Remove LSFG HDR + other optimizations on extra frame

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/model/BooleanSetting.kt

@@ -38,7 +38,6 @@ enum class BooleanSetting(override val key: String) : AbstractBooleanSetting {
     RENDERER_VERTEX_INPUT_DYNAMIC_STATE("vertex_input_dynamic_state"),
     RENDERER_SAMPLE_SHADING("sample_shading"),
     RENDERER_FRAME_GEN("frame_gen"),
-    RENDERER_FRAME_GEN_HDR("frame_gen_hdr"),
     RENDERER_FRAME_GEN_FP16("frame_gen_fp16"),
     RENDERER_FRAME_GEN_DUMP_FLOW("frame_gen_dump_flow"),
     GPU_UNSWIZZLE_ENABLED("gpu_unswizzle_enabled"),

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/model/view/SettingsItem.kt

@@ -633,13 +633,6 @@ abstract class SettingsItem(
                     units = "%"
                 )
             )
-            put(
-                SwitchSetting(
-                    BooleanSetting.RENDERER_FRAME_GEN_HDR,
-                    titleId = R.string.frame_gen_hdr,
-                    descriptionId = R.string.frame_gen_hdr_description
-                )
-            )
             put(
                 SwitchSetting(
                     BooleanSetting.RENDERER_FRAME_GEN_FP16,

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/ui/SettingsFragmentPresenter.kt

@@ -119,7 +119,6 @@ class SettingsFragmentPresenter(
             add(IntSetting.RENDERER_FRAME_GEN_MULTIPLIER.key)
             add(IntSetting.RENDERER_FRAME_GEN_FLOW_SCALE.key)
             add(BooleanSetting.RENDERER_FRAME_GEN_FP16.key)
-            add(BooleanSetting.RENDERER_FRAME_GEN_HDR.key)
             add(BooleanSetting.RENDERER_FRAME_GEN_DUMP_FLOW.key)
         }
     }

src/android/app/src/main/res/values/strings.xml

@@ -309,8 +309,6 @@
     <string name="frame_gen_flow_scale_description">Resolution of the optical flow pass, as a fraction of the output. Lowering it is the cheapest way to reclaim performance.</string>
     <string name="frame_gen_fp16">Half precision shaders</string>
     <string name="frame_gen_fp16_description">Use the 16-bit shader variant shipped in Lossless.dll. Much faster on Adreno. Falls back automatically if the driver or the file lacks it.</string>
-    <string name="frame_gen_hdr">HDR interpolation</string>
-    <string name="frame_gen_hdr_description">Treat frames as HDR when interpolating. Only useful on an HDR output.</string>
     <string name="frame_gen_dump_flow">Dump generated frame</string>
     <string name="frame_gen_dump_flow_description">Write the optical flow mip levels and the interpolated frame to the lossless/debug folder once, for troubleshooting</string>
     <string name="frame_gen_unsupported">Frame generation unavailable</string>

src/common/settings.h

@@ -414,9 +414,6 @@ struct Values {
                                                       true,
                                                       &frame_gen};
 
-    SwitchableSetting<bool> frame_gen_hdr{linkage,      false, "frame_gen_hdr", Category::Renderer,
-                                          Specialization::Default, true, true, &frame_gen};
-
     SwitchableSetting<bool> frame_gen_fp16{linkage,      true,  "frame_gen_fp16", Category::Renderer,
                                            Specialization::Default, true, true, &frame_gen};
 

src/video_core/renderer_vulkan/present/frame_gen.cpp

@@ -182,7 +182,9 @@ FrameGen::FrameGen(MemoryAllocator& memory_allocator_, Scheduler& scheduler_)
 
 FrameGen::~FrameGen() = default;
 
-void FrameGen::Process(const Device& device, Frame* frame, VkFormat format) {
+void FrameGen::Process(const Device& device, Frame* frame, VkFormat format, bool generate) {
+    generated = false;
+
     if (unavailable || !Settings::values.frame_gen.GetValue()) {
         return;
     }
@@ -198,20 +200,23 @@ void FrameGen::Process(const Device& device, Frame* frame, VkFormat format) {
     const VkExtent2D extent{.width = frame->width, .height = frame->height};
     if (!chain || built_extent.width != extent.width || built_extent.height != extent.height ||
         built_format != format || built_flow_scale != ConfiguredFlowScale() ||
-        built_hdr != Settings::values.frame_gen_hdr.GetValue() ||
         built_generations != ConfiguredGenerations()) {
         Rebuild(device, extent, format);
     }
 
     const u64 count = frame_count++;
+    generated = generate && count + 1 >= LSFG_REQUIRED_FRAMES;
 
