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_dim_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>
26 * The scheduling algorithm implemented in this file was inspired by
27 * Bondhugula et al., "Automatic Transformations for Communication-Minimized
28 * Parallelization and Locality Optimization in the Polyhedral Model".
32 /* The schedule for an individual domain, plus information about the bands.
33 * In particular, we keep track of the number of bands and for each
34 * band, the starting position of the next band. The first band starts at
37 struct isl_schedule_node {
43 /* Information about the computed schedule.
44 * n is the number of nodes/domains/statements.
45 * n_band is the maximal number of bands.
46 * n_total_row is the number of coordinates of the schedule.
47 * dim contains a description of the parameters.
55 struct isl_schedule_node node[1];
58 /* Internal information about a node that is used during the construction
60 * dim represents the space in which the domain lives
61 * sched is a matrix representation of the schedule being constructed
63 * sched_map is an isl_map representation of the same (partial) schedule
64 * sched_map may be NULL
65 * rank is the number of linearly independent rows in the linear part
67 * the columns of cmap represent a change of basis for the schedule
68 * coefficients; the first rank columns span the linear part of
70 * start is the first variable in the LP problem in the sequences that
71 * represents the schedule coefficients of this node
72 * nvar is the dimension of the domain
73 * nparam is the number of parameters or 0 if we are not constructing
74 * a parametric schedule
76 * scc is the index of SCC (or WCC) this node belongs to
78 * band contains the band index for each of the rows of the schedule
80 * index, min_index and on_stack are used during the SCC detection
81 * index represents the order in which nodes are visited.
82 * min_index is the index of the root of a (sub)component.
83 * on_stack indicates whether the node is currently on the stack.
85 struct isl_sched_node {
105 static int node_has_dim(const void *entry, const void *val)
107 struct isl_sched_node *node = (struct isl_sched_node *)entry;
108 isl_dim *dim = (isl_dim *)val;
110 return isl_dim_equal(node->dim, dim);
113 /* An edge in the dependence graph. An edge may be used to
114 * ensure validity of the generated schedule, to minimize the dependence
117 * map is the dependence relation
118 * src is the source node
119 * dst is the sink node
120 * validity is set if the edge is used to ensure correctness
121 * proximity is set if the edge is used to minimize dependence distances
123 * For validity edges, start and end mark the sequence of inequality
124 * constraints in the LP problem that encode the validity constraint
125 * corresponding to this edge.
127 struct isl_sched_edge {
130 struct isl_sched_node *src;
131 struct isl_sched_node *dst;
140 /* Internal information about the dependence graph used during
141 * the construction of the schedule.
143 * intra_hmap is a cache, mapping dependence relations to their dual,
144 * for dependences from a node to itself
145 * inter_hmap is a cache, mapping dependence relations to their dual,
146 * for dependences between distinct nodes
148 * n is the number of nodes
149 * node is the list of nodes
150 * maxvar is the maximal number of variables over all nodes
151 * n_row is the current (maximal) number of linearly independent
152 * rows in the node schedules
153 * n_total_row is the current number of rows in the node schedules
154 * n_band is the current number of completed bands
155 * band_start is the starting row in the node schedules of the current band
156 * root is set if this graph is the original dependence graph,
157 * without any splitting
159 * sorted contains a list of node indices sorted according to the
160 * SCC to which a node belongs
162 * n_edge is the number of edges
163 * edge is the list of edges
164 * edge_table contains pointers into the edge array, hashed on the source
165 * and sink spaces; the table only contains edges that represent
166 * validity constraints (and that may or may not also represent proximity
169 * node_table contains pointers into the node array, hashed on the space
171 * region contains a list of variable sequences that should be non-trivial
173 * lp contains the (I)LP problem used to obtain new schedule rows
175 * src_scc and dst_scc are the source and sink SCCs of an edge with
176 * conflicting constraints
178 * scc, sp, index and stack are used during the detection of SCCs
179 * scc is the number of the next SCC
180 * stack contains the nodes on the path from the root to the current node
181 * sp is the stack pointer
182 * index is the index of the last node visited
184 struct isl_sched_graph {
185 isl_hmap_map_basic_set *intra_hmap;
186 isl_hmap_map_basic_set *inter_hmap;
188 struct isl_sched_node *node;
201 struct isl_sched_edge *edge;
203 struct isl_hash_table *edge_table;
205 struct isl_hash_table *node_table;
206 struct isl_region *region;
220 /* Initialize node_table based on the list of nodes.
222 static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph)
226 graph->node_table = isl_hash_table_alloc(ctx, graph->n);
227 if (!graph->node_table)
230 for (i = 0; i < graph->n; ++i) {
231 struct isl_hash_table_entry *entry;
234 hash = isl_dim_get_hash(graph->node[i].dim);
235 entry = isl_hash_table_find(ctx, graph->node_table, hash,
237 graph->node[i].dim, 1);
240 entry->data = &graph->node[i];
246 /* Return a pointer to the node that lives within the given space,
247 * or NULL if there is no such node.
249 static struct isl_sched_node *graph_find_node(isl_ctx *ctx,
250 struct isl_sched_graph *graph, __isl_keep isl_dim *dim)
252 struct isl_hash_table_entry *entry;
255 hash = isl_dim_get_hash(dim);
256 entry = isl_hash_table_find(ctx, graph->node_table, hash,
257 &node_has_dim, dim, 0);
259 return entry ? entry->data : NULL;
262 static int edge_has_src_and_dst(const void *entry, const void *val)
264 const struct isl_sched_edge *edge = entry;
265 const struct isl_sched_edge *temp = val;
267 return edge->src == temp->src && edge->dst == temp->dst;
270 /* Initialize edge_table based on the list of edges.
271 * Only edges with validity set are added to the table.
273 static int graph_init_edge_table(isl_ctx *ctx, struct isl_sched_graph *graph)
277 graph->edge_table = isl_hash_table_alloc(ctx, graph->n_edge);
278 if (!graph->edge_table)
281 for (i = 0; i < graph->n_edge; ++i) {
282 struct isl_hash_table_entry *entry;
285 if (!graph->edge[i].validity)
288 hash = isl_hash_init();
289 hash = isl_hash_builtin(hash, graph->edge[i].src);
290 hash = isl_hash_builtin(hash, graph->edge[i].dst);
291 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
292 &edge_has_src_and_dst,
296 entry->data = &graph->edge[i];
302 /* Check whether the dependence graph has a (validity) edge
303 * between the given two nodes.
305 static int graph_has_edge(struct isl_sched_graph *graph,
306 struct isl_sched_node *src, struct isl_sched_node *dst)
308 isl_ctx *ctx = isl_dim_get_ctx(src->dim);
309 struct isl_hash_table_entry *entry;
311 struct isl_sched_edge temp = { .src = src, .dst = dst };
312 struct isl_sched_edge *edge;
315 hash = isl_hash_init();
316 hash = isl_hash_builtin(hash, temp.src);
317 hash = isl_hash_builtin(hash, temp.dst);
318 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
319 &edge_has_src_and_dst, &temp, 0);
324 empty = isl_map_fast_is_empty(edge->map);
331 static int graph_alloc(isl_ctx *ctx, struct isl_sched_graph *graph,
332 int n_node, int n_edge)
337 graph->n_edge = n_edge;
338 graph->node = isl_calloc_array(ctx, struct isl_sched_node, graph->n);
339 graph->sorted = isl_calloc_array(ctx, int, graph->n);
340 graph->region = isl_alloc_array(ctx, struct isl_region, graph->n);
341 graph->stack = isl_alloc_array(ctx, int, graph->n);
342 graph->edge = isl_calloc_array(ctx,
343 struct isl_sched_edge, graph->n_edge);
345 graph->intra_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
346 graph->inter_hmap = isl_hmap_map_basic_set_alloc(ctx, 2 * n_edge);
348 if (!graph->node || !graph->region || !graph->stack || !graph->edge ||
352 for(i = 0; i < graph->n; ++i)
353 graph->sorted[i] = i;
358 static void graph_free(isl_ctx *ctx, struct isl_sched_graph *graph)
362 isl_hmap_map_basic_set_free(ctx, graph->intra_hmap);
363 isl_hmap_map_basic_set_free(ctx, graph->inter_hmap);
365 for (i = 0; i < graph->n; ++i) {
366 isl_dim_free(graph->node[i].dim);
367 isl_mat_free(graph->node[i].sched);
368 isl_map_free(graph->node[i].sched_map);
369 isl_mat_free(graph->node[i].cmap);
371 free(graph->node[i].band);
375 for (i = 0; i < graph->n_edge; ++i)
376 isl_map_free(graph->edge[i].map);
380 isl_hash_table_free(ctx, graph->edge_table);
381 isl_hash_table_free(ctx, graph->node_table);
382 isl_basic_set_free(graph->lp);
385 /* Add a new node to the graph representing the given set.
