GPU_Async: Make the GPU thread use the Event pattern

This commit makes the GPU use the Event pattern for syncing instead of
spin-looping while waiting for commands. This should make better use of
the GPU's thread time.

src/video_core/gpu_thread.cpp

@@ -19,8 +19,7 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
     MicroProfileOnThreadCreate("GpuThread");
 
     // Wait for first GPU command before acquiring the window context
-    while (state.queue.Empty())
-        ;
+    state.sync_event.Wait();
 
     // If emulation was stopped during disk shader loading, abort before trying to acquire context
     if (!state.is_running) {
@@ -31,22 +30,25 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
 
     CommandDataContainer next;
     while (state.is_running) {
-        next = state.queue.PopWait();
-        if (const auto submit_list = std::get_if<SubmitListCommand>(&next.data)) {
-            dma_pusher.Push(std::move(submit_list->entries));
-            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<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);
-        } else if (std::holds_alternative<EndProcessingCommand>(next.data)) {
-            return;
-        } else {
-            UNREACHABLE();
+        while (!state.queue.Empty()) {
+            state.queue.Pop(next);
+            if (const auto submit_list = std::get_if<SubmitListCommand>(&next.data)) {
+                dma_pusher.Push(std::move(submit_list->entries));
+                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<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);
+            } else if (std::holds_alternative<EndProcessingCommand>(next.data)) {
+                return;
+            } else {
+                UNREACHABLE();
+            }
+            state.signaled_fence.store(next.fence);
         }
-        state.signaled_fence.store(next.fence);
+        state.sync_event.Wait();
     }
 }
 
@@ -96,6 +98,7 @@ void ThreadManager::WaitIdle() const {
 u64 ThreadManager::PushCommand(CommandData&& command_data) {
     const u64 fence{++state.last_fence};
     state.queue.Push(CommandDataContainer(std::move(command_data), fence));
+    state.sync_event.Set();
     return fence;
 }
 

src/video_core/gpu_thread.h

@@ -11,6 +11,7 @@
 #include <thread>
 #include <variant>
 
+#include "common/thread.h"
 #include "common/threadsafe_queue.h"
 #include "video_core/gpu.h"
 
@@ -90,6 +91,7 @@ struct SynchState final {
     CommandQueue queue;
     u64 last_fence{};
     std::atomic<u64> signaled_fence{};
+    Common::Event sync_event{};
 };
 
 /// Class used to manage the GPU thread