Fixed infinite loop bug in Subgraph isomorphism (#1672)

* Experiment to address infinite loop * Experiment #2 to address infinite loop * Experiment #3 to address infinite loop * Temporary debug output added * External limit for number of threads was removed * Experiment #4 to address infinite loop * Experiment #5 to address infinite loop * Experiment #6 to address infinite loop * Experiment #7 to address infinite loop * Experiment #8 to address infinite loop * 20% is set as optimal value * Temporary code removed * Optimal array_size is caclulated based on max thread count * Optimal array_size is decreased * Optimal array_size is adjusted * Optimal array_size is adjusted * Optimal array_size is adjusted * Optimal array_size is adjusted * Optimal array_size is adjusted * Optimal array_size is adjusted * Optimal array_size is adjusted * Experiment #9 to address infinite loop * Optimal array_size is adjusted * Temporary debug output added * Compilation fix * Experiment #10 to address infinite loop * Temporary debug output added * Experiment #11 to address infinite loop * Temporary code removed * Optimal array_size is adjusted
uxlfoundation · Jun 3, 2021 · 658eb24 · 658eb24
1 parent 9b4f2b2
commit 658eb24
Showing 1 changed file with 5 additions and 2 deletions.
diff --git a/cpp/oneapi/dal/algo/subgraph_isomorphism/backend/cpu/matching.hpp b/cpp/oneapi/dal/algo/subgraph_isomorphism/backend/cpu/matching.hpp
@@ -457,15 +457,18 @@ solution<Cpu> engine_bundle<Cpu>::run() {
     std::uint64_t max_threads_count = dal::detail::threader_get_max_threads();
     std::uint64_t possible_first_states_count_per_thread = first_states_count / max_threads_count;
     if (possible_first_states_count_per_thread < 1) {
-        max_threads_count = first_states_count;
         possible_first_states_count_per_thread = 1;
     }
     else {
         possible_first_states_count_per_thread +=
             static_cast<bool>(first_states_count % max_threads_count);
     }
 
-    const std::uint64_t array_size = max_threads_count;
+    const std::uint64_t array_size = (max_threads_count >= 64)   ? max_threads_count * 2 / 10
+                                     : (max_threads_count >= 24) ? max_threads_count * 4 / 10
+                                     : (max_threads_count >= 8)  ? 4
+                                     : (max_threads_count >= 4)  ? 2
+                                                                 : 1;
     auto engine_array_ptr = allocator_.make_shared_memory<matching_engine<Cpu>>(array_size);
     matching_engine<Cpu>* engine_array = engine_array_ptr.get();