#37 Made get_mappings sequential, as it's faster that way.

src/lib/cpp/cpu/connected_components.cc

@@ -739,42 +739,14 @@ void filter_largest(const std::string &base_path, bool *__restrict__ mask, const
 }
 
 std::tuple<mapping_t, mapping_t> get_mappings(const std::vector<int64_t> &a, const int64_t n_labels_a, const std::vector<int64_t> &b, const int64_t n_labels_b, const idx3d &global_shape) {
-    std::vector<mapping_t> mappings_a;
-    std::vector<mapping_t> mappings_b;
-
     mapping_t mapping_a(n_labels_a+1);
     mapping_t mapping_b(n_labels_b+1);
 
-    #pragma omp parallel num_threads(8)
-    {
-        int64_t n_threads = omp_get_num_threads();
-
-        #pragma omp single
-        {
-            mappings_a.resize(n_threads, mapping_t(n_labels_a+1));
-            mappings_b.resize(n_threads, mapping_t(n_labels_b+1));
-        }
-
-        #pragma omp for schedule(static) collapse(2)
-        for (int64_t y = 0; y < global_shape.y; y++) {
-            for (int64_t x = 0; x < global_shape.x; x++) {
-                int64_t i = (y * global_shape.x) + x;
+    const int64_t plane_size = global_shape.y * global_shape.x;
+    for (int64_t i = 0; i < plane_size; i++) {
                 if (a[i] != 0 && b[i] != 0) {
-                    mappings_a[omp_get_thread_num()][a[i]].insert(b[i]);
-                    mappings_b[omp_get_thread_num()][b[i]].insert(a[i]);
-                }
-            }
-        }
-
-        for (int64_t i = 0; i < n_threads; i++) {
-            #pragma omp for schedule(static)
-            for (int64_t j = 1; j < n_labels_a+1; j++) {
-                mapping_a[j].insert(mappings_a[i][j].begin(), mappings_a[i][j].end());
-            }
-            #pragma omp for schedule(static)
-            for (int64_t j = 1; j < n_labels_b+1; j++) {
-                mapping_b[j].insert(mappings_b[i][j].begin(), mappings_b[i][j].end());
-            }
+            mapping_a[a[i]].insert(b[i]);
+            mapping_b[b[i]].insert(a[i]);
         }
     }