Use WorkerThreadPool for Server threads

* Servers now use WorkerThreadPool for background computation. * This helps keep the number of threads used fixed at all times. * It also ensures everything works on HTML5 with threads. * And makes it easier to support disabling threads for also HTML5. The only downside of this is that GLES3 threads likely no longer work because every time the render thread is called, the thread context is different. This needs to be researched how to fix. Maybe if GLES3 is used, threaded rendering server should be disabled for the time being?
godotengine · Jan 26, 2024 · 900dc33 · 900dc33
1 parent 0bcc0e9
commit 900dc33
Show file tree

Hide file tree

Showing 13 changed files with 598 additions and 533 deletions.
diff --git a/core/config/engine.cpp b/core/config/engine.cpp
@@ -82,6 +82,16 @@ int Engine::get_audio_output_latency() const {
 	return _audio_output_latency;
 }
 
+void Engine::increment_frames_drawn() {
+	if (frame_server_synced) {
+		server_syncs++;
+	} else {
+		server_syncs = 0;
+	}
+	frame_server_synced = 0;
+
+	frames_drawn++;
+}
 uint64_t Engine::get_frames_drawn() {
 	return frames_drawn;
 }
@@ -351,6 +361,11 @@ Engine *Engine::get_singleton() {
 	return singleton;
 }
 
+bool Engine::notify_frame_server_synced() {
+	frame_server_synced = true;
+	return server_syncs > server_sync_frame_count_warning;
+}
+
 Engine::Engine() {
 	singleton = this;
 }

diff --git a/core/config/engine.h b/core/config/engine.h
@@ -89,6 +89,10 @@ class Engine {
 	String write_movie_path;
 	String shader_cache_path;
 
+	static constexpr int server_sync_frame_count_warning = 5;
+	int server_syncs = 0;
+	bool frame_server_synced = false;
+
 public:
 	static Engine *get_singleton();
 
@@ -175,6 +179,9 @@ class Engine {
 	bool is_generate_spirv_debug_info_enabled() const;
 	int32_t get_gpu_index() const;
 
+	void increment_frames_drawn();
+	bool notify_frame_server_synced();
+
 	Engine();
 	virtual ~Engine() {}
 };

diff --git a/drivers/gles3/rasterizer_gles3.cpp b/drivers/gles3/rasterizer_gles3.cpp
@@ -115,6 +115,10 @@ void RasterizerGLES3::end_viewport(bool p_swap_buffers) {
 	}
 }
 
+void RasterizerGLES3::context_move_to_current_thread() {
+	// I need to move the OpenGL context thread to this thread
+}
+
 void RasterizerGLES3::clear_depth(float p_depth) {
 #ifdef GL_API_ENABLED
 	if (is_gles_over_gl()) {

diff --git a/drivers/gles3/rasterizer_gles3.h b/drivers/gles3/rasterizer_gles3.h
@@ -115,6 +115,8 @@ class RasterizerGLES3 : public RendererCompositor {
 	_ALWAYS_INLINE_ double get_frame_delta_time() const { return delta; }
 	_ALWAYS_INLINE_ double get_total_time() const { return time_total; }
 
+	virtual void context_move_to_current_thread();
+
 	static RasterizerGLES3 *get_singleton() { return singleton; }
 	RasterizerGLES3();
 	~RasterizerGLES3();

diff --git a/main/main.cpp b/main/main.cpp
@@ -3799,11 +3799,11 @@ bool Main::iteration() {
 		if ((!force_redraw_requested) && OS::get_singleton()->is_in_low_processor_usage_mode()) {
 			if (RenderingServer::get_singleton()->has_changed()) {
 				RenderingServer::get_singleton()->draw(true, scaled_step); // flush visual commands
-				Engine::get_singleton()->frames_drawn++;
+				Engine::get_singleton()->increment_frames_drawn();
 			}
 		} else {
 			RenderingServer::get_singleton()->draw(true, scaled_step); // flush visual commands
-			Engine::get_singleton()->frames_drawn++;
+			Engine::get_singleton()->increment_frames_drawn();
 			force_redraw_requested = false;
 		}
 	}

diff --git a/servers/physics_server_2d_wrap_mt.cpp b/servers/physics_server_2d_wrap_mt.cpp
@@ -32,43 +32,28 @@
 
