2 * Copyright 2011 INRIA Saclay
4 * Use of this software is governed by the GNU LGPLv2.1 license
6 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
7 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
11 #include <isl_ctx_private.h>
12 #include <isl_map_private.h>
13 #include <isl_space_private.h>
15 #include <isl/constraint.h>
16 #include <isl/schedule.h>
17 #include <isl_mat_private.h>
21 #include <isl_dim_map.h>
22 #include <isl_hmap_map_basic_set.h>
23 #include <isl_qsort.h>
24 #include <isl_schedule_private.h>
25 #include <isl_band_private.h>
26 #include <isl_list_private.h>
29 * The scheduling algorithm implemented in this file was inspired by
30 * Bondhugula et al., "Automatic Transformations for Communication-Minimized
31 * Parallelization and Locality Optimization in the Polyhedral Model".
35 /* Internal information about a node that is used during the construction
37 * dim represents the space in which the domain lives
38 * sched is a matrix representation of the schedule being constructed
40 * sched_map is an isl_map representation of the same (partial) schedule
41 * sched_map may be NULL
42 * rank is the number of linearly independent rows in the linear part
44 * the columns of cmap represent a change of basis for the schedule
45 * coefficients; the first rank columns span the linear part of
47 * start is the first variable in the LP problem in the sequences that
48 * represents the schedule coefficients of this node
49 * nvar is the dimension of the domain
50 * nparam is the number of parameters or 0 if we are not constructing
51 * a parametric schedule
53 * scc is the index of SCC (or WCC) this node belongs to
55 * band contains the band index for each of the rows of the schedule.
56 * band_id is used to differentiate between separate bands at the same
57 * level within the same parent band, i.e., bands that are separated
58 * by the parent band or bands that are independent of each other.
59 * zero contains a boolean for each of the rows of the schedule,
60 * indicating whether the corresponding scheduling dimension results
61 * in zero dependence distances within its band and with respect
62 * to the proximity edges.
64 * index, min_index and on_stack are used during the SCC detection
65 * index represents the order in which nodes are visited.
66 * min_index is the index of the root of a (sub)component.
67 * on_stack indicates whether the node is currently on the stack.
69 struct isl_sched_node {
91 static int node_has_dim(const void *entry, const void *val)
93 struct isl_sched_node *node = (struct isl_sched_node *)entry;
94 isl_space *dim = (isl_space *)val;
96 return isl_space_is_equal(node->dim, dim);
99 /* An edge in the dependence graph. An edge may be used to
100 * ensure validity of the generated schedule, to minimize the dependence
103 * map is the dependence relation
104 * src is the source node
105 * dst is the sink node
106 * validity is set if the edge is used to ensure correctness
107 * proximity is set if the edge is used to minimize dependence distances
109 * For validity edges, start and end mark the sequence of inequality
110 * constraints in the LP problem that encode the validity constraint
111 * corresponding to this edge.
113 struct isl_sched_edge {
116 struct isl_sched_node *src;
117 struct isl_sched_node *dst;
126 /* Internal information about the dependence graph used during
127 * the construction of the schedule.
129 * intra_hmap is a cache, mapping dependence relations to their dual,
130 * for dependences from a node to itself
131 * inter_hmap is a cache, mapping dependence relations to their dual,
132 * for dependences between distinct nodes
134 * n is the number of nodes
135 * node is the list of nodes
136 * maxvar is the maximal number of variables over all nodes
137 * n_row is the current (maximal) number of linearly independent
138 * rows in the node schedules
139 * n_total_row is the current number of rows in the node schedules
140 * n_band is the current number of completed bands
141 * band_start is the starting row in the node schedules of the current band
142 * root is set if this graph is the original dependence graph,
143 * without any splitting
145 * sorted contains a list of node indices sorted according to the
146 * SCC to which a node belongs
148 * n_edge is the number of edges
149 * edge is the list of edges
150 * edge_table contains pointers into the edge array, hashed on the source
151 * and sink spaces; the table only contains edges that represent
152 * validity constraints (and that may or may not also represent proximity
155 * node_table contains pointers into the node array, hashed on the space
157 * region contains a list of variable sequences that should be non-trivial
159 * lp contains the (I)LP problem used to obtain new schedule rows
161 * src_scc and dst_scc are the source and sink SCCs of an edge with
162 * conflicting constraints
164 * scc, sp, index and stack are used during the detection of SCCs
165 * scc is the number of the next SCC
166 * stack contains the nodes on the path from the root to the current node
167 * sp is the stack pointer
168 * index is the index of the last node visited
170 struct isl_sched_graph {
171 isl_hmap_map_basic_set *intra_hmap;
172 isl_hmap_map_basic_set *inter_hmap;
174 struct isl_sched_node *node;
187 struct isl_sched_edge *edge;
189 struct isl_hash_table *edge_table;
191 struct isl_hash_table *node_table;
192 struct isl_region *region;
206 /* Initialize node_table based on the list of nodes.
208 static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph)
212 graph->node_table = isl_hash_table_alloc(ctx, graph->n);
213 if (!graph->node_table)
216 for (i = 0; i < graph->n; ++i) {
217 struct isl_hash_table_entry *entry;
220 hash = isl_space_get_hash(graph->node[i].dim);
221 entry = isl_hash_table_find(ctx, graph->node_table, hash,
223 graph->node[i].dim, 1);
226 entry->data = &graph->node[i];
232 /* Return a pointer to the node that lives within the given space,
233 * or NULL if there is no such node.
235 static struct isl_sched_node *graph_find_node(isl_ctx *ctx,
236 struct isl_sched_graph *graph, __isl_keep isl_space *dim)
238 struct isl_hash_table_entry *entry;
241 hash = isl_space_get_hash(dim);
242 entry = isl_hash_table_find(ctx, graph->node_table, hash,
243 &node_has_dim, dim, 0);
245 return entry ? entry->data : NULL;
248 static int edge_has_src_and_dst(const void *entry, const void *val)
250 const struct isl_sched_edge *edge = entry;
251 const struct isl_sched_edge *temp = val;
253 return edge->src == temp->src && edge->dst == temp->dst;
256 /* Initialize edge_table based on the list of edges.
257 * Only edges with validity set are added to the table.
259 static int graph_init_edge_table(isl_ctx *ctx, struct isl_sched_graph *graph)
263 graph->edge_table = isl_hash_table_alloc(ctx, graph->n_edge);
264 if (!graph->edge_table)
267 for (i = 0; i < graph->n_edge; ++i) {
268 struct isl_hash_table_entry *entry;
271 if (!graph->edge[i].validity)
274 hash = isl_hash_init();
275 hash = isl_hash_builtin(hash, graph->edge[i].src);
276 hash = isl_hash_builtin(hash, graph->edge[i].dst);
277 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
278 &edge_has_src_and_dst,
282 entry->data = &graph->edge[i];
288 /* Check whether the dependence graph has a (validity) edge
289 * between the given two nodes.
291 static int graph_has_edge(struct isl_sched_graph *graph,
292 struct isl_sched_node *src, struct isl_sched_node *dst)
294 isl_ctx *ctx = isl_space_get_ctx(src->dim);
295 struct isl_hash_table_entry *entry;
297 struct isl_sched_edge temp = { .src = src, .dst = dst };
298 struct isl_sched_edge *edge;
301 hash = isl_hash_init();
302 hash = isl_hash_builtin(hash, temp.src);
303 hash = isl_hash_builtin(hash, temp.dst);
304 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
305 &edge_has_src_and_dst, &temp, 0);
310 empty = isl_map_plain_is_empty(edge->map);
317 static int graph_alloc(isl_ctx *ctx, struct isl_sched_graph *graph,
318 int n_node, int n_edge)
323 graph->n_edge = n_edge;
324 graph->node = isl_calloc_array(ctx, struct isl_sched_node, graph->n);
325 graph->sorted = isl_calloc_array(ctx, int, graph->n);
326 graph->region = isl_alloc_array(ctx, struct isl_region, graph->n);
327 graph->stack = isl_alloc_array(ctx, int, graph->n);
328 graph->edge = isl_calloc_array(ctx,
329 struct isl_sched_edge, graph->n_edge);
331 graph->intra_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
332 graph->inter_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
334 if (!graph->node || !graph->region || !graph->stack || !graph->edge ||
338 for(i = 0; i < graph->n; ++i)
339 graph->sorted[i] = i;
344 static void graph_free(isl_ctx *ctx, struct isl_sched_graph *graph)
348 isl_hmap_map_basic_set_free(ctx, graph->intra_hmap);
349 isl_hmap_map_basic_set_free(ctx, graph->inter_hmap);
351 for (i = 0; i < graph->n; ++i) {
352 isl_space_free(graph->node[i].dim);
353 isl_mat_free(graph->node[i].sched);
354 isl_map_free(graph->node[i].sched_map);
355 isl_mat_free(graph->node[i].cmap);
357 free(graph->node[i].band);
358 free(graph->node[i].band_id);
359 free(graph->node[i].zero);
364 for (i = 0; i < graph->n_edge; ++i)
365 isl_map_free(graph->edge[i].map);
369 isl_hash_table_free(ctx, graph->edge_table);
370 isl_hash_table_free(ctx, graph->node_table);
371 isl_basic_set_free(graph->lp);
374 /* Add a new node to the graph representing the given set.
