == scc; |
2997 | } |
2998 | |
2999 | static int edge_dst_scc_at_most(struct isl_sched_edge *edge, int scc) |
3000 | { |
3001 | return edge->dst->scc <= scc; |
3002 | } |
3003 | |
3004 | static int edge_src_scc_at_least(struct isl_sched_edge *edge, int scc) |
3005 | { |
3006 | return edge->src->scc >= scc; |
3007 | } |
3008 | |
3009 | /* Reset the current band by dropping all its schedule rows. |
3010 | */ |
3011 | static int reset_band(struct isl_sched_graph *graph) |
3012 | { |
3013 | int i; |
3014 | int drop; |
3015 | |
3016 | drop = graph->n_total_row - graph->band_start; |
3017 | graph->n_total_row -= drop; |
3018 | graph->n_row -= drop; |
3019 | |
3020 | for (i = 0; i < graph->n; ++i) { |
3021 | struct isl_sched_node *node = &graph->node[i]; |
3022 | |
3023 | isl_map_free(node->sched_map); |
3024 | node->sched_map = NULL((void*)0); |
3025 | |
3026 | node->sched = isl_mat_drop_rows(node->sched, |
3027 | graph->band_start, drop); |
3028 | |
3029 | if (!node->sched) |
3030 | return -1; |
3031 | } |
3032 | |
3033 | return 0; |
3034 | } |
3035 | |
3036 | /* Split the current graph into two parts and compute a schedule for each |
3037 | * part individually. In particular, one part consists of all SCCs up |
3038 | * to and including graph->src_scc, while the other part contains the other |
3039 | * SCCS. The split is enforced by a sequence node inserted at position "node" |
3040 | * in the schedule tree. Return the updated schedule node. |
3041 | * |
3042 | * The current band is reset. It would be possible to reuse |
3043 | * the previously computed rows as the first rows in the next |
3044 | * band, but recomputing them may result in better rows as we are looking |
3045 | * at a smaller part of the dependence graph. |
3046 | */ |
3047 | static __isl_give isl_schedule_node *compute_split_schedule( |
3048 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph) |
3049 | { |
3050 | int i, n, e1, e2; |
3051 | isl_ctx *ctx; |
3052 | isl_union_set_list *filters; |
3053 | |
3054 | if (!node) |
3055 | return NULL((void*)0); |
3056 | |
3057 | if (reset_band(graph) < 0) |
3058 | return isl_schedule_node_free(node); |
3059 | |
3060 | n = 0; |
3061 | for (i = 0; i < graph->n; ++i) { |
3062 | struct isl_sched_node *node = &graph->node[i]; |
3063 | int before = node->scc <= graph->src_scc; |
3064 | |
3065 | if (before) |
3066 | n++; |
3067 | } |
3068 | |
3069 | e1 = e2 = 0; |
3070 | for (i = 0; i < graph->n_edge; ++i) { |
3071 | if (graph->edge[i].dst->scc <= graph->src_scc) |
3072 | e1++; |
3073 | if (graph->edge[i].src->scc > graph->src_scc) |
3074 | e2++; |
3075 | } |
3076 | |
3077 | next_band(graph); |
3078 | |
3079 | ctx = isl_schedule_node_get_ctx(node); |
3080 | filters = extract_split(ctx, graph); |
3081 | node = isl_schedule_node_insert_sequence(node, filters); |
3082 | node = isl_schedule_node_child(node, 0); |
3083 | node = isl_schedule_node_child(node, 0); |
3084 | |
3085 | node = compute_sub_schedule(node, ctx, graph, n, e1, |
3086 | &node_scc_at_most, &edge_dst_scc_at_most, |
3087 | graph->src_scc, 0); |
3088 | node = isl_schedule_node_parent(node); |
3089 | node = isl_schedule_node_next_sibling(node); |
3090 | node = isl_schedule_node_child(node, 0); |
3091 | node = compute_sub_schedule(node, ctx, graph, graph->n - n, e2, |
3092 | &node_scc_at_least, &edge_src_scc_at_least, |
3093 | graph->src_scc + 1, 0); |
3094 | node = isl_schedule_node_parent(node); |
3095 | node = isl_schedule_node_parent(node); |
3096 | |
3097 | return node; |
3098 | } |
3099 | |
3100 | /* Insert a band node at position "node" in the schedule tree corresponding |
3101 | * to the current band in "graph". Mark the band node permutable |
3102 | * if "permutable" is set. |
3103 | * The partial schedules and the coincidence property are extracted |
3104 | * from the graph nodes. |
3105 | * Return the updated schedule node. |
3106 | */ |
3107 | static __isl_give isl_schedule_node *insert_current_band( |
3108 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph, |
3109 | int permutable) |
3110 | { |
3111 | int i; |
3112 | int start, end, n; |
3113 | isl_multi_aff *ma; |
3114 | isl_multi_pw_aff *mpa; |
3115 | isl_multi_union_pw_aff *mupa; |
3116 | |
3117 | if (!node) |
3118 | return NULL((void*)0); |
3119 | |
3120 | if (graph->n < 1) |
3121 | isl_die(isl_schedule_node_get_ctx(node), isl_error_internal,do { isl_handle_error(isl_schedule_node_get_ctx(node), isl_error_internal , "graph should have at least one node", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3123); return isl_schedule_node_free(node); } while (0) |
3122 | "graph should have at least one node",do { isl_handle_error(isl_schedule_node_get_ctx(node), isl_error_internal , "graph should have at least one node", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3123); return isl_schedule_node_free(node); } while (0) |
3123 | return isl_schedule_node_free(node))do { isl_handle_error(isl_schedule_node_get_ctx(node), isl_error_internal , "graph should have at least one node", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3123); return isl_schedule_node_free(node); } while (0); |
3124 | |
3125 | start = graph->band_start; |
3126 | end = graph->n_total_row; |
3127 | n = end - start; |
3128 | |
3129 | ma = node_extract_partial_schedule_multi_aff(&graph->node[0], start, n); |
3130 | mpa = isl_multi_pw_aff_from_multi_aff(ma); |
3131 | mupa = isl_multi_union_pw_aff_from_multi_pw_aff(mpa); |
3132 | |
3133 | for (i = 1; i < graph->n; ++i) { |
3134 | isl_multi_union_pw_aff *mupa_i; |
3135 | |
3136 | ma = node_extract_partial_schedule_multi_aff(&graph->node[i], |
3137 | start, n); |
3138 | mpa = isl_multi_pw_aff_from_multi_aff(ma); |
3139 | mupa_i = isl_multi_union_pw_aff_from_multi_pw_aff(mpa); |
3140 | mupa = isl_multi_union_pw_aff_union_add(mupa, mupa_i); |
3141 | } |
3142 | node = isl_schedule_node_insert_partial_schedule(node, mupa); |
3143 | |
3144 | for (i = 0; i < n; ++i) |
3145 | node = isl_schedule_node_band_member_set_coincident(node, i, |
3146 | graph->node[0].coincident[start + i]); |
3147 | node = isl_schedule_node_band_set_permutable(node, permutable); |
3148 | |
3149 | return node; |
3150 | } |
3151 | |
3152 | /* Update the dependence relations based on the current schedule, |
3153 | * add the current band to "node" and then continue with the computation |
3154 | * of the next band. |
3155 | * Return the updated schedule node. |
3156 | */ |
3157 | static __isl_give isl_schedule_node *compute_next_band( |
3158 | __isl_take isl_schedule_node *node, |
3159 | struct isl_sched_graph *graph, int permutable) |
3160 | { |
3161 | isl_ctx *ctx; |
3162 | |
3163 | if (!node) |
3164 | return NULL((void*)0); |
3165 | |
3166 | ctx = isl_schedule_node_get_ctx(node); |
3167 | if (update_edges(ctx, graph) < 0) |
3168 | return isl_schedule_node_free(node); |
3169 | node = insert_current_band(node, graph, permutable); |
