early-access version 3605

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 3604.
+This is the source code for early-access 3605.
 
 ## Legal Notice
 

src/audio_core/renderer/adsp/audio_renderer.cpp

@@ -154,6 +154,11 @@ void AudioRenderer::ThreadFunc() {
             return;
 
         case RenderMessage::AudioRenderer_Render: {
+            if (system.IsShuttingDown()) [[unlikely]] {
+                std::this_thread::sleep_for(std::chrono::milliseconds(5));
+                mailbox->ADSPSendMessage(RenderMessage::AudioRenderer_RenderResponse);
+                continue;
+            }
             std::array<bool, MaxRendererSessions> buffers_reset{};
             std::array<u64, MaxRendererSessions> render_times_taken{};
             const auto start_time{system.CoreTiming().GetClockTicks()};

src/audio_core/renderer/system_manager.cpp

@@ -27,7 +27,7 @@ bool SystemManager::InitializeUnsafe() {
     if (!active) {
         if (adsp.Start()) {
             active = true;
-            thread = std::jthread([this](std::stop_token stop_token) { ThreadFunc(); });
+            thread = std::jthread([this](std::stop_token stop_token) { ThreadFunc(stop_token); });
         }
     }
 
@@ -39,8 +39,7 @@ void SystemManager::Stop() {
         return;
     }
     active = false;
-    update.store(true);
-    update.notify_all();
+    thread.request_stop();
     thread.join();
     adsp.Stop();
 }
@@ -85,12 +84,12 @@ bool SystemManager::Remove(System& system_) {
     return true;
 }
 
-void SystemManager::ThreadFunc() {
+void SystemManager::ThreadFunc(std::stop_token stop_token) {
     static constexpr char name[]{"AudioRenderSystemManager"};
     MicroProfileOnThreadCreate(name);
     Common::SetCurrentThreadName(name);
     Common::SetCurrentThreadPriority(Common::ThreadPriority::High);
-    while (active) {
+    while (active && !stop_token.stop_requested()) {
         {
             std::scoped_lock l{mutex1};
 

src/audio_core/renderer/system_manager.h

@@ -66,13 +66,7 @@ private:
     /**
      * Main thread responsible for command generation.
      */
-    void ThreadFunc();
-
-    enum class StreamState {
-        Filling,
-        Steady,
-        Draining,
-    };
+    void ThreadFunc(std::stop_token stop_token);
 
     /// Core system
     Core::System& core;
@@ -90,8 +84,6 @@ private:
     ADSP::ADSP& adsp;
     /// AudioRenderer mailbox for communication
     ADSP::AudioRenderer_Mailbox* mailbox{};
-    /// Atomic for main thread to wait on
-    std::atomic<bool> update{};
 };
 
 } // namespace AudioCore::AudioRenderer

src/audio_core/sink/sink_stream.cpp

@@ -270,8 +270,8 @@ u64 SinkStream::GetExpectedPlayedSampleCount() {
 
 void SinkStream::WaitFreeSpace() {
     std::unique_lock lk{release_mutex};
-    release_cv.wait(
-        lk, [this]() { return queued_buffers < max_queue_size || system.IsShuttingDown(); });
+    release_cv.wait_for(lk, std::chrono::milliseconds(5),
+                        [this]() { return queued_buffers < max_queue_size; });
 }
 
 } // namespace AudioCore::Sink

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -144,7 +144,7 @@ public:
         return state_tracker;
     }
 
-    void CheckFeedbackLoop(ImageView& image_view) const noexcept {
+    void BarrierFeedbackLoop() const noexcept {
         // OpenGL does not require a barrier for attachment feedback loops.
     }
 

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -175,7 +175,6 @@ public:
     std::array<f32, 4> words{};
 };
 
-template <bool CheckFeedbackLoop = false>
 inline void PushImageDescriptors(TextureCache& texture_cache,
                                  GuestDescriptorQueue& guest_descriptor_queue,
                                  const Shader::Info& info, RescalingPushConstant& rescaling,
@@ -193,9 +192,6 @@ inline void PushImageDescriptors(TextureCache& texture_cache,
             const VkImageView vk_image_view{image_view.Handle(desc.type)};
             guest_descriptor_queue.AddSampledImage(vk_image_view, sampler);
             rescaling.PushTexture(texture_cache.IsRescaling(image_view));
-            if constexpr (CheckFeedbackLoop) {
-                texture_cache.CheckFeedbackLoop(image_view);
-            }
         }
     }
     for (const auto& desc : info.image_descriptors) {
@@ -207,9 +203,6 @@ inline void PushImageDescriptors(TextureCache& texture_cache,
             const VkImageView vk_image_view{image_view.StorageView(desc.type, desc.format)};
             guest_descriptor_queue.AddImage(vk_image_view);
             rescaling.PushImage(texture_cache.IsRescaling(image_view));
-            if constexpr (CheckFeedbackLoop) {
-                texture_cache.CheckFeedbackLoop(image_view);
-            }
         }
     }
 }