376 static int extract_node(__isl_take isl_set *set, void *user)
382 struct isl_sched_graph *graph = user;
383 int *band, *band_id, *zero;
385 ctx = isl_set_get_ctx(set);
386 dim = isl_set_get_space(set);
388 nvar = isl_space_dim(dim, isl_dim_set);
389 nparam = isl_space_dim(dim, isl_dim_param);
390 if (!ctx->opt->schedule_parametric)
392 sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar);
393 graph->node[graph->n].dim = dim;
394 graph->node[graph->n].nvar = nvar;
395 graph->node[graph->n].nparam = nparam;
396 graph->node[graph->n].sched = sched;
397 graph->node[graph->n].sched_map = NULL;
398 band = isl_alloc_array(ctx, int, graph->n_edge + nvar);
399 graph->node[graph->n].band = band;
400 band_id = isl_calloc_array(ctx, int, graph->n_edge + nvar);
401 graph->node[graph->n].band_id = band_id;
402 zero = isl_calloc_array(ctx, int, graph->n_edge + nvar);
403 graph->node[graph->n].zero = zero;
406 if (!sched || !band || !band_id || !zero)
412 /* Add a new edge to the graph based on the given map.
413 * Edges are first extracted from the validity dependences,
414 * from which the edge_table is constructed.
415 * Afterwards, the proximity dependences are added. If a proximity
416 * dependence relation happens to be identical to one of the
417 * validity dependence relations added before, then we don't create
418 * a new edge, but instead mark the original edge as also representing
419 * a proximity dependence.
421 static int extract_edge(__isl_take isl_map *map, void *user)
423 isl_ctx *ctx = isl_map_get_ctx(map);
424 struct isl_sched_graph *graph = user;
425 struct isl_sched_node *src, *dst;
428 dim = isl_space_domain(isl_map_get_space(map));
429 src = graph_find_node(ctx, graph, dim);
431 dim = isl_space_range(isl_map_get_space(map));
432 dst = graph_find_node(ctx, graph, dim);
440 graph->edge[graph->n_edge].src = src;
441 graph->edge[graph->n_edge].dst = dst;
442 graph->edge[graph->n_edge].map = map;
443 graph->edge[graph->n_edge].validity = !graph->edge_table;
444 graph->edge[graph->n_edge].proximity = !!graph->edge_table;
447 if (graph->edge_table) {
449 struct isl_hash_table_entry *entry;
450 struct isl_sched_edge *edge;
453 hash = isl_hash_init();
454 hash = isl_hash_builtin(hash, src);
455 hash = isl_hash_builtin(hash, dst);
456 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
457 &edge_has_src_and_dst,
458 &graph->edge[graph->n_edge - 1], 0);
462 is_equal = isl_map_plain_is_equal(map, edge->map);
476 /* Check whether there is a validity dependence from src to dst,
477 * forcing dst to follow src.
479 static int node_follows(struct isl_sched_graph *graph,
480 struct isl_sched_node *dst, struct isl_sched_node *src)
482 return graph_has_edge(graph, src, dst);
485 /* Perform Tarjan's algorithm for computing the strongly connected components
486 * in the dependence graph (only validity edges).
487 * If directed is not set, we consider the graph to be undirected and
488 * we effectively compute the (weakly) connected components.
490 static int detect_sccs_tarjan(struct isl_sched_graph *g, int i, int directed)
494 g->node[i].index = g->index;
495 g->node[i].min_index = g->index;
496 g->node[i].on_stack = 1;
498 g->stack[g->sp++] = i;
500 for (j = g->n - 1; j >= 0; --j) {
505 if (g->node[j].index >= 0 &&
506 (!g->node[j].on_stack ||
507 g->node[j].index > g->node[i].min_index))
510 f = node_follows(g, &g->node[i], &g->node[j]);
513 if (!f && !directed) {
514 f = node_follows(g, &g->node[j], &g->node[i]);
520 if (g->node[j].index < 0) {
521 detect_sccs_tarjan(g, j, directed);
522 if (g->node[j].min_index < g->node[i].min_index)
523 g->node[i].min_index = g->node[j].min_index;
524 } else if (g->node[j].index < g->node[i].min_index)
525 g->node[i].min_index = g->node[j].index;
528 if (g->node[i].index != g->node[i].min_index)
532 j = g->stack[--g->sp];
533 g->node[j].on_stack = 0;
534 g->node[j].scc = g->scc;
541 static int detect_ccs(struct isl_sched_graph *graph, int directed)
548 for (i = graph->n - 1; i >= 0; --i)
549 graph->node[i].index = -1;
551 for (i = graph->n - 1; i >= 0; --i) {
552 if (graph->node[i].index >= 0)
554 if (detect_sccs_tarjan(graph, i, directed) < 0)
561 /* Apply Tarjan's algorithm to detect the strongly connected components
562 * in the dependence graph.
564 static int detect_sccs(struct isl_sched_graph *graph)
566 return detect_ccs(graph, 1);
569 /* Apply Tarjan's algorithm to detect the (weakly) connected components
570 * in the dependence graph.
572 static int detect_wccs(struct isl_sched_graph *graph)
574 return detect_ccs(graph, 0);
577 static int cmp_scc(const void *a, const void *b, void *data)
579 struct isl_sched_graph *graph = data;
583 return graph->node[*i1].scc - graph->node[*i2].scc;
586 /* Sort the elements of graph->sorted according to the corresponding SCCs.
588 static void sort_sccs(struct isl_sched_graph *graph)
590 isl_quicksort(graph->sorted, graph->n, sizeof(int), &cmp_scc, graph);
593 /* Given a dependence relation R from a node to itself,
594 * construct the set of coefficients of valid constraints for elements
595 * in that dependence relation.
596 * In particular, the result contains tuples of coefficients
597 * c_0, c_n, c_x such that
599 * c_0 + c_n n + c_x y - c_x x >= 0 for each (x,y) in R
603 * c_0 + c_n n + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R }
605 * We choose here to compute the dual of delta R.
606 * Alternatively, we could have computed the dual of R, resulting
607 * in a set of tuples c_0, c_n, c_x, c_y, and then
608 * plugged in (c_0, c_n, c_x, -c_x).
610 static __isl_give isl_basic_set *intra_coefficients(
611 struct isl_sched_graph *graph, __isl_take isl_map *map)
613 isl_ctx *ctx = isl_map_get_ctx(map);
617 if (isl_hmap_map_basic_set_has(ctx, graph->intra_hmap, map))
618 return isl_hmap_map_basic_set_get(ctx, graph->intra_hmap, map);
620 delta = isl_set_remove_divs(isl_map_deltas(isl_map_copy(map)));
621 coef = isl_set_coefficients(delta);
622 isl_hmap_map_basic_set_set(ctx, graph->intra_hmap, map,
623 isl_basic_set_copy(coef));
628 /* Given a dependence relation R, * construct the set of coefficients
629 * of valid constraints for elements in that dependence relation.
630 * In particular, the result contains tuples of coefficients
631 * c_0, c_n, c_x, c_y such that
633 * c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R
636 static __isl_give isl_basic_set *inter_coefficients(
637 struct isl_sched_graph *graph, __isl_take isl_map *map)
639 isl_ctx *ctx = isl_map_get_ctx(map);
643 if (isl_hmap_map_basic_set_has(ctx, graph->inter_hmap, map))
644 return isl_hmap_map_basic_set_get(ctx, graph->inter_hmap, map);
646 set = isl_map_wrap(isl_map_remove_divs(isl_map_copy(map)));
647 coef = isl_set_coefficients(set);
648 isl_hmap_map_basic_set_set(ctx, graph->inter_hmap, map,
649 isl_basic_set_copy(coef));
654 /* Add constraints to graph->lp that force validity for the given
655 * dependence from a node i to itself.
656 * That is, add constraints that enforce
658 * (c_i_0 + c_i_n n + c_i_x y) - (c_i_0 + c_i_n n + c_i_x x)
659 * = c_i_x (y - x) >= 0
661 * for each (x,y) in R.
662 * We obtain general constraints on coefficients (c_0, c_n, c_x)
663 * of valid constraints for (y - x) and then plug in (0, 0, c_i_x^+ - c_i_x^-),
664 * where c_i_x = c_i_x^+ - c_i_x^-, with c_i_x^+ and c_i_x^- non-negative.
665 * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart.
667 * Actually, we do not construct constraints for the c_i_x themselves,
668 * but for the coefficients of c_i_x written as a linear combination
669 * of the columns in node->cmap.
671 static int add_intra_validity_constraints(struct isl_sched_graph *graph,
672 struct isl_sched_edge *edge)
675 isl_map *map = isl_map_copy(edge->map);
676 isl_ctx *ctx = isl_map_get_ctx(map);
678 isl_dim_map *dim_map;
680 struct isl_sched_node *node = edge->src;
682 coef = intra_coefficients(graph, map);
684 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
686 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
687 isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap));
689 total = isl_basic_set_total_dim(graph->lp);
690 dim_map = isl_dim_map_alloc(ctx, total);
691 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
692 isl_space_dim(dim, isl_dim_set), 1,
694 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
695 isl_space_dim(dim, isl_dim_set), 1,
697 graph->lp = isl_basic_set_extend_constraints(graph->lp,
698 coef->n_eq, coef->n_ineq);
699 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
706 /* Add constraints to graph->lp that force validity for the given
707 * dependence from node i to node j.
708 * That is, add constraints that enforce
710 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) >= 0
712 * for each (x,y) in R.
713 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
714 * of valid constraints for R and then plug in
715 * (c_j_0 - c_i_0, c_j_n^+ - c_j_n^- - (c_i_n^+ - c_i_n^-),
716 * c_j_x^+ - c_j_x^- - (c_i_x^+ - c_i_x^-)),
717 * where c_* = c_*^+ - c_*^-, with c_*^+ and c_*^- non-negative.
718 * In graph->lp, the c_*^- appear before their c_*^+ counterpart.
720 * Actually, we do not construct constraints for the c_*_x themselves,
721 * but for the coefficients of c_*_x written as a linear combination
722 * of the columns in node->cmap.
