Merge pull request #2413 from FernandoS27/opt-gpu

Rasterizer Cache: refactor flushing & optimize memory usage of surfaces

src/video_core/rasterizer_cache.h

@@ -37,9 +37,6 @@ public:
     /// Gets the size of the shader in guest memory, required for cache management
     virtual std::size_t GetSizeInBytes() const = 0;
 
-    /// Wriets any cached resources back to memory
-    virtual void Flush() = 0;
-
     /// Sets whether the cached object should be considered registered
     void SetIsRegistered(bool registered) {
         is_registered = registered;
@@ -158,6 +155,8 @@ protected:
         return ++modified_ticks;
     }
 
+    virtual void FlushObjectInner(const T& object) = 0;
+
     /// Flushes the specified object, updating appropriate cache state as needed
     void FlushObject(const T& object) {
         std::lock_guard lock{mutex};
@@ -165,7 +164,7 @@ protected:
         if (!object->IsDirty()) {
             return;
         }
-        object->Flush();
+        FlushObjectInner(object);
         object->MarkAsModified(false, *this);
     }
 

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -42,9 +42,6 @@ public:
         return alignment;
     }
 
-    // We do not have to flush this cache as things in it are never modified by us.
-    void Flush() override {}
-
 private:
     VAddr cpu_addr{};
     std::size_t size{};
@@ -75,6 +72,9 @@ public:
 protected:
     void AlignBuffer(std::size_t alignment);
 
+    // We do not have to flush this cache as things in it are never modified by us.
+    void FlushObjectInner(const std::shared_ptr<CachedBufferEntry>& object) override {}
+
 private:
     OGLStreamBuffer stream_buffer;
 

src/video_core/renderer_opengl/gl_global_cache.h

@@ -46,7 +46,7 @@ public:
     /// Reloads the global region from guest memory
     void Reload(u32 size_);
 
-    void Flush() override;
+    void Flush();
 
 private:
     VAddr cpu_addr{};
@@ -65,6 +65,11 @@ public:
     GlobalRegion GetGlobalRegion(const GLShader::GlobalMemoryEntry& descriptor,
                                  Tegra::Engines::Maxwell3D::Regs::ShaderStage stage);
 
+protected:
+    void FlushObjectInner(const GlobalRegion& object) override {
+        object->Flush();
+    }
+
 private:
     GlobalRegion TryGetReservedGlobalRegion(CacheAddr addr, u32 size) const;
     GlobalRegion GetUncachedGlobalRegion(GPUVAddr addr, u8* host_ptr, u32 size);

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -628,8 +628,10 @@ CachedSurface::CachedSurface(const SurfaceParams& params)
 }
 
 MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 192, 64));
-void CachedSurface::LoadGLBuffer() {
+void CachedSurface::LoadGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem) {
     MICROPROFILE_SCOPE(OpenGL_SurfaceLoad);
+    auto& gl_buffer = res_cache_tmp_mem.gl_buffer;
+    if (gl_buffer.size() < params.max_mip_level)
         gl_buffer.resize(params.max_mip_level);
     for (u32 i = 0; i < params.max_mip_level; i++)
         gl_buffer[i].resize(params.GetMipmapSizeGL(i));
@@ -671,13 +673,13 @@ void CachedSurface::LoadGLBuffer() {
 }
 
 MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64));
-void CachedSurface::FlushGLBuffer() {
+void CachedSurface::FlushGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem) {
     MICROPROFILE_SCOPE(OpenGL_SurfaceFlush);
 
     ASSERT_MSG(!IsPixelFormatASTC(params.pixel_format), "Unimplemented");
 
+    auto& gl_buffer = res_cache_tmp_mem.gl_buffer;
     // OpenGL temporary buffer needs to be big enough to store raw texture size
-    gl_buffer.resize(1);
     gl_buffer[0].resize(GetSizeInBytes());
 
     const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type);
@@ -713,10 +715,12 @@ void CachedSurface::FlushGLBuffer() {
     }
 }
 
-void CachedSurface::UploadGLMipmapTexture(u32 mip_map, GLuint read_fb_handle,
+void CachedSurface::UploadGLMipmapTexture(RasterizerTemporaryMemory& res_cache_tmp_mem, u32 mip_map,
-                                          GLuint draw_fb_handle) {
+                                          GLuint read_fb_handle, GLuint draw_fb_handle) {
     const auto& rect{params.GetRect(mip_map)};
 
+    auto& gl_buffer = res_cache_tmp_mem.gl_buffer;
+
     // Load data from memory to the surface
     const auto x0 = static_cast<GLint>(rect.left);
     const auto y0 = static_cast<GLint>(rect.bottom);
@@ -845,11 +849,12 @@ void CachedSurface::EnsureTextureDiscrepantView() {
 }
 