387 static int extract_node(__isl_take isl_set *set, void *user)
393 struct isl_sched_graph *graph = user;
396 ctx = isl_set_get_ctx(set);
397 dim = isl_set_get_dim(set);
399 nvar = isl_dim_size(dim, isl_dim_set);
400 nparam = isl_dim_size(dim, isl_dim_param);
401 if (!ctx->opt->schedule_parametric)
403 sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar);
404 graph->node[graph->n].dim = dim;
405 graph->node[graph->n].nvar = nvar;
406 graph->node[graph->n].nparam = nparam;
407 graph->node[graph->n].sched = sched;
408 graph->node[graph->n].sched_map = NULL;
409 band = isl_alloc_array(ctx, int, graph->n_edge + nvar);
410 graph->node[graph->n].band = band;
419 /* Add a new edge to the graph based on the given map.
420 * Edges are first extracted from the validity dependences,
421 * from which the edge_table is constructed.
422 * Afterwards, the proximity dependences are added. If a proximity
423 * dependence relation happens to be identical to one of the
424 * validity dependence relations added before, then we don't create
425 * a new edge, but instead mark the original edge as also representing
426 * a proximity dependence.
428 static int extract_edge(__isl_take isl_map *map, void *user)
430 isl_ctx *ctx = isl_map_get_ctx(map);
431 struct isl_sched_graph *graph = user;
432 struct isl_sched_node *src, *dst;
435 dim = isl_dim_domain(isl_map_get_dim(map));
436 src = graph_find_node(ctx, graph, dim);
438 dim = isl_dim_range(isl_map_get_dim(map));
439 dst = graph_find_node(ctx, graph, dim);
447 graph->edge[graph->n_edge].src = src;
448 graph->edge[graph->n_edge].dst = dst;
449 graph->edge[graph->n_edge].map = map;
450 graph->edge[graph->n_edge].validity = !graph->edge_table;
451 graph->edge[graph->n_edge].proximity = !!graph->edge_table;
454 if (graph->edge_table) {
456 struct isl_hash_table_entry *entry;
457 struct isl_sched_edge *edge;
460 hash = isl_hash_init();
461 hash = isl_hash_builtin(hash, src);
462 hash = isl_hash_builtin(hash, dst);
463 entry = isl_hash_table_find(ctx, graph->edge_table, hash,
464 &edge_has_src_and_dst,
465 &graph->edge[graph->n_edge - 1], 0);
469 is_equal = isl_map_fast_is_equal(map, edge->map);
483 /* Check whether there is a validity dependence from src to dst,
484 * forcing dst to follow src.
486 static int node_follows(struct isl_sched_graph *graph,
487 struct isl_sched_node *dst, struct isl_sched_node *src)
489 return graph_has_edge(graph, src, dst);
492 /* Perform Tarjan's algorithm for computing the strongly connected components
493 * in the dependence graph (only validity edges).
494 * If directed is not set, we consider the graph to be undirected and
495 * we effectively compute the (weakly) connected components.
497 static int detect_sccs_tarjan(struct isl_sched_graph *g, int i, int directed)
501 g->node[i].index = g->index;
502 g->node[i].min_index = g->index;
503 g->node[i].on_stack = 1;
505 g->stack[g->sp++] = i;
507 for (j = g->n - 1; j >= 0; --j) {
512 if (g->node[j].index >= 0 &&
513 (!g->node[j].on_stack ||
514 g->node[j].index > g->node[i].min_index))
517 f = node_follows(g, &g->node[i], &g->node[j]);
520 if (!f && !directed) {
521 f = node_follows(g, &g->node[j], &g->node[i]);
527 if (g->node[j].index < 0) {
528 detect_sccs_tarjan(g, j, directed);
529 if (g->node[j].min_index < g->node[i].min_index)
530 g->node[i].min_index = g->node[j].min_index;
531 } else if (g->node[j].index < g->node[i].min_index)
532 g->node[i].min_index = g->node[j].index;
535 if (g->node[i].index != g->node[i].min_index)
539 j = g->stack[--g->sp];
540 g->node[j].on_stack = 0;
541 g->node[j].scc = g->scc;
548 static int detect_ccs(struct isl_sched_graph *graph, int directed)
555 for (i = graph->n - 1; i >= 0; --i)
556 graph->node[i].index = -1;
558 for (i = graph->n - 1; i >= 0; --i) {
559 if (graph->node[i].index >= 0)
561 if (detect_sccs_tarjan(graph, i, directed) < 0)
568 /* Apply Tarjan's algorithm to detect the strongly connected components
569 * in the dependence graph.
571 static int detect_sccs(struct isl_sched_graph *graph)
573 return detect_ccs(graph, 1);
576 /* Apply Tarjan's algorithm to detect the (weakly) connected components
577 * in the dependence graph.
579 static int detect_wccs(struct isl_sched_graph *graph)
581 return detect_ccs(graph, 0);
584 static int cmp_scc(const void *a, const void *b, void *data)
586 struct isl_sched_graph *graph = data;
590 return graph->node[*i1].scc - graph->node[*i2].scc;
593 /* Sort the elements of graph->sorted according to the corresponding SCCs.
595 static void sort_sccs(struct isl_sched_graph *graph)
597 isl_quicksort(graph->sorted, graph->n, sizeof(int), &cmp_scc, graph);
600 /* Given a dependence relation R from a node to itself,
601 * construct the set of coefficients of valid constraints for elements
602 * in that dependence relation.
603 * In particular, the result contains tuples of coefficients
604 * c_0, c_n, c_x such that
606 * c_0 + c_n n + c_x y - c_x x >= 0 for each (x,y) in R
610 * c_0 + c_n n + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R }
612 * We choose here to compute the dual of delta R.
613 * Alternatively, we could have computed the dual of R, resulting
614 * in a set of tuples c_0, c_n, c_x, c_y, and then
615 * plugged in (c_0, c_n, c_x, -c_x).
617 static __isl_give isl_basic_set *intra_coefficients(
618 struct isl_sched_graph *graph, __isl_take isl_map *map)
620 isl_ctx *ctx = isl_map_get_ctx(map);
624 if (isl_hmap_map_basic_set_has(ctx, graph->intra_hmap, map))
625 return isl_hmap_map_basic_set_get(ctx, graph->intra_hmap, map);
627 delta = isl_set_remove_divs(isl_map_deltas(isl_map_copy(map)));
628 coef = isl_set_coefficients(delta);
629 isl_hmap_map_basic_set_set(ctx, graph->intra_hmap, map,
630 isl_basic_set_copy(coef));
635 /* Given a dependence relation R, * construct the set of coefficients
636 * of valid constraints for elements in that dependence relation.
637 * In particular, the result contains tuples of coefficients
638 * c_0, c_n, c_x, c_y such that
640 * c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R
643 static __isl_give isl_basic_set *inter_coefficients(
644 struct isl_sched_graph *graph, __isl_take isl_map *map)
646 isl_ctx *ctx = isl_map_get_ctx(map);
650 if (isl_hmap_map_basic_set_has(ctx, graph->inter_hmap, map))
651 return isl_hmap_map_basic_set_get(ctx, graph->inter_hmap, map);
653 set = isl_map_wrap(isl_map_remove_divs(isl_map_copy(map)));
654 coef = isl_set_coefficients(set);
655 isl_hmap_map_basic_set_set(ctx, graph->inter_hmap, map,
656 isl_basic_set_copy(coef));
661 /* Add constraints to graph->lp that force validity for the given
662 * dependence from a node i to itself.
663 * That is, add constraints that enforce
665 * (c_i_0 + c_i_n n + c_i_x y) - (c_i_0 + c_i_n n + c_i_x x)
666 * = c_i_x (y - x) >= 0
668 * for each (x,y) in R.
