X-Git-Url: http://review.tizen.org/git/?a=blobdiff_plain;f=isl_schedule.c;h=9f261e0ffc210c6696af19f6183fd27193fc887c;hb=f5e4e6f324ebd2eb8edb47e57880be621a3e58e1;hp=925ffe32a22dc4a235c098106bd401249894d192;hpb=fd1fbc971cfbbb85e3bada17c104e469845ab52e;p=platform%2Fupstream%2Fisl.git diff --git a/isl_schedule.c b/isl_schedule.c index 925ffe3..9f261e0 100644 --- a/isl_schedule.c +++ b/isl_schedule.c @@ -10,7 +10,7 @@ #include #include -#include +#include #include #include #include @@ -22,6 +22,9 @@ #include #include #include +#include +#include +#include /* * The scheduling algorithm implemented in this file was inspired by @@ -50,7 +53,14 @@ * * 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. @@ -58,7 +68,7 @@ * on_stack indicates whether the node is currently on the stack. */ struct isl_sched_node { - isl_dim *dim; + isl_space *dim; isl_mat *sched; isl_map *sched_map; int rank; @@ -70,6 +80,8 @@ struct isl_sched_node { int scc; int *band; + int *band_id; + int *zero; /* scc detection */ int index; @@ -80,9 +92,9 @@ struct isl_sched_node { static int node_has_dim(const void *entry, const void *val) { struct isl_sched_node *node = (struct isl_sched_node *)entry; - isl_dim *dim = (isl_dim *)val; + isl_space *dim = (isl_space *)val; - return isl_dim_equal(node->dim, dim); + return isl_space_is_equal(node->dim, dim); } /* An edge in the dependence graph. An edge may be used to @@ -206,7 +218,7 @@ static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph) struct isl_hash_table_entry *entry; uint32_t hash; - hash = isl_dim_get_hash(graph->node[i].dim); + hash = isl_space_get_hash(graph->node[i].dim); entry = isl_hash_table_find(ctx, graph->node_table, hash, &node_has_dim, graph->node[i].dim, 1); @@ -222,12 +234,12 @@ static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph) * or NULL if there is no such node. */ static struct isl_sched_node *graph_find_node(isl_ctx *ctx, - struct isl_sched_graph *graph, __isl_keep isl_dim *dim) + struct isl_sched_graph *graph, __isl_keep isl_space *dim) { struct isl_hash_table_entry *entry; uint32_t hash; - hash = isl_dim_get_hash(dim); + hash = isl_space_get_hash(dim); entry = isl_hash_table_find(ctx, graph->node_table, hash, &node_has_dim, dim, 0); @@ -280,7 +292,7 @@ static int graph_init_edge_table(isl_ctx *ctx, struct isl_sched_graph *graph) static int graph_has_edge(struct isl_sched_graph *graph, struct isl_sched_node *src, struct isl_sched_node *dst) { - isl_ctx *ctx = isl_dim_get_ctx(src->dim); + isl_ctx *ctx = isl_space_get_ctx(src->dim); struct isl_hash_table_entry *entry; uint32_t hash; struct isl_sched_edge temp = { .src = src, .dst = dst }; @@ -338,12 +350,15 @@ static void graph_free(isl_ctx *ctx, struct isl_sched_graph *graph) isl_hmap_map_basic_set_free(ctx, graph->inter_hmap); for (i = 0; i < graph->n; ++i) { - isl_dim_free(graph->node[i].dim); + isl_space_free(graph->node[i].dim); 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); @@ -363,16 +378,16 @@ static int extract_node(__isl_take isl_set *set, void *user) { int nvar, nparam; isl_ctx *ctx; - isl_dim *dim; + isl_space *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); + dim = isl_set_get_space(set); isl_set_free(set); - nvar = isl_dim_size(dim, isl_dim_set); - nparam = isl_dim_size(dim, isl_dim_param); + nvar = isl_space_dim(dim, isl_dim_set); + nparam = isl_space_dim(dim, isl_dim_param); if (!ctx->opt->schedule_parametric) nparam = 0; sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar); @@ -383,9 +398,13 @@ static int extract_node(__isl_take isl_set *set, void *user) 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; @@ -405,14 +424,14 @@ static int extract_edge(__isl_take isl_map *map, void *user) isl_ctx *ctx = isl_map_get_ctx(map); struct isl_sched_graph *graph = user; struct isl_sched_node *src, *dst; - isl_dim *dim; + isl_space *dim; - dim = isl_dim_domain(isl_map_get_dim(map)); + dim = isl_space_domain(isl_map_get_space(map)); src = graph_find_node(ctx, graph, dim); - isl_dim_free(dim); - dim = isl_dim_range(isl_map_get_dim(map)); + isl_space_free(dim); + dim = isl_space_range(isl_map_get_space(map)); dst = graph_find_node(ctx, graph, dim); - isl_dim_free(dim); + isl_space_free(dim); if (!src || !dst) { isl_map_free(map); @@ -656,31 +675,31 @@ static int add_intra_validity_constraints(struct isl_sched_graph *graph, unsigned total; isl_map *map = isl_map_copy(edge->map); isl_ctx *ctx = isl_map_get_ctx(map); - isl_dim *dim; + isl_space *dim; isl_dim_map *dim_map; isl_basic_set *coef; struct isl_sched_node *node = edge->src; coef = intra_coefficients(graph, map); - dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef))); + dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); coef = isl_basic_set_transform_dims(coef, isl_dim_set, - isl_dim_size(dim, isl_dim_set), isl_mat_copy(node->cmap)); + isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap)); total = isl_basic_set_total_dim(graph->lp); dim_map = isl_dim_map_alloc(ctx, total); isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, node->nvar, -1); isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, node->nvar, 1); graph->lp = isl_basic_set_extend_constraints(graph->lp, coef->n_eq, coef->n_ineq); graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, coef, dim_map); - isl_dim_free(dim); + isl_space_free(dim); return 0; } @@ -709,7 +728,7 @@ static int add_inter_validity_constraints(struct isl_sched_graph *graph, unsigned total; isl_map *map = isl_map_copy(edge->map); isl_ctx *ctx = isl_map_get_ctx(map); - isl_dim *dim; + isl_space *dim; isl_dim_map *dim_map; isl_basic_set *coef; struct isl_sched_node *src = edge->src; @@ -717,12 +736,12 @@ static int add_inter_validity_constraints(struct isl_sched_graph *graph, coef = inter_coefficients(graph, map); - dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef))); + dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); coef = isl_basic_set_transform_dims(coef, isl_dim_set, - isl_dim_size(dim, isl_dim_set), isl_mat_copy(src->cmap)); + isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap)); coef = isl_basic_set_transform_dims(coef, isl_dim_set, - isl_dim_size(dim, isl_dim_set) + src->nvar, + isl_space_dim(dim, isl_dim_set) + src->nvar, isl_mat_copy(dst->cmap)); total = isl_basic_set_total_dim(graph->lp); @@ -732,20 +751,20 @@ static int add_inter_validity_constraints(struct isl_sched_graph *graph, isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1); isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1); isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set) + src->nvar, 1, + isl_space_dim(dim, isl_dim_set) + src->nvar, 1, dst->nvar, -1); isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set) + src->nvar, 1, + isl_space_dim(dim, isl_dim_set) + src->nvar, 1, dst->nvar, 1); isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1); isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1); isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1); isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, src->nvar, 1); isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, src->nvar, -1); edge->start = graph->lp->n_ineq; @@ -753,7 +772,7 @@ static int add_inter_validity_constraints(struct isl_sched_graph *graph, coef->n_eq, coef->n_ineq); graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, coef, dim_map); - isl_dim_free(dim); + isl_space_free(dim); edge->end = graph->lp->n_ineq; return 0; @@ -795,35 +814,35 @@ static int add_intra_proximity_constraints(struct isl_sched_graph *graph, unsigned nparam; isl_map *map = isl_map_copy(edge->map); isl_ctx *ctx = isl_map_get_ctx(map); - isl_dim *dim; + isl_space *dim; isl_dim_map *dim_map; isl_basic_set *coef; struct isl_sched_node *node = edge->src; coef = intra_coefficients(graph, map); - dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef))); + dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); coef = isl_basic_set_transform_dims(coef, isl_dim_set, - isl_dim_size(dim, isl_dim_set), isl_mat_copy(node->cmap)); + isl_space_dim(dim, isl_dim_set), isl_mat_copy(node->cmap)); - nparam = isl_dim_size(node->dim, isl_dim_param); + nparam = isl_space_dim(node->dim, isl_dim_param); total = isl_basic_set_total_dim(graph->lp); dim_map = isl_dim_map_alloc(ctx, total); isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1); isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1); isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1); isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, node->nvar, s); isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, node->nvar, -s); graph->lp = isl_basic_set_extend_constraints(graph->lp, coef->n_eq, coef->n_ineq); graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, coef, dim_map); - isl_dim_free(dim); + isl_space_free(dim); return 0; } @@ -870,7 +889,7 @@ static int add_inter_proximity_constraints(struct isl_sched_graph *graph, unsigned nparam; isl_map *map = isl_map_copy(edge->map); isl_ctx *ctx = isl_map_get_ctx(map); - isl_dim *dim; + isl_space *dim; isl_dim_map *dim_map; isl_basic_set *coef; struct isl_sched_node *src = edge->src; @@ -878,15 +897,15 @@ static int add_inter_proximity_constraints(struct isl_sched_graph *graph, coef = inter_coefficients(graph, map); - dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef))); + dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); coef = isl_basic_set_transform_dims(coef, isl_dim_set, - isl_dim_size(dim, isl_dim_set), isl_mat_copy(src->cmap)); + isl_space_dim(dim, isl_dim_set), isl_mat_copy(src->cmap)); coef = isl_basic_set_transform_dims(coef, isl_dim_set, - isl_dim_size(dim, isl_dim_set) + src->nvar, + isl_space_dim(dim, isl_dim_set) + src->nvar, isl_mat_copy(dst->cmap)); - nparam = isl_dim_size(src->dim, isl_dim_param); + nparam = isl_space_dim(src->dim, isl_dim_param); total = isl_basic_set_total_dim(graph->lp); dim_map = isl_dim_map_alloc(ctx, total); @@ -898,27 +917,27 @@ static int add_inter_proximity_constraints(struct isl_sched_graph *graph, isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, s); isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, -s); isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set) + src->nvar, 1, + isl_space_dim(dim, isl_dim_set) + src->nvar, 1, dst->nvar, s); isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set) + src->nvar, 1, + isl_space_dim(dim, isl_dim_set) + src->nvar, 1, dst->nvar, -s); isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, s); isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, -s); isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, s); isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, src->nvar, -s); isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, src->nvar, s); graph->lp = isl_basic_set_extend_constraints(graph->lp, coef->n_eq, coef->n_ineq); graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, coef, dim_map); - isl_dim_free(dim); + isl_space_free(dim); return 0; } @@ -1022,7 +1041,36 @@ static int node_update_cmap(struct isl_sched_node *node) } /* Count the number of equality and inequality constraints - * that will be added. If once is set, then we count + * that will be added for the given map. + * If once is set, then we count + * each edge exactly once. Otherwise, we count as follows + * validity -> 1 (>= 0) + * validity+proximity -> 2 (>= 0 and upper bound) + * proximity -> 2 (lower and upper bound) + */ +static int count_map_constraints(struct isl_sched_graph *graph, + struct isl_sched_edge *edge, __isl_take isl_map *map, + int *n_eq, int *n_ineq, int once) +{ + isl_basic_set *coef; + int f = once ? 1 : edge->proximity ? 