[3.x] Disable NVIDIA's threaded OpenGL optimization on Windows

COPYRIGHT.txt

@@ -348,6 +348,11 @@ Comment: NanoSVG
 Copyright: 2013-2014, Mikko Mononen
 License: Zlib
 
+Files: ./thirdparty/nvapi/nvapi_minimal.h
+Comment: Stripped down version of "nvapi.h" from the NVIDIA NVAPI SDK
+Copyright: 2019-2022, NVIDIA Corporation
+License: Expat
+
 Files: ./thirdparty/opus/
 Comment: Opus
 Copyright: 2001-2011, Xiph.Org, Skype Limited, Octasic,

doc/classes/ProjectSettings.xml

@@ -1637,6 +1637,10 @@
 		<member name="rendering/limits/time/time_rollover_secs" type="float" setter="" getter="" default="3600">
 			Shaders have a time variable that constantly increases. At some point, it needs to be rolled back to zero to avoid precision errors on shader animations. This setting specifies when (in seconds).
 		</member>
+		<member name="rendering/misc/compatibility/nvidia_disable_threaded_optimization" type="bool" setter="" getter="" default="false">
+			If [code]true[/code], disables the threaded optimization feature from the NVIDIA drivers, which are known to cause stuttering in most OpenGL applications.
+			[b]Note:[/b] This setting only works on Windows, as threaded optimization is disabled by default on other platforms.
+		</member>
 		<member name="rendering/misc/lossless_compression/force_png" type="bool" setter="" getter="" default="false">
 			If [code]true[/code], the texture importer will import lossless textures using the PNG format. Otherwise, it will default to using WebP.
 		</member>

main/main.cpp

@@ -1246,6 +1246,7 @@ Error Main::setup(const char *execpath, int argc, char *argv[], bool p_second_ph
 		rtm = GLOBAL_DEF("rendering/threads/thread_model", OS::RENDER_THREAD_SAFE);
 	}
 	GLOBAL_DEF("rendering/threads/thread_safe_bvh", false);
+	GLOBAL_DEF_RST("rendering/misc/compatibility/nvidia_disable_threaded_optimization", false);
 
 	if (rtm >= 0 && rtm < 3) {
 #ifdef NO_THREADS

platform/windows/os_windows.cpp

@@ -32,6 +32,7 @@
 
 #include "core/io/marshalls.h"
 #include "core/math/geometry.h"
+#include "core/project_settings.h"
 #include "core/version_generated.gen.h"
 #include "drivers/gles2/rasterizer_gles2.h"
 #include "drivers/gles3/rasterizer_gles3.h"
@@ -44,6 +45,7 @@
 #include "servers/audio_server.h"
 #include "servers/visual/visual_server_raster.h"
 #include "servers/visual/visual_server_wrap_mt.h"
+#include "thirdparty/nvapi/nvapi_minimal.h"
 #include "windows_terminal_logger.h"
 
 #include <avrt.h>
@@ -60,6 +62,218 @@
 __declspec(dllexport) int AmdPowerXpressRequestHighPerformance = 1;
 }
 