669 * We obtain general constraints on coefficients (c_0, c_n, c_x)
670 * of valid constraints for (y - x) and then plug in (0, 0, c_i_x^+ - c_i_x^-),
671 * where c_i_x = c_i_x^+ - c_i_x^-, with c_i_x^+ and c_i_x^- non-negative.
672 * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart.
674 * Actually, we do not construct constraints for the c_i_x themselves,
675 * but for the coefficients of c_i_x written as a linear combination
676 * of the columns in node->cmap.
678 static int add_intra_validity_constraints(struct isl_sched_graph *graph,
679 struct isl_sched_edge *edge)
682 isl_map *map = isl_map_copy(edge->map);
683 isl_ctx *ctx = isl_map_get_ctx(map);
685 isl_dim_map *dim_map;
687 struct isl_sched_node *node = edge->src;
689 coef = intra_coefficients(graph, map);
691 dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef)));
693 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
694 isl_dim_size(dim, isl_dim_set), isl_mat_copy(node->cmap));
696 total = isl_basic_set_total_dim(graph->lp);
697 dim_map = isl_dim_map_alloc(ctx, total);
698 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
699 isl_dim_size(dim, isl_dim_set), 1,
701 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
702 isl_dim_size(dim, isl_dim_set), 1,
704 graph->lp = isl_basic_set_extend_constraints(graph->lp,
705 coef->n_eq, coef->n_ineq);
706 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
713 /* Add constraints to graph->lp that force validity for the given
714 * dependence from node i to node j.
715 * That is, add constraints that enforce
717 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) >= 0
719 * for each (x,y) in R.
720 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
721 * of valid constraints for R and then plug in
722 * (c_j_0 - c_i_0, c_j_n^+ - c_j_n^- - (c_i_n^+ - c_i_n^-),
723 * c_j_x^+ - c_j_x^- - (c_i_x^+ - c_i_x^-)),
724 * where c_* = c_*^+ - c_*^-, with c_*^+ and c_*^- non-negative.
725 * In graph->lp, the c_*^- appear before their c_*^+ counterpart.
727 * Actually, we do not construct constraints for the c_*_x themselves,
728 * but for the coefficients of c_*_x written as a linear combination
729 * of the columns in node->cmap.
731 static int add_inter_validity_constraints(struct isl_sched_graph *graph,
732 struct isl_sched_edge *edge)
735 isl_map *map = isl_map_copy(edge->map);
736 isl_ctx *ctx = isl_map_get_ctx(map);
738 isl_dim_map *dim_map;
740 struct isl_sched_node *src = edge->src;
741 struct isl_sched_node *dst = edge->dst;
743 coef = inter_coefficients(graph, map);
745 dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef)));
747 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
748 isl_dim_size(dim, isl_dim_set), isl_mat_copy(src->cmap));
749 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
750 isl_dim_size(dim, isl_dim_set) + src->nvar,
751 isl_mat_copy(dst->cmap));
753 total = isl_basic_set_total_dim(graph->lp);
754 dim_map = isl_dim_map_alloc(ctx, total);
756 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
757 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
758 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
759 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
760 isl_dim_size(dim, isl_dim_set) + src->nvar, 1,
762 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
763 isl_dim_size(dim, isl_dim_set) + src->nvar, 1,
766 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
767 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
768 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
769 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
770 isl_dim_size(dim, isl_dim_set), 1,
772 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
773 isl_dim_size(dim, isl_dim_set), 1,
776 edge->start = graph->lp->n_ineq;
777 graph->lp = isl_basic_set_extend_constraints(graph->lp,
778 coef->n_eq, coef->n_ineq);
779 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
782 edge->end = graph->lp->n_ineq;
787 /* Add constraints to graph->lp that bound the dependence distance for the given
788 * dependence from a node i to itself.
789 * If s = 1, we add the constraint
791 * c_i_x (y - x) <= m_0 + m_n n
795 * -c_i_x (y - x) + m_0 + m_n n >= 0
797 * for each (x,y) in R.
798 * If s = -1, we add the constraint
800 * -c_i_x (y - x) <= m_0 + m_n n
804 * c_i_x (y - x) + m_0 + m_n n >= 0
806 * for each (x,y) in R.
807 * We obtain general constraints on coefficients (c_0, c_n, c_x)
808 * of valid constraints for (y - x) and then plug in (m_0, m_n, -s * c_i_x),
809 * with each coefficient (except m_0) represented as a pair of non-negative
812 * Actually, we do not construct constraints for the c_i_x themselves,
813 * but for the coefficients of c_i_x written as a linear combination
814 * of the columns in node->cmap.
816 static int add_intra_proximity_constraints(struct isl_sched_graph *graph,
817 struct isl_sched_edge *edge, int s)
821 isl_map *map = isl_map_copy(edge->map);
822 isl_ctx *ctx = isl_map_get_ctx(map);
824 isl_dim_map *dim_map;
826 struct isl_sched_node *node = edge->src;
828 coef = intra_coefficients(graph, map);
830 dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef)));
832 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
833 isl_dim_size(dim, isl_dim_set), isl_mat_copy(node->cmap));
835 nparam = isl_dim_size(node->dim, isl_dim_param);
836 total = isl_basic_set_total_dim(graph->lp);
837 dim_map = isl_dim_map_alloc(ctx, total);
838 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
839 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
840 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
841 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
842 isl_dim_size(dim, isl_dim_set), 1,
844 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
845 isl_dim_size(dim, isl_dim_set), 1,
847 graph->lp = isl_basic_set_extend_constraints(graph->lp,
848 coef->n_eq, coef->n_ineq);
849 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
856 /* Add constraints to graph->lp that bound the dependence distance for the given
857 * dependence from node i to node j.
858 * If s = 1, we add the constraint
860 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)
865 * -(c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x) +
868 * for each (x,y) in R.
869 * If s = -1, we add the constraint
871 * -((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x))
876 * (c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) +
879 * for each (x,y) in R.
880 * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
881 * of valid constraints for R and then plug in
882 * (m_0 - s*c_j_0 + s*c_i_0, m_n - s*c_j_n + s*c_i_n,
884 * with each coefficient (except m_0, c_j_0 and c_i_0)
885 * represented as a pair of non-negative coefficients.
887 * Actually, we do not construct constraints for the c_*_x themselves,
888 * but for the coefficients of c_*_x written as a linear combination
889 * of the columns in node->cmap.
891 static int add_inter_proximity_constraints(struct isl_sched_graph *graph,
892 struct isl_sched_edge *edge, int s)
896 isl_map *map = isl_map_copy(edge->map);
897 isl_ctx *ctx = isl_map_get_ctx(map);
899 isl_dim_map *dim_map;
901 struct isl_sched_node *src = edge->src;
902 struct isl_sched_node *dst = edge->dst;
904 coef = inter_coefficients(graph, map);
906 dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef)));
908 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
909 isl_dim_size(dim, isl_dim_set), isl_mat_copy(src->cmap));
910 coef = isl_basic_set_transform_dims(coef, isl_dim_set,
911 isl_dim_size(dim, isl_dim_set) + src->nvar,
912 isl_mat_copy(dst->cmap));
914 nparam = isl_dim_size(src->dim, isl_dim_param);
915 total = isl_basic_set_total_dim(graph->lp);
916 dim_map = isl_dim_map_alloc(ctx, total);
918 isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
919 isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
920 isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
922 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, -s);
923 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, s);
924 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, -s);
925 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
926 isl_dim_size(dim, isl_dim_set) + src->nvar, 1,
928 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
929 isl_dim_size(dim, isl_dim_set) + src->nvar, 1,
932 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, s);
933 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, -s);
934 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, s);
935 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
936 isl_dim_size(dim, isl_dim_set), 1,
938 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
939 isl_dim_size(dim, isl_dim_set), 1,
942 graph->lp = isl_basic_set_extend_constraints(graph->lp,
943 coef->n_eq, coef->n_ineq);
944 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
951 static int add_all_validity_constraints(struct isl_sched_graph *graph)
955 for (i = 0; i < graph->n_edge; ++i) {
956 struct isl_sched_edge *edge= &graph->edge[i];
959 if (edge->src != edge->dst)
961 if (add_intra_validity_constraints(graph, edge) < 0)
965 for (i = 0; i < graph->n_edge; ++i) {
966 struct isl_sched_edge *edge = &graph->edge[i];
969 if (edge->src == edge->dst)
971 if (add_inter_validity_constraints(graph, edge) < 0)
978 /* Add constraints to graph->lp that bound the dependence distance
979 * for all dependence relations.
