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>
27 #include <isl_options_private.h>
30 * The scheduling algorithm implemented in this file was inspired by
31 * Bondhugula et al., "Automatic Transformations for Communication-Minimized
32 * Parallelization and Locality Optimization in the Polyhedral Model".
36 /* Internal information about a node that is used during the construction
38 * dim represents the space in which the domain lives
39 * sched is a matrix representation of the schedule being constructed
41 * sched_map is an isl_map representation of the same (partial) schedule
42 * sched_map may be NULL
43 * rank is the number of linearly independent rows in the linear part
45 * the columns of cmap represent a change of basis for the schedule
46 * coefficients; the first rank columns span the linear part of
48 * start is the first variable in the LP problem in the sequences that
49 * represents the schedule coefficients of this node
50 * nvar is the dimension of the domain
51 * nparam is the number of parameters or 0 if we are not constructing
52 * a parametric schedule
54 * scc is the index of SCC (or WCC) this node belongs to
56 * band contains the band index for each of the rows of the schedule.
57 * band_id is used to differentiate between separate bands at the same
58 * level within the same parent band, i.e., bands that are separated
59 * by the parent band or bands that are independent of each other.
60 * zero contains a boolean for each of the rows of the schedule,
61 * indicating whether the corresponding scheduling dimension results
62 * in zero dependence distances within its band and with respect
63 * to the proximity edges.
65 * index, min_index and on_stack are used during the SCC detection
66 * index represents the order in which nodes are visited.
67 * min_index is the index of the root of a (sub)component.
68 * on_stack indicates whether the node is currently on the stack.
70 struct isl_sched_node {
92 static int node_has_dim(const void *entry, const void *val)
94 struct isl_sched_node *node = (struct isl_sched_node *)entry;
95 isl_space *dim = (isl_space *)val;
97 return isl_space_is_equal(node->dim, dim);
100 /* An edge in the dependence graph. An edge may be used to
101 * ensure validity of the generated schedule, to minimize the dependence
104 * map is the dependence relation
105 * src is the source node
106 * dst is the sink node
107 * validity is set if the edge is used to ensure correctness
108 * proximity is set if the edge is used to minimize dependence distances
110 * For validity edges, start and end mark the sequence of inequality
111 * constraints in the LP problem that encode the validity constraint
112 * corresponding to this edge.
114 struct isl_sched_edge {
117 struct isl_sched_node *src;
118 struct isl_sched_node *dst;
128 isl_edge_validity = 0,
130 isl_edge_last = isl_edge_proximity
133 /* Internal information about the dependence graph used during
134 * the construction of the schedule.
136 * intra_hmap is a cache, mapping dependence relations to their dual,
137 * for dependences from a node to itself
138 * inter_hmap is a cache, mapping dependence relations to their dual,
139 * for dependences between distinct nodes
141 * n is the number of nodes
142 * node is the list of nodes
143 * maxvar is the maximal number of variables over all nodes
144 * n_row is the current (maximal) number of linearly independent
145 * rows in the node schedules
146 * n_total_row is the current number of rows in the node schedules
147 * n_band is the current number of completed bands
148 * band_start is the starting row in the node schedules of the current band
149 * root is set if this graph is the original dependence graph,
150 * without any splitting
152 * sorted contains a list of node indices sorted according to the
153 * SCC to which a node belongs
155 * n_edge is the number of edges
156 * edge is the list of edges
157 * max_edge contains the maximal number of edges of each type;
158 * in particular, it contains the number of edges in the inital graph.
159 * edge_table contains pointers into the edge array, hashed on the source
160 * and sink spaces; there is one such table for each type;
161 * a given edge may be referenced from more than one table
162 * if the corresponding relation appears in more than of the
163 * sets of dependences
165 * node_table contains pointers into the node array, hashed on the space
167 * region contains a list of variable sequences that should be non-trivial
169 * lp contains the (I)LP problem used to obtain new schedule rows
171 * src_scc and dst_scc are the source and sink SCCs of an edge with
172 * conflicting constraints
174 * scc, sp, index and stack are used during the detection of SCCs
175 * scc is the number of the next SCC
176 * stack contains the nodes on the path from the root to the current node
177 * sp is the stack pointer
178 * index is the index of the last node visited
180 struct isl_sched_graph {
181 isl_hmap_map_basic_set *intra_hmap;
182 isl_hmap_map_basic_set *inter_hmap;
184 struct isl_sched_node *node;
197 struct isl_sched_edge *edge;
199 int max_edge[isl_edge_last + 1];
200 struct isl_hash_table *edge_table[isl_edge_last + 1];
202 struct isl_hash_table *node_table;
203 struct isl_region *region;
217 /* Initialize node_table based on the list of nodes.
219 static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph)
223 graph->node_table = isl_hash_table_alloc(ctx, graph->n);
224 if (!graph->node_table)
227 for (i = 0; i < graph->n; ++i) {
228 struct isl_hash_table_entry *entry;
231 hash = isl_space_get_hash(graph->node[i].dim);
232 entry = isl_hash_table_find(ctx, graph->node_table, hash,
234 graph->node[i].dim, 1);
237 entry->data = &graph->node[i];
243 /* Return a pointer to the node that lives within the given space,
244 * or NULL if there is no such node.
246 static struct isl_sched_node *graph_find_node(isl_ctx *ctx,
247 struct isl_sched_graph *graph, __isl_keep isl_space *dim)
249 struct isl_hash_table_entry *entry;
252 hash = isl_space_get_hash(dim);
253 entry = isl_hash_table_find(ctx, graph->node_table, hash,
254 &node_has_dim, dim, 0);
256 return entry ? entry->data : NULL;
259 static int edge_has_src_and_dst(const void *entry, const void *val)
261 const struct isl_sched_edge *edge = entry;
262 const struct isl_sched_edge *temp = val;
264 return edge->src == temp->src && edge->dst == temp->dst;
267 /* Add the given edge to graph->edge_table[type].
269 static int graph_edge_table_add(isl_ctx *ctx, struct isl_sched_graph *graph,
270 enum isl_edge_type type, struct isl_sched_edge *edge)
272 struct isl_hash_table_entry *entry;
275 hash = isl_hash_init();
276 hash = isl_hash_builtin(hash, edge->src);
277 hash = isl_hash_builtin(hash, edge->dst);
278 entry = isl_hash_table_find(ctx, graph->edge_table[type], hash,
279 &edge_has_src_and_dst, edge, 1);
287 /* Allocate the edge_tables based on the maximal number of edges of
290 static int graph_init_edge_tables(isl_ctx *ctx, struct isl_sched_graph *graph)
294 for (i = 0; i <= isl_edge_last; ++i) {
295 graph->edge_table[i] = isl_hash_table_alloc(ctx,
297 if (!graph->edge_table[i])
304 /* If graph->edge_table[type] contains an edge from the given source
305 * to the given destination, then return the hash table entry of this edge.
306 * Otherwise, return NULL.
308 static struct isl_hash_table_entry *graph_find_edge_entry(
309 struct isl_sched_graph *graph,
310 enum isl_edge_type type,
311 struct isl_sched_node *src, struct isl_sched_node *dst)
313 isl_ctx *ctx = isl_space_get_ctx(src->dim);
315 struct isl_sched_edge temp = { .src = src, .dst = dst };
317 hash = isl_hash_init();
318 hash = isl_hash_builtin(hash, temp.src);
319 hash = isl_hash_builtin(hash, temp.dst);
320 return isl_hash_table_find(ctx, graph->edge_table[type], hash,
321 &edge_has_src_and_dst, &temp, 0);
325 /* If graph->edge_table[type] contains an edge from the given source
326 * to the given destination, then return this edge.
327 * Otherwise, return NULL.
329 static struct isl_sched_edge *graph_find_edge(struct isl_sched_graph *graph,
330 enum isl_edge_type type,
331 struct isl_sched_node *src, struct isl_sched_node *dst)
333 struct isl_hash_table_entry *entry;
335 entry = graph_find_edge_entry(graph, type, src, dst);
342 /* Check whether the dependence graph has an edge of the give type
343 * between the given two nodes.
345 static int graph_has_edge(struct isl_sched_graph *graph,
346 enum isl_edge_type type,
347 struct isl_sched_node *src, struct isl_sched_node *dst)
349 struct isl_sched_edge *edge;
352 edge = graph_find_edge(graph, type, src, dst);
356 empty = isl_map_plain_is_empty(edge->map);
363 /* If there is an edge from the given source to the given destination
364 * of any type then return this edge.
365 * Otherwise, return NULL.
367 static struct isl_sched_edge *graph_find_any_edge(struct isl_sched_graph *graph,
368 struct isl_sched_node *src, struct isl_sched_node *dst)
371 struct isl_sched_edge *edge;
373 for (i = 0; i <= isl_edge_last; ++i) {
374 edge = graph_find_edge(graph, i, src, dst);
382 /* Remove the given edge from all the edge_tables that refer to it.
384 static void graph_remove_edge(struct isl_sched_graph *graph,
385 struct isl_sched_edge *edge)
387 isl_ctx *ctx = isl_map_get_ctx(edge->map);
390 for (i = 0; i <= isl_edge_last; ++i) {
391 struct isl_hash_table_entry *entry;
393 entry = graph_find_edge_entry(graph, i, edge->src, edge->dst);
396 if (entry->data != edge)
398 isl_hash_table_remove(ctx, graph->edge_table[i], entry);
402 /* Check whether the dependence graph has any edge
403 * between the given two nodes.
405 static int graph_has_any_edge(struct isl_sched_graph *graph,
406 struct isl_sched_node *src, struct isl_sched_node *dst)
411 for (i = 0; i <= isl_edge_last; ++i) {
412 r = graph_has_edge(graph, i, src, dst);
420 /* Check whether the dependence graph has a validity edge
421 * between the given two nodes.
423 static int graph_has_validity_edge(struct isl_sched_graph *graph,
424 struct isl_sched_node *src, struct isl_sched_node *dst)
426 return graph_has_edge(graph, isl_edge_validity, src, dst);
429 static int graph_alloc(isl_ctx *ctx, struct isl_sched_graph *graph,
430 int n_node, int n_edge)
435 graph->n_edge = n_edge;
436 graph->node = isl_calloc_array(ctx, struct isl_sched_node, graph->n);
437 graph->sorted = isl_calloc_array(ctx, int, graph->n);
438 graph->region = isl_alloc_array(ctx, struct isl_region, graph->n);
439 graph->stack = isl_alloc_array(ctx, int, graph->n);
440 graph->edge = isl_calloc_array(ctx,
441 struct isl_sched_edge, graph->n_edge);
443 graph->intra_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
444 graph->inter_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
446 if (!graph->node || !graph->region || !graph->stack || !graph->edge ||
450 for(i = 0; i < graph->n; ++i)
451 graph->sorted[i] = i;
456 static void graph_free(isl_ctx *ctx, struct isl_sched_graph *graph)
460 isl_hmap_map_basic_set_free(ctx, graph->intra_hmap);
461 isl_hmap_map_basic_set_free(ctx, graph->inter_hmap);
463 for (i = 0; i < graph->n; ++i) {
464 isl_space_free(graph->node[i].dim);
465 isl_mat_free(graph->node[i].sched);
466 isl_map_free(graph->node[i].sched_map);
467 isl_mat_free(graph->node[i].cmap);
469 free(graph->node[i].band);
470 free(graph->node[i].band_id);
471 free(graph->node[i].zero);
476 for (i = 0; i < graph->n_edge; ++i)
477 isl_map_free(graph->edge[i].map);
481 for (i = 0; i <= isl_edge_last; ++i)
482 isl_hash_table_free(ctx, graph->edge_table[i]);
483 isl_hash_table_free(ctx, graph->node_table);
484 isl_basic_set_free(graph->lp);
487 /* Add a new node to the graph representing the given set.
