#include <isl_dim_map.h>
#include <isl_hmap_map_basic_set.h>
#include <isl_qsort.h>
+#include <isl_schedule_private.h>
+#include <isl_band_private.h>
+#include <isl_list_private.h>
/*
* The scheduling algorithm implemented in this file was inspired by
*/
-/* The schedule for an individual domain, plus information about the bands.
- * In particular, we keep track of the number of bands and for each
- * band, the starting position of the next band. The first band starts at
- * position 0.
- */
-struct isl_schedule_node {
- isl_map *sched;
- int n_band;
- int *band_end;
-};
-
-/* Information about the computed schedule.
- * n is the number of nodes/domains/statements.
- * n_band is the maximal number of bands.
- * n_total_row is the number of coordinates of the schedule.
- * dim contains a description of the parameters.
- */
-struct isl_schedule {
- int n;
- int n_band;
- int n_total_row;
- isl_dim *dim;
-
- struct isl_schedule_node node[1];
-};
-
/* Internal information about a node that is used during the construction
* of a schedule.
* dim represents the space in which the domain lives
*
* scc is the index of SCC (or WCC) this node belongs to
*
- * band contains the band index for each of the rows of the schedule
+ * band contains the band index for each of the rows of the schedule.
+ * band_id is used to differentiate between separate bands at the same
+ * level within the same parent band, i.e., bands that are separated
+ * by the parent band or bands that are independent of each other.
+ * zero contains a boolean for each of the rows of the schedule,
+ * indicating whether the corresponding scheduling dimension results
+ * in zero dependence distances within its band and with respect
+ * to the proximity edges.
*
* index, min_index and on_stack are used during the SCC detection
* index represents the order in which nodes are visited.
int scc;
int *band;
+ int *band_id;
+ int *zero;
/* scc detection */
int index;
return 0;
edge = entry->data;
- empty = isl_map_fast_is_empty(edge->map);
+ empty = isl_map_plain_is_empty(edge->map);
if (empty < 0)
return -1;
isl_mat_free(graph->node[i].sched);
isl_map_free(graph->node[i].sched_map);
isl_mat_free(graph->node[i].cmap);
- if (graph->root)
+ if (graph->root) {
free(graph->node[i].band);
+ free(graph->node[i].band_id);
+ free(graph->node[i].zero);
+ }
}
free(graph->node);
free(graph->sorted);
isl_dim *dim;
isl_mat *sched;
struct isl_sched_graph *graph = user;
- int *band;
+ int *band, *band_id, *zero;
ctx = isl_set_get_ctx(set);
dim = isl_set_get_dim(set);
graph->node[graph->n].sched_map = NULL;
band = isl_alloc_array(ctx, int, graph->n_edge + nvar);
graph->node[graph->n].band = band;
+ band_id = isl_calloc_array(ctx, int, graph->n_edge + nvar);
+ graph->node[graph->n].band_id = band_id;
+ zero = isl_calloc_array(ctx, int, graph->n_edge + nvar);
+ graph->node[graph->n].zero = zero;
graph->n++;
- if (!sched || !band)
+ if (!sched || !band || !band_id || !zero)
return -1;
return 0;
if (!entry)
return 0;
edge = entry->data;
- is_equal = isl_map_fast_is_equal(map, edge->map);
+ is_equal = isl_map_plain_is_equal(map, edge->map);
if (is_equal < 0)
return -1;
if (!is_equal)
*
* The constraints are those from the edges plus two or three equalities
* to express the sums.
+ *
+ * If force_zero is set, then we add equalities to ensure that
+ * the sum of the m_n coefficients and m_0 are both zero.
