Merge pull request #3677 from FernandoS27/better-sync

Introduce Predictive Flushing and Improve ASYNC GPU

src/core/settings.cpp

@@ -92,7 +92,7 @@ void LogSettings() {
     LogSetting("Renderer_UseFrameLimit", Settings::values.use_frame_limit);
     LogSetting("Renderer_FrameLimit", Settings::values.frame_limit);
     LogSetting("Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache);
-    LogSetting("Renderer_UseAccurateGpuEmulation", Settings::values.use_accurate_gpu_emulation);
+    LogSetting("Renderer_GPUAccuracyLevel", Settings::values.gpu_accuracy);
     LogSetting("Renderer_UseAsynchronousGpuEmulation",
                Settings::values.use_asynchronous_gpu_emulation);
     LogSetting("Renderer_UseVsync", Settings::values.use_vsync);
@@ -109,4 +109,12 @@ void LogSettings() {
     LogSetting("Services_BCATBoxcatLocal", Settings::values.bcat_boxcat_local);
 }
 
+bool IsGPULevelExtreme() {
+    return values.gpu_accuracy == GPUAccuracy::Extreme;
+}
+
+bool IsGPULevelHigh() {
+    return values.gpu_accuracy == GPUAccuracy::Extreme || values.gpu_accuracy == GPUAccuracy::High;
+}
+
 } // namespace Settings

src/core/settings.h

@@ -376,6 +376,12 @@ enum class RendererBackend {
     Vulkan = 1,
 };
 
+enum class GPUAccuracy : u32 {
+    Normal = 0,
+    High = 1,
+    Extreme = 2,
+};
+
 struct Values {
     // System
     bool use_docked_mode;
@@ -436,7 +442,7 @@ struct Values {
     bool use_frame_limit;
     u16 frame_limit;
     bool use_disk_shader_cache;
-    bool use_accurate_gpu_emulation;
+    GPUAccuracy gpu_accuracy;
     bool use_asynchronous_gpu_emulation;
     bool use_vsync;
     bool force_30fps_mode;
@@ -480,6 +486,9 @@ struct Values {
     std::map<u64, std::vector<std::string>> disabled_addons;
 } extern values;
 
+bool IsGPULevelExtreme();
+bool IsGPULevelHigh();
+
 void Apply();
 void LogSettings();
 } // namespace Settings

src/core/telemetry_session.cpp

@@ -56,6 +56,18 @@ static const char* TranslateRenderer(Settings::RendererBackend backend) {
     return "Unknown";
 }
 
+static const char* TranslateGPUAccuracyLevel(Settings::GPUAccuracy backend) {
+    switch (backend) {
+    case Settings::GPUAccuracy::Normal:
+        return "Normal";
+    case Settings::GPUAccuracy::High:
+        return "High";
+    case Settings::GPUAccuracy::Extreme:
+        return "Extreme";
+    }
+    return "Unknown";
+}
+
 u64 GetTelemetryId() {
     u64 telemetry_id{};
     const std::string filename{FileUtil::GetUserPath(FileUtil::UserPath::ConfigDir) +
@@ -184,8 +196,8 @@ void TelemetrySession::AddInitialInfo(Loader::AppLoader& app_loader) {
     AddField(field_type, "Renderer_UseFrameLimit", Settings::values.use_frame_limit);
     AddField(field_type, "Renderer_FrameLimit", Settings::values.frame_limit);
     AddField(field_type, "Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache);
-    AddField(field_type, "Renderer_UseAccurateGpuEmulation",
-             Settings::values.use_accurate_gpu_emulation);
+    AddField(field_type, "Renderer_GPUAccuracyLevel",
+             TranslateGPUAccuracyLevel(Settings::values.gpu_accuracy));
     AddField(field_type, "Renderer_UseAsynchronousGpuEmulation",
              Settings::values.use_asynchronous_gpu_emulation);
     AddField(field_type, "Renderer_UseVsync", Settings::values.use_vsync);

src/video_core/CMakeLists.txt

@@ -23,6 +23,7 @@ add_library(video_core STATIC
     engines/shader_bytecode.h
     engines/shader_header.h
     engines/shader_type.h
+    fence_manager.h
     gpu.cpp
     gpu.h
     gpu_asynch.cpp
@@ -51,6 +52,8 @@ add_library(video_core STATIC
     renderer_opengl/gl_buffer_cache.h
     renderer_opengl/gl_device.cpp
     renderer_opengl/gl_device.h
+    renderer_opengl/gl_fence_manager.cpp
+    renderer_opengl/gl_fence_manager.h
     renderer_opengl/gl_framebuffer_cache.cpp
     renderer_opengl/gl_framebuffer_cache.h
     renderer_opengl/gl_rasterizer.cpp
@@ -176,6 +179,8 @@ if (ENABLE_VULKAN)
         renderer_vulkan/vk_descriptor_pool.h
         renderer_vulkan/vk_device.cpp
         renderer_vulkan/vk_device.h
+        renderer_vulkan/vk_fence_manager.cpp
+        renderer_vulkan/vk_fence_manager.h
         renderer_vulkan/vk_graphics_pipeline.cpp
         renderer_vulkan/vk_graphics_pipeline.h
         renderer_vulkan/vk_image.cpp

src/video_core/buffer_cache/buffer_cache.h

@@ -5,6 +5,7 @@
 #pragma once
 
 #include <array>
+#include <list>
 #include <memory>
 #include <mutex>
 #include <unordered_map>
@@ -18,8 +19,10 @@
 
 #include "common/alignment.h"
 #include "common/common_types.h"
+#include "common/logging/log.h"
 #include "core/core.h"
 #include "core/memory.h"
+#include "core/settings.h"
 #include "video_core/buffer_cache/buffer_block.h"
 #include "video_core/buffer_cache/map_interval.h"
 #include "video_core/memory_manager.h"
@@ -79,6 +82,9 @@ public:
         auto map = MapAddress(block, gpu_addr, cpu_addr, size);
         if (is_written) {
             map->MarkAsModified(true, GetModifiedTicks());
+            if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
+                MarkForAsyncFlush(map);
+            }
             if (!map->IsWritten()) {
                 map->MarkAsWritten(true);
                 MarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
@@ -137,11 +143,22 @@ public:
         });
         for (auto& object : objects) {
             if (object->IsModified() && object->IsRegistered()) {
+                mutex.unlock();
                 FlushMap(object);
+                mutex.lock();
             }
         }
     }
 
+    bool MustFlushRegion(VAddr addr, std::size_t size) {
+        std::lock_guard lock{mutex};
+
+        const std::vector<MapInterval> objects = GetMapsInRange(addr, size);
+        return std::any_of(objects.cbegin(), objects.cend(), [](const MapInterval& map) {
+            return map->IsModified() && map->IsRegistered();
+        });
+    }
+
     /// Mark the specified region as being invalidated
     void InvalidateRegion(VAddr addr, u64 size) {
         std::lock_guard lock{mutex};
@@ -154,6 +171,77 @@ public:
         }
     }
 
+    void OnCPUWrite(VAddr addr, std::size_t size) {
+        std::lock_guard lock{mutex};
+
+        for (const auto& object : GetMapsInRange(addr, size)) {
+            if (object->IsMemoryMarked() && object->IsRegistered()) {
+                UnmarkMemory(object);
+                object->SetSyncPending(true);
+                marked_for_unregister.emplace_back(object);
+            }
+        }
+    }
+
+    void SyncGuestHost() {
+        std::lock_guard lock{mutex};
+
+        for (const auto& object : marked_for_unregister) {
+            if (object->IsRegistered()) {
+                object->SetSyncPending(false);
+                Unregister(object);
+            }
+        }
+        marked_for_unregister.clear();
+    }
+
+    void CommitAsyncFlushes() {
+        if (uncommitted_flushes) {
+            auto commit_list = std::make_shared<std::list<MapInterval>>();
+            for (auto& map : *uncommitted_flushes) {
+                if (map->IsRegistered() && map->IsModified()) {
+                    // TODO(Blinkhawk): Implement backend asynchronous flushing
+                    // AsyncFlushMap(map)
+                    commit_list->push_back(map);
+                }
+            }
+            if (!commit_list->empty()) {
+                committed_flushes.push_back(commit_list);
+            } else {
+                committed_flushes.emplace_back();
+            }
+        } else {
+            committed_flushes.emplace_back();
+        }
+        uncommitted_flushes.reset();
+    }
+
+    bool ShouldWaitAsyncFlushes() const {
+        return !committed_flushes.empty() && committed_flushes.front() != nullptr;
+    }
+
+    bool HasUncommittedFlushes() const {
+        return uncommitted_flushes != nullptr;
+    }
+
+    void PopAsyncFlushes() {
+        if (committed_flushes.empty()) {
+            return;
+        }
+        auto& flush_list = committed_flushes.front();
+        if (!flush_list) {
+            committed_flushes.pop_front();
+            return;
+        }
+        for (MapInterval& map : *flush_list) {
+            if (map->IsRegistered()) {
+                // TODO(Blinkhawk): Replace this for reading the asynchronous flush
+                FlushMap(map);
+            }
+        }
+        committed_flushes.pop_front();
+    }
+
     virtual BufferType GetEmptyBuffer(std::size_t size) = 0;
 
 protected:
@@ -196,17 +284,30 @@ protected:
         const IntervalType interval{new_map->GetStart(), new_map->GetEnd()};
         mapped_addresses.insert({interval, new_map});
         rasterizer.UpdatePagesCachedCount(cpu_addr, size, 1);
+        new_map->SetMemoryMarked(true);
         if (inherit_written) {
             MarkRegionAsWritten(new_map->GetStart(), new_map->GetEnd() - 1);
             new_map->MarkAsWritten(true);
         }
     }
 