+const int OGL_THREAD_CONTROL_ID = 0x20C1221E;
+const int OGL_THREAD_CONTROL_DISABLE = 0x00000002;
+const int OGL_THREAD_CONTROL_ENABLE = 0x00000001;
+
+typedef int(__cdecl *NvAPI_Initialize_t)();
+typedef int(__cdecl *NvAPI_Unload_t)();
+typedef int(__cdecl *NvAPI_GetErrorMessage_t)(unsigned int, NvAPI_ShortString);
+typedef int(__cdecl *NvAPI_DRS_CreateSession_t)(NvDRSSessionHandle *);
+typedef int(__cdecl *NvAPI_DRS_DestroySession_t)(NvDRSSessionHandle);
+typedef int(__cdecl *NvAPI_DRS_LoadSettings_t)(NvDRSSessionHandle);
+typedef int(__cdecl *NvAPI_DRS_CreateProfile_t)(NvDRSSessionHandle, NVDRS_PROFILE *, NvDRSProfileHandle *);
+typedef int(__cdecl *NvAPI_DRS_CreateApplication_t)(NvDRSSessionHandle, NvDRSProfileHandle, NVDRS_APPLICATION *);
+typedef int(__cdecl *NvAPI_DRS_SaveSettings_t)(NvDRSSessionHandle);
+typedef int(__cdecl *NvAPI_DRS_SetSetting_t)(NvDRSSessionHandle, NvDRSProfileHandle, NVDRS_SETTING *);
+typedef int(__cdecl *NvAPI_DRS_FindProfileByName_t)(NvDRSSessionHandle, NvAPI_UnicodeString, NvDRSProfileHandle *);
+typedef int(__cdecl *NvAPI_DRS_GetApplicationInfo_t)(NvDRSSessionHandle, NvDRSProfileHandle, NvAPI_UnicodeString, NVDRS_APPLICATION *);
+typedef int(__cdecl *NvAPI_DRS_DeleteProfile_t)(NvDRSSessionHandle, NvDRSProfileHandle);
+NvAPI_GetErrorMessage_t NvAPI_GetErrorMessage__;
+
+static bool nvapi_err_check(const char *msg, int status) {
+	if (status != 0) {
+		if (OS::get_singleton()->is_stdout_verbose()) {
+			NvAPI_ShortString err_desc = { 0 };
+			NvAPI_GetErrorMessage__(status, err_desc);
+			print_verbose(vformat("%s: %s(code %d)", msg, err_desc, status));
+		}
+		return false;
+	}
+	return true;
+}
+
+// On windows we have to customize the NVIDIA application profile:
+// * disable threaded optimization when using NVIDIA cards to avoid stuttering, see
+//   https://stackoverflow.com/questions/36959508/nvidia-graphics-driver-causing-noticeable-frame-stuttering/37632948
+//   https://github.com/Ryujinx/Ryujinx/blob/master/src/Ryujinx.Common/GraphicsDriver/NVThreadedOptimization.cs
+// * disable G-SYNC in windowed mode, as it results in unstable editor refresh rates
+void GLManagerNative_Windows::_nvapi_setup_profile() {
+	HMODULE nvapi = nullptr;
+#ifdef _WIN64
+	nvapi = LoadLibraryA("nvapi64.dll");
+#else
+	nvapi = LoadLibraryA("nvapi.dll");
+#endif
+
+	if (nvapi == nullptr) {
+		return;
+	}
+
+	void *(__cdecl * NvAPI_QueryInterface)(unsigned int interface_id) = nullptr;
+
+	NvAPI_QueryInterface = (void *(__cdecl *)(unsigned int))(void *)GetProcAddress(nvapi, "nvapi_QueryInterface");
+
+	if (NvAPI_QueryInterface == nullptr) {
+		print_verbose("Error getting NVAPI NvAPI_QueryInterface");
+		return;
+	}
+
+	// Setup NVAPI function pointers
+	NvAPI_Initialize_t NvAPI_Initialize = (NvAPI_Initialize_t)NvAPI_QueryInterface(0x0150E828);
+	NvAPI_GetErrorMessage__ = (NvAPI_GetErrorMessage_t)NvAPI_QueryInterface(0x6C2D048C);
+	NvAPI_DRS_CreateSession_t NvAPI_DRS_CreateSession = (NvAPI_DRS_CreateSession_t)NvAPI_QueryInterface(0x0694D52E);
+	NvAPI_DRS_DestroySession_t NvAPI_DRS_DestroySession = (NvAPI_DRS_DestroySession_t)NvAPI_QueryInterface(0xDAD9CFF8);
+	NvAPI_Unload_t NvAPI_Unload = (NvAPI_Unload_t)NvAPI_QueryInterface(0xD22BDD7E);
+	NvAPI_DRS_LoadSettings_t NvAPI_DRS_LoadSettings = (NvAPI_DRS_LoadSettings_t)NvAPI_QueryInterface(0x375DBD6B);
+	NvAPI_DRS_CreateProfile_t NvAPI_DRS_CreateProfile = (NvAPI_DRS_CreateProfile_t)NvAPI_QueryInterface(0xCC176068);