2 : 1; + + if (edge->src == edge->dst) + coef = intra_coefficients(graph, map); + else + coef = inter_coefficients(graph, map); + if (!coef) + return -1; + *n_eq += f * coef->n_eq; + *n_ineq += f * coef->n_ineq; + isl_basic_set_free(coef); + + return 0; +} + +/* Count the number of equality and inequality constraints + * that will be added to the main lp problem. + * If once is set, then we count * each edge exactly once. Otherwise, we count as follows * validity -> 1 (>= 0) * validity+proximity -> 2 (>= 0 and upper bound) @@ -1032,23 +1080,15 @@ static int count_constraints(struct isl_sched_graph *graph, int *n_eq, int *n_ineq, int once) { int i; - isl_basic_set *coef; *n_eq = *n_ineq = 0; for (i = 0; i < graph->n_edge; ++i) { struct isl_sched_edge *edge= &graph->edge[i]; isl_map *map = isl_map_copy(edge->map); - int f = once ? 1 : edge->proximity ? 2 : 1; - if (edge->src == edge->dst) - coef = intra_coefficients(graph, map); - else - coef = inter_coefficients(graph, map); - if (!coef) + if (count_map_constraints(graph, edge, map, + n_eq, n_ineq, once) < 0) return -1; - *n_eq += f * coef->n_eq; - *n_ineq += f * coef->n_ineq; - isl_basic_set_free(coef); } return 0; @@ -1084,20 +1124,24 @@ static int count_constraints(struct isl_sched_graph *graph, * * 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; unsigned nparam; unsigned total; - isl_dim *dim; + isl_space *dim; int parametric; int param_pos; int n_eq, n_ineq; parametric = ctx->opt->schedule_parametric; - nparam = isl_dim_size(graph->node[0].dim, isl_dim_param); + nparam = isl_space_dim(graph->node[0].dim, isl_dim_param); param_pos = 4; total = param_pos + 2 * nparam; for (i = 0; i < graph->n; ++i) { @@ -1111,19 +1155,28 @@ static int setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph) if (count_constraints(graph, &n_eq, &n_ineq, 0) < 0) return -1; - dim = isl_dim_set_alloc(ctx, 0, total); + dim = isl_space_set_alloc(ctx, 0, total); isl_basic_set_free(graph->lp); - n_eq += 2 + parametric; - graph->lp = isl_basic_set_alloc_dim(dim, 0, n_eq, n_ineq); + n_eq += 2 + parametric + force_zero; + graph->lp = isl_basic_set_alloc_space(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) @@ -1250,11 +1303,17 @@ static __isl_give isl_vec *solve_lp(struct isl_sched_graph *graph) * 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) @@ -1263,6 +1322,10 @@ static int update_schedule(struct isl_sched_graph *graph, 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; @@ -1299,6 +1362,7 @@ static int update_schedule(struct isl_sched_graph *graph, 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); @@ -1324,7 +1388,8 @@ error: static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node) { int i, j; - isl_dim *dim; + isl_space *dim; + isl_local_space *ls; isl_basic_map *bmap; isl_constraint *c; int nrow, ncol; @@ -1335,14 +1400,15 @@ static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node) nrow = isl_mat_rows(node->sched); ncol = isl_mat_cols(node->sched) - 1; - dim = isl_dim_from_domain(isl_dim_copy(node->dim)); - dim = isl_dim_add(dim, isl_dim_out, nrow); - bmap = isl_basic_map_universe(isl_dim_copy(dim)); + dim = isl_space_from_domain(isl_space_copy(node->dim)); + dim = isl_space_add_dims(dim, isl_dim_out, nrow); + bmap = isl_basic_map_universe(isl_space_copy(dim)); + ls = isl_local_space_from_space(dim); isl_int_init(v); for (i = 0; i < nrow; ++i) { - c = isl_equality_alloc(isl_dim_copy(dim)); + c = isl_equality_alloc(isl_local_space_copy(ls)); isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1); isl_mat_get_element(node->sched, i, 0, &v); isl_constraint_set_constant(c, v); @@ -1360,7 +1426,7 @@ static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node) isl_int_clear(v); - isl_dim_free(dim); + isl_local_space_free(ls); node->sched_map = isl_map_from_basic_map(bmap); return isl_map_copy(node->sched_map); @@ -1470,7 +1536,7 @@ static int sort_statements(isl_ctx *ctx, struct isl_sched_graph *graph) * in graph and with parameters specified by dim. */ static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph, - __isl_take isl_dim *dim) + __isl_take