 #include "core/os/os.h"
 
-void PhysicsServer2DWrapMT::thread_exit() {
-	exit.set();
-}
-
 void PhysicsServer2DWrapMT::thread_step(real_t p_delta) {
+	server_thread = Thread::get_caller_id();
+	command_queue.flush_all();
 	physics_server_2d->step(p_delta);
-	step_sem.post();
-}
-
-void PhysicsServer2DWrapMT::_thread_callback(void *_instance) {
-	PhysicsServer2DWrapMT *vsmt = reinterpret_cast<PhysicsServer2DWrapMT *>(_instance);
-
-	vsmt->thread_loop();
+	command_queue.flush_all(); // Flush pending commands before and after
+	server_thread = Thread::UNASSIGNED_ID;
 }
 
-void PhysicsServer2DWrapMT::thread_loop() {
-	server_thread = Thread::get_caller_id();
-
-	physics_server_2d->init();
-
-	exit.clear();
-	step_thread_up.set();
-	while (!exit.is_set()) {
-		// flush commands one by one, until exit is requested
-		command_queue.wait_and_flush();
+void PhysicsServer2DWrapMT::_main_thread_sync() {
+	if (task_id != WorkerThreadPool::INVALID_TASK_ID) {
+		WorkerThreadPool::get_singleton()->wait_for_task_completion(task_id);
+		task_id = WorkerThreadPool::INVALID_TASK_ID;
 	}
 
-	command_queue.flush_all(); // flush all
-
-	physics_server_2d->finish();
+	server_thread = Thread::MAIN_ID;
 }
 
 /* EVENT QUEUING */
 
 void PhysicsServer2DWrapMT::step(real_t p_step) {
 	if (create_thread) {
-		command_queue.push(this, &PhysicsServer2DWrapMT::thread_step, p_step);
+		task_id = WorkerThreadPool::get_singleton()->add_task(callable_mp(this, &PhysicsServer2DWrapMT::thread_step).bind(p_step), true);
 	} else {
 		command_queue.flush_all(); //flush all pending from other threads
 		physics_server_2d->step(p_step);
@@ -77,10 +62,9 @@ void PhysicsServer2DWrapMT::step(real_t p_step) {
 
 void PhysicsServer2DWrapMT::sync() {
 	if (create_thread) {
-		if (first_frame) {
-			first_frame = false;
-		} else {
-			step_sem.wait(); //must not wait if a step was not issued
+		if (task_id != WorkerThreadPool::INVALID_TASK_ID) {
+			WorkerThreadPool::get_singleton()->wait_for_task_completion(task_id);
+			task_id = WorkerThreadPool::INVALID_TASK_ID;
 		}
 	}
 	physics_server_2d->sync();
@@ -95,24 +79,15 @@ void PhysicsServer2DWrapMT::end_sync() {
 }
 
 void PhysicsServer2DWrapMT::init() {
-	if (create_thread) {
-		//OS::get_singleton()->release_rendering_thread();
-		thread.start(_thread_callback, this);
-		while (!step_thread_up.is_set()) {
-			OS::get_singleton()->delay_usec(1000);
-		}
-	} else {
-		physics_server_2d->init();
-	}
+	physics_server_2d->init();
 }
 
 void PhysicsServer2DWrapMT::finish() {
-	if (thread.is_started()) {
-		command_queue.push(this, &PhysicsServer2DWrapMT::thread_exit);
-		thread.wait_to_finish();
-	} else {
-		physics_server_2d->finish();
+	if (task_id != WorkerThreadPool::INVALID_TASK_ID) {
+		WorkerThreadPool::get_singleton()->wait_for_task_completion(task_id);
+		task_id = WorkerThreadPool::INVALID_TASK_ID;
 	}
+	physics_server_2d->finish();
 }
 