+	NvAPI_DRS_CreateApplication_t NvAPI_DRS_CreateApplication = (NvAPI_DRS_CreateApplication_t)NvAPI_QueryInterface(0x4347A9DE);
+	NvAPI_DRS_SaveSettings_t NvAPI_DRS_SaveSettings = (NvAPI_DRS_SaveSettings_t)NvAPI_QueryInterface(0xFCBC7E14);
+	NvAPI_DRS_SetSetting_t NvAPI_DRS_SetSetting = (NvAPI_DRS_SetSetting_t)NvAPI_QueryInterface(0x577DD202);
+	NvAPI_DRS_FindProfileByName_t NvAPI_DRS_FindProfileByName = (NvAPI_DRS_FindProfileByName_t)NvAPI_QueryInterface(0x7E4A9A0B);
+	NvAPI_DRS_GetApplicationInfo_t NvAPI_DRS_GetApplicationInfo = (NvAPI_DRS_GetApplicationInfo_t)NvAPI_QueryInterface(0xED1F8C69);
+	NvAPI_DRS_DeleteProfile_t NvAPI_DRS_DeleteProfile = (NvAPI_DRS_DeleteProfile_t)NvAPI_QueryInterface(0x17093206);
+
+	if (!nvapi_err_check("NVAPI: Init failed", NvAPI_Initialize())) {
+		return;
+	}
+
+	print_verbose("NVAPI: Init OK!");
+
+	NvDRSSessionHandle session_handle;
+
+	if (NvAPI_DRS_CreateSession == nullptr) {
+		return;
+	}
+
+	if (!nvapi_err_check("NVAPI: Error creating DRS session", NvAPI_DRS_CreateSession(&session_handle))) {
+		NvAPI_Unload();
+		return;
+	}
+
+	if (!nvapi_err_check("NVAPI: Error loading DRS settings", NvAPI_DRS_LoadSettings(session_handle))) {
+		NvAPI_DRS_DestroySession(session_handle);
+		NvAPI_Unload();
+		return;
+	}
+
+	String app_executable_name = OS::get_singleton()->get_executable_path().get_file();
+	String app_profile_name = GLOBAL_GET("application/config/name");
+	// We need a name anyways, so let's use the engine name if an application name is not available
+	// (this is used mostly by the Project Manager)
+	if (app_profile_name.is_empty()) {
+		app_profile_name = GODOT_VERSION_NAME;
+	}
+	String old_profile_name = app_profile_name + " Nvidia Profile";
+	Char16String app_profile_name_u16 = app_profile_name.utf16();
+	Char16String old_profile_name_u16 = old_profile_name.utf16();
+	Char16String app_executable_name_u16 = app_executable_name.utf16();
+
+	// A previous error in app creation logic could result in invalid profiles,
+	// clean these if they exist before proceeding.
+	NvDRSProfileHandle old_profile_handle;
+
+	int old_status = NvAPI_DRS_FindProfileByName(session_handle, (NvU16 *)(old_profile_name_u16.ptrw()), &old_profile_handle);
+
+	if (old_status == 0) {
+		print_verbose("NVAPI: Deleting old profile...");
+
+		if (!nvapi_err_check("NVAPI: Error deleting old profile", NvAPI_DRS_DeleteProfile(session_handle, old_profile_handle))) {
+			NvAPI_DRS_DestroySession(session_handle);
+			NvAPI_Unload();
+			return;
+		}
+
+		if (!nvapi_err_check("NVAPI: Error deleting old profile", NvAPI_DRS_SaveSettings(session_handle))) {
+			NvAPI_DRS_DestroySession(session_handle);
+			NvAPI_Unload();
+			return;
+		}
+	}
+
+	NvDRSProfileHandle profile_handle = nullptr;
+
+	int profile_status = NvAPI_DRS_FindProfileByName(session_handle, (NvU16 *)(app_profile_name_u16.ptrw()), &profile_handle);
+
+	if (profile_status != 0) {
+		print_verbose("NVAPI: Profile not found, creating...");
+
+		NVDRS_PROFILE profile_info;
+		profile_info.version = NVDRS_PROFILE_VER;
+		profile_info.isPredefined = 0;
+		memcpy(profile_info.profileName, app_profile_name_u16.get_data(), sizeof(char16_t) * app_profile_name_u16.size());
+
+		if (!nvapi_err_check("NVAPI: Error creating profile", NvAPI_DRS_CreateProfile(session_handle, &profile_info, &profile_handle))) {
+			NvAPI_DRS_DestroySession(session_handle);
+			NvAPI_Unload();
