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;
127 /* Internal information about the dependence graph used during
128 * the construction of the schedule.
130 * intra_hmap is a cache, mapping dependence relations to their dual,
131 * for dependences from a node to itself
132 * inter_hmap is a cache, mapping dependence relations to their dual,
133 * for dependences between distinct nodes
135 * n is the number of nodes
136 * node is the list of nodes
137 * maxvar is the maximal number of variables over all nodes
138 * n_row is the current (maximal) number of linearly independent
139 * rows in the node schedules
140 * n_total_row is the current number of rows in the node schedules
141 * n_band is the current number of completed bands
142 * band_start is the starting row in the node schedules of the current band
143 * root is set if this graph is the original dependence graph,
144 * without any splitting
146 * sorted contains a list of node indices sorted according to the
147 * SCC to which a node belongs
149 * n_edge is the number of edges
150 * edge is the list of edges
151 * edge_table contains pointers into the edge array, hashed on the source
152 * and sink spaces; the table only contains edges that represent
153 * validity constraints (and that may or may not also represent proximity
156 * node_table contains pointers into the node array, hashed on the space
158 * region contains a list of variable sequences that should be non-trivial
160 * lp contains the (I)LP problem used to obtain new schedule rows
162 * src_scc and dst_scc are the source and sink SCCs of an edge with
163 * conflicting constraints
165 * scc, sp, index and stack are used during the detection of SCCs
166 * scc is the number of the next SCC
167 * stack contains the nodes on the path from the root to the current node
168 * sp is the stack pointer
169 * index is the index of the last node visited
171 struct isl_sched_graph {
172 isl_hmap_map_basic_set *intra_hmap;
173 isl_hmap_map_basic_set *inter_hmap;
175 struct isl_sched_node *node;
188 struct isl_sched_edge *edge;
190 struct isl_hash_table *edge_table;
192 struct isl_hash_table *node_table;
193 struct isl_region *region;
207 /* Initialize node_table based on the list of nodes.
209 static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph)
213 graph->node_table = isl_hash_table_alloc(ctx, graph->n);
214 if (!graph->node_table)
217 for (i = 0; i < graph->n; ++i) {
218 struct isl_hash_table_entry *entry;
221 hash = isl_space_get_hash(graph->node[i].dim);
222 entry = isl_hash_table_find(ctx, graph->node_table, hash,
224 graph->node[i].dim, 1);
227 entry->data = &graph->node[i];
233 /* Return a pointer to the node that lives within the given space,
234 * or NULL if there is no such node.
236 static struct isl_sched_node *graph_find_node(isl_ctx *ctx,
237 struct isl_sched_graph *graph, __isl_keep isl_space *dim)
239 struct isl_hash_table_entry *entry;
242 hash = isl_space_get_hash(dim);
243 entry = isl_hash_table_find(ctx, graph->node_table, hash,
244 &node_has_dim, dim, 0);
246 return entry ? entry->data : NULL;
249 static int edge_has_src_and_dst(const void *entry, const void *val)
251 const struct isl_sched_edge *edge = entry;
252 const struct isl_sched_edge *temp = val;
254 return edge->src == temp->src && edge->dst == temp->dst;
257 /* Initialize edge_table based on the list of edges.
258 * Only edges with validity set are added to the table.
260 static int graph_init_edge_table(isl_ctx *ctx, struct isl_sched_graph *graph)
264 graph->edge_table = isl_hash_table_alloc(ctx, graph->n_edge);
265 if (!graph->edge_table)
268 for (i = 0; i < graph->n_edge; ++i) {
269 struct isl_hash_table_entry *entry;
272 if (!graph->edge[i].validity)
275 hash = isl_hash_init();
276 hash = isl_hash_builtin(hash, graph->edge[i].src);
277 hash = isl_hash_builtin(hash, graph->edge[i].dst);
278 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
279 &edge_has_src_and_dst,
283 entry->data = &graph->edge[i];
289 /* Check whether the dependence graph has a (validity) edge
290 * between the given two nodes.
292 static int graph_has_edge(struct isl_sched_graph *graph,
293 struct isl_sched_node *src, struct isl_sched_node *dst)
295 isl_ctx *ctx = isl_space_get_ctx(src->dim);
296 struct isl_hash_table_entry *entry;
298 struct isl_sched_edge temp = { .src = src, .dst = dst };
299 struct isl_sched_edge *edge;
302 hash = isl_hash_init();
303 hash = isl_hash_builtin(hash, temp.src);
304 hash = isl_hash_builtin(hash, temp.dst);
305 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
306 &edge_has_src_and_dst, &temp, 0);
311 empty = isl_map_plain_is_empty(edge->map);
318 static int graph_alloc(isl_ctx *ctx, struct isl_sched_graph *graph,
319 int n_node, int n_edge)
324 graph->n_edge = n_edge;
325 graph->node = isl_calloc_array(ctx, struct isl_sched_node, graph->n);
326 graph->sorted = isl_calloc_array(ctx, int, graph->n);
327 graph->region = isl_alloc_array(ctx, struct isl_region, graph->n);
328 graph->stack = isl_alloc_array(ctx, int, graph->n);
329 graph->edge = isl_calloc_array(ctx,
330 struct isl_sched_edge, graph->n_edge);
332 graph->intra_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
333 graph->inter_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
335 if (!graph->node || !graph->region || !graph->stack || !graph->edge ||
339 for(i = 0; i < graph->n; ++i)
340 graph->sorted[i] = i;
345 static void graph_free(isl_ctx *ctx, struct isl_sched_graph *graph)
349 isl_hmap_map_basic_set_free(ctx, graph->intra_hmap);
350 isl_hmap_map_basic_set_free(ctx, graph->inter_hmap);
352 for (i = 0; i < graph->n; ++i) {
353 isl_space_free(graph->node[i].dim);
354 isl_mat_free(graph->node[i].sched);
355 isl_map_free(graph->node[i].sched_map);
356 isl_mat_free(graph->node[i].cmap);
358 free(graph->node[i].band);
359 free(graph->node[i].band_id);
360 free(graph->node[i].zero);
365 for (i = 0; i < graph->n_edge; ++i)
366 isl_map_free(graph->edge[i].map);
370 isl_hash_table_free(ctx, graph->edge_table);
371 isl_hash_table_free(ctx, graph->node_table);
372 isl_basic_set_free(graph->lp);
375 /* Add a new node to the graph representing the given set.
377 static int extract_node(__isl_take isl_set *set, void *user)
383 struct isl_sched_graph *graph = user;
384 int *band, *band_id, *zero;
386 ctx = isl_set_get_ctx(set);
387 dim = isl_set_get_space(set);
389 nvar = isl_space_dim(dim, isl_dim_set);
390 nparam = isl_space_dim(dim, isl_dim_param);
391 if (!ctx->opt->schedule_parametric)
393 sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar);
394 graph->node[graph->n].dim = dim;
395 graph->node[graph->n].nvar = nvar;
396 graph->node[graph->n].nparam = nparam;
397 graph->node[graph->n].sched = sched;
398 graph->node[graph->n].sched_map = NULL;
399 band = isl_alloc_array(ctx, int, graph->n_edge + nvar);
400 graph->node[graph->n].band = band;
401 band_id = isl_calloc_array(ctx, int, graph->n_edge + nvar);
402 graph->node[graph->n].band_id = band_id;
403 zero = isl_calloc_array(ctx, int, graph->n_edge + nvar);
404 graph->node[graph->n].zero = zero;
407 if (!sched || !band || !band_id || !zero)
413 /* Add a new edge to the graph based on the given map.
414 * Edges are first extracted from the validity dependences,
415 * from which the edge_table is constructed.
416 * Afterwards, the proximity dependences are added. If a proximity
417 * dependence relation happens to be identical to one of the
418 * validity dependence relations added before, then we don't create
419 * a new edge, but instead mark the original edge as also representing
420 * a proximity dependence.
422 static int extract_edge(__isl_take isl_map *map, void *user)
424 isl_ctx *ctx = isl_map_get_ctx(map);
425 struct isl_sched_graph *graph = user;
426 struct isl_sched_node *src, *dst;
429 dim = isl_space_domain(isl_map_get_space(map));
430 src = graph_find_node(ctx, graph, dim);
432 dim = isl_space_range(isl_map_get_space(map));
433 dst = graph_find_node(ctx, graph, dim);
441 graph->edge[graph->n_edge].src = src;
442 graph->edge[graph->n_edge].dst = dst;
443 graph->edge[graph->n_edge].map = map;
444 graph->edge[graph->n_edge].validity = !graph->edge_table;
445 graph->edge[graph->n_edge].proximity = !!graph->edge_table;
448 if (graph->edge_table) {
450 struct isl_hash_table_entry *entry;
451 struct isl_sched_edge *edge;
454 hash = isl_hash_init();
455 hash = isl_hash_builtin(hash, src);
456 hash = isl_hash_builtin(hash, dst);
457 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
458 &edge_has_src_and_dst,
459 &graph->edge[graph->n_edge - 1], 0);
463 is_equal = isl_map_plain_is_equal(map, edge->map);
477 /* Check whether there is a validity dependence from src to dst,
478 * forcing dst to follow src.
480 static int node_follows(struct isl_sched_graph *graph,
481 struct isl_sched_node *dst, struct isl_sched_node *src)
483 return graph_has_edge(graph, src, dst);
486 /* Perform Tarjan's algorithm for computing the strongly connected components
487 * in the dependence graph (only validity edges).
488 * If directed is not set, we consider the graph to be undirected and
489 * we effectively compute the (weakly) connected components.
