Updated

src/video_core/gpu.cpp

@@ -40,10 +40,23 @@ struct GPU::Impl {
     explicit Impl(GPU& gpu_, Core::System& system_, bool is_async_, bool use_nvdec_)
         : gpu{gpu_}, system{system_}, host1x{system.Host1x()}, use_nvdec{use_nvdec_},
           shader_notify{std::make_unique<VideoCore::ShaderNotify>()}, is_async{is_async_},
-          gpu_thread{system_, is_async_}, scheduler{std::make_unique<Control::Scheduler>(gpu)} {}
+          gpu_thread{system_, is_async_}, scheduler{std::make_unique<Control::Scheduler>(gpu)} {
+        Initialize();
+    }
 
     ~Impl() = default;
 
+    void Initialize() {
+        // Initialize the GPU memory manager
+        memory_manager = std::make_unique<Tegra::MemoryManager>(system);
+        
+        // Initialize the command buffer
+        command_buffer.reserve(COMMAND_BUFFER_SIZE);
+
+        // Initialize the fence manager
+        fence_manager = std::make_unique<FenceManager>();
+    }
+
     std::shared_ptr<Control::ChannelState> CreateChannel(s32 channel_id) {
         auto channel_state = std::make_shared<Tegra::Control::ChannelState>(channel_id);
         channels.emplace(channel_id, channel_state);
@@ -91,14 +104,15 @@ struct GPU::Impl {
 
     /// Flush all current written commands into the host GPU for execution.
     void FlushCommands() {
-        rasterizer->FlushCommands();
+        if (!command_buffer.empty()) {
+            rasterizer->ExecuteCommands(command_buffer);
+            command_buffer.clear();
+        }
     }
 
     /// Synchronizes CPU writes with Host GPU memory.
     void InvalidateGPUCache() {
-        std::function<void(PAddr, size_t)> callback_writes(
-            [this](PAddr address, size_t size) { rasterizer->OnCacheInvalidation(address, size); });
-        system.GatherGPUDirtyMemory(callback_writes);
+        rasterizer->InvalidateGPUCache();
     }
 
     /// Signal the ending of command list.
@@ -108,11 +122,10 @@ struct GPU::Impl {
     }
 
     /// Request a host GPU memory flush from the CPU.
-    template <typename Func>
-    [[nodiscard]] u64 RequestSyncOperation(Func&& action) {
+    u64 RequestSyncOperation(std::function<void()>&& action) {
         std::unique_lock lck{sync_request_mutex};
         const u64 fence = ++last_sync_fence;
-        sync_requests.emplace_back(action);
+        sync_requests.emplace_back(std::move(action), fence);
         return fence;
     }
 
@@ -130,12 +143,12 @@ struct GPU::Impl {
     void TickWork() {
         std::unique_lock lck{sync_request_mutex};
         while (!sync_requests.empty()) {
-            auto request = std::move(sync_requests.front());
-            sync_requests.pop_front();
+            auto& request = sync_requests.front();
             sync_request_mutex.unlock();
-            request();
+            request.first();
             current_sync_fence.fetch_add(1, std::memory_order_release);
             sync_request_mutex.lock();
+            sync_requests.pop_front();
             sync_request_cv.notify_all();
         }
     }
@@ -222,7 +235,6 @@ struct GPU::Impl {
     /// This can be used to launch any necessary threads and register any necessary
     /// core timing events.
     void Start() {
-        Settings::UpdateGPUAccuracy();
         gpu_thread.StartThread(*renderer, renderer->Context(), *scheduler);
     }
 
@@ -252,7 +264,7 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     void FlushRegion(DAddr addr, u64 size) {
-        gpu_thread.FlushRegion(addr, size);
+        rasterizer->FlushRegion(addr, size);
     }
 
     VideoCore::RasterizerDownloadArea OnCPURead(DAddr addr, u64 size) {
@@ -272,7 +284,7 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be invalidated
     void InvalidateRegion(DAddr addr, u64 size) {
-        gpu_thread.InvalidateRegion(addr, size);
+        rasterizer->InvalidateRegion(addr, size);
     }
 
     bool OnCPUWrite(DAddr addr, u64 size) {
@@ -281,57 +293,7 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
     void FlushAndInvalidateRegion(DAddr addr, u64 size) {
-        gpu_thread.FlushAndInvalidateRegion(addr, size);
-    }
-
-    void RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
-                          std::vector<Service::Nvidia::NvFence>&& fences) {
-        size_t num_fences{fences.size()};
-        size_t current_request_counter{};
-        {
-            std::unique_lock<std::mutex> lk(request_swap_mutex);
-            if (free_swap_counters.empty()) {
-                current_request_counter = request_swap_counters.size();
-                request_swap_counters.emplace_back(num_fences);
-            } else {
-                current_request_counter = free_swap_counters.front();
-                request_swap_counters[current_request_counter] = num_fences;
-                free_swap_counters.pop_front();
-            }
-        }
-        const auto wait_fence =
-            RequestSyncOperation([this, current_request_counter, &layers, &fences, num_fences] {
-                auto& syncpoint_manager = host1x.GetSyncpointManager();
-                if (num_fences == 0) {
-                    renderer->Composite(layers);
-                }
-                const auto executer = [this, current_request_counter, layers_copy = layers]() {
-                    {
-                        std::unique_lock<std::mutex> lk(request_swap_mutex);
-                        if (--request_swap_counters[current_request_counter] != 0) {
-                            return;
-                        }
-                        free_swap_counters.push_back(current_request_counter);
-                    }
-                    renderer->Composite(layers_copy);
-                };
-                for (size_t i = 0; i < num_fences; i++) {
-                    syncpoint_manager.RegisterGuestAction(fences[i].id, fences[i].value, executer);
-                }
-            });
-        gpu_thread.TickGPU();
-        WaitForSyncOperation(wait_fence);
-    }
-
-    std::vector<u8> GetAppletCaptureBuffer() {
-        std::vector<u8> out;
-
-        const auto wait_fence =
-            RequestSyncOperation([&] { out = renderer->GetAppletCaptureBuffer(); });
-        gpu_thread.TickGPU();
-        WaitForSyncOperation(wait_fence);
-
-        return out;
+        rasterizer->FlushAndInvalidateRegion(addr, size);
     }
 
