Merge pull request #2358 from ReinUsesLisp/parallel-shader

gl_shader_cache: Use shared contexts to build shaders in parallel at boot

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -98,9 +98,11 @@ struct FramebufferCacheKey {
     }
 };
 
-RasterizerOpenGL::RasterizerOpenGL(Core::System& system, ScreenInfo& info)
+RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window,
-    : res_cache{*this}, shader_cache{*this, system, device}, global_cache{*this}, system{system},
+                                   ScreenInfo& info)
-      screen_info{info}, buffer_cache(*this, STREAM_BUFFER_SIZE) {
+    : res_cache{*this}, shader_cache{*this, system, emu_window, device},
+      global_cache{*this}, system{system}, screen_info{info},
+      buffer_cache(*this, STREAM_BUFFER_SIZE) {
     OpenGLState::ApplyDefaultState();
 
     shader_program_manager = std::make_unique<GLShader::ProgramManager>();

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -48,7 +48,8 @@ struct FramebufferCacheKey;
 
 class RasterizerOpenGL : public VideoCore::RasterizerInterface {
 public:
-    explicit RasterizerOpenGL(Core::System& system, ScreenInfo& info);
+    explicit RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window,
+                              ScreenInfo& info);
     ~RasterizerOpenGL() override;
 
     void DrawArrays() override;

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -2,10 +2,14 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <mutex>
+#include <thread>
 #include <boost/functional/hash.hpp>
 #include "common/assert.h"
 #include "common/hash.h"
+#include "common/scope_exit.h"
 #include "core/core.h"
+#include "core/frontend/emu_window.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/memory_manager.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
@@ -344,8 +348,8 @@ ShaderDiskCacheUsage CachedShader::GetUsage(GLenum primitive_mode,
 }
 
 ShaderCacheOpenGL::ShaderCacheOpenGL(RasterizerOpenGL& rasterizer, Core::System& system,
-                                     const Device& device)
+                                     Core::Frontend::EmuWindow& emu_window, const Device& device)
-    : RasterizerCache{rasterizer}, device{device}, disk_cache{system} {}
+    : RasterizerCache{rasterizer}, emu_window{emu_window}, device{device}, disk_cache{system} {}
 
 void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
                                       const VideoCore::DiskResourceLoadCallback& callback) {
@@ -353,62 +357,107 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
     if (!transferable) {
         return;
     }
-    const auto [raws, usages] = *transferable;
+    const auto [raws, shader_usages] = *transferable;
 
     auto [decompiled, dumps] = disk_cache.LoadPrecompiled();
 
     const auto supported_formats{GetSupportedFormats()};
-    const auto unspecialized{
+    const auto unspecialized_shaders{
         GenerateUnspecializedShaders(stop_loading, callback, raws, decompiled)};
-    if (stop_loading)
+    if (stop_loading) {
         return;
+    }
 
     // Track if precompiled cache was altered during loading to know if we have to serialize the
     // virtual precompiled cache file back to the hard drive
     bool precompiled_cache_altered = false;
 
-    // Build shaders
+    // Inform the frontend about shader build initialization
-    if (callback)
+    if (callback) {
-        callback(VideoCore::LoadCallbackStage::Build, 0, usages.size());
+        callback(VideoCore::LoadCallbackStage::Build, 0, shader_usages.size());
-    for (std::size_t i = 0; i < usages.size(); ++i) {
+    }
-        if (stop_loading)
+
+    std::mutex mutex;
+    std::size_t built_shaders = 0; // It doesn't have be atomic since it's used behind a mutex
+    std::atomic_bool compilation_failed = false;
+
+    const auto Worker = [&](Core::Frontend::GraphicsContext* context, std::size_t begin,
+                            std::size_t end, const std::vector<ShaderDiskCacheUsage>& shader_usages,
+                            const ShaderDumpsMap& dumps) {
+        context->MakeCurrent();
+        SCOPE_EXIT({ return context->DoneCurrent(); });
+
+        for (std::size_t i = begin; i < end; ++i) {
+            if (stop_loading || compilation_failed) {
                 return;
+            }
+            const auto& usage{shader_usages[i]};
+            LOG_INFO(Render_OpenGL, "Building shader {:016x} (index {} of {})",
+                     usage.unique_identifier, i, shader_usages.size());
 
-        const auto& usage{usages[i]};
+            const auto& unspecialized{unspecialized_shaders.at(usage.unique_identifier)};
-        LOG_INFO(Render_OpenGL, "Building shader {:016x} ({} of {})", usage.unique_identifier,
+            const auto dump{dumps.find(usage)};
-                 i + 1, usages.size());
-
-        const auto& unspec{unspecialized.at(usage.unique_identifier)};
-        const auto dump_it = dumps.find(usage);
 