-    /// Unregisters an object from the cache
-    void Unregister(MapInterval& map) {
+    void UnmarkMemory(const MapInterval& map) {
+        if (!map->IsMemoryMarked()) {
+            return;
+        }
         const std::size_t size = map->GetEnd() - map->GetStart();
         rasterizer.UpdatePagesCachedCount(map->GetStart(), size, -1);
+        map->SetMemoryMarked(false);
+    }
+
+    /// Unregisters an object from the cache
+    void Unregister(const MapInterval& map) {
+        UnmarkMemory(map);
         map->MarkAsRegistered(false);
+        if (map->IsSyncPending()) {
+            marked_for_unregister.remove(map);
+            map->SetSyncPending(false);
+        }
         if (map->IsWritten()) {
             UnmarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
         }
@@ -264,6 +365,9 @@ private:
         MapInterval new_map = CreateMap(new_start, new_end, new_gpu_addr);
         if (modified_inheritance) {
             new_map->MarkAsModified(true, GetModifiedTicks());
+            if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
+                MarkForAsyncFlush(new_map);
+            }
         }
         Register(new_map, write_inheritance);
         return new_map;
@@ -450,6 +554,13 @@ private:
         return false;
     }
 
+    void MarkForAsyncFlush(MapInterval& map) {
+        if (!uncommitted_flushes) {
+            uncommitted_flushes = std::make_shared<std::unordered_set<MapInterval>>();
+        }
+        uncommitted_flushes->insert(map);
+    }
+
     VideoCore::RasterizerInterface& rasterizer;
     Core::System& system;
 
@@ -479,6 +590,10 @@ private:
     u64 modified_ticks = 0;
 
     std::vector<u8> staging_buffer;
+    std::list<MapInterval> marked_for_unregister;
+
+    std::shared_ptr<std::unordered_set<MapInterval>> uncommitted_flushes{};
+    std::list<std::shared_ptr<std::list<MapInterval>>> committed_flushes;
 
     std::recursive_mutex mutex;
 };

src/video_core/buffer_cache/map_interval.h

@@ -46,6 +46,22 @@ public:
         return is_registered;
     }
 
+    void SetMemoryMarked(bool is_memory_marked_) {
+        is_memory_marked = is_memory_marked_;
+    }
+
+    bool IsMemoryMarked() const {
+        return is_memory_marked;
+    }
+
+    void SetSyncPending(bool is_sync_pending_) {
+        is_sync_pending = is_sync_pending_;
+    }
+
+    bool IsSyncPending() const {
+        return is_sync_pending;
+    }
+
     VAddr GetStart() const {
         return start;
     }
@@ -83,6 +99,8 @@ private:
     bool is_written{};
     bool is_modified{};
     bool is_registered{};
+    bool is_memory_marked{};
+    bool is_sync_pending{};
     u64 ticks{};
 };
 

src/video_core/dma_pusher.cpp

@@ -21,6 +21,7 @@ MICROPROFILE_DEFINE(DispatchCalls, "GPU", "Execute command buffer", MP_RGB(128,
 void DmaPusher::DispatchCalls() {
     MICROPROFILE_SCOPE(DispatchCalls);
 
+    gpu.SyncGuestHost();
     // On entering GPU code, assume all memory may be touched by the ARM core.
     gpu.Maxwell3D().OnMemoryWrite();
 
@@ -32,6 +33,8 @@ void DmaPusher::DispatchCalls() {
         }
     }
     gpu.FlushCommands();
+    gpu.SyncGuestHost();
+    gpu.OnCommandListEnd();
 }
 
 bool DmaPusher::Step() {

src/video_core/engines/maxwell_3d.cpp

@@ -404,7 +404,11 @@ void Maxwell3D::ProcessQueryGet() {
 
     switch (regs.query.query_get.operation) {
     case Regs::QueryOperation::Release:
-        StampQueryResult(regs.query.query_sequence, regs.query.query_get.short_query == 0);
+        if (regs.query.query_get.fence == 1) {
+            rasterizer.SignalSemaphore(regs.query.QueryAddress(), regs.query.query_sequence);
+        } else {
+            StampQueryResult(regs.query.query_sequence, regs.query.query_get.short_query == 0);
+        }
         break;
     case Regs::QueryOperation::Acquire:
         // TODO(Blinkhawk): Under this operation, the GPU waits for the CPU to write a value that
@@ -483,7 +487,7 @@ void Maxwell3D::ProcessSyncPoint() {
     const u32 increment = regs.sync_info.increment.Value();
     [[maybe_unused]] const u32 cache_flush = regs.sync_info.unknown.Value();
     if (increment) {
-        system.GPU().IncrementSyncPoint(sync_point);
+        rasterizer.SignalSyncPoint(sync_point);
     }
 }
 

src/video_core/engines/maxwell_dma.cpp

@@ -104,8 +104,13 @@ void MaxwellDMA::HandleCopy() {
             write_buffer.resize(dst_size);
         }
 
-        memory_manager.ReadBlock(source, read_buffer.data(), src_size);
-        memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+        if (Settings::IsGPULevelExtreme()) {
+            memory_manager.ReadBlock(source, read_buffer.data(), src_size);
+            memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+        } else {
+            memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
+            memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
+        }
 
         Texture::UnswizzleSubrect(
             regs.x_count, regs.y_count, regs.dst_pitch, regs.src_params.size_x, bytes_per_pixel,
@@ -136,7 +141,7 @@ void MaxwellDMA::HandleCopy() {
             write_buffer.resize(dst_size);
         }
 
-        if (Settings::values.use_accurate_gpu_emulation) {
+        if (Settings::IsGPULevelExtreme()) {
             memory_manager.ReadBlock(source, read_buffer.data(), src_size);
             memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
         } else {

src/video_core/fence_manager.h

@@ -0,0 +1,170 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <memory>
+#include <queue>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "core/core.h"
+#include "core/memory.h"
+#include "core/settings.h"
+#include "video_core/gpu.h"
+#include "video_core/memory_manager.h"
+#include "video_core/rasterizer_interface.h"
+
+namespace VideoCommon {
+
+class FenceBase {
+public:
+    FenceBase(u32 payload, bool is_stubbed)
+        : address{}, payload{payload}, is_semaphore{false}, is_stubbed{is_stubbed} {}
+
+    FenceBase(GPUVAddr address, u32 payload, bool is_stubbed)
+        : address{address}, payload{payload}, is_semaphore{true}, is_stubbed{is_stubbed} {}
+
+    GPUVAddr GetAddress() const {
+        return address;
+    }
+
+    u32 GetPayload() const {
+        return payload;
+    }
+
+    bool IsSemaphore() const {
+        return is_semaphore;
+    }
+
+private:
+    GPUVAddr address;
+    u32 payload;
+    bool is_semaphore;
+
+protected:
+    bool is_stubbed;
+};
+
+template <typename TFence, typename TTextureCache, typename TTBufferCache, typename TQueryCache>
+class FenceManager {
+public:
+    void SignalSemaphore(GPUVAddr addr, u32 value) {
+        TryReleasePendingFences();
+        const bool should_flush = ShouldFlush();
+        CommitAsyncFlushes();
+        TFence new_fence = CreateFence(addr, value, !should_flush);
+        fences.push(new_fence);
+        QueueFence(new_fence);
+        if (should_flush) {
+            rasterizer.FlushCommands();
+        }
+        rasterizer.SyncGuestHost();
+    }
+
+    void SignalSyncPoint(u32 value) {
+        TryReleasePendingFences();
+        const bool should_flush = ShouldFlush();
+        CommitAsyncFlushes();
+        TFence new_fence = CreateFence(value, !should_flush);
+        fences.push(new_fence);
+        QueueFence(new_fence);
+        if (should_flush) {
+            rasterizer.FlushCommands();
+        }
+        rasterizer.SyncGuestHost();
+    }
+
+    void WaitPendingFences() {
+        auto& gpu{system.GPU()};
+        auto& memory_manager{gpu.MemoryManager()};
+        while (!fences.empty()) {
+            TFence& current_fence = fences.front();
+            if (ShouldWait()) {
+                WaitFence(current_fence);
+            }
+            PopAsyncFlushes();
+            if (current_fence->IsSemaphore()) {
+                memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+            } else {
+                gpu.IncrementSyncPoint(current_fence->GetPayload());
+            }
+            fences.pop();
+        }
+    }
+
+protected:
+    FenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+                 TTextureCache& texture_cache, TTBufferCache& buffer_cache,
+                 TQueryCache& query_cache)
+        : system{system}, rasterizer{rasterizer}, texture_cache{texture_cache},
+          buffer_cache{buffer_cache}, query_cache{query_cache} {}
+
+    virtual ~FenceManager() {}
+
+    /// Creates a Sync Point Fence Interface, does not create a backend fence if 'is_stubbed' is
+    /// true
+    virtual TFence CreateFence(u32 value, bool is_stubbed) = 0;
+    /// Creates a Semaphore Fence Interface, does not create a backend fence if 'is_stubbed' is true
+    virtual TFence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) = 0;
+    /// Queues a fence into the backend if the fence isn't stubbed.
+    virtual void QueueFence(TFence& fence) = 0;
+    /// Notifies that the backend fence has been signaled/reached in host GPU.
+    virtual bool IsFenceSignaled(TFence& fence) const = 0;
+    /// Waits until a fence has been signalled by the host GPU.
+    virtual void WaitFence(TFence& fence) = 0;
+
+    Core::System& system;
+    VideoCore::RasterizerInterface& rasterizer;
+    TTextureCache& texture_cache;
+    TTBufferCache& buffer_cache;
+    TQueryCache& query_cache;
+
+private:
+    void TryReleasePendingFences() {
+        auto& gpu{system.GPU()};
+        auto& memory_manager{gpu.MemoryManager()};
+        while (!fences.empty()) {
+            TFence& current_fence = fences.front();
+            if (ShouldWait() && !IsFenceSignaled(current_fence)) {
+                return;
+            }
+            PopAsyncFlushes();
+            if (current_fence->IsSemaphore()) {
+                memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+            } else {
+                gpu.IncrementSyncPoint(current_fence->GetPayload());
+            }
+            fences.pop();
+        }
+    }
+
+    bool ShouldWait() const {
+        return texture_cache.ShouldWaitAsyncFlushes() || buffer_cache.ShouldWaitAsyncFlushes() ||
+               query_cache.ShouldWaitAsyncFlushes();
+    }
+
+    bool ShouldFlush() const {
+        return texture_cache.HasUncommittedFlushes() || buffer_cache.HasUncommittedFlushes() ||
+               query_cache.HasUncommittedFlushes();
+    }
+
+    void PopAsyncFlushes() {
+        texture_cache.PopAsyncFlushes();
+        buffer_cache.PopAsyncFlushes();
+        query_cache.PopAsyncFlushes();
+    }
+
+    void CommitAsyncFlushes() {
+        texture_cache.CommitAsyncFlushes();
+        buffer_cache.CommitAsyncFlushes();
+        query_cache.CommitAsyncFlushes();
+    }
+
+    std::queue<TFence> fences;
+};
+
+} // namespace VideoCommon

src/video_core/gpu.cpp

@@ -125,6 +125,28 @@ bool GPU::CancelSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
     return true;
 }
 