489 static int extract_node(__isl_take isl_set *set, void *user)
495 struct isl_sched_graph *graph = user;
496 int *band, *band_id, *zero;
498 ctx = isl_set_get_ctx(set);
499 dim = isl_set_get_space(set);
501 nvar = isl_space_dim(dim, isl_dim_set);
502 nparam = isl_space_dim(dim, isl_dim_param);
503 if (!ctx->opt->schedule_parametric)
505 sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar);
506 graph->node[graph->n].dim = dim;
507 graph->node[graph->n].nvar = nvar;
508 graph->node[graph->n].nparam = nparam;
509 graph->node[graph->n].sched = sched;
510 graph->node[graph->n].sched_map = NULL;
511 band = isl_alloc_array(ctx, int, graph->n_edge + nvar);
512 graph->node[graph->n].band = band;
513 band_id = isl_calloc_array(ctx, int, graph->n_edge + nvar);
514 graph->node[graph->n].band_id = band_id;
515 zero = isl_calloc_array(ctx, int, graph->n_edge + nvar);
516 graph->node[graph->n].zero = zero;
519 if (!sched || !band || !band_id || !zero)
525 struct isl_extract_edge_data {
526 enum isl_edge_type type;
527 struct isl_sched_graph *graph;
530 /* Add a new edge to the graph based on the given map
531 * and add it to data->graph->edge_table[data->type].
532 * If a dependence relation of a given type happens to be identical
533 * to one of the dependence relations of a type that was added before,
534 * then we don't create a new edge, but instead mark the original edge
535 * as also representing a dependence of the current type.
537 static int extract_edge(__isl_take isl_map *map, void *user)
539 isl_ctx *ctx = isl_map_get_ctx(map);
540 struct isl_extract_edge_data *data = user;
541 struct isl_sched_graph *graph = data->graph;
542 struct isl_sched_node *src, *dst;
544 struct isl_sched_edge *edge;
547 dim = isl_space_domain(isl_map_get_space(map));
548 src = graph_find_node(ctx, graph, dim);
550 dim = isl_space_range(isl_map_get_space(map));
551 dst = graph_find_node(ctx, graph, dim);
559 graph->edge[graph->n_edge].src = src;
560 graph->edge[graph->n_edge].dst = dst;
561 graph->edge[graph->n_edge].map = map;
562 if (data->type == isl_edge_validity) {
563 graph->edge[graph->n_edge].validity = 1;
564 graph->edge[graph->n_edge].proximity = 0;
566 if (data->type == isl_edge_proximity) {
567 graph->edge[graph->n_edge].validity = 0;
568 graph->edge[graph->n_edge].proximity = 1;
572 edge = graph_find_any_edge(graph, src, dst);
574 return graph_edge_table_add(ctx, graph, data->type,
575 &graph->edge[graph->n_edge - 1]);
576 is_equal = isl_map_plain_is_equal(map, edge->map);
580 return graph_edge_table_add(ctx, graph, data->type,
581 &graph->edge[graph->n_edge - 1]);
584 edge->validity |= graph->edge[graph->n_edge].validity;
585 edge->proximity |= graph->edge[graph->n_edge].proximity;
588 return graph_edge_table_add(ctx, graph, data->type, edge);
591 /* Check whether there is a validity dependence from src to dst,
592 * forcing dst to follow src (if weak is not set).
593 * If weak is set, then check if there is any dependence from src to dst.
595 static int node_follows(struct isl_sched_graph *graph,
596 struct isl_sched_node *dst, struct isl_sched_node *src, int weak)
599 return graph_has_any_edge(graph, src, dst);
601 return graph_has_validity_edge(graph, src, dst);
604 /* Perform Tarjan's algorithm for computing the strongly connected components
605 * in the dependence graph (only validity edges).
606 * If weak is set, we consider the graph to be undirected and
607 * we effectively compute the (weakly) connected components.
608 * Additionally, we also consider other edges when weak is set.
610 static int detect_sccs_tarjan(struct isl_sched_graph *g, int i, int weak)
614 g->node[i].index = g->index;
615 g->node[i].min_index = g->index;
616 g->node[i].on_stack = 1;
618 g->stack[g->sp++] = i;
620 for (j = g->n - 1; j >= 0; --j) {
625 if (g->node[j].index >= 0 &&
626 (!g->node[j].on_stack ||
627 g->node[j].index > g->node[i].min_index))
630 f = node_follows(g, &g->node[i], &g->node[j], weak);
634 f = node_follows(g, &g->node[j], &g->node[i], weak);
640 if (g->node[j].index < 0) {
641 detect_sccs_tarjan(g, j, weak);
642 if (g->node[j].min_index < g->node[i].min_index)
643 g->node[i].min_index = g->node[j].min_index;
644 } else if (g->node[j].index < g->node[i].min_index)
645 g->node[i].min_index = g->node[j].index;
648 if (g->node[i].index != g->node[i].min_index)
652 j = g->stack[--g->sp];
653 g->node[j].on_stack = 0;
654 g->node[j].scc = g->scc;
661 static int detect_ccs(struct isl_sched_graph *graph, int weak)
668 for (i = graph->n - 1; i >= 0; --i)
669 graph->node[i].index = -1;
671 for (i = graph->n - 1; i >= 0; --i) {
672 if (graph->node[i].index >= 0)
674 if (detect_sccs_tarjan(graph, i, weak) < 0)
681 /* Apply Tarjan's algorithm to detect the strongly connected components
682 * in the dependence graph.
684 static int detect_sccs(struct isl_sched_graph *graph)
686 return detect_ccs(graph, 0);
689 /* Apply Tarjan's algorithm to detect the (weakly) connected components
690 * in the dependence graph.
692 static int detect_wccs(struct isl_sched_graph *graph)
694 return detect_ccs(graph, 1);
697 static int cmp_scc(const void *a, const void *b, void *data)
699 struct isl_sched_graph *graph = data;
703 return graph->node[*i1].scc - graph->node[*i2].scc;
706 /* Sort the elements of graph->sorted according to the corresponding SCCs.
708 static void sort_sccs(struct isl_sched_graph *graph)
710 isl_quicksort(graph->sorted, graph->n, sizeof(int), &cmp_scc, graph);
713 /* Given a dependence relation R from a node to itself,
714 * construct the set of coefficients of valid constraints for elements
715 * in that dependence relation.
716 * In particular, the result contains tuples of coefficients
717 * c_0, c_n, c_x such that
719 * c_0 + c_n n + c_x y - c_x x >= 0 for each (x,y) in R
723 * c_0 + c_n n + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R }
725 * We choose here to compute the dual of delta R.
726 * Alternatively, we could have computed the dual of R, resulting
727 * in a set of tuples c_0, c_n, c_x, c_y, and then
728 * plugged in (c_0, c_n, c_x, -c_x).
730 static __isl_give isl_basic_set *intra_coefficients(
731 struct isl_sched_graph *graph, __isl_take isl_map *map)
733 isl_ctx *ctx = isl_map_get_ctx(map);
737 if (isl_hmap_map_basic_set_has(ctx, graph->intra_hmap, map))
738 return isl_hmap_map_basic_set_get(ctx, graph->intra_hmap, map);
740 delta = isl_set_remove_divs(isl_map_deltas(isl_map_copy(map)));
741 coef = isl_set_coefficients(delta);
742 isl_hmap_map_basic_set_set(ctx, graph->intra_hmap, map,
743 isl_basic_set_copy(coef));
748 /* Given a dependence relation R, * construct the set of coefficients
749 * of valid constraints for elements in that dependence relation.
750 * In particular, the result contains tuples of coefficients
751 * c_0, c_n, c_x, c_y such that
753 * c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R
756 static __isl_give isl_basic_set *inter_coefficients(
757 struct isl_sched_graph *graph, __isl_take isl_map *map)
759 isl_ctx *ctx = isl_map_get_ctx(map);
763 if (isl_hmap_map_basic_set_has(ctx, graph->inter_hmap, map))
764 return isl_hmap_map_basic_set_get(ctx, graph->inter_hmap, map);
766 set = isl_map_wrap(isl_map_remove_divs(isl_map_copy(map)));
767 coef = isl_set_coefficients(set);
768 isl_hmap_map_basic_set_set(ctx, graph->inter_hmap, map,
769 isl_basic_set_copy(coef));
774 /* Add constraints to graph->lp that force validity for the given
775 * dependence from a node i to itself.
776 * That is, add constraints that enforce
778 * (c_i_0 + c_i_n n + c_i_x y) - (c_i_0 + c_i_n n + c_i_x x)
779 * = c_i_x (y - x) >= 0
781 * for each (x,y) in R.
782 * We obtain general constraints on coefficients (c_0, c_n, c_x)
783 * of valid constraints for (y - x) and then plug in (0, 0, c_i_x^+ - c_i_x^-),
784 * where c_i_x = c_i_x^+ - c_i_x^-, with c_i_x^+ and c_i_x^- non-negative.
785 * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart.
787 * Actually, we do not construct constraints for the c_i_x themselves,
788 * but for the coefficients of c_i_x written as a linear combination
789 * of the columns in node->cmap.
791 static int add_intra_validity_constraints(struct isl_sched_graph *graph,
792 struct isl_sched_edge *edge)
795 isl_map *map = isl_map_copy(edge->map);
796 isl_ctx *ctx = isl_map_get_ctx(map);
798 isl_dim_map *dim_map;
800 struct isl_sched_node *node = edge->src;
802 coef = intra_coefficients(graph, map);
804 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
806 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
807 isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap));
809 total = isl_basic_set_total_dim(graph->lp);
810 dim_map = isl_dim_map_alloc(ctx, total);
811 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
812 isl_space_dim(dim, isl_dim_set), 1,
814 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
815 isl_space_dim(dim, isl_dim_set), 1,
817 graph->lp = isl_basic_set_extend_constraints(graph->lp,
818 coef->n_eq, coef->n_ineq);
819 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
826 /* Add constraints to graph->lp that force validity for the given
827 * dependence from node i to node j.
828 * That is, add constraints that enforce
830 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) >= 0
832 * for each (x,y) in R.
833 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
834 * of valid constraints for R and then plug in
835 * (c_j_0 - c_i_0, c_j_n^+ - c_j_n^- - (c_i_n^+ - c_i_n^-),
836 * c_j_x^+ - c_j_x^- - (c_i_x^+ - c_i_x^-)),
837 * where c_* = c_*^+ - c_*^-, with c_*^+ and c_*^- non-negative.
