From baf257758a0ebb54b65c769c3af95407e22090ca Mon Sep 17 00:00:00 2001
From: Christoph Urlacher <christoph.urlacher@protonmail.com>
Date: Sun, 22 Feb 2026 23:54:14 +0100
Subject: [PATCH] replace openmp with thread-pool library bc openmp has larger
 fork boundary overhead

---
 include/physics.hpp | 13 +++++++++++++
 src/main.cpp        |  7 -------
 src/physics.cpp     | 24 +++++++++++++++++++++++-
 src/renderer.cpp    |  7 +++++++
 4 files changed, 43 insertions(+), 8 deletions(-)

diff --git a/include/physics.hpp b/include/physics.hpp
index 153be68..410d436 100644
--- a/include/physics.hpp
+++ b/include/physics.hpp
@@ -13,6 +13,10 @@
 #include "octree.hpp"
 #endif
 
+#ifndef WEB
+#include <BS_thread_pool.hpp>
+#endif
+
 class Mass {
 public:
   const float mass;
@@ -71,6 +75,10 @@ private:
   int last_springs_count;
 #endif
 
+#ifndef WEB
+  BS::thread_pool<BS::tp::none> threads;
+#endif
+
 public:
   // This is the main ownership of all the states/masses/springs.
   // TODO: Everything is stored multiple times but idc (currently).
@@ -82,6 +90,11 @@ public:
 #ifndef BARNES_HUT
     last_build = REPULSION_GRID_REFRESH;
 #endif
+
+#ifndef WEB
+    std::cout << "Thread-Pool: " << threads.get_thread_count() << " threads."
+              << std::endl;
+#endif
   };
 
   MassSpringSystem(const MassSpringSystem &copy) = delete;
diff --git a/src/main.cpp b/src/main.cpp
index 37b2873..254cd23 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -8,9 +8,6 @@
 #include "renderer.hpp"
 #include "state.hpp"
 
-#ifndef WEB
-#include <omp.h>
-#endif
 #ifdef PRINT_TIMINGS
 #include <chrono>
 #include <ratio>
@@ -33,10 +30,6 @@ auto main(int argc, char *argv[]) -> int {
   //   return 1;
   // }
 
-#ifndef WEB
-  std::cout << "OpenMP: " << omp_get_max_threads() << " threads." << std::endl;
-#endif
-
   // RayLib window setup
   SetTraceLogLevel(LOG_ERROR);
   // SetConfigFlags(FLAG_VSYNC_HINT);
diff --git a/src/physics.cpp b/src/physics.cpp
index 82d8813..6c4d899 100644
--- a/src/physics.cpp
+++ b/src/physics.cpp
@@ -138,6 +138,8 @@ auto MassSpringSystem::ClearForces() -> void {
 }
 
 auto MassSpringSystem::CalculateSpringForces() -> void {
+  ZoneScoped;
+
   for (auto &[states, spring] : springs) {
     spring.CalculateSpringForce();
   }
@@ -145,6 +147,8 @@ auto MassSpringSystem::CalculateSpringForces() -> void {
 
 #ifdef BARNES_HUT
 auto MassSpringSystem::BuildOctree() -> void {
+  ZoneScoped;
+
   octree.nodes.clear();
   octree.nodes.reserve(masses.size() * 2);
 
@@ -228,17 +232,33 @@ auto MassSpringSystem::BuildUniformGrid() -> void {
 
 auto MassSpringSystem::CalculateRepulsionForces() -> void {
   ZoneScoped;
+
 #ifdef BARNES_HUT
   BuildOctree();
 
 // Calculate forces using Barnes-Hut
-#pragma omp parallel for schedule(dynamic, 256)
+#ifdef WEB
   for (int i = 0; i < mass_pointers.size(); ++i) {
     int root = 0;
     Vector3 force = octree.CalculateForce(root, mass_pointers[i]->position);
 
     mass_pointers[i]->force = Vector3Add(mass_pointers[i]->force, force);
   }
+#else
+  threads.detach_blocks(
+      0, mass_pointers.size(),
+      [&](int start, int end) {
+        for (int i = start; i < end; ++i) {
+          int root = 0;
+          Vector3 force =
+              octree.CalculateForce(root, mass_pointers[i]->position);
+
+          mass_pointers[i]->force = Vector3Add(mass_pointers[i]->force, force);
+        }
+      },
+      256);
+  threads.wait();
+#endif
 
 #else
 
@@ -312,6 +332,8 @@ auto MassSpringSystem::CalculateRepulsionForces() -> void {
 }
 
 auto MassSpringSystem::VerletUpdate(float delta_time) -> void {
+  ZoneScoped;
+
   for (auto &[state, mass] : masses) {
     mass.VerletUpdate(delta_time);
   }
diff --git a/src/renderer.cpp b/src/renderer.cpp
index bfb303e..285ff5e 100644
--- a/src/renderer.cpp
+++ b/src/renderer.cpp
@@ -5,6 +5,7 @@
 #include <raylib.h>
 #include <raymath.h>
 #include <rlgl.h>
+#include <tracy/Tracy.hpp>
 #include <unordered_set>
 
 #include "config.hpp"
@@ -64,6 +65,8 @@ auto Renderer::DrawMassSprings(const MassSpringSystem &mass_springs,
                                const State &current_state,
                                const std::unordered_set<State> &winning_states)
     -> void {
+  ZoneScoped;
+
   // Prepare cube instancing
   if (transforms == nullptr) {
     AllocateGraphInstancing(mass_springs);
@@ -135,6 +138,8 @@ auto Renderer::DrawMassSprings(const MassSpringSystem &mass_springs,
 auto Renderer::DrawKlotski(const State &state, int hov_x, int hov_y, int sel_x,
                            int sel_y, int block_add_x, int block_add_y,
                            const WinCondition win_condition) -> void {
+  ZoneScoped;
+
   BeginTextureMode(klotski_target);
   ClearBackground(RAYWHITE);
 
@@ -220,6 +225,8 @@ auto Renderer::DrawMenu(const MassSpringSystem &mass_springs,
                         int current_preset, const State &current_state,
                         const std::unordered_set<State> &winning_states)
     -> void {
+  ZoneScoped;
+
   BeginTextureMode(menu_target);
   ClearBackground(RAYWHITE);