Merge pull request #2282 from bunnei/gpu-asynch-v2

gpu_thread: Improve synchronization by using CoreTiming.

src/video_core/gpu_asynch.cpp

@@ -9,7 +9,7 @@
 namespace VideoCommon {
 
 GPUAsynch::GPUAsynch(Core::System& system, VideoCore::RendererBase& renderer)
-    : Tegra::GPU(system, renderer), gpu_thread{renderer, *dma_pusher} {}
+    : Tegra::GPU(system, renderer), gpu_thread{system, renderer, *dma_pusher} {}
 
 GPUAsynch::~GPUAsynch() = default;
 

src/video_core/gpu_thread.cpp

@@ -4,6 +4,9 @@
 
 #include "common/assert.h"
 #include "common/microprofile.h"
+#include "core/core.h"
+#include "core/core_timing.h"
+#include "core/core_timing_util.h"
 #include "core/frontend/scope_acquire_window_context.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/gpu.h"
@@ -36,7 +39,6 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
                 dma_pusher.Push(std::move(submit_list->entries));
                 dma_pusher.DispatchCalls();
             } else if (const auto data = std::get_if<SwapBuffersCommand>(&next.data)) {
-                state.DecrementFramesCounter();
                 renderer.SwapBuffers(std::move(data->framebuffer));
             } else if (const auto data = std::get_if<FlushRegionCommand>(&next.data)) {
                 renderer.Rasterizer().FlushRegion(data->addr, data->size);
@@ -47,13 +49,18 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
             } else {
                 UNREACHABLE();
             }
+            state.signaled_fence = next.fence;
+            state.TrySynchronize();
         }
     }
 }
 
-ThreadManager::ThreadManager(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher)
+ThreadManager::ThreadManager(Core::System& system, VideoCore::RendererBase& renderer,
-    : renderer{renderer}, thread{RunThread, std::ref(renderer), std::ref(dma_pusher),
+                             Tegra::DmaPusher& dma_pusher)
-                                 std::ref(state)} {}
+    : system{system}, thread{RunThread, std::ref(renderer), std::ref(dma_pusher), std::ref(state)} {
+    synchronization_event = system.CoreTiming().RegisterEvent(
+        "GPUThreadSynch", [this](u64 fence, int) { state.WaitForSynchronization(fence); });
+}
 
 ThreadManager::~ThreadManager() {
     // Notify GPU thread that a shutdown is pending
@@ -62,14 +69,14 @@ ThreadManager::~ThreadManager() {
 }
 
 void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
-    PushCommand(SubmitListCommand(std::move(entries)));
+    const u64 fence{PushCommand(SubmitListCommand(std::move(entries)))};
+    const s64 synchronization_ticks{Core::Timing::usToCycles(9000)};
+    system.CoreTiming().ScheduleEvent(synchronization_ticks, synchronization_event, fence);
 }
 
 void ThreadManager::SwapBuffers(
     std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) {
-    state.IncrementFramesCounter();
     PushCommand(SwapBuffersCommand(std::move(framebuffer)));
-    state.WaitForFrames();
 }
 
 void ThreadManager::FlushRegion(CacheAddr addr, u64 size) {
@@ -79,7 +86,7 @@ void ThreadManager::FlushRegion(CacheAddr addr, u64 size) {
 void ThreadManager::InvalidateRegion(CacheAddr addr, u64 size) {
     if (state.queue.Empty()) {
         // It's quicker to invalidate a single region on the CPU if the queue is already empty
-        renderer.Rasterizer().InvalidateRegion(addr, size);
+        system.Renderer().Rasterizer().InvalidateRegion(addr, size);
     } else {
         PushCommand(InvalidateRegionCommand(addr, size));
     }
@@ -90,9 +97,25 @@ void ThreadManager::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
     InvalidateRegion(addr, size);
 }
 
-void ThreadManager::PushCommand(CommandData&& command_data) {
+u64 ThreadManager::PushCommand(CommandData&& command_data) {
-    state.queue.Push(CommandDataContainer(std::move(command_data)));
+    const u64 fence{++state.last_fence};
+    state.queue.Push(CommandDataContainer(std::move(command_data), fence));
     state.SignalCommands();
+    return fence;
+}
+
+MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
+void SynchState::WaitForSynchronization(u64 fence) {
+    if (signaled_fence >= fence) {
+        return;
+    }
+
+    // Wait for the GPU to be idle (all commands to be executed)
+    {
+        MICROPROFILE_SCOPE(GPU_wait);
+        std::unique_lock<std::mutex> lock{synchronization_mutex};
+        synchronization_condition.wait(lock, [this, fence] { return signaled_fence >= fence; });
+    }
 }
 
