2 files changed, 277 insertions, 0 deletions

diff --git a/src/video_core/shader/async_shaders.cpp b/src/video_core/shader/async_shaders.cpp
new file mode 100644
index 000000000..fb94ac2e7
--- /dev/null
+++ b/src/video_core/shader/async_shaders.cpp
@@ -0,0 +1,170 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <chrono>
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/renderer_base.h"
+#include "video_core/renderer_opengl/gl_shader_cache.h"
+#include "video_core/shader/async_shaders.h"
+
+namespace VideoCommon::Shader {
+AsyncShaders::AsyncShaders(Core::Frontend::EmuWindow& emu_window) : emu_window(emu_window) {}
+AsyncShaders::~AsyncShaders() {
+    KillWorkers();
+}
+
+void AsyncShaders::AllocateWorkers(std::size_t num_workers) {
+    // If we're already have workers queued or don't want to queue workers, ignore
+    if (num_workers == worker_threads.size() || num_workers == 0) {
+        return;
+    }
+
+    // If workers already exist, clear them
+    if (!worker_threads.empty()) {
+        FreeWorkers();
+    }
+
+    // Create workers
+    for (std::size_t i = 0; i < num_workers; i++) {
+        context_list.push_back(emu_window.CreateSharedContext());
+        worker_threads.push_back(std::move(
+            std::thread(&AsyncShaders::ShaderCompilerThread, this, context_list[i].get())));
+    }
+}
+
+void AsyncShaders::FreeWorkers() {
+    // Mark all threads to quit
+    is_thread_exiting.store(true);
+    for (auto& thread : worker_threads) {
+        thread.join();
+    }
+    // Clear our shared contexts
+    context_list.clear();
+
+    // Clear our worker threads
+    worker_threads.clear();
+}
+
+void AsyncShaders::KillWorkers() {
+    is_thread_exiting.store(true);
+    for (auto& thread : worker_threads) {
+        thread.detach();
+    }
+    // Clear our shared contexts
+    context_list.clear();
+
+    // Clear our worker threads
+    worker_threads.clear();
+}
+
+bool AsyncShaders::HasWorkQueued() {
+    std::shared_lock lock(queue_mutex);
+    return !pending_queue.empty();
+}
+
+bool AsyncShaders::HasCompletedWork() {
+    std::shared_lock lock(completed_mutex);
+    return !finished_work.empty();
+}
+
+bool AsyncShaders::IsShaderAsync(const Tegra::GPU& gpu) const {
+    const auto& regs = gpu.Maxwell3D().regs;
+
+    // If something is using depth, we can assume that games are not rendering anything which will
+    // be used one time.
+    if (regs.zeta_enable) {
+        return true;
+    }
+
+    // If games are using a small index count, we can assume these are full screen quads. Usually
+    // these shaders are only used once for building textures so we can assume they can't be built
+    // async
+    if (regs.index_array.count <= 6 || regs.vertex_buffer.count <= 6) {
+        return false;
+    }
+
+    return true;
+}
+
+std::vector<AsyncShaders::Result> AsyncShaders::GetCompletedWork() {
+    std::vector<AsyncShaders::Result> results;
+    {
+        std::unique_lock lock(completed_mutex);
+        results.assign(std::make_move_iterator(finished_work.begin()),
+                       std::make_move_iterator(finished_work.end()));
+        finished_work.clear();
+    }
+    return results;
+}
+
+void AsyncShaders::QueueOpenGLShader(const OpenGL::Device& device,
+                                     Tegra::Engines::ShaderType shader_type, u64 uid,
+                                     std::vector<u64> code, std::vector<u64> code_b,
+                                     u32 main_offset,
+                                     VideoCommon::Shader::CompilerSettings compiler_settings,
+                                     const VideoCommon::Shader::Registry& registry,
+                                     VAddr cpu_addr) {
+    WorkerParams params{device.UseAssemblyShaders() ? AsyncShaders::Backend::GLASM
+                                                    : AsyncShaders::Backend::OpenGL,
+                        device,
+                        shader_type,
+                        uid,
+                        std::move(code),
+                        std::move(code_b),
+                        main_offset,
+                        compiler_settings,
+                        registry,
+                        cpu_addr};
+    std::unique_lock lock(queue_mutex);
+    pending_queue.push_back(std::move(params));
+}
+
+void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context) {
+    using namespace std::chrono_literals;
+    while (!is_thread_exiting.load(std::memory_order_relaxed)) {
+        // Partial lock to allow all threads to read at the same time
+        if (!HasWorkQueued()) {
+            continue;
+        }
+        // Complete lock for pulling workload
+        queue_mutex.lock();
+        // Another thread beat us, just unlock and wait for the next load
+        if (pending_queue.empty()) {
+            queue_mutex.unlock();