*/
-static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
+static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph,
+ int force_zero)
{
int i, j;
int k;
dim = isl_dim_set_alloc(ctx, 0, total);
isl_basic_set_free(graph->lp);
- n_eq += 2 + parametric;
+ n_eq += 2 + parametric + force_zero;
graph->lp = isl_basic_set_alloc_dim(dim, 0, n_eq, n_ineq);
k = isl_basic_set_alloc_equality(graph->lp);
if (k < 0)
return -1;
isl_seq_clr(graph->lp->eq[k], 1 + total);
- isl_int_set_si(graph->lp->eq[k][1], -1);
+ if (!force_zero)
+ isl_int_set_si(graph->lp->eq[k][1], -1);
for (i = 0; i < 2 * nparam; ++i)
isl_int_set_si(graph->lp->eq[k][1 + param_pos + i], 1);
+ if (force_zero) {
+ k = isl_basic_set_alloc_equality(graph->lp);
+ if (k < 0)
+ return -1;
+ isl_seq_clr(graph->lp->eq[k], 1 + total);
+ isl_int_set_si(graph->lp->eq[k][2], -1);
+ }
+
if (parametric) {
k = isl_basic_set_alloc_equality(graph->lp);
if (k < 0)
* t_i_x such that c_i_x = Q t_i_x and Q is equal to node->cmap.
* In this case, we then also need to perform this multiplication
* to obtain the values of c_i_x.
+ *
+ * If check_zero is set, then the first two coordinates of sol are
+ * assumed to correspond to the dependence distance. If these two
+ * coordinates are zero, then the corresponding scheduling dimension
+ * is marked as being zero distance.
*/
static int update_schedule(struct isl_sched_graph *graph,
- __isl_take isl_vec *sol, int use_cmap)
+ __isl_take isl_vec *sol, int use_cmap, int check_zero)
{
int i, j;
+ int zero = 0;
isl_vec *csol = NULL;
if (!sol)
isl_die(sol->ctx, isl_error_internal,
"no solution found", goto error);
+ if (check_zero)
+ zero = isl_int_is_zero(sol->el[1]) &&
+ isl_int_is_zero(sol->el[2]);
+
for (i = 0; i < graph->n; ++i) {
struct isl_sched_node *node = &graph->node[i];
int pos = node->start;
node->sched = isl_mat_set_element(node->sched,
row, 1 + node->nparam + j, csol->el[j]);
node->band[graph->n_total_row] = graph->n_band;
+ node->zero[graph->n_total_row] = zero;
}
isl_vec_free(sol);
isl_vec_free(csol);
if (!edge->map)
return -1;
- if (isl_map_fast_is_empty(edge->map)) {
+ if (isl_map_plain_is_empty(edge->map)) {
reset_table = 1;
isl_map_free(edge->map);
if (i != graph->n_edge - 1)
if (!sched)
goto error;
+ sched->ref = 1;
sched->n = graph->n;
sched->n_band = graph->n_band;
sched->n_total_row = graph->n_total_row;
for (i = 0; i < sched->n; ++i) {
int r, b;
- int *band_end;
+ int *band_end, *band_id, *zero;
band_end = isl_alloc_array(ctx, int, graph->n_band);
- if (!band_end)
- goto error;
+ band_id = isl_alloc_array(ctx, int, graph->n_band);
+ zero = isl_alloc_array(ctx, int, graph->n_total_row);
sched->node[i].sched = node_extract_schedule(&graph->node[i]);
sched->node[i].band_end = band_end;
+ sched->node[i].band_id = band_id;
+ sched->node[i].zero = zero;
+ if (!band_end || !band_id || !zero)
+ goto error;
+ for (r = 0; r < graph->n_total_row; ++r)
+ zero[r] = graph->node[i].zero[r];
for (r = b = 0; r < graph->n_total_row; ++r) {
if (graph->node[i].band[r] == b)
continue;
if (r == graph->n_total_row)
band_end[b++] = r;
sched->node[i].n_band = b;
+ for (--b; b >= 0; --b)
+ band_id[b] = graph->node[i].band_id[b];
}
sched->dim = dim;
dst->node[dst->n].sched_map =
isl_map_copy(src->node[i].sched_map);
dst->node[dst->n].band = src->node[i].band;
+ dst->node[dst->n].band_id = src->node[i].band_id;
+ dst->node[dst->n].zero = src->node[i].zero;
dst->n++;
}
if (!edge_pred(edge, data))
continue;
- if (isl_map_fast_is_empty(edge->map))
+ if (isl_map_plain_is_empty(edge->map))
continue;
map = isl_map_copy(edge->map);
* It would be possible to reuse them as the first rows in the next
* band, but recomputing them may result in better rows as we are looking
* at a smaller part of the dependence graph.