724 static int add_inter_validity_constraints(struct isl_sched_graph *graph,
725 struct isl_sched_edge *edge)
728 isl_map *map = isl_map_copy(edge->map);
729 isl_ctx *ctx = isl_map_get_ctx(map);
731 isl_dim_map *dim_map;
733 struct isl_sched_node *src = edge->src;
734 struct isl_sched_node *dst = edge->dst;
736 coef = inter_coefficients(graph, map);
738 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
740 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
741 isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap));
742 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
743 isl_space_dim(dim, isl_dim_set) + src->nvar,
744 isl_mat_copy(dst->cmap));
746 total = isl_basic_set_total_dim(graph->lp);
747 dim_map = isl_dim_map_alloc(ctx, total);
749 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
750 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
751 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
752 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
753 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
755 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
756 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
759 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
760 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
761 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
762 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
763 isl_space_dim(dim, isl_dim_set), 1,
765 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
766 isl_space_dim(dim, isl_dim_set), 1,
769 edge->start = graph->lp->n_ineq;
770 graph->lp = isl_basic_set_extend_constraints(graph->lp,
771 coef->n_eq, coef->n_ineq);
772 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
775 edge->end = graph->lp->n_ineq;
780 /* Add constraints to graph->lp that bound the dependence distance for the given
781 * dependence from a node i to itself.
782 * If s = 1, we add the constraint
784 * c_i_x (y - x) <= m_0 + m_n n
788 * -c_i_x (y - x) + m_0 + m_n n >= 0
790 * for each (x,y) in R.
791 * If s = -1, we add the constraint
793 * -c_i_x (y - x) <= m_0 + m_n n
797 * c_i_x (y - x) + m_0 + m_n n >= 0
799 * for each (x,y) in R.
800 * We obtain general constraints on coefficients (c_0, c_n, c_x)
801 * of valid constraints for (y - x) and then plug in (m_0, m_n, -s * c_i_x),
802 * with each coefficient (except m_0) represented as a pair of non-negative
805 * Actually, we do not construct constraints for the c_i_x themselves,
806 * but for the coefficients of c_i_x written as a linear combination
807 * of the columns in node->cmap.
809 static int add_intra_proximity_constraints(struct isl_sched_graph *graph,
810 struct isl_sched_edge *edge, int s)
814 isl_map *map = isl_map_copy(edge->map);
815 isl_ctx *ctx = isl_map_get_ctx(map);
817 isl_dim_map *dim_map;
819 struct isl_sched_node *node = edge->src;
821 coef = intra_coefficients(graph, map);
823 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
825 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
826 isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap));
828 nparam = isl_space_dim(node->dim, isl_dim_param);
829 total = isl_basic_set_total_dim(graph->lp);
830 dim_map = isl_dim_map_alloc(ctx, total);
831 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
832 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
833 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
834 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
835 isl_space_dim(dim, isl_dim_set), 1,
837 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
838 isl_space_dim(dim, isl_dim_set), 1,
840 graph->lp = isl_basic_set_extend_constraints(graph->lp,
841 coef->n_eq, coef->n_ineq);
842 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
849 /* Add constraints to graph->lp that bound the dependence distance for the given
850 * dependence from node i to node j.
851 * If s = 1, we add the constraint
853 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)
858 * -(c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x) +
861 * for each (x,y) in R.
862 * If s = -1, we add the constraint
864 * -((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x))
869 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) +
872 * for each (x,y) in R.
873 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
874 * of valid constraints for R and then plug in
875 * (m_0 - s*c_j_0 + s*c_i_0, m_n - s*c_j_n + s*c_i_n,
877 * with each coefficient (except m_0, c_j_0 and c_i_0)
878 * represented as a pair of non-negative coefficients.
880 * Actually, we do not construct constraints for the c_*_x themselves,
881 * but for the coefficients of c_*_x written as a linear combination
882 * of the columns in node->cmap.
884 static int add_inter_proximity_constraints(struct isl_sched_graph *graph,
885 struct isl_sched_edge *edge, int s)
889 isl_map *map = isl_map_copy(edge->map);
890 isl_ctx *ctx = isl_map_get_ctx(map);
892 isl_dim_map *dim_map;
894 struct isl_sched_node *src = edge->src;
895 struct isl_sched_node *dst = edge->dst;
897 coef = inter_coefficients(graph, map);
899 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
901 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
902 isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap));
903 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
904 isl_space_dim(dim, isl_dim_set) + src->nvar,
905 isl_mat_copy(dst->cmap));
907 nparam = isl_space_dim(src->dim, isl_dim_param);
908 total = isl_basic_set_total_dim(graph->lp);
909 dim_map = isl_dim_map_alloc(ctx, total);
911 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
912 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
913 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
915 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, -s);
916 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, s);
917 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, -s);
918 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
919 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
921 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
922 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
925 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, s);
926 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, -s);
927 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, s);
928 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
929 isl_space_dim(dim, isl_dim_set), 1,
931 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
932 isl_space_dim(dim, isl_dim_set), 1,
935 graph->lp = isl_basic_set_extend_constraints(graph->lp,
936 coef->n_eq, coef->n_ineq);
937 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
944 static int add_all_validity_constraints(struct isl_sched_graph *graph)
948 for (i = 0; i < graph->n_edge; ++i) {
949 struct isl_sched_edge *edge= &graph->edge[i];
952 if (edge->src != edge->dst)
954 if (add_intra_validity_constraints(graph, edge) < 0)
958 for (i = 0; i < graph->n_edge; ++i) {
959 struct isl_sched_edge *edge = &graph->edge[i];
962 if (edge->src == edge->dst)
964 if (add_inter_validity_constraints(graph, edge) < 0)
971 /* Add constraints to graph->lp that bound the dependence distance
972 * for all dependence relations.
973 * If a given proximity dependence is identical to a validity
974 * dependence, then the dependence distance is already bounded
975 * from below (by zero), so we only need to bound the distance
977 * Otherwise, we need to bound the distance both from above and from below.
979 static int add_all_proximity_constraints(struct isl_sched_graph *graph)
983 for (i = 0; i < graph->n_edge; ++i) {
984 struct isl_sched_edge *edge= &graph->edge[i];
985 if (!edge->proximity)
987 if (edge->src == edge->dst &&
988 add_intra_proximity_constraints(graph, edge, 1) < 0)
990 if (edge->src != edge->dst &&
991 add_inter_proximity_constraints(graph, edge, 1) < 0)
995 if (edge->src == edge->dst &&
996 add_intra_proximity_constraints(graph, edge, -1) < 0)
998 if (edge->src != edge->dst &&
999 add_inter_proximity_constraints(graph, edge, -1) < 0)
1006 /* Compute a basis for the rows in the linear part of the schedule
1007 * and extend this basis to a full basis. The remaining rows
1008 * can then be used to force linear independence from the rows
1011 * In particular, given the schedule rows S, we compute
1015 * with H the Hermite normal form of S. That is, all but the
1016 * first rank columns of Q are zero and so each row in S is
1017 * a linear combination of the first rank rows of Q.
1018 * The matrix Q is then transposed because we will write the
1019 * coefficients of the next schedule row as a column vector s
1020 * and express this s as a linear combination s = Q c of the
1023 static int node_update_cmap(struct isl_sched_node *node)
1026 int n_row = isl_mat_rows(node->sched);
1028 H = isl_mat_sub_alloc(node->sched, 0, n_row,
1029 1 + node->nparam, node->nvar);
1031 H = isl_mat_left_hermite(H, 0, NULL, &Q);
1032 isl_mat_free(node->cmap);
1033 node->cmap = isl_mat_transpose(Q);
1034 node->rank = isl_mat_initial_non_zero_cols(H);
1037 if (!node->cmap || node->rank < 0)
1042 /* Count the number of equality and inequality constraints
1043 * that will be added for the given map.
1044 * If once is set, then we count
1045 * each edge exactly once. Otherwise, we count as follows
1046 * validity -> 1 (>= 0)
1047 * validity+proximity -> 2 (>= 0 and upper bound)
1048 * proximity -> 2 (lower and upper bound)
1050 static int count_map_constraints(struct isl_sched_graph *graph,
1051 struct isl_sched_edge *edge, __isl_take isl_map *map,
1052 int *n_eq, int *n_ineq, int once)
1054 isl_basic_set *coef;
1055 int f = once ? 1 : edge->proximity ? 2 : 1;
1057 if (edge->src == edge->dst)
1058 coef = intra_coefficients(graph, map);
1060 coef = inter_coefficients(graph, map);
1063 *n_eq += f * coef->n_eq;
1064 *n_ineq += f * coef->n_ineq;
1065 isl_basic_set_free(coef);
1070 /* Count the number of equality and inequality constraints
1071 * that will be added to the main lp problem.
1072 * If once is set, then we count
1073 * each edge exactly once. Otherwise, we count as follows
1074 * validity -> 1 (>= 0)
1075 * validity+proximity -> 2 (>= 0 and upper bound)
1076 * proximity -> 2 (lower and upper bound)
1078 static int count_constraints(struct isl_sched_graph *graph,
1079 int *n_eq, int *n_ineq, int once)
1083 *n_eq = *n_ineq = 0;
1084 for (i = 0; i < graph->n_edge; ++i) {
1085 struct isl_sched_edge *edge= &graph->edge[i];
1086 isl_map *map = isl_map_copy(edge->map);
1088 if (count_map_constraints(graph, edge, map,
1089 n_eq, n_ineq, once) < 0)
1096 /* Construct an ILP problem for finding schedule coefficients
1097 * that result in non-negative, but small dependence distances
1098 * over all dependences.
1099 * In particular, the dependence distances over proximity edges
1100 * are bounded by m_0 + m_n n and we compute schedule coefficients
1101 * with small values (preferably zero) of m_n and m_0.
1103 * All variables of the ILP are non-negative. The actual coefficients
1104 * may be negative, so each coefficient is represented as the difference
1105 * of two non-negative variables. The negative part always appears
1106 * immediately before the positive part.