980 * If a given proximity dependence is identical to a validity
981 * dependence, then the dependence distance is already bounded
982 * from below (by zero), so we only need to bound the distance
984 * Otherwise, we need to bound the distance both from above and from below.
986 static int add_all_proximity_constraints(struct isl_sched_graph *graph)
990 for (i = 0; i < graph->n_edge; ++i) {
991 struct isl_sched_edge *edge= &graph->edge[i];
992 if (!edge->proximity)
994 if (edge->src == edge->dst &&
995 add_intra_proximity_constraints(graph, edge, 1) < 0)
997 if (edge->src != edge->dst &&
998 add_inter_proximity_constraints(graph, edge, 1) < 0)
1002 if (edge->src == edge->dst &&
1003 add_intra_proximity_constraints(graph, edge, -1) < 0)
1005 if (edge->src != edge->dst &&
1006 add_inter_proximity_constraints(graph, edge, -1) < 0)
1013 /* Compute a basis for the rows in the linear part of the schedule
1014 * and extend this basis to a full basis. The remaining rows
1015 * can then be used to force linear independence from the rows
1018 * In particular, given the schedule rows S, we compute
1022 * with H the Hermite normal form of S. That is, all but the
1023 * first rank columns of Q are zero and so each row in S is
1024 * a linear combination of the first rank rows of Q.
1025 * The matrix Q is then transposed because we will write the
1026 * coefficients of the next schedule row as a column vector s
1027 * and express this s as a linear combination s = Q c of the
1030 static int node_update_cmap(struct isl_sched_node *node)
1033 int n_row = isl_mat_rows(node->sched);
1035 H = isl_mat_sub_alloc(node->sched, 0, n_row,
1036 1 + node->nparam, node->nvar);
1038 H = isl_mat_left_hermite(H, 0, NULL, &Q);
1039 isl_mat_free(node->cmap);
1040 node->cmap = isl_mat_transpose(Q);
1041 node->rank = isl_mat_initial_non_zero_cols(H);
1044 if (!node->cmap || node->rank < 0)
1049 /* Count the number of equality and inequality constraints
1050 * that will be added. If once is set, then we count
1051 * each edge exactly once. Otherwise, we count as follows
1052 * validity -> 1 (>= 0)
1053 * validity+proximity -> 2 (>= 0 and upper bound)
1054 * proximity -> 2 (lower and upper bound)
1056 static int count_constraints(struct isl_sched_graph *graph,
1057 int *n_eq, int *n_ineq, int once)
1060 isl_basic_set *coef;
1062 *n_eq = *n_ineq = 0;
1063 for (i = 0; i < graph->n_edge; ++i) {
1064 struct isl_sched_edge *edge= &graph->edge[i];
1065 isl_map *map = isl_map_copy(edge->map);
1066 int f = once ? 1 : edge->proximity ? 2 : 1;
1068 if (edge->src == edge->dst)
1069 coef = intra_coefficients(graph, map);
1071 coef = inter_coefficients(graph, map);
1074 *n_eq += f * coef->n_eq;
1075 *n_ineq += f * coef->n_ineq;
1076 isl_basic_set_free(coef);
1082 /* Construct an ILP problem for finding schedule coefficients
1083 * that result in non-negative, but small dependence distances
1084 * over all dependences.
1085 * In particular, the dependence distances over proximity edges
1086 * are bounded by m_0 + m_n n and we compute schedule coefficients
1087 * with small values (preferably zero) of m_n and m_0.
1089 * All variables of the ILP are non-negative. The actual coefficients
1090 * may be negative, so each coefficient is represented as the difference
1091 * of two non-negative variables. The negative part always appears
1092 * immediately before the positive part.
1093 * Other than that, the variables have the following order
1095 * - sum of positive and negative parts of m_n coefficients
1097 * - sum of positive and negative parts of all c_n coefficients
1098 * (unconstrained when computing non-parametric schedules)
1099 * - sum of positive and negative parts of all c_x coefficients
1100 * - positive and negative parts of m_n coefficients
1103 * - positive and negative parts of c_i_n (if parametric)
1104 * - positive and negative parts of c_i_x
1106 * The c_i_x are not represented directly, but through the columns of
1107 * node->cmap. That is, the computed values are for variable t_i_x
1108 * such that c_i_x = Q t_i_x with Q equal to node->cmap.
1110 * The constraints are those from the edges plus two or three equalities
1111 * to express the sums.
1113 static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
1124 parametric = ctx->opt->schedule_parametric;
1125 nparam = isl_dim_size(graph->node[0].dim, isl_dim_param);
1127 total = param_pos + 2 * nparam;
1128 for (i = 0; i < graph->n; ++i) {
1129 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
1130 if (node_update_cmap(node) < 0)
1132 node->start = total;
1133 total += 1 + 2 * (node->nparam + node->nvar);
1136 if (count_constraints(graph, &n_eq, &n_ineq, 0) < 0)
1139 dim = isl_dim_set_alloc(ctx, 0, total);
1140 isl_basic_set_free(graph->lp);
1141 n_eq += 2 + parametric;
1142 graph->lp = isl_basic_set_alloc_dim(dim, 0, n_eq, n_ineq);
1144 k = isl_basic_set_alloc_equality(graph->lp);
1147 isl_seq_clr(graph->lp->eq[k], 1 + total);
1148 isl_int_set_si(graph->lp->eq[k][1], -1);
1149 for (i = 0; i < 2 * nparam; ++i)
1150 isl_int_set_si(graph->lp->eq[k][1 + param_pos + i], 1);
1153 k = isl_basic_set_alloc_equality(graph->lp);
1156 isl_seq_clr(graph->lp->eq[k], 1 + total);
1157 isl_int_set_si(graph->lp->eq[k][3], -1);
1158 for (i = 0; i < graph->n; ++i) {
1159 int pos = 1 + graph->node[i].start + 1;
1161 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
1162 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1166 k = isl_basic_set_alloc_equality(graph->lp);
1169 isl_seq_clr(graph->lp->eq[k], 1 + total);
1170 isl_int_set_si(graph->lp->eq[k][4], -1);
1171 for (i = 0; i < graph->n; ++i) {
1172 struct isl_sched_node *node = &graph->node[i];
1173 int pos = 1 + node->start + 1 + 2 * node->nparam;
1175 for (j = 0; j < 2 * node->nvar; ++j)
1176 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
1179 if (add_all_validity_constraints(graph) < 0)
1181 if (add_all_proximity_constraints(graph) < 0)
1187 /* Analyze the conflicting constraint found by
1188 * isl_tab_basic_set_non_trivial_lexmin. If it corresponds to the validity
1189 * constraint of one of the edges between distinct nodes, living, moreover
1190 * in distinct SCCs, then record the source and sink SCC as this may
1191 * be a good place to cut between SCCs.
1193 static int check_conflict(int con, void *user)
1196 struct isl_sched_graph *graph = user;
1198 if (graph->src_scc >= 0)
1201 con -= graph->lp->n_eq;
1203 if (con >= graph->lp->n_ineq)
1206 for (i = 0; i < graph->n_edge; ++i) {
1207 if (!graph->edge[i].validity)
1209 if (graph->edge[i].src == graph->edge[i].dst)
1211 if (graph->edge[i].src->scc == graph->edge[i].dst->scc)
1213 if (graph->edge[i].start > con)
1215 if (graph->edge[i].end <= con)
1217 graph->src_scc = graph->edge[i].src->scc;
1218 graph->dst_scc = graph->edge[i].dst->scc;
1224 /* Check whether the next schedule row of the given node needs to be
1225 * non-trivial. Lower-dimensional domains may have some trivial rows,
1226 * but as soon as the number of remaining required non-trivial rows
1227 * is as large as the number or remaining rows to be computed,
1228 * all remaining rows need to be non-trivial.
1230 static int needs_row(struct isl_sched_graph *graph, struct isl_sched_node *node)
1232 return node->nvar - node->rank >= graph->maxvar - graph->n_row;
1235 /* Solve the ILP problem constructed in setup_lp.
1236 * For each node such that all the remaining rows of its schedule
1237 * need to be non-trivial, we construct a non-triviality region.
1238 * This region imposes that the next row is independent of previous rows.