838 * In graph->lp, the c_*^- appear before their c_*^+ counterpart.
840 * Actually, we do not construct constraints for the c_*_x themselves,
841 * but for the coefficients of c_*_x written as a linear combination
842 * of the columns in node->cmap.
844 static int add_inter_validity_constraints(struct isl_sched_graph *graph,
845 struct isl_sched_edge *edge)
848 isl_map *map = isl_map_copy(edge->map);
849 isl_ctx *ctx = isl_map_get_ctx(map);
851 isl_dim_map *dim_map;
853 struct isl_sched_node *src = edge->src;
854 struct isl_sched_node *dst = edge->dst;
856 coef = inter_coefficients(graph, map);
858 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
860 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
861 isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap));
862 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
863 isl_space_dim(dim, isl_dim_set) + src->nvar,
864 isl_mat_copy(dst->cmap));
866 total = isl_basic_set_total_dim(graph->lp);
867 dim_map = isl_dim_map_alloc(ctx, total);
869 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
870 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
871 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
872 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
873 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
875 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
876 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
879 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
880 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
881 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
882 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
883 isl_space_dim(dim, isl_dim_set), 1,
885 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
886 isl_space_dim(dim, isl_dim_set), 1,
889 edge->start = graph->lp->n_ineq;
890 graph->lp = isl_basic_set_extend_constraints(graph->lp,
891 coef->n_eq, coef->n_ineq);
892 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
895 edge->end = graph->lp->n_ineq;
900 /* Add constraints to graph->lp that bound the dependence distance for the given
901 * dependence from a node i to itself.
902 * If s = 1, we add the constraint
904 * c_i_x (y - x) <= m_0 + m_n n
908 * -c_i_x (y - x) + m_0 + m_n n >= 0
910 * for each (x,y) in R.
911 * If s = -1, we add the constraint
913 * -c_i_x (y - x) <= m_0 + m_n n
917 * c_i_x (y - x) + m_0 + m_n n >= 0
919 * for each (x,y) in R.
920 * We obtain general constraints on coefficients (c_0, c_n, c_x)
921 * of valid constraints for (y - x) and then plug in (m_0, m_n, -s * c_i_x),
922 * with each coefficient (except m_0) represented as a pair of non-negative
925 * Actually, we do not construct constraints for the c_i_x themselves,
926 * but for the coefficients of c_i_x written as a linear combination
927 * of the columns in node->cmap.
929 static int add_intra_proximity_constraints(struct isl_sched_graph *graph,
930 struct isl_sched_edge *edge, int s)
934 isl_map *map = isl_map_copy(edge->map);
935 isl_ctx *ctx = isl_map_get_ctx(map);
937 isl_dim_map *dim_map;
939 struct isl_sched_node *node = edge->src;
941 coef = intra_coefficients(graph, map);
943 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
945 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
946 isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap));
948 nparam = isl_space_dim(node->dim, isl_dim_param);
949 total = isl_basic_set_total_dim(graph->lp);
950 dim_map = isl_dim_map_alloc(ctx, total);
951 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
952 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
953 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
954 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
955 isl_space_dim(dim, isl_dim_set), 1,
957 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
958 isl_space_dim(dim, isl_dim_set), 1,
960 graph->lp = isl_basic_set_extend_constraints(graph->lp,
961 coef->n_eq, coef->n_ineq);
962 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
969 /* Add constraints to graph->lp that bound the dependence distance for the given
970 * dependence from node i to node j.
971 * If s = 1, we add the constraint
973 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)
978 * -(c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x) +
981 * for each (x,y) in R.
982 * If s = -1, we add the constraint
984 * -((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x))
989 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) +
992 * for each (x,y) in R.
993 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
994 * of valid constraints for R and then plug in
995 * (m_0 - s*c_j_0 + s*c_i_0, m_n - s*c_j_n + s*c_i_n,
997 * with each coefficient (except m_0, c_j_0 and c_i_0)
998 * represented as a pair of non-negative coefficients.
1000 * Actually, we do not construct constraints for the c_*_x themselves,
1001 * but for the coefficients of c_*_x written as a linear combination
1002 * of the columns in node->cmap.
1004 static int add_inter_proximity_constraints(struct isl_sched_graph *graph,
1005 struct isl_sched_edge *edge, int s)
1009 isl_map *map = isl_map_copy(edge->map);
1010 isl_ctx *ctx = isl_map_get_ctx(map);
1012 isl_dim_map *dim_map;
1013 isl_basic_set *coef;
1014 struct isl_sched_node *src = edge->src;
1015 struct isl_sched_node *dst = edge->dst;
1017 coef = inter_coefficients(graph, map);
1019 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
1021 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
1022 isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap));
1023 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
1024 isl_space_dim(dim, isl_dim_set) + src->nvar,
1025 isl_mat_copy(dst->cmap));
1027 nparam = isl_space_dim(src->dim, isl_dim_param);
1028 total = isl_basic_set_total_dim(graph->lp);
1029 dim_map = isl_dim_map_alloc(ctx, total);
1031 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
1032 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
1033 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
1035 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, -s);
1036 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, s);
1037 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, -s);
1038 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
1039 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
1041 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
1042 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
1045 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, s);
1046 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, -s);
1047 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, s);
1048 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
1049 isl_space_dim(dim, isl_dim_set), 1,
1051 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
1052 isl_space_dim(dim, isl_dim_set), 1,
1055 graph->lp = isl_basic_set_extend_constraints(graph->lp,
1056 coef->n_eq, coef->n_ineq);
1057 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
1059 isl_space_free(dim);
1064 static int add_all_validity_constraints(struct isl_sched_graph *graph)
1068 for (i = 0; i < graph->n_edge; ++i) {
1069 struct isl_sched_edge *edge= &graph->edge[i];
1070 if (!edge->validity)
1072 if (edge->src != edge->dst)
1074 if (add_intra_validity_constraints(graph, edge) < 0)
1078 for (i = 0; i < graph->n_edge; ++i) {
1079 struct isl_sched_edge *edge = &graph->edge[i];
1080 if (!edge->validity)
1082 if (edge->src == edge->dst)
1084 if (add_inter_validity_constraints(graph, edge) < 0)
1091 /* Add constraints to graph->lp that bound the dependence distance
1092 * for all dependence relations.
1093 * If a given proximity dependence is identical to a validity
1094 * dependence, then the dependence distance is already bounded
1095 * from below (by zero), so we only need to bound the distance
1097 * Otherwise, we need to bound the distance both from above and from below.
1099 static int add_all_proximity_constraints(struct isl_sched_graph *graph)
1103 for (i = 0; i < graph->n_edge; ++i) {
1104 struct isl_sched_edge *edge= &graph->edge[i];
1105 if (!edge->proximity)
1107 if (edge->src == edge->dst &&
1108 add_intra_proximity_constraints(graph, edge, 1) < 0)
1110 if (edge->src != edge->dst &&
1111 add_inter_proximity_constraints(graph, edge, 1) < 0)
1115 if (edge->src == edge->dst &&
1116 add_intra_proximity_constraints(graph, edge, -1) < 0)
1118 if (edge->src != edge->dst &&
1119 add_inter_proximity_constraints(graph, edge, -1) < 0)
1126 /* Compute a basis for the rows in the linear part of the schedule
1127 * and extend this basis to a full basis. The remaining rows
1128 * can then be used to force linear independence from the rows
1131 * In particular, given the schedule rows S, we compute
1135 * with H the Hermite normal form of S. That is, all but the
1136 * first rank columns of Q are zero and so each row in S is
1137 * a linear combination of the first rank rows of Q.
1138 * The matrix Q is then transposed because we will write the
1139 * coefficients of the next schedule row as a column vector s
1140 * and express this s as a linear combination s = Q c of the
1143 static int node_update_cmap(struct isl_sched_node *node)
1146 int n_row = isl_mat_rows(node->sched);
1148 H = isl_mat_sub_alloc(node->sched, 0, n_row,
1149 1 + node->nparam, node->nvar);
1151 H = isl_mat_left_hermite(H, 0, NULL, &Q);
1152 isl_mat_free(node->cmap);
1153 node->cmap = isl_mat_transpose(Q);
1154 node->rank = isl_mat_initial_non_zero_cols(H);
1157 if (!node->cmap || node->rank < 0)
1162 /* Count the number of equality and inequality constraints
1163 * that will be added for the given map.
1164 * If carry is set, then we are counting the number of (validity)
1165 * constraints that will be added in setup_carry_lp and we count
1166 * each edge exactly once. Otherwise, we count as follows
1167 * validity -> 1 (>= 0)
1168 * validity+proximity -> 2 (>= 0 and upper bound)
1169 * proximity -> 2 (lower and upper bound)
1171 static int count_map_constraints(struct isl_sched_graph *graph,
1172 struct isl_sched_edge *edge, __isl_take isl_map *map,
1173 int *n_eq, int *n_ineq, int carry)
1175 isl_basic_set *coef;
1176 int f = carry ? 1 : edge->proximity ? 2 : 1;
1178 if (carry && !edge->validity) {
1183 if (edge->src == edge->dst)
1184 coef = intra_coefficients(graph, map);
1186 coef = inter_coefficients(graph, map);
1189 *n_eq += f * coef->n_eq;
1190 *n_ineq += f * coef->n_ineq;
1191 isl_basic_set_free(coef);
1196 /* Count the number of equality and inequality constraints
1197 * that will be added to the main lp problem.
1198 * We count as follows
1199 * validity -> 1 (>= 0)
1200 * validity+proximity -> 2 (>= 0 and upper bound)
1201 * proximity -> 2 (lower and upper bound)
1203 static int count_constraints(struct isl_sched_graph *graph,
1204 int *n_eq, int *n_ineq)
1208 *n_eq = *n_ineq = 0;
1209 for (i = 0; i < graph->n_edge; ++i) {
1210 struct isl_sched_edge *edge= &graph->edge[i];
1211 isl_map *map = isl_map_copy(edge->map);
1213 if (count_map_constraints(graph, edge, map,
1214 n_eq, n_ineq, 0) < 0)
1221 /* Add constraints that bound the values of the variable and parameter
1222 * coefficients of the schedule.
1224 * The maximal value of the coefficients is defined by the option
1225 * 'schedule_max_coefficient'.
1227 static int add_bound_coefficient_constraints(isl_ctx *ctx,
1228 struct isl_sched_graph *graph)
1231 int max_coefficient;
1234 max_coefficient = ctx->opt->schedule_max_coefficient;
1236 if (max_coefficient == -1)
1239 total = isl_basic_set_total_dim(graph->lp);
1241 for (i = 0; i < graph->n; ++i) {
1242 struct isl_sched_node *node = &graph->node[i];
1243 for (j = 0; j < 2 * node->nparam + 2 * node->nvar; ++j) {
1245 k = isl_basic_set_alloc_inequality(graph->lp);
1248 dim = 1 + node->start + 1 + j;
1249 isl_seq_clr(graph->lp->ineq[k], 1 + total);
1250 isl_int_set_si(graph->lp->ineq[k][dim], -1);
1251 isl_int_set_si(graph->lp->ineq[k][0], max_coefficient);
1258 /* Construct an ILP problem for finding schedule coefficients
1259 * that result in non-negative, but small dependence distances
1260 * over all dependences.
