annoying

CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.25)
-project(blt-gp VERSION 0.1.23)
+project(blt-gp VERSION 0.1.26)
 
 include(CTest)
 

examples/symbolic_regression.cpp

@@ -170,6 +170,11 @@ int main()
     
     BLT_PRINT_PROFILE("Symbolic Regression", blt::PRINT_CYCLES | blt::PRINT_THREAD | blt::PRINT_WALL);
 
+#ifdef BLT_TRACK_ALLOCATIONS
+    BLT_TRACE("Total Allocations: %ld times with a total of %s", blt::gp::tracker.getAllocations(),
+              blt::byte_convert_t(blt::gp::tracker.getAllocatedBytes()).convert_to_nearest_type().to_pretty_string().c_str());
+#endif
+
 //    BLT_TRACE("Allocations:");
 //    auto h = static_cast<blt::ptrdiff_t>(blt::gp::hello.load());
 //    auto u = static_cast<blt::ptrdiff_t>(blt::gp::unhello.load());

include/blt/gp/generators.h

@@ -51,7 +51,7 @@ namespace blt::gp
     class tree_generator_t
     {
         public:
-            virtual tree_t generate(const generator_arguments& args) = 0;
+            virtual void generate(tree_t& out, const generator_arguments& args) = 0;
             
             virtual ~tree_generator_t() = default;
     };
@@ -59,13 +59,13 @@ namespace blt::gp
     class grow_generator_t : public tree_generator_t
     {
         public:
-            tree_t generate(const generator_arguments& args) final;
+            void generate(tree_t& out, const generator_arguments& args) final;
     };
     
     class full_generator_t : public tree_generator_t
     {
         public:
-            tree_t generate(const generator_arguments& args) final;
+            void generate(tree_t& out, const generator_arguments& args) final;
     };
     
     class population_initializer_t

include/blt/gp/program.h

@@ -123,7 +123,6 @@ namespace blt::gp
                     static thread_local evaluation_context results{};
                     results.values.reset();
                     results.values.reserve(largest);
-//                    BLT_DEBUG("%ld stored %ld", largest, results.values.internal_storage_size());
                     
                     blt::size_t total_so_far = 0;
                     
@@ -301,7 +300,7 @@ namespace blt::gp
             
             template<typename... Operators, size_t... operator_ids>
             static inline void call_jmp_table_internal(size_t op, void* context, stack_allocator& write_stack, stack_allocator& read_stack,
-                                                       std::integer_sequence<size_t, operator_ids...>, Operators&... operators)
+                                                       std::integer_sequence<size_t, operator_ids...>, Operators& ... operators)
             {
                 if (op >= sizeof...(operator_ids))
                 {
@@ -502,22 +501,16 @@ namespace blt::gp
                                     while (thread_helper.next_gen_left > 0)
                                     {
                                         blt::size_t size = 0;
-                                        blt::size_t begin = 0;
                                         blt::size_t end = thread_helper.next_gen_left.load(std::memory_order_relaxed);
                                         do
                                         {
                                             size = std::min(end, config.evaluation_size);
-                                            begin = end - size;
                                         } while (!thread_helper.next_gen_left.compare_exchange_weak(end, end - size,
                                                                                                     std::memory_order::memory_order_relaxed,
                                                                                                     std::memory_order::memory_order_relaxed));
                                         
-                                        //auto measure = tracker.start_measurement();
+                                        while (new_children.size() < size)
-                                        for (blt::size_t i = begin; i < end; i++)
                                             func(args, crossover_selection, mutation_selection, reproduction_selection);
-                                        //tracker.stop_measurement(measure);
-                                        //BLT_TRACE("Allocated %ld times with a total of %s", measure.getAllocationDifference(),
-                                        //          blt::byte_convert_t(measure.getAllocatedByteDifference()).convert_to_nearest_type().to_pretty_string().c_str());
                                         