             CachedProgram shader;
-        if (dump_it != dumps.end()) {
+            if (dump != dumps.end()) {
                 // If the shader is dumped, attempt to load it with
-            shader = GeneratePrecompiledProgram(dump_it->second, supported_formats);
+                shader = GeneratePrecompiledProgram(dump->second, supported_formats);
                 if (!shader) {
+                    compilation_failed = true;
+                    return;
+                }
+            }
+            if (!shader) {
+                shader = SpecializeShader(unspecialized.code, unspecialized.entries,
+                                          unspecialized.program_type, usage.bindings,
+                                          usage.primitive, true);
+            }
+
+            std::scoped_lock lock(mutex);
+            if (callback) {
+                callback(VideoCore::LoadCallbackStage::Build, ++built_shaders,
+                         shader_usages.size());
+            }
+
+            precompiled_programs.emplace(usage, std::move(shader));
+        }
+    };
+
+    const auto num_workers{static_cast<std::size_t>(std::thread::hardware_concurrency() + 1)};
+    const std::size_t bucket_size{shader_usages.size() / num_workers};
+    std::vector<std::unique_ptr<Core::Frontend::GraphicsContext>> contexts(num_workers);
+    std::vector<std::thread> threads(num_workers);
+    for (std::size_t i = 0; i < num_workers; ++i) {
+        const bool is_last_worker = i + 1 == num_workers;
+        const std::size_t start{bucket_size * i};
+        const std::size_t end{is_last_worker ? shader_usages.size() : start + bucket_size};
+
+        // On some platforms the shared context has to be created from the GUI thread
+        contexts[i] = emu_window.CreateSharedContext();
+        threads[i] = std::thread(Worker, contexts[i].get(), start, end, shader_usages, dumps);
+    }
+    for (auto& thread : threads) {
+        thread.join();
+    }
+
+    if (compilation_failed) {
         // Invalidate the precompiled cache if a shader dumped shader was rejected
         disk_cache.InvalidatePrecompiled();
-                precompiled_cache_altered = true;
         dumps.clear();
+        precompiled_cache_altered = true;
+        return;
     }
-        }
+    if (stop_loading) {
-        if (!shader) {
+        return;
-            shader = SpecializeShader(unspec.code, unspec.entries, unspec.program_type,
-                                      usage.bindings, usage.primitive, true);
-        }
-        precompiled_programs.insert({usage, std::move(shader)});
-
-        if (callback)
-            callback(VideoCore::LoadCallbackStage::Build, i + 1, usages.size());
     }
 
     // TODO(Rodrigo): Do state tracking for transferable shaders and do a dummy draw before
     // precompiling them
 
-    for (std::size_t i = 0; i < usages.size(); ++i) {
+    for (std::size_t i = 0; i < shader_usages.size(); ++i) {
-        const auto& usage{usages[i]};
+        const auto& usage{shader_usages[i]};
         if (dumps.find(usage) == dumps.end()) {
-            const auto& program = precompiled_programs.at(usage);
+            const auto& program{precompiled_programs.at(usage)};
             disk_cache.SaveDump(usage, program->handle);
             precompiled_cache_altered = true;
         }

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -22,7 +22,11 @@
 
 namespace Core {
 class System;
-} // namespace Core
+}
+
+namespace Core::Frontend {
+class EmuWindow;
+}
 
 namespace OpenGL {
 
@@ -111,7 +115,7 @@ private:
 class ShaderCacheOpenGL final : public RasterizerCache<Shader> {
 public:
     explicit ShaderCacheOpenGL(RasterizerOpenGL& rasterizer, Core::System& system,
-                               const Device& device);
+                               Core::Frontend::EmuWindow& emu_window, const Device& device);
 
     /// Loads disk cache for the current game
     void LoadDiskCache(const std::atomic_bool& stop_loading,
@@ -133,13 +137,13 @@ private:
     CachedProgram GeneratePrecompiledProgram(const ShaderDiskCacheDump& dump,
                                              const std::set<GLenum>& supported_formats);
 
+    Core::Frontend::EmuWindow& emu_window;
     const Device& device;
-
-    std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
-
     ShaderDiskCacheOpenGL disk_cache;
+
     PrecompiledShaders precompiled_shaders;
     PrecompiledPrograms precompiled_programs;
+    std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -183,8 +183,7 @@ ShaderDiskCacheOpenGL::LoadTransferable() {
     return {{raws, usages}};
 }
 
-std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>,
+std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>, ShaderDumpsMap>
-          std::unordered_map<ShaderDiskCacheUsage, ShaderDiskCacheDump>>
 ShaderDiskCacheOpenGL::LoadPrecompiled() {
     if (!IsUsable())
         return {};
@@ -208,8 +207,7 @@ ShaderDiskCacheOpenGL::LoadPrecompiled() {
     return *result;
 }
 