+            continue;
+        }
+        // Pull work from queue
+        WorkerParams work = std::move(pending_queue.front());
+        pending_queue.pop_front();
+        queue_mutex.unlock();
+
+        if (work.backend == AsyncShaders::Backend::OpenGL ||
+            work.backend == AsyncShaders::Backend::GLASM) {
+            const ShaderIR ir(work.code, work.main_offset, work.compiler_settings, work.registry);
+            const auto scope = context->Acquire();
+            auto program =
+                OpenGL::BuildShader(work.device, work.shader_type, work.uid, ir, work.registry);
+            Result result{};
+            result.backend = work.backend;
+            result.cpu_address = work.cpu_address;
+            result.uid = work.uid;
+            result.code = std::move(work.code);
+            result.code_b = std::move(work.code_b);
+            result.shader_type = work.shader_type;
+
+            if (work.backend == AsyncShaders::Backend::OpenGL) {
+                result.program.opengl = std::move(program->source_program);
+            } else if (work.backend == AsyncShaders::Backend::GLASM) {
+                result.program.glasm = std::move(program->assembly_program);
+            }
+
+            {
+                std::unique_lock complete_lock(completed_mutex);
+                finished_work.push_back(std::move(result));
+            }
+        }
+    }
+}
+
+} // namespace VideoCommon::Shader
diff --git a/src/video_core/shader/async_shaders.h b/src/video_core/shader/async_shaders.h
new file mode 100644
index 000000000..26bc38326
--- /dev/null
+++ b/src/video_core/shader/async_shaders.h
@@ -0,0 +1,107 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <deque>
+#include <memory>
+#include <shared_mutex>
+#include <thread>
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "video_core/renderer_opengl/gl_device.h"
+#include "video_core/renderer_opengl/gl_resource_manager.h"
+#include "video_core/renderer_opengl/gl_shader_decompiler.h"
+
+namespace Core::Frontend {
+class EmuWindow;
+class GraphicsContext;
+} // namespace Core::Frontend
+
+namespace Tegra {
+class GPU;
+}
+
+namespace VideoCommon::Shader {
+
+class AsyncShaders {
+public:
+    enum class Backend {
+        OpenGL,
+        GLASM,
+    };
+
+    struct ResultPrograms {
+        OpenGL::OGLProgram opengl;
+        OpenGL::OGLAssemblyProgram glasm;
+    };
+
+    struct Result {
+        u64 uid;
+        VAddr cpu_address;
+        Backend backend;
+        ResultPrograms program;
+        std::vector<u64> code;
+        std::vector<u64> code_b;
+        Tegra::Engines::ShaderType shader_type;
+    };
+
+    explicit AsyncShaders(Core::Frontend::EmuWindow& emu_window);
+    ~AsyncShaders();
+
+    /// Start up shader worker threads
+    void AllocateWorkers(std::size_t num_workers);
+
+    /// Clear the shader queue and kill all worker threads
+    void FreeWorkers();
+
+    // Force end all threads
+    void KillWorkers();
+
+    /// Check our worker queue to see if we have any work queued already
+    bool HasWorkQueued();
+
+    /// Check to see if any shaders have actually been compiled
+    bool HasCompletedWork();
+
+    /// Deduce if a shader can be build on another thread of MUST be built in sync. We cannot build
+    /// every shader async as some shaders are only built and executed once. We try to "guess" which
+    /// shader would be used only once
+    bool IsShaderAsync(const Tegra::GPU& gpu) const;
+
+    /// Pulls completed compiled shaders
+    std::vector<Result> GetCompletedWork();
+
+    void QueueOpenGLShader(const OpenGL::Device& device, Tegra::Engines::ShaderType shader_type,
+                           u64 uid, std::vector<u64> code, std::vector<u64> code_b, u32 main_offset,
+                           VideoCommon::Shader::CompilerSettings compiler_settings,
+                           const VideoCommon::Shader::Registry& registry, VAddr cpu_addr);
+
+private:
+    void ShaderCompilerThread(Core::Frontend::GraphicsContext* context);
+
+    struct WorkerParams {
+        AsyncShaders::Backend backend;
+        OpenGL::Device device;
+        Tegra::Engines::ShaderType shader_type;
+        u64 uid;
+        std::vector<u64> code;
+        std::vector<u64> code_b;
+        u32 main_offset;
+        VideoCommon::Shader::CompilerSettings compiler_settings;
+        VideoCommon::Shader::Registry registry;
+        VAddr cpu_address;
+    };
+
+    std::shared_mutex queue_mutex;
+    std::shared_mutex completed_mutex;
+    std::atomic<bool> is_thread_exiting{};
+    std::vector<std::unique_ptr<Core::Frontend::GraphicsContext>> context_list;
+    std::vector<std::thread> worker_threads;
+    std::deque<WorkerParams> pending_queue;
+    std::vector<AsyncShaders::Result> finished_work;
+    Core::Frontend::EmuWindow& emu_window;
+};
+
+} // namespace VideoCommon::Shader