3170 | next_band(graph); |
3171 | |
3172 | node = isl_schedule_node_child(node, 0); |
3173 | node = compute_schedule(node, graph); |
3174 | node = isl_schedule_node_parent(node); |
3175 | |
3176 | return node; |
3177 | } |
3178 | |
3179 | /* Add constraints to graph->lp that force the dependence "map" (which |
3180 | * is part of the dependence relation of "edge") |
3181 | * to be respected and attempt to carry it, where the edge is one from |
3182 | * a node j to itself. "pos" is the sequence number of the given map. |
3183 | * That is, add constraints that enforce |
3184 | * |
3185 | * (c_j_0 + c_j_n n + c_j_x y) - (c_j_0 + c_j_n n + c_j_x x) |
3186 | * = c_j_x (y - x) >= e_i |
3187 | * |
3188 | * for each (x,y) in R. |
3189 | * We obtain general constraints on coefficients (c_0, c_n, c_x) |
3190 | * of valid constraints for (y - x) and then plug in (-e_i, 0, c_j_x), |
3191 | * with each coefficient in c_j_x represented as a pair of non-negative |
3192 | * coefficients. |
3193 | */ |
3194 | static int add_intra_constraints(struct isl_sched_graph *graph, |
3195 | struct isl_sched_edge *edge, __isl_take isl_map *map, int pos) |
3196 | { |
3197 | unsigned total; |
3198 | isl_ctx *ctx = isl_map_get_ctx(map); |
3199 | isl_space *dim; |
3200 | isl_dim_map *dim_map; |
3201 | isl_basic_setisl_basic_map *coef; |
3202 | struct isl_sched_node *node = edge->src; |
3203 | |
3204 | coef = intra_coefficients(graph, node, map); |
3205 | if (!coef) |
3206 | return -1; |
3207 | |
3208 | dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); |
3209 | |
3210 | total = isl_basic_set_total_dim(graph->lp); |
3211 | dim_map = isl_dim_map_alloc(ctx, total); |
3212 | isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1); |
3213 | isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2, |
3214 | isl_space_dim(dim, isl_dim_set), 1, |
3215 | node->nvar, -1); |
3216 | isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2, |
3217 | isl_space_dim(dim, isl_dim_set), 1, |
3218 | node->nvar, 1); |
3219 | graph->lp = isl_basic_set_extend_constraints(graph->lp, |
3220 | coef->n_eq, coef->n_ineq); |
3221 | graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, |
3222 | coef, dim_map); |
3223 | isl_space_free(dim); |
3224 | |
3225 | return 0; |
3226 | } |
3227 | |
3228 | /* Add constraints to graph->lp that force the dependence "map" (which |
3229 | * is part of the dependence relation of "edge") |
3230 | * to be respected and attempt to carry it, where the edge is one from |
3231 | * node j to node k. "pos" is the sequence number of the given map. |
3232 | * That is, add constraints that enforce |
3233 | * |
3234 | * (c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i |
3235 | * |
3236 | * for each (x,y) in R. |
3237 | * We obtain general constraints on coefficients (c_0, c_n, c_x) |
3238 | * of valid constraints for R and then plug in |
3239 | * (-e_i + c_k_0 - c_j_0, c_k_n - c_j_n, c_k_x - c_j_x) |
3240 | * with each coefficient (except e_i, c_k_0 and c_j_0) |
3241 | * represented as a pair of non-negative coefficients. |
3242 | */ |
3243 | static int add_inter_constraints(struct isl_sched_graph *graph, |
3244 | struct isl_sched_edge *edge, __isl_take isl_map *map, int pos) |
3245 | { |
3246 | unsigned total; |
3247 | isl_ctx *ctx = isl_map_get_ctx(map); |
3248 | isl_space *dim; |
3249 | isl_dim_map *dim_map; |
3250 | isl_basic_setisl_basic_map *coef; |
3251 | struct isl_sched_node *src = edge->src; |
3252 | struct isl_sched_node *dst = edge->dst; |
3253 | |
3254 | coef = inter_coefficients(graph, edge, map); |
3255 | if (!coef) |
3256 | return -1; |
3257 | |
3258 | dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); |
3259 | |
3260 | total = isl_basic_set_total_dim(graph->lp); |
3261 | dim_map = isl_dim_map_alloc(ctx, total); |
3262 | |
3263 | isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1); |
3264 | |
3265 | isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1); |
3266 | isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1); |
3267 | isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1); |
3268 | isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2, |
3269 | isl_space_dim(dim, isl_dim_set) + src->nvar, 1, |
3270 | dst->nvar, -1); |
3271 | isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2, |
3272 | isl_space_dim(dim, isl_dim_set) + src->nvar, 1, |
3273 | dst->nvar, 1); |
3274 | |
3275 | isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1); |
3276 | isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1); |
3277 | isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1); |
3278 | isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2, |
3279 | isl_space_dim(dim, isl_dim_set), 1, |
3280 | src->nvar, 1); |
3281 | isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2, |
3282 | isl_space_dim(dim, isl_dim_set), 1, |
3283 | src->nvar, -1); |
3284 | |
3285 | graph->lp = isl_basic_set_extend_constraints(graph->lp, |
3286 | coef->n_eq, coef->n_ineq); |
3287 | graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, |
3288 | coef, dim_map); |
3289 | isl_space_free(dim); |
3290 | |
3291 | return 0; |
3292 | } |
3293 | |
3294 | /* Add constraints to graph->lp that force all (conditional) validity |
3295 | * dependences to be respected and attempt to carry them. |
3296 | */ |
3297 | static int add_all_constraints(struct isl_sched_graph *graph) |
3298 | { |
3299 | int i, j; |
3300 | int pos; |
3301 | |
3302 | pos = 0; |
3303 | for (i = 0; i < graph->n_edge; ++i) { |
3304 | struct isl_sched_edge *edge= &graph->edge[i]; |
3305 | |
3306 | if (!edge->validity && !edge->conditional_validity) |
3307 | continue; |
3308 | |
3309 | for (j = 0; j < edge->map->n; ++j) { |
3310 | isl_basic_map *bmap; |
3311 | isl_map *map; |
3312 | |
3313 | bmap = isl_basic_map_copy(edge->map->p[j]); |
3314 | map = isl_map_from_basic_map(bmap); |
3315 | |
3316 | if (edge->src == edge->dst && |
3317 | add_intra_constraints(graph, edge, map, pos) < 0) |
3318 | return -1; |
3319 | if (edge->src != edge->dst && |
3320 | add_inter_constraints(graph, edge, map, pos) < 0) |
3321 | return -1; |
3322 | ++pos; |
3323 | } |
3324 | } |
3325 | |
3326 | return 0; |
3327 | } |
3328 | |
3329 | /* Count the number of equality and inequality constraints |
3330 | * that will be added to the carry_lp problem. |
3331 | * We count each edge exactly once. |
3332 | */ |
3333 | static int count_all_constraints(struct isl_sched_graph *graph, |
3334 | int *n_eq, int *n_ineq) |
3335 | { |
3336 | int i, j; |
3337 | |
3338 | *n_eq = *n_ineq = 0; |
3339 | for (i = 0; i < graph->n_edge; ++i) { |
3340 | struct isl_sched_edge *edge= &graph->edge[i]; |
3341 | for (j = 0; j < edge->map->n; ++j) { |
3342 | isl_basic_map *bmap; |
3343 | isl_map *map; |
3344 | |
3345 | bmap = isl_basic_map_copy(edge->map->p[j]); |
3346 | map = isl_map_from_basic_map(bmap); |
3347 | |
3348 | if (count_map_constraints(graph, edge, map, |
