#include "timing.h"
pbqp_edge **edge_bucket;
+pbqp_edge **rm_bucket;
pbqp_node **node_buckets[4];
pbqp_node **reduced_bucket = NULL;
static int buckets_filled = 0;
edge_bucket_insert(&edge_bucket, edge);
}
+static void insert_into_rm_bucket(pbqp_edge *edge)
+{
+ if (edge_bucket_contains(rm_bucket, edge)) {
+ /* Edge is already inserted. */
+ return;
+ }
+
+ edge_bucket_insert(&rm_bucket, edge);
+}
+
static void init_buckets(void)
{
int i;
edge_bucket_init(&edge_bucket);
+ edge_bucket_init(&rm_bucket);
node_bucket_init(&reduced_bucket);
for (i = 0; i < 4; ++i) {
}
edge_bucket_free(&edge_bucket);
+ edge_bucket_free(&rm_bucket);
node_bucket_free(&reduced_bucket);
buckets_filled = 0;
node_len = pbqp->num_nodes;
#if KAPS_TIMING
- ir_timer_t *t_fill_buckets = ir_timer_register("be_pbqp_fill_buckets", "PBQP Fill Nodes into buckets");
- ir_timer_reset_and_start(t_fill_buckets);
+ ir_timer_t *t_fill_buckets = ir_timer_new();
+ ir_timer_start(t_fill_buckets);
#endif
for (node_index = 0; node_index < node_len; ++node_index) {
#if KAPS_TIMING
ir_timer_stop(t_fill_buckets);
- printf("%-20s: %8.3lf msec\n", ir_timer_get_description(t_fill_buckets), (double)ir_timer_elapsed_usec(t_fill_buckets) / 1000.0);
+ printf("PBQP Fill Nodes into buckets: %8.3lf msec\n", (double)ir_timer_elapsed_usec(t_fill_buckets) / 1000.0);
#endif
}
mat = edge->costs;
assert(mat);
- mapping = NEW_ARR_F(unsigned, src_len);
+ mapping = NEW_ARR_F(unsigned, tgt_len);
/* Check that each column has at most one zero entry. */
for (tgt_index = 0; tgt_index < tgt_len; ++tgt_index) {
unsigned onlyOneZero = 0;
+
if (tgt_vec->entries[tgt_index].data == INF_COSTS)
continue;
/* Reconnect the source's edges with the target node. */
for (edge_index = 0; edge_index < edge_len; ++edge_index) {
pbqp_edge *old_edge = src_node->edges[edge_index];
+ pbqp_edge *new_edge;
pbqp_matrix *old_matrix;
pbqp_matrix *new_matrix;
pbqp_node *other_node;
+ vector *other_vec;
unsigned other_len;
unsigned other_index;
unsigned tgt_index;
assert(old_matrix);
if (old_edge->tgt == src_node) {
- other_node = edge->src;
+ other_node = old_edge->src;
other_len = old_matrix->rows;
}
else {
- other_node = edge->tgt;
+ other_node = old_edge->tgt;
other_len = old_matrix->cols;
}
assert(other_node);
+ other_vec = other_node->costs;
new_matrix = pbqp_matrix_alloc(pbqp, tgt_len, other_len);
/* Source node selects the column of the old_matrix. */
if (old_edge->tgt == src_node) {
for (tgt_index = 0; tgt_index < tgt_len; ++tgt_index) {
- unsigned old_index = mapping[tgt_index];
+ unsigned src_index = mapping[tgt_index];
+
+ if (tgt_vec->entries[tgt_index].data == INF_COSTS)
+ continue;
+
for (other_index = 0; other_index < other_len; ++other_index) {
- new_matrix->entries[tgt_index*other_len+other_index] = old_matrix->entries[other_index*src_len+old_index];
+ if (other_vec->entries[other_index].data == INF_COSTS)
+ continue;
+
+ new_matrix->entries[tgt_index*other_len+other_index] = old_matrix->entries[other_index*src_len+src_index];
}
}
}
/* Source node selects the row of the old_matrix. */
else {
for (tgt_index = 0; tgt_index < tgt_len; ++tgt_index) {
- unsigned old_index = mapping[tgt_index];
+ unsigned src_index = mapping[tgt_index];
