works

CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.25)
-project(blt-gp VERSION 0.1.4)
+project(blt-gp VERSION 0.1.8)
 
 include(CTest)
 

examples/symbolic_regression.cpp

@@ -117,15 +117,15 @@ int main()
     program.set_operations(builder.build());
     
     BLT_DEBUG("Generate Initial Population");
-    program.generate_population(type_system.get_type<float>().id(), fitness_function);
+    auto sel = blt::gp::select_fitness_proportionate_t{};
+    program.generate_population(type_system.get_type<float>().id(), fitness_function, sel, sel, sel);
     
     BLT_DEBUG("Begin Generation Loop");
     while (!program.should_terminate())
     {
         BLT_TRACE("------------{Begin Generation %ld}------------", program.get_current_generation());
         BLT_START_INTERVAL("Symbolic Regression", "Gen");
-        auto sel = blt::gp::select_fitness_proportionate_t{};
+        program.create_next_generation();
-        program.create_next_generation(sel, sel, sel);
         BLT_END_INTERVAL("Symbolic Regression", "Gen");
         BLT_TRACE("Move to next generation");
         BLT_START_INTERVAL("Symbolic Regression", "Fitness");

include/blt/gp/program.h

@@ -264,19 +264,12 @@ namespace blt::gp
                     system(system), seed(seed), config(config)
             { create_threads(); }
             
-            template<typename Crossover, typename Mutation, typename Reproduction, typename CreationFunc = decltype(default_next_pop_creator<Crossover, Mutation, Reproduction>)>
+            void create_next_generation()
-            void create_next_generation(Crossover&& crossover_selection, Mutation&& mutation_selection, Reproduction&& reproduction_selection,
-                                        CreationFunc& func = default_next_pop_creator<Crossover, Mutation, Reproduction>)
             {
                 // should already be empty
                 next_pop.clear();
-                crossover_selection.pre_process(*this, current_pop, current_stats);
+                thread_helper.next_gen_left.store(config.population_size, std::memory_order_release);
-                mutation_selection.pre_process(*this, current_pop, current_stats);
+                (*thread_execution_service)(0);
-                reproduction_selection.pre_process(*this, current_pop, current_stats);
-                
-                auto args = get_selector_args();
-                func(args, std::forward<Crossover>(crossover_selection), std::forward<Mutation>(mutation_selection),
-                     std::forward<Reproduction>(reproduction_selection));
             }
             
             void evaluate_fitness()
@@ -294,8 +287,10 @@ namespace blt::gp
              *
              * NOTE: 0 is considered the best, in terms of standardized fitness
              */
-            template<typename FitnessFunc>
+            template<typename FitnessFunc, typename Crossover, typename Mutation, typename Reproduction, typename CreationFunc = decltype(default_next_pop_creator<Crossover, Mutation, Reproduction>)>
-            void generate_population(type_id root_type, FitnessFunc& fitness_function, bool eval_fitness_now = true)
+            void generate_population(type_id root_type, FitnessFunc& fitness_function,
+                                     Crossover& crossover_selection, Mutation& mutation_selection, Reproduction& reproduction_selection,
+                                     CreationFunc& func = default_next_pop_creator<Crossover, Mutation, Reproduction>, bool eval_fitness_now = true)
             {
                 using LambdaReturn = typename decltype(blt::meta::lambda_helper(fitness_function))::Return;
                 current_pop = config.pop_initializer.get().generate(
@@ -303,7 +298,10 @@ namespace blt::gp
                 if (config.threads == 1)
                 {
                     BLT_INFO("Starting with single thread variant!");
-                    thread_execution_service = new std::function([this, &fitness_function](blt::size_t) {
+                    thread_execution_service = new std::function(
+                            [this, &fitness_function, &crossover_selection, &mutation_selection, &reproduction_selection, &func](blt::size_t) {
+                                if (thread_helper.evaluation_left > 0)
+                                {
                                     for (const auto& ind : blt::enumerate(current_pop.get_individuals()))
                                     {
                                         if constexpr (std::is_same_v<LambdaReturn, bool> || std::is_convertible_v<LambdaReturn, bool>)
@@ -324,12 +322,35 @@ namespace blt::gp
                                         