3349 | n_eq, n_ineq, 1, 0) < 0) |
3350 | return -1; |
3351 | } |
3352 | } |
3353 | |
3354 | return 0; |
3355 | } |
3356 | |
3357 | /* Construct an LP problem for finding schedule coefficients |
3358 | * such that the schedule carries as many dependences as possible. |
3359 | * In particular, for each dependence i, we bound the dependence distance |
3360 | * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum |
3361 | * of all e_i's. Dependences with e_i = 0 in the solution are simply |
3362 | * respected, while those with e_i > 0 (in practice e_i = 1) are carried. |
3363 | * Note that if the dependence relation is a union of basic maps, |
3364 | * then we have to consider each basic map individually as it may only |
3365 | * be possible to carry the dependences expressed by some of those |
3366 | * basic maps and not all of them. |
3367 | * Below, we consider each of those basic maps as a separate "edge". |
3368 | * |
3369 | * All variables of the LP are non-negative. The actual coefficients |
3370 | * may be negative, so each coefficient is represented as the difference |
3371 | * of two non-negative variables. The negative part always appears |
3372 | * immediately before the positive part. |
3373 | * Other than that, the variables have the following order |
3374 | * |
3375 | * - sum of (1 - e_i) over all edges |
3376 | * - sum of positive and negative parts of all c_n coefficients |
3377 | * (unconstrained when computing non-parametric schedules) |
3378 | * - sum of positive and negative parts of all c_x coefficients |
3379 | * - for each edge |
3380 | * - e_i |
3381 | * - for each node |
3382 | * - c_i_0 |
3383 | * - positive and negative parts of c_i_n (if parametric) |
3384 | * - positive and negative parts of c_i_x |
3385 | * |
3386 | * The constraints are those from the (validity) edges plus three equalities |
3387 | * to express the sums and n_edge inequalities to express e_i <= 1. |
3388 | */ |
3389 | static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph) |
3390 | { |
3391 | int i, j; |
3392 | int k; |
3393 | isl_space *dim; |
3394 | unsigned total; |
3395 | int n_eq, n_ineq; |
3396 | int n_edge; |
3397 | |
3398 | n_edge = 0; |
3399 | for (i = 0; i < graph->n_edge; ++i) |
3400 | n_edge += graph->edge[i].map->n; |
3401 | |
3402 | total = 3 + n_edge; |
3403 | for (i = 0; i < graph->n; ++i) { |
3404 | struct isl_sched_node *node = &graph->node[graph->sorted[i]]; |
3405 | node->start = total; |
3406 | total += 1 + 2 * (node->nparam + node->nvar); |
3407 | } |
3408 | |
3409 | if (count_all_constraints(graph, &n_eq, &n_ineq) < 0) |
3410 | return -1; |
3411 | |
3412 | dim = isl_space_set_alloc(ctx, 0, total); |
3413 | isl_basic_set_free(graph->lp); |
3414 | n_eq += 3; |
3415 | n_ineq += n_edge; |
3416 | graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq); |
3417 | graph->lp = isl_basic_set_set_rational(graph->lp); |
3418 | |
3419 | k = isl_basic_set_alloc_equality(graph->lp); |
3420 | if (k < 0) |
3421 | return -1; |
3422 | isl_seq_clr(graph->lp->eq[k], 1 + total); |
3423 | isl_int_set_si(graph->lp->eq[k][0], -n_edge)isl_sioimath_set_si((graph->lp->eq[k][0]), -n_edge); |
3424 | isl_int_set_si(graph->lp->eq[k][1], 1)isl_sioimath_set_si((graph->lp->eq[k][1]), 1); |
3425 | for (i = 0; i < n_edge; ++i) |
3426 | isl_int_set_si(graph->lp->eq[k][4 + i], 1)isl_sioimath_set_si((graph->lp->eq[k][4 + i]), 1); |
3427 | |
3428 | k = isl_basic_set_alloc_equality(graph->lp); |
3429 | if (k < 0) |
3430 | return -1; |
3431 | isl_seq_clr(graph->lp->eq[k], 1 + total); |
3432 | isl_int_set_si(graph->lp->eq[k][2], -1)isl_sioimath_set_si((graph->lp->eq[k][2]), -1); |
3433 | for (i = 0; i < graph->n; ++i) { |
3434 | int pos = 1 + graph->node[i].start + 1; |
3435 | |
3436 | for (j = 0; j < 2 * graph->node[i].nparam; ++j) |
3437 | isl_int_set_si(graph->lp->eq[k][pos + j], 1)isl_sioimath_set_si((graph->lp->eq[k][pos + j]), 1); |
3438 | } |
3439 | |
3440 | k = isl_basic_set_alloc_equality(graph->lp); |
3441 | if (k < 0) |
3442 | return -1; |
3443 | isl_seq_clr(graph->lp->eq[k], 1 + total); |
3444 | isl_int_set_si(graph->lp->eq[k][3], -1)isl_sioimath_set_si((graph->lp->eq[k][3]), -1); |
3445 | for (i = 0; i < graph->n; ++i) { |
3446 | struct isl_sched_node *node = &graph->node[i]; |
3447 | int pos = 1 + node->start + 1 + 2 * node->nparam; |
3448 | |
3449 | for (j = 0; j < 2 * node->nvar; ++j) |
3450 | isl_int_set_si(graph->lp->eq[k][pos + j], 1)isl_sioimath_set_si((graph->lp->eq[k][pos + j]), 1); |
3451 | } |
3452 | |
3453 | for (i = 0; i < n_edge; ++i) { |
3454 | k = isl_basic_set_alloc_inequality(graph->lp); |
3455 | if (k < 0) |
3456 | return -1; |
3457 | isl_seq_clr(graph->lp->ineq[k], 1 + total); |
3458 | isl_int_set_si(graph->lp->ineq[k][4 + i], -1)isl_sioimath_set_si((graph->lp->ineq[k][4 + i]), -1); |
3459 | isl_int_set_si(graph->lp->ineq[k][0], 1)isl_sioimath_set_si((graph->lp->ineq[k][0]), 1); |
3460 | } |
3461 | |
3462 | if (add_all_constraints(graph) < 0) |
3463 | return -1; |
3464 | |
3465 | return 0; |
3466 | } |
3467 | |
3468 | static __isl_give isl_schedule_node *compute_component_schedule( |
3469 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph, |
3470 | int wcc); |
3471 | |
3472 | /* Comparison function for sorting the statements based on |
3473 | * the corresponding value in "r". |
3474 | */ |
3475 | static int smaller_value(const void *a, const void *b, void *data) |
3476 | { |
3477 | isl_vec *r = data; |
3478 | const int *i1 = a; |
3479 | const int *i2 = b; |
3480 | |
3481 | return isl_int_cmp(r->el[*i1], r->el[*i2])isl_sioimath_cmp(*(r->el[*i1]), *(r->el[*i2])); |
3482 | } |
3483 | |
3484 | /* If the schedule_split_scaled option is set and if the linear |
3485 | * parts of the scheduling rows for all nodes in the graphs have |
3486 | * a non-trivial common divisor, then split off the remainder of the |
3487 | * constant term modulo this common divisor from the linear part. |
3488 | * Otherwise, insert a band node directly and continue with |
3489 | * the construction of the schedule. |
3490 | * |
3491 | * If a non-trivial common divisor is found, then |
3492 | * the linear part is reduced and the remainder is enforced |
3493 | * by a sequence node with the children placed in the order |
3494 | * of this remainder. |
3495 | * In particular, we assign an scc index based on the remainder and |
3496 | * then rely on compute_component_schedule to insert the sequence and |
3497 | * to continue the schedule construction on each part. |
3498 | */ |
3499 | static __isl_give isl_schedule_node *split_scaled( |
3500 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph) |
3501 | { |
3502 | int i; |
3503 | int row; |
3504 | int scc; |
3505 | isl_ctx *ctx; |
3506 | isl_int gcd, gcd_i; |
3507 | isl_vec *r; |
3508 | int *order; |
3509 | |
3510 | if (!node) |
3511 | return NULL((void*)0); |
3512 | |
3513 | ctx = isl_schedule_node_get_ctx(node); |