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([this, source = *frame->image, extent, count](vk::CommandBuffer cmdbuf) {
+    scheduler.Record([this, source = *frame->image, extent, count,
+                      dispatch = generated](vk::CommandBuffer cmdbuf) {
         CopyPresentedFrame(cmdbuf, source, chain->Input(count), extent);
-        chain->Dispatch(cmdbuf, count);
+        if (dispatch) {
+            chain->Dispatch(cmdbuf, count);
+        }
     });
 
-    const bool dump_requested = Settings::values.frame_gen_dump_flow.GetValue();
+    const bool dump_requested = generated && Settings::values.frame_gen_dump_flow.GetValue();
     if (!dump_requested) {
         dumped = false;
     } else if (!dumped) {
@@ -220,12 +225,15 @@ void FrameGen::Process(const Device& device, Frame* frame, VkFormat format) {
     }
 }
 
-size_t FrameGen::GeneratedFrameCount() const {
-    if (!chain || frame_count < LSFG_REQUIRED_FRAMES ||
-        !Settings::values.frame_gen.GetValue()) {
+size_t FrameGen::WantedGenerations() const {
+    if (unavailable || !Settings::values.frame_gen.GetValue()) {
         return 0;
     }
-    return chain->GenerationCount();
+    return ConfiguredGenerations();
+}
+
+size_t FrameGen::GeneratedFrameCount() const {
+    return generated && chain ? chain->GenerationCount() : 0;
 }
 
 void FrameGen::CopyToFrame(Frame* destination, size_t generation) {
@@ -273,14 +281,14 @@ void FrameGen::Rebuild(const Device& device, VkExtent2D extent, VkFormat format)
     chain.reset();
 
     built_flow_scale = ConfiguredFlowScale();
-    built_hdr = Settings::values.frame_gen_hdr.GetValue();
     built_generations = ConfiguredGenerations();
 
-    chain.emplace(device, memory_allocator, *shaders, extent, format, built_flow_scale, built_hdr,
+    chain.emplace(device, memory_allocator, *shaders, extent, format, built_flow_scale,
                   built_generations);
     built_extent = extent;
     built_format = format;
     frame_count = 0;
+    generated = false;
 }
 
 void FrameGen::DumpDebugImages(u64 count) {

src/video_core/renderer_vulkan/present/frame_gen.h

@@ -21,7 +21,9 @@ public:
     explicit FrameGen(MemoryAllocator& memory_allocator, Scheduler& scheduler);
     ~FrameGen();
 
-    void Process(const Device& device, Frame* frame, VkFormat format);
+    void Process(const Device& device, Frame* frame, VkFormat format, bool generate);
+
+    [[nodiscard]] size_t WantedGenerations() const;
 
     [[nodiscard]] size_t GeneratedFrameCount() const;
 
@@ -39,9 +41,9 @@ private:
     VkExtent2D built_extent{};
     VkFormat built_format{VK_FORMAT_UNDEFINED};
     f32 built_flow_scale{};
-    bool built_hdr{};
     size_t built_generations{};
     u64 frame_count{};
+    bool generated{};
     bool unavailable{};
     bool dumped{};
 };

src/video_core/renderer_vulkan/present/lsfg_chain.cpp

@@ -19,9 +19,9 @@ constexpr size_t FIRST_DELTA_LEVEL = 4;
 
 LsfgChain::LsfgChain(const Device& device, MemoryAllocator& memory_allocator,
                      const LsfgShaders& shaders, VkExtent2D extent, VkFormat format,
-                     f32 flow_scale, bool is_hdr, size_t generation_count_)
+                     f32 flow_scale, size_t generation_count_)
     : generation_count{generation_count_},
-      resources{device, memory_allocator, flow_scale, is_hdr},
+      resources{device, memory_allocator, flow_scale},
       descriptor_pool{CreateLsfgDescriptorPool(
           device, FIXED_DESCRIPTOR_SETS +
                       DESCRIPTOR_SETS_PER_GENERATION * static_cast<u32>(generation_count))} {

src/video_core/renderer_vulkan/present/lsfg_chain.h

@@ -24,8 +24,7 @@ constexpr size_t LSFG_DELTA_INSTANCES = 3;
 class LsfgChain {
 public:
     LsfgChain(const Device& device, MemoryAllocator& memory_allocator, const LsfgShaders& shaders,
-              VkExtent2D extent, VkFormat format, f32 flow_scale, bool is_hdr,
-              size_t generation_count_);
+              VkExtent2D extent, VkFormat format, f32 flow_scale, size_t generation_count_);
 