+u64 GPU::RequestFlush(VAddr addr, std::size_t size) {
+    std::unique_lock lck{flush_request_mutex};
+    const u64 fence = ++last_flush_fence;
+    flush_requests.emplace_back(fence, addr, size);
+    return fence;
+}
+
+void GPU::TickWork() {
+    std::unique_lock lck{flush_request_mutex};
+    while (!flush_requests.empty()) {
+        auto& request = flush_requests.front();
+        const u64 fence = request.fence;
+        const VAddr addr = request.addr;
+        const std::size_t size = request.size;
+        flush_requests.pop_front();
+        flush_request_mutex.unlock();
+        renderer->Rasterizer().FlushRegion(addr, size);
+        current_flush_fence.store(fence);
+        flush_request_mutex.lock();
+    }
+}
+
 u64 GPU::GetTicks() const {
     // This values were reversed engineered by fincs from NVN
     // The gpu clock is reported in units of 385/625 nanoseconds
@@ -142,6 +164,13 @@ void GPU::FlushCommands() {
     renderer->Rasterizer().FlushCommands();
 }
 
+void GPU::SyncGuestHost() {
+    renderer->Rasterizer().SyncGuestHost();
+}
+
+void GPU::OnCommandListEnd() {
+    renderer->Rasterizer().ReleaseFences();
+}
 // Note that, traditionally, methods are treated as 4-byte addressable locations, and hence
 // their numbers are written down multiplied by 4 in Docs. Here we are not multiply by 4.
 // So the values you see in docs might be multiplied by 4.

src/video_core/gpu.h

@@ -155,7 +155,23 @@ public:
     /// Calls a GPU method.
     void CallMethod(const MethodCall& method_call);
 
+    /// Flush all current written commands into the host GPU for execution.
     void FlushCommands();
+    /// Synchronizes CPU writes with Host GPU memory.
+    void SyncGuestHost();
+    /// Signal the ending of command list.
+    virtual void OnCommandListEnd();
+
+    /// Request a host GPU memory flush from the CPU.
+    u64 RequestFlush(VAddr addr, std::size_t size);
+
+    /// Obtains current flush request fence id.
+    u64 CurrentFlushRequestFence() const {
+        return current_flush_fence.load(std::memory_order_relaxed);
+    }
+
+    /// Tick pending requests within the GPU.
+    void TickWork();
 
     /// Returns a reference to the Maxwell3D GPU engine.
     Engines::Maxwell3D& Maxwell3D();
@@ -325,6 +341,19 @@ private:
 
     std::condition_variable sync_cv;
 
+    struct FlushRequest {
+        FlushRequest(u64 fence, VAddr addr, std::size_t size)
+            : fence{fence}, addr{addr}, size{size} {}
+        u64 fence;
+        VAddr addr;
+        std::size_t size;
+    };
+
+    std::list<FlushRequest> flush_requests;
+    std::atomic<u64> current_flush_fence{};
+    u64 last_flush_fence{};
+    std::mutex flush_request_mutex;
+
     const bool is_async;
 };
 

src/video_core/gpu_asynch.cpp

@@ -52,4 +52,8 @@ void GPUAsynch::WaitIdle() const {
     gpu_thread.WaitIdle();
 }
 
+void GPUAsynch::OnCommandListEnd() {
+    gpu_thread.OnCommandListEnd();
+}
+
 } // namespace VideoCommon

src/video_core/gpu_asynch.h

@@ -32,6 +32,8 @@ public:
     void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
     void WaitIdle() const override;
 
+    void OnCommandListEnd() override;
+
 protected:
     void TriggerCpuInterrupt(u32 syncpoint_id, u32 value) const override;
 

src/video_core/gpu_thread.cpp

@@ -6,6 +6,7 @@
 #include "common/microprofile.h"
 #include "core/core.h"
 #include "core/frontend/emu_window.h"
+#include "core/settings.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/gpu.h"
 #include "video_core/gpu_thread.h"
@@ -14,8 +15,9 @@
 namespace VideoCommon::GPUThread {
 
 /// Runs the GPU thread
-static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::GraphicsContext& context,
-                      Tegra::DmaPusher& dma_pusher, SynchState& state) {
+static void RunThread(Core::System& system, VideoCore::RendererBase& renderer,
+                      Core::Frontend::GraphicsContext& context, Tegra::DmaPusher& dma_pusher,
+                      SynchState& state) {
     MicroProfileOnThreadCreate("GpuThread");
 
     // Wait for first GPU command before acquiring the window context
@@ -37,10 +39,14 @@ static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::Graphic
             dma_pusher.DispatchCalls();
         } else if (const auto data = std::get_if<SwapBuffersCommand>(&next.data)) {
             renderer.SwapBuffers(data->framebuffer ? &*data->framebuffer : nullptr);
+        } else if (const auto data = std::get_if<OnCommandListEndCommand>(&next.data)) {
+            renderer.Rasterizer().ReleaseFences();
+        } else if (const auto data = std::get_if<GPUTickCommand>(&next.data)) {
+            system.GPU().TickWork();
         } else if (const auto data = std::get_if<FlushRegionCommand>(&next.data)) {
             renderer.Rasterizer().FlushRegion(data->addr, data->size);
         } else if (const auto data = std::get_if<InvalidateRegionCommand>(&next.data)) {
-            renderer.Rasterizer().InvalidateRegion(data->addr, data->size);
+            renderer.Rasterizer().OnCPUWrite(data->addr, data->size);
         } else if (std::holds_alternative<EndProcessingCommand>(next.data)) {
             return;
         } else {
@@ -65,8 +71,8 @@ ThreadManager::~ThreadManager() {
 void ThreadManager::StartThread(VideoCore::RendererBase& renderer,
                                 Core::Frontend::GraphicsContext& context,
                                 Tegra::DmaPusher& dma_pusher) {
-    thread = std::thread{RunThread, std::ref(renderer), std::ref(context), std::ref(dma_pusher),
-                         std::ref(state)};
+    thread = std::thread{RunThread,         std::ref(system),     std::ref(renderer),
+                         std::ref(context), std::ref(dma_pusher), std::ref(state)};
 }
 
 void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
@@ -78,16 +84,29 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
 }
 
 void ThreadManager::FlushRegion(VAddr addr, u64 size) {
-    PushCommand(FlushRegionCommand(addr, size));
+    if (!Settings::IsGPULevelHigh()) {
+        PushCommand(FlushRegionCommand(addr, size));
+        return;
+    }
+    if (!Settings::IsGPULevelExtreme()) {
+        return;
+    }
+    if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
+        auto& gpu = system.GPU();
+        u64 fence = gpu.RequestFlush(addr, size);
+        PushCommand(GPUTickCommand());
+        while (fence > gpu.CurrentFlushRequestFence()) {
+        }
+    }
 }
 
 void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
-    system.Renderer().Rasterizer().InvalidateRegion(addr, size);
+    system.Renderer().Rasterizer().OnCPUWrite(addr, size);
 }
 
 void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) {
     // Skip flush on asynch mode, as FlushAndInvalidateRegion is not used for anything too important
-    InvalidateRegion(addr, size);
+    system.Renderer().Rasterizer().OnCPUWrite(addr, size);
 }
 
 void ThreadManager::WaitIdle() const {
@@ -95,6 +114,10 @@ void ThreadManager::WaitIdle() const {
     }
 }
 
+void ThreadManager::OnCommandListEnd() {
+    PushCommand(OnCommandListEndCommand());
+}
+
 u64 ThreadManager::PushCommand(CommandData&& command_data) {
     const u64 fence{++state.last_fence};
     state.queue.Push(CommandDataContainer(std::move(command_data), fence));

src/video_core/gpu_thread.h

@@ -70,9 +70,16 @@ struct FlushAndInvalidateRegionCommand final {
     u64 size;
 };
 
+/// Command called within the gpu, to schedule actions after a command list end
+struct OnCommandListEndCommand final {};
+
+/// Command to make the gpu look into pending requests
+struct GPUTickCommand final {};
+
 using CommandData =
     std::variant<EndProcessingCommand, SubmitListCommand, SwapBuffersCommand, FlushRegionCommand,
-                 InvalidateRegionCommand, FlushAndInvalidateRegionCommand>;
+                 InvalidateRegionCommand, FlushAndInvalidateRegionCommand, OnCommandListEndCommand,
+                 GPUTickCommand>;
 
 struct CommandDataContainer {
     CommandDataContainer() = default;
@@ -122,6 +129,8 @@ public:
     // Wait until the gpu thread is idle.
     void WaitIdle() const;
 