-std::optional<std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>,
+std::optional<std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>, ShaderDumpsMap>>
-                        std::unordered_map<ShaderDiskCacheUsage, ShaderDiskCacheDump>>>
 ShaderDiskCacheOpenGL::LoadPrecompiledFile(FileUtil::IOFile& file) {
     // Read compressed file from disk and decompress to virtual precompiled cache file
     std::vector<u8> compressed(file.GetSize());
@@ -230,7 +228,7 @@ ShaderDiskCacheOpenGL::LoadPrecompiledFile(FileUtil::IOFile& file) {
     }
 
     std::unordered_map<u64, ShaderDiskCacheDecompiled> decompiled;
-    std::unordered_map<ShaderDiskCacheUsage, ShaderDiskCacheDump> dumps;
+    ShaderDumpsMap dumps;
     while (precompiled_cache_virtual_file_offset < precompiled_cache_virtual_file.GetSize()) {
         PrecompiledEntryKind kind{};
         if (!LoadObjectFromPrecompiled(kind)) {

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -33,6 +33,11 @@ namespace OpenGL {
 using ProgramCode = std::vector<u64>;
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
+struct ShaderDiskCacheUsage;
+struct ShaderDiskCacheDump;
+
+using ShaderDumpsMap = std::unordered_map<ShaderDiskCacheUsage, ShaderDiskCacheDump>;
+
 /// Allocated bindings used by an OpenGL shader program
 struct BaseBindings {
     u32 cbuf{};

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -97,8 +97,8 @@ static std::array<GLfloat, 3 * 2> MakeOrthographicMatrix(const float width, cons
     return matrix;
 }
 
-RendererOpenGL::RendererOpenGL(Core::Frontend::EmuWindow& window, Core::System& system)
+RendererOpenGL::RendererOpenGL(Core::Frontend::EmuWindow& emu_window, Core::System& system)
-    : VideoCore::RendererBase{window}, system{system} {}
+    : VideoCore::RendererBase{emu_window}, emu_window{emu_window}, system{system} {}
 
 RendererOpenGL::~RendererOpenGL() = default;
 
@@ -265,7 +265,7 @@ void RendererOpenGL::CreateRasterizer() {
     }
     // Initialize sRGB Usage
     OpenGLState::ClearsRGBUsed();
-    rasterizer = std::make_unique<RasterizerOpenGL>(system, screen_info);
+    rasterizer = std::make_unique<RasterizerOpenGL>(system, emu_window, screen_info);
 }
 
 void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,

src/video_core/renderer_opengl/renderer_opengl.h

@@ -45,7 +45,7 @@ struct ScreenInfo {
 
 class RendererOpenGL : public VideoCore::RendererBase {
 public:
-    explicit RendererOpenGL(Core::Frontend::EmuWindow& window, Core::System& system);
+    explicit RendererOpenGL(Core::Frontend::EmuWindow& emu_window, Core::System& system);
     ~RendererOpenGL() override;
 
     /// Swap buffers (render frame)
@@ -77,6 +77,7 @@ private:
     void LoadColorToActiveGLTexture(u8 color_r, u8 color_g, u8 color_b, u8 color_a,
                                     const TextureInfo& texture);
 
+    Core::Frontend::EmuWindow& emu_window;
     Core::System& system;
 
     OpenGLState state;

src/yuzu/bootmanager.cpp

@@ -91,25 +91,25 @@ void EmuThread::run() {
 
 class GGLContext : public Core::Frontend::GraphicsContext {
 public:
-    explicit GGLContext(QOpenGLContext* shared_context)
+    explicit GGLContext(QOpenGLContext* shared_context) : shared_context{shared_context} {
-        : context{std::make_unique<QOpenGLContext>(shared_context)} {
+        context.setFormat(shared_context->format());
-        surface.setFormat(shared_context->format());
+        context.setShareContext(shared_context);
-        surface.create();
+        context.create();
     }
 
     void MakeCurrent() override {
-        context->makeCurrent(&surface);
+        context.makeCurrent(shared_context->surface());
     }
 
     void DoneCurrent() override {
-        context->doneCurrent();
+        context.doneCurrent();
     }
 
     void SwapBuffers() override {}
 
 private:
-    std::unique_ptr<QOpenGLContext> context;
+    QOpenGLContext* shared_context;
-    QOffscreenSurface surface;
+    QOpenGLContext context;
 };
 
 // This class overrides paintEvent and resizeEvent to prevent the GUI thread from stealing GL
@@ -358,7 +358,7 @@ void GRenderWindow::OnClientAreaResized(unsigned width, unsigned height) {
 }
 
 std::unique_ptr<Core::Frontend::GraphicsContext> GRenderWindow::CreateSharedContext() const {
-    return std::make_unique<GGLContext>(shared_context.get());
+    return std::make_unique<GGLContext>(context.get());
 }
 
 void GRenderWindow::InitRenderTarget() {