temp2

.github/workflows/linux.yml

@@ -8,12 +8,12 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        build-type: [ Release, RelWithDebInfo ]
+        build-type: [ Release, RelWithDebInfo, Debug ]
         compiler: [
-            { cxx: g++-11, pkgs: g++-11 libtbb-dev gdb },
-            { cxx: g++-12, pkgs: g++-12 libtbb-dev gdb },
-            { cxx: clang++-14, pkgs: clang-14 libtbb-dev gdb },
-            { cxx: clang++-15, pkgs: clang-15 libtbb-dev gdb },
+            { cxx: g++-11, pkgs: g++-11 libtbb-dev },
+            { cxx: g++-12, pkgs: g++-12 libtbb-dev },
+            { cxx: clang++-14, pkgs: clang-14 libtbb-dev },
+            { cxx: clang++-15, pkgs: clang-15 libtbb-dev },
         ]
 
     defaults:
@@ -43,7 +43,7 @@ jobs:
       run: cmake --build . --parallel
 
     - name: run-tests
-      run: gdb -ex run ./unit_tests --args schedule-random
+      run: ctest --output-on-failure --schedule-random
 
     - name: install
       run: sudo cmake --install .

.tsan-supressions

@@ -8,4 +8,7 @@ race:^tbb::detail::d1::auto_partition_type::is_divisible
 race:tbb::detail::d1::small_object_allocator::new_object
 race:tbb::detail::d1::small_object_allocator::delete_object
 
-race:tbb::detail::d1::dynamic_grainsize_mode<*>::check_being_stolen
+race:tbb::detail::d1::dynamic_grainsize_mode<*>::check_being_stolen
+
+race:tbb::detail::d1::start_for<*>::finalize
+race:tbb::detail::d1::start_for<*>::offer_work_impl

CMakeLists.txt

@@ -185,6 +185,7 @@ if(GAPP_BUILD_TESTS AND BUILD_TESTING AND PROJECT_IS_TOP_LEVEL)
     add_subdirectory("${CMAKE_CURRENT_SOURCE_DIR}/test/unit")
     add_subdirectory("${CMAKE_CURRENT_SOURCE_DIR}/test/integration")
     add_subdirectory("${CMAKE_CURRENT_SOURCE_DIR}/test/misc")
+    add_subdirectory("${CMAKE_CURRENT_SOURCE_DIR}/test/tsan")
 endif()
 
 if(GAPP_BUILD_BENCHMARKS AND BUILD_TESTING AND PROJECT_IS_TOP_LEVEL)

src/algorithm/nd_sort.cpp

@@ -188,8 +188,10 @@ namespace gapp::algorithm::dtl
         const size_t popsize = std::distance(first, last);
         DominanceMatrix dmat(popsize, popsize /*, MAXIMAL */);
 
+        GAPP_BARRIER();
         std::for_each(GAPP_EXEC, detail::iota_iterator(0_sz), detail::iota_iterator(first->size()), [&](size_t obj)
         {
+            GAPP_BARRIER();
             FitnessVector fvec(popsize);
             std::transform(first, last, fvec.begin(), detail::element_at(obj));
 
@@ -210,10 +212,13 @@ namespace gapp::algorithm::dtl
                     }
                 });
             });
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
 
         std::for_each(GAPP_EXEC, detail::iota_iterator(0_sz), detail::iota_iterator(popsize), [&](size_t row) noexcept
         {
+            GAPP_BARRIER();
             dmat(row, row).store(NONMAXIMAL, std::memory_order_relaxed); // diagonal is all nonmax
 
             for (size_t col = row + 1; col < popsize; col++)
@@ -224,7 +229,9 @@ namespace gapp::algorithm::dtl
                     dmat(col, row).store(NONMAXIMAL, std::memory_order_relaxed);
                 }
             }
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
 
         return dmat;
     }

src/algorithm/nsga3.cpp

@@ -239,8 +239,10 @@ namespace gapp::algorithm
 
         sol_info_.resize(last - first);
 
+        GAPP_BARRIER();
         std::for_each(GAPP_EXEC, pfirst, plast, [&](const FrontInfo& sol)
         {
+            GAPP_BARRIER();
             const FitnessVector fnorm = normalizeFitnessVec(first[sol.idx], ideal_point_, nadir_point_);
 
             auto idistance = [&](const auto& line) { return std::inner_product(fnorm.begin(), fnorm.end(), line.begin(), 0.0); };
@@ -249,7 +251,9 @@ namespace gapp::algorithm
 
             sol_info_[sol.idx].ref_idx  = std::distance(ref_lines_.begin(), closest);
             sol_info_[sol.idx].ref_dist = math::perpendicularDistanceSq(*closest, fnorm);
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
     }
 
     inline bool NSGA3::Impl::nichedCompare(size_t lhs, size_t rhs) const noexcept

src/algorithm/reference_lines.cpp

@@ -183,12 +183,11 @@ namespace gapp::algorithm::reflines
             min_distances.pop_back();
 