3514 | if (!ctx->opt->schedule_split_scaled) |
3515 | return compute_next_band(node, graph, 0); |
3516 | if (graph->n <= 1) |
3517 | return compute_next_band(node, graph, 0); |
3518 | |
3519 | isl_int_init(gcd)isl_sioimath_init((gcd)); |
3520 | isl_int_init(gcd_i)isl_sioimath_init((gcd_i)); |
3521 | |
3522 | isl_int_set_si(gcd, 0)isl_sioimath_set_si((gcd), 0); |
3523 | |
3524 | row = isl_mat_rows(graph->node[0].sched) - 1; |
3525 | |
3526 | for (i = 0; i < graph->n; ++i) { |
3527 | struct isl_sched_node *node = &graph->node[i]; |
3528 | int cols = isl_mat_cols(node->sched); |
3529 | |
3530 | isl_seq_gcd(node->sched->row[row] + 1, cols - 1, &gcd_i); |
3531 | isl_int_gcd(gcd, gcd, gcd_i)isl_sioimath_gcd((gcd), *(gcd), *(gcd_i)); |
3532 | } |
3533 | |
3534 | isl_int_clear(gcd_i)isl_sioimath_clear((gcd_i)); |
3535 | |
3536 | if (isl_int_cmp_si(gcd, 1)isl_sioimath_cmp_si(*(gcd), 1) <= 0) { |
3537 | isl_int_clear(gcd)isl_sioimath_clear((gcd)); |
3538 | return compute_next_band(node, graph, 0); |
3539 | } |
3540 | |
3541 | r = isl_vec_alloc(ctx, graph->n); |
3542 | order = isl_calloc_array(ctx, int, graph->n)((int *)isl_calloc_or_die(ctx, graph->n, sizeof(int))); |
3543 | if (!r || !order) |
3544 | goto error; |
3545 | |
3546 | for (i = 0; i < graph->n; ++i) { |
3547 | struct isl_sched_node *node = &graph->node[i]; |
3548 | |
3549 | order[i] = i; |
3550 | isl_int_fdiv_r(r->el[i], node->sched->row[row][0], gcd)isl_sioimath_fdiv_r((r->el[i]), *(node->sched->row[row ][0]), *(gcd)); |
3551 | isl_int_fdiv_q(node->sched->row[row][0],isl_sioimath_fdiv_q((node->sched->row[row][0]), *(node-> sched->row[row][0]), *(gcd)) |
3552 | node->sched->row[row][0], gcd)isl_sioimath_fdiv_q((node->sched->row[row][0]), *(node-> sched->row[row][0]), *(gcd)); |
3553 | isl_int_mul(node->sched->row[row][0],isl_sioimath_mul((node->sched->row[row][0]), *(node-> sched->row[row][0]), *(gcd)) |
3554 | node->sched->row[row][0], gcd)isl_sioimath_mul((node->sched->row[row][0]), *(node-> sched->row[row][0]), *(gcd)); |
3555 | node->sched = isl_mat_scale_down_row(node->sched, row, gcd); |
3556 | if (!node->sched) |
3557 | goto error; |
3558 | } |
3559 | |
3560 | if (isl_sort(order, graph->n, sizeof(order[0]), &smaller_value, r) < 0) |
3561 | goto error; |
3562 | |
3563 | scc = 0; |
3564 | for (i = 0; i < graph->n; ++i) { |
3565 | if (i > 0 && isl_int_ne(r->el[order[i - 1]], r->el[order[i]])(isl_sioimath_cmp(*(r->el[order[i - 1]]), *(r->el[order [i]])) != 0)) |
3566 | ++scc; |
3567 | graph->node[order[i]].scc = scc; |
3568 | } |
3569 | graph->scc = ++scc; |
3570 | graph->weak = 0; |
3571 | |
3572 | isl_int_clear(gcd)isl_sioimath_clear((gcd)); |
3573 | isl_vec_free(r); |
3574 | free(order); |
3575 | |
3576 | if (update_edges(ctx, graph) < 0) |
3577 | return isl_schedule_node_free(node); |
3578 | node = insert_current_band(node, graph, 0); |
3579 | next_band(graph); |
3580 | |
3581 | node = isl_schedule_node_child(node, 0); |
3582 | node = compute_component_schedule(node, graph, 0); |
3583 | node = isl_schedule_node_parent(node); |
3584 | |
3585 | return node; |
3586 | error: |
3587 | isl_vec_free(r); |
3588 | free(order); |
3589 | isl_int_clear(gcd)isl_sioimath_clear((gcd)); |
3590 | return isl_schedule_node_free(node); |
3591 | } |
3592 | |
3593 | /* Is the schedule row "sol" trivial on node "node"? |
3594 | * That is, is the solution zero on the dimensions orthogonal to |
3595 | * the previously found solutions? |
3596 | * Return 1 if the solution is trivial, 0 if it is not and -1 on error. |
3597 | * |
3598 | * Each coefficient is represented as the difference between |
3599 | * two non-negative values in "sol". "sol" has been computed |
3600 | * in terms of the original iterators (i.e., without use of cmap). |
3601 | * We construct the schedule row s and write it as a linear |
3602 | * combination of (linear combinations of) previously computed schedule rows. |
3603 | * s = Q c or c = U s. |
3604 | * If the final entries of c are all zero, then the solution is trivial. |
3605 | */ |
3606 | static int is_trivial(struct isl_sched_node *node, __isl_keep isl_vec *sol) |
3607 | { |
3608 | int i; |
3609 | int pos; |
3610 | int trivial; |
3611 | isl_ctx *ctx; |
3612 | isl_vec *node_sol; |
3613 | |
3614 | if (!sol) |
3615 | return -1; |
3616 | if (node->nvar == node->rank) |
3617 | return 0; |
3618 | |
3619 | ctx = isl_vec_get_ctx(sol); |
3620 | node_sol = isl_vec_alloc(ctx, node->nvar); |
3621 | if (!node_sol) |
3622 | return -1; |
3623 | |
3624 | pos = 1 + node->start + 1 + 2 * node->nparam; |
3625 | |
3626 | for (i = 0; i < node->nvar; ++i) |
3627 | isl_int_sub(node_sol->el[i],isl_sioimath_sub((node_sol->el[i]), *(sol->el[pos + 2 * i + 1]), *(sol->el[pos + 2 * i])) |
3628 | sol->el[pos + 2 * i + 1], sol->el[pos + 2 * i])isl_sioimath_sub((node_sol->el[i]), *(sol->el[pos + 2 * i + 1]), *(sol->el[pos + 2 * i])); |
3629 | |
3630 | node_sol = isl_mat_vec_product(isl_mat_copy(node->cinv), node_sol); |
3631 | |
3632 | if (!node_sol) |
3633 | return -1; |
3634 | |
3635 | trivial = isl_seq_first_non_zero(node_sol->el + node->rank, |
3636 | node->nvar - node->rank) == -1; |
3637 | |
3638 | isl_vec_free(node_sol); |
3639 | |
3640 | return trivial; |
3641 | } |
3642 | |
3643 | /* Is the schedule row "sol" trivial on any node where it should |
3644 | * not be trivial? |
3645 | * "sol" has been computed in terms of the original iterators |
3646 | * (i.e., without use of cmap). |
3647 | * Return 1 if any solution is trivial, 0 if they are not and -1 on error. |
3648 | */ |
3649 | static int is_any_trivial(struct isl_sched_graph *graph, |
3650 | __isl_keep isl_vec *sol) |
3651 | { |
3652 | int i; |
3653 | |
3654 | for (i = 0; i < graph->n; ++i) { |
3655 | struct isl_sched_node *node = &graph->node[i]; |
3656 | int trivial; |
3657 | |
3658 | if (!needs_row(graph, node)) |
3659 | continue; |
3660 | trivial = is_trivial(node, sol); |
3661 | if (trivial < 0 || trivial) |
3662 | return trivial; |
3663 | } |
3664 | |
3665 | return 0; |
3666 | } |
3667 | |
3668 | /* Construct a schedule row for each node such that as many dependences |
3669 | * as possible are carried and then continue with the next band. |
3670 | * |
3671 | * If the computed schedule row turns out to be trivial on one or |
3672 | * more nodes where it should not be trivial, then we throw it away |
3673 | * and try again on each component separately. |
3674 | * |
3675 | * If there is only one component, then we accept the schedule row anyway, |
3676 | * but we do not consider it as a complete row and therefore do not |
3677 | * increment graph->n_row. Note that the ranks of the nodes that |
3678 | * do get a non-trivial schedule part will get updated regardless and |
3679 | * graph->maxvar is computed based on these ranks. The test for |
3680 | * whether more schedule rows are required in compute_schedule_wcc |
3681 | * is therefore not affected. |