491 static int detect_sccs_tarjan(struct isl_sched_graph *g, int i, int directed)
495 g->node[i].index = g->index;
496 g->node[i].min_index = g->index;
497 g->node[i].on_stack = 1;
499 g->stack[g->sp++] = i;
501 for (j = g->n - 1; j >= 0; --j) {
506 if (g->node[j].index >= 0 &&
507 (!g->node[j].on_stack ||
508 g->node[j].index > g->node[i].min_index))
511 f = node_follows(g, &g->node[i], &g->node[j]);
514 if (!f && !directed) {
515 f = node_follows(g, &g->node[j], &g->node[i]);
521 if (g->node[j].index < 0) {
522 detect_sccs_tarjan(g, j, directed);
523 if (g->node[j].min_index < g->node[i].min_index)
524 g->node[i].min_index = g->node[j].min_index;
525 } else if (g->node[j].index < g->node[i].min_index)
526 g->node[i].min_index = g->node[j].index;
529 if (g->node[i].index != g->node[i].min_index)
533 j = g->stack[--g->sp];
534 g->node[j].on_stack = 0;
535 g->node[j].scc = g->scc;
542 static int detect_ccs(struct isl_sched_graph *graph, int directed)
549 for (i = graph->n - 1; i >= 0; --i)
550 graph->node[i].index = -1;
552 for (i = graph->n - 1; i >= 0; --i) {
553 if (graph->node[i].index >= 0)
555 if (detect_sccs_tarjan(graph, i, directed) < 0)
562 /* Apply Tarjan's algorithm to detect the strongly connected components
563 * in the dependence graph.
565 static int detect_sccs(struct isl_sched_graph *graph)
567 return detect_ccs(graph, 1);
570 /* Apply Tarjan's algorithm to detect the (weakly) connected components
571 * in the dependence graph.
573 static int detect_wccs(struct isl_sched_graph *graph)
575 return detect_ccs(graph, 0);
578 static int cmp_scc(const void *a, const void *b, void *data)
580 struct isl_sched_graph *graph = data;
584 return graph->node[*i1].scc - graph->node[*i2].scc;
587 /* Sort the elements of graph->sorted according to the corresponding SCCs.
589 static void sort_sccs(struct isl_sched_graph *graph)
591 isl_quicksort(graph->sorted, graph->n, sizeof(int), &cmp_scc, graph);
594 /* Given a dependence relation R from a node to itself,
595 * construct the set of coefficients of valid constraints for elements
596 * in that dependence relation.
597 * In particular, the result contains tuples of coefficients
598 * c_0, c_n, c_x such that
600 * c_0 + c_n n + c_x y - c_x x >= 0 for each (x,y) in R
604 * c_0 + c_n n + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R }
606 * We choose here to compute the dual of delta R.
607 * Alternatively, we could have computed the dual of R, resulting
608 * in a set of tuples c_0, c_n, c_x, c_y, and then
609 * plugged in (c_0, c_n, c_x, -c_x).
611 static __isl_give isl_basic_set *intra_coefficients(
612 struct isl_sched_graph *graph, __isl_take isl_map *map)
614 isl_ctx *ctx = isl_map_get_ctx(map);
618 if (isl_hmap_map_basic_set_has(ctx, graph->intra_hmap, map))
619 return isl_hmap_map_basic_set_get(ctx, graph->intra_hmap, map);
621 delta = isl_set_remove_divs(isl_map_deltas(isl_map_copy(map)));
622 coef = isl_set_coefficients(delta);
623 isl_hmap_map_basic_set_set(ctx, graph->intra_hmap, map,
624 isl_basic_set_copy(coef));
629 /* Given a dependence relation R, * construct the set of coefficients
630 * of valid constraints for elements in that dependence relation.
631 * In particular, the result contains tuples of coefficients
632 * c_0, c_n, c_x, c_y such that
634 * c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R
637 static __isl_give isl_basic_set *inter_coefficients(
638 struct isl_sched_graph *graph, __isl_take isl_map *map)
640 isl_ctx *ctx = isl_map_get_ctx(map);
644 if (isl_hmap_map_basic_set_has(ctx, graph->inter_hmap, map))
645 return isl_hmap_map_basic_set_get(ctx, graph->inter_hmap, map);
647 set = isl_map_wrap(isl_map_remove_divs(isl_map_copy(map)));
648 coef = isl_set_coefficients(set);
649 isl_hmap_map_basic_set_set(ctx, graph->inter_hmap, map,
650 isl_basic_set_copy(coef));
655 /* Add constraints to graph->lp that force validity for the given
656 * dependence from a node i to itself.
657 * That is, add constraints that enforce
659 * (c_i_0 + c_i_n n + c_i_x y) - (c_i_0 + c_i_n n + c_i_x x)
660 * = c_i_x (y - x) >= 0
662 * for each (x,y) in R.
663 * We obtain general constraints on coefficients (c_0, c_n, c_x)
664 * of valid constraints for (y - x) and then plug in (0, 0, c_i_x^+ - c_i_x^-),
665 * where c_i_x = c_i_x^+ - c_i_x^-, with c_i_x^+ and c_i_x^- non-negative.
666 * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart.
668 * Actually, we do not construct constraints for the c_i_x themselves,
669 * but for the coefficients of c_i_x written as a linear combination
670 * of the columns in node->cmap.
672 static int add_intra_validity_constraints(struct isl_sched_graph *graph,
673 struct isl_sched_edge *edge)
676 isl_map *map = isl_map_copy(edge->map);
677 isl_ctx *ctx = isl_map_get_ctx(map);
679 isl_dim_map *dim_map;
681 struct isl_sched_node *node = edge->src;
683 coef = intra_coefficients(graph, map);
685 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
687 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
688 isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap));
690 total = isl_basic_set_total_dim(graph->lp);
691 dim_map = isl_dim_map_alloc(ctx, total);
692 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
693 isl_space_dim(dim, isl_dim_set), 1,
695 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
696 isl_space_dim(dim, isl_dim_set), 1,
698 graph->lp = isl_basic_set_extend_constraints(graph->lp,
699 coef->n_eq, coef->n_ineq);
700 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
707 /* Add constraints to graph->lp that force validity for the given
708 * dependence from node i to node j.
709 * That is, add constraints that enforce
711 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) >= 0
713 * for each (x,y) in R.
714 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
715 * of valid constraints for R and then plug in
716 * (c_j_0 - c_i_0, c_j_n^+ - c_j_n^- - (c_i_n^+ - c_i_n^-),
717 * c_j_x^+ - c_j_x^- - (c_i_x^+ - c_i_x^-)),
718 * where c_* = c_*^+ - c_*^-, with c_*^+ and c_*^- non-negative.
719 * In graph->lp, the c_*^- appear before their c_*^+ counterpart.
721 * Actually, we do not construct constraints for the c_*_x themselves,
722 * but for the coefficients of c_*_x written as a linear combination
723 * of the columns in node->cmap.
725 static int add_inter_validity_constraints(struct isl_sched_graph *graph,
726 struct isl_sched_edge *edge)
729 isl_map *map = isl_map_copy(edge->map);
730 isl_ctx *ctx = isl_map_get_ctx(map);
732 isl_dim_map *dim_map;
734 struct isl_sched_node *src = edge->src;
735 struct isl_sched_node *dst = edge->dst;
737 coef = inter_coefficients(graph, map);
739 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
741 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
742 isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap));
743 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
744 isl_space_dim(dim, isl_dim_set) + src->nvar,
745 isl_mat_copy(dst->cmap));
747 total = isl_basic_set_total_dim(graph->lp);
748 dim_map = isl_dim_map_alloc(ctx, total);
750 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
751 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
752 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
753 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
754 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
756 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
757 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
760 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
761 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
762 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
763 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
764 isl_space_dim(dim, isl_dim_set), 1,
766 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
767 isl_space_dim(dim, isl_dim_set), 1,
770 edge->start = graph->lp->n_ineq;
771 graph->lp = isl_basic_set_extend_constraints(graph->lp,
772 coef->n_eq, coef->n_ineq);
773 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
776 edge->end = graph->lp->n_ineq;
781 /* Add constraints to graph->lp that bound the dependence distance for the given
782 * dependence from a node i to itself.
783 * If s = 1, we add the constraint
785 * c_i_x (y - x) <= m_0 + m_n n
789 * -c_i_x (y - x) + m_0 + m_n n >= 0
791 * for each (x,y) in R.
792 * If s = -1, we add the constraint
794 * -c_i_x (y - x) <= m_0 + m_n n
798 * c_i_x (y - x) + m_0 + m_n n >= 0
800 * for each (x,y) in R.
801 * We obtain general constraints on coefficients (c_0, c_n, c_x)
802 * of valid constraints for (y - x) and then plug in (m_0, m_n, -s * c_i_x),
803 * with each coefficient (except m_0) represented as a pair of non-negative
806 * Actually, we do not construct constraints for the c_i_x themselves,
807 * but for the coefficients of c_i_x written as a linear combination
808 * of the columns in node->cmap.
810 static int add_intra_proximity_constraints(struct isl_sched_graph *graph,
811 struct isl_sched_edge *edge, int s)
815 isl_map *map = isl_map_copy(edge->map);
816 isl_ctx *ctx = isl_map_get_ctx(map);
818 isl_dim_map *dim_map;
820 struct isl_sched_node *node = edge->src;
822 coef = intra_coefficients(graph, map);
824 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
826 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
827 isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap));
829 nparam = isl_space_dim(node->dim, isl_dim_param);
830 total = isl_basic_set_total_dim(graph->lp);
831 dim_map = isl_dim_map_alloc(ctx, total);
832 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
833 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
834 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
835 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
836 isl_space_dim(dim, isl_dim_set), 1,
838 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
839 isl_space_dim(dim, isl_dim_set), 1,
841 graph->lp = isl_basic_set_extend_constraints(graph->lp,
842 coef->n_eq, coef->n_ineq);
843 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
850 /* Add constraints to graph->lp that bound the dependence distance for the given
851 * dependence from node i to node j.
852 * If s = 1, we add the constraint
854 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)
859 * -(c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x) +
862 * for each (x,y) in R.