             /* Calc the distance of each candidate to the closest ref point. */
-            std::transform(GAPP_EXEC, candidate_points.begin(), candidate_points.end(), min_distances.begin(), min_distances.begin(),
-            [&](const Point& candidate, double current_min) noexcept
+            for (size_t i = 0; i < candidate_points.size(); i++)
             {
-                const double dist = math::euclideanDistanceSq(candidate, points.back());
-                return std::min(current_min, dist);
-            });
+                double dist = math::euclideanDistanceSq(candidate_points[i], points.back());
+                min_distances[i] = std::min(min_distances[i], dist);
+            }
         }
 
         return points;

src/core/ga_base.impl.hpp

@@ -327,7 +327,7 @@ namespace gapp
 
         /* Reset state in case solve() has already been called before. */
         generation_cntr_ = 0;
-        num_fitness_evals_ = 0;
+        num_fitness_evals_->store(0, std::memory_order_release);
         solutions_.clear();
         population_.clear();
 
@@ -339,7 +339,9 @@ namespace gapp
         /* Create and evaluate the initial population of the algorithm. */
         num_objectives_ = findNumberOfObjectives();
         population_ = generatePopulation(population_size_, std::move(initial_population));
-        std::for_each(GAPP_EXEC, population_.begin(), population_.end(), [this](Candidate<T>& sol) { evaluate(sol); });
+        GAPP_BARRIER();
+        std::for_each(GAPP_EXEC, population_.begin(), population_.end(), [this](Candidate<T>& sol) { GAPP_BARRIER(); evaluate(sol); GAPP_BARRIER(); });
+        GAPP_BARRIER();
         fitness_matrix_ = detail::toFitnessMatrix(population_);
         if (keep_all_optimal_sols_) solutions_ = detail::findParetoFront(population_);
 
@@ -465,11 +467,10 @@ namespace gapp
          * is no point doing it again. */
         if (!sol.is_evaluated || fitness_function_->dynamic())
         {
+            num_fitness_evals_->fetch_add(1, std::memory_order_acq_rel);
+
             sol.fitness = (*fitness_function_)(sol.chromosome);
             sol.is_evaluated = true;
-
-            std::atomic_ref num_evals{ num_fitness_evals_ };
-            num_evals.fetch_add(1_sz, std::memory_order_acq_rel);
         }
 
         GAPP_ASSERT(hasValidFitness(sol));
@@ -504,27 +505,34 @@ namespace gapp
         std::vector<CandidatePair<T>> child_pairs(num_children / 2);
 
         prepareSelections();
-        std::generate(GAPP_EXEC, child_pairs.begin(), child_pairs.end(),
-        [this]
+        GAPP_BARRIER();
+        std::for_each(GAPP_EXEC, child_pairs.begin(), child_pairs.end(),
+        [this](CandidatePair<T>& children)
         {
+            GAPP_BARRIER();
             const auto& parent1 = select();
             const auto& parent2 = select();
-
-            return crossover(parent1, parent2);
+            children = crossover(parent1, parent2);
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
 
         auto children = detail::flatten(std::move(child_pairs));
 
         /* If the population size is odd, one too many child candidates were generated by the crossovers. */
         if (children.size() > population_size_) children.pop_back();
 
+        GAPP_BARRIER();
         std::for_each(GAPP_EXEC, children.begin(), children.end(),
         [this](Candidate<T>& child)
         {
+            GAPP_BARRIER();
             mutate(child);
             repair(child);
             evaluate(child);
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
 
         updatePopulation(std::move(children));
         if (keep_all_optimal_sols_) updateOptimalSolutions(solutions_, population_);

src/core/ga_info.cpp

@@ -4,7 +4,6 @@
 #include "../algorithm/single_objective.hpp"
 #include "../stop_condition/stop_condition.hpp"
 #include "../utility/utility.hpp"
-#include <atomic>
 #include <memory>
 #include <utility>
 
@@ -13,7 +12,7 @@ namespace gapp
     GaInfo::GaInfo(GaInfo&&) noexcept            = default;
     GaInfo& GaInfo::operator=(GaInfo&&) noexcept = default;
 
-    GaInfo::~GaInfo()                            = default;
+    GaInfo::~GaInfo() noexcept                   = default;
 