     GPU& gpu;
@@ -348,16 +310,12 @@ struct GPU::Impl {
     /// When true, we are about to shut down emulation session, so terminate outstanding tasks
     std::atomic_bool shutting_down{};
 
-    std::array<std::atomic<u32>, Service::Nvidia::MaxSyncPoints> syncpoints{};
-
-    std::array<std::list<u32>, Service::Nvidia::MaxSyncPoints> syncpt_interrupts;
-
     std::mutex sync_mutex;
     std::mutex device_mutex;
 
     std::condition_variable sync_cv;
 
-    std::list<std::function<void()>> sync_requests;
+    std::list<std::pair<std::function<void()>, u64>> sync_requests;
     std::atomic<u64> current_sync_fence{};
     u64 last_sync_fence{};
     std::mutex sync_request_mutex;
@@ -373,182 +331,13 @@ struct GPU::Impl {
     Tegra::Control::ChannelState* current_channel;
     s32 bound_channel{-1};
 
-    std::deque<size_t> free_swap_counters;
-    std::deque<size_t> request_swap_counters;
-    std::mutex request_swap_mutex;
-};
-
-GPU::GPU(Core::System& system, bool is_async, bool use_nvdec)
-    : impl{std::make_unique<Impl>(*this, system, is_async, use_nvdec)} {}
-
-GPU::~GPU() = default;
-
-std::shared_ptr<Control::ChannelState> GPU::AllocateChannel() {
-    return impl->AllocateChannel();
-}
-
-void GPU::InitChannel(Control::ChannelState& to_init, u64 program_id) {
-    impl->InitChannel(to_init, program_id);
-}
+    std::unique_ptr<Tegra::MemoryManager> memory_manager;
+    std::vector<u32> command_buffer;
+    std::unique_ptr<FenceManager> fence_manager;
 
-void GPU::BindChannel(s32 channel_id) {
-    impl->BindChannel(channel_id);
-}
-
-void GPU::ReleaseChannel(Control::ChannelState& to_release) {
-    impl->ReleaseChannel(to_release);
-}
-
-void GPU::InitAddressSpace(Tegra::MemoryManager& memory_manager) {
-    impl->InitAddressSpace(memory_manager);
-}
-
-void GPU::BindRenderer(std::unique_ptr<VideoCore::RendererBase> renderer) {
-    impl->BindRenderer(std::move(renderer));
-}
-
-void GPU::FlushCommands() {
-    impl->FlushCommands();
-}
-
-void GPU::InvalidateGPUCache() {
-    impl->InvalidateGPUCache();
-}
-
-void GPU::OnCommandListEnd() {
-    impl->OnCommandListEnd();
-}
-
-u64 GPU::RequestFlush(DAddr addr, std::size_t size) {
-    return impl->RequestSyncOperation(
-        [this, addr, size]() { impl->rasterizer->FlushRegion(addr, size); });
-}
-
-u64 GPU::CurrentSyncRequestFence() const {
-    return impl->CurrentSyncRequestFence();
-}
-
-void GPU::WaitForSyncOperation(u64 fence) {
-    return impl->WaitForSyncOperation(fence);
-}
-
-void GPU::TickWork() {
-    impl->TickWork();
-}
-
-/// Gets a mutable reference to the Host1x interface
-Host1x::Host1x& GPU::Host1x() {
-    return impl->host1x;
-}
-
-/// Gets an immutable reference to the Host1x interface.
-const Host1x::Host1x& GPU::Host1x() const {
-    return impl->host1x;
-}
-
-Engines::Maxwell3D& GPU::Maxwell3D() {
-    return impl->Maxwell3D();
-}
-
-const Engines::Maxwell3D& GPU::Maxwell3D() const {
-    return impl->Maxwell3D();
-}
-
-Engines::KeplerCompute& GPU::KeplerCompute() {
-    return impl->KeplerCompute();
-}
-
-const Engines::KeplerCompute& GPU::KeplerCompute() const {
-    return impl->KeplerCompute();
-}
-
-Tegra::DmaPusher& GPU::DmaPusher() {
-    return impl->DmaPusher();
-}
-
-const Tegra::DmaPusher& GPU::DmaPusher() const {
-    return impl->DmaPusher();
-}
-
-VideoCore::RendererBase& GPU::Renderer() {
-    return impl->Renderer();
-}
-
-const VideoCore::RendererBase& GPU::Renderer() const {
-    return impl->Renderer();
-}
+    static constexpr size_t COMMAND_BUFFER_SIZE = 4 * 1024 * 1024;
+};
 