1261 * In particular, the dependence distances over proximity edges
1262 * are bounded by m_0 + m_n n and we compute schedule coefficients
1263 * with small values (preferably zero) of m_n and m_0.
1265 * All variables of the ILP are non-negative. The actual coefficients
1266 * may be negative, so each coefficient is represented as the difference
1267 * of two non-negative variables. The negative part always appears
1268 * immediately before the positive part.
1269 * Other than that, the variables have the following order
1271 * - sum of positive and negative parts of m_n coefficients
1273 * - sum of positive and negative parts of all c_n coefficients
1274 * (unconstrained when computing non-parametric schedules)
1275 * - sum of positive and negative parts of all c_x coefficients
1276 * - positive and negative parts of m_n coefficients
1279 * - positive and negative parts of c_i_n (if parametric)
1280 * - positive and negative parts of c_i_x
1282 * The c_i_x are not represented directly, but through the columns of
1283 * node->cmap. That is, the computed values are for variable t_i_x
1284 * such that c_i_x = Q t_i_x with Q equal to node->cmap.
1286 * The constraints are those from the edges plus two or three equalities
1287 * to express the sums.
1289 * If force_zero is set, then we add equalities to ensure that
1290 * the sum of the m_n coefficients and m_0 are both zero.
1292 static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph,
1303 int max_constant_term;
1304 int max_coefficient;
1306 max_constant_term = ctx->opt->schedule_max_constant_term;
1307 max_coefficient = ctx->opt->schedule_max_coefficient;
1309 parametric = ctx->opt->schedule_parametric;
1310 nparam = isl_space_dim(graph->node[0].dim, isl_dim_param);
1312 total = param_pos + 2 * nparam;
1313 for (i = 0; i < graph->n; ++i) {
1314 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
1315 if (node_update_cmap(node) < 0)
1317 node->start = total;
1318 total += 1 + 2 * (node->nparam + node->nvar);
1321 if (count_constraints(graph, &n_eq, &n_ineq) < 0)
1324 dim = isl_space_set_alloc(ctx, 0, total);
1325 isl_basic_set_free(graph->lp);
1326 n_eq += 2 + parametric + force_zero;
1327 if (max_constant_term != -1)
1329 if (max_coefficient != -1)
1330 for (i = 0; i < graph->n; ++i)
1331 n_ineq += 2 * graph->node[i].nparam +
1332 2 * graph->node[i].nvar;
1334 graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq);
1336 k = isl_basic_set_alloc_equality(graph->lp);
1339 isl_seq_clr(graph->lp->eq[k], 1 + total);
1341 isl_int_set_si(graph->lp->eq[k][1], -1);
1342 for (i = 0; i < 2 * nparam; ++i)
1343 isl_int_set_si(graph->lp->eq[k][1 + param_pos + i], 1);
1346 k = isl_basic_set_alloc_equality(graph->lp);
1349 isl_seq_clr(graph->lp->eq[k], 1 + total);
1350 isl_int_set_si(graph->lp->eq[k][2], -1);
1354 k = isl_basic_set_alloc_equality(graph->lp);
1357 isl_seq_clr(graph->lp->eq[k], 1 + total);
1358 isl_int_set_si(graph->lp->eq[k][3], -1);
1359 for (i = 0; i < graph->n; ++i) {
1360 int pos = 1 + graph->node[i].start + 1;
1362 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
1363 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1367 k = isl_basic_set_alloc_equality(graph->lp);
1370 isl_seq_clr(graph->lp->eq[k], 1 + total);
1371 isl_int_set_si(graph->lp->eq[k][4], -1);
1372 for (i = 0; i < graph->n; ++i) {
1373 struct isl_sched_node *node = &graph->node[i];
1374 int pos = 1 + node->start + 1 + 2 * node->nparam;
1376 for (j = 0; j < 2 * node->nvar; ++j)
1377 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1380 if (max_constant_term != -1)
1381 for (i = 0; i < graph->n; ++i) {
1382 struct isl_sched_node *node = &graph->node[i];
1383 k = isl_basic_set_alloc_inequality(graph->lp);
1386 isl_seq_clr(graph->lp->ineq[k], 1 + total);
1387 isl_int_set_si(graph->lp->ineq[k][1 + node->start], -1);
1388 isl_int_set_si(graph->lp->ineq[k][0], max_constant_term);
1391 if (add_bound_coefficient_constraints(ctx, graph) < 0)
1393 if (add_all_validity_constraints(graph) < 0)
1395 if (add_all_proximity_constraints(graph) < 0)
1401 /* Analyze the conflicting constraint found by
1402 * isl_tab_basic_set_non_trivial_lexmin. If it corresponds to the validity
1403 * constraint of one of the edges between distinct nodes, living, moreover
1404 * in distinct SCCs, then record the source and sink SCC as this may
1405 * be a good place to cut between SCCs.
1407 static int check_conflict(int con, void *user)
1410 struct isl_sched_graph *graph = user;
1412 if (graph->src_scc >= 0)
1415 con -= graph->lp->n_eq;
1417 if (con >= graph->lp->n_ineq)
1420 for (i = 0; i < graph->n_edge; ++i) {
1421 if (!graph->edge[i].validity)
1423 if (graph->edge[i].src == graph->edge[i].dst)
1425 if (graph->edge[i].src->scc == graph->edge[i].dst->scc)
1427 if (graph->edge[i].start > con)
1429 if (graph->edge[i].end <= con)
1431 graph->src_scc = graph->edge[i].src->scc;
1432 graph->dst_scc = graph->edge[i].dst->scc;
1438 /* Check whether the next schedule row of the given node needs to be
1439 * non-trivial. Lower-dimensional domains may have some trivial rows,
1440 * but as soon as the number of remaining required non-trivial rows
1441 * is as large as the number or remaining rows to be computed,
1442 * all remaining rows need to be non-trivial.
1444 static int needs_row(struct isl_sched_graph *graph, struct isl_sched_node *node)
1446 return node->nvar - node->rank >= graph->maxvar - graph->n_row;
1449 /* Solve the ILP problem constructed in setup_lp.
1450 * For each node such that all the remaining rows of its schedule
1451 * need to be non-trivial, we construct a non-triviality region.
1452 * This region imposes that the next row is independent of previous rows.
1453 * In particular the coefficients c_i_x are represented by t_i_x
1454 * variables with c_i_x = Q t_i_x and Q a unimodular matrix such that
1455 * its first columns span the rows of the previously computed part
1456 * of the schedule. The non-triviality region enforces that at least
1457 * one of the remaining components of t_i_x is non-zero, i.e.,
1458 * that the new schedule row depends on at least one of the remaining
1461 static __isl_give isl_vec *solve_lp(struct isl_sched_graph *graph)
1467 for (i = 0; i < graph->n; ++i) {
1468 struct isl_sched_node *node = &graph->node[i];
1469 int skip = node->rank;
1470 graph->region[i].pos = node->start + 1 + 2*(node->nparam+skip);
1471 if (needs_row(graph, node))
1472 graph->region[i].len = 2 * (node->nvar - skip);
1474 graph->region[i].len = 0;
1476 lp = isl_basic_set_copy(graph->lp);
1477 sol = isl_tab_basic_set_non_trivial_lexmin(lp, 2, graph->n,
1478 graph->region, &check_conflict, graph);
1482 /* Update the schedules of all nodes based on the given solution
1483 * of the LP problem.
1484 * The new row is added to the current band.
1485 * All possibly negative coefficients are encoded as a difference
1486 * of two non-negative variables, so we need to perform the subtraction
1487 * here. Moreover, if use_cmap is set, then the solution does
1488 * not refer to the actual coefficients c_i_x, but instead to variables
1489 * t_i_x such that c_i_x = Q t_i_x and Q is equal to node->cmap.
1490 * In this case, we then also need to perform this multiplication
1491 * to obtain the values of c_i_x.
1493 * If check_zero is set, then the first two coordinates of sol are
1494 * assumed to correspond to the dependence distance. If these two
1495 * coordinates are zero, then the corresponding scheduling dimension
1496 * is marked as being zero distance.
1498 static int update_schedule(struct isl_sched_graph *graph,
1499 __isl_take isl_vec *sol, int use_cmap, int check_zero)
1503 isl_vec *csol = NULL;
1508 isl_die(sol->ctx, isl_error_internal,
1509 "no solution found", goto error);
1512 zero = isl_int_is_zero(sol->el[1]) &&
1513 isl_int_is_zero(sol->el[2]);
1515 for (i = 0; i < graph->n; ++i) {
1516 struct isl_sched_node *node = &graph->node[i];
1517 int pos = node->start;
1518 int row = isl_mat_rows(node->sched);
1521 csol = isl_vec_alloc(sol->ctx, node->nvar);
1525 isl_map_free(node->sched_map);
1526 node->sched_map = NULL;
1527 node->sched = isl_mat_add_rows(node->sched, 1);
1530 node->sched = isl_mat_set_element(node->sched, row, 0,
1532 for (j = 0; j < node->nparam + node->nvar; ++j)
1533 isl_int_sub(sol->el[1 + pos + 1 + 2 * j + 1],
1534 sol->el[1 + pos + 1 + 2 * j + 1],
1535 sol->el[1 + pos + 1 + 2 * j]);
1536 for (j = 0; j < node->nparam; ++j)
1537 node->sched = isl_mat_set_element(node->sched,
1538 row, 1 + j, sol->el[1+pos+1+2*j+1]);
1539 for (j = 0; j < node->nvar; ++j)
1540 isl_int_set(csol->el[j],
1541 sol->el[1+pos+1+2*(node->nparam+j)+1]);
1543 csol = isl_mat_vec_product(isl_mat_copy(node->cmap),
1547 for (j = 0; j < node->nvar; ++j)
1548 node->sched = isl_mat_set_element(node->sched,
1549 row, 1 + node->nparam + j, csol->el[j]);
1550 node->band[graph->n_total_row] = graph->n_band;
1551 node->zero[graph->n_total_row] = zero;
1557 graph->n_total_row++;
1566 /* Convert node->sched into a map and return this map.
1567 * We simply add equality constraints that express each output variable
1568 * as the affine combination of parameters and input variables specified
1569 * by the schedule matrix.
1571 * The result is cached in node->sched_map, which needs to be released
1572 * whenever node->sched is updated.
