Fifo: Replace busy loop with condition variable

Source/Core/VideoCommon/AsyncRequests.cpp

@@ -1,4 +1,5 @@
 #include "VideoCommon/AsyncRequests.h"
+#include "VideoCommon/Fifo.h"
 #include "VideoCommon/RenderBase.h"
 
 AsyncRequests AsyncRequests::s_singleton;
@@ -49,6 +50,7 @@ void AsyncRequests::PushEvent(const AsyncRequests::Event& event, bool blocking)
 
 	m_queue.push(event);
 
+	RunGpu();
 	if (blocking)
 	{
 		m_cond.wait(lock, [this]{return m_queue.empty();});

Source/Core/VideoCommon/CommandProcessor.cpp

@@ -322,10 +322,7 @@ void GatherPipeBursted()
 				ProcessFifoAllDistance();
 			}
 		}
-		else
-		{
 		RunGpu();
-		}
 		return;
 	}
 
@@ -375,6 +372,7 @@ void UpdateInterrupts(u64 userdata)
 	}
 	CoreTiming::ForceExceptionCheck(0);
 	interruptWaiting = false;
+	RunGpu();
 }
 
 void UpdateInterruptsFromVideoBackend(u64 userdata)
@@ -551,5 +549,7 @@ void Update()
 
 	if (fifo.isGpuReadingData)
 		Common::AtomicAdd(VITicks, SystemTimers::GetTicksPerSecond() / 10000);
+
+	RunGpu();
 }
 } // end of namespace CommandProcessor

Source/Core/VideoCommon/Fifo.cpp

@@ -5,6 +5,7 @@
 #include "Common/Atomic.h"
 #include "Common/ChunkFile.h"
 #include "Common/CPUDetect.h"
+#include "Common/Event.h"
 #include "Common/FPURoundMode.h"
 #include "Common/MemoryUtil.h"
 #include "Common/Thread.h"
@@ -58,6 +59,9 @@ static u8* s_video_buffer_pp_read_ptr;
 // polls, it's just atomic.
 // - The pp_read_ptr is the CPU preprocessing version of the read_ptr.
 
+static Common::Flag s_gpu_is_running; // If this one is set, the gpu loop will be called at least once again
+static Common::Event s_gpu_new_work_event;
+
 void Fifo_DoState(PointerWrap &p)
 {
 	p.DoArray(s_video_buffer, FIFO_SIZE);
@@ -133,11 +137,13 @@ void ExitGpuLoop()
 	// Terminate GPU thread loop
 	GpuRunningState = false;
 	EmuRunningState = true;
+	s_gpu_new_work_event.Set();
 }
 
 void EmulatorState(bool running)
 {
 	EmuRunningState = running;
+	s_gpu_new_work_event.Set();
 }
 
 void SyncGPU(SyncGPUReason reason, bool may_move_read_ptr)
@@ -271,10 +277,6 @@ void RunGpuLoop()
 	SCPFifoStruct &fifo = CommandProcessor::fifo;
 	u32 cyclesExecuted = 0;
 
-	// If the host CPU has only two cores, idle loop instead of busy loop
-	// This allows a system that we are maxing out in dual core mode to do other things
-	bool yield_cpu = cpu_info.num_cores <= 2;
-
 	AsyncRequests::GetInstance()->SetEnable(true);
 	AsyncRequests::GetInstance()->SetPassthrough(false);
 
@@ -353,11 +355,15 @@ void RunGpuLoop()
 
 		if (EmuRunningState)
 		{
-			// NOTE(jsd): Calling SwitchToThread() on Windows 7 x64 is a hot spot, according to profiler.
+			if (s_gpu_is_running.IsSet())
-			// See https://docs.google.com/spreadsheet/ccc?key=0Ah4nh0yGtjrgdFpDeF9pS3V6RUotRVE3S3J4TGM1NlE#gid=0
+			{
-			// for benchmark details.
+				// reset the atomic flag. But as the CPU thread might have pushed some new data, we have to rerun the GPU loop
-			if (yield_cpu)
+				s_gpu_is_running.Clear();
-				Common::YieldCPU();
+			}
+			else
+			{
+				s_gpu_new_work_event.WaitFor(std::chrono::milliseconds(100));
+			}
 		}
 		else
 		{
@@ -386,11 +392,11 @@ bool AtBreakpoint()
 
 void RunGpu()
 {
-	if (SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread &&
-	    !g_use_deterministic_gpu_thread)
-		return;
-
 	SCPFifoStruct &fifo = CommandProcessor::fifo;
+
+	// execute GPU
+	if (!SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread || g_use_deterministic_gpu_thread)
+	{
 		while (fifo.bFF_GPReadEnable && fifo.CPReadWriteDistance && !AtBreakpoint() )
 		{
 			if (g_use_deterministic_gpu_thread)
@@ -416,6 +422,14 @@ void RunGpu()
 			fifo.CPReadWriteDistance -= 32;
 		}
 		CommandProcessor::SetCPStatusFromGPU();
+	}
+
+	// wake up GPU thread
+	if (SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread && !s_gpu_is_running.IsSet())
+	{
+		s_gpu_is_running.Set();
+		s_gpu_new_work_event.Set();
+	}
 }
 
 void Fifo_UpdateWantDeterminism(bool want)