                                         current_stats.overall_fitness = current_stats.overall_fitness + ind.second.fitness.adjusted_fitness;
                                     }
+                                    thread_helper.evaluation_left = 0;
+                                }
+                                if (thread_helper.next_gen_left > 0)
+                                {
+                                    static thread_local std::vector<tree_t> new_children;
+                                    new_children.clear();
+                                    auto args = get_selector_args(new_children);
+                                    
+                                    crossover_selection.pre_process(*this, current_pop, current_stats);
+                                    mutation_selection.pre_process(*this, current_pop, current_stats);
+                                    reproduction_selection.pre_process(*this, current_pop, current_stats);
+                                    
+                                    perform_elitism(args);
+                                    
+                                    while (new_children.size() < config.population_size)
+                                        func(args, crossover_selection, mutation_selection, reproduction_selection);
+                                    
+                                    for (auto& i : new_children)
+                                        next_pop.get_individuals().emplace_back(std::move(i));
+                                    
+                                    thread_helper.next_gen_left = 0;
+                                }
                             });
                 } else
                 {
                     BLT_INFO("Starting thread execution service!");
                     std::scoped_lock lock(thread_helper.thread_function_control);
-                    thread_execution_service = new std::function([this, &fitness_function](blt::size_t) {
+                    thread_execution_service = new std::function(
+                            [this, &fitness_function, &crossover_selection, &mutation_selection, &reproduction_selection, &func](blt::size_t id) {
                                 thread_helper.barrier.wait();
                                 if (thread_helper.evaluation_left > 0)
                                 {
@@ -363,12 +384,14 @@ namespace blt::gp
                                             auto old_best = current_stats.best_fitness.load(std::memory_order_relaxed);
                                             while (ind.fitness.adjusted_fitness > old_best &&
                                                    !current_stats.best_fitness.compare_exchange_weak(old_best, ind.fitness.adjusted_fitness,
-                                                                                             std::memory_order_relaxed, std::memory_order_relaxed));
+                                                                                                     std::memory_order_relaxed,
+                                                                                                     std::memory_order_relaxed));
                                             
                                             auto old_worst = current_stats.worst_fitness.load(std::memory_order_relaxed);
                                             while (ind.fitness.adjusted_fitness < old_worst &&
                                                    !current_stats.worst_fitness.compare_exchange_weak(old_worst, ind.fitness.adjusted_fitness,
-                                                                                              std::memory_order_relaxed, std::memory_order_relaxed));
+                                                                                                      std::memory_order_relaxed,
+                                                                                                      std::memory_order_relaxed));
                                             
                                             auto old_overall = current_stats.overall_fitness.load(std::memory_order_relaxed);
                                             while (!current_stats.overall_fitness.compare_exchange_weak(old_overall,
@@ -380,6 +403,26 @@ namespace blt::gp
                                 }
                                 if (thread_helper.next_gen_left > 0)
                                 {
+                                    static thread_local std::vector<tree_t> new_children;
+                                    new_children.clear();
+                                    auto args = get_selector_args(new_children);
+                                    if (id == 0)
+                                    {
+                                        crossover_selection.pre_process(*this, current_pop, current_stats);
+                                        if (&crossover_selection != &mutation_selection)
+                                            mutation_selection.pre_process(*this, current_pop, current_stats);
+                                        if (&crossover_selection != &reproduction_selection)
+                                            reproduction_selection.pre_process(*this, current_pop, current_stats);
+                                        
+                                        perform_elitism(args);
+                                        
+                                        for (auto& i : new_children)
+                                            next_pop.get_individuals().emplace_back(std::move(i));
+                                        thread_helper.next_gen_left -= new_children.size();
+                                        new_children.clear();
+                                    }
+                                    thread_helper.barrier.wait();
+                                    
                                     while (thread_helper.next_gen_left > 0)
                                     {
                                         blt::size_t size = 0;
@@ -393,12 +436,16 @@ namespace blt::gp
                                                                                                     std::memory_order::memory_order_relaxed,
                                                                                                     std::memory_order::memory_order_relaxed));
                                         