1107 * Other than that, the variables have the following order
1109 * - sum of positive and negative parts of m_n coefficients
1111 * - sum of positive and negative parts of all c_n coefficients
1112 * (unconstrained when computing non-parametric schedules)
1113 * - sum of positive and negative parts of all c_x coefficients
1114 * - positive and negative parts of m_n coefficients
1117 * - positive and negative parts of c_i_n (if parametric)
1118 * - positive and negative parts of c_i_x
1120 * The c_i_x are not represented directly, but through the columns of
1121 * node->cmap. That is, the computed values are for variable t_i_x
1122 * such that c_i_x = Q t_i_x with Q equal to node->cmap.
1124 * The constraints are those from the edges plus two or three equalities
1125 * to express the sums.
1127 * If force_zero is set, then we add equalities to ensure that
1128 * the sum of the m_n coefficients and m_0 are both zero.
1130 static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph,
1142 parametric = ctx->opt->schedule_parametric;
1143 nparam = isl_space_dim(graph->node[0].dim, isl_dim_param);
1145 total = param_pos + 2 * nparam;
1146 for (i = 0; i < graph->n; ++i) {
1147 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
1148 if (node_update_cmap(node) < 0)
1150 node->start = total;
1151 total += 1 + 2 * (node->nparam + node->nvar);
1154 if (count_constraints(graph, &n_eq, &n_ineq, 0) < 0)
1157 dim = isl_space_set_alloc(ctx, 0, total);
1158 isl_basic_set_free(graph->lp);
1159 n_eq += 2 + parametric + force_zero;
1160 graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq);
1162 k = isl_basic_set_alloc_equality(graph->lp);
1165 isl_seq_clr(graph->lp->eq[k], 1 + total);
1167 isl_int_set_si(graph->lp->eq[k][1], -1);
1168 for (i = 0; i < 2 * nparam; ++i)
1169 isl_int_set_si(graph->lp->eq[k][1 + param_pos + i], 1);
1172 k = isl_basic_set_alloc_equality(graph->lp);
1175 isl_seq_clr(graph->lp->eq[k], 1 + total);
1176 isl_int_set_si(graph->lp->eq[k][2], -1);
1180 k = isl_basic_set_alloc_equality(graph->lp);
1183 isl_seq_clr(graph->lp->eq[k], 1 + total);
1184 isl_int_set_si(graph->lp->eq[k][3], -1);
1185 for (i = 0; i < graph->n; ++i) {
1186 int pos = 1 + graph->node[i].start + 1;
1188 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
1189 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1193 k = isl_basic_set_alloc_equality(graph->lp);
1196 isl_seq_clr(graph->lp->eq[k], 1 + total);
1197 isl_int_set_si(graph->lp->eq[k][4], -1);
1198 for (i = 0; i < graph->n; ++i) {
1199 struct isl_sched_node *node = &graph->node[i];
1200 int pos = 1 + node->start + 1 + 2 * node->nparam;
1202 for (j = 0; j < 2 * node->nvar; ++j)
1203 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1206 if (add_all_validity_constraints(graph) < 0)
1208 if (add_all_proximity_constraints(graph) < 0)
1214 /* Analyze the conflicting constraint found by
1215 * isl_tab_basic_set_non_trivial_lexmin. If it corresponds to the validity
1216 * constraint of one of the edges between distinct nodes, living, moreover
1217 * in distinct SCCs, then record the source and sink SCC as this may
1218 * be a good place to cut between SCCs.
1220 static int check_conflict(int con, void *user)
1223 struct isl_sched_graph *graph = user;
1225 if (graph->src_scc >= 0)
1228 con -= graph->lp->n_eq;
1230 if (con >= graph->lp->n_ineq)
1233 for (i = 0; i < graph->n_edge; ++i) {
1234 if (!graph->edge[i].validity)
1236 if (graph->edge[i].src == graph->edge[i].dst)
1238 if (graph->edge[i].src->scc == graph->edge[i].dst->scc)
1240 if (graph->edge[i].start > con)
1242 if (graph->edge[i].end <= con)
1244 graph->src_scc = graph->edge[i].src->scc;
1245 graph->dst_scc = graph->edge[i].dst->scc;
1251 /* Check whether the next schedule row of the given node needs to be
1252 * non-trivial. Lower-dimensional domains may have some trivial rows,
1253 * but as soon as the number of remaining required non-trivial rows
1254 * is as large as the number or remaining rows to be computed,
1255 * all remaining rows need to be non-trivial.
1257 static int needs_row(struct isl_sched_graph *graph, struct isl_sched_node *node)
1259 return node->nvar - node->rank >= graph->maxvar - graph->n_row;
1262 /* Solve the ILP problem constructed in setup_lp.
1263 * For each node such that all the remaining rows of its schedule
1264 * need to be non-trivial, we construct a non-triviality region.
1265 * This region imposes that the next row is independent of previous rows.
1266 * In particular the coefficients c_i_x are represented by t_i_x
1267 * variables with c_i_x = Q t_i_x and Q a unimodular matrix such that
1268 * its first columns span the rows of the previously computed part
1269 * of the schedule. The non-triviality region enforces that at least
1270 * one of the remaining components of t_i_x is non-zero, i.e.,
1271 * that the new schedule row depends on at least one of the remaining
1274 static __isl_give isl_vec *solve_lp(struct isl_sched_graph *graph)
1280 for (i = 0; i < graph->n; ++i) {
1281 struct isl_sched_node *node = &graph->node[i];
1282 int skip = node->rank;
1283 graph->region[i].pos = node->start + 1 + 2*(node->nparam+skip);
1284 if (needs_row(graph, node))
1285 graph->region[i].len = 2 * (node->nvar - skip);
1287 graph->region[i].len = 0;
1289 lp = isl_basic_set_copy(graph->lp);
1290 sol = isl_tab_basic_set_non_trivial_lexmin(lp, 2, graph->n,
1291 graph->region, &check_conflict, graph);
1295 /* Update the schedules of all nodes based on the given solution
1296 * of the LP problem.
1297 * The new row is added to the current band.
1298 * All possibly negative coefficients are encoded as a difference
1299 * of two non-negative variables, so we need to perform the subtraction
1300 * here. Moreover, if use_cmap is set, then the solution does
1301 * not refer to the actual coefficients c_i_x, but instead to variables
1302 * t_i_x such that c_i_x = Q t_i_x and Q is equal to node->cmap.
1303 * In this case, we then also need to perform this multiplication
1304 * to obtain the values of c_i_x.
1306 * If check_zero is set, then the first two coordinates of sol are
1307 * assumed to correspond to the dependence distance. If these two
1308 * coordinates are zero, then the corresponding scheduling dimension
1309 * is marked as being zero distance.
1311 static int update_schedule(struct isl_sched_graph *graph,
1312 __isl_take isl_vec *sol, int use_cmap, int check_zero)
1316 isl_vec *csol = NULL;
1321 isl_die(sol->ctx, isl_error_internal,
1322 "no solution found", goto error);
1325 zero = isl_int_is_zero(sol->el[1]) &&
1326 isl_int_is_zero(sol->el[2]);
1328 for (i = 0; i < graph->n; ++i) {
1329 struct isl_sched_node *node = &graph->node[i];
1330 int pos = node->start;
1331 int row = isl_mat_rows(node->sched);
1334 csol = isl_vec_alloc(sol->ctx, node->nvar);
1338 isl_map_free(node->sched_map);
1339 node->sched_map = NULL;
1340 node->sched = isl_mat_add_rows(node->sched, 1);
1343 node->sched = isl_mat_set_element(node->sched, row, 0,
1345 for (j = 0; j < node->nparam + node->nvar; ++j)
1346 isl_int_sub(sol->el[1 + pos + 1 + 2 * j + 1],
1347 sol->el[1 + pos + 1 + 2 * j + 1],
1348 sol->el[1 + pos + 1 + 2 * j]);
1349 for (j = 0; j < node->nparam; ++j)
1350 node->sched = isl_mat_set_element(node->sched,
1351 row, 1 + j, sol->el[1+pos+1+2*j+1]);
1352 for (j = 0; j < node->nvar; ++j)
1353 isl_int_set(csol->el[j],
1354 sol->el[1+pos+1+2*(node->nparam+j)+1]);
1356 csol = isl_mat_vec_product(isl_mat_copy(node->cmap),
1360 for (j = 0; j < node->nvar; ++j)
1361 node->sched = isl_mat_set_element(node->sched,
1362 row, 1 + node->nparam + j, csol->el[j]);
1363 node->band[graph->n_total_row] = graph->n_band;
1364 node->zero[graph->n_total_row] = zero;
1370 graph->n_total_row++;
1379 /* Convert node->sched into a map and return this map.
1380 * We simply add equality constraints that express each output variable
1381 * as the affine combination of parameters and input variables specified
1382 * by the schedule matrix.
1384 * The result is cached in node->sched_map, which needs to be released
1385 * whenever node->sched is updated.
1387 static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node)
1391 isl_local_space *ls;
1392 isl_basic_map *bmap;
1397 if (node->sched_map)
1398 return isl_map_copy(node->sched_map);
1400 nrow = isl_mat_rows(node->sched);
1401 ncol = isl_mat_cols(node->sched) - 1;
1402 dim = isl_space_from_domain(isl_space_copy(node->dim));
1403 dim = isl_space_add_dims(dim, isl_dim_out, nrow);
1404 bmap = isl_basic_map_universe(isl_space_copy(dim));
1405 ls = isl_local_space_from_space(dim);
1409 for (i = 0; i < nrow; ++i) {
1410 c = isl_equality_alloc(isl_local_space_copy(ls));
1411 isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
1412 isl_mat_get_element(node->sched, i, 0, &v);
1413 isl_constraint_set_constant(c, v);
1414 for (j = 0; j < node->nparam; ++j) {
1415 isl_mat_get_element(node->sched, i, 1 + j, &v);
1416 isl_constraint_set_coefficient(c, isl_dim_param, j, v);
1418 for (j = 0; j < node->nvar; ++j) {
1419 isl_mat_get_element(node->sched,
1420 i, 1 + node->nparam + j, &v);
1421 isl_constraint_set_coefficient(c, isl_dim_in, j, v);
1423 bmap = isl_basic_map_add_constraint(bmap, c);
1428 isl_local_space_free(ls);
1430 node->sched_map = isl_map_from_basic_map(bmap);
1431 return isl_map_copy(node->sched_map);
1434 /* Update the given dependence relation based on the current schedule.