                                         {
                                             std::scoped_lock lock(thread_helper.thread_generation_lock);
@@ -548,6 +541,7 @@ namespace blt::gp
             
             void next_generation()
             {
+                BLT_ASSERT_MSG(next_pop.get_individuals().size() == config.population_size, ("pop size: " + std::to_string(next_pop.get_individuals().size())).c_str());
                 current_pop = std::move(next_pop);
                 current_generation++;
             }

include/blt/gp/selection.h

@@ -24,6 +24,7 @@
 #include <blt/gp/config.h>
 #include <blt/gp/random.h>
 #include <blt/std/assert.h>
+#include "blt/std/format.h"
 
 namespace blt::gp
 {
@@ -84,6 +85,7 @@ namespace blt::gp
                 // everyone gets a chance once per loop.
                 if (random.choice(config.crossover_chance))
                 {
+//                    auto state = tracker.start_measurement();
                     // crossover
                     auto& p1 = crossover_selection.select(program, current_pop, current_stats);
                     auto& p2 = crossover_selection.select(program, current_pop, current_stats);
@@ -94,24 +96,36 @@ namespace blt::gp
                     if (results)
                     {
                         next_pop.push_back(std::move(results->child1));
+                        if (next_pop.size() != config.population_size)
                             next_pop.push_back(std::move(results->child2));
                     }
+//                    tracker.stop_measurement(state);
+//                    BLT_TRACE("Crossover Allocated %ld times with a total of %s", state.getAllocationDifference(),
+//                              blt::byte_convert_t(state.getAllocatedByteDifference()).convert_to_nearest_type().to_pretty_string().c_str());
                 }
                 break;
             case 1:
                 if (random.choice(config.mutation_chance))
                 {
+//                    auto state = tracker.start_measurement();
                     // mutation
                     auto& p = mutation_selection.select(program, current_pop, current_stats);
                     next_pop.push_back(std::move(config.mutator.get().apply(program, p)));
+//                    tracker.stop_measurement(state);
+//                    BLT_TRACE("Mutation Allocated %ld times with a total of %s", state.getAllocationDifference(),
+//                              blt::byte_convert_t(state.getAllocatedByteDifference()).convert_to_nearest_type().to_pretty_string().c_str());
                 }
                 break;
             case 2:
                 if (config.reproduction_chance > 0 && random.choice(config.reproduction_chance))
                 {
+//                    auto state = tracker.start_measurement();
                     // reproduction
                     auto& p = reproduction_selection.select(program, current_pop, current_stats);
                     next_pop.push_back(p);
+//                    tracker.stop_measurement(state);
+//                    BLT_TRACE("Reproduction Allocated %ld times with a total of %s", state.getAllocationDifference(),
+//                              blt::byte_convert_t(state.getAllocatedByteDifference()).convert_to_nearest_type().to_pretty_string().c_str());
                 }
                 break;
             default:

include/blt/gp/stack.h

@@ -121,7 +121,7 @@ namespace blt::gp
                 return getAllocatedBytes() - getDeallocatedBytes();
             }
             
-            allocation_data_t start_measurement()
+            [[nodiscard]] allocation_data_t start_measurement() const
             {
                 allocation_data_t data{};
                 data.start_allocations = allocations;
@@ -131,7 +131,7 @@ namespace blt::gp
                 return data;
             }
             
-            void stop_measurement(allocation_data_t& data)
+            void stop_measurement(allocation_data_t& data) const
             {
                 data.end_allocations = allocations;
                 data.end_deallocations = deallocations;

include/blt/gp/tree.h

@@ -57,6 +57,8 @@ namespace blt::gp
         public:
             explicit tree_t(gp_program& program);
             
+            void clear(gp_program& program);
+            
             struct child_t
             {
                 blt::ptrdiff_t start;

src/generators.cpp

@@ -49,11 +49,10 @@ namespace blt::gp
     }
     
     template<typename Func>
-    inline tree_t create_tree(Func&& perChild, const generator_arguments& args)
+    inline void create_tree(tree_t& tree, Func&& perChild, const generator_arguments& args)
     {
         std::stack<stack> tree_generator = get_initial_stack(args.program, args.root_type);
         blt::size_t max_depth = 0;
-        tree_t tree{args.program};
         
         while (!tree_generator.empty())
         {
@@ -77,13 +76,11 @@ namespace blt::gp
             for (const auto& child : info.argument_types)
                 std::forward<Func>(perChild)(args.program, tree_generator, child, top.depth + 1);
         }
-        
-        return tree;
     }
     