863 * If s = -1, we add the constraint
865 * -((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x))
870 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) +
873 * for each (x,y) in R.
874 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
875 * of valid constraints for R and then plug in
876 * (m_0 - s*c_j_0 + s*c_i_0, m_n - s*c_j_n + s*c_i_n,
878 * with each coefficient (except m_0, c_j_0 and c_i_0)
879 * represented as a pair of non-negative coefficients.
881 * Actually, we do not construct constraints for the c_*_x themselves,
882 * but for the coefficients of c_*_x written as a linear combination
883 * of the columns in node->cmap.
885 static int add_inter_proximity_constraints(struct isl_sched_graph *graph,
886 struct isl_sched_edge *edge, int s)
890 isl_map *map = isl_map_copy(edge->map);
891 isl_ctx *ctx = isl_map_get_ctx(map);
893 isl_dim_map *dim_map;
895 struct isl_sched_node *src = edge->src;
896 struct isl_sched_node *dst = edge->dst;
898 coef = inter_coefficients(graph, map);
900 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
902 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
903 isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap));
904 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
905 isl_space_dim(dim, isl_dim_set) + src->nvar,
906 isl_mat_copy(dst->cmap));
908 nparam = isl_space_dim(src->dim, isl_dim_param);
909 total = isl_basic_set_total_dim(graph->lp);
910 dim_map = isl_dim_map_alloc(ctx, total);
912 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
913 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
914 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
916 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, -s);
917 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, s);
918 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, -s);
919 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
920 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
922 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
923 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
926 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, s);
927 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, -s);
928 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, s);
929 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
930 isl_space_dim(dim, isl_dim_set), 1,
932 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
933 isl_space_dim(dim, isl_dim_set), 1,
936 graph->lp = isl_basic_set_extend_constraints(graph->lp,
937 coef->n_eq, coef->n_ineq);
938 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
945 static int add_all_validity_constraints(struct isl_sched_graph *graph)
949 for (i = 0; i < graph->n_edge; ++i) {
950 struct isl_sched_edge *edge= &graph->edge[i];
953 if (edge->src != edge->dst)
955 if (add_intra_validity_constraints(graph, edge) < 0)
959 for (i = 0; i < graph->n_edge; ++i) {
960 struct isl_sched_edge *edge = &graph->edge[i];
963 if (edge->src == edge->dst)
965 if (add_inter_validity_constraints(graph, edge) < 0)
972 /* Add constraints to graph->lp that bound the dependence distance
973 * for all dependence relations.
974 * If a given proximity dependence is identical to a validity
975 * dependence, then the dependence distance is already bounded
976 * from below (by zero), so we only need to bound the distance
978 * Otherwise, we need to bound the distance both from above and from below.
980 static int add_all_proximity_constraints(struct isl_sched_graph *graph)
984 for (i = 0; i < graph->n_edge; ++i) {
985 struct isl_sched_edge *edge= &graph->edge[i];
986 if (!edge->proximity)
988 if (edge->src == edge->dst &&
989 add_intra_proximity_constraints(graph, edge, 1) < 0)
991 if (edge->src != edge->dst &&
992 add_inter_proximity_constraints(graph, edge, 1) < 0)
996 if (edge->src == edge->dst &&
997 add_intra_proximity_constraints(graph, edge, -1) < 0)
999 if (edge->src != edge->dst &&
1000 add_inter_proximity_constraints(graph, edge, -1) < 0)
1007 /* Compute a basis for the rows in the linear part of the schedule
1008 * and extend this basis to a full basis. The remaining rows
1009 * can then be used to force linear independence from the rows
1012 * In particular, given the schedule rows S, we compute
1016 * with H the Hermite normal form of S. That is, all but the
1017 * first rank columns of Q are zero and so each row in S is
1018 * a linear combination of the first rank rows of Q.
1019 * The matrix Q is then transposed because we will write the
1020 * coefficients of the next schedule row as a column vector s
1021 * and express this s as a linear combination s = Q c of the
1024 static int node_update_cmap(struct isl_sched_node *node)
1027 int n_row = isl_mat_rows(node->sched);
1029 H = isl_mat_sub_alloc(node->sched, 0, n_row,
1030 1 + node->nparam, node->nvar);
1032 H = isl_mat_left_hermite(H, 0, NULL, &Q);
1033 isl_mat_free(node->cmap);
1034 node->cmap = isl_mat_transpose(Q);
1035 node->rank = isl_mat_initial_non_zero_cols(H);
1038 if (!node->cmap || node->rank < 0)
1043 /* Count the number of equality and inequality constraints
1044 * that will be added for the given map.
1045 * If once is set, then we count
1046 * each edge exactly once. Otherwise, we count as follows
1047 * validity -> 1 (>= 0)
1048 * validity+proximity -> 2 (>= 0 and upper bound)
1049 * proximity -> 2 (lower and upper bound)
1051 static int count_map_constraints(struct isl_sched_graph *graph,
1052 struct isl_sched_edge *edge, __isl_take isl_map *map,
1053 int *n_eq, int *n_ineq, int once)
1055 isl_basic_set *coef;
1056 int f = once ? 1 : edge->proximity ? 2 : 1;
1058 if (edge->src == edge->dst)
1059 coef = intra_coefficients(graph, map);
1061 coef = inter_coefficients(graph, map);
1064 *n_eq += f * coef->n_eq;
1065 *n_ineq += f * coef->n_ineq;
1066 isl_basic_set_free(coef);
1071 /* Count the number of equality and inequality constraints
1072 * that will be added to the main lp problem.
1073 * If once is set, then we count
1074 * each edge exactly once. Otherwise, we count as follows
1075 * validity -> 1 (>= 0)
1076 * validity+proximity -> 2 (>= 0 and upper bound)
1077 * proximity -> 2 (lower and upper bound)
1079 static int count_constraints(struct isl_sched_graph *graph,
1080 int *n_eq, int *n_ineq, int once)
1084 *n_eq = *n_ineq = 0;
1085 for (i = 0; i < graph->n_edge; ++i) {
1086 struct isl_sched_edge *edge= &graph->edge[i];
1087 isl_map *map = isl_map_copy(edge->map);
1089 if (count_map_constraints(graph, edge, map,
1090 n_eq, n_ineq, once) < 0)
1097 /* Construct an ILP problem for finding schedule coefficients
1098 * that result in non-negative, but small dependence distances
1099 * over all dependences.
1100 * In particular, the dependence distances over proximity edges
1101 * are bounded by m_0 + m_n n and we compute schedule coefficients
1102 * with small values (preferably zero) of m_n and m_0.
1104 * All variables of the ILP are non-negative. The actual coefficients
1105 * may be negative, so each coefficient is represented as the difference
1106 * of two non-negative variables. The negative part always appears
1107 * immediately before the positive part.
1108 * Other than that, the variables have the following order
1110 * - sum of positive and negative parts of m_n coefficients
1112 * - sum of positive and negative parts of all c_n coefficients
1113 * (unconstrained when computing non-parametric schedules)
1114 * - sum of positive and negative parts of all c_x coefficients
1115 * - positive and negative parts of m_n coefficients
1118 * - positive and negative parts of c_i_n (if parametric)
1119 * - positive and negative parts of c_i_x
1121 * The c_i_x are not represented directly, but through the columns of
1122 * node->cmap. That is, the computed values are for variable t_i_x
1123 * such that c_i_x = Q t_i_x with Q equal to node->cmap.
1125 * The constraints are those from the edges plus two or three equalities
1126 * to express the sums.
1128 * If force_zero is set, then we add equalities to ensure that
1129 * the sum of the m_n coefficients and m_0 are both zero.
1131 static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph,
1142 int max_constant_term;
1144 max_constant_term = ctx->opt->schedule_max_constant_term;
1146 parametric = ctx->opt->schedule_parametric;
1147 nparam = isl_space_dim(graph->node[0].dim, isl_dim_param);
1149 total = param_pos + 2 * nparam;
1150 for (i = 0; i < graph->n; ++i) {
1151 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
1152 if (node_update_cmap(node) < 0)
1154 node->start = total;
1155 total += 1 + 2 * (node->nparam + node->nvar);
1158 if (count_constraints(graph, &n_eq, &n_ineq, 0) < 0)
1161 dim = isl_space_set_alloc(ctx, 0, total);
1162 isl_basic_set_free(graph->lp);
1163 n_eq += 2 + parametric + force_zero;
1164 if (max_constant_term != -1)
1167 graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq);
1169 k = isl_basic_set_alloc_equality(graph->lp);
1172 isl_seq_clr(graph->lp->eq[k], 1 + total);
1174 isl_int_set_si(graph->lp->eq[k][1], -1);
1175 for (i = 0; i < 2 * nparam; ++i)
1176 isl_int_set_si(graph->lp->eq[k][1 + param_pos + i], 1);
1179 k = isl_basic_set_alloc_equality(graph->lp);
1182 isl_seq_clr(graph->lp->eq[k], 1 + total);
1183 isl_int_set_si(graph->lp->eq[k][2], -1);
1187 k = isl_basic_set_alloc_equality(graph->lp);
1190 isl_seq_clr(graph->lp->eq[k], 1 + total);
1191 isl_int_set_si(graph->lp->eq[k][3], -1);
1192 for (i = 0; i < graph->n; ++i) {
1193 int pos = 1 + graph->node[i].start + 1;
1195 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
1196 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1200 k = isl_basic_set_alloc_equality(graph->lp);
1203 isl_seq_clr(graph->lp->eq[k], 1 + total);
1204 isl_int_set_si(graph->lp->eq[k][4], -1);
1205 for (i = 0; i < graph->n; ++i) {
1206 struct isl_sched_node *node = &graph->node[i];
1207 int pos = 1 + node->start + 1 + 2 * node->nparam;
1209 for (j = 0; j < 2 * node->nvar; ++j)
1210 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1213 if (max_constant_term != -1)
1214 for (i = 0; i < graph->n; ++i) {
1215 struct isl_sched_node *node = &graph->node[i];
1216 k = isl_basic_set_alloc_inequality(graph->lp);
1219 isl_seq_clr(graph->lp->ineq[k], 1 + total);
1220 isl_int_set_si(graph->lp->ineq[k][1 + node->start], -1);
1221 isl_int_set_si(graph->lp->ineq[k][0], max_constant_term);
1224 if (add_all_validity_constraints(graph) < 0)
1226 if (add_all_proximity_constraints(graph) < 0)
1232 /* Analyze the conflicting constraint found by
1233 * isl_tab_basic_set_non_trivial_lexmin. If it corresponds to the validity
1234 * constraint of one of the edges between distinct nodes, living, moreover
1235 * in distinct SCCs, then record the source and sink SCC as this may
1236 * be a good place to cut between SCCs.