+    void OnCommandListEnd();
+
 private:
     /// Pushes a command to be executed by the GPU thread
     u64 PushCommand(CommandData&& command_data);

src/video_core/query_cache.h

@@ -12,10 +12,12 @@
 #include <mutex>
 #include <optional>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 #include "common/assert.h"
 #include "core/core.h"
+#include "core/settings.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/gpu.h"
 #include "video_core/memory_manager.h"
@@ -130,6 +132,9 @@ public:
         }
 
         query->BindCounter(Stream(type).Current(), timestamp);
+        if (Settings::values.use_asynchronous_gpu_emulation) {
+            AsyncFlushQuery(cpu_addr);
+        }
     }
 
     /// Updates counters from GPU state. Expected to be called once per draw, clear or dispatch.
@@ -170,6 +175,37 @@ public:
         return streams[static_cast<std::size_t>(type)];
     }
 
+    void CommitAsyncFlushes() {
+        committed_flushes.push_back(uncommitted_flushes);
+        uncommitted_flushes.reset();
+    }
+
+    bool HasUncommittedFlushes() const {
+        return uncommitted_flushes != nullptr;
+    }
+
+    bool ShouldWaitAsyncFlushes() const {
+        if (committed_flushes.empty()) {
+            return false;
+        }
+        return committed_flushes.front() != nullptr;
+    }
+
+    void PopAsyncFlushes() {
+        if (committed_flushes.empty()) {
+            return;
+        }
+        auto& flush_list = committed_flushes.front();
+        if (!flush_list) {
+            committed_flushes.pop_front();
+            return;
+        }
+        for (VAddr query_address : *flush_list) {
+            FlushAndRemoveRegion(query_address, 4);
+        }
+        committed_flushes.pop_front();
+    }
+
 protected:
     std::array<QueryPool, VideoCore::NumQueryTypes> query_pools;
 
@@ -224,6 +260,13 @@ private:
         return found != std::end(contents) ? &*found : nullptr;
     }
 
+    void AsyncFlushQuery(VAddr addr) {
+        if (!uncommitted_flushes) {
+            uncommitted_flushes = std::make_shared<std::unordered_set<VAddr>>();
+        }
+        uncommitted_flushes->insert(addr);
+    }
+
     static constexpr std::uintptr_t PAGE_SIZE = 4096;
     static constexpr unsigned PAGE_SHIFT = 12;
 
@@ -235,6 +278,9 @@ private:
     std::unordered_map<u64, std::vector<CachedQuery>> cached_queries;
 
     std::array<CounterStream, VideoCore::NumQueryTypes> streams;
+
+    std::shared_ptr<std::unordered_set<VAddr>> uncommitted_flushes{};
+    std::list<std::shared_ptr<std::unordered_set<VAddr>>> committed_flushes;
 };
 
 template <class QueryCache, class HostCounter>

src/video_core/rasterizer_interface.h

@@ -49,15 +49,33 @@ public:
     /// Records a GPU query and caches it
     virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
 
+    /// Signal a GPU based semaphore as a fence
+    virtual void SignalSemaphore(GPUVAddr addr, u32 value) = 0;
+
+    /// Signal a GPU based syncpoint as a fence
+    virtual void SignalSyncPoint(u32 value) = 0;
+
+    /// Release all pending fences.
+    virtual void ReleaseFences() = 0;
+
     /// Notify rasterizer that all caches should be flushed to Switch memory
     virtual void FlushAll() = 0;
 
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     virtual void FlushRegion(VAddr addr, u64 size) = 0;
 
+    /// Check if the the specified memory area requires flushing to CPU Memory.
+    virtual bool MustFlushRegion(VAddr addr, u64 size) = 0;
+
     /// Notify rasterizer that any caches of the specified region should be invalidated
     virtual void InvalidateRegion(VAddr addr, u64 size) = 0;
 
+    /// Notify rasterizer that any caches of the specified region are desync with guest
+    virtual void OnCPUWrite(VAddr addr, u64 size) = 0;
+
+    /// Sync memory between guest and host.
+    virtual void SyncGuestHost() = 0;
+
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     /// and invalidated
     virtual void FlushAndInvalidateRegion(VAddr addr, u64 size) = 0;

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -52,7 +52,7 @@ Buffer OGLBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
 }
 
 void OGLBufferCache::WriteBarrier() {
-    glMemoryBarrier(GL_ALL_BARRIER_BITS);
+    glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
 }
 
 GLuint OGLBufferCache::ToHandle(const Buffer& buffer) {
@@ -72,6 +72,7 @@ void OGLBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, s
 void OGLBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
                                        u8* data) {
     MICROPROFILE_SCOPE(OpenGL_Buffer_Download);
+    glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
     glGetNamedBufferSubData(buffer->GetHandle(), static_cast<GLintptr>(offset),
                             static_cast<GLsizeiptr>(size), data);
 }

src/video_core/renderer_opengl/gl_fence_manager.cpp

@@ -0,0 +1,72 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/assert.h"
+
+#include "video_core/renderer_opengl/gl_fence_manager.h"
+
+namespace OpenGL {
+
+GLInnerFence::GLInnerFence(u32 payload, bool is_stubbed)
+    : VideoCommon::FenceBase(payload, is_stubbed), sync_object{} {}
+
+GLInnerFence::GLInnerFence(GPUVAddr address, u32 payload, bool is_stubbed)
+    : VideoCommon::FenceBase(address, payload, is_stubbed), sync_object{} {}
+
+GLInnerFence::~GLInnerFence() = default;
+
+void GLInnerFence::Queue() {
+    if (is_stubbed) {
+        return;
+    }
+    ASSERT(sync_object.handle == 0);
+    sync_object.Create();
+}
+
+bool GLInnerFence::IsSignaled() const {
+    if (is_stubbed) {
+        return true;
+    }
+    ASSERT(sync_object.handle != 0);
+    GLsizei length;
+    GLint sync_status;
+    glGetSynciv(sync_object.handle, GL_SYNC_STATUS, sizeof(GLint), &length, &sync_status);
+    return sync_status == GL_SIGNALED;
+}
+
+void GLInnerFence::Wait() {
+    if (is_stubbed) {
+        return;
+    }
+    ASSERT(sync_object.handle != 0);
+    glClientWaitSync(sync_object.handle, 0, GL_TIMEOUT_IGNORED);
+}
+
+FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system,
+                                       VideoCore::RasterizerInterface& rasterizer,
+                                       TextureCacheOpenGL& texture_cache,
+                                       OGLBufferCache& buffer_cache, QueryCache& query_cache)
+    : GenericFenceManager(system, rasterizer, texture_cache, buffer_cache, query_cache) {}
+
+Fence FenceManagerOpenGL::CreateFence(u32 value, bool is_stubbed) {
+    return std::make_shared<GLInnerFence>(value, is_stubbed);
+}
+
+Fence FenceManagerOpenGL::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) {
+    return std::make_shared<GLInnerFence>(addr, value, is_stubbed);
+}
+
+void FenceManagerOpenGL::QueueFence(Fence& fence) {
+    fence->Queue();
+}
+
+bool FenceManagerOpenGL::IsFenceSignaled(Fence& fence) const {
+    return fence->IsSignaled();
+}
+
+void FenceManagerOpenGL::WaitFence(Fence& fence) {
+    fence->Wait();
+}
+
+} // namespace OpenGL

src/video_core/renderer_opengl/gl_fence_manager.h

@@ -0,0 +1,53 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <glad/glad.h>
+
+#include "common/common_types.h"
+#include "video_core/fence_manager.h"
+#include "video_core/renderer_opengl/gl_buffer_cache.h"
+#include "video_core/renderer_opengl/gl_query_cache.h"
+#include "video_core/renderer_opengl/gl_resource_manager.h"
+#include "video_core/renderer_opengl/gl_texture_cache.h"
+
+namespace OpenGL {
+
+class GLInnerFence : public VideoCommon::FenceBase {
+public:
+    GLInnerFence(u32 payload, bool is_stubbed);
+    GLInnerFence(GPUVAddr address, u32 payload, bool is_stubbed);
+    ~GLInnerFence();
+
+    void Queue();
+
+    bool IsSignaled() const;
+
+    void Wait();
+
+private:
+    OGLSync sync_object;
+};
+
+using Fence = std::shared_ptr<GLInnerFence>;
+using GenericFenceManager =
+    VideoCommon::FenceManager<Fence, TextureCacheOpenGL, OGLBufferCache, QueryCache>;
+
+class FenceManagerOpenGL final : public GenericFenceManager {
+public:
+    FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+                       TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache,
+                       QueryCache& query_cache);
+
+protected:
+    Fence CreateFence(u32 value, bool is_stubbed) override;
+    Fence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) override;
+    void QueueFence(Fence& fence) override;
+    bool IsFenceSignaled(Fence& fence) const override;
+    void WaitFence(Fence& fence) override;
+};
+
+} // namespace OpenGL

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -99,9 +99,10 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
                                    ScreenInfo& info, GLShader::ProgramManager& program_manager,
                                    StateTracker& state_tracker)
     : RasterizerAccelerated{system.Memory()}, texture_cache{system, *this, device, state_tracker},
-      shader_cache{*this, system, emu_window, device}, query_cache{system, *this}, system{system},
-      screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker},
-      buffer_cache{*this, system, device, STREAM_BUFFER_SIZE} {
+      shader_cache{*this, system, emu_window, device}, query_cache{system, *this},
+      buffer_cache{*this, system, device, STREAM_BUFFER_SIZE},
+      fence_manager{system, *this, texture_cache, buffer_cache, query_cache}, system{system},
+      screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker} {
     CheckExtensions();
 }
 
@@ -599,6 +600,8 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
     EndTransformFeedback();
 
     ++num_queued_commands;
+
+    system.GPU().TickWork();
 }
 
 void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
@@ -649,6 +652,13 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
     query_cache.FlushRegion(addr, size);
 }
 
+bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
+    if (!Settings::IsGPULevelHigh()) {
+        return buffer_cache.MustFlushRegion(addr, size);
+    }
+    return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
+}
+
 void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
     if (addr == 0 || size == 0) {
@@ -660,8 +670,52 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
     query_cache.InvalidateRegion(addr, size);
 }
 
+void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
+    MICROPROFILE_SCOPE(OpenGL_CacheManagement);
+    if (addr == 0 || size == 0) {
+        return;
+    }
+    texture_cache.OnCPUWrite(addr, size);
+    shader_cache.InvalidateRegion(addr, size);
+    buffer_cache.OnCPUWrite(addr, size);
+    query_cache.InvalidateRegion(addr, size);
+}
+
+void RasterizerOpenGL::SyncGuestHost() {
+    MICROPROFILE_SCOPE(OpenGL_CacheManagement);
+    texture_cache.SyncGuestHost();
+    buffer_cache.SyncGuestHost();
+}
+
+void RasterizerOpenGL::SignalSemaphore(GPUVAddr addr, u32 value) {
+    auto& gpu{system.GPU()};
+    if (!gpu.IsAsync()) {
+        auto& memory_manager{gpu.MemoryManager()};
+        memory_manager.Write<u32>(addr, value);
+        return;
+    }
+    fence_manager.SignalSemaphore(addr, value);
+}
+
+void RasterizerOpenGL::SignalSyncPoint(u32 value) {
+    auto& gpu{system.GPU()};
+    if (!gpu.IsAsync()) {
+        gpu.IncrementSyncPoint(value);
+        return;
+    }
+    fence_manager.SignalSyncPoint(value);
+}
+
+void RasterizerOpenGL::ReleaseFences() {
+    auto& gpu{system.GPU()};
+    if (!gpu.IsAsync()) {
+        return;
+    }
+    fence_manager.WaitPendingFences();
+}
+
 void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
-    if (Settings::values.use_accurate_gpu_emulation) {
+    if (Settings::IsGPULevelExtreme()) {
         FlushRegion(addr, size);
     }
     InvalidateRegion(addr, size);

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -23,6 +23,7 @@
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_opengl/gl_buffer_cache.h"
 #include "video_core/renderer_opengl/gl_device.h"
+#include "video_core/renderer_opengl/gl_fence_manager.h"
 #include "video_core/renderer_opengl/gl_framebuffer_cache.h"
 #include "video_core/renderer_opengl/gl_query_cache.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
@@ -66,7 +67,13 @@ public:
     void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
     void FlushAll() override;
     void FlushRegion(VAddr addr, u64 size) override;
+    bool MustFlushRegion(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size) override;
+    void OnCPUWrite(VAddr addr, u64 size) override;
+    void SyncGuestHost() override;
+    void SignalSemaphore(GPUVAddr addr, u32 value) override;
+    void SignalSyncPoint(u32 value) override;
+    void ReleaseFences() override;
     void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
     void FlushCommands() override;
     void TickFrame() override;
@@ -222,6 +229,8 @@ private:
     SamplerCacheOpenGL sampler_cache;
     FramebufferCacheOpenGL framebuffer_cache;
     QueryCache query_cache;
+    OGLBufferCache buffer_cache;
+    FenceManagerOpenGL fence_manager;
 
     Core::System& system;
     ScreenInfo& screen_info;
@@ -229,7 +238,6 @@ private:
     StateTracker& state_tracker;
 
     static constexpr std::size_t STREAM_BUFFER_SIZE = 128 * 1024 * 1024;
-    OGLBufferCache buffer_cache;
 
     GLint vertex_binding = 0;
 

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -448,7 +448,7 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
 
     // Look up shader in the cache based on address
     const auto cpu_addr{memory_manager.GpuToCpuAddress(address)};
-    Shader shader{cpu_addr ? TryGet(*cpu_addr) : nullptr};
+    Shader shader{cpu_addr ? TryGet(*cpu_addr) : null_shader};
     if (shader) {
         return last_shaders[static_cast<std::size_t>(program)] = shader;
     }
@@ -477,7 +477,12 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
         const std::size_t size_in_bytes = code.size() * sizeof(u64);
         shader = CachedShader::CreateFromCache(params, found->second, size_in_bytes);
     }
-    Register(shader);
+
+    if (cpu_addr) {
+        Register(shader);
+    } else {
+        null_shader = shader;
+    }
 
     return last_shaders[static_cast<std::size_t>(program)] = shader;
 }
@@ -486,7 +491,7 @@ Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
     auto& memory_manager{system.GPU().MemoryManager()};
     const auto cpu_addr{memory_manager.GpuToCpuAddress(code_addr)};
 
-    auto kernel = cpu_addr ? TryGet(*cpu_addr) : nullptr;
+    auto kernel = cpu_addr ? TryGet(*cpu_addr) : null_kernel;
     if (kernel) {
         return kernel;
     }
@@ -507,7 +512,11 @@ Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
         kernel = CachedShader::CreateFromCache(params, found->second, size_in_bytes);
     }
 
-    Register(kernel);
+    if (cpu_addr) {
+        Register(kernel);
+    } else {
+        null_kernel = kernel;
+    }
     return kernel;
 }
 

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -125,6 +125,9 @@ private:
     ShaderDiskCacheOpenGL disk_cache;
     std::unordered_map<u64, PrecompiledShader> runtime_cache;
 
+    Shader null_shader{};
+    Shader null_kernel{};
+
     std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
 };
 

src/video_core/renderer_vulkan/vk_fence_manager.cpp

@@ -0,0 +1,101 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <memory>
+#include <thread>
+
+#include "video_core/renderer_vulkan/vk_buffer_cache.h"
+#include "video_core/renderer_vulkan/vk_device.h"
+#include "video_core/renderer_vulkan/vk_fence_manager.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_texture_cache.h"
+#include "video_core/renderer_vulkan/wrapper.h"
+
+namespace Vulkan {
+
+InnerFence::InnerFence(const VKDevice& device, VKScheduler& scheduler, u32 payload, bool is_stubbed)
+    : VideoCommon::FenceBase(payload, is_stubbed), device{device}, scheduler{scheduler} {}
+
+InnerFence::InnerFence(const VKDevice& device, VKScheduler& scheduler, GPUVAddr address,
+                       u32 payload, bool is_stubbed)
+    : VideoCommon::FenceBase(address, payload, is_stubbed), device{device}, scheduler{scheduler} {}
+
+InnerFence::~InnerFence() = default;
+
+void InnerFence::Queue() {
+    if (is_stubbed) {
+        return;
+    }
+    ASSERT(!event);
+
+    event = device.GetLogical().CreateEvent();
+    ticks = scheduler.Ticks();
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([event = *event](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetEvent(event, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
+    });
+}
+
+bool InnerFence::IsSignaled() const {
+    if (is_stubbed) {
+        return true;
+    }
+    ASSERT(event);
+    return IsEventSignalled();
+}
+
+void InnerFence::Wait() {
+    if (is_stubbed) {
+        return;
+    }
+    ASSERT(event);
+
+    if (ticks >= scheduler.Ticks()) {
+        scheduler.Flush();
+    }
+    while (!IsEventSignalled()) {
+        std::this_thread::yield();
+    }
+}
+
+bool InnerFence::IsEventSignalled() const {
+    switch (const VkResult result = event.GetStatus()) {
+    case VK_EVENT_SET:
+        return true;
+    case VK_EVENT_RESET:
+        return false;
+    default:
+        throw vk::Exception(result);
+    }
+}
+
+VKFenceManager::VKFenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+                               const VKDevice& device, VKScheduler& scheduler,
+                               VKTextureCache& texture_cache, VKBufferCache& buffer_cache,
+                               VKQueryCache& query_cache)
+    : GenericFenceManager(system, rasterizer, texture_cache, buffer_cache, query_cache),
+      device{device}, scheduler{scheduler} {}
+
+Fence VKFenceManager::CreateFence(u32 value, bool is_stubbed) {
+    return std::make_shared<InnerFence>(device, scheduler, value, is_stubbed);
+}
+
+Fence VKFenceManager::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) {
+    return std::make_shared<InnerFence>(device, scheduler, addr, value, is_stubbed);
+}
+
+void VKFenceManager::QueueFence(Fence& fence) {
+    fence->Queue();
+}
+
+bool VKFenceManager::IsFenceSignaled(Fence& fence) const {
+    return fence->IsSignaled();
+}
+
+void VKFenceManager::WaitFence(Fence& fence) {
+    fence->Wait();
+}
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_fence_manager.h

@@ -0,0 +1,74 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+
+#include "video_core/fence_manager.h"
+#include "video_core/renderer_vulkan/wrapper.h"
+
+namespace Core {
+class System;
+}
+
+namespace VideoCore {
+class RasterizerInterface;
+}
+
+namespace Vulkan {
+
+class VKBufferCache;
+class VKDevice;
+class VKQueryCache;
+class VKScheduler;
+class VKTextureCache;
+
+class InnerFence : public VideoCommon::FenceBase {
+public:
+    explicit InnerFence(const VKDevice& device, VKScheduler& scheduler, u32 payload,
+                        bool is_stubbed);
+    explicit InnerFence(const VKDevice& device, VKScheduler& scheduler, GPUVAddr address,
+                        u32 payload, bool is_stubbed);
+    ~InnerFence();
+
+    void Queue();
+
+    bool IsSignaled() const;
+
+    void Wait();
+
+private:
+    bool IsEventSignalled() const;
+
+    const VKDevice& device;
+    VKScheduler& scheduler;
+    vk::Event event;
+    u64 ticks = 0;
+};
+using Fence = std::shared_ptr<InnerFence>;
+
+using GenericFenceManager =
+    VideoCommon::FenceManager<Fence, VKTextureCache, VKBufferCache, VKQueryCache>;
+
+class VKFenceManager final : public GenericFenceManager {
+public:
+    explicit VKFenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+                            const VKDevice& device, VKScheduler& scheduler,
+                            VKTextureCache& texture_cache, VKBufferCache& buffer_cache,
+                            VKQueryCache& query_cache);
+
+protected:
+    Fence CreateFence(u32 value, bool is_stubbed) override;
+    Fence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) override;
+    void QueueFence(Fence& fence) override;
+    bool IsFenceSignaled(Fence& fence) const override;
+    void WaitFence(Fence& fence) override;
+
+private:
+    const VKDevice& device;
+    VKScheduler& scheduler;
+};
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -207,7 +207,7 @@ std::array<Shader, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
         const GPUVAddr program_addr{GetShaderAddress(system, program)};
         const std::optional cpu_addr = memory_manager.GpuToCpuAddress(program_addr);
         ASSERT(cpu_addr);
-        auto shader = cpu_addr ? TryGet(*cpu_addr) : nullptr;
+        auto shader = cpu_addr ? TryGet(*cpu_addr) : null_shader;
         if (!shader) {
             const auto host_ptr{memory_manager.GetPointer(program_addr)};
 