-                                static thread_local std::vector<tree_t> new_children;
-                                new_children.clear();
-                                
                                         for (blt::size_t i = begin; i < end; i++)
-                                {
+                                            func(args, crossover_selection, mutation_selection, reproduction_selection);
                                         
+                                        {
+                                            std::scoped_lock lock(thread_helper.thread_generation_lock);
+                                            for (auto& i : new_children)
+                                            {
+                                                if (next_pop.get_individuals().size() < config.population_size)
+                                                    next_pop.get_individuals().emplace_back(i);
+                                            }
                                         }
                                     }
                                 }
@@ -605,6 +652,7 @@ namespace blt::gp
                 std::vector<std::unique_ptr<std::thread>> threads;
                 
                 std::mutex thread_function_control;
+                std::mutex thread_generation_lock;
                 std::condition_variable thread_function_condition{};
                 
                 std::atomic_uint64_t evaluation_left = 0;
@@ -620,9 +668,9 @@ namespace blt::gp
             // for convenience, shouldn't decrease performance too much
             std::atomic<std::function<void(blt::size_t)>*> thread_execution_service = nullptr;
             
-            inline selector_args get_selector_args()
+            inline selector_args get_selector_args(std::vector<tree_t>& next_pop_trees)
             {
-                return {*this, next_pop, current_pop, current_stats, config, get_random()};
+                return {*this, next_pop_trees, current_pop, current_stats, config, get_random()};
             }
             
             template<typename Return, blt::size_t size, typename Accessor, blt::size_t... indexes>
@@ -637,7 +685,6 @@ namespace blt::gp
             void evaluate_fitness_internal()
             {
                 current_stats.clear();
-                if (config.threads != 1)
                 thread_helper.evaluation_left.store(current_pop.get_individuals().size(), std::memory_order_release);
                 (*thread_execution_service)(0);
                 

include/blt/gp/selection.h

@@ -31,7 +31,7 @@ namespace blt::gp
     struct selector_args
     {
         gp_program& program;
-        population_t& next_pop;
+        std::vector<tree_t>& next_pop;
         population_t& current_pop;
         population_stats& current_stats;
         prog_config_t& config;
@@ -52,8 +52,6 @@ namespace blt::gp
             {
                 for (blt::size_t i = 0; i < config.elites; i++)
                 {
-//                    BLT_INFO("%lf >= %lf? // %lf (indexes: %ld %ld)", ind.second.fitness.adjusted_fitness, values[i].second,
-//                             ind.second.fitness.raw_fitness, ind.first, values[i].first);
                     if (ind.second.fitness.adjusted_fitness >= values[i].second)
                     {
                         bool doesnt_contain = true;
@@ -70,77 +68,15 @@ namespace blt::gp
             }
             