 
     GaInfo::GaInfo(Positive<size_t> population_size, std::unique_ptr<algorithm::Algorithm> algorithm, std::unique_ptr<stopping::StopCondition> stop_condition) noexcept :
@@ -26,8 +25,7 @@ namespace gapp
 
     size_t GaInfo::num_fitness_evals() const noexcept
     {
-        std::atomic_ref num_fitness_evals{ num_fitness_evals_ };
-        return num_fitness_evals.load(std::memory_order_acquire);
+        return num_fitness_evals_->load(std::memory_order_acquire);
     }
 
     void GaInfo::algorithm(std::unique_ptr<algorithm::Algorithm> f)

src/core/ga_info.hpp

@@ -5,6 +5,7 @@
 
 #include "../population/population.hpp"
 #include "../utility/bounded_value.hpp"
+#include "../utility/atomic.hpp"
 #include "../utility/utility.hpp"
 #include "../metrics/metric_set.hpp"
 #include <functional>
@@ -341,7 +342,7 @@ namespace gapp
         GaInfo& operator=(const GaInfo&) = delete;
 
         /** Destructor. */
-        virtual ~GaInfo();
+        virtual ~GaInfo() noexcept;
 
     protected:
 
@@ -359,7 +360,7 @@ namespace gapp
         Positive<size_t> max_gen_ = 500;
         size_t num_objectives_ = 0;
         size_t generation_cntr_ = 0;
-        size_t num_fitness_evals_ = 0;
+        detail::atomic<size_t> num_fitness_evals_ = 0;
 
         bool keep_all_optimal_sols_ = false;
         bool use_default_algorithm_ = false;

src/metrics/pop_stats.cpp

@@ -154,15 +154,19 @@ namespace gapp::detail
         const FitnessMatrix front = uniqueSortedParetoFront(fmat);
 
         std::atomic<double> hypervolume = 0.0;
+        GAPP_BARRIER();
         std::for_each(GAPP_EXEC, detail::iota_iterator(0_sz), detail::iota_iterator(front.size()), [&](size_t idx)
         {
+            GAPP_BARRIER();
             const auto point = front[idx];
             const FitnessMatrix rest = { front.begin() + idx + 1, front.end() };
 
             const double exclusive_hypervolume = exclusiveHypervolume(point, rest, ref_point);
 
             hypervolume.fetch_add(exclusive_hypervolume, std::memory_order_acq_rel);
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
 
         return hypervolume.load(std::memory_order_acquire);
     }

src/population/population.hpp

@@ -106,8 +106,10 @@ namespace gapp::detail
         std::vector<Dominance> lhs_state(lhs.size());
         std::vector<std::atomic<Dominance>> rhs_state(rhs.size());
 
+        GAPP_BARRIER();
         std::for_each(GAPP_EXEC, iota_iterator(0_sz), iota_iterator(lhs.size()), [&](size_t i) noexcept
         {
+            GAPP_BARRIER();
             for (size_t j = 0; j < rhs.size(); j++)
             {
                 const Dominance rhs_state_j = rhs_state[j].load(std::memory_order_acquire);
@@ -145,7 +147,9 @@ namespace gapp::detail
                 }
                 // comp == 0 --> both are OPTIMAL or DOMINATED, can't know
             }
+            GAPP_BARRIER();
         });
+        GAPP_BARRIER();
 
         Candidates<T> optimal_solutions;
         optimal_solutions.reserve(lhs.size() + rhs.size());

src/utility/atomic.hpp

@@ -0,0 +1,64 @@
+/* Copyright (c) 2023 Krisztián Rugási. Subject to the MIT License. */
+
+#ifndef GA_UTILITY_ATOMIC_HPP
+#define GA_UTILITY_ATOMIC_HPP
+
+#include <atomic>
+#include <memory>
+#include <utility>
+
+namespace gapp::detail
+{
+    /*
+    * This is a simple wrapper class around std::atomic with atomic
+    * initialization, in order to prevent data races between the
+    * initialization of the variable and later accesses to it.
+    * 
+    * The wrapper also adds move operators for convenience.
+    */
+    template<typename T>
+    class atomic
+    {
+    public:
+        atomic(std::memory_order order = std::memory_order_seq_cst) noexcept
+        {
+            data_.store(T{}, order);
+        }
+
+        atomic(T value, std::memory_order order = std::memory_order_seq_cst) noexcept
+        {
+            data_.store(std::move(value), order);
+        }
+
+        atomic(atomic<T>&& other) noexcept
+        {
+            data_.store(other->load());
+        }
+
+        atomic<T>& operator=(T value) noexcept
+        {
+            data_.store(std::move(value));
+            return *this;
+        }
+
+        atomic<T>& operator=(atomic<T>&& other) noexcept
+        {
+            data_.store(other->load());
+            return *this;
+        }
+
+        ~atomic() noexcept = default; // maybe needs release?
+
+        std::atomic<T>& operator*() noexcept  { return data_; }
+        std::atomic<T>* operator->() noexcept { return std::addressof(data_); }
+        const std::atomic<T>& operator*() const noexcept  { return data_; }
+        const std::atomic<T>* operator->() const noexcept { return std::addressof(data_); }
+
+    private:
+        std::atomic<T> data_;
+    };
+
+
+} // namespace gapp::detail
+
+#endif // !GA_UTILITY_ATOMIC_HPP