isl_space *dim) { int i; isl_ctx *ctx; @@ -1479,27 +1545,34 @@ static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph, if (!dim) return NULL; - ctx = isl_dim_get_ctx(dim); + ctx = isl_space_get_ctx(dim); sched = isl_calloc(ctx, struct isl_schedule, sizeof(struct isl_schedule) + (graph->n - 1) * sizeof(struct isl_schedule_node)); 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; @@ -1510,13 +1583,15 @@ static __isl_give isl_schedule *extract_schedule(struct isl_sched_graph *graph, 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; return sched; error: - isl_dim_free(dim); + isl_space_free(dim); isl_schedule_free(sched); return NULL; } @@ -1533,13 +1608,15 @@ static int copy_nodes(struct isl_sched_graph *dst, struct isl_sched_graph *src, for (i = 0; i < src->n; ++i) { if (!node_pred(&src->node[i], data)) continue; - dst->node[dst->n].dim = isl_dim_copy(src->node[i].dim); + dst->node[dst->n].dim = isl_space_copy(src->node[i].dim); dst->node[dst->n].nvar = src->node[i].nvar; dst->node[dst->n].nparam = src->node[i].nparam; dst->node[dst->n].sched = isl_mat_copy(src->node[i].sched); 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++; } @@ -1752,6 +1829,11 @@ static int pad_schedule(struct isl_sched_graph *graph) * 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) { @@ -1800,6 +1882,12 @@ 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, @@ -1834,9 +1922,10 @@ static int compute_next_band(isl_ctx *ctx, struct isl_sched_graph *graph) return compute_schedule(ctx, graph); } -/* Add constraints to graph->lp that force the dependence of edge i - * to be respected and attempt to carry it, where edge i is one from - * a node j to itself. +/* Add constraints to graph->lp that force the dependence "map" (which + * is part of the dependence relation of "edge") + * to be respected and attempt to carry it, where the edge is one from + * a node j to itself. "pos" is the sequence number of the given map. * That is, add constraints that enforce * * (c_j_0 + c_j_n n + c_j_x y) - (c_j_0 + c_j_n n + c_j_x x) @@ -1848,42 +1937,42 @@ static int compute_next_band(isl_ctx *ctx, struct isl_sched_graph *graph) * with each coefficient in c_j_x represented as a pair of non-negative * coefficients. */ -static int add_intra_constraints(struct isl_sched_graph *graph, int i) +static int add_intra_constraints(struct isl_sched_graph *graph, + struct isl_sched_edge *edge, __isl_take isl_map *map, int pos) { unsigned total; - struct isl_sched_edge *edge= &graph->edge[i]; - isl_map *map = isl_map_copy(edge->map); isl_ctx *ctx = isl_map_get_ctx(map); - isl_dim *dim; + isl_space *dim; isl_dim_map *dim_map; isl_basic_set *coef; struct isl_sched_node *node = edge->src; coef = intra_coefficients(graph, map); - dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef))); + dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); total = isl_basic_set_total_dim(graph->lp); dim_map = isl_dim_map_alloc(ctx, total); - isl_dim_map_range(dim_map, 3 + i, 0, 0, 0, 1, -1); + isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1); isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, node->nvar, -1); isl_dim_map_range(dim_map, node->start + 2 * node->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, node->nvar, 1); graph->lp = isl_basic_set_extend_constraints(graph->lp, coef->n_eq, coef->n_ineq); graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, coef, dim_map); - isl_dim_free(dim); + isl_space_free(dim); return 0; } -/* Add constraints to graph->lp that force the dependence of edge i - * to be respected and attempt to carry it, where edge i is one from - * node j to node k. +/* Add constraints to graph->lp that force the dependence "map" (which + * is part of the dependence relation of "edge") + * to be respected and attempt to carry it, where the edge is one from + * node j to node k. "pos" is the sequence number of the given map. * That is, add constraints