1239 * In particular the coefficients c_i_x are represented by t_i_x
1240 * variables with c_i_x = Q t_i_x and Q a unimodular matrix such that
1241 * its first columns span the rows of the previously computed part
1242 * of the schedule. The non-triviality region enforces that at least
1243 * one of the remaining components of t_i_x is non-zero, i.e.,
1244 * that the new schedule row depends on at least one of the remaining
1247 static __isl_give isl_vec *solve_lp(struct isl_sched_graph *graph)
1253 for (i = 0; i < graph->n; ++i) {
1254 struct isl_sched_node *node = &graph->node[i];
1255 int skip = node->rank;
1256 graph->region[i].pos = node->start + 1 + 2*(node->nparam+skip);
1257 if (needs_row(graph, node))
1258 graph->region[i].len = 2 * (node->nvar - skip);
1260 graph->region[i].len = 0;
1262 lp = isl_basic_set_copy(graph->lp);
1263 sol = isl_tab_basic_set_non_trivial_lexmin(lp, 2, graph->n,
1264 graph->region, &check_conflict, graph);
1268 /* Update the schedules of all nodes based on the given solution
1269 * of the LP problem.
1270 * The new row is added to the current band.
1271 * All possibly negative coefficients are encoded as a difference
1272 * of two non-negative variables, so we need to perform the subtraction
1273 * here. Moreover, if use_cmap is set, then the solution does
1274 * not refer to the actual coefficients c_i_x, but instead to variables
1275 * t_i_x such that c_i_x = Q t_i_x and Q is equal to node->cmap.
1276 * In this case, we then also need to perform this multiplication
1277 * to obtain the values of c_i_x.
1279 static int update_schedule(struct isl_sched_graph *graph,
1280 __isl_take isl_vec *sol, int use_cmap)
1283 isl_vec *csol = NULL;
1288 isl_die(sol->ctx, isl_error_internal,
1289 "no solution found", goto error);
1291 for (i = 0; i < graph->n; ++i) {
1292 struct isl_sched_node *node = &graph->node[i];
1293 int pos = node->start;
1294 int row = isl_mat_rows(node->sched);
1297 csol = isl_vec_alloc(sol->ctx, node->nvar);
1301 isl_map_free(node->sched_map);
1302 node->sched_map = NULL;
1303 node->sched = isl_mat_add_rows(node->sched, 1);
1306 node->sched = isl_mat_set_element(node->sched, row, 0,
1308 for (j = 0; j < node->nparam + node->nvar; ++j)
1309 isl_int_sub(sol->el[1 + pos + 1 + 2 * j + 1],
1310 sol->el[1 + pos + 1 + 2 * j + 1],
1311 sol->el[1 + pos + 1 + 2 * j]);
1312 for (j = 0; j < node->nparam; ++j)
1313 node->sched = isl_mat_set_element(node->sched,
1314 row, 1 + j, sol->el[1+pos+1+2*j+1]);
1315 for (j = 0; j < node->nvar; ++j)
1316 isl_int_set(csol->el[j],
1317 sol->el[1+pos+1+2*(node->nparam+j)+1]);
1319 csol = isl_mat_vec_product(isl_mat_copy(node->cmap),
1323 for (j = 0; j < node->nvar; ++j)
1324 node->sched = isl_mat_set_element(node->sched,
1325 row, 1 + node->nparam + j, csol->el[j]);
1326 node->band[graph->n_total_row] = graph->n_band;
1332 graph->n_total_row++;
1341 /* Convert node->sched into a map and return this map.
1342 * We simply add equality constraints that express each output variable
1343 * as the affine combination of parameters and input variables specified
1344 * by the schedule matrix.
1346 * The result is cached in node->sched_map, which needs to be released
1347 * whenever node->sched is updated.
1349 static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node)
1353 isl_basic_map *bmap;
1358 if (node->sched_map)
1359 return isl_map_copy(node->sched_map);
1361 nrow = isl_mat_rows(node->sched);
1362 ncol = isl_mat_cols(node->sched) - 1;
1363 dim = isl_dim_from_domain(isl_dim_copy(node->dim));
1364 dim = isl_dim_add(dim, isl_dim_out, nrow);
1365 bmap = isl_basic_map_universe(isl_dim_copy(dim));
1369 for (i = 0; i < nrow; ++i) {
1370 c = isl_equality_alloc(isl_dim_copy(dim));
1371 isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
1372 isl_mat_get_element(node->sched, i, 0, &v);
1373 isl_constraint_set_constant(c, v);
1374 for (j = 0; j < node->nparam; ++j) {
1375 isl_mat_get_element(node->sched, i, 1 + j, &v);
1376 isl_constraint_set_coefficient(c, isl_dim_param, j, v);
1378 for (j = 0; j < node->nvar; ++j) {
1379 isl_mat_get_element(node->sched,
1380 i, 1 + node->nparam + j, &v);
1381 isl_constraint_set_coefficient(c, isl_dim_in, j, v);
1383 bmap = isl_basic_map_add_constraint(bmap, c);
1390 node->sched_map = isl_map_from_basic_map(bmap);
1391 return isl_map_copy(node->sched_map);
1394 /* Update the given dependence relation based on the current schedule.
1395 * That is, intersect the dependence relation with a map expressing
1396 * that source and sink are executed within the same iteration of
1397 * the current schedule.
1398 * This is not the most efficient way, but this shouldn't be a critical
1401 static __isl_give isl_map *specialize(__isl_take isl_map *map,
1402 struct isl_sched_node *src, struct isl_sched_node *dst)
1404 isl_map *src_sched, *dst_sched, *id;
1406 src_sched = node_extract_schedule(src);
1407 dst_sched = node_extract_schedule(dst);
1408 id = isl_map_apply_range(src_sched, isl_map_reverse(dst_sched));
1409 return isl_map_intersect(map, id);
1412 /* Update the dependence relations of all edges based on the current schedule.
1413 * If a dependence is carried completely by the current schedule, then
1414 * it is removed and edge_table is updated accordingly.
1416 static int update_edges(isl_ctx *ctx, struct isl_sched_graph *graph)
1419 int reset_table = 0;
1421 for (i = graph->n_edge - 1; i >= 0; --i) {
1422 struct isl_sched_edge *edge = &graph->edge[i];
1423 edge->map = specialize(edge->map, edge->src, edge->dst);
1427 if (isl_map_fast_is_empty(edge->map)) {
1429 isl_map_free(edge->map);
1430 if (i != graph->n_edge - 1)
1431 graph->edge[i] = graph->edge[graph->n_edge - 1];
1437 isl_hash_table_free(ctx, graph->edge_table);
1438 graph->edge_table = NULL;
1439 return graph_init_edge_table(ctx, graph);
1445 static void next_band(struct isl_sched_graph *graph)
1447 graph->band_start = graph->n_total_row;
1451 /* Topologically sort statements mapped to same schedule iteration
1452 * and add a row to the schedule corresponding to this order.
1454 static int sort_statements(isl_ctx *ctx, struct isl_sched_graph *graph)
1461 if (update_edges(ctx, graph) < 0)
1464 if (graph->n_edge == 0)
1467 if (detect_sccs(graph) < 0)
1470 for (i = 0; i < graph->n; ++i) {
1471 struct isl_sched_node *node = &graph->node[i];
1472 int row = isl_mat_rows(node->sched);
1473 int cols = isl_mat_cols(node->sched);
1475 isl_map_free(node->sched_map);
1476 node->sched_map = NULL;
1477 node->sched = isl_mat_add_rows(node->sched, 1);
1480 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1482 for (j = 1; j < cols; ++j)
1483 node->sched = isl_mat_set_element_si(node->sched,
1485 node->band[graph->n_total_row] = graph->n_band;
1488 graph->n_total_row++;
1494 /* Construct an isl_schedule based on the computed schedule stored
1495 * in graph and with parameters specified by dim.
1497 static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph,
1498 __isl_take isl_dim *dim)
1502 isl_schedule *sched = NULL;
1507 ctx = isl_dim_get_ctx(dim);
1508 sched = isl_calloc(ctx, struct isl_schedule,
1509 sizeof(struct isl_schedule) +
1510 (graph->n - 1) * sizeof(struct isl_schedule_node));
1514 sched->n = graph->n;
1515 sched->n_band = graph->n_band;
1516 sched->n_total_row = graph->n_total_row;
1518 for (i = 0; i < sched->n; ++i) {
1522 band_end = isl_alloc_array(ctx, int, graph->n_band);
1525 sched->node[i].sched = node_extract_schedule(&graph->node[i]);
1526 sched->node[i].band_end = band_end;
1528 for (r = b = 0; r < graph->n_total_row; ++r) {
1529 if (graph->node[i].band[r] == b)
1532 if (graph->node[i].band[r] == -1)
1535 if (r == graph->n_total_row)
1537 sched->node[i].n_band = b;
1545 isl_schedule_free(sched);
1549 /* Copy nodes that satisfy node_pred from the src dependence graph
1550 * to the dst dependence graph.