+			return;
+		}
+	}
+
+	NVDRS_APPLICATION_V4 app;
+	app.version = NVDRS_APPLICATION_VER_V4;
+
+	int app_status = NvAPI_DRS_GetApplicationInfo(session_handle, profile_handle, (NvU16 *)(app_executable_name_u16.ptrw()), &app);
+
+	if (app_status != 0) {
+		print_verbose("NVAPI: Application not found in profile, creating...");
+
+		app.isPredefined = 0;
+		memcpy(app.appName, app_executable_name_u16.get_data(), sizeof(char16_t) * app_executable_name_u16.size());
+		memcpy(app.launcher, L"", sizeof(wchar_t));
+		memcpy(app.fileInFolder, L"", sizeof(wchar_t));
+
+		if (!nvapi_err_check("NVAPI: Error creating application", NvAPI_DRS_CreateApplication(session_handle, profile_handle, &app))) {
+			NvAPI_DRS_DestroySession(session_handle);
+			NvAPI_Unload();
+			return;
+		}
+	}
+
+	NVDRS_SETTING ogl_thread_control_setting = {};
+	ogl_thread_control_setting.version = NVDRS_SETTING_VER;
+	ogl_thread_control_setting.settingId = OGL_THREAD_CONTROL_ID;
+	ogl_thread_control_setting.settingType = NVDRS_DWORD_TYPE;
+	int thread_control_val = OGL_THREAD_CONTROL_DISABLE;
+	if (!GLOBAL_GET("rendering/misc/compatibility/nvidia_disable_threaded_optimization")) {
+		thread_control_val = OGL_THREAD_CONTROL_ENABLE;
+	}
+	ogl_thread_control_setting.u32CurrentValue = thread_control_val;
+
+	if (!nvapi_err_check("NVAPI: Error calling NvAPI_DRS_SetSetting", NvAPI_DRS_SetSetting(session_handle, profile_handle, &ogl_thread_control_setting))) {
+		NvAPI_DRS_DestroySession(session_handle);
+		NvAPI_Unload();
+		return;
+	}
+
+	NVDRS_SETTING vrr_mode_setting = {};
+	vrr_mode_setting.version = NVDRS_SETTING_VER;
+	vrr_mode_setting.settingId = VRR_MODE_ID;
+	vrr_mode_setting.settingType = NVDRS_DWORD_TYPE;
+	vrr_mode_setting.u32CurrentValue = VRR_MODE_FULLSCREEN_ONLY;
+
+	if (!nvapi_err_check("NVAPI: Error calling NvAPI_DRS_SetSetting", NvAPI_DRS_SetSetting(session_handle, profile_handle, &vrr_mode_setting))) {
+		NvAPI_DRS_DestroySession(session_handle);
+		NvAPI_Unload();
+		return;
+	}
+
+	if (!nvapi_err_check("NVAPI: Error saving settings", NvAPI_DRS_SaveSettings(session_handle))) {
+		NvAPI_DRS_DestroySession(session_handle);
+		NvAPI_Unload();
+		return;
+	}
+
+	if (thread_control_val == OGL_THREAD_CONTROL_DISABLE) {
+		print_verbose("NVAPI: Disabled OpenGL threaded optimization successfully");
+	} else {
+		print_verbose("NVAPI: Enabled OpenGL threaded optimization successfully");
+	}
+	print_verbose("NVAPI: Disabled G-SYNC for windowed mode successfully");
+
+	NvAPI_DRS_DestroySession(session_handle);
+}
+
 // Workaround mingw-w64 < 4.0 bug
 #ifndef WM_TOUCH
 #define WM_TOUCH 576
@@ -1705,7 +1919,7 @@
 	}
 
 	update_real_mouse_position();
-
+	_nvapi_disable_threaded_optimization();
 	return OK;
 }
 

platform/windows/os_windows.h

@@ -397,6 +397,7 @@ class OS_Windows : public OS {
 	void _update_window_mouse_passthrough();
 
 	void _set_mouse_mode_impl(MouseMode p_mode);
+	void _nvapi_disable_threaded_optimization();
 
 	// functions used by main to initialize/deinitialize the OS
 protected:

thirdparty/README.md

@@ -417,6 +417,15 @@ Files extracted from the upstream source:
 library.
 
 
+## nvapi
+
+- Upstream: http://download.nvidia.com/XFree86/nvapi-open-source-sdk
+- Version: R525
+- License: MIT
+
+- `nvapi_minimal.h` was created by using `nvapi.h` from upstream and removing unnecessary code.
+
+
 ## oidn
 
 - Upstream: https://github.com/OpenImageDenoise/oidn