1574 static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node)
1578 isl_local_space *ls;
1579 isl_basic_map *bmap;
1584 if (node->sched_map)
1585 return isl_map_copy(node->sched_map);
1587 nrow = isl_mat_rows(node->sched);
1588 ncol = isl_mat_cols(node->sched) - 1;
1589 dim = isl_space_from_domain(isl_space_copy(node->dim));
1590 dim = isl_space_add_dims(dim, isl_dim_out, nrow);
1591 bmap = isl_basic_map_universe(isl_space_copy(dim));
1592 ls = isl_local_space_from_space(dim);
1596 for (i = 0; i < nrow; ++i) {
1597 c = isl_equality_alloc(isl_local_space_copy(ls));
1598 isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
1599 isl_mat_get_element(node->sched, i, 0, &v);
1600 isl_constraint_set_constant(c, v);
1601 for (j = 0; j < node->nparam; ++j) {
1602 isl_mat_get_element(node->sched, i, 1 + j, &v);
1603 isl_constraint_set_coefficient(c, isl_dim_param, j, v);
1605 for (j = 0; j < node->nvar; ++j) {
1606 isl_mat_get_element(node->sched,
1607 i, 1 + node->nparam + j, &v);
1608 isl_constraint_set_coefficient(c, isl_dim_in, j, v);
1610 bmap = isl_basic_map_add_constraint(bmap, c);
1615 isl_local_space_free(ls);
1617 node->sched_map = isl_map_from_basic_map(bmap);
1618 return isl_map_copy(node->sched_map);
1621 /* Update the given dependence relation based on the current schedule.
1622 * That is, intersect the dependence relation with a map expressing
1623 * that source and sink are executed within the same iteration of
1624 * the current schedule.
1625 * This is not the most efficient way, but this shouldn't be a critical
1628 static __isl_give isl_map *specialize(__isl_take isl_map *map,
1629 struct isl_sched_node *src, struct isl_sched_node *dst)
1631 isl_map *src_sched, *dst_sched, *id;
1633 src_sched = node_extract_schedule(src);
1634 dst_sched = node_extract_schedule(dst);
1635 id = isl_map_apply_range(src_sched, isl_map_reverse(dst_sched));
1636 return isl_map_intersect(map, id);
1639 /* Update the dependence relations of all edges based on the current schedule.
1640 * If a dependence is carried completely by the current schedule, then
1641 * it is removed from the edge_tables. It is kept in the list of edges
1642 * as otherwise all edge_tables would have to be recomputed.
1644 static int update_edges(isl_ctx *ctx, struct isl_sched_graph *graph)
1648 for (i = graph->n_edge - 1; i >= 0; --i) {
1649 struct isl_sched_edge *edge = &graph->edge[i];
1650 edge->map = specialize(edge->map, edge->src, edge->dst);
1654 if (isl_map_plain_is_empty(edge->map))
1655 graph_remove_edge(graph, edge);
1661 static void next_band(struct isl_sched_graph *graph)
1663 graph->band_start = graph->n_total_row;
1667 /* Topologically sort statements mapped to the same schedule iteration
1668 * and add a row to the schedule corresponding to this order.
1670 static int sort_statements(isl_ctx *ctx, struct isl_sched_graph *graph)
1677 if (update_edges(ctx, graph) < 0)
1680 if (graph->n_edge == 0)
1683 if (detect_sccs(graph) < 0)
1686 for (i = 0; i < graph->n; ++i) {
1687 struct isl_sched_node *node = &graph->node[i];
1688 int row = isl_mat_rows(node->sched);
1689 int cols = isl_mat_cols(node->sched);
1691 isl_map_free(node->sched_map);
1692 node->sched_map = NULL;
1693 node->sched = isl_mat_add_rows(node->sched, 1);
1696 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1698 for (j = 1; j < cols; ++j)
1699 node->sched = isl_mat_set_element_si(node->sched,
1701 node->band[graph->n_total_row] = graph->n_band;
1704 graph->n_total_row++;
1710 /* Construct an isl_schedule based on the computed schedule stored
1711 * in graph and with parameters specified by dim.
1713 static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph,
1714 __isl_take isl_space *dim)
1718 isl_schedule *sched = NULL;
1723 ctx = isl_space_get_ctx(dim);
1724 sched = isl_calloc(ctx, struct isl_schedule,
1725 sizeof(struct isl_schedule) +
1726 (graph->n - 1) * sizeof(struct isl_schedule_node));
1731 sched->n = graph->n;
1732 sched->n_band = graph->n_band;
1733 sched->n_total_row = graph->n_total_row;
1735 for (i = 0; i < sched->n; ++i) {
1737 int *band_end, *band_id, *zero;
1739 band_end = isl_alloc_array(ctx, int, graph->n_band);
1740 band_id = isl_alloc_array(ctx, int, graph->n_band);
1741 zero = isl_alloc_array(ctx, int, graph->n_total_row);
1742 sched->node[i].sched = node_extract_schedule(&graph->node[i]);
1743 sched->node[i].band_end = band_end;
1744 sched->node[i].band_id = band_id;
1745 sched->node[i].zero = zero;
1746 if (!band_end || !band_id || !zero)
1749 for (r = 0; r < graph->n_total_row; ++r)
1750 zero[r] = graph->node[i].zero[r];
1751 for (r = b = 0; r < graph->n_total_row; ++r) {
1752 if (graph->node[i].band[r] == b)
1755 if (graph->node[i].band[r] == -1)
1758 if (r == graph->n_total_row)
1760 sched->node[i].n_band = b;
1761 for (--b; b >= 0; --b)
1762 band_id[b] = graph->node[i].band_id[b];
1769 isl_space_free(dim);
1770 isl_schedule_free(sched);
1774 /* Copy nodes that satisfy node_pred from the src dependence graph
1775 * to the dst dependence graph.
1777 static int copy_nodes(struct isl_sched_graph *dst, struct isl_sched_graph *src,
1778 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1783 for (i = 0; i < src->n; ++i) {
1784 if (!node_pred(&src->node[i], data))
1786 dst->node[dst->n].dim = isl_space_copy(src->node[i].dim);
1787 dst->node[dst->n].nvar = src->node[i].nvar;
1788 dst->node[dst->n].nparam = src->node[i].nparam;
1789 dst->node[dst->n].sched = isl_mat_copy(src->node[i].sched);
1790 dst->node[dst->n].sched_map =
1791 isl_map_copy(src->node[i].sched_map);
1792 dst->node[dst->n].band = src->node[i].band;
1793 dst->node[dst->n].band_id = src->node[i].band_id;
1794 dst->node[dst->n].zero = src->node[i].zero;
1801 /* Copy non-empty edges that satisfy edge_pred from the src dependence graph
1802 * to the dst dependence graph.
1803 * If the source or destination node of the edge is not in the destination
1804 * graph, then it must be a backward proximity edge and it should simply
1807 static int copy_edges(isl_ctx *ctx, struct isl_sched_graph *dst,
1808 struct isl_sched_graph *src,
1809 int (*edge_pred)(struct isl_sched_edge *edge, int data), int data)
1815 for (i = 0; i < src->n_edge; ++i) {
1816 struct isl_sched_edge *edge = &src->edge[i];
1818 struct isl_sched_node *dst_src, *dst_dst;
1820 if (!edge_pred(edge, data))
1823 if (isl_map_plain_is_empty(edge->map))
1826 dst_src = graph_find_node(ctx, dst, edge->src->dim);
1827 dst_dst = graph_find_node(ctx, dst, edge->dst->dim);
1828 if (!dst_src || !dst_dst) {
1830 isl_die(ctx, isl_error_internal,
1831 "backward validity edge", return -1);
1835 map = isl_map_copy(edge->map);
1837 dst->edge[dst->n_edge].src = dst_src;
1838 dst->edge[dst->n_edge].dst = dst_dst;
1839 dst->edge[dst->n_edge].map = map;
1840 dst->edge[dst->n_edge].validity = edge->validity;
1841 dst->edge[dst->n_edge].proximity = edge->proximity;
1844 for (t = 0; t <= isl_edge_last; ++t) {
1846 graph_find_edge(src, t, edge->src, edge->dst))
1848 if (graph_edge_table_add(ctx, dst, t,
1849 &dst->edge[dst->n_edge - 1]) < 0)
1857 /* Given a "src" dependence graph that contains the nodes from "dst"
1858 * that satisfy node_pred, copy the schedule computed in "src"
1859 * for those nodes back to "dst".
1861 static int copy_schedule(struct isl_sched_graph *dst,
1862 struct isl_sched_graph *src,
1863 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1868 for (i = 0; i < dst->n; ++i) {
1869 if (!node_pred(&dst->node[i], data))
1871 isl_mat_free(dst->node[i].sched);
1872 isl_map_free(dst->node[i].sched_map);
1873 dst->node[i].sched = isl_mat_copy(src->node[src->n].sched);
1874 dst->node[i].sched_map =
1875 isl_map_copy(src->node[src->n].sched_map);
1879 dst->n_total_row = src->n_total_row;
1880 dst->n_band = src->n_band;
1885 /* Compute the maximal number of variables over all nodes.
1886 * This is the maximal number of linearly independent schedule
1887 * rows that we need to compute.
1888 * Just in case we end up in a part of the dependence graph
1889 * with only lower-dimensional domains, we make sure we will
1890 * compute the required amount of extra linearly independent rows.
1892 static int compute_maxvar(struct isl_sched_graph *graph)
1897 for (i = 0; i < graph->n; ++i) {
1898 struct isl_sched_node *node = &graph->node[i];
1901 if (node_update_cmap(node) < 0)
1903 nvar = node->nvar + graph->n_row - node->rank;
1904 if (nvar > graph->maxvar)
1905 graph->maxvar = nvar;
1911 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph);
1912 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph);
1914 /* Compute a schedule for a subgraph of "graph". In particular, for
1915 * the graph composed of nodes that satisfy node_pred and edges that
1916 * that satisfy edge_pred. The caller should precompute the number
1917 * of nodes and edges that satisfy these predicates and pass them along
1918 * as "n" and "n_edge".
1919 * If the subgraph is known to consist of a single component, then wcc should
1920 * be set and then we call compute_schedule_wcc on the constructed subgraph.
1921 * Otherwise, we call compute_schedule, which will check whether the subgraph
1924 static int compute_sub_schedule(isl_ctx *ctx,
1925 struct isl_sched_graph *graph, int n, int n_edge,
1926 int (*node_pred)(struct isl_sched_node *node, int data),
1927 int (*edge_pred)(struct isl_sched_edge *edge, int data),
1930 struct isl_sched_graph split = { 0 };
1933 if (graph_alloc(ctx, &split, n, n_edge) < 0)
1935 if (copy_nodes(&split, graph, node_pred, data) < 0)
1937 if (graph_init_table(ctx, &split) < 0)
1939 for (t = 0; t <= isl_edge_last; ++t)
1940 split.max_edge[t] = graph->max_edge[t];
1941 if (graph_init_edge_tables(ctx, &split) < 0)
1943 if (copy_edges(ctx, &split, graph, edge_pred, data) < 0)
1945 split.n_row = graph->n_row;
1946 split.n_total_row = graph->n_total_row;
1947 split.n_band = graph->n_band;
1948 split.band_start = graph->band_start;
1950 if (wcc && compute_schedule_wcc(ctx, &split) < 0)
1952 if (!wcc && compute_schedule(ctx, &split) < 0)
1955 copy_schedule(graph, &split, node_pred, data);
1957 graph_free(ctx, &split);
1960 graph_free(ctx, &split);
1964 static int node_scc_exactly(struct isl_sched_node *node, int scc)
1966 return node->scc == scc;
1969 static int node_scc_at_most(struct isl_sched_node *node, int scc)
1971 return node->scc <= scc;
1974 static int node_scc_at_least(struct isl_sched_node *node, int scc)
1976 return node->scc >= scc;
1979 static int edge_scc_exactly(struct isl_sched_edge *edge, int scc)
1981 return edge->src->scc == scc && edge->dst->scc == scc;
1984 static int edge_dst_scc_at_most(struct isl_sched_edge *edge, int scc)
1986 return edge->dst->scc <= scc;
1989 static int edge_src_scc_at_least(struct isl_sched_edge *edge, int scc)
1991 return edge->src->scc >= scc;
1994 /* Pad the schedules of all nodes with zero rows such that in the end
1995 * they all have graph->n_total_row rows.