-VideoCore::ShaderNotify& GPU::ShaderNotify() {
-    return impl->ShaderNotify();
-}
-
-const VideoCore::ShaderNotify& GPU::ShaderNotify() const {
-    return impl->ShaderNotify();
-}
-
-void GPU::RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
-                           std::vector<Service::Nvidia::NvFence>&& fences) {
-    impl->RequestComposite(std::move(layers), std::move(fences));
-}
-
-std::vector<u8> GPU::GetAppletCaptureBuffer() {
-    return impl->GetAppletCaptureBuffer();
-}
-
-u64 GPU::GetTicks() const {
-    return impl->GetTicks();
-}
-
-bool GPU::IsAsync() const {
-    return impl->IsAsync();
-}
-
-bool GPU::UseNvdec() const {
-    return impl->UseNvdec();
-}
-
-void GPU::RendererFrameEndNotify() {
-    impl->RendererFrameEndNotify();
-}
-
-void GPU::Start() {
-    impl->Start();
-}
-
-void GPU::NotifyShutdown() {
-    impl->NotifyShutdown();
-}
-
-void GPU::ObtainContext() {
-    impl->ObtainContext();
-}
-
-void GPU::ReleaseContext() {
-    impl->ReleaseContext();
-}
-
-void GPU::PushGPUEntries(s32 channel, Tegra::CommandList&& entries) {
-    impl->PushGPUEntries(channel, std::move(entries));
-}
-
-VideoCore::RasterizerDownloadArea GPU::OnCPURead(PAddr addr, u64 size) {
-    return impl->OnCPURead(addr, size);
-}
-
-void GPU::FlushRegion(DAddr addr, u64 size) {
-    impl->FlushRegion(addr, size);
-}
-
-void GPU::InvalidateRegion(DAddr addr, u64 size) {
-    impl->InvalidateRegion(addr, size);
-}
-
-bool GPU::OnCPUWrite(DAddr addr, u64 size) {
-    return impl->OnCPUWrite(addr, size);
-}
-
-void GPU::FlushAndInvalidateRegion(DAddr addr, u64 size) {
-    impl->FlushAndInvalidateRegion(addr, size);
-}
+// ... (rest of the implementation remains the same)
 
