6 files changed, 70 insertions, 8 deletions

diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 19d3bd305..85a6c7bb5 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -125,6 +125,28 @@ bool GPU::CancelSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
     return true;
 }
 
+u64 GPU::RequestFlush(CacheAddr 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 CacheAddr 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
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index fa9991c87..943a5b110 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -159,6 +159,14 @@ public:
     void SyncGuestHost();
     virtual void OnCommandListEnd();
 
+    u64 RequestFlush(CacheAddr addr, std::size_t size);
+
+    u64 CurrentFlushRequestFence() const {
+        return current_flush_fence.load(std::memory_order_relaxed);
+    }
+
+    void TickWork();
+
     /// Returns a reference to the Maxwell3D GPU engine.
     Engines::Maxwell3D& Maxwell3D();
 
@@ -327,6 +335,19 @@ private:
 
     std::condition_variable sync_cv;
 
+    struct FlushRequest {
+        FlushRequest(u64 fence, CacheAddr addr, std::size_t size)
+            : fence{fence}, addr{addr}, size{size} {}
+        u64 fence;
+        CacheAddr 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;
 };
 
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 3e2be00e9..9460364a3 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -15,8 +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
@@ -40,6 +41,8 @@ static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::Graphic
             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)) {
@@ -68,8 +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) {
@@ -85,8 +88,10 @@ void ThreadManager::FlushRegion(VAddr addr, u64 size) {
         return;
     }
     if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
-        u64 fence = PushCommand(FlushRegionCommand(addr, size));
-        while (fence > state.signaled_fence.load(std::memory_order_relaxed)) {
+        auto& gpu = system.GPU();
+        u64 fence = gpu.RequestFlush(addr, size);
+        PushCommand(GPUTickCommand());
+        while (fence > gpu.CurrentFlushRequestFence()) {
         }
     }
 }
diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index 9d0877921..5a28335d6 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -70,12 +70,16 @@ struct FlushAndInvalidateRegionCommand final {
     u64 size;
 };
 
-/// Command to signal to the GPU thread that processing has ended
+/// 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, OnCommandListEndCommand>;
+                 InvalidateRegionCommand, FlushAndInvalidateRegionCommand, OnCommandListEndCommand,
+                 GPUTickCommand>;
 
 struct CommandDataContainer {
     CommandDataContainer() = default;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index e52e5961f..fbd81b895 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -601,6 +601,8 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
     EndTransformFeedback();
 
     ++num_queued_commands;
+
+    system.GPU().TickWork();
 }
 
 void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
@@ -628,6 +630,7 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
     const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
     glDispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
     ++num_queued_commands;
+    system.GPU().TickWork();
 }
 
 void RasterizerOpenGL::ResetCounter(VideoCore::QueryType type) {
@@ -652,6 +655,9 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
 }
 
 bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
+    if (!Settings::IsGPULevelExtreme()) {
+        return buffer_cache.MustFlushRegion(addr, size);
+    }
     return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
 }
 
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 507262c8f..926ecf38e 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -365,6 +365,8 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
     });
 
     EndTransformFeedback();
+
+    system.GPU().TickWork();
 }
 
 void RasterizerVulkan::Clear() {
@@ -492,6 +494,8 @@ void RasterizerVulkan::DispatchCompute(GPUVAddr code_addr) {
                                   descriptor_set, {});
         cmdbuf.Dispatch(grid_x, grid_y, grid_z);
     });
+
+    system.GPU().TickWork();
 }
 
 void RasterizerVulkan::ResetCounter(VideoCore::QueryType type) {