1996 * The extra rows don't belong to any band, so they get assigned band number -1.
1998 static int pad_schedule(struct isl_sched_graph *graph)
2002 for (i = 0; i < graph->n; ++i) {
2003 struct isl_sched_node *node = &graph->node[i];
2004 int row = isl_mat_rows(node->sched);
2005 if (graph->n_total_row > row) {
2006 isl_map_free(node->sched_map);
2007 node->sched_map = NULL;
2009 node->sched = isl_mat_add_zero_rows(node->sched,
2010 graph->n_total_row - row);
2013 for (j = row; j < graph->n_total_row; ++j)
2020 /* Split the current graph into two parts and compute a schedule for each
2021 * part individually. In particular, one part consists of all SCCs up
2022 * to and including graph->src_scc, while the other part contains the other
2025 * The split is enforced in the schedule by constant rows with two different
2026 * values (0 and 1). These constant rows replace the previously computed rows
2027 * in the current band.
2028 * It would be possible to reuse them as the first rows in the next
2029 * band, but recomputing them may result in better rows as we are looking
2030 * at a smaller part of the dependence graph.
2031 * compute_split_schedule is only called when no zero-distance schedule row
2032 * could be found on the entire graph, so we wark the splitting row as
2033 * non zero-distance.
2035 * The band_id of the second group is set to n, where n is the number
2036 * of nodes in the first group. This ensures that the band_ids over
2037 * the two groups remain disjoint, even if either or both of the two
2038 * groups contain independent components.
2040 static int compute_split_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
2042 int i, j, n, e1, e2;
2043 int n_total_row, orig_total_row;
2044 int n_band, orig_band;
2047 drop = graph->n_total_row - graph->band_start;
2048 graph->n_total_row -= drop;
2049 graph->n_row -= drop;
2052 for (i = 0; i < graph->n; ++i) {
2053 struct isl_sched_node *node = &graph->node[i];
2054 int row = isl_mat_rows(node->sched) - drop;
2055 int cols = isl_mat_cols(node->sched);
2056 int before = node->scc <= graph->src_scc;
2061 isl_map_free(node->sched_map);
2062 node->sched_map = NULL;
2063 node->sched = isl_mat_drop_rows(node->sched,
2064 graph->band_start, drop);
2065 node->sched = isl_mat_add_rows(node->sched, 1);
2068 node->sched = isl_mat_set_element_si(node->sched, row, 0,
2070 for (j = 1; j < cols; ++j)
2071 node->sched = isl_mat_set_element_si(node->sched,
2073 node->band[graph->n_total_row] = graph->n_band;
2074 node->zero[graph->n_total_row] = 0;
2078 for (i = 0; i < graph->n_edge; ++i) {
2079 if (graph->edge[i].dst->scc <= graph->src_scc)
2081 if (graph->edge[i].src->scc > graph->src_scc)
2085 graph->n_total_row++;
2088 for (i = 0; i < graph->n; ++i) {
2089 struct isl_sched_node *node = &graph->node[i];
2090 if (node->scc > graph->src_scc)
2091 node->band_id[graph->n_band] = n;
2094 orig_total_row = graph->n_total_row;
2095 orig_band = graph->n_band;
2096 if (compute_sub_schedule(ctx, graph, n, e1,
2097 &node_scc_at_most, &edge_dst_scc_at_most,
2098 graph->src_scc, 0) < 0)
2100 n_total_row = graph->n_total_row;
2101 graph->n_total_row = orig_total_row;
2102 n_band = graph->n_band;
2103 graph->n_band = orig_band;
2104 if (compute_sub_schedule(ctx, graph, graph->n - n, e2,
2105 &node_scc_at_least, &edge_src_scc_at_least,
2106 graph->src_scc + 1, 0) < 0)
2108 if (n_total_row > graph->n_total_row)
2109 graph->n_total_row = n_total_row;
2110 if (n_band > graph->n_band)
2111 graph->n_band = n_band;
2113 return pad_schedule(graph);
2116 /* Compute the next band of the schedule after updating the dependence
2117 * relations based on the the current schedule.
2119 static int compute_next_band(isl_ctx *ctx, struct isl_sched_graph *graph)
2121 if (update_edges(ctx, graph) < 0)
2125 return compute_schedule(ctx, graph);
2128 /* Add constraints to graph->lp that force the dependence "map" (which
2129 * is part of the dependence relation of "edge")
2130 * to be respected and attempt to carry it, where the edge is one from
2131 * a node j to itself. "pos" is the sequence number of the given map.
2132 * That is, add constraints that enforce
2134 * (c_j_0 + c_j_n n + c_j_x y) - (c_j_0 + c_j_n n + c_j_x x)
2135 * = c_j_x (y - x) >= e_i
2137 * for each (x,y) in R.
2138 * We obtain general constraints on coefficients (c_0, c_n, c_x)
2139 * of valid constraints for (y - x) and then plug in (-e_i, 0, c_j_x),
2140 * with each coefficient in c_j_x represented as a pair of non-negative
2143 static int add_intra_constraints(struct isl_sched_graph *graph,
2144 struct isl_sched_edge *edge, __isl_take isl_map *map, int pos)
2147 isl_ctx *ctx = isl_map_get_ctx(map);
2149 isl_dim_map *dim_map;
2150 isl_basic_set *coef;
2151 struct isl_sched_node *node = edge->src;
2153 coef = intra_coefficients(graph, map);
2155 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
2157 total = isl_basic_set_total_dim(graph->lp);
2158 dim_map = isl_dim_map_alloc(ctx, total);
2159 isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
2160 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
2161 isl_space_dim(dim, isl_dim_set), 1,
2163 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
2164 isl_space_dim(dim, isl_dim_set), 1,
2166 graph->lp = isl_basic_set_extend_constraints(graph->lp,
2167 coef->n_eq, coef->n_ineq);
2168 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
2170 isl_space_free(dim);
2175 /* Add constraints to graph->lp that force the dependence "map" (which
2176 * is part of the dependence relation of "edge")
2177 * to be respected and attempt to carry it, where the edge is one from
2178 * node j to node k. "pos" is the sequence number of the given map.
2179 * That is, add constraints that enforce
2181 * (c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i
2183 * for each (x,y) in R.
2184 * We obtain general constraints on coefficients (c_0, c_n, c_x)
2185 * of valid constraints for R and then plug in
2186 * (-e_i + c_k_0 - c_j_0, c_k_n - c_j_n, c_k_x - c_j_x)
2187 * with each coefficient (except e_i, c_k_0 and c_j_0)
2188 * represented as a pair of non-negative coefficients.
2190 static int add_inter_constraints(struct isl_sched_graph *graph,
2191 struct isl_sched_edge *edge, __isl_take isl_map *map, int pos)
2194 isl_ctx *ctx = isl_map_get_ctx(map);
2196 isl_dim_map *dim_map;
2197 isl_basic_set *coef;
2198 struct isl_sched_node *src = edge->src;
2199 struct isl_sched_node *dst = edge->dst;
2201 coef = inter_coefficients(graph, map);
2203 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
2205 total = isl_basic_set_total_dim(graph->lp);
2206 dim_map = isl_dim_map_alloc(ctx, total);
2208 isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
2210 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
2211 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
2212 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
2213 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
2214 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
2216 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
2217 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
2220 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
2221 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
2222 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
2223 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
2224 isl_space_dim(dim, isl_dim_set), 1,
2226 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
2227 isl_space_dim(dim, isl_dim_set), 1,
2230 graph->lp = isl_basic_set_extend_constraints(graph->lp,
2231 coef->n_eq, coef->n_ineq);
2232 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
2234 isl_space_free(dim);
2239 /* Add constraints to graph->lp that force all validity dependences
2240 * to be respected and attempt to carry them.
2242 static int add_all_constraints(struct isl_sched_graph *graph)
2248 for (i = 0; i < graph->n_edge; ++i) {
2249 struct isl_sched_edge *edge= &graph->edge[i];
2251 if (!edge->validity)
2254 for (j = 0; j < edge->map->n; ++j) {
2255 isl_basic_map *bmap;
2258 bmap = isl_basic_map_copy(edge->map->p[j]);
2259 map = isl_map_from_basic_map(bmap);
2261 if (edge->src == edge->dst &&
2262 add_intra_constraints(graph, edge, map, pos) < 0)
2264 if (edge->src != edge->dst &&
2265 add_inter_constraints(graph, edge, map, pos) < 0)
2274 /* Count the number of equality and inequality constraints
2275 * that will be added to the carry_lp problem.
2276 * We count each edge exactly once.
2278 static int count_all_constraints(struct isl_sched_graph *graph,
2279 int *n_eq, int *n_ineq)
2283 *n_eq = *n_ineq = 0;
2284 for (i = 0; i < graph->n_edge; ++i) {
2285 struct isl_sched_edge *edge= &graph->edge[i];
2286 for (j = 0; j < edge->map->n; ++j) {
2287 isl_basic_map *bmap;
2290 bmap = isl_basic_map_copy(edge->map->p[j]);
2291 map = isl_map_from_basic_map(bmap);
2293 if (count_map_constraints(graph, edge, map,
2294 n_eq, n_ineq, 1) < 0)
2302 /* Construct an LP problem for finding schedule coefficients
2303 * such that the schedule carries as many dependences as possible.
2304 * In particular, for each dependence i, we bound the dependence distance
2305 * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum
2306 * of all e_i's. Dependence with e_i = 0 in the solution are simply
2307 * respected, while those with e_i > 0 (in practice e_i = 1) are carried.
2308 * Note that if the dependence relation is a union of basic maps,
2309 * then we have to consider each basic map individually as it may only
2310 * be possible to carry the dependences expressed by some of those
2311 * basic maps and not all off them.
2312 * Below, we consider each of those basic maps as a separate "edge".
2314 * All variables of the LP are non-negative. The actual coefficients
2315 * may be negative, so each coefficient is represented as the difference
2316 * of two non-negative variables. The negative part always appears
2317 * immediately before the positive part.