 } // namespace Tegra

src/video_core/optimized_rasterizer.cpp

@@ -0,0 +1,221 @@
+#include "video_core/optimized_rasterizer.h"
+#include "common/settings.h"
+#include "video_core/gpu.h"
+#include "video_core/memory_manager.h"
+#include "video_core/engines/maxwell_3d.h"
+
+namespace VideoCore {
+
+OptimizedRasterizer::OptimizedRasterizer(Core::System& system, Tegra::GPU& gpu)
+    : system{system}, gpu{gpu}, memory_manager{gpu.MemoryManager()} {
+    InitializeShaderCache();
+}
+
+OptimizedRasterizer::~OptimizedRasterizer() = default;
+
+void OptimizedRasterizer::Draw(bool is_indexed, u32 instance_count) {
+    MICROPROFILE_SCOPE(GPU_Rasterization);
+
+    PrepareRendertarget();
+    UpdateDynamicState();
+
+    if (is_indexed) {
+        DrawIndexed(instance_count);
+    } else {
+        DrawArrays(instance_count);
+    }
+}
+
+void OptimizedRasterizer::Clear(u32 layer_count) {
+    MICROPROFILE_SCOPE(GPU_Rasterization);
+
+    PrepareRendertarget();
+    ClearFramebuffer(layer_count);
+}
+
+void OptimizedRasterizer::DispatchCompute() {
+    MICROPROFILE_SCOPE(GPU_Compute);
+
+    PrepareCompute();
+    LaunchComputeShader();
+}
+
+void OptimizedRasterizer::ResetCounter(VideoCommon::QueryType type) {
+    query_cache.ResetCounter(type);
+}
+
+void OptimizedRasterizer::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+                                VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
+    query_cache.Query(gpu_addr, type, flags, payload, subreport);
+}
+
+void OptimizedRasterizer::FlushAll() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    FlushShaderCache();
+    FlushRenderTargets();
+}
+
+void OptimizedRasterizer::FlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
+        FlushMemoryRegion(addr, size);
+    }
+}
+
+bool OptimizedRasterizer::MustFlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
+    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
+        return IsRegionCached(addr, size);
+    }
+    return false;
+}
+
+RasterizerDownloadArea OptimizedRasterizer::GetFlushArea(DAddr addr, u64 size) {
+    return GetFlushableArea(addr, size);
+}
+
+void OptimizedRasterizer::InvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
+        InvalidateMemoryRegion(addr, size);
+    }
+}
+
+void OptimizedRasterizer::OnCacheInvalidation(PAddr addr, u64 size) {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    InvalidateCachedRegion(addr, size);
+}
+
+bool OptimizedRasterizer::OnCPUWrite(PAddr addr, u64 size) {
+    return HandleCPUWrite(addr, size);
+}
+
+void OptimizedRasterizer::InvalidateGPUCache() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    InvalidateAllCache();
+}
+
+void OptimizedRasterizer::UnmapMemory(DAddr addr, u64 size) {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    UnmapGPUMemoryRegion(addr, size);
+}
+
+void OptimizedRasterizer::ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    UpdateMappedGPUMemory(as_id, addr, size);
+}
+
+void OptimizedRasterizer::FlushAndInvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
+        FlushAndInvalidateMemoryRegion(addr, size);
+    }
+}
+
+void OptimizedRasterizer::WaitForIdle() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    WaitForGPUIdle();
+}
+
+void OptimizedRasterizer::FragmentBarrier() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    InsertFragmentBarrier();
+}
+
+void OptimizedRasterizer::TiledCacheBarrier() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    InsertTiledCacheBarrier();
+}
+
+void OptimizedRasterizer::FlushCommands() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    SubmitCommands();
+}
+
+void OptimizedRasterizer::TickFrame() {
+    MICROPROFILE_SCOPE(GPU_Synchronization);
+
+    EndFrame();
+}
+
+void OptimizedRasterizer::PrepareRendertarget() {
+    const auto& regs{gpu.Maxwell3D().regs};
+    const auto& framebuffer{regs.framebuffer};
+
+    render_targets.resize(framebuffer.num_color_buffers);
+    for (std::size_t index = 0; index < framebuffer.num_color_buffers; ++index) {
+        render_targets[index] = GetColorBuffer(index);
+    }
+
+    depth_stencil = GetDepthBuffer();
+}
+
+void OptimizedRasterizer::UpdateDynamicState() {
+    const auto& regs{gpu.Maxwell3D().regs};
+
+    UpdateViewport(regs.viewport_transform);
+    UpdateScissor(regs.scissor_test);
+    UpdateDepthBias(regs.polygon_offset_units, regs.polygon_offset_clamp, regs.polygon_offset_factor);
+    UpdateBlendConstants(regs.blend_color);
+    UpdateStencilFaceMask(regs.stencil_front_func_mask, regs.stencil_back_func_mask);
+}
+
+void OptimizedRasterizer::DrawIndexed(u32 instance_count) {
+    const auto& draw_state{gpu.Maxwell3D().draw_manager->GetDrawState()};
+    const auto& index_buffer{memory_manager.ReadBlockUnsafe(draw_state.index_buffer.Address(),
+                                                            draw_state.index_buffer.size)};
+
+    shader_cache.BindComputeShader();
+    shader_cache.BindGraphicsShader();
+
+    DrawElementsInstanced(draw_state.topology, draw_state.index_buffer.count,
+                          draw_state.index_buffer.format, index_buffer.data(), instance_count);
+}
+
+void OptimizedRasterizer::DrawArrays(u32 instance_count) {
+    const auto& draw_state{gpu.Maxwell3D().draw_manager->GetDrawState()};
+
+    shader_cache.BindComputeShader();
+    shader_cache.BindGraphicsShader();
+
+    DrawArraysInstanced(draw_state.topology, draw_state.vertex_buffer.first,
+                        draw_state.vertex_buffer.count, instance_count);
+}
+
+void OptimizedRasterizer::ClearFramebuffer(u32 layer_count) {
+    const auto& regs{gpu.Maxwell3D().regs};
+    const auto& clear_state{regs.clear_buffers};
+
+    if (clear_state.R || clear_state.G || clear_state.B || clear_state.A) {
+        ClearColorBuffers(clear_state.R, clear_state.G, clear_state.B, clear_state.A,
+                          regs.clear_color[0], regs.clear_color[1], regs.clear_color[2],
+                          regs.clear_color[3], layer_count);
+    }
+
+    if (clear_state.Z || clear_state.S) {
+        ClearDepthStencilBuffer(clear_state.Z, clear_state.S, regs.clear_depth, regs.clear_stencil,
+                                layer_count);
+    }
+}
+
+void OptimizedRasterizer::PrepareCompute() {
+    shader_cache.BindComputeShader();
+}
+
+void OptimizedRasterizer::LaunchComputeShader() {
+    const auto& launch_desc{gpu.KeplerCompute().launch_description};
+    DispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
+}
+
+} // namespace VideoCore

src/video_core/optimized_rasterizer.h

@@ -0,0 +1,73 @@
+#pragma once
+
+#include <memory>
+#include <vector>
+#include "common/common_types.h"
+#include "video_core/rasterizer_interface.h"
+#include "video_core/engines/maxwell_3d.h"
+
+namespace Core {
+class System;
+}
+
+namespace Tegra {
+class GPU;
+class MemoryManager;
+}
+
+namespace VideoCore {
+
+class ShaderCache;
+class QueryCache;
+
+class OptimizedRasterizer final : public RasterizerInterface {
+public:
+    explicit OptimizedRasterizer(Core::System& system, Tegra::GPU& gpu);
+    ~OptimizedRasterizer() override;
+
+    void Draw(bool is_indexed, u32 instance_count) override;
+    void Clear(u32 layer_count) override;
+    void DispatchCompute() override;
+    void ResetCounter(VideoCommon::QueryType type) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
+    void FlushAll() override;
+    void FlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
+    bool MustFlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
+    RasterizerDownloadArea GetFlushArea(DAddr addr, u64 size) override;
+    void InvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
+    void OnCacheInvalidation(PAddr addr, u64 size) override;
+    bool OnCPUWrite(PAddr addr, u64 size) override;
+    void InvalidateGPUCache() override;
+    void UnmapMemory(DAddr addr, u64 size) override;
+    void ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) override;
+    void FlushAndInvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
+    void WaitForIdle() override;
+    void FragmentBarrier() override;
+    void TiledCacheBarrier() override;
+    void FlushCommands() override;
+    void TickFrame() override;
+
+private:
+    void PrepareRendertarget();
+    void UpdateDynamicState();
+    void DrawIndexed(u32 instance_count);
+    void DrawArrays(u32 instance_count);
+    void ClearFramebuffer(u32 layer_count);
+    void PrepareCompute();
+    void LaunchComputeShader();
+
+    Core::System& system;
+    Tegra::GPU& gpu;
+    Tegra::MemoryManager& memory_manager;
+
+    std::unique_ptr<ShaderCache> shader_cache;
+    std::unique_ptr<QueryCache> query_cache;
+
+    std::vector<RenderTargetConfig> render_targets;
+    DepthStencilConfig depth_stencil;
+
+    // Add any additional member variables needed for the optimized rasterizer
+};
+
+} // namespace VideoCore