@@ -218,7 +218,11 @@ std::array<Shader, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
 
             shader = std::make_shared<CachedShader>(system, stage, program_addr, *cpu_addr,
                                                     std::move(code), stage_offset);
-            Register(shader);
+            if (cpu_addr) {
+                Register(shader);
+            } else {
+                null_shader = shader;
+            }
         }
         shaders[index] = std::move(shader);
     }
@@ -261,7 +265,7 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
     const auto cpu_addr = memory_manager.GpuToCpuAddress(program_addr);
     ASSERT(cpu_addr);
 
-    auto shader = cpu_addr ? TryGet(*cpu_addr) : nullptr;
+    auto shader = cpu_addr ? TryGet(*cpu_addr) : null_kernel;
     if (!shader) {
         // No shader found - create a new one
         const auto host_ptr = memory_manager.GetPointer(program_addr);
@@ -271,7 +275,11 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
         shader = std::make_shared<CachedShader>(system, Tegra::Engines::ShaderType::Compute,
                                                 program_addr, *cpu_addr, std::move(code),
                                                 kernel_main_offset);
-        Register(shader);
+        if (cpu_addr) {
+            Register(shader);
+        } else {
+            null_kernel = shader;
+        }
     }
 
     Specialization specialization;

src/video_core/renderer_vulkan/vk_pipeline_cache.h

@@ -182,6 +182,9 @@ private:
     VKUpdateDescriptorQueue& update_descriptor_queue;
     VKRenderPassCache& renderpass_cache;
 
+    Shader null_shader{};
+    Shader null_kernel{};
+
     std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
 
     GraphicsPipelineCacheKey last_graphics_key;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -17,6 +17,7 @@
 #include "common/microprofile.h"
 #include "core/core.h"
 #include "core/memory.h"
+#include "core/settings.h"
 #include "video_core/engines/kepler_compute.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/fixed_pipeline_state.h"
@@ -299,7 +300,9 @@ RasterizerVulkan::RasterizerVulkan(Core::System& system, Core::Frontend::EmuWind
       pipeline_cache(system, *this, device, scheduler, descriptor_pool, update_descriptor_queue,
                      renderpass_cache),
       buffer_cache(*this, system, device, memory_manager, scheduler, staging_pool),
-      sampler_cache(device), query_cache(system, *this, device, scheduler) {
+      sampler_cache(device),
+      fence_manager(system, *this, device, scheduler, texture_cache, buffer_cache, query_cache),
+      query_cache(system, *this, device, scheduler) {
     scheduler.SetQueryCache(query_cache);
 }
 
@@ -360,6 +363,8 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
     });
 
     EndTransformFeedback();
+
+    system.GPU().TickWork();
 }
 
 void RasterizerVulkan::Clear() {
@@ -504,6 +509,13 @@ void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) {
     query_cache.FlushRegion(addr, size);
 }
 
+bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size) {
+    if (!Settings::IsGPULevelHigh()) {
+        return buffer_cache.MustFlushRegion(addr, size);
+    }
+    return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
+}
+
 void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
     if (addr == 0 || size == 0) {
         return;
@@ -514,6 +526,47 @@ void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
     query_cache.InvalidateRegion(addr, size);
 }
 
+void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
+    if (addr == 0 || size == 0) {
+        return;
+    }
+    texture_cache.OnCPUWrite(addr, size);
+    pipeline_cache.InvalidateRegion(addr, size);
+    buffer_cache.OnCPUWrite(addr, size);
+    query_cache.InvalidateRegion(addr, size);
+}
+
+void RasterizerVulkan::SyncGuestHost() {
+    texture_cache.SyncGuestHost();
+    buffer_cache.SyncGuestHost();
+}
+
+void RasterizerVulkan::SignalSemaphore(GPUVAddr addr, u32 value) {
+    auto& gpu{system.GPU()};
+    if (!gpu.IsAsync()) {
+        gpu.MemoryManager().Write<u32>(addr, value);
+        return;
+    }
+    fence_manager.SignalSemaphore(addr, value);
+}
+
+void RasterizerVulkan::SignalSyncPoint(u32 value) {
+    auto& gpu{system.GPU()};
+    if (!gpu.IsAsync()) {
+        gpu.IncrementSyncPoint(value);
+        return;
+    }
+    fence_manager.SignalSyncPoint(value);
+}
+
+void RasterizerVulkan::ReleaseFences() {
+    auto& gpu{system.GPU()};
+    if (!gpu.IsAsync()) {
+        return;
+    }
+    fence_manager.WaitPendingFences();
+}
+
 void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size) {
     FlushRegion(addr, size);
     InvalidateRegion(addr, size);

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -21,6 +21,7 @@
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
+#include "video_core/renderer_vulkan/vk_fence_manager.h"
 #include "video_core/renderer_vulkan/vk_memory_manager.h"
 #include "video_core/renderer_vulkan/vk_pipeline_cache.h"
 #include "video_core/renderer_vulkan/vk_query_cache.h"
@@ -118,7 +119,13 @@ public:
     void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
     void FlushAll() override;
     void FlushRegion(VAddr addr, u64 size) override;
+    bool MustFlushRegion(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size) override;
+    void OnCPUWrite(VAddr addr, u64 size) override;
+    void SyncGuestHost() override;
+    void SignalSemaphore(GPUVAddr addr, u32 value) override;
+    void SignalSyncPoint(u32 value) override;
+    void ReleaseFences() override;
     void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
     void FlushCommands() override;
     void TickFrame() override;
@@ -261,6 +268,7 @@ private:
     VKPipelineCache pipeline_cache;
     VKBufferCache buffer_cache;
     VKSamplerCache sampler_cache;
+    VKFenceManager fence_manager;
     VKQueryCache query_cache;
 
     std::array<View, Maxwell::NumRenderTargets> color_attachments;

src/video_core/renderer_vulkan/wrapper.cpp

@@ -63,6 +63,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdSetBlendConstants);
     X(vkCmdSetDepthBias);
     X(vkCmdSetDepthBounds);
+    X(vkCmdSetEvent);
     X(vkCmdSetScissor);
     X(vkCmdSetStencilCompareMask);
     X(vkCmdSetStencilReference);
@@ -75,6 +76,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCreateDescriptorPool);
     X(vkCreateDescriptorSetLayout);
     X(vkCreateDescriptorUpdateTemplateKHR);
+    X(vkCreateEvent);
     X(vkCreateFence);
     X(vkCreateFramebuffer);
     X(vkCreateGraphicsPipelines);
@@ -93,6 +95,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkDestroyDescriptorPool);
     X(vkDestroyDescriptorSetLayout);
     X(vkDestroyDescriptorUpdateTemplateKHR);
+    X(vkDestroyEvent);
     X(vkDestroyFence);
     X(vkDestroyFramebuffer);
     X(vkDestroyImage);
@@ -112,6 +115,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkFreeMemory);
     X(vkGetBufferMemoryRequirements);
     X(vkGetDeviceQueue);
+    X(vkGetEventStatus);
     X(vkGetFenceStatus);
     X(vkGetImageMemoryRequirements);
     X(vkGetQueryPoolResults);
@@ -269,6 +273,10 @@ void Destroy(VkDevice device, VkDeviceMemory handle, const DeviceDispatch& dld)
     dld.vkFreeMemory(device, handle, nullptr);
 }
 
+void Destroy(VkDevice device, VkEvent handle, const DeviceDispatch& dld) noexcept {
+    dld.vkDestroyEvent(device, handle, nullptr);
+}
+
 void Destroy(VkDevice device, VkFence handle, const DeviceDispatch& dld) noexcept {
     dld.vkDestroyFence(device, handle, nullptr);
 }
@@ -599,6 +607,16 @@ ShaderModule Device::CreateShaderModule(const VkShaderModuleCreateInfo& ci) cons
     return ShaderModule(object, handle, *dld);
 }
 