1435 * That is, intersect the dependence relation with a map expressing
1436 * that source and sink are executed within the same iteration of
1437 * the current schedule.
1438 * This is not the most efficient way, but this shouldn't be a critical
1441 static __isl_give isl_map *specialize(__isl_take isl_map *map,
1442 struct isl_sched_node *src, struct isl_sched_node *dst)
1444 isl_map *src_sched, *dst_sched, *id;
1446 src_sched = node_extract_schedule(src);
1447 dst_sched = node_extract_schedule(dst);
1448 id = isl_map_apply_range(src_sched, isl_map_reverse(dst_sched));
1449 return isl_map_intersect(map, id);
1452 /* Update the dependence relations of all edges based on the current schedule.
1453 * If a dependence is carried completely by the current schedule, then
1454 * it is removed and edge_table is updated accordingly.
1456 static int update_edges(isl_ctx *ctx, struct isl_sched_graph *graph)
1459 int reset_table = 0;
1461 for (i = graph->n_edge - 1; i >= 0; --i) {
1462 struct isl_sched_edge *edge = &graph->edge[i];
1463 edge->map = specialize(edge->map, edge->src, edge->dst);
1467 if (isl_map_plain_is_empty(edge->map)) {
1469 isl_map_free(edge->map);
1470 if (i != graph->n_edge - 1)
1471 graph->edge[i] = graph->edge[graph->n_edge - 1];
1477 isl_hash_table_free(ctx, graph->edge_table);
1478 graph->edge_table = NULL;
1479 return graph_init_edge_table(ctx, graph);
1485 static void next_band(struct isl_sched_graph *graph)
1487 graph->band_start = graph->n_total_row;
1491 /* Topologically sort statements mapped to same schedule iteration
1492 * and add a row to the schedule corresponding to this order.
1494 static int sort_statements(isl_ctx *ctx, struct isl_sched_graph *graph)
1501 if (update_edges(ctx, graph) < 0)
1504 if (graph->n_edge == 0)
1507 if (detect_sccs(graph) < 0)
1510 for (i = 0; i < graph->n; ++i) {
1511 struct isl_sched_node *node = &graph->node[i];
1512 int row = isl_mat_rows(node->sched);
1513 int cols = isl_mat_cols(node->sched);
1515 isl_map_free(node->sched_map);
1516 node->sched_map = NULL;
1517 node->sched = isl_mat_add_rows(node->sched, 1);
1520 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1522 for (j = 1; j < cols; ++j)
1523 node->sched = isl_mat_set_element_si(node->sched,
1525 node->band[graph->n_total_row] = graph->n_band;
1528 graph->n_total_row++;
1534 /* Construct an isl_schedule based on the computed schedule stored
1535 * in graph and with parameters specified by dim.
1537 static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph,
1538 __isl_take isl_space *dim)
1542 isl_schedule *sched = NULL;
1547 ctx = isl_space_get_ctx(dim);
1548 sched = isl_calloc(ctx, struct isl_schedule,
1549 sizeof(struct isl_schedule) +
1550 (graph->n - 1) * sizeof(struct isl_schedule_node));
1555 sched->n = graph->n;
1556 sched->n_band = graph->n_band;
1557 sched->n_total_row = graph->n_total_row;
1559 for (i = 0; i < sched->n; ++i) {
1561 int *band_end, *band_id, *zero;
1563 band_end = isl_alloc_array(ctx, int, graph->n_band);
1564 band_id = isl_alloc_array(ctx, int, graph->n_band);
1565 zero = isl_alloc_array(ctx, int, graph->n_total_row);
1566 sched->node[i].sched = node_extract_schedule(&graph->node[i]);
1567 sched->node[i].band_end = band_end;
1568 sched->node[i].band_id = band_id;
1569 sched->node[i].zero = zero;
1570 if (!band_end || !band_id || !zero)
1573 for (r = 0; r < graph->n_total_row; ++r)
1574 zero[r] = graph->node[i].zero[r];
1575 for (r = b = 0; r < graph->n_total_row; ++r) {
1576 if (graph->node[i].band[r] == b)
1579 if (graph->node[i].band[r] == -1)
1582 if (r == graph->n_total_row)
1584 sched->node[i].n_band = b;
1585 for (--b; b >= 0; --b)
1586 band_id[b] = graph->node[i].band_id[b];
1593 isl_space_free(dim);
1594 isl_schedule_free(sched);
1598 /* Copy nodes that satisfy node_pred from the src dependence graph
1599 * to the dst dependence graph.
1601 static int copy_nodes(struct isl_sched_graph *dst, struct isl_sched_graph *src,
1602 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1607 for (i = 0; i < src->n; ++i) {
1608 if (!node_pred(&src->node[i], data))
1610 dst->node[dst->n].dim = isl_space_copy(src->node[i].dim);
1611 dst->node[dst->n].nvar = src->node[i].nvar;
1612 dst->node[dst->n].nparam = src->node[i].nparam;
1613 dst->node[dst->n].sched = isl_mat_copy(src->node[i].sched);
1614 dst->node[dst->n].sched_map =
1615 isl_map_copy(src->node[i].sched_map);
1616 dst->node[dst->n].band = src->node[i].band;
1617 dst->node[dst->n].band_id = src->node[i].band_id;
1618 dst->node[dst->n].zero = src->node[i].zero;
1625 /* Copy non-empty edges that satisfy edge_pred from the src dependence graph
1626 * to the dst dependence graph.
1628 static int copy_edges(isl_ctx *ctx, struct isl_sched_graph *dst,
1629 struct isl_sched_graph *src,
1630 int (*edge_pred)(struct isl_sched_edge *edge, int data), int data)
1635 for (i = 0; i < src->n_edge; ++i) {
1636 struct isl_sched_edge *edge = &src->edge[i];
1639 if (!edge_pred(edge, data))
1642 if (isl_map_plain_is_empty(edge->map))
1645 map = isl_map_copy(edge->map);
1647 dst->edge[dst->n_edge].src =
1648 graph_find_node(ctx, dst, edge->src->dim);
1649 dst->edge[dst->n_edge].dst =
1650 graph_find_node(ctx, dst, edge->dst->dim);
1651 dst->edge[dst->n_edge].map = map;
1652 dst->edge[dst->n_edge].validity = edge->validity;
1653 dst->edge[dst->n_edge].proximity = edge->proximity;
1660 /* Given a "src" dependence graph that contains the nodes from "dst"
1661 * that satisfy node_pred, copy the schedule computed in "src"
1662 * for those nodes back to "dst".
1664 static int copy_schedule(struct isl_sched_graph *dst,
1665 struct isl_sched_graph *src,
1666 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1671 for (i = 0; i < dst->n; ++i) {
1672 if (!node_pred(&dst->node[i], data))
1674 isl_mat_free(dst->node[i].sched);
1675 isl_map_free(dst->node[i].sched_map);
1676 dst->node[i].sched = isl_mat_copy(src->node[src->n].sched);
1677 dst->node[i].sched_map =
1678 isl_map_copy(src->node[src->n].sched_map);
1682 dst->n_total_row = src->n_total_row;
1683 dst->n_band = src->n_band;
1688 /* Compute the maximal number of variables over all nodes.
1689 * This is the maximal number of linearly independent schedule
1690 * rows that we need to compute.
1691 * Just in case we end up in a part of the dependence graph
1692 * with only lower-dimensional domains, we make sure we will
1693 * compute the required amount of extra linearly independent rows.
1695 static int compute_maxvar(struct isl_sched_graph *graph)
1700 for (i = 0; i < graph->n; ++i) {
1701 struct isl_sched_node *node = &graph->node[i];
1704 if (node_update_cmap(node) < 0)
1706 nvar = node->nvar + graph->n_row - node->rank;
1707 if (nvar > graph->maxvar)
1708 graph->maxvar = nvar;
1714 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph);
1715 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph);
1717 /* Compute a schedule for a subgraph of "graph". In particular, for
1718 * the graph composed of nodes that satisfy node_pred and edges that
1719 * that satisfy edge_pred. The caller should precompute the number
1720 * of nodes and edges that satisfy these predicates and pass them along
1721 * as "n" and "n_edge".
1722 * If the subgraph is known to consist of a single component, then wcc should
1723 * be set and then we call compute_schedule_wcc on the constructed subgraph.