1552 static int copy_nodes(struct isl_sched_graph *dst, struct isl_sched_graph *src,
1553 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1558 for (i = 0; i < src->n; ++i) {
1559 if (!node_pred(&src->node[i], data))
1561 dst->node[dst->n].dim = isl_dim_copy(src->node[i].dim);
1562 dst->node[dst->n].nvar = src->node[i].nvar;
1563 dst->node[dst->n].nparam = src->node[i].nparam;
1564 dst->node[dst->n].sched = isl_mat_copy(src->node[i].sched);
1565 dst->node[dst->n].sched_map =
1566 isl_map_copy(src->node[i].sched_map);
1567 dst->node[dst->n].band = src->node[i].band;
1574 /* Copy non-empty edges that satisfy edge_pred from the src dependence graph
1575 * to the dst dependence graph.
1577 static int copy_edges(isl_ctx *ctx, struct isl_sched_graph *dst,
1578 struct isl_sched_graph *src,
1579 int (*edge_pred)(struct isl_sched_edge *edge, int data), int data)
1584 for (i = 0; i < src->n_edge; ++i) {
1585 struct isl_sched_edge *edge = &src->edge[i];
1588 if (!edge_pred(edge, data))
1591 if (isl_map_fast_is_empty(edge->map))
1594 map = isl_map_copy(edge->map);
1596 dst->edge[dst->n_edge].src =
1597 graph_find_node(ctx, dst, edge->src->dim);
1598 dst->edge[dst->n_edge].dst =
1599 graph_find_node(ctx, dst, edge->dst->dim);
1600 dst->edge[dst->n_edge].map = map;
1601 dst->edge[dst->n_edge].validity = edge->validity;
1602 dst->edge[dst->n_edge].proximity = edge->proximity;
1609 /* Given a "src" dependence graph that contains the nodes from "dst"
1610 * that satisfy node_pred, copy the schedule computed in "src"
1611 * for those nodes back to "dst".
1613 static int copy_schedule(struct isl_sched_graph *dst,
1614 struct isl_sched_graph *src,
1615 int (*node_pred)(struct isl_sched_node *node, int data), int data)
1620 for (i = 0; i < dst->n; ++i) {
1621 if (!node_pred(&dst->node[i], data))
1623 isl_mat_free(dst->node[i].sched);
1624 isl_map_free(dst->node[i].sched_map);
1625 dst->node[i].sched = isl_mat_copy(src->node[src->n].sched);
1626 dst->node[i].sched_map =
1627 isl_map_copy(src->node[src->n].sched_map);
1631 dst->n_total_row = src->n_total_row;
1632 dst->n_band = src->n_band;
1637 /* Compute the maximal number of variables over all nodes.
1638 * This is the maximal number of linearly independent schedule
1639 * rows that we need to compute.
1640 * Just in case we end up in a part of the dependence graph
1641 * with only lower-dimensional domains, we make sure we will
1642 * compute the required amount of extra linearly independent rows.
1644 static int compute_maxvar(struct isl_sched_graph *graph)
1649 for (i = 0; i < graph->n; ++i) {
1650 struct isl_sched_node *node = &graph->node[i];
1653 if (node_update_cmap(node) < 0)
1655 nvar = node->nvar + graph->n_row - node->rank;
1656 if (nvar > graph->maxvar)
1657 graph->maxvar = nvar;
1663 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph);
1664 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph);
1666 /* Compute a schedule for a subgraph of "graph". In particular, for
1667 * the graph composed of nodes that satisfy node_pred and edges that
1668 * that satisfy edge_pred. The caller should precompute the number
1669 * of nodes and edges that satisfy these predicates and pass them along
1670 * as "n" and "n_edge".
1671 * If the subgraph is known to consist of a single component, then wcc should
1672 * be set and then we call compute_schedule_wcc on the constructed subgraph.
1673 * Otherwise, we call compute_schedule, which will check whether the subgraph
1676 static int compute_sub_schedule(isl_ctx *ctx,
1677 struct isl_sched_graph *graph, int n, int n_edge,
1678 int (*node_pred)(struct isl_sched_node *node, int data),
1679 int (*edge_pred)(struct isl_sched_edge *edge, int data),
1682 struct isl_sched_graph split = { 0 };
1684 if (graph_alloc(ctx, &split, n, n_edge) < 0)
1686 if (copy_nodes(&split, graph, node_pred, data) < 0)
1688 if (graph_init_table(ctx, &split) < 0)
1690 if (copy_edges(ctx, &split, graph, edge_pred, data) < 0)
1692 if (graph_init_edge_table(ctx, &split) < 0)
1694 split.n_row = graph->n_row;
1695 split.n_total_row = graph->n_total_row;
1696 split.n_band = graph->n_band;
1697 split.band_start = graph->band_start;
1699 if (wcc && compute_schedule_wcc(ctx, &split) < 0)
1701 if (!wcc && compute_schedule(ctx, &split) < 0)
1704 copy_schedule(graph, &split, node_pred, data);
1706 graph_free(ctx, &split);
1709 graph_free(ctx, &split);
1713 static int node_scc_exactly(struct isl_sched_node *node, int scc)
1715 return node->scc == scc;
1718 static int node_scc_at_most(struct isl_sched_node *node, int scc)
1720 return node->scc <= scc;
1723 static int node_scc_at_least(struct isl_sched_node *node, int scc)
1725 return node->scc >= scc;
1728 static int edge_src_scc_exactly(struct isl_sched_edge *edge, int scc)
1730 return edge->src->scc == scc;
1733 static int edge_dst_scc_at_most(struct isl_sched_edge *edge, int scc)
1735 return edge->dst->scc <= scc;
1738 static int edge_src_scc_at_least(struct isl_sched_edge *edge, int scc)
1740 return edge->src->scc >= scc;
1743 /* Pad the schedules of all nodes with zero rows such that in the end
1744 * they all have graph->n_total_row rows.
1745 * The extra rows don't belong to any band, so they get assigned band number -1.
1747 static int pad_schedule(struct isl_sched_graph *graph)
1751 for (i = 0; i < graph->n; ++i) {
1752 struct isl_sched_node *node = &graph->node[i];
1753 int row = isl_mat_rows(node->sched);
1754 if (graph->n_total_row > row) {
1755 isl_map_free(node->sched_map);
1756 node->sched_map = NULL;
1758 node->sched = isl_mat_add_zero_rows(node->sched,
1759 graph->n_total_row - row);
1762 for (j = row; j < graph->n_total_row; ++j)
1769 /* Split the current graph into two parts and compute a schedule for each
1770 * part individually. In particular, one part consists of all SCCs up
1771 * to and including graph->src_scc, while the other part contains the other
1774 * The split is enforced in the schedule by constant rows with two different
1775 * values (0 and 1). These constant rows replace the previously computed rows
1776 * in the current band.
1777 * It would be possible to reuse them as the first rows in the next
1778 * band, but recomputing them may result in better rows as we are looking
1779 * at a smaller part of the dependence graph.