+ *
+ * The band_id of the second group is set to n, where n is the number
+ * of nodes in the first group. This ensures that the band_ids over
+ * the two groups remain disjoint, even if either or both of the two
+ * groups contain independent components.
*/
static int compute_split_schedule(isl_ctx *ctx, struct isl_sched_graph *graph)
{
graph->n_total_row++;
next_band(graph);
+ for (i = 0; i < graph->n; ++i) {
+ struct isl_sched_node *node = &graph->node[i];
+ if (node->scc > graph->src_scc)
+ node->band_id[graph->n_band] = n;
+ }
+
orig_total_row = graph->n_total_row;
orig_band = graph->n_band;
if (compute_sub_schedule(ctx, graph, n, e1,
return 0;
}
+/* If the schedule_split_parallel option is set and if the linear
+ * parts of the scheduling rows for all nodes in the graphs are the same,
+ * then split off the constant term from the linear part.
+ * The constant term is then placed in a separate band and
+ * the linear part is simplified.
+ */
+static int split_parallel(isl_ctx *ctx, struct isl_sched_graph *graph)
+{
+ int i;
+ int equal = 1;
+ int row, cols;
+ struct isl_sched_node *node0;
+
+ if (!ctx->opt->schedule_split_parallel)
+ return 0;
+ if (graph->n <= 1)
+ return 0;
+
+ node0 = &graph->node[0];
+ row = isl_mat_rows(node0->sched) - 1;
+ cols = isl_mat_cols(node0->sched);
+ for (i = 1; i < graph->n; ++i) {
+ struct isl_sched_node *node = &graph->node[i];
+
+ if (!isl_seq_eq(node0->sched->row[row] + 1,
+ node->sched->row[row] + 1, cols - 1))
+ return 0;
+ if (equal &&
+ isl_int_ne(node0->sched->row[row][0],
+ node->sched->row[row][0]))
+ equal = 0;
+ }
+ if (equal)
+ return 0;
+
+ next_band(graph);
+
+ for (i = 0; i < graph->n; ++i) {
+ struct isl_sched_node *node = &graph->node[i];
+
+ isl_map_free(node->sched_map);
+ node->sched_map = NULL;
+ node->sched = isl_mat_add_zero_rows(node->sched, 1);
+ if (!node->sched)
+ return -1;
+ isl_int_set(node->sched->row[row + 1][0],
+ node->sched->row[row][0]);
+ isl_int_set_si(node->sched->row[row][0], 0);
+ node->sched = isl_mat_normalize_row(node->sched, row);
+ if (!node->sched)
+ return -1;
+ node->band[graph->n_total_row] = graph->n_band;
+ }
+
+ graph->n_total_row++;
+
+ return 0;
+}
+
/* Construct a schedule row for each node such that as many dependences
* as possible are carried and then continue with the next band.
*/
"unable to carry dependences", return -1);
}
- if (update_schedule(graph, sol, 0) < 0)
+ if (update_schedule(graph, sol, 0, 0) < 0)
+ return -1;
+
+ if (split_parallel(ctx, graph) < 0)
return -1;
return compute_next_band(ctx, graph);
*
* If we manage to complete the schedule, we finish off by topologically
* sorting the statements based on the remaining dependences.
+ *
+ * If ctx->opt->schedule_outer_zero_distance is set, then we force the
+ * outermost dimension in the current band to be zero distance. If this
+ * turns out to be impossible, we fall back on the general scheme above
+ * and try to carry as many dependences as possible.