-    tree_t grow_generator_t::generate(const generator_arguments& args)
+    void grow_generator_t::generate(tree_t& tree, const generator_arguments& args)
     {
-        return create_tree([args](gp_program& program, std::stack<stack>& tree_generator, type_id type, blt::size_t new_depth) {
+        return create_tree(tree, [args](gp_program& program, std::stack<stack>& tree_generator, type_id type, blt::size_t new_depth) {
             if (new_depth >= args.max_depth)
             {
                 if (program.get_type_terminals(type).empty())
@@ -99,9 +96,9 @@ namespace blt::gp
         }, args);
     }
     
-    tree_t full_generator_t::generate(const generator_arguments& args)
+    void full_generator_t::generate(tree_t& tree, const generator_arguments& args)
     {
-        return create_tree([args](gp_program& program, std::stack<stack>& tree_generator, type_id type, blt::size_t new_depth) {
+        return create_tree(tree, [args](gp_program& program, std::stack<stack>& tree_generator, type_id type, blt::size_t new_depth) {
             if (new_depth >= args.max_depth)
             {
                 if (program.get_type_terminals(type).empty())
@@ -119,7 +116,11 @@ namespace blt::gp
         population_t pop;
         
         for (auto i = 0ul; i < args.size; i++)
-            pop.get_individuals().emplace_back(grow.generate(args.to_gen_args()));
+        {
+            tree_t tree{args.program};
+            grow.generate(tree, args.to_gen_args());
+            pop.get_individuals().emplace_back(tree);
+        }
         
         return pop;
     }
@@ -129,7 +130,11 @@ namespace blt::gp
         population_t pop;
         
         for (auto i = 0ul; i < args.size; i++)
-            pop.get_individuals().emplace_back(full.generate(args.to_gen_args()));
+        {
+            tree_t tree{args.program};
+            full.generate(tree, args.to_gen_args());
+            pop.get_individuals().emplace_back(tree);
+        }
         
         return pop;
     }
@@ -140,10 +145,12 @@ namespace blt::gp
         
         for (auto i = 0ul; i < args.size; i++)
         {
+            tree_t tree{args.program};
             if (args.program.get_random().choice())
-                pop.get_individuals().emplace_back(full.generate(args.to_gen_args()));
+                full.generate(tree, args.to_gen_args());
             else
-                pop.get_individuals().emplace_back(grow.generate(args.to_gen_args()));
+                grow.generate(tree, args.to_gen_args());
+            pop.get_individuals().emplace_back(tree);
         }
         
         return pop;
@@ -160,22 +167,26 @@ namespace blt::gp
         {
             for (auto i = 0ul; i < per_step; i++)
             {
+                tree_t tree{args.program};
                 if (args.program.get_random().choice())
-                    pop.get_individuals().emplace_back(full.generate({args.program, args.root_type, args.min_depth, depth}));
+                    full.generate(tree, {args.program, args.root_type, args.min_depth, depth});
                 else
-                    pop.get_individuals().emplace_back(grow.generate({args.program, args.root_type, args.min_depth, depth}));
+                    grow.generate(tree, {args.program, args.root_type, args.min_depth, depth});
+                pop.get_individuals().emplace_back(tree);
             }
         }
         
         for (auto i = 0ul; i < remainder; i++)
         {
+            tree_t tree{args.program};
             if (args.program.get_random().choice())
-                pop.get_individuals().emplace_back(full.generate(args.to_gen_args()));
+                full.generate(tree, args.to_gen_args());
             else
-                pop.get_individuals().emplace_back(grow.generate(args.to_gen_args()));
+                grow.generate(tree, args.to_gen_args());
+            pop.get_individuals().emplace_back(tree);
         }
         