src/video_core/shader_cache.cpp

@@ -3,9 +3,18 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
+#include <filesystem>
+#include <fstream>
+#include <mutex>
+#include <thread>
 #include <vector>
 
 #include "common/assert.h"
+#include "common/fs/file.h"
+#include "common/fs/path_util.h"
+#include "common/logging/log.h"
+#include "common/thread_worker.h"
 #include "shader_recompiler/frontend/maxwell/control_flow.h"
 #include "shader_recompiler/object_pool.h"
 #include "video_core/control/channel_state.h"
@@ -19,233 +28,288 @@
 
 namespace VideoCommon {
 
-void ShaderCache::InvalidateRegion(VAddr addr, size_t size) {
-    std::scoped_lock lock{invalidation_mutex};
-    InvalidatePagesInRegion(addr, size);
-    RemovePendingShaders();
-}
+constexpr size_t MAX_SHADER_CACHE_SIZE = 1024 * 1024 * 1024; // 1GB
 
-void ShaderCache::OnCacheInvalidation(VAddr addr, size_t size) {
-    std::scoped_lock lock{invalidation_mutex};
-    InvalidatePagesInRegion(addr, size);
-}
+class ShaderCacheWorker : public Common::ThreadWorker {
+public:
+    explicit ShaderCacheWorker(const std::string& name) : ThreadWorker(name) {}
+    ~ShaderCacheWorker() = default;
 
-void ShaderCache::SyncGuestHost() {
-    std::scoped_lock lock{invalidation_mutex};
-    RemovePendingShaders();
-}
+    void CompileShader(ShaderInfo* shader) {
+        Push([shader]() {
+            // Compile shader here
+            // This is a placeholder for the actual compilation process
+            std::this_thread::sleep_for(std::chrono::milliseconds(10));
+            shader->is_compiled.store(true, std::memory_order_release);
+        });
+    }
+};
 
-ShaderCache::ShaderCache(Tegra::MaxwellDeviceMemoryManager& device_memory_)
-    : device_memory{device_memory_} {}
+class ShaderCache::Impl {
+public:
+    explicit Impl(Tegra::MaxwellDeviceMemoryManager& device_memory_)
+        : device_memory{device_memory_}, workers{CreateWorkers()} {
+        LoadCache();
+    }
 
-bool ShaderCache::RefreshStages(std::array<u64, 6>& unique_hashes) {
-    auto& dirty{maxwell3d->dirty.flags};
-    if (!dirty[VideoCommon::Dirty::Shaders]) {
-        return last_shaders_valid;
+    ~Impl() {
+        SaveCache();
     }
-    dirty[VideoCommon::Dirty::Shaders] = false;
 
-    const GPUVAddr base_addr{maxwell3d->regs.program_region.Address()};
-    for (size_t index = 0; index < Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram; ++index) {
-        if (!maxwell3d->regs.IsShaderConfigEnabled(index)) {
-            unique_hashes[index] = 0;
-            continue;
-        }
-        const auto& shader_config{maxwell3d->regs.pipelines[index]};
-        const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderType>(index)};
-        if (program == Tegra::Engines::Maxwell3D::Regs::ShaderType::Pixel &&
-            !maxwell3d->regs.rasterize_enable) {
-            unique_hashes[index] = 0;
-            continue;
+    void InvalidateRegion(VAddr addr, size_t size) {
+        std::scoped_lock lock{invalidation_mutex};
+        InvalidatePagesInRegion(addr, size);
+        RemovePendingShaders();
+    }
+
+    void OnCacheInvalidation(VAddr addr, size_t size) {
+        std::scoped_lock lock{invalidation_mutex};
+        InvalidatePagesInRegion(addr, size);
+    }
+
+    void SyncGuestHost() {
+        std::scoped_lock lock{invalidation_mutex};
+        RemovePendingShaders();
+    }
+
+    bool RefreshStages(std::array<u64, 6>& unique_hashes);
+    const ShaderInfo* ComputeShader();
+    void GetGraphicsEnvironments(GraphicsEnvironments& result, const std::array<u64, NUM_PROGRAMS>& unique_hashes);
+
+    ShaderInfo* TryGet(VAddr addr) const {
+        std::scoped_lock lock{lookup_mutex};
+
+        const auto it = lookup_cache.find(addr);
+        if (it == lookup_cache.end()) {
+            return nullptr;
         }
-        const GPUVAddr shader_addr{base_addr + shader_config.offset};
-        const std::optional<VAddr> cpu_shader_addr{gpu_memory->GpuToCpuAddress(shader_addr)};
-        if (!cpu_shader_addr) {
-            LOG_ERROR(HW_GPU, "Invalid GPU address for shader 0x{:016x}", shader_addr);
-            last_shaders_valid = false;
-            return false;
+        return it->second->data;
+    }
+
+    void Register(std::unique_ptr<ShaderInfo> data, VAddr addr, size_t size) {
+        std::scoped_lock lock{invalidation_mutex, lookup_mutex};
+
+        const VAddr addr_end = addr + size;
+        Entry* const entry = NewEntry(addr, addr_end, data.get());
+
+        const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
+        for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
+            invalidation_cache[page].push_back(entry);
         }
-        const ShaderInfo* shader_info{TryGet(*cpu_shader_addr)};
-        if (!shader_info) {
-            const u32 start_address{shader_config.offset};
-            GraphicsEnvironment env{*maxwell3d, *gpu_memory, program, base_addr, start_address};
-            shader_info = MakeShaderInfo(env, *cpu_shader_addr);
+
+        storage.push_back(std::move(data));
+
+        device_memory.UpdatePagesCachedCount(addr, size, 1);
+    }
+
+private:
+    std::vector<std::unique_ptr<ShaderCacheWorker>> CreateWorkers() {
+        const size_t num_workers = std::thread::hardware_concurrency();
+        std::vector<std::unique_ptr<ShaderCacheWorker>> workers;
+        workers.reserve(num_workers);
+        for (size_t i = 0; i < num_workers; ++i) {
+            workers.emplace_back(std::make_unique<ShaderCacheWorker>(fmt::format("ShaderWorker{}", i)));
         }
-        shader_infos[index] = shader_info;
-        unique_hashes[index] = shader_info->unique_hash;
+        return workers;
     }
-    last_shaders_valid = true;
-    return true;
-}
 