 MICROPROFILE_DEFINE(OpenGL_TextureUL, "OpenGL", "Texture Upload", MP_RGB(128, 192, 64));
-void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle) {
+void CachedSurface::UploadGLTexture(RasterizerTemporaryMemory& res_cache_tmp_mem,
+                                    GLuint read_fb_handle, GLuint draw_fb_handle) {
     MICROPROFILE_SCOPE(OpenGL_TextureUL);
 
     for (u32 i = 0; i < params.max_mip_level; i++)
-        UploadGLMipmapTexture(i, read_fb_handle, draw_fb_handle);
+        UploadGLMipmapTexture(res_cache_tmp_mem, i, read_fb_handle, draw_fb_handle);
 }
 
 void CachedSurface::UpdateSwizzle(Tegra::Texture::SwizzleSource swizzle_x,
@@ -929,8 +934,8 @@ Surface RasterizerCacheOpenGL::GetColorBufferSurface(std::size_t index, bool pre
 }
 
 void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) {
-    surface->LoadGLBuffer();
+    surface->LoadGLBuffer(temporal_memory);
-    surface->UploadGLTexture(read_framebuffer.handle, draw_framebuffer.handle);
+    surface->UploadGLTexture(temporal_memory, read_framebuffer.handle, draw_framebuffer.handle);
     surface->MarkAsModified(false, *this);
     surface->MarkForReload(false);
 }

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -355,6 +355,12 @@ namespace OpenGL {
 
 class RasterizerOpenGL;
 
+// This is used to store temporary big buffers,
+// instead of creating/destroying all the time
+struct RasterizerTemporaryMemory {
+    std::vector<std::vector<u8>> gl_buffer;
+};
+
 class CachedSurface final : public RasterizerCacheObject {
 public:
     explicit CachedSurface(const SurfaceParams& params);
@@ -371,10 +377,6 @@ public:
         return memory_size;
     }
 
-    void Flush() override {
-        FlushGLBuffer();
-    }
-
     const OGLTexture& Texture() const {
         return texture;
     }
@@ -397,11 +399,12 @@ public:
     }
 
     // Read/Write data in Switch memory to/from gl_buffer
-    void LoadGLBuffer();
+    void LoadGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem);
-    void FlushGLBuffer();
+    void FlushGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem);
 
     // Upload data in gl_buffer to this surface's texture
-    void UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle);
+    void UploadGLTexture(RasterizerTemporaryMemory& res_cache_tmp_mem, GLuint read_fb_handle,
+                         GLuint draw_fb_handle);
 
     void UpdateSwizzle(Tegra::Texture::SwizzleSource swizzle_x,
                        Tegra::Texture::SwizzleSource swizzle_y,
@@ -429,13 +432,13 @@ public:
     }
 
 private:
-    void UploadGLMipmapTexture(u32 mip_map, GLuint read_fb_handle, GLuint draw_fb_handle);
+    void UploadGLMipmapTexture(RasterizerTemporaryMemory& res_cache_tmp_mem, u32 mip_map,
+                               GLuint read_fb_handle, GLuint draw_fb_handle);
 
     void EnsureTextureDiscrepantView();
 
     OGLTexture texture;
     OGLTexture discrepant_view;
-    std::vector<std::vector<u8>> gl_buffer;
     SurfaceParams params{};
     GLenum gl_target{};
     GLenum gl_internal_format{};
@@ -473,6 +476,11 @@ public:
     void SignalPreDrawCall();
     void SignalPostDrawCall();
 
+protected:
+    void FlushObjectInner(const Surface& object) override {
+        object->FlushGLBuffer(temporal_memory);
+    }
+
 private:
     void LoadSurface(const Surface& surface);
     Surface GetSurface(const SurfaceParams& params, bool preserve_contents = true);
@@ -519,6 +527,8 @@ private:
     std::array<Surface, Maxwell::NumRenderTargets> current_color_buffers;
     Surface last_depth_buffer;
 
+    RasterizerTemporaryMemory temporal_memory;
+
     using SurfaceIntervalCache = boost::icl::interval_map<CacheAddr, Surface>;
     using SurfaceInterval = typename SurfaceIntervalCache::interval_type;
 

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -57,9 +57,6 @@ public:
         return shader_length;
     }
 
-    // We do not have to flush this cache as things in it are never modified by us.
-    void Flush() override {}
-
     /// Gets the shader entries for the shader
     const GLShader::ShaderEntries& GetShaderEntries() const {
         return entries;
@@ -123,6 +120,10 @@ public:
     /// Gets the current specified shader stage program
     Shader GetStageProgram(Maxwell::ShaderProgram program);
 
+protected:
+    // We do not have to flush this cache as things in it are never modified by us.
+    void FlushObjectInner(const Shader& object) override {}
+
 private:
     std::unordered_map<u64, UnspecializedShader> GenerateUnspecializedShaders(
         const std::atomic_bool& stop_loading, const VideoCore::DiskResourceLoadCallback& callback,

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -49,9 +49,6 @@ public:
         return alignment;
     }
 
-    // We do not have to flush this cache as things in it are never modified by us.
-    void Flush() override {}
-
 private:
     VAddr cpu_addr{};
     std::size_t size{};
@@ -87,6 +84,10 @@ public:
         return buffer_handle;
     }
 
+protected:
+    // We do not have to flush this cache as things in it are never modified by us.
+    void FlushObjectInner(const std::shared_ptr<CachedBufferEntry>& object) override {}
+
 private:
     void AlignBuffer(std::size_t alignment);