3682 | * |
3683 | * Insert a band corresponding to the schedule row at position "node" |
3684 | * of the schedule tree and continue with the construction of the schedule. |
3685 | * This insertion and the continued construction is performed by split_scaled |
3686 | * after optionally checking for non-trivial common divisors. |
3687 | */ |
3688 | static __isl_give isl_schedule_node *carry_dependences( |
3689 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph) |
3690 | { |
3691 | int i; |
3692 | int n_edge; |
3693 | int trivial; |
3694 | isl_ctx *ctx; |
3695 | isl_vec *sol; |
3696 | isl_basic_setisl_basic_map *lp; |
3697 | |
3698 | if (!node) |
3699 | return NULL((void*)0); |
3700 | |
3701 | n_edge = 0; |
3702 | for (i = 0; i < graph->n_edge; ++i) |
3703 | n_edge += graph->edge[i].map->n; |
3704 | |
3705 | ctx = isl_schedule_node_get_ctx(node); |
3706 | if (setup_carry_lp(ctx, graph) < 0) |
3707 | return isl_schedule_node_free(node); |
3708 | |
3709 | lp = isl_basic_set_copy(graph->lp); |
3710 | sol = isl_tab_basic_set_non_neg_lexmin(lp); |
3711 | if (!sol) |
3712 | return isl_schedule_node_free(node); |
3713 | |
3714 | if (sol->size == 0) { |
3715 | isl_vec_free(sol); |
3716 | isl_die(ctx, isl_error_internal,do { isl_handle_error(ctx, isl_error_internal, "error in schedule construction" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3718); return isl_schedule_node_free(node); } while (0) |
3717 | "error in schedule construction",do { isl_handle_error(ctx, isl_error_internal, "error in schedule construction" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3718); return isl_schedule_node_free(node); } while (0) |
3718 | return isl_schedule_node_free(node))do { isl_handle_error(ctx, isl_error_internal, "error in schedule construction" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3718); return isl_schedule_node_free(node); } while (0); |
3719 | } |
3720 | |
3721 | isl_int_divexact(sol->el[1], sol->el[1], sol->el[0])isl_sioimath_tdiv_q((sol->el[1]), *(sol->el[1]), *(sol-> el[0])); |
3722 | if (isl_int_cmp_si(sol->el[1], n_edge)isl_sioimath_cmp_si(*(sol->el[1]), n_edge) >= 0) { |
3723 | isl_vec_free(sol); |
3724 | isl_die(ctx, isl_error_unknown,do { isl_handle_error(ctx, isl_error_unknown, "unable to carry dependences" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3726); return isl_schedule_node_free(node); } while (0) |
3725 | "unable to carry dependences",do { isl_handle_error(ctx, isl_error_unknown, "unable to carry dependences" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3726); return isl_schedule_node_free(node); } while (0) |
3726 | return isl_schedule_node_free(node))do { isl_handle_error(ctx, isl_error_unknown, "unable to carry dependences" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3726); return isl_schedule_node_free(node); } while (0); |
3727 | } |
3728 | |
3729 | trivial = is_any_trivial(graph, sol); |
3730 | if (trivial < 0) { |
3731 | sol = isl_vec_free(sol); |
3732 | } else if (trivial && graph->scc > 1) { |
3733 | isl_vec_free(sol); |
3734 | return compute_component_schedule(node, graph, 1); |
3735 | } |
3736 | |
3737 | if (update_schedule(graph, sol, 0, 0) < 0) |
3738 | return isl_schedule_node_free(node); |
3739 | if (trivial) |
3740 | graph->n_row--; |
3741 | |
3742 | return split_scaled(node, graph); |
3743 | } |
3744 | |
3745 | /* Topologically sort statements mapped to the same schedule iteration |
3746 | * and add insert a sequence node in front of "node" |
3747 | * corresponding to this order. |
3748 | * |
3749 | * If it turns out to be impossible to sort the statements apart, |
3750 | * because different dependences impose different orderings |
3751 | * on the statements, then we extend the schedule such that |
3752 | * it carries at least one more dependence. |
3753 | */ |
3754 | static __isl_give isl_schedule_node *sort_statements( |
3755 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph) |
3756 | { |
3757 | isl_ctx *ctx; |
3758 | isl_union_set_list *filters; |
3759 | |
3760 | if (!node) |
3761 | return NULL((void*)0); |
3762 | |
3763 | ctx = isl_schedule_node_get_ctx(node); |
3764 | if (graph->n < 1) |
3765 | isl_die(ctx, isl_error_internal,do { isl_handle_error(ctx, isl_error_internal, "graph should have at least one node" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3767); return isl_schedule_node_free(node); } while (0) |
3766 | "graph should have at least one node",do { isl_handle_error(ctx, isl_error_internal, "graph should have at least one node" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3767); return isl_schedule_node_free(node); } while (0) |
3767 | return isl_schedule_node_free(node))do { isl_handle_error(ctx, isl_error_internal, "graph should have at least one node" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/polly/lib/External/isl/isl_scheduler.c" , 3767); return isl_schedule_node_free(node); } while (0); |
3768 | |
3769 | if (graph->n == 1) |
3770 | return node; |
3771 | |
3772 | if (update_edges(ctx, graph) < 0) |
3773 | return isl_schedule_node_free(node); |
3774 | |
3775 | if (graph->n_edge == 0) |
3776 | return node; |
3777 | |
3778 | if (detect_sccs(ctx, graph) < 0) |
3779 | return isl_schedule_node_free(node); |
3780 | |
3781 | next_band(graph); |
3782 | if (graph->scc < graph->n) |
3783 | return carry_dependences(node, graph); |
3784 | |
3785 | filters = extract_sccs(ctx, graph); |
3786 | node = isl_schedule_node_insert_sequence(node, filters); |
3787 | |
3788 | return node; |
3789 | } |
3790 | |
3791 | /* Are there any (non-empty) (conditional) validity edges in the graph? |
3792 | */ |
3793 | static int has_validity_edges(struct isl_sched_graph *graph) |
3794 | { |
3795 | int i; |
3796 | |
3797 | for (i = 0; i < graph->n_edge; ++i) { |
3798 | int empty; |
3799 | |
3800 | empty = isl_map_plain_is_empty(graph->edge[i].map); |
3801 | if (empty < 0) |
3802 | return -1; |
3803 | if (empty) |
3804 | continue; |
3805 | if (graph->edge[i].validity || |
3806 | graph->edge[i].conditional_validity) |
3807 | return 1; |
3808 | } |
3809 | |
3810 | return 0; |
3811 | } |
3812 | |
3813 | /* Should we apply a Feautrier step? |
3814 | * That is, did the user request the Feautrier algorithm and are |
3815 | * there any validity dependences (left)? |
3816 | */ |
3817 | static int need_feautrier_step(isl_ctx *ctx, struct isl_sched_graph *graph) |
3818 | { |
3819 | if (ctx->opt->schedule_algorithm != ISL_SCHEDULE_ALGORITHM_FEAUTRIER1) |
3820 | return 0; |
3821 | |
3822 | return has_validity_edges(graph); |
3823 | } |
3824 | |
3825 | /* Compute a schedule for a connected dependence graph using Feautrier's |
3826 | * multi-dimensional scheduling algorithm and return the updated schedule node. |
3827 | * |
3828 | * The original algorithm is described in [1]. |
3829 | * The main idea is to minimize the number of scheduling dimensions, by |