 } // namespace VideoCommon::GPUThread

src/video_core/gpu_thread.h

@@ -19,9 +19,12 @@ struct FramebufferConfig;
 class DmaPusher;
 } // namespace Tegra
 
-namespace VideoCore {
+namespace Core {
-class RendererBase;
+class System;
-} // namespace VideoCore
+namespace Timing {
+struct EventType;
+} // namespace Timing
+} // namespace Core
 
 namespace VideoCommon::GPUThread {
 
@@ -75,64 +78,48 @@ using CommandData =
 struct CommandDataContainer {
     CommandDataContainer() = default;
 
-    CommandDataContainer(CommandData&& data) : data{std::move(data)} {}
+    CommandDataContainer(CommandData&& data, u64 next_fence)
+        : data{std::move(data)}, fence{next_fence} {}
 
     CommandDataContainer& operator=(const CommandDataContainer& t) {
         data = std::move(t.data);
+        fence = t.fence;
         return *this;
     }
 
     CommandData data;
+    u64 fence{};
 };
 
 /// Struct used to synchronize the GPU thread
 struct SynchState final {
     std::atomic_bool is_running{true};
     std::atomic_int queued_frame_count{};
-    std::mutex frames_mutex;
+    std::mutex synchronization_mutex;
     std::mutex commands_mutex;
     std::condition_variable commands_condition;
-    std::condition_variable frames_condition;
+    std::condition_variable synchronization_condition;
 
-    void IncrementFramesCounter() {
+    /// Returns true if the gap in GPU commands is small enough that we can consider the CPU and GPU
-        std::lock_guard lock{frames_mutex};
+    /// synchronized. This is entirely empirical.
-        ++queued_frame_count;
+    bool IsSynchronized() const {
+        constexpr std::size_t max_queue_gap{5};
+        return queue.Size() <= max_queue_gap;
     }
 
-    void DecrementFramesCounter() {
+    void TrySynchronize() {
-        {
+        if (IsSynchronized()) {
-            std::lock_guard lock{frames_mutex};
+            std::lock_guard<std::mutex> lock{synchronization_mutex};
-            --queued_frame_count;
+            synchronization_condition.notify_one();
-
-            if (queued_frame_count) {
-                return;
-            }
-        }
-        frames_condition.notify_one();
-    }
-
-    void WaitForFrames() {
-        {
-            std::lock_guard lock{frames_mutex};
-            if (!queued_frame_count) {
-                return;
         }
     }
 
-        // Wait for the GPU to be idle (all commands to be executed)
+    void WaitForSynchronization(u64 fence);
-        {
-            std::unique_lock lock{frames_mutex};
-            frames_condition.wait(lock, [this] { return !queued_frame_count; });
-        }
-    }
 
     void SignalCommands() {
-        {
-            std::unique_lock lock{commands_mutex};
         if (queue.Empty()) {
             return;
         }
-        }
 
         commands_condition.notify_one();
     }
@@ -144,12 +131,15 @@ struct SynchState final {
 
     using CommandQueue = Common::SPSCQueue<CommandDataContainer>;
     CommandQueue queue;
+    u64 last_fence{};
+    std::atomic<u64> signaled_fence{};
 };
 
 /// Class used to manage the GPU thread
 class ThreadManager final {
 public:
-    explicit ThreadManager(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher);
+    explicit ThreadManager(Core::System& system, VideoCore::RendererBase& renderer,
+                           Tegra::DmaPusher& dma_pusher);
     ~ThreadManager();
 
     /// Push GPU command entries to be processed
@@ -170,11 +160,12 @@ public:
 
 private:
     /// Pushes a command to be executed by the GPU thread
-    void PushCommand(CommandData&& command_data);
+    u64 PushCommand(CommandData&& command_data);
 
 private:
     SynchState state;
-    VideoCore::RendererBase& renderer;
+    Core::System& system;
+    Core::Timing::EventType* synchronization_event{};
     std::thread thread;
     std::thread::id thread_id;
 };