2318 * Other than that, the variables have the following order
2320 * - sum of (1 - e_i) over all edges
2321 * - sum of positive and negative parts of all c_n coefficients
2322 * (unconstrained when computing non-parametric schedules)
2323 * - sum of positive and negative parts of all c_x coefficients
2328 * - positive and negative parts of c_i_n (if parametric)
2329 * - positive and negative parts of c_i_x
2331 * The constraints are those from the (validity) edges plus three equalities
2332 * to express the sums and n_edge inequalities to express e_i <= 1.
2334 static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
2344 for (i = 0; i < graph->n_edge; ++i)
2345 n_edge += graph->edge[i].map->n;
2348 for (i = 0; i < graph->n; ++i) {
2349 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
2350 node->start = total;
2351 total += 1 + 2 * (node->nparam + node->nvar);
2354 if (count_all_constraints(graph, &n_eq, &n_ineq) < 0)
2357 dim = isl_space_set_alloc(ctx, 0, total);
2358 isl_basic_set_free(graph->lp);
2361 graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq);
2362 graph->lp = isl_basic_set_set_rational(graph->lp);
2364 k = isl_basic_set_alloc_equality(graph->lp);
2367 isl_seq_clr(graph->lp->eq[k], 1 + total);
2368 isl_int_set_si(graph->lp->eq[k][0], -n_edge);
2369 isl_int_set_si(graph->lp->eq[k][1], 1);
2370 for (i = 0; i < n_edge; ++i)
2371 isl_int_set_si(graph->lp->eq[k][4 + i], 1);
2373 k = isl_basic_set_alloc_equality(graph->lp);
2376 isl_seq_clr(graph->lp->eq[k], 1 + total);
2377 isl_int_set_si(graph->lp->eq[k][2], -1);
2378 for (i = 0; i < graph->n; ++i) {
2379 int pos = 1 + graph->node[i].start + 1;
2381 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
2382 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2385 k = isl_basic_set_alloc_equality(graph->lp);
2388 isl_seq_clr(graph->lp->eq[k], 1 + total);
2389 isl_int_set_si(graph->lp->eq[k][3], -1);
2390 for (i = 0; i < graph->n; ++i) {
2391 struct isl_sched_node *node = &graph->node[i];
2392 int pos = 1 + node->start + 1 + 2 * node->nparam;
2394 for (j = 0; j < 2 * node->nvar; ++j)
2395 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2398 for (i = 0; i < n_edge; ++i) {
2399 k = isl_basic_set_alloc_inequality(graph->lp);
2402 isl_seq_clr(graph->lp->ineq[k], 1 + total);
2403 isl_int_set_si(graph->lp->ineq[k][4 + i], -1);
2404 isl_int_set_si(graph->lp->ineq[k][0], 1);
2407 if (add_all_constraints(graph) < 0)
2413 /* If the schedule_split_scaled option is set and if the linear
2414 * parts of the scheduling rows for all nodes in the graphs have
2415 * non-trivial common divisor, then split off the constant term
2416 * from the linear part.
2417 * The constant term is then placed in a separate band and
2418 * the linear part is reduced.
2420 static int split_scaled(isl_ctx *ctx, struct isl_sched_graph *graph)
2426 if (!ctx->opt->schedule_split_scaled)
2432 isl_int_init(gcd_i);
2434 isl_int_set_si(gcd, 0);
2436 row = isl_mat_rows(graph->node[0].sched) - 1;
2438 for (i = 0; i < graph->n; ++i) {
2439 struct isl_sched_node *node = &graph->node[i];
2440 int cols = isl_mat_cols(node->sched);
2442 isl_seq_gcd(node->sched->row[row] + 1, cols - 1, &gcd_i);
2443 isl_int_gcd(gcd, gcd, gcd_i);
2446 isl_int_clear(gcd_i);
2448 if (isl_int_cmp_si(gcd, 1) <= 0) {
2455 for (i = 0; i < graph->n; ++i) {
2456 struct isl_sched_node *node = &graph->node[i];
2458 isl_map_free(node->sched_map);
2459 node->sched_map = NULL;
2460 node->sched = isl_mat_add_zero_rows(node->sched, 1);
2463 isl_int_fdiv_r(node->sched->row[row + 1][0],
2464 node->sched->row[row][0], gcd);
2465 isl_int_fdiv_q(node->sched->row[row][0],
2466 node->sched->row[row][0], gcd);
2467 isl_int_mul(node->sched->row[row][0],
2468 node->sched->row[row][0], gcd);
2469 node->sched = isl_mat_scale_down_row(node->sched, row, gcd);
2472 node->band[graph->n_total_row] = graph->n_band;
2475 graph->n_total_row++;
2484 /* Construct a schedule row for each node such that as many dependences
2485 * as possible are carried and then continue with the next band.
2487 static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *graph)
2495 for (i = 0; i < graph->n_edge; ++i)
2496 n_edge += graph->edge[i].map->n;
2498 if (setup_carry_lp(ctx, graph) < 0)
2501 lp = isl_basic_set_copy(graph->lp);
2502 sol = isl_tab_basic_set_non_neg_lexmin(lp);
2506 if (sol->size == 0) {
2508 isl_die(ctx, isl_error_internal,
2509 "error in schedule construction", return -1);
2512 if (isl_int_cmp_si(sol->el[1], n_edge) >= 0) {
2514 isl_die(ctx, isl_error_unknown,
2515 "unable to carry dependences", return -1);
2518 if (update_schedule(graph, sol, 0, 0) < 0)
2521 if (split_scaled(ctx, graph) < 0)
2524 return compute_next_band(ctx, graph);
2527 /* Are there any (non-empty) validity edges in the graph?
2529 static int has_validity_edges(struct isl_sched_graph *graph)
2533 for (i = 0; i < graph->n_edge; ++i) {
2536 empty = isl_map_plain_is_empty(graph->edge[i].map);
2541 if (graph->edge[i].validity)
2548 /* Should we apply a Feautrier step?
2549 * That is, did the user request the Feautrier algorithm and are
2550 * there any validity dependences (left)?
2552 static int need_feautrier_step(isl_ctx *ctx, struct isl_sched_graph *graph)
2554 if (ctx->opt->schedule_algorithm != ISL_SCHEDULE_ALGORITHM_FEAUTRIER)
2557 return has_validity_edges(graph);
2560 /* Compute a schedule for a connected dependence graph using Feautrier's
2561 * multi-dimensional scheduling algorithm.
2562 * The original algorithm is described in [1].
2563 * The main idea is to minimize the number of scheduling dimensions, by
2564 * trying to satisfy as many dependences as possible per scheduling dimension.
2566 * [1] P. Feautrier, Some Efficient Solutions to the Affine Scheduling
2567 * Problem, Part II: Multi-Dimensional Time.
2568 * In Intl. Journal of Parallel Programming, 1992.
2570 static int compute_schedule_wcc_feautrier(isl_ctx *ctx,
2571 struct isl_sched_graph *graph)
2573 return carry_dependences(ctx, graph);
2576 /* Compute a schedule for a connected dependence graph.
2577 * We try to find a sequence of as many schedule rows as possible that result
2578 * in non-negative dependence distances (independent of the previous rows
2579 * in the sequence, i.e., such that the sequence is tilable).
2580 * If we can't find any more rows we either
2581 * - split between SCCs and start over (assuming we found an interesting
2582 * pair of SCCs between which to split)
2583 * - continue with the next band (assuming the current band has at least
2585 * - try to carry as many dependences as possible and continue with the next
2588 * If Feautrier's algorithm is selected, we first recursively try to satisfy
2589 * as many validity dependences as possible. When all validity dependences
2590 * are satisfied we extend the schedule to a full-dimensional schedule.
2592 * If we manage to complete the schedule, we finish off by topologically
2593 * sorting the statements based on the remaining dependences.
2595 * If ctx->opt->schedule_outer_zero_distance is set, then we force the
2596 * outermost dimension in the current band to be zero distance. If this
2597 * turns out to be impossible, we fall back on the general scheme above
2598 * and try to carry as many dependences as possible.
2600 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph)
2604 if (detect_sccs(graph) < 0)
2608 if (compute_maxvar(graph) < 0)
2611 if (need_feautrier_step(ctx, graph))
2612 return compute_schedule_wcc_feautrier(ctx, graph);
2614 if (ctx->opt->schedule_outer_zero_distance)
2617 while (graph->n_row < graph->maxvar) {
2620 graph->src_scc = -1;
2621 graph->dst_scc = -1;
2623 if (setup_lp(ctx, graph, force_zero) < 0)
2625 sol = solve_lp(graph);
2628 if (sol->size == 0) {
2630 if (!ctx->opt->schedule_maximize_band_depth &&
2631 graph->n_total_row > graph->band_start)
2632 return compute_next_band(ctx, graph);
2633 if (graph->src_scc >= 0)
2634 return compute_split_schedule(ctx, graph);
2635 if (graph->n_total_row > graph->band_start)
2636 return compute_next_band(ctx, graph);
2637 return carry_dependences(ctx, graph);
2639 if (update_schedule(graph, sol, 1, 1) < 0)
2644 if (graph->n_total_row > graph->band_start)
2646 return sort_statements(ctx, graph);
2649 /* Add a row to the schedules that separates the SCCs and move
2652 static int split_on_scc(struct isl_sched_graph *graph)
2656 for (i = 0; i < graph->n; ++i) {
2657 struct isl_sched_node *node = &graph->node[i];
2658 int row = isl_mat_rows(node->sched);
2660 isl_map_free(node->sched_map);
2661 node->sched_map = NULL;
2662 node->sched = isl_mat_add_zero_rows(node->sched, 1);
2663 node->sched = isl_mat_set_element_si(node->sched, row, 0,
2667 node->band[graph->n_total_row] = graph->n_band;
2670 graph->n_total_row++;
2676 /* Compute a schedule for each component (identified by node->scc)
2677 * of the dependence graph separately and then combine the results.
2678 * Depending on the setting of schedule_fuse, a component may be
2679 * either weakly or strongly connected.
2681 * The band_id is adjusted such that each component has a separate id.
2682 * Note that the band_id may have already been set to a value different
2683 * from zero by compute_split_schedule.
2685 static int compute_component_schedule(isl_ctx *ctx,
2686 struct isl_sched_graph *graph)
2690 int n_total_row, orig_total_row;
2691 int n_band, orig_band;
2693 if (ctx->opt->schedule_fuse == ISL_SCHEDULE_FUSE_MIN)
2694 split_on_scc(graph);
2697 orig_total_row = graph->n_total_row;
2699 orig_band = graph->n_band;
2700 for (i = 0; i < graph->n; ++i)
2701 graph->node[i].band_id[graph->n_band] += graph->node[i].scc;
2702 for (wcc = 0; wcc < graph->scc; ++wcc) {
2704 for (i = 0; i < graph->n; ++i)
2705 if (graph->node[i].scc == wcc)
2708 for (i = 0; i < graph->n_edge; ++i)
2709 if (graph->edge[i].src->scc == wcc &&
2710 graph->edge[i].dst->scc == wcc)
2713 if (compute_sub_schedule(ctx, graph, n, n_edge,
2715 &edge_scc_exactly, wcc, 1) < 0)
2717 if (graph->n_total_row > n_total_row)
2718 n_total_row = graph->n_total_row;
2719 graph->n_total_row = orig_total_row;
2720 if (graph->n_band > n_band)
2721 n_band = graph->n_band;
2722 graph->n_band = orig_band;
2725 graph->n_total_row = n_total_row;
2726 graph->n_band = n_band;
2728 return pad_schedule(graph);
2731 /* Compute a schedule for the given dependence graph.