3830 | * trying to satisfy as many dependences as possible per scheduling dimension. |
3831 | * |
3832 | * [1] P. Feautrier, Some Efficient Solutions to the Affine Scheduling |
3833 | * Problem, Part II: Multi-Dimensional Time. |
3834 | * In Intl. Journal of Parallel Programming, 1992. |
3835 | */ |
3836 | static __isl_give isl_schedule_node *compute_schedule_wcc_feautrier( |
3837 | isl_schedule_node *node, struct isl_sched_graph *graph) |
3838 | { |
3839 | return carry_dependences(node, graph); |
3840 | } |
3841 | |
3842 | /* Turn off the "local" bit on all (condition) edges. |
3843 | */ |
3844 | static void clear_local_edges(struct isl_sched_graph *graph) |
3845 | { |
3846 | int i; |
3847 | |
3848 | for (i = 0; i < graph->n_edge; ++i) |
3849 | if (graph->edge[i].condition) |
3850 | graph->edge[i].local = 0; |
3851 | } |
3852 | |
3853 | /* Does "graph" have both condition and conditional validity edges? |
3854 | */ |
3855 | static int need_condition_check(struct isl_sched_graph *graph) |
3856 | { |
3857 | int i; |
3858 | int any_condition = 0; |
3859 | int any_conditional_validity = 0; |
3860 | |
3861 | for (i = 0; i < graph->n_edge; ++i) { |
3862 | if (graph->edge[i].condition) |
3863 | any_condition = 1; |
3864 | if (graph->edge[i].conditional_validity) |
3865 | any_conditional_validity = 1; |
3866 | } |
3867 | |
3868 | return any_condition && any_conditional_validity; |
3869 | } |
3870 | |
3871 | /* Does "graph" contain any coincidence edge? |
3872 | */ |
3873 | static int has_any_coincidence(struct isl_sched_graph *graph) |
3874 | { |
3875 | int i; |
3876 | |
3877 | for (i = 0; i < graph->n_edge; ++i) |
3878 | if (graph->edge[i].coincidence) |
3879 | return 1; |
3880 | |
3881 | return 0; |
3882 | } |
3883 | |
3884 | /* Extract the final schedule row as a map with the iteration domain |
3885 | * of "node" as domain. |
3886 | */ |
3887 | static __isl_give isl_map *final_row(struct isl_sched_node *node) |
3888 | { |
3889 | isl_local_space *ls; |
3890 | isl_aff *aff; |
3891 | int row; |
3892 | |
3893 | row = isl_mat_rows(node->sched) - 1; |
3894 | ls = isl_local_space_from_space(isl_space_copy(node->space)); |
3895 | aff = extract_schedule_row(ls, node, row); |
3896 | return isl_map_from_aff(aff); |
3897 | } |
3898 | |
3899 | /* Is the conditional validity dependence in the edge with index "edge_index" |
3900 | * violated by the latest (i.e., final) row of the schedule? |
3901 | * That is, is i scheduled after j |
3902 | * for any conditional validity dependence i -> j? |
3903 | */ |
3904 | static int is_violated(struct isl_sched_graph *graph, int edge_index) |
3905 | { |
3906 | isl_map *src_sched, *dst_sched, *map; |
3907 | struct isl_sched_edge *edge = &graph->edge[edge_index]; |
3908 | int empty; |
3909 | |
3910 | src_sched = final_row(edge->src); |
3911 | dst_sched = final_row(edge->dst); |
3912 | map = isl_map_copy(edge->map); |
3913 | map = isl_map_apply_domain(map, src_sched); |
3914 | map = isl_map_apply_range(map, dst_sched); |
3915 | map = isl_map_order_gt(map, isl_dim_in, 0, isl_dim_out, 0); |
3916 | empty = isl_map_is_empty(map); |
3917 | isl_map_free(map); |
3918 | |
3919 | if (empty < 0) |
3920 | return -1; |
3921 | |
3922 | return !empty; |
3923 | } |
3924 | |
3925 | /* Does "graph" have any satisfied condition edges that |
3926 | * are adjacent to the conditional validity constraint with |
3927 | * domain "conditional_source" and range "conditional_sink"? |
3928 | * |
3929 | * A satisfied condition is one that is not local. |
3930 | * If a condition was forced to be local already (i.e., marked as local) |
3931 | * then there is no need to check if it is in fact local. |
3932 | * |
3933 | * Additionally, mark all adjacent condition edges found as local. |
3934 | */ |
3935 | static int has_adjacent_true_conditions(struct isl_sched_graph *graph, |
3936 | __isl_keep isl_union_set *conditional_source, |
3937 | __isl_keep isl_union_set *conditional_sink) |
3938 | { |
3939 | int i; |
3940 | int any = 0; |
3941 | |
3942 | for (i = 0; i < graph->n_edge; ++i) { |
3943 | int adjacent, local; |
3944 | isl_union_map *condition; |
3945 | |
3946 | if (!graph->edge[i].condition) |
3947 | continue; |
3948 | if (graph->edge[i].local) |
3949 | continue; |
3950 | |
3951 | condition = graph->edge[i].tagged_condition; |
3952 | adjacent = domain_intersects(condition, conditional_sink); |
3953 | if (adjacent >= 0 && !adjacent) |
3954 | adjacent = range_intersects(condition, |
3955 | conditional_source); |
3956 | if (adjacent < 0) |
3957 | return -1; |
3958 | if (!adjacent) |
3959 | continue; |
3960 | |
3961 | graph->edge[i].local = 1; |
3962 | |
3963 | local = is_condition_false(&graph->edge[i]); |
3964 | if (local < 0) |
3965 | return -1; |
3966 | if (!local) |
3967 | any = 1; |
3968 | } |
3969 | |
3970 | return any; |
3971 | } |
3972 | |
3973 | /* Are there any violated conditional validity dependences with |
3974 | * adjacent condition dependences that are not local with respect |
3975 | * to the current schedule? |
3976 | * That is, is the conditional validity constraint violated? |
3977 | * |
3978 | * Additionally, mark all those adjacent condition dependences as local. |
3979 | * We also mark those adjacent condition dependences that were not marked |
3980 | * as local before, but just happened to be local already. This ensures |
3981 | * that they remain local if the schedule is recomputed. |
3982 | * |
3983 | * We first collect domain and range of all violated conditional validity |
3984 | * dependences and then check if there are any adjacent non-local |
3985 | * condition dependences. |
3986 | */ |
3987 | static int has_violated_conditional_constraint(isl_ctx *ctx, |
3988 | struct isl_sched_graph *graph) |
3989 | { |
3990 | int i; |
3991 | int any = 0; |
3992 | isl_union_set *source, *sink; |
3993 | |
3994 | source = isl_union_set_empty(isl_space_params_alloc(ctx, 0)); |
3995 | sink = isl_union_set_empty(isl_space_params_alloc(ctx, 0)); |
3996 | for (i = 0; i < graph->n_edge; ++i) { |
3997 | isl_union_set *uset; |
3998 | isl_union_map *umap; |
3999 | int violated; |
4000 | |
4001 | if (!graph->edge[i].conditional_validity) |
4002 | continue; |
4003 | |
4004 | violated = is_violated(graph, i); |
4005 | if (violated < 0) |
4006 | goto error; |
4007 | if (!violated) |
4008 | continue; |
4009 | |
4010 | any = 1; |
4011 | |
4012 | umap = isl_union_map_copy(graph->edge[i].tagged_validity); |
4013 | uset = isl_union_map_domain(umap); |
4014 | source = isl_union_set_union(source, uset); |
4015 | source = isl_union_set_coalesce(source); |
4016 | |
4017 | umap = isl_union_map_copy(graph->edge[i].tagged_validity); |
4018 | uset = isl_union_map_range(umap); |
4019 | sink = isl_union_set_union(sink, uset); |
4020 | sink = isl_union_set_coalesce(sink); |
4021 | } |
4022 | |
4023 | if (any) |
4024 | any = has_adjacent_true_conditions(graph, source, sink); |
4025 | |