that enforce * * (c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i @@ -1895,13 +1984,12 @@ static int add_intra_constraints(struct isl_sched_graph *graph, int i) * with each coefficient (except e_i, c_k_0 and c_j_0) * represented as a pair of non-negative coefficients. */ -static int add_inter_constraints(struct isl_sched_graph *graph, int i) +static int add_inter_constraints(struct isl_sched_graph *graph, + struct isl_sched_edge *edge, __isl_take isl_map *map, int pos) { unsigned total; - struct isl_sched_edge *edge= &graph->edge[i]; - isl_map *map = isl_map_copy(edge->map); isl_ctx *ctx = isl_map_get_ctx(map); - isl_dim *dim; + isl_space *dim; isl_dim_map *dim_map; isl_basic_set *coef; struct isl_sched_node *src = edge->src; @@ -1909,38 +1997,38 @@ static int add_inter_constraints(struct isl_sched_graph *graph, int i) coef = inter_coefficients(graph, map); - dim = isl_dim_domain(isl_dim_unwrap(isl_basic_set_get_dim(coef))); + dim = isl_space_domain(isl_space_unwrap(isl_basic_set_get_space(coef))); total = isl_basic_set_total_dim(graph->lp); dim_map = isl_dim_map_alloc(ctx, total); - isl_dim_map_range(dim_map, 3 + i, 0, 0, 0, 1, -1); + isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1); isl_dim_map_range(dim_map, dst->start, 0, 0, 0, 1, 1); isl_dim_map_range(dim_map, dst->start + 1, 2, 1, 1, dst->nparam, -1); isl_dim_map_range(dim_map, dst->start + 2, 2, 1, 1, dst->nparam, 1); isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set) + src->nvar, 1, + isl_space_dim(dim, isl_dim_set) + src->nvar, 1, dst->nvar, -1); isl_dim_map_range(dim_map, dst->start + 2 * dst->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set) + src->nvar, 1, + isl_space_dim(dim, isl_dim_set) + src->nvar, 1, dst->nvar, 1); isl_dim_map_range(dim_map, src->start, 0, 0, 0, 1, -1); isl_dim_map_range(dim_map, src->start + 1, 2, 1, 1, src->nparam, 1); isl_dim_map_range(dim_map, src->start + 2, 2, 1, 1, src->nparam, -1); isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 1, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, src->nvar, 1); isl_dim_map_range(dim_map, src->start + 2 * src->nparam + 2, 2, - isl_dim_size(dim, isl_dim_set), 1, + isl_space_dim(dim, isl_dim_set), 1, src->nvar, -1); graph->lp = isl_basic_set_extend_constraints(graph->lp, coef->n_eq, coef->n_ineq); graph->lp = isl_basic_set_add_constraints_dim_map(graph->lp, coef, dim_map); - isl_dim_free(dim); + isl_space_free(dim); return 0; } @@ -1950,16 +2038,59 @@ static int add_inter_constraints(struct isl_sched_graph *graph, int i) */ static int add_all_constraints(struct isl_sched_graph *graph) { - int i; + int i, j; + int pos; + pos = 0; for (i = 0; i < graph->n_edge; ++i) { struct isl_sched_edge *edge= &graph->edge[i]; - if (edge->src == edge->dst && - add_intra_constraints(graph, i) < 0) - return -1; - if (edge->src != edge->dst && - add_inter_constraints(graph, i) < 0) - return -1; + for (j = 0; j < edge->map->n; ++j) { + isl_basic_map *bmap; + isl_map *map; + + bmap = isl_basic_map_copy(edge->map->p[j]); + map = isl_map_from_basic_map(bmap); + + if (edge->src == edge->dst && + add_intra_constraints(graph, edge, map, pos) < 0) + return -1; + if (edge->src != edge->dst && + add_inter_constraints(graph, edge, map, pos) < 0) + return -1; + ++pos; + } + } + + return 0; +} + +/* Count the number of equality and inequality constraints + * that will be added to the carry_lp problem. + * If once is set, then we count + * each edge exactly once. Otherwise, we count as follows + * validity -> 1 (>= 0) + * validity+proximity -> 2 (>= 0 and upper bound) + * proximity -> 2 (lower and upper bound) + */ +static int count_all_constraints(struct isl_sched_graph *graph, + int *n_eq, int *n_ineq, int once) +{ + int i, j; + + *n_eq = *n_ineq = 0; + for (i = 0; i < graph->n_edge; ++i) { + struct isl_sched_edge *edge= &graph->edge[i]; + for (j = 0; j < edge->map->n; ++j) { + isl_basic_map *bmap; + isl_map *map; + + bmap = isl_basic_map_copy(edge->map->p[j]); + map = isl_map_from_basic_map(bmap); + + if (count_map_constraints(graph, edge, map, + n_eq, n_ineq, once) < 0) + return -1; + } } return 0; @@ -1971,6 +2102,11 @@ static int add_all_constraints(struct isl_sched_graph *graph) * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum * of all e_i's. Dependence with e_i = 0 in the solution are simply * respected, while those with e_i > 0 (in practice e_i = 1) are carried. + * Note that if the dependence relation is a union of basic maps, + * then we have to consider each basic map individually as it may only + * be possible to carry the dependences expressed by some of those + * basic maps and not all off them. + * Below, we consider each of those basic maps as a separate "edge". * * All variables of the LP are non-negative. The actual coefficients * may be negative, so each coefficient is represented as the difference @@ -1996,34 +2132,39 @@ static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph) { int i, j; int k; - isl_dim *dim; + isl_space *dim; unsigned total; int n_eq, n_ineq; + int n_edge; - total = 3 + graph->n_edge; + n_edge = 0; + for (i = 0; i < graph->n_edge; ++i) + n_edge += graph->edge[i].map->n; + + total = 3 + n_edge; for (i = 0; i < graph->n; ++i) { struct isl_sched_node *node = &graph->node[graph->sorted[i]]; node->start = total; total += 1 + 2 * (node->nparam + node->nvar); } - if (count_constraints(graph, &n_eq, &n_ineq, 1) < 0) + if (count_all_constraints(graph, &n_eq, &n_ineq, 1) < 0) return -1; - dim = isl_dim_set_alloc(ctx, 0, total); + dim = isl_space_set_alloc(ctx, 0, total); isl_basic_set_free(graph->lp); n_eq += 3; - n_ineq += graph->n_edge; - graph->lp = isl_basic_set_alloc_dim(dim, 0, n_eq, n_ineq); + n_ineq += n_edge; + graph->lp = isl_basic_set_alloc_space(dim, 0, n_eq, n_ineq); graph->lp = isl_basic_set_set_rational(graph->lp); 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][0], -graph->n_edge); + isl_int_set_si(graph->lp->eq[k][0], -n_edge); isl_int_set_si(graph->lp->eq[k][1], 1); - for (i = 0; i < graph->n_edge; ++i) + for (i = 0; i < n_edge; ++i) isl_int_set_si(graph->lp->eq[k][4 + i], 1); k = isl_basic_set_alloc_equality(graph->lp); @@ -2051,7 +2192,7 @@ static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph) isl_int_set_si(graph->lp->eq[k][pos + j], 1); } - for (i = 0; i < graph->n_edge; ++i) { + for (i = 0; i < n_edge; ++i) { k = isl_basic_set_alloc_inequality(graph->lp); if (k < 0) return -1; @@ -2066,14 +2207,79 @@ static int setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph) 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. */ static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *graph) { + int i; + int n_edge; isl_vec *sol; isl_basic_set *lp; + n_edge = 0; + for (i = 0; i < graph->n_edge; ++i) + n_edge += graph->edge[i].map->n; + if (setup_carry_lp(ctx, graph) < 0) return -1; @@ -2088,13 +2294,16 @@ static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *graph) "error in schedule construction", return -1); } - if (isl_int_cmp_si(sol->el[1], graph->n_edge) >= 0) { + if (isl_int_cmp_si(sol->el[1], n_edge) >= 0) { isl_vec_free(sol); isl_die(ctx, isl_error_unknown, "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); @@ -2114,9 +2323,16 @@ static int carry_dependences(isl_ctx *ctx, struct isl_sched_graph *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); @@ -2124,27 +2340,34 @@ static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *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) @@ -2154,6 +2377,10 @@ static int compute_schedule_wcc(isl_ctx *ctx, struct isl_sched_graph *graph) /* 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) @@ -2167,6 +2394,8 @@ static int compute_component_schedule(isl_ctx *ctx, 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) @@ -2197,7 +2426,7 @@ static int compute_component_schedule(isl_ctx *ctx, /* Compute a schedule