1238 static int check_conflict(int con, void *user)
1241 struct isl_sched_graph *graph = user;
1243 if (graph->src_scc >= 0)
1246 con -= graph->lp->n_eq;
1248 if (con >= graph->lp->n_ineq)
1251 for (i = 0; i < graph->n_edge; ++i) {
1252 if (!graph->edge[i].validity)
1254 if (graph->edge[i].src == graph->edge[i].dst)
1256 if (graph->edge[i].src->scc == graph->edge[i].dst->scc)
1258 if (graph->edge[i].start > con)
1260 if (graph->edge[i].end <= con)
1262 graph->src_scc = graph->edge[i].src->scc;
1263 graph->dst_scc = graph->edge[i].dst->scc;
1269 /* Check whether the next schedule row of the given node needs to be
1270 * non-trivial. Lower-dimensional domains may have some trivial rows,
1271 * but as soon as the number of remaining required non-trivial rows
1272 * is as large as the number or remaining rows to be computed,
1273 * all remaining rows need to be non-trivial.
1275 static int needs_row(struct isl_sched_graph *graph, struct isl_sched_node *node)
1277 return node->nvar - node->rank >= graph->maxvar - graph->n_row;
1280 /* Solve the ILP problem constructed in setup_lp.
1281 * For each node such that all the remaining rows of its schedule
1282 * need to be non-trivial, we construct a non-triviality region.
1283 * This region imposes that the next row is independent of previous rows.
1284 * In particular the coefficients c_i_x are represented by t_i_x
1285 * variables with c_i_x = Q t_i_x and Q a unimodular matrix such that
1286 * its first columns span the rows of the previously computed part
1287 * of the schedule. The non-triviality region enforces that at least
1288 * one of the remaining components of t_i_x is non-zero, i.e.,
1289 * that the new schedule row depends on at least one of the remaining
1292 static __isl_give isl_vec *solve_lp(struct isl_sched_graph *graph)
1298 for (i = 0; i < graph->n; ++i) {
1299 struct isl_sched_node *node = &graph->node[i];
1300 int skip = node->rank;
1301 graph->region[i].pos = node->start + 1 + 2*(node->nparam+skip);
1302 if (needs_row(graph, node))
1303 graph->region[i].len = 2 * (node->nvar - skip);
1305 graph->region[i].len = 0;
1307 lp = isl_basic_set_copy(graph->lp);
1308 sol = isl_tab_basic_set_non_trivial_lexmin(lp, 2, graph->n,
1309 graph->region, &check_conflict, graph);
1313 /* Update the schedules of all nodes based on the given solution
1314 * of the LP problem.
1315 * The new row is added to the current band.
1316 * All possibly negative coefficients are encoded as a difference
1317 * of two non-negative variables, so we need to perform the subtraction
1318 * here. Moreover, if use_cmap is set, then the solution does
1319 * not refer to the actual coefficients c_i_x, but instead to variables
1320 * t_i_x such that c_i_x = Q t_i_x and Q is equal to node->cmap.
1321 * In this case, we then also need to perform this multiplication
1322 * to obtain the values of c_i_x.
1324 * If check_zero is set, then the first two coordinates of sol are
1325 * assumed to correspond to the dependence distance. If these two
1326 * coordinates are zero, then the corresponding scheduling dimension
1327 * is marked as being zero distance.
1329 static int update_schedule(struct isl_sched_graph *graph,
1330 __isl_take isl_vec *sol, int use_cmap, int check_zero)
1334 isl_vec *csol = NULL;
1339 isl_die(sol->ctx, isl_error_internal,
1340 "no solution found", goto error);
1343 zero = isl_int_is_zero(sol->el[1]) &&
1344 isl_int_is_zero(sol->el[2]);
1346 for (i = 0; i < graph->n; ++i) {
1347 struct isl_sched_node *node = &graph->node[i];
1348 int pos = node->start;
1349 int row = isl_mat_rows(node->sched);
1352 csol = isl_vec_alloc(sol->ctx, node->nvar);
1356 isl_map_free(node->sched_map);
1357 node->sched_map = NULL;
1358 node->sched = isl_mat_add_rows(node->sched, 1);
1361 node->sched = isl_mat_set_element(node->sched, row, 0,
1363 for (j = 0; j < node->nparam + node->nvar; ++j)
1364 isl_int_sub(sol->el[1 + pos + 1 + 2 * j + 1],
1365 sol->el[1 + pos + 1 + 2 * j + 1],
1366 sol->el[1 + pos + 1 + 2 * j]);
1367 for (j = 0; j < node->nparam; ++j)
1368 node->sched = isl_mat_set_element(node->sched,
1369 row, 1 + j, sol->el[1+pos+1+2*j+1]);
1370 for (j = 0; j < node->nvar; ++j)
1371 isl_int_set(csol->el[j],
1372 sol->el[1+pos+1+2*(node->nparam+j)+1]);
1374 csol = isl_mat_vec_product(isl_mat_copy(node->cmap),
1378 for (j = 0; j < node->nvar; ++j)
1379 node->sched = isl_mat_set_element(node->sched,
1380 row, 1 + node->nparam + j, csol->el[j]);
1381 node->band[graph->n_total_row] = graph->n_band;
1382 node->zero[graph->n_total_row] = zero;
1388 graph->n_total_row++;
1397 /* Convert node->sched into a map and return this map.
1398 * We simply add equality constraints that express each output variable
1399 * as the affine combination of parameters and input variables specified
1400 * by the schedule matrix.
1402 * The result is cached in node->sched_map, which needs to be released
1403 * whenever node->sched is updated.
1405 static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node)
1409 isl_local_space *ls;
1410 isl_basic_map *bmap;
1415 if (node->sched_map)
1416 return isl_map_copy(node->sched_map);
1418 nrow = isl_mat_rows(node->sched);
1419 ncol = isl_mat_cols(node->sched) - 1;
1420 dim = isl_space_from_domain(isl_space_copy(node->dim));
1421 dim = isl_space_add_dims(dim, isl_dim_out, nrow);
1422 bmap = isl_basic_map_universe(isl_space_copy(dim));
1423 ls = isl_local_space_from_space(dim);
1427 for (i = 0; i < nrow; ++i) {
1428 c = isl_equality_alloc(isl_local_space_copy(ls));
1429 isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
1430 isl_mat_get_element(node->sched, i, 0, &v);
1431 isl_constraint_set_constant(c, v);
1432 for (j = 0; j < node->nparam; ++j) {
1433 isl_mat_get_element(node->sched, i, 1 + j, &v);
1434 isl_constraint_set_coefficient(c, isl_dim_param, j, v);
1436 for (j = 0; j < node->nvar; ++j) {
1437 isl_mat_get_element(node->sched,
1438 i, 1 + node->nparam + j, &v);
1439 isl_constraint_set_coefficient(c, isl_dim_in, j, v);
1441 bmap = isl_basic_map_add_constraint(bmap, c);
1446 isl_local_space_free(ls);
1448 node->sched_map = isl_map_from_basic_map(bmap);
1449 return isl_map_copy(node->sched_map);
1452 /* Update the given dependence relation based on the current schedule.
1453 * That is, intersect the dependence relation with a map expressing
1454 * that source and sink are executed within the same iteration of
1455 * the current schedule.
1456 * This is not the most efficient way, but this shouldn't be a critical
1459 static __isl_give isl_map *specialize(__isl_take isl_map *map,
1460 struct isl_sched_node *src, struct isl_sched_node *dst)
1462 isl_map *src_sched, *dst_sched, *id;
1464 src_sched = node_extract_schedule(src);
1465 dst_sched = node_extract_schedule(dst);
1466 id = isl_map_apply_range(src_sched, isl_map_reverse(dst_sched));
1467 return isl_map_intersect(map, id);
1470 /* Update the dependence relations of all edges based on the current schedule.
1471 * If a dependence is carried completely by the current schedule, then
1472 * it is removed and edge_table is updated accordingly.
1474 static int update_edges(isl_ctx *ctx, struct isl_sched_graph *graph)
1477 int reset_table = 0;
1479 for (i = graph->n_edge - 1; i >= 0; --i) {
1480 struct isl_sched_edge *edge = &graph->edge[i];
1481 edge->map = specialize(edge->map, edge->src, edge->dst);
1485 if (isl_map_plain_is_empty(edge->map)) {
1487 isl_map_free(edge->map);
1488 if (i != graph->n_edge - 1)
1489 graph->edge[i] = graph->edge[graph->n_edge - 1];
1495 isl_hash_table_free(ctx, graph->edge_table);
1496 graph->edge_table = NULL;
1497 return graph_init_edge_table(ctx, graph);
1503 static void next_band(struct isl_sched_graph *graph)
1505 graph->band_start = graph->n_total_row;
1509 /* Topologically sort statements mapped to same schedule iteration
1510 * and add a row to the schedule corresponding to this order.