1781 static int compute_split_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
1783 int i, j, n, e1, e2;
1784 int n_total_row, orig_total_row;
1785 int n_band, orig_band;
1788 drop = graph->n_total_row - graph->band_start;
1789 graph->n_total_row -= drop;
1790 graph->n_row -= drop;
1793 for (i = 0; i < graph->n; ++i) {
1794 struct isl_sched_node *node = &graph->node[i];
1795 int row = isl_mat_rows(node->sched) - drop;
1796 int cols = isl_mat_cols(node->sched);
1797 int before = node->scc <= graph->src_scc;
1802 isl_map_free(node->sched_map);
1803 node->sched_map = NULL;
1804 node->sched = isl_mat_drop_rows(node->sched,
1805 graph->band_start, drop);
1806 node->sched = isl_mat_add_rows(node->sched, 1);
1809 node->sched = isl_mat_set_element_si(node->sched, row, 0,
1811 for (j = 1; j < cols; ++j)
1812 node->sched = isl_mat_set_element_si(node->sched,
1814 node->band[graph->n_total_row] = graph->n_band;
1818 for (i = 0; i < graph->n_edge; ++i) {
1819 if (graph->edge[i].dst->scc <= graph->src_scc)
1821 if (graph->edge[i].src->scc > graph->src_scc)
1825 graph->n_total_row++;
1828 orig_total_row = graph->n_total_row;
1829 orig_band = graph->n_band;
1830 if (compute_sub_schedule(ctx, graph, n, e1,
1831 &node_scc_at_most, &edge_dst_scc_at_most,
1832 graph->src_scc, 0) < 0)
1834 n_total_row = graph->n_total_row;
1835 graph->n_total_row = orig_total_row;
1836 n_band = graph->n_band;
1837 graph->n_band = orig_band;
1838 if (compute_sub_schedule(ctx, graph, graph->n - n, e2,
1839 &node_scc_at_least, &edge_src_scc_at_least,
1840 graph->src_scc + 1, 0) < 0)
1842 if (n_total_row > graph->n_total_row)
1843 graph->n_total_row = n_total_row;
1844 if (n_band > graph->n_band)
1845 graph->n_band = n_band;
1847 return pad_schedule(graph);
1850 /* Compute the next band of the schedule after updating the dependence
1851 * relations based on the the current schedule.
1853 static int compute_next_band(isl_ctx *ctx, struct isl_sched_graph *graph)
1855 if (update_edges(ctx, graph) < 0)
1859 return compute_schedule(ctx, graph);
1862 /* Add constraints to graph->lp that force the dependence of edge i
1863 * to be respected and attempt to carry it, where edge i is one from
1864 * a node j to itself.
1865 * That is, add constraints that enforce
1867 * (c_j_0 + c_j_n n + c_j_x y) - (c_j_0 + c_j_n n + c_j_x x)
1868 * = c_j_x (y - x) >= e_i
1870 * for each (x,y) in R.
1871 * We obtain general constraints on coefficients (c_0, c_n, c_x)
1872 * of valid constraints for (y - x) and then plug in (-e_i, 0, c_j_x),
1873 * with each coefficient in c_j_x represented as a pair of non-negative
1876 static int add_intra_constraints(struct isl_sched_graph *graph, int i)
1879 struct isl_sched_edge *edge= &graph->edge[i];
1880 isl_map *map = isl_map_copy(edge->map);
1881 isl_ctx *ctx = isl_map_get_ctx(map);
1883 isl_dim_map *dim_map;
1884 isl_basic_set *coef;
1885 struct isl_sched_node *node = edge->src;
1887 coef = intra_coefficients(graph, map);
1889 dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef)));
1891 total = isl_basic_set_total_dim(graph->lp);
1892 dim_map = isl_dim_map_alloc(ctx, total);
1893 isl_dim_map_range(dim_map, 3 + i, 0, 0, 0, 1, -1);
1894 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2,
1895 isl_dim_size(dim, isl_dim_set), 1,
1897 isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2,
1898 isl_dim_size(dim, isl_dim_set), 1,
1900 graph->lp = isl_basic_set_extend_constraints(graph->lp,
1901 coef->n_eq, coef->n_ineq);
1902 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
1909 /* Add constraints to graph->lp that force the dependence of edge i
1910 * to be respected and attempt to carry it, where edge i is one from
1912 * That is, add constraints that enforce
1914 * (c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i
1916 * for each (x,y) in R.
1917 * We obtain general constraints on coefficients (c_0, c_n, c_x)
1918 * of valid constraints for R and then plug in
1919 * (-e_i + c_k_0 - c_j_0, c_k_n - c_j_n, c_k_x - c_j_x)
1920 * with each coefficient (except e_i, c_k_0 and c_j_0)
1921 * represented as a pair of non-negative coefficients.
1923 static int add_inter_constraints(struct isl_sched_graph *graph, int i)
1926 struct isl_sched_edge *edge= &graph->edge[i];
1927 isl_map *map = isl_map_copy(edge->map);
1928 isl_ctx *ctx = isl_map_get_ctx(map);
1930 isl_dim_map *dim_map;
1931 isl_basic_set *coef;
1932 struct isl_sched_node *src = edge->src;
1933 struct isl_sched_node *dst = edge->dst;
1935 coef = inter_coefficients(graph, map);
1937 dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef)));
1939 total = isl_basic_set_total_dim(graph->lp);
1940 dim_map = isl_dim_map_alloc(ctx, total);
1942 isl_dim_map_range(dim_map, 3 + i, 0, 0, 0, 1, -1);
1944 isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1);
1945 isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1);
1946 isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1);
1947 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2,
1948 isl_dim_size(dim, isl_dim_set) + src->nvar, 1,
1950 isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2,
1951 isl_dim_size(dim, isl_dim_set) + src->nvar, 1,
1954 isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1);
1955 isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1);
1956 isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1);
1957 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2,
1958 isl_dim_size(dim, isl_dim_set), 1,
1960 isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2,
1961 isl_dim_size(dim, isl_dim_set), 1,
1964 graph->lp = isl_basic_set_extend_constraints(graph->lp,
1965 coef->n_eq, coef->n_ineq);
1966 graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp,
1973 /* Add constraints to graph->lp that force all dependence
1974 * to be respected and attempt to carry it.
1976 static int add_all_constraints(struct isl_sched_graph *graph)
1980 for (i = 0; i < graph->n_edge; ++i) {
1981 struct isl_sched_edge *edge= &graph->edge[i];
1982 if (edge->src == edge->dst &&
1983 add_intra_constraints(graph, i) < 0)
1985 if (edge->src != edge->dst &&
1986 add_inter_constraints(graph, i) < 0)
1993 /* Construct an LP problem for finding schedule coefficients
1994 * such that the schedule carries as many dependences as possible.
1995 * In particular, for each dependence i, we bound the dependence distance
1996 * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum
1997 * of all e_i's. Dependence with e_i = 0 in the solution are simply
1998 * respected, while those with e_i > 0 (in practice e_i = 1) are carried.
2000 * All variables of the LP are non-negative. The actual coefficients
2001 * may be negative, so each coefficient is represented as the difference
2002 * of two non-negative variables. The negative part always appears
2003 * immediately before the positive part.
2004 * Other than that, the variables have the following order
2006 * - sum of (1 - e_i) over all edges
2007 * - sum of positive and negative parts of all c_n coefficients
2008 * (unconstrained when computing non-parametric schedules)
2009 * - sum of positive and negative parts of all c_x coefficients
2014 * - positive and negative parts of c_i_n (if parametric)
2015 * - positive and negative parts of c_i_x
2017 * The constraints are those from the edges plus three equalities
2018 * to express the sums and n_edge inequalities to express e_i <= 1.
2020 static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
2028 total = 3 + graph->n_edge;
2029 for (i = 0; i < graph->n; ++i) {
2030 struct isl_sched_node *node = &graph->node[graph->sorted[i]];
2031 node->start = total;
2032 total += 1 + 2 * (node->nparam + node->nvar);
2035 if (count_constraints(graph, &n_eq, &n_ineq, 1) < 0)
2038 dim = isl_dim_set_alloc(ctx, 0, total);
2039 isl_basic_set_free(graph->lp);
2041 n_ineq += graph->n_edge;
2042 graph->lp = isl_basic_set_alloc_dim(dim, 0, n_eq, n_ineq);
2043 graph->lp = isl_basic_set_set_rational(graph->lp);
2045 k = isl_basic_set_alloc_equality(graph->lp);
2048 isl_seq_clr(graph->lp->eq[k], 1 + total);
2049 isl_int_set_si(graph->lp->eq[k][0], -graph->n_edge);
2050 isl_int_set_si(graph->lp->eq[k][1], 1);
2051 for (i = 0; i < graph->n_edge; ++i)
2052 isl_int_set_si(graph->lp->eq[k][4 + i], 1);
2054 k = isl_basic_set_alloc_equality(graph->lp);
2057 isl_seq_clr(graph->lp->eq[k], 1 + total);
2058 isl_int_set_si(graph->lp->eq[k][2], -1);
2059 for (i = 0; i < graph->n; ++i) {
2060 int pos = 1 + graph->node[i].start + 1;
2062 for (j = 0; j < 2 * graph->node[i].nparam; ++j)
2063 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2066 k = isl_basic_set_alloc_equality(graph->lp);
2069 isl_seq_clr(graph->lp->eq[k], 1 + total);
2070 isl_int_set_si(graph->lp->eq[k][3], -1);
2071 for (i = 0; i < graph->n; ++i) {
2072 struct isl_sched_node *node = &graph->node[i];
2073 int pos = 1 + node->start + 1 + 2 * node->nparam;
2075 for (j = 0; j < 2 * node->nvar; ++j)
2076 isl_int_set_si(graph->lp->eq[k][pos + j], 1);
2079 for (i = 0; i < graph->n_edge; ++i) {
2080 k = isl_basic_set_alloc_inequality(graph->lp);
2083 isl_seq_clr(graph->lp->ineq[k], 1 + total);
2084 isl_int_set_si(graph->lp->ineq[k][4 + i], -1);
2085 isl_int_set_si(graph->lp->ineq[k][0], 1);
2088 if (add_all_constraints(graph) < 0)
2094 /* Construct a schedule row for each node such that as many dependences
2095 * as possible are carried and then continue with the next band.