     LsfgChain(const LsfgChain&) = delete;
     LsfgChain& operator=(const LsfgChain&) = delete;

src/video_core/renderer_vulkan/present/lsfg_common.cpp

@@ -154,8 +154,8 @@ VkBuffer LsfgResources::GetBuffer(f32 timestamp, bool first_iter, bool first_ite
         .input_offset = {0, 0},
         .first_iter = first_iter ? 1u : 0u,
         .first_iter_s = first_iter_s ? 1u : 0u,
-        .advanced_color_kind = is_hdr ? 2u : 0u,
-        .hdr_support = is_hdr ? 1u : 0u,
+        .advanced_color_kind = 0,
+        .hdr_support = 0,
         .resolution_inv_scale = 1.0f / flow_scale,
         .timestamp = timestamp,
         .ui_threshold = 0.5f,

src/video_core/renderer_vulkan/present/lsfg_common.h

@@ -76,10 +76,8 @@ using LsfgImageHistory = std::array<LsfgImagePair, LSFG_HISTORY_SLOTS>;
 class LsfgResources {
 public:
     LsfgResources() = default;
-    LsfgResources(const Device& device_, MemoryAllocator& memory_allocator_, f32 flow_scale_,
-                  bool is_hdr_)
-        : device{&device_}, memory_allocator{&memory_allocator_}, flow_scale{flow_scale_},
-          is_hdr{is_hdr_} {}
+    LsfgResources(const Device& device_, MemoryAllocator& memory_allocator_, f32 flow_scale_)
+        : device{&device_}, memory_allocator{&memory_allocator_}, flow_scale{flow_scale_} {}
 
     [[nodiscard]] VkSampler GetSampler(
         VkSamplerAddressMode address_mode = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER,
@@ -94,7 +92,6 @@ private:
     const Device* device{};
     MemoryAllocator* memory_allocator{};
     f32 flow_scale{1.0f};
-    bool is_hdr{};
 
     std::map<u64, vk::Sampler> samplers;
     std::map<u64, vk::Buffer> buffers;

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -193,14 +193,13 @@ void RendererVulkan::Composite(std::span<const Tegra::FramebufferConfig> framebu
                                render_window.GetFramebufferLayout(), swapchain.GetImageCount(),
                                swapchain.GetImageViewFormat());
 
-    frame_gen.Process(device, frame, swapchain.GetImageFormat());
+    const size_t wanted = frame_gen.WantedGenerations();
+    const bool can_present_all = wanted > 0 && present_manager.AvailableExtraFrames() >= wanted;
+
+    frame_gen.Process(device, frame, swapchain.GetImageFormat(), can_present_all);
 
     const size_t generated_frames = frame_gen.GeneratedFrameCount();
     for (size_t generation = 0; generation < generated_frames; ++generation) {
-        if (!present_manager.CanQueueExtraFrame()) {
-            break;
-        }
-
         Frame* generated = present_manager.GetRenderFrame();
         blit_swapchain.PrepareFrame(device, generated, render_window.GetFramebufferLayout());
         frame_gen.CopyToFrame(generated, generation);

src/video_core/renderer_vulkan/vk_present_manager.cpp

@@ -181,9 +181,9 @@ void PresentManager::Present(Frame* frame) {
     }
 }
 
-bool PresentManager::CanQueueExtraFrame() {
+size_t PresentManager::AvailableExtraFrames() {
     std::scoped_lock lock{free_mutex};
-    return !free_queue.empty();
+    return free_queue.size();
 }
 
 void PresentManager::RecreateFrame(Frame* frame, u32 width, u32 height, VkFormat image_view_format,

src/video_core/renderer_vulkan/vk_present_manager.h

@@ -60,8 +60,8 @@ public:
     /// Waits for the present thread to finish presenting all queued frames.
     void WaitPresent();
 
-    /// Whether an additional frame can be queued without stalling the render thread
-    bool CanQueueExtraFrame();
+    /// How many additional frames can be queued without stalling the render thread
+    size_t AvailableExtraFrames();
 
 private:
     void PresentThread(std::stop_token token);