1512 static int sort_statements(isl_ctx *ctx, struct isl_sched_graph *graph)
1519 if (update_edges(ctx, graph) < 0)
1522 if (graph->n_edge == 0)
1525 if (detect_sccs(graph) < 0)
1528 for (i = 0; i < graph->n; ++i) {
1529 struct isl_sched_node *node = &graph->node[i];
1530 int row = isl_mat_rows(node->sched);
1531 int cols = isl_mat_cols(node->sched);
1533 isl_map_free(node->sched_map);
1534 node->sched_map = NULL;
1535 node->sched = isl_mat_add_rows(node->sched, 1);
1538 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1540 for (j = 1; j < cols; ++j)
1541 node->sched = isl_mat_set_element_si(node->sched,
1543 node->band[graph->n_total_row] = graph->n_band;
1546 graph->n_total_row++;
1552 /* Construct an isl_schedule based on the computed schedule stored
1553 * in graph and with parameters specified by dim.
1555 static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph,
1556 __isl_take isl_space *dim)
1560 isl_schedule *sched = NULL;
1565 ctx = isl_space_get_ctx(dim);
1566 sched = isl_calloc(ctx, struct isl_schedule,
1567 sizeof(struct isl_schedule) +
1568 (graph->n - 1) * sizeof(struct isl_schedule_node));
1573 sched->n = graph->n;
1574 sched->n_band = graph->n_band;
1575 sched->n_total_row = graph->n_total_row;
1577 for (i = 0; i < sched->n; ++i) {
1579 int *band_end, *band_id, *zero;
1581 band_end = isl_alloc_array(ctx, int, graph->n_band);
1582 band_id = isl_alloc_array(ctx, int, graph->n_band);
1583 zero = isl_alloc_array(ctx, int, graph->n_total_row);
1584 sched->node[i].sched = node_extract_schedule(&graph->node[i]);
1585 sched->node[i].band_end = band_end;
1586 sched->node[i].band_id = band_id;
1587 sched->node[i].zero = zero;
1588 if (!band_end || !band_id || !zero)
1591 for (r = 0; r < graph->n_total_row; ++r)
1592 zero[r] = graph->node[i].zero[r];
1593 for (r = b = 0; r < graph->n_total_row; ++r) {
1594 if (graph->node[i].band[r] == b)
1597 if (graph->node[i].band[r] == -1)
1600 if (r == graph->n_total_row)
1602 sched->node[i].n_band = b;
1603 for (--b; b >= 0; --b)
1604 band_id[b] = graph->node[i].band_id[b];
1611 isl_space_free(dim);
1612 isl_schedule_free(sched);
1616 /* Copy nodes that satisfy node_pred from the src dependence graph
1617 * to the dst dependence graph.
1619 static int copy_nodes(struct isl_sched_graph *dst, struct isl_sched_graph *src,
1620 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1625 for (i = 0; i < src->n; ++i) {
1626 if (!node_pred(&src->node[i], data))
1628 dst->node[dst->n].dim = isl_space_copy(src->node[i].dim);
1629 dst->node[dst->n].nvar = src->node[i].nvar;
1630 dst->node[dst->n].nparam = src->node[i].nparam;
1631 dst->node[dst->n].sched = isl_mat_copy(src->node[i].sched);
1632 dst->node[dst->n].sched_map =
1633 isl_map_copy(src->node[i].sched_map);
1634 dst->node[dst->n].band = src->node[i].band;
1635 dst->node[dst->n].band_id = src->node[i].band_id;
1636 dst->node[dst->n].zero = src->node[i].zero;
1643 /* Copy non-empty edges that satisfy edge_pred from the src dependence graph
1644 * to the dst dependence graph.
1645 * If the source or destination node of the edge is not in the destination
1646 * graph, then it must be a backward proximity edge and it should simply
1649 static int copy_edges(isl_ctx *ctx, struct isl_sched_graph *dst,
1650 struct isl_sched_graph *src,
1651 int (*edge_pred)(struct isl_sched_edge *edge, int data), int data)
1656 for (i = 0; i < src->n_edge; ++i) {
1657 struct isl_sched_edge *edge = &src->edge[i];
1659 struct isl_sched_node *dst_src, *dst_dst;
1661 if (!edge_pred(edge, data))
1664 if (isl_map_plain_is_empty(edge->map))
1667 dst_src = graph_find_node(ctx, dst, edge->src->dim);
1668 dst_dst = graph_find_node(ctx, dst, edge->dst->dim);
1669 if (!dst_src || !dst_dst) {
1671 isl_die(ctx, isl_error_internal,
1672 "backward validity edge", return -1);
1676 map = isl_map_copy(edge->map);
1678 dst->edge[dst->n_edge].src = dst_src;
1679 dst->edge[dst->n_edge].dst = dst_dst;
1680 dst->edge[dst->n_edge].map = map;
1681 dst->edge[dst->n_edge].validity = edge->validity;
1682 dst->edge[dst->n_edge].proximity = edge->proximity;
1689 /* Given a "src" dependence graph that contains the nodes from "dst"
1690 * that satisfy node_pred, copy the schedule computed in "src"
1691 * for those nodes back to "dst".
1693 static int copy_schedule(struct isl_sched_graph *dst,
1694 struct isl_sched_graph *src,
1695 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1700 for (i = 0; i < dst->n; ++i) {
1701 if (!node_pred(&dst->node[i], data))
1703 isl_mat_free(dst->node[i].sched);
1704 isl_map_free(dst->node[i].sched_map);
1705 dst->node[i].sched = isl_mat_copy(src->node[src->n].sched);
1706 dst->node[i].sched_map =
1707 isl_map_copy(src->node[src->n].sched_map);
1711 dst->n_total_row = src->n_total_row;
1712 dst->n_band = src->n_band;
1717 /* Compute the maximal number of variables over all nodes.
1718 * This is the maximal number of linearly independent schedule
1719 * rows that we need to compute.
1720 * Just in case we end up in a part of the dependence graph
1721 * with only lower-dimensional domains, we make sure we will
1722 * compute the required amount of extra linearly independent rows.
1724 static int compute_maxvar(struct isl_sched_graph *graph)
1729 for (i = 0; i < graph->n; ++i) {
1730 struct isl_sched_node *node = &graph->node[i];
1733 if (node_update_cmap(node) < 0)
1735 nvar = node->nvar + graph->n_row - node->rank;
1736 if (nvar > graph->maxvar)
1737 graph->maxvar = nvar;
1743 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph);
1744 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph);
1746 /* Compute a schedule for a subgraph of "graph". In particular, for
1747 * the graph composed of nodes that satisfy node_pred and edges that
1748 * that satisfy edge_pred. The caller should precompute the number
1749 * of nodes and edges that satisfy these predicates and pass them along
1750 * as "n" and "n_edge".
1751 * If the subgraph is known to consist of a single component, then wcc should
1752 * be set and then we call compute_schedule_wcc on the constructed subgraph.
1753 * Otherwise, we call compute_schedule, which will check whether the subgraph
1756 static int compute_sub_schedule(isl_ctx *ctx,
1757 struct isl_sched_graph *graph, int n, int n_edge,
1758 int (*node_pred)(struct isl_sched_node *node, int data),
1759 int (*edge_pred)(struct isl_sched_edge *edge, int data),
1762 struct isl_sched_graph split = { 0 };
1764 if (graph_alloc(ctx, &split, n, n_edge) < 0)
1766 if (copy_nodes(&split, graph, node_pred, data) < 0)
1768 if (graph_init_table(ctx, &split) < 0)
1770 if (copy_edges(ctx, &split, graph, edge_pred, data) < 0)
1772 if (graph_init_edge_table(ctx, &split) < 0)
1774 split.n_row = graph->n_row;
1775 split.n_total_row = graph->n_total_row;
1776 split.n_band = graph->n_band;
1777 split.band_start = graph->band_start;
1779 if (wcc && compute_schedule_wcc(ctx, &split) < 0)
1781 if (!wcc && compute_schedule(ctx, &split) < 0)
1784 copy_schedule(graph, &split, node_pred, data);
1786 graph_free(ctx, &split);
1789 graph_free(ctx, &split);
1793 static int node_scc_exactly(struct isl_sched_node *node, int scc)
1795 return node->scc == scc;
1798 static int node_scc_at_most(struct isl_sched_node *node, int scc)
1800 return node->scc <= scc;
1803 static int node_scc_at_least(struct isl_sched_node *node, int scc)
1805 return node->scc >= scc;
1808 static int edge_src_scc_exactly(struct isl_sched_edge *edge, int scc)
1810 return edge->src->scc == scc;
1813 static int edge_dst_scc_at_most(struct isl_sched_edge *edge, int scc)
1815 return edge->dst->scc <= scc;
1818 static int edge_src_scc_at_least(struct isl_sched_edge *edge, int scc)
1820 return edge->src->scc >= scc;
1823 /* Pad the schedules of all nodes with zero rows such that in the end
1824 * they all have graph->n_total_row rows.
1825 * The extra rows don't belong to any band, so they get assigned band number -1.
1827 static int pad_schedule(struct isl_sched_graph *graph)
1831 for (i = 0; i < graph->n; ++i) {
1832 struct isl_sched_node *node = &graph->node[i];
1833 int row = isl_mat_rows(node->sched);
1834 if (graph->n_total_row > row) {
1835 isl_map_free(node->sched_map);
1836 node->sched_map = NULL;
1838 node->sched = isl_mat_add_zero_rows(node->sched,
1839 graph->n_total_row - row);
1842 for (j = row; j < graph->n_total_row; ++j)
1849 /* Split the current graph into two parts and compute a schedule for each
1850 * part individually. In particular, one part consists of all SCCs up
1851 * to and including graph->src_scc, while the other part contains the other
1854 * The split is enforced in the schedule by constant rows with two different
1855 * values (0 and 1). These constant rows replace the previously computed rows
1856 * in the current band.
1857 * It would be possible to reuse them as the first rows in the next
1858 * band, but recomputing them may result in better rows as we are looking
1859 * at a smaller part of the dependence graph.