             for (blt::size_t i = 0; i < config.elites; i++)
-                next_pop.get_individuals().push_back(current_pop.get_individuals()[values[i].first]);
+                next_pop.push_back(current_pop.get_individuals()[values[i].first].tree);
-        }
-    };
-    
-    template<typename Crossover, typename Mutation, typename Reproduction>
-    constexpr inline auto proportionate_next_pop_creator = [](
-            const selector_args& args, Crossover crossover_selection, Mutation mutation_selection, Reproduction reproduction_selection) {
-        auto& [program, next_pop, current_pop, current_stats, config, random] = args;
-        
-        double total_prob = config.mutation_chance + config.crossover_chance;
-        double crossover_chance = config.crossover_chance / total_prob;
-        double mutation_chance = crossover_chance + config.mutation_chance / total_prob;
-        
-        perform_elitism(args);
-        
-        while (next_pop.get_individuals().size() < config.population_size)
-        {
-            auto type = random.get_double();
-            if (type > crossover_chance && type < mutation_chance)
-            {
-                // crossover
-                auto& p1 = crossover_selection.select(program, current_pop, current_stats);
-                auto& p2 = crossover_selection.select(program, current_pop, current_stats);
-                
-                auto results = config.crossover.get().apply(program, p1, p2);
-                
-                // if crossover fails, we can check for mutation on these guys. otherwise straight copy them into the next pop
-                if (results)
-                {
-                    next_pop.get_individuals().emplace_back(std::move(results->child1));
-                    // annoying check
-                    if (next_pop.get_individuals().size() < config.population_size)
-                        next_pop.get_individuals().emplace_back(std::move(results->child2));
-                } else
-                {
-                    if (config.try_mutation_on_crossover_failure && random.choice(config.mutation_chance))
-                        next_pop.get_individuals().emplace_back(std::move(config.mutator.get().apply(program, p1)));
-                    else
-                        next_pop.get_individuals().push_back(individual{p1});
-                    // annoying check.
-                    if (next_pop.get_individuals().size() < config.population_size)
-                    {
-                        if (config.try_mutation_on_crossover_failure && random.choice(config.mutation_chance))
-                            next_pop.get_individuals().emplace_back(std::move(config.mutator.get().apply(program, p2)));
-                        else
-                            next_pop.get_individuals().push_back(individual{p2});
-                    }
-                }
-            } else if (type > mutation_chance)
-            {
-                // mutation
-                auto& p = mutation_selection.select(program, current_pop, current_stats);
-                next_pop.get_individuals().emplace_back(std::move(config.mutator.get().apply(program, p)));
-            } else
-            {
-                // reproduction
-                auto& p = reproduction_selection.select(program, current_pop, current_stats);
-                next_pop.get_individuals().push_back(individual{p});
-            }
         }
     };
     
     template<typename Crossover, typename Mutation, typename Reproduction>
     constexpr inline auto default_next_pop_creator = [](
-            const blt::gp::selector_args& args, Crossover crossover_selection, Mutation mutation_selection, Reproduction reproduction_selection) {
+            blt::gp::selector_args& args, Crossover& crossover_selection, Mutation& mutation_selection, Reproduction& reproduction_selection) {
         auto& [program, next_pop, current_pop, current_stats, config, random] = args;
         
-        perform_elitism(args);
-        
-        while (next_pop.get_individuals().size() < config.population_size)
-        {
         int sel = random.get_i32(0, 3);
         switch (sel)
         {
@@ -157,10 +93,8 @@ namespace blt::gp
                     // if crossover fails, we can check for mutation on these guys. otherwise straight copy them into the next pop
                     if (results)
                     {
-                            next_pop.get_individuals().emplace_back(std::move(results->child1));
+                        next_pop.push_back(std::move(results->child1));
-                            // annoying check
+                        next_pop.push_back(std::move(results->child2));
-                            if (next_pop.get_individuals().size() < config.population_size)
-                                next_pop.get_individuals().emplace_back(std::move(results->child2));
                     }
                 }
                 break;
@@ -169,7 +103,7 @@ namespace blt::gp
                 {
                     // mutation
                     auto& p = mutation_selection.select(program, current_pop, current_stats);
-                        next_pop.get_individuals().emplace_back(std::move(config.mutator.get().apply(program, p)));
+                    next_pop.push_back(std::move(config.mutator.get().apply(program, p)));
                 }
                 break;
             case 2:
@@ -177,12 +111,15 @@ namespace blt::gp
                 {
                     // reproduction
                     auto& p = reproduction_selection.select(program, current_pop, current_stats);
-                        next_pop.get_individuals().push_back(individual{p});
+                    next_pop.push_back(p);
                 }
                 break;
             default:
+#if BLT_DEBUG_LEVEL > 0
                 BLT_ABORT("This is not possible!");
-            }
+#else
+                BLT_UNREACHABLE;
+#endif
         }
     };
     