-const ShaderInfo* ShaderCache::ComputeShader() {
-    const GPUVAddr program_base{kepler_compute->regs.code_loc.Address()};
-    const auto& qmd{kepler_compute->launch_description};
-    const GPUVAddr shader_addr{program_base + qmd.program_start};
-    const std::optional<VAddr> cpu_shader_addr{gpu_memory->GpuToCpuAddress(shader_addr)};
-    if (!cpu_shader_addr) {
-        LOG_ERROR(HW_GPU, "Invalid GPU address for shader 0x{:016x}", shader_addr);
-        return nullptr;
-    }
-    if (const ShaderInfo* const shader = TryGet(*cpu_shader_addr)) {
-        return shader;
-    }
-    ComputeEnvironment env{*kepler_compute, *gpu_memory, program_base, qmd.program_start};
-    return MakeShaderInfo(env, *cpu_shader_addr);
-}
+    void LoadCache() {
+        const auto cache_dir = Common::FS::GetSuyuPath(Common::FS::SuyuPath::ShaderDir);
+        std::filesystem::create_directories(cache_dir);
 
-void ShaderCache::GetGraphicsEnvironments(GraphicsEnvironments& result,
-                                          const std::array<u64, NUM_PROGRAMS>& unique_hashes) {
-    size_t env_index{};
-    const GPUVAddr base_addr{maxwell3d->regs.program_region.Address()};
-    for (size_t index = 0; index < NUM_PROGRAMS; ++index) {
-        if (unique_hashes[index] == 0) {
-            continue;
+        const auto cache_file = cache_dir / "shader_cache.bin";
+        if (!std::filesystem::exists(cache_file)) {
+            return;
         }
-        const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderType>(index)};
-        auto& env{result.envs[index]};
-        const u32 start_address{maxwell3d->regs.pipelines[index].offset};
-        env = GraphicsEnvironment{*maxwell3d, *gpu_memory, program, base_addr, start_address};
-        env.SetCachedSize(shader_infos[index]->size_bytes);
-        result.env_ptrs[env_index++] = &env;
-    }
-}
 
-ShaderInfo* ShaderCache::TryGet(VAddr addr) const {
-    std::scoped_lock lock{lookup_mutex};
+        std::ifstream file(cache_file, std::ios::binary);
+        if (!file) {
+            LOG_ERROR(Render_Vulkan, "Failed to open shader cache file for reading");
+            return;
+        }
 
-    const auto it = lookup_cache.find(addr);
-    if (it == lookup_cache.end()) {
-        return nullptr;
-    }
-    return it->second->data;
-}
+        size_t num_entries;
+        file.read(reinterpret_cast<char*>(&num_entries), sizeof(num_entries));
 
-void ShaderCache::Register(std::unique_ptr<ShaderInfo> data, VAddr addr, size_t size) {
-    std::scoped_lock lock{invalidation_mutex, lookup_mutex};
+        for (size_t i = 0; i < num_entries; ++i) {
+            VAddr addr;
+            size_t size;
+            file.read(reinterpret_cast<char*>(&addr), sizeof(addr));
+            file.read(reinterpret_cast<char*>(&size), sizeof(size));
 
-    const VAddr addr_end = addr + size;
-    Entry* const entry = NewEntry(addr, addr_end, data.get());
+            auto info = std::make_unique<ShaderInfo>();
+            file.read(reinterpret_cast<char*>(info.get()), sizeof(ShaderInfo));
 
-    const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
-    for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
-        invalidation_cache[page].push_back(entry);
+            Register(std::move(info), addr, size);
+        }
     }
 
-    storage.push_back(std::move(data));
+    void SaveCache() {
+        const auto cache_dir = Common::FS::GetSuyuPath(Common::FS::SuyuPath::ShaderDir);
+        std::filesystem::create_directories(cache_dir);
 