4026 | isl_union_set_free(source); |
4027 | isl_union_set_free(sink); |
4028 | return any; |
4029 | error: |
4030 | isl_union_set_free(source); |
4031 | isl_union_set_free(sink); |
4032 | return -1; |
4033 | } |
4034 | |
4035 | /* Compute a schedule for a connected dependence graph and return |
4036 | * the updated schedule node. |
4037 | * |
4038 | * We try to find a sequence of as many schedule rows as possible that result |
4039 | * in non-negative dependence distances (independent of the previous rows |
4040 | * in the sequence, i.e., such that the sequence is tilable), with as |
4041 | * many of the initial rows as possible satisfying the coincidence constraints. |
4042 | * If we can't find any more rows we either |
4043 | * - split between SCCs and start over (assuming we found an interesting |
4044 | * pair of SCCs between which to split) |
4045 | * - continue with the next band (assuming the current band has at least |
4046 | * one row) |
4047 | * - try to carry as many dependences as possible and continue with the next |
4048 | * band |
4049 | * In each case, we first insert a band node in the schedule tree |
4050 | * if any rows have been computed. |
4051 | * |
4052 | * If Feautrier's algorithm is selected, we first recursively try to satisfy |
4053 | * as many validity dependences as possible. When all validity dependences |
4054 | * are satisfied we extend the schedule to a full-dimensional schedule. |
4055 | * |
4056 | * If we manage to complete the schedule, we insert a band node |
4057 | * (if any schedule rows were computed) and we finish off by topologically |
4058 | * sorting the statements based on the remaining dependences. |
4059 | * |
4060 | * If ctx->opt->schedule_outer_coincidence is set, then we force the |
4061 | * outermost dimension to satisfy the coincidence constraints. If this |
4062 | * turns out to be impossible, we fall back on the general scheme above |
4063 | * and try to carry as many dependences as possible. |
4064 | * |
4065 | * If "graph" contains both condition and conditional validity dependences, |
4066 | * then we need to check that that the conditional schedule constraint |
4067 | * is satisfied, i.e., there are no violated conditional validity dependences |
4068 | * that are adjacent to any non-local condition dependences. |
4069 | * If there are, then we mark all those adjacent condition dependences |
4070 | * as local and recompute the current band. Those dependences that |
4071 | * are marked local will then be forced to be local. |
4072 | * The initial computation is performed with no dependences marked as local. |
4073 | * If we are lucky, then there will be no violated conditional validity |
4074 | * dependences adjacent to any non-local condition dependences. |
4075 | * Otherwise, we mark some additional condition dependences as local and |
4076 | * recompute. We continue this process until there are no violations left or |
4077 | * until we are no longer able to compute a schedule. |
4078 | * Since there are only a finite number of dependences, |
4079 | * there will only be a finite number of iterations. |
4080 | */ |
4081 | static __isl_give isl_schedule_node *compute_schedule_wcc( |
4082 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph) |
4083 | { |
4084 | int has_coincidence; |
4085 | int use_coincidence; |
4086 | int force_coincidence = 0; |
4087 | int check_conditional; |
4088 | int insert; |
4089 | isl_ctx *ctx; |
4090 | |
4091 | if (!node) |
4092 | return NULL((void*)0); |
4093 | |
4094 | ctx = isl_schedule_node_get_ctx(node); |
4095 | if (detect_sccs(ctx, graph) < 0) |
4096 | return isl_schedule_node_free(node); |
4097 | if (sort_sccs(graph) < 0) |
4098 | return isl_schedule_node_free(node); |
4099 | |
4100 | if (compute_maxvar(graph) < 0) |
4101 | return isl_schedule_node_free(node); |
4102 | |
4103 | if (need_feautrier_step(ctx, graph)) |
4104 | return compute_schedule_wcc_feautrier(node, graph); |
4105 | |
4106 | clear_local_edges(graph); |
4107 | check_conditional = need_condition_check(graph); |
4108 | has_coincidence = has_any_coincidence(graph); |
4109 | |
4110 | if (ctx->opt->schedule_outer_coincidence) |
4111 | force_coincidence = 1; |
4112 | |
4113 | use_coincidence = has_coincidence; |
4114 | while (graph->n_row < graph->maxvar) { |
4115 | isl_vec *sol; |
4116 | int violated; |
4117 | int coincident; |
4118 | |
4119 | graph->src_scc = -1; |
4120 | graph->dst_scc = -1; |
4121 | |
4122 | if (setup_lp(ctx, graph, use_coincidence) < 0) |
4123 | return isl_schedule_node_free(node); |
4124 | sol = solve_lp(graph); |
4125 | if (!sol) |
4126 | return isl_schedule_node_free(node); |
4127 | if (sol->size == 0) { |
4128 | int empty = graph->n_total_row == graph->band_start; |
4129 | |
4130 | isl_vec_free(sol); |
4131 | if (use_coincidence && (!force_coincidence || !empty)) { |
4132 | use_coincidence = 0; |
4133 | continue; |
4134 | } |
4135 | if (!ctx->opt->schedule_maximize_band_depth && !empty) |
4136 | return compute_next_band(node, graph, 1); |
4137 | if (graph->src_scc >= 0) |
4138 | return compute_split_schedule(node, graph); |
4139 | if (!empty) |
4140 | return compute_next_band(node, graph, 1); |
4141 | return carry_dependences(node, graph); |
4142 | } |
4143 | coincident = !has_coincidence || use_coincidence; |
4144 | if (update_schedule(graph, sol, 1, coincident) < 0) |
4145 | return isl_schedule_node_free(node); |
4146 | |
4147 | if (!check_conditional) |
4148 | continue; |
4149 | violated = has_violated_conditional_constraint(ctx, graph); |
4150 | if (violated < 0) |
4151 | return isl_schedule_node_free(node); |
4152 | if (!violated) |
4153 | continue; |
4154 | if (reset_band(graph) < 0) |
4155 | return isl_schedule_node_free(node); |
4156 | use_coincidence = has_coincidence; |
4157 | } |
4158 | |
4159 | insert = graph->n_total_row > graph->band_start; |
4160 | if (insert) { |
4161 | node = insert_current_band(node, graph, 1); |
4162 | node = isl_schedule_node_child(node, 0); |
4163 | } |
4164 | node = sort_statements(node, graph); |
4165 | if (insert) |
4166 | node = isl_schedule_node_parent(node); |
4167 | |
4168 | return node; |
4169 | } |
4170 | |
4171 | /* Compute a schedule for each group of nodes identified by node->scc |
4172 | * separately and then combine them in a sequence node (or as set node |
4173 | * if graph->weak is set) inserted at position "node" of the schedule tree. |
4174 | * Return the updated schedule node. |
4175 | * |
4176 | * If "wcc" is set then each of the groups belongs to a single |
4177 | * weakly connected component in the dependence graph so that |
4178 | * there is no need for compute_sub_schedule to look for weakly |
4179 | * connected components. |
4180 | */ |
4181 | static __isl_give isl_schedule_node *compute_component_schedule( |
4182 | __isl_take isl_schedule_node *node, struct isl_sched_graph *graph, |
4183 | int wcc) |
4184 | { |