1724 * Otherwise, we call compute_schedule, which will check whether the subgraph
1727 static int compute_sub_schedule(isl_ctx *ctx,
1728 struct isl_sched_graph *graph, int n, int n_edge,
1729 int (*node_pred)(struct isl_sched_node *node, int data),
1730 int (*edge_pred)(struct isl_sched_edge *edge, int data),
1733 struct isl_sched_graph split = { 0 };
1735 if (graph_alloc(ctx, &split, n, n_edge) < 0)
1737 if (copy_nodes(&split, graph, node_pred, data) < 0)
1739 if (graph_init_table(ctx, &split) < 0)
1741 if (copy_edges(ctx, &split, graph, edge_pred, data) < 0)
1743 if (graph_init_edge_table(ctx, &split) < 0)
1745 split.n_row = graph->n_row;
1746 split.n_total_row = graph->n_total_row;
1747 split.n_band = graph->n_band;
1748 split.band_start = graph->band_start;
1750 if (wcc && compute_schedule_wcc(ctx, &split) < 0)
1752 if (!wcc && compute_schedule(ctx, &split) < 0)
1755 copy_schedule(graph, &split, node_pred, data);
1757 graph_free(ctx, &split);
1760 graph_free(ctx, &split);
1764 static int node_scc_exactly(struct isl_sched_node *node, int scc)
1766 return node->scc == scc;
1769 static int node_scc_at_most(struct isl_sched_node *node, int scc)
1771 return node->scc <= scc;
1774 static int node_scc_at_least(struct isl_sched_node *node, int scc)
1776 return node->scc >= scc;
1779 static int edge_src_scc_exactly(struct isl_sched_edge *edge, int scc)
1781 return edge->src->scc == scc;
1784 static int edge_dst_scc_at_most(struct isl_sched_edge *edge, int scc)
1786 return edge->dst->scc <= scc;
1789 static int edge_src_scc_at_least(struct isl_sched_edge *edge, int scc)
1791 return edge->src->scc >= scc;
1794 /* Pad the schedules of all nodes with zero rows such that in the end
1795 * they all have graph->n_total_row rows.
1796 * The extra rows don't belong to any band, so they get assigned band number -1.
1798 static int pad_schedule(struct isl_sched_graph *graph)
1802 for (i = 0; i < graph->n; ++i) {
1803 struct isl_sched_node *node = &graph->node[i];
1804 int row = isl_mat_rows(node->sched);
1805 if (graph->n_total_row > row) {
1806 isl_map_free(node->sched_map);
1807 node->sched_map = NULL;
1809 node->sched = isl_mat_add_zero_rows(node->sched,
1810 graph->n_total_row - row);
1813 for (j = row; j < graph->n_total_row; ++j)
1820 /* Split the current graph into two parts and compute a schedule for each
1821 * part individually. In particular, one part consists of all SCCs up
1822 * to and including graph->src_scc, while the other part contains the other
1825 * The split is enforced in the schedule by constant rows with two different
1826 * values (0 and 1). These constant rows replace the previously computed rows
1827 * in the current band.
1828 * It would be possible to reuse them as the first rows in the next
1829 * band, but recomputing them may result in better rows as we are looking
1830 * at a smaller part of the dependence graph.
1832 * The band_id of the second group is set to n, where n is the number
1833 * of nodes in the first group. This ensures that the band_ids over
1834 * the two groups remain disjoint, even if either or both of the two
1835 * groups contain independent components.
1837 static int compute_split_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
1839 int i, j, n, e1, e2;
1840 int n_total_row, orig_total_row;
1841 int n_band, orig_band;
1844 drop = graph->n_total_row - graph->band_start;
1845 graph->n_total_row -= drop;
1846 graph->n_row -= drop;
1849 for (i = 0; i < graph->n; ++i) {
1850 struct isl_sched_node *node = &graph->node[i];
1851 int row = isl_mat_rows(node->sched) - drop;
1852 int cols = isl_mat_cols(node->sched);
1853 int before = node->scc <= graph->src_scc;
1858 isl_map_free(node->sched_map);
1859 node->sched_map = NULL;
1860 node->sched = isl_mat_drop_rows(node->sched,
1861 graph->band_start, drop);
1862 node->sched = isl_mat_add_rows(node->sched, 1);
1865 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1867 for (j = 1; j < cols; ++j)
1868 node->sched = isl_mat_set_element_si(node->sched,
1870 node->band[graph->n_total_row] = graph->n_band;
1874 for (i = 0; i < graph->n_edge; ++i) {
1875 if (graph->edge[i].dst->scc <= graph->src_scc)
1877 if (graph->edge[i].src->scc > graph->src_scc)
1881 graph->n_total_row++;
1884 for (i = 0; i < graph->n; ++i) {
1885 struct isl_sched_node *node = &graph->node[i];
1886 if (node->scc > graph->src_scc)
1887 node->band_id[graph->n_band] = n;
1890 orig_total_row = graph->n_total_row;
1891 orig_band = graph->n_band;
1892 if (compute_sub_schedule(ctx, graph, n, e1,
1893 &node_scc_at_most, &edge_dst_scc_at_most,
1894 graph->src_scc, 0) < 0)
1896 n_total_row = graph->n_total_row;
1897 graph->n_total_row = orig_total_row;
1898 n_band = graph->n_band;
1899 graph->n_band = orig_band;
1900 if (compute_sub_schedule(ctx, graph, graph->n - n, e2,
1901 &node_scc_at_least, &edge_src_scc_at_least,
1902 graph->src_scc + 1, 0) < 0)
1904 if (n_total_row > graph->n_total_row)
1905 graph->n_total_row = n_total_row;
1906 if (n_band > graph->n_band)
1907 graph->n_band = n_band;
1909 return pad_schedule(graph);
1912 /* Compute the next band of the schedule after updating the dependence
1913 * relations based on the the current schedule.
1915 static int compute_next_band(isl_ctx *ctx, struct isl_sched_graph *graph)
1917 if (update_edges(ctx, graph) < 0)
1921 return compute_schedule(ctx, graph);
1924 /* Add constraints to graph->lp that force the dependence "map" (which
1925 * is part of the dependence relation of "edge")
1926 * to be respected and attempt to carry it, where the edge is one from
1927 * a node j to itself. "pos" is the sequence number of the given map.
1928 * That is, add constraints that enforce
1930 * (c_j_0 + c_j_n n + c_j_x y) - (c_j_0 + c_j_n n + c_j_x x)
1931 * = c_j_x (y - x) >= e_i
1933 * for each (x,y) in R.
1934 * We obtain general constraints on coefficients (c_0, c_n, c_x)
1935 * of valid constraints for (y - x) and then plug in (-e_i, 0, c_j_x),
1936 * with each coefficient in c_j_x represented as a pair of non-negative
1939 static int add_intra_constraints(struct isl_sched_graph *graph,
1940 struct isl_sched_edge *edge, __isl_take isl_map *map, int pos)
1943 isl_ctx *ctx = isl_map_get_ctx(map);
1945 isl_dim_map *dim_map;
1946 isl_basic_set *coef;
1947 struct isl_sched_node *node = edge->src;
1949 coef = intra_coefficients(graph, map);
1951 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
1953 total = isl_basic_set_total_dim(graph->lp);
1954 dim_map = isl_dim_map_alloc(ctx, total);
1955 isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
1956 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
1957 isl_space_dim(dim, isl_dim_set), 1,
1959 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
1960 isl_space_dim(dim, isl_dim_set), 1,
1962 graph->lp = isl_basic_set_extend_constraints(graph->lp,
1963 coef->n_eq, coef->n_ineq);
1964 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
1966 isl_space_free(dim);
1971 /* Add constraints to graph->lp that force the dependence "map" (which
1972 * is part of the dependence relation of "edge")
1973 * to be respected and attempt to carry it, where the edge is one from
1974 * node j to node k. "pos" is the sequence number of the given map.
1975 * That is, add constraints that enforce
1977 * (c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i
1979 * for each (x,y) in R.
1980 * We obtain general constraints on coefficients (c_0, c_n, c_x)
1981 * of valid constraints for R and then plug in
1982 * (-e_i + c_k_0 - c_j_0, c_k_n - c_j_n, c_k_x - c_j_x)
1983 * with each coefficient (except e_i, c_k_0 and c_j_0)
1984 * represented as a pair of non-negative coefficients.
1986 static int add_inter_constraints(struct isl_sched_graph *graph,
1987 struct isl_sched_edge *edge, __isl_take isl_map *map, int pos)
1990 isl_ctx *ctx = isl_map_get_ctx(map);
1992 isl_dim_map *dim_map;
1993 isl_basic_set *coef;
1994 struct isl_sched_node *src = edge->src;
1995 struct isl_sched_node *dst = edge->dst;
1997 coef = inter_coefficients(graph, map);
1999 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
2001 total = isl_basic_set_total_dim(graph->lp);
2002 dim_map = isl_dim_map_alloc(ctx, total);
2004 isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
2006 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
2007 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
2008 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
2009 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
2010 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
2012 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
2013 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
2016 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
2017 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
2018 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
2019 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
2020 isl_space_dim(dim, isl_dim_set), 1,
2022 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
2023 isl_space_dim(dim, isl_dim_set), 1,
2026 graph->lp = isl_basic_set_extend_constraints(graph->lp,
2027 coef->n_eq, coef->n_ineq);
2028 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
2030 isl_space_free(dim);
2035 /* Add constraints to graph->lp that force all dependence
2036 * to be respected and attempt to carry it.
2038 static int add_all_constraints(struct isl_sched_graph *graph)
2044 for (i = 0; i < graph->n_edge; ++i) {
2045 struct isl_sched_edge *edge= &graph->edge[i];
2046 for (j = 0; j < edge->map->n; ++j) {
2047 isl_basic_map *bmap;
2050 bmap = isl_basic_map_copy(edge->map->p[j]);
2051 map = isl_map_from_basic_map(bmap);
2053 if (edge->src == edge->dst &&
2054 add_intra_constraints(graph, edge, map, pos) < 0)
2056 if (edge->src != edge->dst &&
2057 add_inter_constraints(graph, edge, map, pos) < 0)
2066 /* Count the number of equality and inequality constraints
2067 * that will be added to the carry_lp problem.
2068 * If once is set, then we count
2069 * each edge exactly once. Otherwise, we count as follows
2070 * validity -> 1 (>= 0)
2071 * validity+proximity -> 2 (>= 0 and upper bound)
2072 * proximity -> 2 (lower and upper bound)
2074 static int count_all_constraints(struct isl_sched_graph *graph,
2075 int *n_eq, int *n_ineq, int once)
2079 *n_eq = *n_ineq = 0;
2080 for (i = 0; i < graph->n_edge; ++i) {
2081 struct isl_sched_edge *edge= &graph->edge[i];
2082 for (j = 0; j < edge->map->n; ++j) {
2083 isl_basic_map *bmap;
2086 bmap = isl_basic_map_copy(edge->map->p[j]);
2087 map = isl_map_from_basic_map(bmap);
2089 if (count_map_constraints(graph, edge, map,
2090 n_eq, n_ineq, once) < 0)
2098 /* Construct an LP problem for finding schedule coefficients
2099 * such that the schedule carries as many dependences as possible.