-    device_memory.UpdatePagesCachedCount(addr, size, 1);
-}
+        const auto cache_file = cache_dir / "shader_cache.bin";
+        std::ofstream file(cache_file, std::ios::binary | std::ios::trunc);
+        if (!file) {
+            LOG_ERROR(Render_Vulkan, "Failed to open shader cache file for writing");
+            return;
+        }
+
+        const size_t num_entries = storage.size();
+        file.write(reinterpret_cast<const char*>(&num_entries), sizeof(num_entries));
 
-void ShaderCache::InvalidatePagesInRegion(VAddr addr, size_t size) {
-    const VAddr addr_end = addr + size;
-    const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
-    for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
-        auto it = invalidation_cache.find(page);
-        if (it == invalidation_cache.end()) {
-            continue;
+        for (const auto& shader : storage) {
+            const VAddr addr = shader->addr;
+            const size_t size = shader->size_bytes;
+            file.write(reinterpret_cast<const char*>(&addr), sizeof(addr));
+            file.write(reinterpret_cast<const char*>(&size), sizeof(size));
+            file.write(reinterpret_cast<const char*>(shader.get()), sizeof(ShaderInfo));
         }
-        InvalidatePageEntries(it->second, addr, addr_end);
     }
-}
 
-void ShaderCache::RemovePendingShaders() {
-    if (marked_for_removal.empty()) {
-        return;
+    void InvalidatePagesInRegion(VAddr addr, size_t size) {
+        const VAddr addr_end = addr + size;
+        const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
+        for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
+            auto it = invalidation_cache.find(page);
+            if (it == invalidation_cache.end()) {
+                continue;
+            }
+            InvalidatePageEntries(it->second, addr, addr_end);
+        }
     }
-    // Remove duplicates
-    std::ranges::sort(marked_for_removal);
-    marked_for_removal.erase(std::unique(marked_for_removal.begin(), marked_for_removal.end()),
-                             marked_for_removal.end());
 
-    boost::container::small_vector<ShaderInfo*, 16> removed_shaders;
+    void RemovePendingShaders() {
+        if (marked_for_removal.empty()) {
+            return;
+        }
+        // Remove duplicates
+        std::sort(marked_for_removal.begin(), marked_for_removal.end());
+        marked_for_removal.erase(std::unique(marked_for_removal.begin(), marked_for_removal.end()),
+                                 marked_for_removal.end());
+
+        std::vector<ShaderInfo*> removed_shaders;
+
+        std::scoped_lock lock{lookup_mutex};
+        for (Entry* const entry : marked_for_removal) {
+            removed_shaders.push_back(entry->data);
+
+            const auto it = lookup_cache.find(entry->addr_start);
+            ASSERT(it != lookup_cache.end());
+            lookup_cache.erase(it);
+        }
+        marked_for_removal.clear();
 
-    std::scoped_lock lock{lookup_mutex};
-    for (Entry* const entry : marked_for_removal) {
-        removed_shaders.push_back(entry->data);
+        if (!removed_shaders.empty()) {
+            RemoveShadersFromStorage(removed_shaders);
+        }
+    }
 
-        const auto it = lookup_cache.find(entry->addr_start);
-        ASSERT(it != lookup_cache.end());
-        lookup_cache.erase(it);
+    void InvalidatePageEntries(std::vector<Entry*>& entries, VAddr addr, VAddr addr_end) {
+        size_t index = 0;
+        while (index < entries.size()) {
+            Entry* const entry = entries[index];
+            if (!entry->Overlaps(addr, addr_end)) {
+                ++index;
+                continue;
+            }
+
+            UnmarkMemory(entry);
+            RemoveEntryFromInvalidationCache(entry);
+            marked_for_removal.push_back(entry);
+        }
     }
-    marked_for_removal.clear();
 
-    if (!removed_shaders.empty()) {
-        RemoveShadersFromStorage(removed_shaders);
+    void RemoveEntryFromInvalidationCache(const Entry* entry) {
+        const u64 page_end = (entry->addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
+        for (u64 page = entry->addr_start >> SUYU_PAGEBITS; page < page_end; ++page) {
+            const auto entries_it = invalidation_cache.find(page);
+            ASSERT(entries_it != invalidation_cache.end());
+            std::vector<Entry*>& entries = entries_it->second;
+
+            const auto entry_it = std::find(entries.begin(), entries.end(), entry);
+            ASSERT(entry_it != entries.end());
+            entries.erase(entry_it);
+        }
     }
-}
 
-void ShaderCache::InvalidatePageEntries(std::vector<Entry*>& entries, VAddr addr, VAddr addr_end) {
-    size_t index = 0;
-    while (index < entries.size()) {
-        Entry* const entry = entries[index];
-        if (!entry->Overlaps(addr, addr_end)) {
-            ++index;
-            continue;
+    void UnmarkMemory(Entry* entry) {
+        if (!entry->is_memory_marked) {
+            return;
         }
+        entry->is_memory_marked = false;
 
-        UnmarkMemory(entry);
-        RemoveEntryFromInvalidationCache(entry);
-        marked_for_removal.push_back(entry);
+        const VAddr addr = entry->addr_start;
+        const size_t size = entry->addr_end - addr;
+        device_memory.UpdatePagesCachedCount(addr, size, -1);
     }
-}
 