 PhysicsServer2DWrapMT::PhysicsServer2DWrapMT(PhysicsServer2D *p_contained, bool p_create_thread) :
@@ -121,12 +96,12 @@ PhysicsServer2DWrapMT::PhysicsServer2DWrapMT(PhysicsServer2D *p_contained, bool
 	create_thread = p_create_thread;
 
 	if (!p_create_thread) {
-		server_thread = Thread::get_caller_id();
+		server_thread = Thread::MAIN_ID;
 	} else {
-		server_thread = 0;
+		server_thread = Thread::UNASSIGNED_ID;
 	}
 
-	main_thread = Thread::get_caller_id();
+	main_thread = Thread::MAIN_ID;
 }
 
 PhysicsServer2DWrapMT::~PhysicsServer2DWrapMT() {

diff --git a/servers/physics_server_2d_wrap_mt.h b/servers/physics_server_2d_wrap_mt.h
@@ -32,6 +32,7 @@
 #define PHYSICS_SERVER_2D_WRAP_MT_H
 
 #include "core/config/project_settings.h"
+#include "core/object/worker_thread_pool.h"
 #include "core/os/thread.h"
 #include "core/templates/command_queue_mt.h"
 #include "core/templates/safe_refcount.h"
@@ -43,29 +44,38 @@
 #define SYNC_DEBUG
 #endif
 
+#ifdef DEBUG_ENABLED
+#define MAIN_THREAD_SYNC                                                                                                                                     \
+	if (Engine::get_singleton()->notify_frame_server_synced()) {                                                                                             \
+		WARN_PRINT("Call to " + String(__FUNCTION__) + " causing PhysicsServer2D synchronizations on every frame. This significantly affects performance."); \
+	}                                                                                                                                                        \
+	const_cast<PhysicsServer2DWrapMT *>(this)->_main_thread_sync();
+#else
+#define MAIN_THREAD_SYNC const_cast<PhysicsServer2DWrapMT *>(this)->_main_thread_sync();
+#endif
+
 class PhysicsServer2DWrapMT : public PhysicsServer2D {
 	mutable PhysicsServer2D *physics_server_2d;
 
 	mutable CommandQueueMT command_queue;
 
 	static void _thread_callback(void *_instance);
-	void thread_loop();
 
 	Thread::ID server_thread;
 	Thread::ID main_thread;
-	SafeFlag exit;
-	Thread thread;
-	SafeFlag step_thread_up;
 	bool create_thread = false;
+	WorkerThreadPool::TaskID task_id = WorkerThreadPool::INVALID_TASK_ID;
 
-	Semaphore step_sem;
 	void thread_step(real_t p_delta);
 
 	void thread_exit();
 
+	void _main_thread_sync();
+
 	bool first_frame = true;
 
 	Mutex alloc_mutex;
+	int pool_max_size = 0;
 
 public:
 #define ServerName PhysicsServer2D
@@ -337,5 +347,6 @@ class PhysicsServer2DWrapMT : public PhysicsServer2D {
 #undef DEBUG_SYNC
 #endif
 #undef SYNC_DEBUG
+#undef MAIN_THREAD_SYNC
 
 #endif // PHYSICS_SERVER_2D_WRAP_MT_H
diff --git a/servers/physics_server_3d_wrap_mt.cpp b/servers/physics_server_3d_wrap_mt.cpp
@@ -32,43 +32,28 @@
 
 #include "core/os/os.h"
 
-void PhysicsServer3DWrapMT::thread_exit() {
-	exit = true;
-}
-
 void PhysicsServer3DWrapMT::thread_step(real_t p_delta) {
+	server_thread = Thread::get_caller_id();
+	command_queue.flush_all();
 	physics_server_3d->step(p_delta);
-	step_sem.post();
-}
-
-void PhysicsServer3DWrapMT::_thread_callback(void *_instance) {
-	PhysicsServer3DWrapMT *vsmt = reinterpret_cast<PhysicsServer3DWrapMT *>(_instance);
-
-	vsmt->thread_loop();
+	command_queue.flush_all(); // Flush pending commands before and after
+	server_thread = Thread::UNASSIGNED_ID;
 }
 