1860 * compute_split_schedule is only called when no zero-distance schedule row
1861 * could be found on the entire graph, so we wark the splitting row as
1862 * non zero-distance.
1864 * The band_id of the second group is set to n, where n is the number
1865 * of nodes in the first group. This ensures that the band_ids over
1866 * the two groups remain disjoint, even if either or both of the two
1867 * groups contain independent components.
1869 static int compute_split_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
1871 int i, j, n, e1, e2;
1872 int n_total_row, orig_total_row;
1873 int n_band, orig_band;
1876 drop = graph->n_total_row - graph->band_start;
1877 graph->n_total_row -= drop;
1878 graph->n_row -= drop;
1881 for (i = 0; i < graph->n; ++i) {
1882 struct isl_sched_node *node = &graph->node[i];
1883 int row = isl_mat_rows(node->sched) - drop;
1884 int cols = isl_mat_cols(node->sched);
1885 int before = node->scc <= graph->src_scc;
1890 isl_map_free(node->sched_map);
1891 node->sched_map = NULL;
1892 node->sched = isl_mat_drop_rows(node->sched,
1893 graph->band_start, drop);
1894 node->sched = isl_mat_add_rows(node->sched, 1);
1897 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1899 for (j = 1; j < cols; ++j)
1900 node->sched = isl_mat_set_element_si(node->sched,
1902 node->band[graph->n_total_row] = graph->n_band;
1903 node->zero[graph->n_total_row] = 0;
1907 for (i = 0; i < graph->n_edge; ++i) {
1908 if (graph->edge[i].dst->scc <= graph->src_scc)
1910 if (graph->edge[i].src->scc > graph->src_scc)
1914 graph->n_total_row++;
1917 for (i = 0; i < graph->n; ++i) {
1918 struct isl_sched_node *node = &graph->node[i];
1919 if (node->scc > graph->src_scc)
1920 node->band_id[graph->n_band] = n;
1923 orig_total_row = graph->n_total_row;
1924 orig_band = graph->n_band;
1925 if (compute_sub_schedule(ctx, graph, n, e1,
1926 &node_scc_at_most, &edge_dst_scc_at_most,
1927 graph->src_scc, 0) < 0)
1929 n_total_row = graph->n_total_row;
1930 graph->n_total_row = orig_total_row;
1931 n_band = graph->n_band;
1932 graph->n_band = orig_band;
1933 if (compute_sub_schedule(ctx, graph, graph->n - n, e2,
1934 &node_scc_at_least, &edge_src_scc_at_least,
1935 graph->src_scc + 1, 0) < 0)
1937 if (n_total_row > graph->n_total_row)
1938 graph->n_total_row = n_total_row;
1939 if (n_band > graph->n_band)
1940 graph->n_band = n_band;
1942 return pad_schedule(graph);
1945 /* Compute the next band of the schedule after updating the dependence
1946 * relations based on the the current schedule.
1948 static int compute_next_band(isl_ctx *ctx, struct isl_sched_graph *graph)
1950 if (update_edges(ctx, graph) < 0)
1954 return compute_schedule(ctx, graph);
1957 /* Add constraints to graph->lp that force the dependence "map" (which
1958 * is part of the dependence relation of "edge")
1959 * to be respected and attempt to carry it, where the edge is one from
1960 * a node j to itself. "pos" is the sequence number of the given map.
1961 * That is, add constraints that enforce
1963 * (c_j_0 + c_j_n n + c_j_x y) - (c_j_0 + c_j_n n + c_j_x x)
1964 * = c_j_x (y - x) >= e_i
1966 * for each (x,y) in R.
1967 * We obtain general constraints on coefficients (c_0, c_n, c_x)
1968 * of valid constraints for (y - x) and then plug in (-e_i, 0, c_j_x),
1969 * with each coefficient in c_j_x represented as a pair of non-negative
1972 static int add_intra_constraints(struct isl_sched_graph *graph,
1973 struct isl_sched_edge *edge, __isl_take isl_map *map, int pos)
1976 isl_ctx *ctx = isl_map_get_ctx(map);
1978 isl_dim_map *dim_map;
1979 isl_basic_set *coef;
1980 struct isl_sched_node *node = edge->src;
1982 coef = intra_coefficients(graph, map);
1984 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
1986 total = isl_basic_set_total_dim(graph->lp);
1987 dim_map = isl_dim_map_alloc(ctx, total);
1988 isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
1989 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
1990 isl_space_dim(dim, isl_dim_set), 1,
1992 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
1993 isl_space_dim(dim, isl_dim_set), 1,
1995 graph->lp = isl_basic_set_extend_constraints(graph->lp,
1996 coef->n_eq, coef->n_ineq);
1997 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
1999 isl_space_free(dim);
2004 /* Add constraints to graph->lp that force the dependence "map" (which
2005 * is part of the dependence relation of "edge")
2006 * to be respected and attempt to carry it, where the edge is one from
2007 * node j to node k. "pos" is the sequence number of the given map.
2008 * That is, add constraints that enforce
2010 * (c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i
2012 * for each (x,y) in R.
2013 * We obtain general constraints on coefficients (c_0, c_n, c_x)
2014 * of valid constraints for R and then plug in
2015 * (-e_i + c_k_0 - c_j_0, c_k_n - c_j_n, c_k_x - c_j_x)
2016 * with each coefficient (except e_i, c_k_0 and c_j_0)
2017 * represented as a pair of non-negative coefficients.
2019 static int add_inter_constraints(struct isl_sched_graph *graph,
2020 struct isl_sched_edge *edge, __isl_take isl_map *map, int pos)
2023 isl_ctx *ctx = isl_map_get_ctx(map);
2025 isl_dim_map *dim_map;
2026 isl_basic_set *coef;
2027 struct isl_sched_node *src = edge->src;
2028 struct isl_sched_node *dst = edge->dst;
2030 coef = inter_coefficients(graph, map);
2032 dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef)));
2034 total = isl_basic_set_total_dim(graph->lp);
2035 dim_map = isl_dim_map_alloc(ctx, total);
2037 isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
2039 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
2040 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
2041 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
2042 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
2043 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
2045 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
2046 isl_space_dim(dim, isl_dim_set) + src->nvar, 1,
2049 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
2050 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
2051 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
2052 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
2053 isl_space_dim(dim, isl_dim_set), 1,
2055 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
2056 isl_space_dim(dim, isl_dim_set), 1,
2059 graph->lp = isl_basic_set_extend_constraints(graph->lp,
2060 coef->n_eq, coef->n_ineq);
2061 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
2063 isl_space_free(dim);
2068 /* Add constraints to graph->lp that force all dependence
2069 * to be respected and attempt to carry it.
2071 static int add_all_constraints(struct isl_sched_graph *graph)
2077 for (i = 0; i < graph->n_edge; ++i) {
2078 struct isl_sched_edge *edge= &graph->edge[i];
2079 for (j = 0; j < edge->map->n; ++j) {
2080 isl_basic_map *bmap;
2083 bmap = isl_basic_map_copy(edge->map->p[j]);
2084 map = isl_map_from_basic_map(bmap);
2086 if (edge->src == edge->dst &&
2087 add_intra_constraints(graph, edge, map, pos) < 0)
2089 if (edge->src != edge->dst &&
2090 add_inter_constraints(graph, edge, map, pos) < 0)
2099 /* Count the number of equality and inequality constraints
2100 * that will be added to the carry_lp problem.
2101 * If once is set, then we count
2102 * each edge exactly once. Otherwise, we count as follows
2103 * validity -> 1 (>= 0)
2104 * validity+proximity -> 2 (>= 0 and upper bound)
2105 * proximity -> 2 (lower and upper bound)
2107 static int count_all_constraints(struct isl_sched_graph *graph,
2108 int *n_eq, int *n_ineq, int once)
2112 *n_eq = *n_ineq = 0;
2113 for (i = 0; i < graph->n_edge; ++i) {
2114 struct isl_sched_edge *edge= &graph->edge[i];
2115 for (j = 0; j < edge->map->n; ++j) {
2116 isl_basic_map *bmap;
2119 bmap = isl_basic_map_copy(edge->map->p[j]);
2120 map = isl_map_from_basic_map(bmap);
2122 if (count_map_constraints(graph, edge, map,
2123 n_eq, n_ineq, once) < 0)
2131 /* Construct an LP problem for finding schedule coefficients
2132 * such that the schedule carries as many dependences as possible.
2133 * In particular, for each dependence i, we bound the dependence distance
2134 * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum
2135 * of all e_i's. Dependence with e_i = 0 in the solution are simply
2136 * respected, while those with e_i > 0 (in practice e_i = 1) are carried.
2137 * Note that if the dependence relation is a union of basic maps,
2138 * then we have to consider each basic map individually as it may only
2139 * be possible to carry the dependences expressed by some of those
2140 * basic maps and not all off them.
2141 * Below, we consider each of those basic maps as a separate "edge".
2143 * All variables of the LP are non-negative. The actual coefficients
2144 * may be negative, so each coefficient is represented as the difference
2145 * of two non-negative variables. The negative part always appears
2146 * immediately before the positive part.
2147 * Other than that, the variables have the following order
2149 * - sum of (1 - e_i) over all edges
2150 * - sum of positive and negative parts of all c_n coefficients
2151 * (unconstrained when computing non-parametric schedules)
2152 * - sum of positive and negative parts of all c_x coefficients
2157 * - positive and negative parts of c_i_n (if parametric)
2158 * - positive and negative parts of c_i_x
2160 * The constraints are those from the edges plus three equalities
2161 * to express the sums and n_edge inequalities to express e_i <= 1.