2097 static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *graph)
2102 if (setup_carry_lp(ctx, graph) < 0)
2105 lp = isl_basic_set_copy(graph->lp);
2106 sol = isl_tab_basic_set_non_neg_lexmin(lp);
2110 if (sol->size == 0) {
2112 isl_die(ctx, isl_error_internal,
2113 "error in schedule construction", return -1);
2116 if (isl_int_cmp_si(sol->el[1], graph->n_edge) >= 0) {
2118 isl_die(ctx, isl_error_unknown,
2119 "unable to carry dependences", return -1);
2122 if (update_schedule(graph, sol, 0) < 0)
2125 return compute_next_band(ctx, graph);
2128 /* Compute a schedule for a connected dependence graph.
2129 * We try to find a sequence of as many schedule rows as possible that result
2130 * in non-negative dependence distances (independent of the previous rows
2131 * in the sequence, i.e., such that the sequence is tilable).
2132 * If we can't find any more rows we either
2133 * - split between SCCs and start over (assuming we found an interesting
2134 * pair of SCCs between which to split)
2135 * - continue with the next band (assuming the current band has at least
2137 * - try to carry as many dependences as possible and continue with the next
2140 * If we manage to complete the schedule, we finish off by topologically
2141 * sorting the statements based on the remaining dependences.
2143 static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph)
2145 if (detect_sccs(graph) < 0)
2149 if (compute_maxvar(graph) < 0)
2152 while (graph->n_row < graph->maxvar) {
2155 graph->src_scc = -1;
2156 graph->dst_scc = -1;
2158 if (setup_lp(ctx, graph) < 0)
2160 sol = solve_lp(graph);
2163 if (sol->size == 0) {
2165 if (graph->src_scc >= 0)
2166 return compute_split_schedule(ctx, graph);
2167 if (graph->n_total_row > graph->band_start)
2168 return compute_next_band(ctx, graph);
2169 return carry_dependences(ctx, graph);
2171 if (update_schedule(graph, sol, 1) < 0)
2175 if (graph->n_total_row > graph->band_start)
2177 return sort_statements(ctx, graph);
2180 /* Compute a schedule for each component (identified by node->scc)
2181 * of the dependence graph separately and then combine the results.
2183 static int compute_component_schedule(isl_ctx *ctx,
2184 struct isl_sched_graph *graph)
2188 int n_total_row, orig_total_row;
2189 int n_band, orig_band;
2192 orig_total_row = graph->n_total_row;
2194 orig_band = graph->n_band;
2195 for (wcc = 0; wcc < graph->scc; ++wcc) {
2197 for (i = 0; i < graph->n; ++i)
2198 if (graph->node[i].scc == wcc)
2201 for (i = 0; i < graph->n_edge; ++i)
2202 if (graph->edge[i].src->scc == wcc)
2205 if (compute_sub_schedule(ctx, graph, n, n_edge,
2207 &edge_src_scc_exactly, wcc, 1) < 0)
2209 if (graph->n_total_row > n_total_row)
2210 n_total_row = graph->n_total_row;
2211 graph->n_total_row = orig_total_row;
2212 if (graph->n_band > n_band)
2213 n_band = graph->n_band;
2214 graph->n_band = orig_band;
2217 graph->n_total_row = n_total_row;
2218 graph->n_band = n_band;
2220 return pad_schedule(graph);
2223 /* Compute a schedule for the given dependence graph.
2224 * We first check if the graph is connected (through validity dependences)
2225 * and if so compute a schedule for each component separately.
2227 static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
2229 if (detect_wccs(graph) < 0)
2233 return compute_component_schedule(ctx, graph);
2235 return compute_schedule_wcc(ctx, graph);
2238 /* Compute a schedule for the given union of domains that respects
2239 * all the validity dependences and tries to minimize the dependence
2240 * distances over the proximity dependences.
2242 __isl_give isl_schedule *isl_union_set_compute_schedule(
2243 __isl_take isl_union_set *domain,
2244 __isl_take isl_union_map *validity,
2245 __isl_take isl_union_map *proximity)
2247 isl_ctx *ctx = isl_union_set_get_ctx(domain);
2249 struct isl_sched_graph graph = { 0 };
2250 isl_schedule *sched;
2252 domain = isl_union_set_align_params(domain,
2253 isl_union_map_get_dim(validity));
2254 domain = isl_union_set_align_params(domain,
2255 isl_union_map_get_dim(proximity));
2256 dim = isl_union_set_get_dim(domain);
2257 validity = isl_union_map_align_params(validity, isl_dim_copy(dim));
2258 proximity = isl_union_map_align_params(proximity, dim);
2263 graph.n = isl_union_set_n_set(domain);
2266 if (graph_alloc(ctx, &graph, graph.n,
2267 isl_union_map_n_map(validity) + isl_union_map_n_map(proximity)) < 0)
2271 if (isl_union_set_foreach_set(domain, &extract_node, &graph) < 0)
2273 if (graph_init_table(ctx, &graph) < 0)
2276 if (isl_union_map_foreach_map(validity, &extract_edge, &graph) < 0)
2278 if (graph_init_edge_table(ctx, &graph) < 0)
2280 if (isl_union_map_foreach_map(proximity, &extract_edge, &graph) < 0)
2283 if (compute_schedule(ctx, &graph) < 0)
2287 sched = extract_schedule(&graph, isl_union_set_get_dim(domain));
2289 graph_free(ctx, &graph);
2290 isl_union_set_free(domain);
2291 isl_union_map_free(validity);
2292 isl_union_map_free(proximity);
2296 graph_free(ctx, &graph);
2297 isl_union_set_free(domain);
2298 isl_union_map_free(validity);
2299 isl_union_map_free(proximity);
2303 void *isl_schedule_free(__isl_take isl_schedule *sched)
2308 for (i = 0; i < sched->n; ++i) {
2309 isl_map_free(sched->node[i].sched);
2310 free(sched->node[i].band_end);
2312 isl_dim_free(sched->dim);
2317 __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched)
2320 isl_union_map *umap;
2325 umap = isl_union_map_empty(isl_dim_copy(sched->dim));
2326 for (i = 0; i < sched->n; ++i)
2327 umap = isl_union_map_add_map(umap,
2328 isl_map_copy(sched->node[i].sched));
2333 int isl_schedule_n_band(__isl_keep isl_schedule *sched)
2335 return sched ? sched->n_band : 0;
2338 /* Construct a mapping that maps each domain to the band in its schedule
2339 * with the specified band index. Note that bands with the same index
2340 * but for different domains do not need to be related.
2342 __isl_give isl_union_map *isl_schedule_get_band(__isl_keep isl_schedule *sched,
2346 isl_union_map *umap;
2351 umap = isl_union_map_empty(isl_dim_copy(sched->dim));
2352 for (i = 0; i < sched->n; ++i) {
2356 if (band >= sched->node[i].n_band)
2359 start = band > 0 ? sched->node[i].band_end[band - 1] : 0;
2360 end = sched->node[i].band_end[band];
2362 map = isl_map_copy(sched->node[i].sched);
2364 map = isl_map_project_out(map, isl_dim_out, end,
2365 sched->n_total_row - end);
2366 map = isl_map_project_out(map, isl_dim_out, 0, start);
2368 umap = isl_union_map_add_map(umap, map);