early-access version 3795

src/video_core/renderer_opengl/gl_device.cpp

@@ -106,6 +106,43 @@ bool IsASTCSupported() {
     return true;
 }
 
+static bool HasSlowSoftwareAstc(std::string_view vendor_name, std::string_view renderer) {
+// ifdef for Unix reduces string comparisons for non-Windows drivers, and Intel
+#ifdef YUZU_UNIX
+    // Sorted vaguely by how likely a vendor is to appear
+    if (vendor_name == "AMD") {
+        // RadeonSI
+        return true;
+    }
+    if (vendor_name == "Intel") {
+        // Must be inside YUZU_UNIX ifdef as the Windows driver uses the same vendor string
+        // iris, crocus
+        const bool is_intel_dg = (renderer.find("DG") != std::string_view::npos);
+        return is_intel_dg;
+    }
+    if (vendor_name == "nouveau") {
+        return true;
+    }
+    if (vendor_name == "X.Org") {
+        // R600
+        return true;
+    }
+#endif
+    if (vendor_name == "Collabora Ltd") {
+        // Zink
+        return true;
+    }
+    if (vendor_name == "Microsoft Corporation") {
+        // d3d12
+        return true;
+    }
+    if (vendor_name == "Mesa/X.org") {
+        // llvmpipe, softpipe, virgl
+        return true;
+    }
+    return false;
+}
+
 [[nodiscard]] bool IsDebugToolAttached(std::span<const std::string_view> extensions) {
     const bool nsight = std::getenv("NVTX_INJECTION64_PATH") || std::getenv("NSIGHT_LAUNCHED");
     return nsight || HasExtension(extensions, "GL_EXT_debug_tool") ||
@@ -120,12 +157,16 @@ Device::Device(Core::Frontend::EmuWindow& emu_window) {
     }
     vendor_name = reinterpret_cast<const char*>(glGetString(GL_VENDOR));
     const std::string_view version = reinterpret_cast<const char*>(glGetString(GL_VERSION));
+    const std::string_view renderer = reinterpret_cast<const char*>(glGetString(GL_RENDERER));
     const std::vector extensions = GetExtensions();
 
     const bool is_nvidia = vendor_name == "NVIDIA Corporation";
     const bool is_amd = vendor_name == "ATI Technologies Inc.";
     const bool is_intel = vendor_name == "Intel";
 
+    const bool has_slow_software_astc =
+        !is_nvidia && !is_amd && HasSlowSoftwareAstc(vendor_name, renderer);
+
 #ifdef __unix__
     constexpr bool is_linux = true;
 #else
@@ -152,7 +193,7 @@ Device::Device(Core::Frontend::EmuWindow& emu_window) {
     has_vertex_viewport_layer = GLAD_GL_ARB_shader_viewport_layer_array;
     has_image_load_formatted = HasExtension(extensions, "GL_EXT_shader_image_load_formatted");
     has_texture_shadow_lod = HasExtension(extensions, "GL_EXT_texture_shadow_lod");
-    has_astc = IsASTCSupported();
+    has_astc = !has_slow_software_astc && IsASTCSupported();
     has_variable_aoffi = TestVariableAoffi();
     has_component_indexing_bug = is_amd;
     has_precise_bug = TestPreciseBug();

src/video_core/renderer_opengl/gl_query_cache.cpp

@@ -27,7 +27,7 @@ constexpr GLenum GetTarget(VideoCore::QueryType type) {
 } // Anonymous namespace
 
 QueryCache::QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_)
-    : QueryCacheBase(rasterizer_, cpu_memory_), gl_rasterizer{rasterizer_} {}
+    : QueryCacheLegacy(rasterizer_, cpu_memory_), gl_rasterizer{rasterizer_} {}
 
 QueryCache::~QueryCache() = default;
 

src/video_core/renderer_opengl/gl_query_cache.h

@@ -26,7 +26,7 @@ class RasterizerOpenGL;
 using CounterStream = VideoCommon::CounterStreamBase<QueryCache, HostCounter>;
 
 class QueryCache final
-    : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
+    : public VideoCommon::QueryCacheLegacy<QueryCache, CachedQuery, CounterStream, HostCounter> {
 public:
     explicit QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_);
     ~QueryCache();

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -385,13 +385,39 @@ void RasterizerOpenGL::DispatchCompute() {
     has_written_global_memory |= pipeline->WritesGlobalMemory();
 }
 
-void RasterizerOpenGL::ResetCounter(VideoCore::QueryType type) {
-    query_cache.ResetCounter(type);
+void RasterizerOpenGL::ResetCounter(VideoCommon::QueryType type) {
+    if (type == VideoCommon::QueryType::ZPassPixelCount64) {
+        query_cache.ResetCounter(VideoCore::QueryType::SamplesPassed);
+    }
 }
 
-void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCore::QueryType type,
-                             std::optional<u64> timestamp) {
-    query_cache.Query(gpu_addr, type, timestamp);
+void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+                             VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
+    if (type == VideoCommon::QueryType::ZPassPixelCount64) {
+        if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
+            query_cache.Query(gpu_addr, VideoCore::QueryType::SamplesPassed, {gpu.GetTicks()});
+        } else {
+            query_cache.Query(gpu_addr, VideoCore::QueryType::SamplesPassed, std::nullopt);
+        }
+        return;
+    }
+    if (type != VideoCommon::QueryType::Payload) {
+        payload = 1u;
+    }
+    std::function<void()> func([this, gpu_addr, flags, memory_manager = gpu_memory, payload]() {
+        if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
+            u64 ticks = gpu.GetTicks();
+            memory_manager->Write<u64>(gpu_addr + 8, ticks);
+            memory_manager->Write<u64>(gpu_addr, static_cast<u64>(payload));
+        } else {
+            memory_manager->Write<u32>(gpu_addr, payload);
+        }
+    });
+    if (True(flags & VideoCommon::QueryPropertiesFlags::IsAFence)) {
+        SignalFence(std::move(func));
+        return;
+    }
+    func();
 }
 
 void RasterizerOpenGL::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
@@ -562,8 +588,8 @@ void RasterizerOpenGL::SignalReference() {
     fence_manager.SignalOrdering();
 }
 
-void RasterizerOpenGL::ReleaseFences() {
-    fence_manager.WaitPendingFences();
+void RasterizerOpenGL::ReleaseFences(bool force) {
+    fence_manager.WaitPendingFences(force);
 }
 
 void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size,

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -86,8 +86,9 @@ public:
     void DrawTexture() override;
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
-    void ResetCounter(VideoCore::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
+    void ResetCounter(VideoCommon::QueryType type) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
@@ -107,7 +108,7 @@ public:
     void SyncOperation(std::function<void()>&& func) override;
     void SignalSyncPoint(u32 value) override;
     void SignalReference() override;
-    void ReleaseFences() override;
+    void ReleaseFences(bool force = true) override;
     void FlushAndInvalidateRegion(
         VAddr addr, u64 size, VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void WaitForIdle() override;