include/blt/gp/stack.h

@@ -54,37 +54,14 @@ namespace blt::gp
                 blt::size_t total_size_bytes = 0;
                 blt::size_t total_used_bytes = 0;
                 blt::size_t total_remaining_bytes = 0;
-                blt::size_t total_no_meta_bytes = 0;
-                
-                blt::size_t total_dealloc = 0;
-                blt::size_t total_dealloc_used = 0;
-                blt::size_t total_dealloc_remaining = 0;
-                blt::size_t total_dealloc_no_meta = 0;
-                
-                blt::size_t blocks = 0;
                 
                 friend std::ostream& operator<<(std::ostream& stream, const size_data_t& data)
                 {
                     stream << "[";
-                    stream << data.total_used_bytes << "/";
+                    stream << data.total_used_bytes << " / " << data.total_size_bytes;
-                    stream << data.total_size_bytes << "(";
+                    stream << " ("
-                    stream << (static_cast<double>(data.total_used_bytes) / static_cast<double>(data.total_size_bytes) * 100) << "%), ";
+                           << (data.total_size_bytes != 0 ? (static_cast<double>(data.total_used_bytes) / static_cast<double>(data.total_size_bytes) *
-                    stream << data.total_used_bytes << "/";
+                                                             100) : 0) << "%); space left: " << data.total_remaining_bytes << "]";
-                    stream << data.total_no_meta_bytes << "(";
-                    stream << (static_cast<double>(data.total_used_bytes) / static_cast<double>(data.total_no_meta_bytes) * 100)
-                           << "%), (empty space: ";
-                    stream << data.total_remaining_bytes << ") blocks: " << data.blocks << " || unallocated space: ";
-                    stream << data.total_dealloc_used << "/";
-                    stream << data.total_dealloc;
-                    if (static_cast<double>(data.total_dealloc) > 0)
-                        stream << "(" << (static_cast<double>(data.total_dealloc_used) / static_cast<double>(data.total_dealloc) * 100) << "%)";
-                    stream << ", ";
-                    stream << data.total_dealloc_used << "/";
-                    stream << data.total_dealloc_no_meta;
-                    if (data.total_dealloc_no_meta > 0)
-                        stream << "(" << (static_cast<double>(data.total_dealloc_used) / static_cast<double>(data.total_dealloc_no_meta * 100))
-                               << "%)";
-                    stream << ", (empty space: " << data.total_dealloc_remaining << ")]";
                     return stream;
                 }
             };
@@ -132,6 +109,8 @@ namespace blt::gp
             
             void insert(const stack_allocator& stack)
             {
+                if (stack.empty())
+                    return;
                 if (size_ < stack.bytes_stored + bytes_stored)
                     expand(stack.bytes_stored + bytes_stored);
                 std::memcpy(data_ + bytes_stored, stack.data_, stack.bytes_stored);
@@ -140,6 +119,8 @@ namespace blt::gp
             
             void copy_from(const stack_allocator& stack, blt::size_t bytes)
             {
+                if (bytes == 0)
+                    return;
                 if (size_ < bytes + bytes_stored)
                     expand(bytes + bytes_stored);
                 std::memcpy(data_ + bytes_stored, stack.data_ + (stack.bytes_stored - bytes), bytes);
@@ -148,6 +129,8 @@ namespace blt::gp
             
             void copy_from(blt::u8* data, blt::size_t bytes)
             {
+                if (bytes == 0 || data == nullptr)
+                    return;
                 if (size_ < bytes + bytes_stored)
                     expand(bytes + bytes_stored);
                 std::memcpy(data_ + bytes_stored, data, bytes);
@@ -156,6 +139,8 @@ namespace blt::gp
             
             void copy_to(blt::u8* data, blt::size_t bytes)
             {
+                if (bytes == 0 || data == nullptr)
+                    return;
                 std::memcpy(data, data_ + (bytes_stored - bytes), bytes);
             }
             