*/
static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph)
{
+ int force_zero = 0;
+
if (detect_sccs(graph) < 0)
return -1;
sort_sccs(graph);
if (compute_maxvar(graph) < 0)
return -1;
+ if (ctx->opt->schedule_outer_zero_distance)
+ force_zero = 1;
+
while (graph->n_row < graph->maxvar) {
isl_vec *sol;
graph->src_scc = -1;
graph->dst_scc = -1;
- if (setup_lp(ctx, graph) < 0)
+ if (setup_lp(ctx, graph, force_zero) < 0)
return -1;
sol = solve_lp(graph);
if (!sol)
return -1;
if (sol->size == 0) {
isl_vec_free(sol);
+ if (!ctx->opt->schedule_maximize_band_depth &&
+ graph->n_total_row > graph->band_start)
+ return compute_next_band(ctx, graph);
if (graph->src_scc >= 0)
return compute_split_schedule(ctx, graph);
if (graph->n_total_row > graph->band_start)
return compute_next_band(ctx, graph);
return carry_dependences(ctx, graph);
}
- if (update_schedule(graph, sol, 1) < 0)
+ if (update_schedule(graph, sol, 1, 1) < 0)
return -1;
+ force_zero = 0;
}
if (graph->n_total_row > graph->band_start)
/* Compute a schedule for each component (identified by node->scc)
* of the dependence graph separately and then combine the results.
+ *
+ * The band_id is adjusted such that each component has a separate id.
+ * Note that the band_id may have already been set to a value different
+ * from zero by compute_split_schedule.
*/
static int compute_component_schedule(isl_ctx *ctx,
struct isl_sched_graph *graph)
orig_total_row = graph->n_total_row;
n_band = 0;
orig_band = graph->n_band;
+ for (i = 0; i < graph->n; ++i)
+ graph->node[i].band_id[graph->n_band] += graph->node[i].scc;
for (wcc = 0; wcc < graph->scc; ++wcc) {
n = 0;
for (i = 0; i < graph->n; ++i)
int i;
if (!sched)
return NULL;
+
+ if (--sched->ref > 0)
+ return NULL;
+
for (i = 0; i < sched->n; ++i) {
isl_map_free(sched->node[i].sched);
free(sched->node[i].band_end);
+ free(sched->node[i].band_id);
+ free(sched->node[i].zero);
}
isl_dim_free(sched->dim);
+ isl_band_list_free(sched->band_forest);
free(sched);
return NULL;
}
+isl_ctx *isl_schedule_get_ctx(__isl_keep isl_schedule *schedule)
+{
+ return schedule ? isl_dim_get_ctx(schedule->dim) : NULL;
+}
+
__isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched)
{
int i;
return umap;
}
-int isl_schedule_n_band(__isl_keep isl_schedule *sched)
+static __isl_give isl_band_list *construct_band_list(
+ __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
+ int band_nr, int *parent_active, int n_active);
+
+/* Construct an isl_band structure for the band in the given schedule
+ * with sequence number band_nr for the n_active nodes marked by active.
+ * If the nodes don't have a band with the given sequence number,
+ * then a band without members is created.
+ *
+ * Because of the way the schedule is constructed, we know that
+ * the position of the band inside the schedule of a node is the same
+ * for all active nodes.
+ */
+static __isl_give isl_band *construct_band(__isl_keep isl_schedule *schedule,
+ __isl_keep isl_band *parent,
+ int band_nr, int *active, int n_active)
+{
+ int i, j;
+ isl_ctx *ctx = isl_schedule_get_ctx(schedule);
+ isl_band *band;
+ unsigned start, end;
+
+ band = isl_calloc_type(ctx, isl_band);
+ if (!band)
+ return NULL;
+
+ band->ref = 1;
+ band->schedule = schedule;
+ band->parent = parent;
+
+ for (i = 0; i < schedule->n; ++i)
+ if (active[i] && schedule->node[i].n_band > band_nr + 1)
+ break;
+
+ if (i < schedule->n) {
+ band->children = construct_band_list(schedule, band,
+ band_nr + 1, active, n_active);
+ if (!band->children)
+ goto error;
+ }
+
+ for (i = 0; i < schedule->n; ++i)
+ if (active[i])
+ break;
+
+ if (i >= schedule->n)
+ isl_die(ctx, isl_error_internal,
+ "band without active statements", goto error);
+
+ start = band_nr ? schedule->node[i].band_end[band_nr - 1] : 0;
+ end = band_nr < schedule->node[i].n_band ?