+
+ if (tgt_vec->entries[tgt_index].data == INF_COSTS)
+ continue;
+
for (other_index = 0; other_index < other_len; ++other_index) {
- new_matrix->entries[tgt_index*other_len+other_index] = old_matrix->entries[old_index*other_len+other_index];
+ if (other_vec->entries[other_index].data == INF_COSTS)
+ continue;
+
+ new_matrix->entries[tgt_index*other_len+other_index] = old_matrix->entries[src_index*other_len+other_index];
}
}
}
add_edge_costs(pbqp, tgt_node->index, other_node->index, new_matrix);
- disconnect_edge(src_node, old_edge);
- disconnect_edge(other_node, old_edge);
+ delete_edge(old_edge);
+ reorder_node(src_node);
+ reorder_node(other_node);
+
+ new_edge = get_edge(pbqp, tgt_node->index, other_node->index);
+ insert_into_rm_bucket(new_edge);
}
/* Reduce the remaining source node via RI. */
apply_RI(pbqp);
}
-
-void reorder_node(pbqp_node *node)
-{
- unsigned degree = pbqp_node_get_degree(node);
- /* Assume node lost one incident edge. */
- unsigned old_degree = degree + 1;
-
- if (!buckets_filled) return;
-
- /* Same bucket as before */
- if (degree > 2) return;
-
- if (!node_bucket_contains(node_buckets[old_degree], node)) {
- /* Old arity is new arity, so we have nothing to do. */
- assert(node_bucket_contains(node_buckets[degree], node));
- return;
- }
-
- /* Delete node from old bucket... */
- node_bucket_remove(&node_buckets[old_degree], node);
-
- /* ..and add to new one. */
- node_bucket_insert(&node_buckets[degree], node);
-}
-
-#if KAPS_STATISTIC
-void check_melting_possibility(pbqp *pbqp, pbqp_edge *edge)
+/**
+ * Tries to apply RM for the target node of the given edge.
+ *
+ * Checks whether the target node of edge can be merged into the source node of
+ * edge, and performs the merge, if possible.
+ */
+static void merge_target_into_source(pbqp *pbqp, pbqp_edge *edge)
{
pbqp_matrix *mat;
pbqp_node *src_node;
pbqp_node *tgt_node;
vector *src_vec;
vector *tgt_vec;
- int src_len;
- int tgt_len;
- int src_index;
- int tgt_index;
+ unsigned *mapping;
+ unsigned src_len;
+ unsigned tgt_len;
+ unsigned src_index;
+ unsigned tgt_index;
+ unsigned edge_index;
+ unsigned edge_len;
assert(pbqp);
assert(edge);
src_len = src_vec->len;
tgt_len = tgt_vec->len;
- assert(src_len > 0);
- assert(tgt_len > 0);
+
+ /* Matrizes are normalized. */
+ assert(src_len > 1);
+ assert(tgt_len > 1);
mat = edge->costs;
assert(mat);
- if (src_len == 1 && tgt_len == 1) {
- //panic("Something is wrong");
- }
+ mapping = NEW_ARR_F(unsigned, src_len);
- int allRowsOk = 1;
+ /* Check that each row has at most one zero entry. */
for (src_index = 0; src_index < src_len; ++src_index) {
- int onlyOneZero = 0;
- if (src_vec->entries[src_index].data == INF_COSTS) {
+ unsigned onlyOneZero = 0;
+
+ if (src_vec->entries[src_index].data == INF_COSTS)
continue;
- }
+
for (tgt_index = 0; tgt_index < tgt_len; ++tgt_index) {
- if (tgt_vec->entries[tgt_index].data == INF_COSTS) {
+ if (tgt_vec->entries[tgt_index].data == INF_COSTS)
continue;
- }
- if (mat->entries[src_index * tgt_len + tgt_index] == 0) {
- if (onlyOneZero) {
- onlyOneZero = 0;
- break;
- } else {
- onlyOneZero = 1;
- continue;
- }
- }
- if (mat->entries[src_index * tgt_len + tgt_index] == INF_COSTS) {