+Event Device::CreateEvent() const {
+    VkEventCreateInfo ci;
+    ci.sType = VK_STRUCTURE_TYPE_EVENT_CREATE_INFO;
+    ci.pNext = nullptr;
+    ci.flags = 0;
+    VkEvent object;
+    Check(dld->vkCreateEvent(handle, &ci, nullptr, &object));
+    return Event(object, handle, *dld);
+}
+
 SwapchainKHR Device::CreateSwapchainKHR(const VkSwapchainCreateInfoKHR& ci) const {
     VkSwapchainKHR object;
     Check(dld->vkCreateSwapchainKHR(handle, &ci, nullptr, &object));

src/video_core/renderer_vulkan/wrapper.h

@@ -199,6 +199,7 @@ struct DeviceDispatch : public InstanceDispatch {
     PFN_vkCmdSetBlendConstants vkCmdSetBlendConstants;
     PFN_vkCmdSetDepthBias vkCmdSetDepthBias;
     PFN_vkCmdSetDepthBounds vkCmdSetDepthBounds;
+    PFN_vkCmdSetEvent vkCmdSetEvent;
     PFN_vkCmdSetScissor vkCmdSetScissor;
     PFN_vkCmdSetStencilCompareMask vkCmdSetStencilCompareMask;
     PFN_vkCmdSetStencilReference vkCmdSetStencilReference;
@@ -211,6 +212,7 @@ struct DeviceDispatch : public InstanceDispatch {
     PFN_vkCreateDescriptorPool vkCreateDescriptorPool;
     PFN_vkCreateDescriptorSetLayout vkCreateDescriptorSetLayout;
     PFN_vkCreateDescriptorUpdateTemplateKHR vkCreateDescriptorUpdateTemplateKHR;
+    PFN_vkCreateEvent vkCreateEvent;
     PFN_vkCreateFence vkCreateFence;
     PFN_vkCreateFramebuffer vkCreateFramebuffer;
     PFN_vkCreateGraphicsPipelines vkCreateGraphicsPipelines;
@@ -229,6 +231,7 @@ struct DeviceDispatch : public InstanceDispatch {
     PFN_vkDestroyDescriptorPool vkDestroyDescriptorPool;
     PFN_vkDestroyDescriptorSetLayout vkDestroyDescriptorSetLayout;
     PFN_vkDestroyDescriptorUpdateTemplateKHR vkDestroyDescriptorUpdateTemplateKHR;
+    PFN_vkDestroyEvent vkDestroyEvent;
     PFN_vkDestroyFence vkDestroyFence;
     PFN_vkDestroyFramebuffer vkDestroyFramebuffer;
     PFN_vkDestroyImage vkDestroyImage;
@@ -248,6 +251,7 @@ struct DeviceDispatch : public InstanceDispatch {
     PFN_vkFreeMemory vkFreeMemory;
     PFN_vkGetBufferMemoryRequirements vkGetBufferMemoryRequirements;
     PFN_vkGetDeviceQueue vkGetDeviceQueue;
+    PFN_vkGetEventStatus vkGetEventStatus;
     PFN_vkGetFenceStatus vkGetFenceStatus;
     PFN_vkGetImageMemoryRequirements vkGetImageMemoryRequirements;
     PFN_vkGetQueryPoolResults vkGetQueryPoolResults;
@@ -279,6 +283,7 @@ void Destroy(VkDevice, VkDescriptorPool, const DeviceDispatch&) noexcept;
 void Destroy(VkDevice, VkDescriptorSetLayout, const DeviceDispatch&) noexcept;
 void Destroy(VkDevice, VkDescriptorUpdateTemplateKHR, const DeviceDispatch&) noexcept;
 void Destroy(VkDevice, VkDeviceMemory, const DeviceDispatch&) noexcept;
+void Destroy(VkDevice, VkEvent, const DeviceDispatch&) noexcept;
 void Destroy(VkDevice, VkFence, const DeviceDispatch&) noexcept;
 void Destroy(VkDevice, VkFramebuffer, const DeviceDispatch&) noexcept;
 void Destroy(VkDevice, VkImage, const DeviceDispatch&) noexcept;
@@ -648,6 +653,15 @@ public:
     std::vector<VkImage> GetImages() const;
 };
 
+class Event : public Handle<VkEvent, VkDevice, DeviceDispatch> {
+    using Handle<VkEvent, VkDevice, DeviceDispatch>::Handle;
+
+public:
+    VkResult GetStatus() const noexcept {
+        return dld->vkGetEventStatus(owner, handle);
+    }
+};
+
 class Device : public Handle<VkDevice, NoOwner, DeviceDispatch> {
     using Handle<VkDevice, NoOwner, DeviceDispatch>::Handle;
 
@@ -695,6 +709,8 @@ public:
 
     ShaderModule CreateShaderModule(const VkShaderModuleCreateInfo& ci) const;
 
+    Event CreateEvent() const;
+
     SwapchainKHR CreateSwapchainKHR(const VkSwapchainCreateInfoKHR& ci) const;
 
     DeviceMemory TryAllocateMemory(const VkMemoryAllocateInfo& ai) const noexcept;
@@ -938,6 +954,10 @@ public:
         dld->vkCmdSetDepthBounds(handle, min_depth_bounds, max_depth_bounds);
     }
 
+    void SetEvent(VkEvent event, VkPipelineStageFlags stage_flags) const noexcept {
+        dld->vkCmdSetEvent(handle, event, stage_flags);
+    }
+
     void BindTransformFeedbackBuffersEXT(u32 first, u32 count, const VkBuffer* buffers,
                                          const VkDeviceSize* offsets,
                                          const VkDeviceSize* sizes) const noexcept {

src/video_core/texture_cache/surface_base.h

@@ -192,6 +192,22 @@ public:
         index = index_;
     }
 
+    void SetMemoryMarked(bool is_memory_marked_) {
+        is_memory_marked = is_memory_marked_;
+    }
+
+    bool IsMemoryMarked() const {
+        return is_memory_marked;
+    }
+
+    void SetSyncPending(bool is_sync_pending_) {
+        is_sync_pending = is_sync_pending_;
+    }
+
+    bool IsSyncPending() const {
+        return is_sync_pending;
+    }
+
     void MarkAsPicked(bool is_picked_) {
         is_picked = is_picked_;
     }
@@ -303,6 +319,8 @@ private:
     bool is_target{};
     bool is_registered{};
     bool is_picked{};
+    bool is_memory_marked{};
+    bool is_sync_pending{};
     u32 index{NO_RT};
     u64 modification_tick{};
 };

src/video_core/texture_cache/texture_cache.h

@@ -6,6 +6,7 @@
 
 #include <algorithm>
 #include <array>
+#include <list>
 #include <memory>
 #include <mutex>
 #include <set>
@@ -62,6 +63,30 @@ public:
         }
     }
 
+    void OnCPUWrite(VAddr addr, std::size_t size) {
+        std::lock_guard lock{mutex};
+
+        for (const auto& surface : GetSurfacesInRegion(addr, size)) {
+            if (surface->IsMemoryMarked()) {
+                UnmarkMemory(surface);
+                surface->SetSyncPending(true);
+                marked_for_unregister.emplace_back(surface);
+            }
+        }
+    }
+
+    void SyncGuestHost() {
+        std::lock_guard lock{mutex};
+
+        for (const auto& surface : marked_for_unregister) {
+            if (surface->IsRegistered()) {
+                surface->SetSyncPending(false);
+                Unregister(surface);
+            }
+        }
+        marked_for_unregister.clear();
+    }
+
     /**
      * Guarantees that rendertargets don't unregister themselves if the
      * collide. Protection is currently only done on 3D slices.
@@ -85,10 +110,20 @@ public:
             return a->GetModificationTick() < b->GetModificationTick();
         });
         for (const auto& surface : surfaces) {
+            mutex.unlock();
             FlushSurface(surface);
+            mutex.lock();
         }
     }
 
+    bool MustFlushRegion(VAddr addr, std::size_t size) {
+        std::lock_guard lock{mutex};
+
+        const auto surfaces = GetSurfacesInRegion(addr, size);
+        return std::any_of(surfaces.cbegin(), surfaces.cend(),
+                           [](const TSurface& surface) { return surface->IsModified(); });
+    }
+
     TView GetTextureSurface(const Tegra::Texture::TICEntry& tic,
                             const VideoCommon::Shader::Sampler& entry) {
         std::lock_guard lock{mutex};
@@ -206,8 +241,14 @@ public:
 
         auto surface_view = GetSurface(gpu_addr, *cpu_addr,
                                        SurfaceParams::CreateForFramebuffer(system, index), true);
-        if (render_targets[index].target)
-            render_targets[index].target->MarkAsRenderTarget(false, NO_RT);
+        if (render_targets[index].target) {
+            auto& surface = render_targets[index].target;
+            surface->MarkAsRenderTarget(false, NO_RT);
+            const auto& cr_params = surface->GetSurfaceParams();
+            if (!cr_params.is_tiled && Settings::values.use_asynchronous_gpu_emulation) {
+                AsyncFlushSurface(surface);
+            }
+        }
         render_targets[index].target = surface_view.first;
         render_targets[index].view = surface_view.second;
         if (render_targets[index].target)
@@ -284,6 +325,34 @@ public:
         return ++ticks;
     }
 
+    void CommitAsyncFlushes() {
+        committed_flushes.push_back(uncommitted_flushes);
+        uncommitted_flushes.reset();
+    }
+
+    bool HasUncommittedFlushes() const {
+        return uncommitted_flushes != nullptr;
+    }
+
+    bool ShouldWaitAsyncFlushes() const {
+        return !committed_flushes.empty() && committed_flushes.front() != nullptr;
+    }
+
+    void PopAsyncFlushes() {
+        if (committed_flushes.empty()) {
+            return;
+        }
+        auto& flush_list = committed_flushes.front();
+        if (!flush_list) {
+            committed_flushes.pop_front();
+            return;
+        }
+        for (TSurface& surface : *flush_list) {
+            FlushSurface(surface);
+        }
+        committed_flushes.pop_front();
+    }
+
 protected:
     explicit TextureCache(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
                           bool is_astc_supported)
@@ -345,9 +414,20 @@ protected:
         surface->SetCpuAddr(*cpu_addr);
         RegisterInnerCache(surface);
         surface->MarkAsRegistered(true);
+        surface->SetMemoryMarked(true);
         rasterizer.UpdatePagesCachedCount(*cpu_addr, size, 1);
     }
 