+ schedule->node[i].band_end[band_nr] : start;
+ band->n = end - start;
+
+ band->zero = isl_alloc_array(ctx, int, band->n);
+ if (!band->zero)
+ goto error;
+
+ for (j = 0; j < band->n; ++j)
+ band->zero[j] = schedule->node[i].zero[start + j];
+
+ band->map = isl_union_map_empty(isl_dim_copy(schedule->dim));
+ for (i = 0; i < schedule->n; ++i) {
+ isl_map *map;
+ unsigned n_out;
+
+ if (!active[i])
+ continue;
+
+ map = isl_map_copy(schedule->node[i].sched);
+ n_out = isl_map_dim(map, isl_dim_out);
+ map = isl_map_project_out(map, isl_dim_out, end, n_out - end);
+ map = isl_map_project_out(map, isl_dim_out, 0, start);
+ band->map = isl_union_map_union(band->map,
+ isl_union_map_from_map(map));
+ }
+ if (!band->map)
+ goto error;
+
+ return band;
+error:
+ isl_band_free(band);
+ return NULL;
+}
+
+/* Construct a list of bands that start at the same position (with
+ * sequence number band_nr) in the schedules of the nodes that
+ * were active in the parent band.
+ *
+ * A separate isl_band structure is created for each band_id
+ * and for each node that does not have a band with sequence
+ * number band_nr. In the latter case, a band without members
+ * is created.
+ * This ensures that if a band has any children, then each node
+ * that was active in the band is active in exactly one of the children.
+ */
+static __isl_give isl_band_list *construct_band_list(
+ __isl_keep isl_schedule *schedule, __isl_keep isl_band *parent,
+ int band_nr, int *parent_active, int n_active)
{
- return sched ? sched->n_band : 0;
+ int i, j;
+ isl_ctx *ctx = isl_schedule_get_ctx(schedule);
+ int *active;
+ int n_band;
+ isl_band_list *list;
+
+ n_band = 0;
+ for (i = 0; i < n_active; ++i) {
+ for (j = 0; j < schedule->n; ++j) {
+ if (!parent_active[j])
+ continue;
+ if (schedule->node[j].n_band <= band_nr)
+ continue;
+ if (schedule->node[j].band_id[band_nr] == i) {
+ n_band++;
+ break;
+ }
+ }
+ }
+ for (j = 0; j < schedule->n; ++j)
+ if (schedule->node[j].n_band <= band_nr)
+ n_band++;
+
+ if (n_band == 1) {
+ isl_band *band;
+ list = isl_band_list_alloc(ctx, n_band);
+ band = construct_band(schedule, parent, band_nr,
+ parent_active, n_active);
+ return isl_band_list_add(list, band);
+ }
+
+ active = isl_alloc_array(ctx, int, schedule->n);
+ if (!active)
+ return NULL;
+
+ list = isl_band_list_alloc(ctx, n_band);
+
+ for (i = 0; i < n_active; ++i) {
+ int n = 0;
+ isl_band *band;
+
+ for (j = 0; j < schedule->n; ++j) {
+ active[j] = parent_active[j] &&
+ schedule->node[j].n_band > band_nr &&
+ schedule->node[j].band_id[band_nr] == i;
+ if (active[j])
+ n++;
+ }
+ if (n == 0)
+ continue;
+
+ band = construct_band(schedule, parent, band_nr, active, n);
+
+ list = isl_band_list_add(list, band);
+ }
+ for (i = 0; i < schedule->n; ++i) {
+ isl_band *band;
+ if (!parent_active[i])
+ continue;
+ if (schedule->node[i].n_band > band_nr)
+ continue;
+ for (j = 0; j < schedule->n; ++j)
+ active[j] = j == i;
+ band = construct_band(schedule, parent, band_nr, active, 1);
+ list = isl_band_list_add(list, band);
+ }
+
+ free(active);
+
+ return list;
}
-/* Construct a mapping that maps each domain to the band in its schedule
- * with the specified band index. Note that bands with the same index
- * but for different domains do not need to be related.