+
+ if (mat->entries[src_index * tgt_len + tgt_index] == INF_COSTS)
continue;
+
+ /* Matrix entry is finite. */
+ if (onlyOneZero) {
+ DEL_ARR_F(mapping);
+ return;
}
- onlyOneZero = 0;
- break;
+
+ onlyOneZero = 1;
+ mapping[src_index] = tgt_index;
}
- allRowsOk &= onlyOneZero;
}
- int allColsOk = 1;
- for (tgt_index = 0; tgt_index < tgt_len; ++tgt_index) {
- int onlyOneZero = 0;
- if (tgt_vec->entries[tgt_index].data == INF_COSTS) {
+ /* We know that we can merge the target node into the source node. */
+ edge_len = pbqp_node_get_degree(tgt_node);
+
+#if KAPS_STATISTIC
+ pbqp->num_rm++;
+#endif
+
+ /* Reconnect the target's edges with the source node. */
+ for (edge_index = 0; edge_index < edge_len; ++edge_index) {
+ pbqp_edge *old_edge = tgt_node->edges[edge_index];
+ pbqp_edge *new_edge;
+ pbqp_matrix *old_matrix;
+ pbqp_matrix *new_matrix;
+ pbqp_node *other_node;
+ vector *other_vec;
+ unsigned other_len;
+ unsigned other_index;
+ unsigned src_index;
+
+ assert(old_edge);
+
+ if (old_edge == edge)
continue;
+
+ old_matrix = old_edge->costs;
+ assert(old_matrix);
+
+ if (old_edge->tgt == tgt_node) {
+ other_node = old_edge->src;
+ other_len = old_matrix->rows;
}
- for (src_index = 0; src_index < src_len; ++src_index) {
- if (src_vec->entries[src_index].data == INF_COSTS) {
- continue;
- }
- if (mat->entries[src_index * tgt_len + tgt_index] == 0) {
- if (onlyOneZero) {
- onlyOneZero = 0;
- break;
- } else {
- onlyOneZero = 1;
+ else {
+ other_node = old_edge->tgt;
+ other_len = old_matrix->cols;
+ }
+ assert(other_node);
+ other_vec = other_node->costs;
+
+ new_matrix = pbqp_matrix_alloc(pbqp, src_len, other_len);
+
+ /* Target node selects the column of the old_matrix. */
+ if (old_edge->tgt == tgt_node) {
+ for (src_index = 0; src_index < src_len; ++src_index) {
+ unsigned tgt_index = mapping[src_index];
+
+ if (src_vec->entries[src_index].data == INF_COSTS)
continue;
+
+ for (other_index = 0; other_index < other_len; ++other_index) {
+ if (other_vec->entries[other_index].data == INF_COSTS)
+ continue;
+
+ new_matrix->entries[src_index*other_len+other_index] = old_matrix->entries[other_index*tgt_len+tgt_index];
}
}
- if (mat->entries[src_index * tgt_len + tgt_index] == INF_COSTS) {
- continue;
+ }
+ /* Source node selects the row of the old_matrix. */
+ else {
+ for (src_index = 0; src_index < src_len; ++src_index) {
+ unsigned tgt_index = mapping[src_index];
+
+ if (src_vec->entries[src_index].data == INF_COSTS)
+ continue;
+
+ for (other_index = 0; other_index < other_len; ++other_index) {
+ if (other_vec->entries[other_index].data == INF_COSTS)
+ continue;
+
+ new_matrix->entries[src_index*other_len+other_index] = old_matrix->entries[tgt_index*other_len+other_index];
+ }
}
- onlyOneZero = 0;
- break;
}
- allColsOk &= onlyOneZero;
+
+ add_edge_costs(pbqp, src_node->index, other_node->index, new_matrix);
+
+ delete_edge(old_edge);
+ reorder_node(tgt_node);
+ reorder_node(other_node);
+
+ new_edge = get_edge(pbqp, src_node->index, other_node->index);
+ insert_into_rm_bucket(new_edge);
+ }
+
+ /* Reduce the remaining source node via RI. */
+ apply_RI(pbqp);
+}
+
+/**
+ * Merge neighbors into the given node.