2163 static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
2173 for (i = 0; i < graph->n_edge; ++i)
2174 n_edge += graph->edge[i].map->n;
2177 for (i = 0; i < graph->n; ++i) {
2178 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
2179 node->start = total;
2180 total += 1 + 2 * (node->nparam + node->nvar);
2183 if (count_all_constraints(graph, &n_eq, &n_ineq, 1) < 0)
2186 dim = isl_space_set_alloc(ctx, 0, total);
2187 isl_basic_set_free(graph->lp);
2190 graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq);
2191 graph->lp = isl_basic_set_set_rational(graph->lp);
2193 k = isl_basic_set_alloc_equality(graph->lp);
2196 isl_seq_clr(graph->lp->eq[k], 1 + total);
2197 isl_int_set_si(graph->lp->eq[k][0], -n_edge);
2198 isl_int_set_si(graph->lp->eq[k][1], 1);
2199 for (i = 0; i < n_edge; ++i)
2200 isl_int_set_si(graph->lp->eq[k][4 + i], 1);
2202 k = isl_basic_set_alloc_equality(graph->lp);
2205 isl_seq_clr(graph->lp->eq[k], 1 + total);
2206 isl_int_set_si(graph->lp->eq[k][2], -1);
2207 for (i = 0; i < graph->n; ++i) {
2208 int pos = 1 + graph->node[i].start + 1;
2210 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
2211 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2214 k = isl_basic_set_alloc_equality(graph->lp);
2217 isl_seq_clr(graph->lp->eq[k], 1 + total);
2218 isl_int_set_si(graph->lp->eq[k][3], -1);
2219 for (i = 0; i < graph->n; ++i) {
2220 struct isl_sched_node *node = &graph->node[i];
2221 int pos = 1 + node->start + 1 + 2 * node->nparam;
2223 for (j = 0; j < 2 * node->nvar; ++j)
2224 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2227 for (i = 0; i < n_edge; ++i) {
2228 k = isl_basic_set_alloc_inequality(graph->lp);
2231 isl_seq_clr(graph->lp->ineq[k], 1 + total);
2232 isl_int_set_si(graph->lp->ineq[k][4 + i], -1);
2233 isl_int_set_si(graph->lp->ineq[k][0], 1);
2236 if (add_all_constraints(graph) < 0)
2242 /* If the schedule_split_parallel option is set and if the linear
2243 * parts of the scheduling rows for all nodes in the graphs are the same,
2244 * then split off the constant term from the linear part.
2245 * The constant term is then placed in a separate band and
2246 * the linear part is simplified.
2248 static int split_parallel(isl_ctx *ctx, struct isl_sched_graph *graph)
2253 struct isl_sched_node *node0;
2255 if (!ctx->opt->schedule_split_parallel)
2260 node0 = &graph->node[0];
2261 row = isl_mat_rows(node0->sched) - 1;
2262 cols = isl_mat_cols(node0->sched);
2263 for (i = 1; i < graph->n; ++i) {
2264 struct isl_sched_node *node = &graph->node[i];
2266 if (isl_mat_cols(node->sched) != cols)
2268 if (!isl_seq_eq(node0->sched->row[row] + 1,
2269 node->sched->row[row] + 1, cols - 1))
2272 isl_int_ne(node0->sched->row[row][0],
2273 node->sched->row[row][0]))
2281 for (i = 0; i < graph->n; ++i) {
2282 struct isl_sched_node *node = &graph->node[i];
2284 isl_map_free(node->sched_map);
2285 node->sched_map = NULL;
2286 node->sched = isl_mat_add_zero_rows(node->sched, 1);
2289 isl_int_set(node->sched->row[row + 1][0],
2290 node->sched->row[row][0]);
2291 isl_int_set_si(node->sched->row[row][0], 0);
2292 node->sched = isl_mat_normalize_row(node->sched, row);
2295 node->band[graph->n_total_row] = graph->n_band;
2298 graph->n_total_row++;
2303 /* Construct a schedule row for each node such that as many dependences
2304 * as possible are carried and then continue with the next band.
2306 static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *graph)
2314 for (i = 0; i < graph->n_edge; ++i)
2315 n_edge += graph->edge[i].map->n;
2317 if (setup_carry_lp(ctx, graph) < 0)
2320 lp = isl_basic_set_copy(graph->lp);
2321 sol = isl_tab_basic_set_non_neg_lexmin(lp);
2325 if (sol->size == 0) {
2327 isl_die(ctx, isl_error_internal,
2328 "error in schedule construction", return -1);
2331 if (isl_int_cmp_si(sol->el[1], n_edge) >= 0) {
2333 isl_die(ctx, isl_error_unknown,
2334 "unable to carry dependences", return -1);
2337 if (update_schedule(graph, sol, 0, 0) < 0)
2340 if (split_parallel(ctx, graph) < 0)
2343 return compute_next_band(ctx, graph);
2346 /* Are there any validity edges in the graph?
2348 static int has_validity_edges(struct isl_sched_graph *graph)
2352 for (i = 0; i < graph->n_edge; ++i)
2353 if (graph->edge[i].validity)
2359 /* Should we apply a Feautrier step?
2360 * That is, did the user request the Feautrier algorithm and are
2361 * there any validity dependences (left)?
2363 static int need_feautrier_step(isl_ctx *ctx, struct isl_sched_graph *graph)
2365 if (ctx->opt->schedule_algorithm != ISL_SCHEDULE_ALGORITHM_FEAUTRIER)
2368 return has_validity_edges(graph);
2371 /* Compute a schedule for a connected dependence graph using Feautrier's
2372 * multi-dimensional scheduling algorithm.
2373 * The original algorithm is described in [1].
2374 * The main idea is to minimize the number of scheduling dimensions, by
2375 * trying to satisfy as many dependences as possible per scheduling dimension.
2377 * [1] P. Feautrier, Some Efficient Solutions to the Affine Scheduling
2378 * Problem, Part II: Multi-Dimensional Time.
2379 * In Intl. Journal of Parallel Programming, 1992.
2381 static int compute_schedule_wcc_feautrier(isl_ctx *ctx,
2382 struct isl_sched_graph *graph)
2384 return carry_dependences(ctx, graph);
2387 /* Compute a schedule for a connected dependence graph.
2388 * We try to find a sequence of as many schedule rows as possible that result
2389 * in non-negative dependence distances (independent of the previous rows
2390 * in the sequence, i.e., such that the sequence is tilable).
2391 * If we can't find any more rows we either
2392 * - split between SCCs and start over (assuming we found an interesting
2393 * pair of SCCs between which to split)
2394 * - continue with the next band (assuming the current band has at least
2396 * - try to carry as many dependences as possible and continue with the next
2399 * If Feautrier's algorithm is selected, we first recursively try to satisfy
2400 * as many validity dependences as possible. When all validity dependences
2401 * are satisfied we extend the schedule to a full-dimensional schedule.
2403 * If we manage to complete the schedule, we finish off by topologically
2404 * sorting the statements based on the remaining dependences.
2406 * If ctx->opt->schedule_outer_zero_distance is set, then we force the
2407 * outermost dimension in the current band to be zero distance. If this
2408 * turns out to be impossible, we fall back on the general scheme above
2409 * and try to carry as many dependences as possible.
2411 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph)
2415 if (detect_sccs(graph) < 0)
2419 if (compute_maxvar(graph) < 0)
2422 if (need_feautrier_step(ctx, graph))
2423 return compute_schedule_wcc_feautrier(ctx, graph);
2425 if (ctx->opt->schedule_outer_zero_distance)
2428 while (graph->n_row < graph->maxvar) {
2431 graph->src_scc = -1;
2432 graph->dst_scc = -1;
2434 if (setup_lp(ctx, graph, force_zero) < 0)
2436 sol = solve_lp(graph);
2439 if (sol->size == 0) {
2441 if (!ctx->opt->schedule_maximize_band_depth &&
2442 graph->n_total_row > graph->band_start)
2443 return compute_next_band(ctx, graph);
2444 if (graph->src_scc >= 0)
2445 return compute_split_schedule(ctx, graph);
2446 if (graph->n_total_row > graph->band_start)
2447 return compute_next_band(ctx, graph);
2448 return carry_dependences(ctx, graph);
2450 if (update_schedule(graph, sol, 1, 1) < 0)
2455 if (graph->n_total_row > graph->band_start)
2457 return sort_statements(ctx, graph);
2460 /* Compute a schedule for each component (identified by node->scc)
2461 * of the dependence graph separately and then combine the results.
2463 * The band_id is adjusted such that each component has a separate id.
2464 * Note that the band_id may have already been set to a value different
2465 * from zero by compute_split_schedule.
2467 static int compute_component_schedule(isl_ctx *ctx,
2468 struct isl_sched_graph *graph)
2472 int n_total_row, orig_total_row;
2473 int n_band, orig_band;
2476 orig_total_row = graph->n_total_row;
2478 orig_band = graph->n_band;
2479 for (i = 0; i < graph->n; ++i)
2480 graph->node[i].band_id[graph->n_band] += graph->node[i].scc;
2481 for (wcc = 0; wcc < graph->scc; ++wcc) {
2483 for (i = 0; i < graph->n; ++i)
2484 if (graph->node[i].scc == wcc)
2487 for (i = 0; i < graph->n_edge; ++i)
2488 if (graph->edge[i].src->scc == wcc)
2491 if (compute_sub_schedule(ctx, graph, n, n_edge,
2493 &edge_src_scc_exactly, wcc, 1) < 0)
2495 if (graph->n_total_row > n_total_row)
2496 n_total_row = graph->n_total_row;
2497 graph->n_total_row = orig_total_row;
2498 if (graph->n_band > n_band)
2499 n_band = graph->n_band;
2500 graph->n_band = orig_band;
2503 graph->n_total_row = n_total_row;
2504 graph->n_band = n_band;
2506 return pad_schedule(graph);
2509 /* Compute a schedule for the given dependence graph.
2510 * We first check if the graph is connected (through validity dependences)
2511 * and, if not, compute a schedule for each component separately.