@@ -174,6 +159,10 @@ namespace blt::gp
                 static_assert(std::is_trivially_copyable_v<NO_REF> && "Type must be bitwise copyable!");
                 static_assert(alignof(NO_REF) <= MAX_ALIGNMENT && "Type alignment must not be greater than the max alignment!");
                 constexpr auto size = aligned_size(sizeof(NO_REF));
+#if BLT_DEBUG_LEVEL > 0
+                if (bytes_stored < size)
+                    BLT_ABORT("Not enough bytes left to pop!");
+#endif
                 bytes_stored -= size;
                 return *reinterpret_cast<T*>(data_ + bytes_stored);
             }
@@ -184,16 +173,31 @@ namespace blt::gp
                 static_assert(std::is_trivially_copyable_v<NO_REF> && "Type must be bitwise copyable!");
                 static_assert(alignof(NO_REF) <= MAX_ALIGNMENT && "Type alignment must not be greater than the max alignment!");
                 auto size = aligned_size(sizeof(NO_REF)) + bytes;
+#if BLT_DEBUG_LEVEL > 0
+                if (bytes_stored < size)
+                    BLT_ABORT(("Not enough bytes in stack to reference " + std::to_string(size) + " bytes requested but " + std::to_string(bytes) +
+                               " bytes stored!").c_str());
+#endif
                 return *reinterpret_cast<NO_REF*>(data_ + bytes_stored - size);
             }
             
             void pop_bytes(blt::size_t bytes)
             {
+#if BLT_DEBUG_LEVEL > 0
+                if (bytes_stored < bytes)
+                    BLT_ABORT(("Not enough bytes in stack to pop " + std::to_string(bytes) + " bytes requested but " + std::to_string(bytes) +
+                               " bytes stored!").c_str());
+#endif
                 bytes_stored -= bytes;
             }
             
             void transfer_bytes(stack_allocator& to, blt::size_t bytes)
             {
+#if BLT_DEBUG_LEVEL > 0
+                if (bytes_stored < bytes)
+                    BLT_ABORT(("Not enough bytes in stack to transfer " + std::to_string(bytes) + " bytes requested but " + std::to_string(bytes) +
+                               " bytes stored!").c_str());
+#endif
                 to.copy_from(*this, aligned_size(bytes));
                 pop_bytes(bytes);
             }
@@ -297,7 +301,7 @@ namespace blt::gp
                         if (!mask_r[index])
                             return;
                     }
-                    from<NO_REF_T<T>>(offset).drop();
+                    from<NO_REF_T<T >>(offset).drop();
                 }
             }
             

lib/blt

@@ -1 +1 @@
-Subproject commit 941aa6809c92f05c64ca6624d5898958cfac496d
+Subproject commit 97990401e2332276b5397060a3ccaf19f07fb999

src/transformers.cpp

@@ -243,7 +243,7 @@ namespace blt::gp
         vals_r.pop_bytes(static_cast<blt::ptrdiff_t>(total_bytes_after + accumulate_type_sizes(begin_itr, end_itr)));
         
         // insert the new tree then move back the data from after the original mutation point.
-        vals_r.insert(std::move(new_vals_r));
+        vals_r.insert(new_vals_r);
         vals_r.copy_from(stack_after_data, total_bytes_after);
         
         auto before = begin_itr - 1;
@@ -252,7 +252,7 @@ namespace blt::gp
         
         // this will check to make sure that the tree is in a correct and executable state. it requires that the evaluation is context free!
 #if BLT_DEBUG_LEVEL >= 2
-        BLT_ASSERT(new_vals_r.empty());
+//        BLT_ASSERT(new_vals_r.empty());
         //BLT_ASSERT(stack_after.empty());
         blt::size_t bytes_expected = 0;
         auto bytes_size = vals_r.size().total_used_bytes;
@@ -690,7 +690,7 @@ namespace blt::gp
                     vals.copy_from(from_ptr, from_bytes);
                     vals.copy_from(after_ptr, after_to_bytes);
                     
