Merge pull request #2912 from FernandoS27/async-fixes

General fixes to Async GPU

src/video_core/gpu.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "common/assert.h"
+#include "common/microprofile.h"
 #include "core/core.h"
 #include "core/core_timing.h"
 #include "core/memory.h"
@@ -17,6 +18,8 @@
 
 namespace Tegra {
 
+MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
+
 GPU::GPU(Core::System& system, VideoCore::RendererBase& renderer, bool is_async)
     : system{system}, renderer{renderer}, is_async{is_async} {
     auto& rasterizer{renderer.Rasterizer()};
@@ -63,6 +66,16 @@ const DmaPusher& GPU::DmaPusher() const {
     return *dma_pusher;
 }
 
+void GPU::WaitFence(u32 syncpoint_id, u32 value) const {
+    // Synced GPU, is always in sync
+    if (!is_async) {
+        return;
+    }
+    MICROPROFILE_SCOPE(GPU_wait);
+    while (syncpoints[syncpoint_id].load(std::memory_order_relaxed) < value) {
+    }
+}
+
 void GPU::IncrementSyncPoint(const u32 syncpoint_id) {
     syncpoints[syncpoint_id]++;
     std::lock_guard lock{sync_mutex};

src/video_core/gpu.h

@@ -177,6 +177,12 @@ public:
     /// Returns a reference to the GPU DMA pusher.
     Tegra::DmaPusher& DmaPusher();
 
+    // Waits for the GPU to finish working
+    virtual void WaitIdle() const = 0;
+
+    /// Allows the CPU/NvFlinger to wait on the GPU before presenting a frame.
+    void WaitFence(u32 syncpoint_id, u32 value) const;
+
     void IncrementSyncPoint(u32 syncpoint_id);
 
     u32 GetSyncpointValue(u32 syncpoint_id) const;

src/video_core/gpu_asynch.cpp

@@ -44,4 +44,8 @@ void GPUAsynch::TriggerCpuInterrupt(const u32 syncpoint_id, const u32 value) con
     interrupt_manager.GPUInterruptSyncpt(syncpoint_id, value);
 }
 
+void GPUAsynch::WaitIdle() const {
+    gpu_thread.WaitIdle();
+}
+
 } // namespace VideoCommon

src/video_core/gpu_asynch.h

@@ -25,6 +25,7 @@ public:
     void FlushRegion(CacheAddr addr, u64 size) override;
     void InvalidateRegion(CacheAddr addr, u64 size) override;
     void FlushAndInvalidateRegion(CacheAddr addr, u64 size) override;
+    void WaitIdle() const override;
 
 protected:
     void TriggerCpuInterrupt(u32 syncpoint_id, u32 value) const override;

src/video_core/gpu_synch.h

@@ -24,6 +24,7 @@ public:
     void FlushRegion(CacheAddr addr, u64 size) override;
     void InvalidateRegion(CacheAddr addr, u64 size) override;
     void FlushAndInvalidateRegion(CacheAddr addr, u64 size) override;
+    void WaitIdle() const override {}
 
 protected:
     void TriggerCpuInterrupt([[maybe_unused]] u32 syncpoint_id,

src/video_core/gpu_thread.cpp

@@ -5,8 +5,6 @@
 #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"
@@ -68,14 +66,10 @@ ThreadManager::~ThreadManager() {
 
 void ThreadManager::StartThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher) {
     thread = std::thread{RunThread, std::ref(renderer), std::ref(dma_pusher), std::ref(state)};
-    synchronization_event = system.CoreTiming().RegisterEvent(
-        "GPUThreadSynch", [this](u64 fence, s64) { state.WaitForSynchronization(fence); });
 }
 
 void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
-    const u64 fence{PushCommand(SubmitListCommand(std::move(entries)))};
-    const s64 synchronization_ticks{Core::Timing::usToCycles(std::chrono::microseconds{9000})};
-    system.CoreTiming().ScheduleEvent(synchronization_ticks, synchronization_event, fence);
+    PushCommand(SubmitListCommand(std::move(entries)));
 }
 
 void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
@@ -96,16 +90,15 @@ void ThreadManager::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
     InvalidateRegion(addr, size);
 }
 
+void ThreadManager::WaitIdle() const {
+    while (state.last_fence > state.signaled_fence.load(std::memory_order_relaxed)) {
+    }
+}
+
 u64 ThreadManager::PushCommand(CommandData&& command_data) {
     const u64 fence{++state.last_fence};
     state.queue.Push(CommandDataContainer(std::move(command_data), fence));
     return fence;
 }
 
-MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
-void SynchState::WaitForSynchronization(u64 fence) {
-    while (signaled_fence.load() < fence)
-        ;
-}
-
 } // namespace VideoCommon::GPUThread

src/video_core/gpu_thread.h

@@ -21,9 +21,6 @@ class DmaPusher;
 
 namespace Core {
 class System;
-namespace Timing {
-struct EventType;
-} // namespace Timing
 } // namespace Core
 
 namespace VideoCommon::GPUThread {
@@ -89,8 +86,6 @@ struct CommandDataContainer {
 struct SynchState final {
     std::atomic_bool is_running{true};
 
-    void WaitForSynchronization(u64 fence);
-
     using CommandQueue = Common::SPSCQueue<CommandDataContainer>;
     CommandQueue queue;
     u64 last_fence{};
@@ -121,6 +116,9 @@ public:
     /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
     void FlushAndInvalidateRegion(CacheAddr addr, u64 size);
 
+    // Wait until the gpu thread is idle.
+    void WaitIdle() const;
+
 private:
     /// Pushes a command to be executed by the GPU thread
     u64 PushCommand(CommandData&& command_data);
@@ -128,7 +126,6 @@ private:
 private:
     SynchState state;
     Core::System& system;
-    Core::Timing::EventType* synchronization_event{};
     std::thread thread;
     std::thread::id thread_id;
 };

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -348,6 +348,7 @@ static constexpr auto RangeFromInterval(Map& map, const Interval& interval) {
 }
 
 void RasterizerOpenGL::UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
+    std::lock_guard lock{pages_mutex};
     const u64 page_start{addr >> Memory::PAGE_BITS};
     const u64 page_end{(addr + size + Memory::PAGE_SIZE - 1) >> Memory::PAGE_BITS};
 

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -9,6 +9,7 @@
 #include <cstddef>
 #include <map>
 #include <memory>
+#include <mutex>
 #include <optional>
 #include <tuple>
 #include <utility>
@@ -230,6 +231,8 @@ private:
 
     using CachedPageMap = boost::icl::interval_map<u64, int>;
     CachedPageMap cached_pages;
+
+    std::mutex pages_mutex;
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -102,8 +102,6 @@ RendererOpenGL::RendererOpenGL(Core::Frontend::EmuWindow& emu_window, Core::Syst
 RendererOpenGL::~RendererOpenGL() = default;
 
 void RendererOpenGL::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
-    system.GetPerfStats().EndSystemFrame();
-
     // Maintain the rasterizer's state as a priority
     OpenGLState prev_state = OpenGLState::GetCurState();
     state.AllDirty();
@@ -135,9 +133,6 @@ void RendererOpenGL::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
 
     render_window.PollEvents();
 
-    system.FrameLimiter().DoFrameLimiting(system.CoreTiming().GetGlobalTimeUs());
-    system.GetPerfStats().BeginSystemFrame();
-
     // Restore the rasterizer state
     prev_state.AllDirty();
     prev_state.Apply();