2100 * In particular, for each dependence i, we bound the dependence distance
2101 * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum
2102 * of all e_i's. Dependence with e_i = 0 in the solution are simply
2103 * respected, while those with e_i > 0 (in practice e_i = 1) are carried.
2104 * Note that if the dependence relation is a union of basic maps,
2105 * then we have to consider each basic map individually as it may only
2106 * be possible to carry the dependences expressed by some of those
2107 * basic maps and not all off them.
2108 * Below, we consider each of those basic maps as a separate "edge".
2110 * All variables of the LP are non-negative. The actual coefficients
2111 * may be negative, so each coefficient is represented as the difference
2112 * of two non-negative variables. The negative part always appears
2113 * immediately before the positive part.
2114 * Other than that, the variables have the following order
2116 * - sum of (1 - e_i) over all edges
2117 * - sum of positive and negative parts of all c_n coefficients
2118 * (unconstrained when computing non-parametric schedules)
2119 * - sum of positive and negative parts of all c_x coefficients
2124 * - positive and negative parts of c_i_n (if parametric)
2125 * - positive and negative parts of c_i_x
2127 * The constraints are those from the edges plus three equalities
2128 * to express the sums and n_edge inequalities to express e_i <= 1.
2130 static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
2140 for (i = 0; i < graph->n_edge; ++i)
2141 n_edge += graph->edge[i].map->n;
2144 for (i = 0; i < graph->n; ++i) {
2145 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
2146 node->start = total;
2147 total += 1 + 2 * (node->nparam + node->nvar);
2150 if (count_all_constraints(graph, &n_eq, &n_ineq, 1) < 0)
2153 dim = isl_space_set_alloc(ctx, 0, total);
2154 isl_basic_set_free(graph->lp);
2157 graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq);
2158 graph->lp = isl_basic_set_set_rational(graph->lp);
2160 k = isl_basic_set_alloc_equality(graph->lp);
2163 isl_seq_clr(graph->lp->eq[k], 1 + total);
2164 isl_int_set_si(graph->lp->eq[k][0], -n_edge);
2165 isl_int_set_si(graph->lp->eq[k][1], 1);
2166 for (i = 0; i < n_edge; ++i)
2167 isl_int_set_si(graph->lp->eq[k][4 + i], 1);
2169 k = isl_basic_set_alloc_equality(graph->lp);
2172 isl_seq_clr(graph->lp->eq[k], 1 + total);
2173 isl_int_set_si(graph->lp->eq[k][2], -1);
2174 for (i = 0; i < graph->n; ++i) {
2175 int pos = 1 + graph->node[i].start + 1;
2177 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
2178 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2181 k = isl_basic_set_alloc_equality(graph->lp);
2184 isl_seq_clr(graph->lp->eq[k], 1 + total);
2185 isl_int_set_si(graph->lp->eq[k][3], -1);
2186 for (i = 0; i < graph->n; ++i) {
2187 struct isl_sched_node *node = &graph->node[i];
2188 int pos = 1 + node->start + 1 + 2 * node->nparam;
2190 for (j = 0; j < 2 * node->nvar; ++j)
2191 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2194 for (i = 0; i < n_edge; ++i) {
2195 k = isl_basic_set_alloc_inequality(graph->lp);
2198 isl_seq_clr(graph->lp->ineq[k], 1 + total);
2199 isl_int_set_si(graph->lp->ineq[k][4 + i], -1);
2200 isl_int_set_si(graph->lp->ineq[k][0], 1);
2203 if (add_all_constraints(graph) < 0)
2209 /* If the schedule_split_parallel option is set and if the linear
2210 * parts of the scheduling rows for all nodes in the graphs are the same,
2211 * then split off the constant term from the linear part.
2212 * The constant term is then placed in a separate band and
2213 * the linear part is simplified.
2215 static int split_parallel(isl_ctx *ctx, struct isl_sched_graph *graph)
2220 struct isl_sched_node *node0;
2222 if (!ctx->opt->schedule_split_parallel)
2227 node0 = &graph->node[0];
2228 row = isl_mat_rows(node0->sched) - 1;
2229 cols = isl_mat_cols(node0->sched);
2230 for (i = 1; i < graph->n; ++i) {
2231 struct isl_sched_node *node = &graph->node[i];
2233 if (!isl_seq_eq(node0->sched->row[row] + 1,
2234 node->sched->row[row] + 1, cols - 1))
2237 isl_int_ne(node0->sched->row[row][0],
2238 node->sched->row[row][0]))
2246 for (i = 0; i < graph->n; ++i) {
2247 struct isl_sched_node *node = &graph->node[i];
2249 isl_map_free(node->sched_map);
2250 node->sched_map = NULL;
2251 node->sched = isl_mat_add_zero_rows(node->sched, 1);
2254 isl_int_set(node->sched->row[row + 1][0],
2255 node->sched->row[row][0]);
2256 isl_int_set_si(node->sched->row[row][0], 0);
2257 node->sched = isl_mat_normalize_row(node->sched, row);
2260 node->band[graph->n_total_row] = graph->n_band;
2263 graph->n_total_row++;
2268 /* Construct a schedule row for each node such that as many dependences
2269 * as possible are carried and then continue with the next band.
2271 static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *graph)
2279 for (i = 0; i < graph->n_edge; ++i)
2280 n_edge += graph->edge[i].map->n;
2282 if (setup_carry_lp(ctx, graph) < 0)
2285 lp = isl_basic_set_copy(graph->lp);
2286 sol = isl_tab_basic_set_non_neg_lexmin(lp);
2290 if (sol->size == 0) {
2292 isl_die(ctx, isl_error_internal,
2293 "error in schedule construction", return -1);
2296 if (isl_int_cmp_si(sol->el[1], n_edge) >= 0) {
2298 isl_die(ctx, isl_error_unknown,
2299 "unable to carry dependences", return -1);
2302 if (update_schedule(graph, sol, 0, 0) < 0)
2305 if (split_parallel(ctx, graph) < 0)
2308 return compute_next_band(ctx, graph);
2311 /* Compute a schedule for a connected dependence graph.
2312 * We try to find a sequence of as many schedule rows as possible that result
2313 * in non-negative dependence distances (independent of the previous rows
2314 * in the sequence, i.e., such that the sequence is tilable).
2315 * If we can't find any more rows we either
2316 * - split between SCCs and start over (assuming we found an interesting
2317 * pair of SCCs between which to split)
2318 * - continue with the next band (assuming the current band has at least
2320 * - try to carry as many dependences as possible and continue with the next
2323 * If we manage to complete the schedule, we finish off by topologically
2324 * sorting the statements based on the remaining dependences.
2326 * If ctx->opt->schedule_outer_zero_distance is set, then we force the
2327 * outermost dimension in the current band to be zero distance. If this
2328 * turns out to be impossible, we fall back on the general scheme above
2329 * and try to carry as many dependences as possible.
2331 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph)
2335 if (detect_sccs(graph) < 0)
2339 if (compute_maxvar(graph) < 0)
2342 if (ctx->opt->schedule_outer_zero_distance)
2345 while (graph->n_row < graph->maxvar) {
2348 graph->src_scc = -1;
2349 graph->dst_scc = -1;
2351 if (setup_lp(ctx, graph, force_zero) < 0)
2353 sol = solve_lp(graph);
2356 if (sol->size == 0) {
2358 if (!ctx->opt->schedule_maximize_band_depth &&
2359 graph->n_total_row > graph->band_start)
2360 return compute_next_band(ctx, graph);
2361 if (graph->src_scc >= 0)
2362 return compute_split_schedule(ctx, graph);
2363 if (graph->n_total_row > graph->band_start)
2364 return compute_next_band(ctx, graph);
2365 return carry_dependences(ctx, graph);
2367 if (update_schedule(graph, sol, 1, 1) < 0)
2372 if (graph->n_total_row > graph->band_start)
2374 return sort_statements(ctx, graph);
2377 /* Compute a schedule for each component (identified by node->scc)
2378 * of the dependence graph separately and then combine the results.
2380 * The band_id is adjusted such that each component has a separate id.
2381 * Note that the band_id may have already been set to a value different
2382 * from zero by compute_split_schedule.
2384 static int compute_component_schedule(isl_ctx *ctx,
2385 struct isl_sched_graph *graph)
2389 int n_total_row, orig_total_row;
2390 int n_band, orig_band;
2393 orig_total_row = graph->n_total_row;
2395 orig_band = graph->n_band;
2396 for (i = 0; i < graph->n; ++i)
2397 graph->node[i].band_id[graph->n_band] += graph->node[i].scc;
2398 for (wcc = 0; wcc < graph->scc; ++wcc) {
2400 for (i = 0; i < graph->n; ++i)
2401 if (graph->node[i].scc == wcc)
2404 for (i = 0; i < graph->n_edge; ++i)
2405 if (graph->edge[i].src->scc == wcc)
2408 if (compute_sub_schedule(ctx, graph, n, n_edge,
2410 &edge_src_scc_exactly, wcc, 1) < 0)
2412 if (graph->n_total_row > n_total_row)
2413 n_total_row = graph->n_total_row;
2414 graph->n_total_row = orig_total_row;
2415 if (graph->n_band > n_band)
2416 n_band = graph->n_band;
2417 graph->n_band = orig_band;
2420 graph->n_total_row = n_total_row;
2421 graph->n_band = n_band;
2423 return pad_schedule(graph);
2426 /* Compute a schedule for the given dependence graph.