2732 * We first check if the graph is connected (through validity dependences)
2733 * and, if not, compute a schedule for each component separately.
2734 * If schedule_fuse is set to minimal fusion, then we check for strongly
2735 * connected components instead and compute a separate schedule for
2736 * each such strongly connected component.
2738 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
2740 if (ctx->opt->schedule_fuse == ISL_SCHEDULE_FUSE_MIN) {
2741 if (detect_sccs(graph) < 0)
2744 if (detect_wccs(graph) < 0)
2749 return compute_component_schedule(ctx, graph);
2751 return compute_schedule_wcc(ctx, graph);
2754 /* Compute a schedule for the given union of domains that respects
2755 * all the validity dependences.
2756 * If the default isl scheduling algorithm is used, it tries to minimize
2757 * the dependence distances over the proximity dependences.
2758 * If Feautrier's scheduling algorithm is used, the proximity dependence
2759 * distances are only minimized during the extension to a full-dimensional
2762 __isl_give isl_schedule *isl_union_set_compute_schedule(
2763 __isl_take isl_union_set *domain,
2764 __isl_take isl_union_map *validity,
2765 __isl_take isl_union_map *proximity)
2767 isl_ctx *ctx = isl_union_set_get_ctx(domain);
2769 struct isl_sched_graph graph = { 0 };
2770 isl_schedule *sched;
2771 struct isl_extract_edge_data data;
2773 domain = isl_union_set_align_params(domain,
2774 isl_union_map_get_space(validity));
2775 domain = isl_union_set_align_params(domain,
2776 isl_union_map_get_space(proximity));
2777 dim = isl_union_set_get_space(domain);
2778 validity = isl_union_map_align_params(validity, isl_space_copy(dim));
2779 proximity = isl_union_map_align_params(proximity, dim);
2784 graph.n = isl_union_set_n_set(domain);
2787 if (graph_alloc(ctx, &graph, graph.n,
2788 isl_union_map_n_map(validity) + isl_union_map_n_map(proximity)) < 0)
2792 if (isl_union_set_foreach_set(domain, &extract_node, &graph) < 0)
2794 if (graph_init_table(ctx, &graph) < 0)
2796 graph.max_edge[isl_edge_validity] = isl_union_map_n_map(validity);
2797 graph.max_edge[isl_edge_proximity] = isl_union_map_n_map(proximity);
2798 if (graph_init_edge_tables(ctx, &graph) < 0)
2801 data.graph = &graph;
2802 data.type = isl_edge_validity;
2803 if (isl_union_map_foreach_map(validity, &extract_edge, &data) < 0)
2805 data.type = isl_edge_proximity;
2806 if (isl_union_map_foreach_map(proximity, &extract_edge, &data) < 0)
2809 if (compute_schedule(ctx, &graph) < 0)
2813 sched = extract_schedule(&graph, isl_union_set_get_space(domain));
2815 graph_free(ctx, &graph);
2816 isl_union_set_free(domain);
2817 isl_union_map_free(validity);
2818 isl_union_map_free(proximity);
2822 graph_free(ctx, &graph);
2823 isl_union_set_free(domain);
2824 isl_union_map_free(validity);
2825 isl_union_map_free(proximity);
2829 void *isl_schedule_free(__isl_take isl_schedule *sched)
2835 if (--sched->ref > 0)
2838 for (i = 0; i < sched->n; ++i) {
2839 isl_map_free(sched->node[i].sched);
2840 free(sched->node[i].band_end);
2841 free(sched->node[i].band_id);
2842 free(sched->node[i].zero);
2844 isl_space_free(sched->dim);
2845 isl_band_list_free(sched->band_forest);
2850 isl_ctx *isl_schedule_get_ctx(__isl_keep isl_schedule *schedule)
2852 return schedule ? isl_space_get_ctx(schedule->dim) : NULL;
2855 __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched)
2858 isl_union_map *umap;
2863 umap = isl_union_map_empty(isl_space_copy(sched->dim));
2864 for (i = 0; i < sched->n; ++i)
2865 umap = isl_union_map_add_map(umap,
2866 isl_map_copy(sched->node[i].sched));
2871 static __isl_give isl_band_list *construct_band_list(
2872 __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
2873 int band_nr, int *parent_active, int n_active);
2875 /* Construct an isl_band structure for the band in the given schedule
2876 * with sequence number band_nr for the n_active nodes marked by active.
2877 * If the nodes don't have a band with the given sequence number,
2878 * then a band without members is created.
2880 * Because of the way the schedule is constructed, we know that
2881 * the position of the band inside the schedule of a node is the same
2882 * for all active nodes.
2884 static __isl_give isl_band *construct_band(__isl_keep isl_schedule *schedule,
2885 __isl_keep isl_band *parent,
2886 int band_nr, int *active, int n_active)
2889 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2891 unsigned start, end;
2893 band = isl_calloc_type(ctx, isl_band);
2898 band->schedule = schedule;
2899 band->parent = parent;
2901 for (i = 0; i < schedule->n; ++i)
2902 if (active[i] && schedule->node[i].n_band > band_nr + 1)
2905 if (i < schedule->n) {
2906 band->children = construct_band_list(schedule, band,
2907 band_nr + 1, active, n_active);
2908 if (!band->children)
2912 for (i = 0; i < schedule->n; ++i)
2916 if (i >= schedule->n)
2917 isl_die(ctx, isl_error_internal,
2918 "band without active statements", goto error);
2920 start = band_nr ? schedule->node[i].band_end[band_nr - 1] : 0;
2921 end = band_nr < schedule->node[i].n_band ?
2922 schedule->node[i].band_end[band_nr] : start;
2923 band->n = end - start;
2925 band->zero = isl_alloc_array(ctx, int, band->n);
2929 for (j = 0; j < band->n; ++j)
2930 band->zero[j] = schedule->node[i].zero[start + j];
2932 band->map = isl_union_map_empty(isl_space_copy(schedule->dim));
2933 for (i = 0; i < schedule->n; ++i) {
2940 map = isl_map_copy(schedule->node[i].sched);
2941 n_out = isl_map_dim(map, isl_dim_out);
2942 map = isl_map_project_out(map, isl_dim_out, end, n_out - end);
2943 map = isl_map_project_out(map, isl_dim_out, 0, start);
2944 band->map = isl_union_map_union(band->map,
2945 isl_union_map_from_map(map));
2952 isl_band_free(band);
2956 /* Construct a list of bands that start at the same position (with
2957 * sequence number band_nr) in the schedules of the nodes that
2958 * were active in the parent band.
2960 * A separate isl_band structure is created for each band_id
2961 * and for each node that does not have a band with sequence
2962 * number band_nr. In the latter case, a band without members
2964 * This ensures that if a band has any children, then each node
2965 * that was active in the band is active in exactly one of the children.
2967 static __isl_give isl_band_list *construct_band_list(
2968 __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
2969 int band_nr, int *parent_active, int n_active)
2972 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2975 isl_band_list *list;
2978 for (i = 0; i < n_active; ++i) {
2979 for (j = 0; j < schedule->n; ++j) {
2980 if (!parent_active[j])
2982 if (schedule->node[j].n_band <= band_nr)
2984 if (schedule->node[j].band_id[band_nr] == i) {
2990 for (j = 0; j < schedule->n; ++j)
2991 if (schedule->node[j].n_band <= band_nr)
2996 list = isl_band_list_alloc(ctx, n_band);
2997 band = construct_band(schedule, parent, band_nr,
2998 parent_active, n_active);
2999 return isl_band_list_add(list, band);
3002 active = isl_alloc_array(ctx, int, schedule->n);
3006 list = isl_band_list_alloc(ctx, n_band);
3008 for (i = 0; i < n_active; ++i) {
3012 for (j = 0; j < schedule->n; ++j) {
3013 active[j] = parent_active[j] &&
3014 schedule->node[j].n_band > band_nr &&
3015 schedule->node[j].band_id[band_nr] == i;
3022 band = construct_band(schedule, parent, band_nr, active, n);
3024 list = isl_band_list_add(list, band);
3026 for (i = 0; i < schedule->n; ++i) {
3028 if (!parent_active[i])
3030 if (schedule->node[i].n_band > band_nr)
3032 for (j = 0; j < schedule->n; ++j)
3034 band = construct_band(schedule, parent, band_nr, active, 1);
3035 list = isl_band_list_add(list, band);
3043 /* Construct a band forest representation of the schedule and
3044 * return the list of roots.
3046 static __isl_give isl_band_list *construct_forest(
3047 __isl_keep isl_schedule *schedule)
3050 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
3051 isl_band_list *forest;
3054 active = isl_alloc_array(ctx, int, schedule->n);
3058 for (i = 0; i < schedule->n; ++i)
3061 forest = construct_band_list(schedule, NULL, 0, active, schedule->n);
3068 /* Return the roots of a band forest representation of the schedule.
3070 __isl_give isl_band_list *isl_schedule_get_band_forest(
3071 __isl_keep isl_schedule *schedule)
3075 if (!schedule->band_forest)
3076 schedule->band_forest = construct_forest(schedule);
3077 return isl_band_list_dup(schedule->band_forest);
3080 static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
3081 __isl_keep isl_band_list *list);
3083 static __isl_give isl_printer *print_band(__isl_take isl_printer *p,
3084 __isl_keep isl_band *band)
3086 isl_band_list *children;
3088 p = isl_printer_start_line(p);
3089 p = isl_printer_print_union_map(p, band->map);
3090 p = isl_printer_end_line(p);
3092 if (!isl_band_has_children(band))
3095 children = isl_band_get_children(band);
3097 p = isl_printer_indent(p, 4);
3098 p = print_band_list(p, children);
3099 p = isl_printer_indent(p, -4);
3101 isl_band_list_free(children);
3106 static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
3107 __isl_keep isl_band_list *list)
3111 n = isl_band_list_n_band(list);
3112 for (i = 0; i < n; ++i) {
3114 band = isl_band_list_get_band(list, i);
3115 p = print_band(p, band);
3116 isl_band_free(band);
3122 __isl_give isl_printer *isl_printer_print_schedule(__isl_take isl_printer *p,
3123 __isl_keep isl_schedule *schedule)
3125 isl_band_list *forest;
3127 forest = isl_schedule_get_band_forest(schedule);
3129 p = print_band_list(p, forest);
3131 isl_band_list_free(forest);
3136 void isl_schedule_dump(__isl_keep isl_schedule *schedule)
3138 isl_printer *printer;
3143 printer = isl_printer_to_file(isl_schedule_get_ctx(schedule), stderr);
3144 printer = isl_printer_print_schedule(printer, schedule);
3146 isl_printer_free(printer);