for the given dependence graph. * We first check if the graph is connected (through validity dependences) - * and if so compute a schedule for each component separately. + * and, if not, compute a schedule for each component separately. */ static int compute_schedule(isl_ctx *ctx, struct isl_sched_graph *graph) { @@ -2220,16 +2449,16 @@ __isl_give isl_schedule *isl_union_set_compute_schedule( __isl_take isl_union_map *proximity) { isl_ctx *ctx = isl_union_set_get_ctx(domain); - isl_dim *dim; + isl_space *dim; struct isl_sched_graph graph = { 0 }; isl_schedule *sched; domain = isl_union_set_align_params(domain, - isl_union_map_get_dim(validity)); + isl_union_map_get_space(validity)); domain = isl_union_set_align_params(domain, - isl_union_map_get_dim(proximity)); - dim = isl_union_set_get_dim(domain); - validity = isl_union_map_align_params(validity, isl_dim_copy(dim)); + isl_union_map_get_space(proximity)); + dim = isl_union_set_get_space(domain); + validity = isl_union_map_align_params(validity, isl_space_copy(dim)); proximity = isl_union_map_align_params(proximity, dim); if (!domain) @@ -2259,7 +2488,7 @@ __isl_give isl_schedule *isl_union_set_compute_schedule( goto error; empty: - sched = extract_schedule(&graph, isl_union_set_get_dim(domain)); + sched = extract_schedule(&graph, isl_union_set_get_space(domain)); graph_free(ctx, &graph); isl_union_set_free(domain); @@ -2280,15 +2509,27 @@ void *isl_schedule_free(__isl_take isl_schedule *sched) 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_space_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_space_get_ctx(schedule->dim) : NULL; +} + __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched) { int i; @@ -2297,7 +2538,7 @@ __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched) if (!sched) return NULL; - umap = isl_union_map_empty(isl_dim_copy(sched->dim)); + umap = isl_union_map_empty(isl_space_copy(sched->dim)); for (i = 0; i < sched->n; ++i) umap = isl_union_map_add_map(umap, isl_map_copy(sched->node[i].sched)); @@ -2305,43 +2546,280 @@ __isl_give isl_union_map *isl_schedule_get_map(__isl_keep isl_schedule *sched) 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) { - return sched ? sched->n_band : 0; + 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_space_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 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 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. */ -__isl_give isl_union_map *isl_schedule_get_band(__isl_keep isl_schedule *sched, - unsigned band) +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) { - int i; - isl_union_map *umap; + int i, j; + isl_ctx *ctx = isl_schedule_get_ctx(schedule); + int *active; + int n_band; + isl_band_list *list; - if (!sched) + 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; - umap = isl_union_map_empty(isl_dim_copy(sched->dim)); - for (i = 0; i < sched->n; ++i) { - int start, end; - isl_map *map; + 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); - if (band >= sched->node[i].n_band) + 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); + } - 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 list; +} - 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); +/* Construct a band forest representation of the schedule and + * return the list of roots. + */ +static __isl_give isl_band_list *construct_forest( + __isl_keep isl_schedule *schedule) +{ + int i; + isl_ctx *ctx = isl_schedule_get_ctx(schedule); + isl_band_list *forest; + int *active; + + active = isl_alloc_array(ctx, int, schedule->n); + if (!active) + return NULL; + + for (i = 0; i < schedule->n; ++i) + active[i] = 1; + + forest = construct_band_list(schedule, NULL, 0, active, schedule->n); - umap = isl_union_map_add_map(umap, map); + free(active); + + return forest; +} + +/* 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; + + 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); }