-void PhysicsServer3DWrapMT::thread_loop() {
-	server_thread = Thread::get_caller_id();
-
-	physics_server_3d->init();
-
-	exit = false;
-	step_thread_up = true;
-	while (!exit) {
-		// flush commands one by one, until exit is requested
-		command_queue.wait_and_flush();
+void PhysicsServer3DWrapMT::_main_thread_sync() {
+	if (task_id != WorkerThreadPool::INVALID_TASK_ID) {
+		WorkerThreadPool::get_singleton()->wait_for_task_completion(task_id);
+		task_id = WorkerThreadPool::INVALID_TASK_ID;
 	}
 
-	command_queue.flush_all(); // flush all
-
-	physics_server_3d->finish();
+	server_thread = Thread::MAIN_ID;
 }
 
 /* EVENT QUEUING */
 
 void PhysicsServer3DWrapMT::step(real_t p_step) {
 	if (create_thread) {
-		command_queue.push(this, &PhysicsServer3DWrapMT::thread_step, p_step);
+		task_id = WorkerThreadPool::get_singleton()->add_task(callable_mp(this, &PhysicsServer3DWrapMT::thread_step).bind(p_step), true);
 	} else {
 		command_queue.flush_all(); //flush all pending from other threads
 		physics_server_3d->step(p_step);
@@ -77,10 +62,9 @@ void PhysicsServer3DWrapMT::step(real_t p_step) {
 
 void PhysicsServer3DWrapMT::sync() {
 	if (create_thread) {
-		if (first_frame) {
-			first_frame = false;
-		} else {
-			step_sem.wait(); //must not wait if a step was not issued
+		if (task_id != WorkerThreadPool::INVALID_TASK_ID) {
+			WorkerThreadPool::get_singleton()->wait_for_task_completion(task_id);
+			task_id = WorkerThreadPool::INVALID_TASK_ID;
 		}
 	}
 	physics_server_3d->sync();
@@ -95,24 +79,15 @@ void PhysicsServer3DWrapMT::end_sync() {
 }
 
 void PhysicsServer3DWrapMT::init() {
-	if (create_thread) {
-		//OS::get_singleton()->release_rendering_thread();
-		thread.start(_thread_callback, this);
-		while (!step_thread_up) {
-			OS::get_singleton()->delay_usec(1000);
-		}
-	} else {
-		physics_server_3d->init();
-	}
+	physics_server_3d->init();
 }
 
 void PhysicsServer3DWrapMT::finish() {
-	if (thread.is_started()) {
-		command_queue.push(this, &PhysicsServer3DWrapMT::thread_exit);
-		thread.wait_to_finish();
-	} else {
-		physics_server_3d->finish();
+	if (task_id != WorkerThreadPool::INVALID_TASK_ID) {
+		WorkerThreadPool::get_singleton()->wait_for_task_completion(task_id);
+		task_id = WorkerThreadPool::INVALID_TASK_ID;
 	}
+	physics_server_3d->finish();
 }
 
 PhysicsServer3DWrapMT::PhysicsServer3DWrapMT(PhysicsServer3D *p_contained, bool p_create_thread) :
@@ -121,12 +96,12 @@ PhysicsServer3DWrapMT::PhysicsServer3DWrapMT(PhysicsServer3D *p_contained, bool
 	create_thread = p_create_thread;
 
 	if (!p_create_thread) {
-		server_thread = Thread::get_caller_id();
+		server_thread = Thread::MAIN_ID;
 	} else {
-		server_thread = 0;
+		server_thread = Thread::UNASSIGNED_ID;
 	}
 
-	main_thread = Thread::get_caller_id();
+	main_thread = Thread::MAIN_ID;
 }
 
 PhysicsServer3DWrapMT::~PhysicsServer3DWrapMT() {