+/* Construct a band forest representation of the schedule and
+ * return the list of roots.
*/
-__isl_give isl_union_map *isl_schedule_get_band(__isl_keep isl_schedule *sched,
- unsigned band)
+static __isl_give isl_band_list *construct_forest(
+ __isl_keep isl_schedule *schedule)
{
int i;
- isl_union_map *umap;
+ isl_ctx *ctx = isl_schedule_get_ctx(schedule);
+ isl_band_list *forest;
+ int *active;
- if (!sched)
+ active = isl_alloc_array(ctx, int, schedule->n);
+ if (!active)
return NULL;
- umap = isl_union_map_empty(isl_dim_copy(sched->dim));
- for (i = 0; i < sched->n; ++i) {
- int start, end;
- isl_map *map;
+ for (i = 0; i < schedule->n; ++i)
+ active[i] = 1;
- if (band >= sched->node[i].n_band)
- continue;
+ forest = construct_band_list(schedule, NULL, 0, active, schedule->n);
- start = band > 0 ? sched->node[i].band_end[band - 1] : 0;
- end = sched->node[i].band_end[band];
+ free(active);
- map = isl_map_copy(sched->node[i].sched);
+ return forest;
+}
- map = isl_map_project_out(map, isl_dim_out, end,
- sched->n_total_row - end);
- map = isl_map_project_out(map, isl_dim_out, 0, start);
+/* Return the roots of a band forest representation of the schedule.
+ */
+__isl_give isl_band_list *isl_schedule_get_band_forest(
+ __isl_keep isl_schedule *schedule)
+{
+ if (!schedule)
+ return NULL;
+ if (!schedule->band_forest)
+ schedule->band_forest = construct_forest(schedule);
+ return isl_band_list_dup(schedule->band_forest);
+}
+
+static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
+ __isl_keep isl_band_list *list);
+
+static __isl_give isl_printer *print_band(__isl_take isl_printer *p,
+ __isl_keep isl_band *band)
+{
+ isl_band_list *children;
+
+ p = isl_printer_start_line(p);
+ p = isl_printer_print_union_map(p, band->map);
+ p = isl_printer_end_line(p);
+
+ if (!isl_band_has_children(band))
+ return p;
- umap = isl_union_map_add_map(umap, map);
+ children = isl_band_get_children(band);
+
+ p = isl_printer_indent(p, 4);
+ p = print_band_list(p, children);
+ p = isl_printer_indent(p, -4);
+
+ isl_band_list_free(children);
+
+ return p;
+}
+
+static __isl_give isl_printer *print_band_list(__isl_take isl_printer *p,
+ __isl_keep isl_band_list *list)
+{
+ int i, n;
+
+ n = isl_band_list_n_band(list);
+ for (i = 0; i < n; ++i) {
+ isl_band *band;
+ band = isl_band_list_get_band(list, i);
+ p = print_band(p, band);
+ isl_band_free(band);
}
- return umap;
+ return p;
+}
+
+__isl_give isl_printer *isl_printer_print_schedule(__isl_take isl_printer *p,
+ __isl_keep isl_schedule *schedule)
+{
+ isl_band_list *forest;
+
+ forest = isl_schedule_get_band_forest(schedule);
+
+ p = print_band_list(p, forest);
+
+ isl_band_list_free(forest);
+
+ return p;
+}
+
+void isl_schedule_dump(__isl_keep isl_schedule *schedule)
+{
+ isl_printer *printer;
+
+ if (!schedule)
+ return;
+
+ printer = isl_printer_to_file(isl_schedule_get_ctx(schedule), stderr);
+ printer = isl_printer_print_schedule(printer, schedule);
+
+ isl_printer_free(printer);
}