+    void UnmarkMemory(TSurface surface) {
+        if (!surface->IsMemoryMarked()) {
+            return;
+        }
+        const std::size_t size = surface->GetSizeInBytes();
+        const VAddr cpu_addr = surface->GetCpuAddr();
+        rasterizer.UpdatePagesCachedCount(cpu_addr, size, -1);
+        surface->SetMemoryMarked(false);
+    }
+
     void Unregister(TSurface surface) {
         if (guard_render_targets && surface->IsProtected()) {
             return;
@@ -355,9 +435,11 @@ protected:
         if (!guard_render_targets && surface->IsRenderTarget()) {
             ManageRenderTargetUnregister(surface);
         }
-        const std::size_t size = surface->GetSizeInBytes();
-        const VAddr cpu_addr = surface->GetCpuAddr();
-        rasterizer.UpdatePagesCachedCount(cpu_addr, size, -1);
+        UnmarkMemory(surface);
+        if (surface->IsSyncPending()) {
+            marked_for_unregister.remove(surface);
+            surface->SetSyncPending(false);
+        }
         UnregisterInnerCache(surface);
         surface->MarkAsRegistered(false);
         ReserveSurface(surface->GetSurfaceParams(), surface);
@@ -417,7 +499,7 @@ private:
      **/
     RecycleStrategy PickStrategy(std::vector<TSurface>& overlaps, const SurfaceParams& params,
                                  const GPUVAddr gpu_addr, const MatchTopologyResult untopological) {
-        if (Settings::values.use_accurate_gpu_emulation) {
+        if (Settings::IsGPULevelExtreme()) {
             return RecycleStrategy::Flush;
         }
         // 3D Textures decision
@@ -461,7 +543,7 @@ private:
         }
         switch (PickStrategy(overlaps, params, gpu_addr, untopological)) {
         case RecycleStrategy::Ignore: {
-            return InitializeSurface(gpu_addr, params, Settings::values.use_accurate_gpu_emulation);
+            return InitializeSurface(gpu_addr, params, Settings::IsGPULevelExtreme());
         }
         case RecycleStrategy::Flush: {
             std::sort(overlaps.begin(), overlaps.end(),
@@ -509,7 +591,7 @@ private:
         }
         const auto& final_params = new_surface->GetSurfaceParams();
         if (cr_params.type != final_params.type) {
-            if (Settings::values.use_accurate_gpu_emulation) {
+            if (Settings::IsGPULevelExtreme()) {
                 BufferCopy(current_surface, new_surface);
             }
         } else {
@@ -598,7 +680,7 @@ private:
         if (passed_tests == 0) {
             return {};
             // In Accurate GPU all tests should pass, else we recycle
-        } else if (Settings::values.use_accurate_gpu_emulation && passed_tests != overlaps.size()) {
+        } else if (Settings::IsGPULevelExtreme() && passed_tests != overlaps.size()) {
             return {};
         }
         for (const auto& surface : overlaps) {
@@ -668,7 +750,7 @@ private:
             for (const auto& surface : overlaps) {
                 if (!surface->MatchTarget(params.target)) {
                     if (overlaps.size() == 1 && surface->GetCpuAddr() == cpu_addr) {
-                        if (Settings::values.use_accurate_gpu_emulation) {
+                        if (Settings::IsGPULevelExtreme()) {
                             return std::nullopt;
                         }
                         Unregister(surface);
@@ -1106,6 +1188,13 @@ private:
         TView view;
     };
 
+    void AsyncFlushSurface(TSurface& surface) {
+        if (!uncommitted_flushes) {
+            uncommitted_flushes = std::make_shared<std::list<TSurface>>();
+        }
+        uncommitted_flushes->push_back(surface);
+    }
+
     VideoCore::RasterizerInterface& rasterizer;
 
     FormatLookupTable format_lookup_table;
@@ -1150,6 +1239,11 @@ private:
     std::unordered_map<u32, TSurface> invalid_cache;
     std::vector<u8> invalid_memory;
 
+    std::list<TSurface> marked_for_unregister;
+
+    std::shared_ptr<std::list<TSurface>> uncommitted_flushes{};
+    std::list<std::shared_ptr<std::list<TSurface>>> committed_flushes;
+
     StagingCache staging_cache;
     std::recursive_mutex mutex;
 };

src/yuzu/configuration/config.cpp

@@ -639,8 +639,8 @@ void Config::ReadRendererValues() {
     Settings::values.frame_limit = ReadSetting(QStringLiteral("frame_limit"), 100).toInt();
     Settings::values.use_disk_shader_cache =
         ReadSetting(QStringLiteral("use_disk_shader_cache"), true).toBool();
-    Settings::values.use_accurate_gpu_emulation =
-        ReadSetting(QStringLiteral("use_accurate_gpu_emulation"), false).toBool();
+    const int gpu_accuracy_level = ReadSetting(QStringLiteral("gpu_accuracy"), 0).toInt();
+    Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
     Settings::values.use_asynchronous_gpu_emulation =
         ReadSetting(QStringLiteral("use_asynchronous_gpu_emulation"), false).toBool();
     Settings::values.use_vsync = ReadSetting(QStringLiteral("use_vsync"), true).toBool();
@@ -1080,8 +1080,8 @@ void Config::SaveRendererValues() {
     WriteSetting(QStringLiteral("frame_limit"), Settings::values.frame_limit, 100);
     WriteSetting(QStringLiteral("use_disk_shader_cache"), Settings::values.use_disk_shader_cache,
                  true);
-    WriteSetting(QStringLiteral("use_accurate_gpu_emulation"),
-                 Settings::values.use_accurate_gpu_emulation, false);
+    WriteSetting(QStringLiteral("gpu_accuracy"), static_cast<int>(Settings::values.gpu_accuracy),
+                 0);
     WriteSetting(QStringLiteral("use_asynchronous_gpu_emulation"),
                  Settings::values.use_asynchronous_gpu_emulation, false);
     WriteSetting(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);

src/yuzu/configuration/configure_graphics_advanced.cpp

@@ -19,7 +19,7 @@ ConfigureGraphicsAdvanced::~ConfigureGraphicsAdvanced() = default;
 
 void ConfigureGraphicsAdvanced::SetConfiguration() {
     const bool runtime_lock = !Core::System::GetInstance().IsPoweredOn();
-    ui->use_accurate_gpu_emulation->setChecked(Settings::values.use_accurate_gpu_emulation);
+    ui->gpu_accuracy->setCurrentIndex(static_cast<int>(Settings::values.gpu_accuracy));
     ui->use_vsync->setEnabled(runtime_lock);
     ui->use_vsync->setChecked(Settings::values.use_vsync);
     ui->force_30fps_mode->setEnabled(runtime_lock);
@@ -29,7 +29,8 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
 }
 
 void ConfigureGraphicsAdvanced::ApplyConfiguration() {
-    Settings::values.use_accurate_gpu_emulation = ui->use_accurate_gpu_emulation->isChecked();
+    auto gpu_accuracy = static_cast<Settings::GPUAccuracy>(ui->gpu_accuracy->currentIndex());
+    Settings::values.gpu_accuracy = gpu_accuracy;
     Settings::values.use_vsync = ui->use_vsync->isChecked();
     Settings::values.force_30fps_mode = ui->force_30fps_mode->isChecked();
     Settings::values.max_anisotropy = ui->anisotropic_filtering_combobox->currentIndex();

src/yuzu/configuration/configure_graphics_advanced.ui

@@ -23,11 +23,34 @@
        </property>
        <layout class="QVBoxLayout" name="verticalLayout_3">
         <item>
-         <widget class="QCheckBox" name="use_accurate_gpu_emulation">
-          <property name="text">
-           <string>Use accurate GPU emulation (slow)</string>
-          </property>
-         </widget>
+         <layout class="QHBoxLayout" name="horizontalLayout_2">
+          <item>
+           <widget class="QLabel" name="label_gpu_accuracy">
+            <property name="text">
+             <string>Accuracy Level:</string>
+            </property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QComboBox" name="gpu_accuracy">
+            <item>
+             <property name="text">
+              <string notr="true">Normal</string>
+             </property>
+            </item>
+            <item>
+             <property name="text">
+              <string notr="true">High</string>
+             </property>
+            </item>
+            <item>
+             <property name="text">
+              <string notr="true">Extreme(very slow)</string>
+             </property>
+            </item>
+           </widget>
+          </item>
+         </layout>
         </item>
         <item>
          <widget class="QCheckBox" name="use_vsync">

src/yuzu_cmd/config.cpp

@@ -388,8 +388,8 @@ void Config::ReadValues() {
         static_cast<u16>(sdl2_config->GetInteger("Renderer", "frame_limit", 100));
     Settings::values.use_disk_shader_cache =
         sdl2_config->GetBoolean("Renderer", "use_disk_shader_cache", false);
-    Settings::values.use_accurate_gpu_emulation =
-        sdl2_config->GetBoolean("Renderer", "use_accurate_gpu_emulation", false);
+    const int gpu_accuracy_level = sdl2_config->GetInteger("Renderer", "gpu_accuracy", 0);
+    Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
     Settings::values.use_asynchronous_gpu_emulation =
         sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);
     Settings::values.use_vsync =

src/yuzu_cmd/default_ini.h

@@ -146,9 +146,9 @@ frame_limit =
 # 0 (default): Off, 1 : On
 use_disk_shader_cache =
 
-# Whether to use accurate GPU emulation
-# 0 (default): Off (fast), 1 : On (slow)
-use_accurate_gpu_emulation =
+# Which gpu accuracy level to use
+# 0 (Normal), 1 (High), 2 (Extreme)
+gpu_accuracy =
 
 # Whether to use asynchronous GPU emulation
 # 0 : Off (slow), 1 (default): On (fast)

src/yuzu_tester/config.cpp

@@ -126,8 +126,8 @@ void Config::ReadValues() {
     Settings::values.frame_limit = 100;
     Settings::values.use_disk_shader_cache =
         sdl2_config->GetBoolean("Renderer", "use_disk_shader_cache", false);
-    Settings::values.use_accurate_gpu_emulation =
-        sdl2_config->GetBoolean("Renderer", "use_accurate_gpu_emulation", false);
+    const int gpu_accuracy_level = sdl2_config->GetInteger("Renderer", "gpu_accuracy", 0);
+    Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
     Settings::values.use_asynchronous_gpu_emulation =
         sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);