-                    static std::vector<op_container_t> op_copy;
+                    static thread_local std::vector<op_container_t> op_copy;
                     op_copy.clear();
                     op_copy.insert(op_copy.begin(), ops.begin() + from_child.start, ops.begin() + from_child.end);
                     

thread_branch.txt

@@ -0,0 +1,41 @@
+ Performance counter stats for './cmake-build-release/blt-symbolic-regression-example' (30 runs):
+
+    35,671,860,546      branches                                                                ( +-  5.05% )  (20.11%)
+       130,603,525      branch-misses                    #    0.37% of all branches             ( +-  4.61% )  (20.67%)
+        43,684,408      cache-misses                     #    9.61% of all cache refs           ( +-  3.08% )  (20.97%)
+       454,604,804      cache-references                                                        ( +-  4.53% )  (21.30%)
+    72,861,649,501      cycles                                                                  ( +-  5.33% )  (22.00%)
+   170,811,735,018      instructions                     #    2.34  insn per cycle              ( +-  5.59% )  (22.84%)
+                 0      alignment-faults                                                      
+            33,002      cgroup-switches                                                         ( +-  1.71% )
+           293,932      faults                                                                  ( +-  4.09% )
+     1,130,322,318 ns   duration_time                                                           ( +-  3.73% )
+    16,750,942,537 ns   user_time                                                               ( +-  1.71% )
+     1,165,192,903 ns   system_time                                                             ( +-  0.87% )
+    57,551,179,178      L1-dcache-loads                                                         ( +-  5.63% )  (22.36%)
+       214,283,064      L1-dcache-load-misses            #    0.37% of all L1-dcache accesses   ( +-  5.58% )  (22.13%)
+        75,685,527      L1-dcache-prefetches                                                    ( +-  7.55% )  (22.07%)
+     1,115,360,458      L1-icache-loads                                                         ( +-  3.91% )  (21.67%)
+         2,868,754      L1-icache-load-misses            #    0.26% of all L1-icache accesses   ( +-  3.34% )  (21.34%)
+        65,107,178      dTLB-loads                                                              ( +-  8.94% )  (21.00%)
+         4,971,480      dTLB-load-misses                 #    7.64% of all dTLB cache accesses  ( +-  3.70% )  (20.90%)
+           452,351      iTLB-loads                                                              ( +-  4.80% )  (20.62%)
+         1,600,933      iTLB-load-misses                 #  353.91% of all iTLB cache accesses  ( +-  3.68% )  (20.62%)
+       332,075,460      l2_request_g1.all_no_prefetch                                           ( +-  4.59% )  (20.73%)
+           293,932      page-faults                                                             ( +-  4.09% )
+           293,928      page-faults:u                                                           ( +-  4.09% )
+                 3      page-faults:k                                                           ( +-  4.92% )
+    58,806,652,381      L1-dcache-loads                                                         ( +-  5.44% )  (20.61%)
+       216,591,223      L1-dcache-load-misses            #    0.38% of all L1-dcache accesses   ( +-  5.39% )  (21.02%)
+   <not supported>      LLC-loads                                                             
+   <not supported>      LLC-load-misses                                                       
+     1,059,748,012      L1-icache-loads                                                         ( +-  4.29% )  (21.55%)
+         2,615,017      L1-icache-load-misses            #    0.23% of all L1-icache accesses   ( +-  3.34% )  (21.85%)
+        65,917,126      dTLB-loads                                                              ( +-  8.89% )  (21.78%)
+         4,717,351      dTLB-load-misses                 #    7.25% of all dTLB cache accesses  ( +-  3.52% )  (22.05%)
+           459,796      iTLB-loads                                                              ( +-  5.92% )  (21.77%)
+         1,512,986      iTLB-load-misses                 #  334.47% of all iTLB cache accesses  ( +-  3.64% )  (21.26%)
+        74,656,433      L1-dcache-prefetches                                                    ( +-  7.94% )  (20.50%)
+   <not supported>      L1-dcache-prefetch-misses                                             
+
+            1.1303 +- 0.0422 seconds time elapsed  ( +-  3.73% )