2513 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
2515 if (detect_wccs(graph) < 0)
2519 return compute_component_schedule(ctx, graph);
2521 return compute_schedule_wcc(ctx, graph);
2524 /* Compute a schedule for the given union of domains that respects
2525 * all the validity dependences.
2526 * If the default isl scheduling algorithm is used, it tries to minimize
2527 * the dependence distances over the proximity dependences.
2528 * If Feautrier's scheduling algorithm is used, the proximity dependence
2529 * distances are only minimized during the extension to a full-dimensional
2532 __isl_give isl_schedule *isl_union_set_compute_schedule(
2533 __isl_take isl_union_set *domain,
2534 __isl_take isl_union_map *validity,
2535 __isl_take isl_union_map *proximity)
2537 isl_ctx *ctx = isl_union_set_get_ctx(domain);
2539 struct isl_sched_graph graph = { 0 };
2540 isl_schedule *sched;
2542 domain = isl_union_set_align_params(domain,
2543 isl_union_map_get_space(validity));
2544 domain = isl_union_set_align_params(domain,
2545 isl_union_map_get_space(proximity));
2546 dim = isl_union_set_get_space(domain);
2547 validity = isl_union_map_align_params(validity, isl_space_copy(dim));
2548 proximity = isl_union_map_align_params(proximity, dim);
2553 graph.n = isl_union_set_n_set(domain);
2556 if (graph_alloc(ctx, &graph, graph.n,
2557 isl_union_map_n_map(validity) + isl_union_map_n_map(proximity)) < 0)
2561 if (isl_union_set_foreach_set(domain, &extract_node, &graph) < 0)
2563 if (graph_init_table(ctx, &graph) < 0)
2566 if (isl_union_map_foreach_map(validity, &extract_edge, &graph) < 0)
2568 if (graph_init_edge_table(ctx, &graph) < 0)
2570 if (isl_union_map_foreach_map(proximity, &extract_edge, &graph) < 0)
2573 if (compute_schedule(ctx, &graph) < 0)
2577 sched = extract_schedule(&graph, isl_union_set_get_space(domain));
2579 graph_free(ctx, &graph);
2580 isl_union_set_free(domain);
2581 isl_union_map_free(validity);
2582 isl_union_map_free(proximity);
2586 graph_free(ctx, &graph);
2587 isl_union_set_free(domain);
2588 isl_union_map_free(validity);
2589 isl_union_map_free(proximity);
2593 void *isl_schedule_free(__isl_take isl_schedule *sched)
2599 if (--sched->ref > 0)
2602 for (i = 0; i < sched->n; ++i) {
2603 isl_map_free(sched->node[i].sched);
2604 free(sched->node[i].band_end);
2605 free(sched->node[i].band_id);
2606 free(sched->node[i].zero);
2608 isl_space_free(sched->dim);
2609 isl_band_list_free(sched->band_forest);
2614 isl_ctx *isl_schedule_get_ctx(__isl_keep isl_schedule *schedule)
2616 return schedule ? isl_space_get_ctx(schedule->dim) : NULL;
2619 __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched)
2622 isl_union_map *umap;
2627 umap = isl_union_map_empty(isl_space_copy(sched->dim));
2628 for (i = 0; i < sched->n; ++i)
2629 umap = isl_union_map_add_map(umap,
2630 isl_map_copy(sched->node[i].sched));
2635 static __isl_give isl_band_list *construct_band_list(
2636 __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
2637 int band_nr, int *parent_active, int n_active);
2639 /* Construct an isl_band structure for the band in the given schedule
2640 * with sequence number band_nr for the n_active nodes marked by active.
2641 * If the nodes don't have a band with the given sequence number,
2642 * then a band without members is created.
2644 * Because of the way the schedule is constructed, we know that
2645 * the position of the band inside the schedule of a node is the same
2646 * for all active nodes.
2648 static __isl_give isl_band *construct_band(__isl_keep isl_schedule *schedule,
2649 __isl_keep isl_band *parent,
2650 int band_nr, int *active, int n_active)
2653 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2655 unsigned start, end;
2657 band = isl_calloc_type(ctx, isl_band);
2662 band->schedule = schedule;
2663 band->parent = parent;
2665 for (i = 0; i < schedule->n; ++i)
2666 if (active[i] && schedule->node[i].n_band > band_nr + 1)
2669 if (i < schedule->n) {
2670 band->children = construct_band_list(schedule, band,
2671 band_nr + 1, active, n_active);
2672 if (!band->children)
2676 for (i = 0; i < schedule->n; ++i)
2680 if (i >= schedule->n)
2681 isl_die(ctx, isl_error_internal,
2682 "band without active statements", goto error);
2684 start = band_nr ? schedule->node[i].band_end[band_nr - 1] : 0;
2685 end = band_nr < schedule->node[i].n_band ?
2686 schedule->node[i].band_end[band_nr] : start;
2687 band->n = end - start;
2689 band->zero = isl_alloc_array(ctx, int, band->n);
2693 for (j = 0; j < band->n; ++j)
2694 band->zero[j] = schedule->node[i].zero[start + j];
2696 band->map = isl_union_map_empty(isl_space_copy(schedule->dim));
2697 for (i = 0; i < schedule->n; ++i) {
2704 map = isl_map_copy(schedule->node[i].sched);
2705 n_out = isl_map_dim(map, isl_dim_out);
2706 map = isl_map_project_out(map, isl_dim_out, end, n_out - end);
2707 map = isl_map_project_out(map, isl_dim_out, 0, start);
2708 band->map = isl_union_map_union(band->map,
2709 isl_union_map_from_map(map));
2716 isl_band_free(band);
2720 /* Construct a list of bands that start at the same position (with
2721 * sequence number band_nr) in the schedules of the nodes that
2722 * were active in the parent band.
2724 * A separate isl_band structure is created for each band_id
2725 * and for each node that does not have a band with sequence
2726 * number band_nr. In the latter case, a band without members
2728 * This ensures that if a band has any children, then each node
2729 * that was active in the band is active in exactly one of the children.
2731 static __isl_give isl_band_list *construct_band_list(
2732 __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
2733 int band_nr, int *parent_active, int n_active)
2736 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2739 isl_band_list *list;
2742 for (i = 0; i < n_active; ++i) {
2743 for (j = 0; j < schedule->n; ++j) {
2744 if (!parent_active[j])
2746 if (schedule->node[j].n_band <= band_nr)
2748 if (schedule->node[j].band_id[band_nr] == i) {
2754 for (j = 0; j < schedule->n; ++j)
2755 if (schedule->node[j].n_band <= band_nr)
2760 list = isl_band_list_alloc(ctx, n_band);
2761 band = construct_band(schedule, parent, band_nr,
2762 parent_active, n_active);
2763 return isl_band_list_add(list, band);
2766 active = isl_alloc_array(ctx, int, schedule->n);
2770 list = isl_band_list_alloc(ctx, n_band);
2772 for (i = 0; i < n_active; ++i) {
2776 for (j = 0; j < schedule->n; ++j) {
2777 active[j] = parent_active[j] &&
2778 schedule->node[j].n_band > band_nr &&
2779 schedule->node[j].band_id[band_nr] == i;
2786 band = construct_band(schedule, parent, band_nr, active, n);
2788 list = isl_band_list_add(list, band);
2790 for (i = 0; i < schedule->n; ++i) {
2792 if (!parent_active[i])
2794 if (schedule->node[i].n_band > band_nr)
2796 for (j = 0; j < schedule->n; ++j)
2798 band = construct_band(schedule, parent, band_nr, active, 1);
2799 list = isl_band_list_add(list, band);
2807 /* Construct a band forest representation of the schedule and
2808 * return the list of roots.
2810 static __isl_give isl_band_list *construct_forest(
2811 __isl_keep isl_schedule *schedule)
2814 isl_ctx *ctx = isl_schedule_get_ctx(schedule);
2815 isl_band_list *forest;
2818 active = isl_alloc_array(ctx, int, schedule->n);
2822 for (i = 0; i < schedule->n; ++i)
2825 forest = construct_band_list(schedule, NULL, 0, active, schedule->n);
2832 /* Return the roots of a band forest representation of the schedule.
2834 __isl_give isl_band_list *isl_schedule_get_band_forest(
2835 __isl_keep isl_schedule *schedule)
2839 if (!schedule->band_forest)
2840 schedule->band_forest = construct_forest(schedule);
2841 return isl_band_list_dup(schedule->band_forest);
2844 static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
2845 __isl_keep isl_band_list *list);
2847 static __isl_give isl_printer *print_band(__isl_take isl_printer *p,
2848 __isl_keep isl_band *band)
2850 isl_band_list *children;
2852 p = isl_printer_start_line(p);
2853 p = isl_printer_print_union_map(p, band->map);
2854 p = isl_printer_end_line(p);
2856 if (!isl_band_has_children(band))
2859 children = isl_band_get_children(band);
2861 p = isl_printer_indent(p, 4);
2862 p = print_band_list(p, children);
2863 p = isl_printer_indent(p, -4);
2865 isl_band_list_free(children);
2870 static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
2871 __isl_keep isl_band_list *list)
2875 n = isl_band_list_n_band(list);
2876 for (i = 0; i < n; ++i) {
2878 band = isl_band_list_get_band(list, i);
2879 p = print_band(p, band);
2880 isl_band_free(band);
2886 __isl_give isl_printer *isl_printer_print_schedule(__isl_take isl_printer *p,
2887 __isl_keep isl_schedule *schedule)
2889 isl_band_list *forest;
2891 forest = isl_schedule_get_band_forest(schedule);
2893 p = print_band_list(p, forest);
2895 isl_band_list_free(forest);
2900 void isl_schedule_dump(__isl_keep isl_schedule *schedule)
2902 isl_printer *printer;
2907 printer = isl_printer_to_file(isl_schedule_get_ctx(schedule), stderr);
2908 printer = isl_printer_print_schedule(printer, schedule);
2910 isl_printer_free(printer);