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -457,8 +457,8 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     const VideoCommon::ImageViewInOut* views_it{views.data()};
     const auto prepare_stage{[&](size_t stage) LAMBDA_FORCEINLINE {
         buffer_cache.BindHostStageBuffers(stage);
-        PushImageDescriptors<true>(texture_cache, guest_descriptor_queue, stage_infos[stage],
-                                   rescaling, samplers_it, views_it);
+        PushImageDescriptors(texture_cache, guest_descriptor_queue, stage_infos[stage], rescaling,
+                             samplers_it, views_it);
         const auto& info{stage_infos[0]};
         if (info.uses_render_area) {
             render_area.uses_render_area = true;
@@ -481,12 +481,13 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     if constexpr (Spec::enabled_stages[4]) {
         prepare_stage(4);
     }
+    texture_cache.UpdateRenderTargets(false);
+    texture_cache.CheckFeedbackLoop(views);
     ConfigureDraw(rescaling, render_area);
 }
 
 void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
                                      const RenderAreaPushConstant& render_area) {
-    texture_cache.UpdateRenderTargets(false);
     scheduler.RequestRenderpass(texture_cache.GetFramebuffer());
 
     if (!is_built.load(std::memory_order::relaxed)) {

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -106,15 +106,6 @@ public:
         return *master_semaphore;
     }
 
-    [[nodiscard]] bool IsRenderpassImage(VkImage image) const noexcept {
-        for (u32 i = 0; i < num_renderpass_images; i++) {
-            if (image == renderpass_images[i]) {
-                return true;
-            }
-        }
-        return false;
-    }
-
     std::mutex submit_mutex;
 
 private:

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -861,10 +861,8 @@ VkBuffer TextureCacheRuntime::GetTemporaryBuffer(size_t needed_size) {
     return *buffers[level];
 }
 
-void TextureCacheRuntime::CheckFeedbackLoop(ImageView& image_view) {
-    if (scheduler.IsRenderpassImage(image_view.ImageHandle())) {
-        scheduler.RequestOutsideRenderPassOperationContext();
-    }
+void TextureCacheRuntime::BarrierFeedbackLoop() {
+    scheduler.RequestOutsideRenderPassOperationContext();
 }
 
 void TextureCacheRuntime::ReinterpretImage(Image& dst, Image& src,

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -103,7 +103,7 @@ public:
 
     [[nodiscard]] VkBuffer GetTemporaryBuffer(size_t needed_size);
 
-    void CheckFeedbackLoop(ImageView& image_view);
+    void BarrierFeedbackLoop();
 
     const Device& device;
     Scheduler& scheduler;

src/video_core/texture_cache/texture_cache.h

@@ -182,6 +182,42 @@ void TextureCache<P>::FillComputeImageViews(std::span<ImageViewInOut> views) {
                          views);
 }
 
+template <class P>
+void TextureCache<P>::CheckFeedbackLoop(std::span<const ImageViewInOut> views) {
+    const bool requires_barrier = [&] {
+        for (const auto& view : views) {
+            if (!view.id) {
+                continue;
+            }
+            auto& image_view = slot_image_views[view.id];
+
+            // Check color targets
+            for (const auto& ct_view_id : render_targets.color_buffer_ids) {
+                if (ct_view_id) {
+                    auto& ct_view = slot_image_views[ct_view_id];
+                    if (image_view.image_id == ct_view.image_id) {
+                        return true;
+                    }
+                }
+            }
+
+            // Check zeta target
+            if (render_targets.depth_buffer_id) {
+                auto& zt_view = slot_image_views[render_targets.depth_buffer_id];
+                if (image_view.image_id == zt_view.image_id) {
+                    return true;
+                }
+            }
+        }
+
+        return false;
+    }();
+
+    if (requires_barrier) {
+        runtime.BarrierFeedbackLoop();
+    }
+}
+
 template <class P>
 typename P::Sampler* TextureCache<P>::GetGraphicsSampler(u32 index) {
     if (index > channel_state->graphics_sampler_table.Limit()) {
@@ -2374,11 +2410,6 @@ bool TextureCache<P>::IsFullClear(ImageViewId id) {
            scissor.max_y >= size.height;
 }
 
-template <class P>
-void TextureCache<P>::CheckFeedbackLoop(ImageView& image_view) {
-    runtime.CheckFeedbackLoop(image_view);
-}
-
 template <class P>
 void TextureCache<P>::CreateChannel(struct Tegra::Control::ChannelState& channel) {
     VideoCommon::ChannelSetupCaches<TextureCacheChannelInfo>::CreateChannel(channel);

src/video_core/texture_cache/texture_cache_base.h

@@ -148,6 +148,9 @@ public:
     /// Fill image_view_ids with the compute images in indices
     void FillComputeImageViews(std::span<ImageViewInOut> views);
 
+    /// Handle feedback loops during draws.
+    void CheckFeedbackLoop(std::span<const ImageViewInOut> views);
+
     /// Get the sampler from the graphics descriptor table in the specified index
     Sampler* GetGraphicsSampler(u32 index);
 
@@ -224,9 +227,6 @@ public:
 
     [[nodiscard]] bool IsRescaling(const ImageViewBase& image_view) const noexcept;
 
-    /// Handle feedback loops during draws.
-    void CheckFeedbackLoop(ImageView& image_view);
-
     /// Create channel state.
     void CreateChannel(Tegra::Control::ChannelState& channel) final override;
 

src/video_core/textures/bcn.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <stb_dxt.h>
+#include <string.h>
 
 #include "common/alignment.h"
 #include "video_core/textures/bcn.h"