2427 * We first check if the graph is connected (through validity dependences)
2428 * and if so compute a schedule for each component separately.
2430 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
2432 if (detect_wccs(graph) < 0)
2436 return compute_component_schedule(ctx, graph);
2438 return compute_schedule_wcc(ctx, graph);
2441 /* Compute a schedule for the given union of domains that respects
2442 * all the validity dependences and tries to minimize the dependence
2443 * distances over the proximity dependences.
2445 __isl_give isl_schedule *isl_union_set_compute_schedule(
2446 __isl_take isl_union_set *domain,
2447 __isl_take isl_union_map *validity,
2448 __isl_take isl_union_map *proximity)
2450 isl_ctx *ctx = isl_union_set_get_ctx(domain);
2452 struct isl_sched_graph graph = { 0 };
2453 isl_schedule *sched;
2455 domain = isl_union_set_align_params(domain,
2456 isl_union_map_get_space(validity));
2457 domain = isl_union_set_align_params(domain,
2458 isl_union_map_get_space(proximity));
2459 dim = isl_union_set_get_space(domain);
2460 validity = isl_union_map_align_params(validity, isl_space_copy(dim));
2461 proximity = isl_union_map_align_params(proximity, dim);
2466 graph.n = isl_union_set_n_set(domain);
2469 if (graph_alloc(ctx, &graph, graph.n,
2470 isl_union_map_n_map(validity) + isl_union_map_n_map(proximity)) < 0)
2474 if (isl_union_set_foreach_set(domain, &extract_node, &graph) < 0)
2476 if (graph_init_table(ctx, &graph) < 0)
2479 if (isl_union_map_foreach_map(validity, &extract_edge, &graph) < 0)
2481 if (graph_init_edge_table(ctx, &graph) < 0)
2483 if (isl_union_map_foreach_map(proximity, &extract_edge, &graph) < 0)
2486 if (compute_schedule(ctx, &graph) < 0)
2490 sched = extract_schedule(&graph, isl_union_set_get_space(domain));
2492 graph_free(ctx, &graph);
2493 isl_union_set_free(domain);
2494 isl_union_map_free(validity);
2495 isl_union_map_free(proximity);
2499 graph_free(ctx, &graph);
2500 isl_union_set_free(domain);
2501 isl_union_map_free(validity);
2502 isl_union_map_free(proximity);
2506 void *isl_schedule_free(__isl_take isl_schedule *sched)
2512 if (--sched->ref > 0)
2515 for (i = 0; i < sched->n; ++i) {
2516 isl_map_free(sched->node[i].sched);
2517 free(sched->node[i].band_end);
2518 free(sched->node[i].band_id);
2519 free(sched->node[i].zero);
2521 isl_space_free(sched->dim);
2522 isl_band_list_free(sched->band_forest);
2527 isl_ctx *isl_schedule_get_ctx(__isl_keep isl_schedule *schedule)
2529 return schedule ? isl_space_get_ctx(schedule->dim) : NULL;
2532 __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched)
2535 isl_union_map *umap;
2540 umap = isl_union_map_empty(isl_space_copy(sched->dim));
2541 for (i = 0; i < sched->n; ++i)
2542 umap = isl_union_map_add_map(umap,
2543 isl_map_copy(sched->node[i].sched));
2548 static __isl_give isl_band_list *construct_band_list(
2549 __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
2550 int band_nr, int *parent_active, int n_active);
2552 /* Construct an isl_band structure for the band in the given schedule
2553 * with sequence number band_nr for the n_active nodes marked by active.
2554 * If the nodes don't have a band with the given sequence number,
2555 * then a band without members is created.
2557 * Because of the way the schedule is constructed, we know that
2558 * the position of the band inside the schedule of a node is the same
2559 * for all active nodes.
2561 static __isl_give isl_band *construct_band(__isl_keep isl_schedule *schedule,
2562 __isl_keep isl_band *parent,
2563 int band_nr, int *active, int n_active)
2566 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2568 unsigned start, end;
2570 band = isl_calloc_type(ctx, isl_band);
2575 band->schedule = schedule;
2576 band->parent = parent;
2578 for (i = 0; i < schedule->n; ++i)
2579 if (active[i] && schedule->node[i].n_band > band_nr + 1)
2582 if (i < schedule->n) {
2583 band->children = construct_band_list(schedule, band,
2584 band_nr + 1, active, n_active);
2585 if (!band->children)
2589 for (i = 0; i < schedule->n; ++i)
2593 if (i >= schedule->n)
2594 isl_die(ctx, isl_error_internal,
2595 "band without active statements", goto error);
2597 start = band_nr ? schedule->node[i].band_end[band_nr - 1] : 0;
2598 end = band_nr < schedule->node[i].n_band ?
2599 schedule->node[i].band_end[band_nr] : start;
2600 band->n = end - start;
2602 band->zero = isl_alloc_array(ctx, int, band->n);
2606 for (j = 0; j < band->n; ++j)
2607 band->zero[j] = schedule->node[i].zero[start + j];
2609 band->map = isl_union_map_empty(isl_space_copy(schedule->dim));
2610 for (i = 0; i < schedule->n; ++i) {
2617 map = isl_map_copy(schedule->node[i].sched);
2618 n_out = isl_map_dim(map, isl_dim_out);
2619 map = isl_map_project_out(map, isl_dim_out, end, n_out - end);
2620 map = isl_map_project_out(map, isl_dim_out, 0, start);
2621 band->map = isl_union_map_union(band->map,
2622 isl_union_map_from_map(map));
2629 isl_band_free(band);
2633 /* Construct a list of bands that start at the same position (with
2634 * sequence number band_nr) in the schedules of the nodes that
2635 * were active in the parent band.
2637 * A separate isl_band structure is created for each band_id
2638 * and for each node that does not have a band with sequence
2639 * number band_nr. In the latter case, a band without members
2641 * This ensures that if a band has any children, then each node
2642 * that was active in the band is active in exactly one of the children.
2644 static __isl_give isl_band_list *construct_band_list(
2645 __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
2646 int band_nr, int *parent_active, int n_active)
2649 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2652 isl_band_list *list;
2655 for (i = 0; i < n_active; ++i) {
2656 for (j = 0; j < schedule->n; ++j) {
2657 if (!parent_active[j])
2659 if (schedule->node[j].n_band <= band_nr)
2661 if (schedule->node[j].band_id[band_nr] == i) {
2667 for (j = 0; j < schedule->n; ++j)
2668 if (schedule->node[j].n_band <= band_nr)
2673 list = isl_band_list_alloc(ctx, n_band);
2674 band = construct_band(schedule, parent, band_nr,
2675 parent_active, n_active);
2676 return isl_band_list_add(list, band);
2679 active = isl_alloc_array(ctx, int, schedule->n);
2683 list = isl_band_list_alloc(ctx, n_band);
2685 for (i = 0; i < n_active; ++i) {
2689 for (j = 0; j < schedule->n; ++j) {
2690 active[j] = parent_active[j] &&
2691 schedule->node[j].n_band > band_nr &&
2692 schedule->node[j].band_id[band_nr] == i;
2699 band = construct_band(schedule, parent, band_nr, active, n);
2701 list = isl_band_list_add(list, band);
2703 for (i = 0; i < schedule->n; ++i) {
2705 if (!parent_active[i])
2707 if (schedule->node[i].n_band > band_nr)
2709 for (j = 0; j < schedule->n; ++j)
2711 band = construct_band(schedule, parent, band_nr, active, 1);
2712 list = isl_band_list_add(list, band);
2720 /* Construct a band forest representation of the schedule and
2721 * return the list of roots.
2723 static __isl_give isl_band_list *construct_forest(
2724 __isl_keep isl_schedule *schedule)
2727 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2728 isl_band_list *forest;
2731 active = isl_alloc_array(ctx, int, schedule->n);
2735 for (i = 0; i < schedule->n; ++i)
2738 forest = construct_band_list(schedule, NULL, 0, active, schedule->n);
2745 /* Return the roots of a band forest representation of the schedule.
2747 __isl_give isl_band_list *isl_schedule_get_band_forest(
2748 __isl_keep isl_schedule *schedule)
2752 if (!schedule->band_forest)
2753 schedule->band_forest = construct_forest(schedule);
2754 return isl_band_list_dup(schedule->band_forest);
2757 static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
2758 __isl_keep isl_band_list *list);
2760 static __isl_give isl_printer *print_band(__isl_take isl_printer *p,
2761 __isl_keep isl_band *band)
2763 isl_band_list *children;
2765 p = isl_printer_start_line(p);
2766 p = isl_printer_print_union_map(p, band->map);
2767 p = isl_printer_end_line(p);
2769 if (!isl_band_has_children(band))
2772 children = isl_band_get_children(band);
2774 p = isl_printer_indent(p, 4);
2775 p = print_band_list(p, children);
2776 p = isl_printer_indent(p, -4);
2778 isl_band_list_free(children);
2783 static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
2784 __isl_keep isl_band_list *list)
2788 n = isl_band_list_n_band(list);
2789 for (i = 0; i < n; ++i) {
2791 band = isl_band_list_get_band(list, i);
2792 p = print_band(p, band);
2793 isl_band_free(band);
2799 __isl_give isl_printer *isl_printer_print_schedule(__isl_take isl_printer *p,
2800 __isl_keep isl_schedule *schedule)
2802 isl_band_list *forest;
2804 forest = isl_schedule_get_band_forest(schedule);
2806 p = print_band_list(p, forest);
2808 isl_band_list_free(forest);
2813 void isl_schedule_dump(__isl_keep isl_schedule *schedule)
2815 isl_printer *printer;
2820 printer = isl_printer_to_file(isl_schedule_get_ctx(schedule), stderr);
2821 printer = isl_printer_print_schedule(printer, schedule);
2823 isl_printer_free(printer);