4185 | int component, i; |
4186 | int n, n_edge; |
4187 | isl_ctx *ctx; |
4188 | isl_union_set_list *filters; |
4189 | |
4190 | if (!node) |
4191 | return NULL((void*)0); |
4192 | ctx = isl_schedule_node_get_ctx(node); |
4193 | |
4194 | filters = extract_sccs(ctx, graph); |
4195 | if (graph->weak) |
4196 | node = isl_schedule_node_insert_set(node, filters); |
4197 | else |
4198 | node = isl_schedule_node_insert_sequence(node, filters); |
4199 | |
4200 | for (component = 0; component < graph->scc; ++component) { |
4201 | n = 0; |
4202 | for (i = 0; i < graph->n; ++i) |
4203 | if (graph->node[i].scc == component) |
4204 | n++; |
4205 | n_edge = 0; |
4206 | for (i = 0; i < graph->n_edge; ++i) |
4207 | if (graph->edge[i].src->scc == component && |
4208 | graph->edge[i].dst->scc == component) |
4209 | n_edge++; |
4210 | |
4211 | node = isl_schedule_node_child(node, component); |
4212 | node = isl_schedule_node_child(node, 0); |
4213 | node = compute_sub_schedule(node, ctx, graph, n, n_edge, |
4214 | &node_scc_exactly, |
4215 | &edge_scc_exactly, component, wcc); |
4216 | node = isl_schedule_node_parent(node); |
4217 | node = isl_schedule_node_parent(node); |
4218 | } |
4219 | |
4220 | return node; |
4221 | } |
4222 | |
4223 | /* Compute a schedule for the given dependence graph and insert it at "node". |
4224 | * Return the updated schedule node. |
4225 | * |
4226 | * We first check if the graph is connected (through validity and conditional |
4227 | * validity dependences) and, if not, compute a schedule |
4228 | * for each component separately. |
4229 | * If the schedule_serialize_sccs option is set, then we check for strongly |
4230 | * connected components instead and compute a separate schedule for |
4231 | * each such strongly connected component. |
4232 | */ |
4233 | static __isl_give isl_schedule_node *compute_schedule(isl_schedule_node *node, |
4234 | struct isl_sched_graph *graph) |
4235 | { |
4236 | isl_ctx *ctx; |
4237 | |
4238 | if (!node) |
4239 | return NULL((void*)0); |
4240 | |
4241 | ctx = isl_schedule_node_get_ctx(node); |
4242 | if (isl_options_get_schedule_serialize_sccs(ctx)) { |
4243 | if (detect_sccs(ctx, graph) < 0) |
4244 | return isl_schedule_node_free(node); |
4245 | } else { |
4246 | if (detect_wccs(ctx, graph) < 0) |
4247 | return isl_schedule_node_free(node); |
4248 | } |
4249 | |
4250 | if (graph->scc > 1) |
4251 | return compute_component_schedule(node, graph, 1); |
4252 | |
4253 | return compute_schedule_wcc(node, graph); |
4254 | } |
4255 | |
4256 | /* Compute a schedule on sc->domain that respects the given schedule |
4257 | * constraints. |
4258 | * |
4259 | * In particular, the schedule respects all the validity dependences. |
4260 | * If the default isl scheduling algorithm is used, it tries to minimize |
4261 | * the dependence distances over the proximity dependences. |
4262 | * If Feautrier's scheduling algorithm is used, the proximity dependence |
4263 | * distances are only minimized during the extension to a full-dimensional |
4264 | * schedule. |
4265 | * |
4266 | * If there are any condition and conditional validity dependences, |
4267 | * then the conditional validity dependences may be violated inside |
4268 | * a tilable band, provided they have no adjacent non-local |
4269 | * condition dependences. |
4270 | * |
4271 | * The context is included in the domain before the nodes of |
4272 | * the graphs are extracted in order to be able to exploit |
4273 | * any possible additional equalities. |
4274 | * However, the returned schedule contains the original domain |
4275 | * (before this intersection). |
4276 | */ |
4277 | __isl_give isl_schedule *isl_schedule_constraints_compute_schedule( |
4278 | __isl_take isl_schedule_constraints *sc) |
4279 | { |
4280 | isl_ctx *ctx = isl_schedule_constraints_get_ctx(sc); |
4281 | struct isl_sched_graph graph = { 0 }; |
4282 | isl_schedule *sched; |
4283 | isl_schedule_node *node; |
4284 | isl_union_set *domain; |
4285 | struct isl_extract_edge_data data; |
4286 | enum isl_edge_type i; |
4287 | int r; |
4288 | |
4289 | sc = isl_schedule_constraints_align_params(sc); |
4290 | if (!sc) |
4291 | return NULL((void*)0); |
4292 | |
4293 | graph.n = isl_union_set_n_set(sc->domain); |
4294 | if (graph.n == 0) { |
4295 | isl_union_set *domain = isl_union_set_copy(sc->domain); |
4296 | sched = isl_schedule_from_domain(domain); |
4297 | goto done; |
4298 | } |
4299 | if (graph_alloc(ctx, &graph, graph.n, |
4300 | isl_schedule_constraints_n_map(sc)) < 0) |
4301 | goto error; |
4302 | if (compute_max_row(&graph, sc) < 0) |
4303 | goto error; |
4304 | graph.root = 1; |
4305 | graph.n = 0; |
4306 | domain = isl_union_set_copy(sc->domain); |
4307 | domain = isl_union_set_intersect_params(domain, |
4308 | isl_set_copy(sc->context)); |
4309 | r = isl_union_set_foreach_set(domain, &extract_node, &graph); |
4310 | isl_union_set_free(domain); |
4311 | if (r < 0) |
4312 | goto error; |
4313 | if (graph_init_table(ctx, &graph) < 0) |
4314 | goto error; |
4315 | for (i = isl_edge_first; i <= isl_edge_last; ++i) |
4316 | graph.max_edge[i] = isl_union_map_n_map(sc->constraint[i]); |
4317 | if (graph_init_edge_tables(ctx, &graph) < 0) |
4318 | goto error; |
4319 | graph.n_edge = 0; |
4320 | data.graph = &graph; |
4321 | for (i = isl_edge_first; i <= isl_edge_last; ++i) { |
4322 | data.type = i; |
4323 | if (isl_union_map_foreach_map(sc->constraint[i], |
4324 | &extract_edge, &data) < 0) |
4325 | goto error; |
4326 | } |
4327 | |
4328 | node = isl_schedule_node_from_domain(isl_union_set_copy(sc->domain)); |
4329 | node = isl_schedule_node_child(node, 0); |
4330 | if (graph.n > 0) |
4331 | node = compute_schedule(node, &graph); |
4332 | sched = isl_schedule_node_get_schedule(node); |
4333 | isl_schedule_node_free(node); |
4334 | |
4335 | done: |
4336 | graph_free(ctx, &graph); |
4337 | isl_schedule_constraints_free(sc); |
4338 | |
4339 | return sched; |
4340 | error: |
4341 | graph_free(ctx, &graph); |
4342 | isl_schedule_constraints_free(sc); |
4343 | return NULL((void*)0); |
4344 | } |
4345 | |
4346 | /* Compute a schedule for the given union of domains that respects |
4347 | * all the validity dependences and minimizes |
4348 | * the dependence distances over the proximity dependences. |
4349 | * |
4350 | * This function is kept for backward compatibility. |
4351 | */ |
4352 | __isl_give isl_schedule *isl_union_set_compute_schedule( |
4353 | __isl_take isl_union_set *domain, |
4354 | __isl_take isl_union_map *validity, |
4355 | __isl_take isl_union_map *proximity) |
4356 | { |
4357 | isl_schedule_constraints *sc; |
4358 | |
4359 | sc = isl_schedule_constraints_on_domain(domain); |
4360 | sc = isl_schedule_constraints_set_validity(sc, validity); |
4361 | sc = isl_schedule_constraints_set_proximity(sc, proximity); |
4362 | |
4363 | return isl_schedule_constraints_compute_schedule(sc); |
4364 | } |