feat(fp16): add flag for overriding fp16

framegen/include/core/device.hpp

@@ -21,10 +21,11 @@ namespace LSFG::Core {
         ///
         /// @param instance Vulkan instance
         /// @param deviceUUID The UUID of the Vulkan device to use.
+        /// @param forceDisableFp16 Force-disable FP16 shaders.
         ///
         /// @throws LSFG::vulkan_error if object creation fails.
         ///
-        Device(const Instance& instance, uint64_t deviceUUID);
+        Device(const Instance& instance, uint64_t deviceUUID, bool forceDisableFp16);
 
         /// Get the Vulkan handle.
         [[nodiscard]] auto handle() const { return *this->device; }

framegen/public/lsfg_3_1.hpp

@@ -16,6 +16,7 @@ namespace LSFG_3_1 {
     /// @param isHdr Whether the images are in HDR format.
     /// @param flowScale Internal flow scale factor.
     /// @param generationCount Number of frames to generate.
+    /// @param forceDisableFp16 Whether to force-disable FP16 optimizations.
     /// @param loader Function to load shader source code by name.
     ///
     /// @throws LSFG::vulkan_error if Vulkan objects fail to initialize.
@@ -23,6 +24,7 @@ namespace LSFG_3_1 {
     [[gnu::visibility("default")]]
     void initialize(uint64_t deviceUUID,
         bool isHdr, float flowScale, uint64_t generationCount,
+        bool forceDisableFp16,
         const std::function<std::vector<uint8_t>(const std::string&, bool)>& loader);
 
     ///

framegen/public/lsfg_3_1p.hpp

@@ -16,6 +16,7 @@ namespace LSFG_3_1P {
     /// @param isHdr Whether the images are in HDR format.
     /// @param flowScale Internal flow scale factor.
     /// @param generationCount Number of frames to generate.
+    /// @param forceDisableFp16 Whether to force-disable FP16 optimizations.
     /// @param loader Function to load shader source code by name.
     ///
     /// @throws LSFG::vulkan_error if Vulkan objects fail to initialize.
@@ -23,6 +24,7 @@ namespace LSFG_3_1P {
     [[gnu::visibility("default")]]
     void initialize(uint64_t deviceUUID,
         bool isHdr, float flowScale, uint64_t generationCount,
+        bool forceDisableFp16,
         const std::function<std::vector<uint8_t>(const std::string&, bool)>& loader);
 
     ///

framegen/src/core/device.cpp

@@ -19,7 +19,7 @@ const std::vector<const char*> requiredExtensions = {
     "VK_EXT_robustness2"
 };
 
-Device::Device(const Instance& instance, uint64_t deviceUUID) {
+Device::Device(const Instance& instance, uint64_t deviceUUID, bool forceDisableFp16) {
     // get all physical devices
     uint32_t deviceCount{};
     auto res = vkEnumeratePhysicalDevices(instance.handle(), &deviceCount, nullptr);
@@ -72,10 +72,10 @@ Device::Device(const Instance& instance, uint64_t deviceUUID) {
         .pNext = &supported12Features
     };
     vkGetPhysicalDeviceFeatures2(*physicalDevice, &supportedFeatures);
-    this->supportsFP16 = supported12Features.shaderFloat16;
+    this->supportsFP16 = !forceDisableFp16 && supported12Features.shaderFloat16;
     if (this->supportsFP16)
         std::cerr << "lsfg-vk: Using FP16 acceleration" << '\n';
-    else
+    else if (!forceDisableFp16)
         std::cerr << "lsfg-vk: FP16 acceleration not supported, using FP32" << '\n';
 