-void ShaderCache::RemoveEntryFromInvalidationCache(const Entry* entry) {
-    const u64 page_end = (entry->addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
-    for (u64 page = entry->addr_start >> SUYU_PAGEBITS; page < page_end; ++page) {
-        const auto entries_it = invalidation_cache.find(page);
-        ASSERT(entries_it != invalidation_cache.end());
-        std::vector<Entry*>& entries = entries_it->second;
+    void RemoveShadersFromStorage(const std::vector<ShaderInfo*>& removed_shaders) {
+        storage.erase(
+            std::remove_if(storage.begin(), storage.end(),
+                           [&removed_shaders](const std::unique_ptr<ShaderInfo>& shader) {
+                               return std::find(removed_shaders.begin(), removed_shaders.end(),
+                                                shader.get()) != removed_shaders.end();
+                           }),
+            storage.end());
+    }
 
-        const auto entry_it = std::ranges::find(entries, entry);
-        ASSERT(entry_it != entries.end());
-        entries.erase(entry_it);
+    Entry* NewEntry(VAddr addr, VAddr addr_end, ShaderInfo* data) {
+        auto entry = std::make_unique<Entry>(Entry{addr, addr_end, data});
+        Entry* const entry_pointer = entry.get();
+
+        lookup_cache.emplace(addr, std::move(entry));
+        return entry_pointer;
     }
+
+    Tegra::MaxwellDeviceMemoryManager& device_memory;
+    std::vector<std::unique_ptr<ShaderCacheWorker>> workers;
+
+    mutable std::mutex lookup_mutex;
+    std::mutex invalidation_mutex;
+
+    std::unordered_map<VAddr, std::unique_ptr<Entry>> lookup_cache;
+    std::unordered_map<u64, std::vector<Entry*>> invalidation_cache;
+    std::vector<std::unique_ptr<ShaderInfo>> storage;
+    std::vector<Entry*> marked_for_removal;
+};
+
+ShaderCache::ShaderCache(Tegra::MaxwellDeviceMemoryManager& device_memory_)
+    : impl{std::make_unique<Impl>(device_memory_)} {}
+
+ShaderCache::~ShaderCache() = default;
+
+void ShaderCache::InvalidateRegion(VAddr addr, size_t size) {
+    impl->InvalidateRegion(addr, size);
 }
 
-void ShaderCache::UnmarkMemory(Entry* entry) {
-    if (!entry->is_memory_marked) {
-        return;
-    }
-    entry->is_memory_marked = false;
+void ShaderCache::OnCacheInvalidation(VAddr addr, size_t size) {
+    impl->OnCacheInvalidation(addr, size);
+}
 
-    const VAddr addr = entry->addr_start;
-    const size_t size = entry->addr_end - addr;
-    device_memory.UpdatePagesCachedCount(addr, size, -1);
+void ShaderCache::SyncGuestHost() {
+    impl->SyncGuestHost();
 }
 
-void ShaderCache::RemoveShadersFromStorage(std::span<ShaderInfo*> removed_shaders) {
-    // Remove them from the cache
-    std::erase_if(storage, [&removed_shaders](const std::unique_ptr<ShaderInfo>& shader) {
-        return std::ranges::find(removed_shaders, shader.get()) != removed_shaders.end();
-    });
+bool ShaderCache::RefreshStages(std::array<u64, 6>& unique_hashes) {
+    return impl->RefreshStages(unique_hashes);
+}
+
+const ShaderInfo* ShaderCache::ComputeShader() {
+    return impl->ComputeShader();
 }
 
-ShaderCache::Entry* ShaderCache::NewEntry(VAddr addr, VAddr addr_end, ShaderInfo* data) {
-    auto entry = std::make_unique<Entry>(Entry{addr, addr_end, data});
-    Entry* const entry_pointer = entry.get();
+void ShaderCache::GetGraphicsEnvironments(GraphicsEnvironments& result,
+                                          const std::array<u64, NUM_PROGRAMS>& unique_hashes) {
+    impl->GetGraphicsEnvironments(result, unique_hashes);
+}
 
-    lookup_cache.emplace(addr, std::move(entry));
-    return entry_pointer;
+ShaderInfo* ShaderCache::TryGet(VAddr addr) const {
+    return impl->TryGet(addr);
 }
 
-const ShaderInfo* ShaderCache::MakeShaderInfo(GenericEnvironment& env, VAddr cpu_addr) {
-    auto info = std::make_unique<ShaderInfo>();
-    if (const std::optional<u64> cached_hash{env.Analyze()}) {
-        info->unique_hash = *cached_hash;
-        info->size_bytes = env.CachedSizeBytes();
-    } else {
-        // Slow path, not really hit on commercial games
-        // Build a control flow graph to get the real shader size
-        Shader::ObjectPool<Shader::Maxwell::Flow::Block> flow_block;
-        Shader::Maxwell::Flow::CFG cfg{env, flow_block, env.StartAddress()};
-        info->unique_hash = env.CalculateHash();
-        info->size_bytes = env.ReadSizeBytes();
-    }
-    const size_t size_bytes{info->size_bytes};
-    const ShaderInfo* const result{info.get()};
-    Register(std::move(info), cpu_addr, size_bytes);
-    return result;
+void ShaderCache::Register(std::unique_ptr<ShaderInfo> data, VAddr addr, size_t size) {
+    impl->Register(std::move(data), addr, size);
 }
 
 } // namespace VideoCommon