-        blt_assert(pop.get_individuals().size() == args.size);
+        BLT_ASSERT(pop.get_individuals().size() == args.size);
         
         return pop;
     }

src/transformers.cpp

@@ -27,6 +27,15 @@
 namespace blt::gp
 {
     
+    grow_generator_t grow_generator;
+    
+    inline tree_t& get_static_tree_tl(gp_program& program)
+    {
+        static thread_local tree_t new_tree{program};
+        new_tree.clear(program);
+        return new_tree;
+    }
+    
     inline blt::size_t accumulate_type_sizes(detail::op_iter begin, detail::op_iter end)
     {
         blt::size_t total = 0;
@@ -47,8 +56,6 @@ namespace blt::gp
         return buffer.data();
     }
     
-    grow_generator_t grow_generator;
-    
     mutation_t::config_t::config_t(): generator(grow_generator)
     {}
     
@@ -228,7 +235,8 @@ namespace blt::gp
         auto begin_itr = ops_r.begin() + begin_point;
         auto end_itr = ops_r.begin() + end_point;
         
-        auto new_tree = config.generator.get().generate({program, type_info.return_type, config.replacement_min_depth, config.replacement_max_depth});
+        auto& new_tree = get_static_tree_tl(program);
+        config.generator.get().generate(new_tree, {program, type_info.return_type, config.replacement_min_depth, config.replacement_max_depth});
         
         auto& new_ops_r = new_tree.get_operations();
         auto& new_vals_r = new_tree.get_values();
@@ -364,7 +372,8 @@ namespace blt::gp
                             if (index < current_func_info.argument_types.size() && val.id != current_func_info.argument_types[index].id)
                             {
                                 // TODO: new config?
-                                auto tree = config.generator.get().generate(
+                                auto& tree = get_static_tree_tl(program);
+                                config.generator.get().generate(tree,
                                         {program, val.id, config.replacement_min_depth, config.replacement_max_depth});
                                 
                                 auto& child = children_data[children_data.size() - 1 - index];
@@ -445,7 +454,8 @@ namespace blt::gp
                             for (blt::ptrdiff_t i = static_cast<blt::ptrdiff_t>(replacement_func_info.argc.argc) - 1;
                                  i >= current_func_info.argc.argc; i--)
                             {
-                                auto tree = config.generator.get().generate(
+                                auto& tree = get_static_tree_tl(program);
+                                config.generator.get().generate(tree,
                                         {program, replacement_func_info.argument_types[i].id, config.replacement_min_depth,
                                          config.replacement_max_depth});
                                 blt::size_t total_bytes_for = tree.total_value_bytes();
@@ -518,7 +528,8 @@ namespace blt::gp
                     blt::size_t start_index = c_node;
                     for (blt::ptrdiff_t i = new_argc - 1; i > static_cast<blt::ptrdiff_t>(arg_position); i--)
                     {
-                        auto tree = config.generator.get().generate(
+                        auto& tree = get_static_tree_tl(program);
+                        config.generator.get().generate(tree,
                                 {program, replacement_func_info.argument_types[i].id, config.replacement_min_depth,
                                  config.replacement_max_depth});
                         blt::size_t total_bytes_for = tree.total_value_bytes();
@@ -531,7 +542,8 @@ namespace blt::gp
                     vals.copy_from(combined_ptr, for_bytes);
                     for (blt::ptrdiff_t i = static_cast<blt::ptrdiff_t>(arg_position) - 1; i >= 0; i--)
                     {
-                        auto tree = config.generator.get().generate(
+                        auto& tree = get_static_tree_tl(program);
+                        config.generator.get().generate(tree,
                                 {program, replacement_func_info.argument_types[i].id, config.replacement_min_depth,
                                  config.replacement_max_depth});
                         blt::size_t total_bytes_for = tree.total_value_bytes();

src/tree.cpp

@@ -299,4 +299,13 @@ namespace blt::gp
     {
     
     }
+    
+    void tree_t::clear(gp_program& program)
+    {
+        auto* f = &program.get_eval_func();
+        if (f != func)
+            func = f;
+        operations.clear();
+        values.reset();
+    }
 }