+ */
+void apply_RM(pbqp *pbqp, pbqp_node *node)
+{
+ pbqp_edge **edges;
+ unsigned edge_index;
+ unsigned edge_len;
+
+ assert(node);
+ assert(pbqp);
+
+ edges = node->edges;
+ edge_len = pbqp_node_get_degree(node);
+
+ /* Check all incident edges. */
+ for (edge_index = 0; edge_index < edge_len; ++edge_index) {
+ pbqp_edge *edge = edges[edge_index];
+
+ insert_into_rm_bucket(edge);
}
- if (allRowsOk || allColsOk) {
- pbqp->num_rm++;
+ /* ALAP: Merge neighbors into given node. */
+ while(edge_bucket_get_length(rm_bucket) > 0) {
+ pbqp_edge *edge = edge_bucket_pop(&rm_bucket);
+ assert(edge);
+
+ if (edge->src == node)
+ merge_target_into_source(pbqp, edge);
+ else
+ merge_source_into_target(pbqp, edge);
}
}
-#endif
+
+void reorder_node(pbqp_node *node)
+{
+ unsigned degree = pbqp_node_get_degree(node);
+ /* Assume node lost one incident edge. */
+ unsigned old_degree = degree + 1;
+
+ if (!buckets_filled) return;
+
+ /* Same bucket as before */
+ if (degree > 2) return;
+
+ if (!node_bucket_contains(node_buckets[old_degree], node)) {
+ /* Old arity is new arity, so we have nothing to do. */
+ assert(node_bucket_contains(node_buckets[degree], node));
+ return;
+ }
+
+ /* Delete node from old bucket... */
+ node_bucket_remove(&node_buckets[old_degree], node);
+
+ /* ..and add to new one. */
+ node_bucket_insert(&node_buckets[degree], node);
+}
void simplify_edge(pbqp *pbqp, pbqp_edge *edge)
{
assert(pbqp);
#if KAPS_TIMING
- ir_timer_t *t_int_simpl = ir_timer_register("be_pbqp_init_simp", "PBQP Initial simplify edges");
- ir_timer_reset_and_start(t_int_simpl);
+ ir_timer_t *t_int_simpl = ir_timer_new();
+ ir_timer_start(t_int_simpl);
#endif
#if KAPS_DUMP
#if KAPS_TIMING
ir_timer_stop(t_int_simpl);
- printf("%-20s: %8.3lf msec\n", ir_timer_get_description(t_int_simpl), (double)ir_timer_elapsed_usec(t_int_simpl) / 1000.0);
+ printf("PBQP Initial simplify edges: %8.3lf msec\n", (double)ir_timer_elapsed_usec(t_int_simpl) / 1000.0);
#endif
}
num solution = 0;
#if KAPS_TIMING
- ir_timer_t *t_det_solution = ir_timer_register("be_det_solution", "PBQP Determine Solution");
+ ir_timer_t *t_det_solution = ir_timer_new();
ir_timer_reset_and_start(t_det_solution);
#endif
#if KAPS_TIMING
ir_timer_stop(t_det_solution);
- printf("%-20s: %8.3lf msec\n", ir_timer_get_description(t_det_solution), (double)ir_timer_elapsed_usec(t_det_solution) / 1000.0);
+ printf("PBQP Determine Solution: %8.3lf msec\n", (double)ir_timer_elapsed_usec(t_det_solution) / 1000.0);
#endif
return solution;
projects / libfirm / blobdiff