     // create logical device

framegen/v3.1_src/lsfg.cpp

@@ -35,13 +35,14 @@ namespace {
 
 void LSFG_3_1::initialize(uint64_t deviceUUID,
         bool isHdr, float flowScale, uint64_t generationCount,
+        bool forceDisableFp16,
         const std::function<std::vector<uint8_t>(const std::string&, bool)>& loader) {
     if (instance.has_value() || device.has_value())
         return;
 
     instance.emplace();
     device.emplace(Vulkan {
-        .device{*instance, deviceUUID},
+        .device{*instance, deviceUUID, forceDisableFp16},
         .generationCount = generationCount,
         .flowScale = flowScale,
         .isHdr = isHdr

framegen/v3.1p_src/lsfg.cpp

@@ -35,13 +35,14 @@ namespace {
 
 void LSFG_3_1P::initialize(uint64_t deviceUUID,
         bool isHdr, float flowScale, uint64_t generationCount,
+        bool forceDisableFp16,
         const std::function<std::vector<uint8_t>(const std::string&, bool)>& loader) {
     if (instance.has_value() || device.has_value())
         return;
 
     instance.emplace();
     device.emplace(Vulkan {
-        .device{*instance, deviceUUID},
+        .device{*instance, deviceUUID, forceDisableFp16},
         .generationCount = generationCount,
         .flowScale = flowScale,
         .isHdr = isHdr

include/config/config.hpp

@@ -15,6 +15,8 @@ namespace Config {
         bool enable{false};
         /// Path to Lossless.dll.
         std::string dll;
+        /// Whether FP16 is force-disabled
+        bool no_fp16{false};
 
         /// The frame generation muliplier
         size_t multiplier{2};

src/config/config.cpp

@@ -74,6 +74,7 @@ void Config::updateConfig(const std::string& file) {
     const toml::value globalTable = toml::find_or_default<toml::table>(toml, "global");
     const Configuration global{
         .dll =     toml::find_or(globalTable, "dll", std::string()),
+        .no_fp16 = toml::find_or(globalTable, "no_fp16", false),
         .config_file = file,
         .timestamp = std::filesystem::last_write_time(file)
     };
@@ -97,6 +98,7 @@ void Config::updateConfig(const std::string& file) {
         Configuration game{
             .enable = true,
             .dll = global.dll,
+            .no_fp16 = global.no_fp16,
             .multiplier = toml::find_or(gameTable, "multiplier", 2U),
             .flowScale = toml::find_or(gameTable, "flow_scale", 1.0F),
             .performance = toml::find_or(gameTable, "performance_mode", false),

src/context.cpp

@@ -53,6 +53,7 @@ LsContext::LsContext(const Hooks::DeviceInfo& info, VkSwapchainKHR swapchain,
         // print config
         std::cerr << "lsfg-vk: Reloaded configuration for " << name.second << ":\n";
         if (!conf.dll.empty()) std::cerr << "  Using DLL from: " << conf.dll << '\n';
+        if (conf.no_fp16) std::cerr << "  FP16 Acceleration: Force-disabled\n";
         std::cerr << "  Multiplier: " << conf.multiplier << '\n';
         std::cerr << "  Flow Scale: " << conf.flowScale << '\n';
         std::cerr << "  Performance Mode: " << (conf.performance ? "Enabled" : "Disabled") << '\n';
@@ -98,6 +99,7 @@ LsContext::LsContext(const Hooks::DeviceInfo& info, VkSwapchainKHR swapchain,
     lsfgInitialize(
         Utils::getDeviceUUID(info.physicalDevice),
         conf.hdr, 1.0F / conf.flowScale, conf.multiplier - 1,
+        conf.no_fp16,
         Extract::getShader
     );
 

src/main.cpp

@@ -47,6 +47,7 @@ namespace {
         // print config
         std::cerr << "lsfg-vk: Loaded configuration for " << name.second << ":\n";
         if (!conf.dll.empty()) std::cerr << "  Using DLL from: " << conf.dll << '\n';
+        if (conf.no_fp16) std::cerr << "  FP16 Acceleration: Force-disabled\n";
         std::cerr << "  Multiplier: " << conf.multiplier << '\n';
         std::cerr << "  Flow Scale: " << conf.flowScale << '\n';
         std::cerr << "  Performance Mode: " << (conf.performance ? "Enabled" : "Disabled") << '\n';

src/utils/benchmark.cpp

@@ -41,6 +41,7 @@ void Benchmark::run(uint32_t width, uint32_t height) {
     lsfgInitialize(
         deviceUUID, // some magic number if not given
         conf.hdr, 1.0F / conf.flowScale, conf.multiplier - 1,
+        conf.no_fp16,
         Extract::getShader
     );
     const int32_t ctx = lsfgCreateContext(-1, -1, {},