This could alleviate the suffering of dual core synchronization a bit.

But I doubt you would notice it in most cases.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@4830 8ced0084-cf51-0410-be5f-012b33b47a6e

11 files changed, 108 insertions, 49 deletions

diff --git a/Source/Core/AudioCommon/Src/DSoundStream.cpp b/Source/Core/AudioCommon/Src/DSoundStream.cpp
index 2e6cd69146..3d4c868cdb 100644
--- a/Source/Core/AudioCommon/Src/DSoundStream.cpp
+++ b/Source/Core/AudioCommon/Src/DSoundStream.cpp
@@ -106,7 +106,6 @@ void DSound::SoundLoop()
 	while (!threadData)
 	{
 		// No blocking inside the csection
-		soundCriticalSection.Enter();
 		dsBuffer->GetCurrentPosition((DWORD*)&currentPos, 0);
 		int numBytesToRender = FIX128(ModBufferSize(currentPos - lastPos));
 		if (numBytesToRender >= 256)
@@ -117,7 +116,6 @@ void DSound::SoundLoop()
 			WriteDataToBuffer(lastPos, (char*)realtimeBuffer, numBytesToRender);
 			lastPos = ModBufferSize(lastPos + numBytesToRender);
 		}
-		soundCriticalSection.Leave();
 		soundSyncEvent.Wait();
 	}
 }
@@ -149,10 +147,8 @@ void DSound::SetVolume(int volume)
 	// This is in "dBA attenuation" from 0 to -10000, logarithmic
 	m_volume = (int)floor(log10((float)volume) * 5000.0f) - 10000;
 
-	soundCriticalSection.Enter();
 	if (dsBuffer != NULL)
 		dsBuffer->SetVolume(m_volume);
-	soundCriticalSection.Leave();
 }
 
 void DSound::Update()
@@ -164,7 +160,6 @@ void DSound::Clear(bool mute)
 {
 	m_muted = mute;
 
-	soundCriticalSection.Enter();
 	if (m_muted)
 	{
 		dsBuffer->Stop();
@@ -173,7 +168,6 @@ void DSound::Clear(bool mute)
 	{
 		dsBuffer->Play(0, 0, DSBPLAY_LOOPING);
 	}
-	soundCriticalSection.Leave();
 }
 
 void DSound::Stop()
@@ -182,13 +176,11 @@ void DSound::Stop()
 	// kick the thread if it's waiting
 	soundSyncEvent.Set();
 
-	soundCriticalSection.Enter();
 	delete thread;
 	thread = NULL;
 	dsBuffer->Stop();
 	dsBuffer->Release();
 	ds->Release();
-	soundCriticalSection.Leave();
 
 	soundSyncEvent.Shutdown();
 }
diff --git a/Source/Core/AudioCommon/Src/DSoundStream.h b/Source/Core/AudioCommon/Src/DSoundStream.h
index fcf0413ee7..adbeaeb468 100644
--- a/Source/Core/AudioCommon/Src/DSoundStream.h
+++ b/Source/Core/AudioCommon/Src/DSoundStream.h
@@ -32,8 +32,7 @@ class DSound : public SoundStream
 {
 #ifdef _WIN32
     Common::Thread *thread;
-    Common::CriticalSection soundCriticalSection;
-    Common::Event soundSyncEvent;
+    Common::EventEx soundSyncEvent;
     void  *hWnd;
 
     IDirectSound8* ds;
diff --git a/Source/Core/AudioCommon/Src/OpenALStream.h b/Source/Core/AudioCommon/Src/OpenALStream.h
index 57bf2e34b7..8aebf32068 100644
--- a/Source/Core/AudioCommon/Src/OpenALStream.h
+++ b/Source/Core/AudioCommon/Src/OpenALStream.h
@@ -64,8 +64,7 @@ public:
 
 private:
 	Common::Thread *thread;
-	Common::CriticalSection soundCriticalSection;
-	Common::Event soundSyncEvent;
+	Common::EventEx soundSyncEvent;
 	
 	short realtimeBuffer[OAL_MAX_SAMPLES * 2];
 	ALuint uiBuffers[OAL_NUM_BUFFERS];
diff --git a/Source/Core/Common/Src/Thread.cpp b/Source/Core/Common/Src/Thread.cpp
index bb0d591771..37f52e9ce9 100644
--- a/Source/Core/Common/Src/Thread.cpp
+++ b/Source/Core/Common/Src/Thread.cpp
@@ -117,6 +117,60 @@ void Thread::SetCurrentThreadAffinity(int mask)
 	SetThreadAffinityMask(GetCurrentThread(), mask);
 }
 
+#ifdef _WIN32
+EventEx::EventEx()
+{
+	InterlockedExchange(&m_Lock, 1);
+}
+
+void EventEx::Init()
+{
+	InterlockedExchange(&m_Lock, 1);
+}
+
+void EventEx::Shutdown()
+{
+	InterlockedExchange(&m_Lock, 0);
+}
+
+void EventEx::Set()
+{
+	InterlockedExchange(&m_Lock, 0);
+}
+
+void EventEx::Spin()
+{
+	while (InterlockedCompareExchange(&m_Lock, 1, 0))
+		// This only yields when there is a runnable thread on this core
+		// If not, spin
+		SwitchToThread();
+}
+
+void EventEx::Wait()
+{
+	while (InterlockedCompareExchange(&m_Lock, 1, 0))
+		// This directly enters Ring0 and enforces a sleep about 15ms
+		SleepCurrentThread(1);
+}
+
+bool EventEx::MsgWait()
+{
+	while (InterlockedCompareExchange(&m_Lock, 1, 0))
+	{
+		MSG msg;
+		while (PeekMessage(&msg, 0, 0, 0, PM_REMOVE))
+		{
+			if (msg.message == WM_QUIT) return false;
+			TranslateMessage(&msg);
+			DispatchMessage(&msg);
+		}
+		// This directly enters Ring0 and enforces a sleep about 15ms
+		SleepCurrentThread(1);
+	}
+	return true;
+}
+#endif
+
 // Regular same thread loop based waiting
 Event::Event()
 {
@@ -164,6 +218,7 @@ void Event::MsgWait()
 			if (msg.message == WM_QUIT)  
 				return; 
 			// Otherwise, dispatch the message.
+			TranslateMessage(&msg);
 			DispatchMessage(&msg); 
 		}
 
diff --git a/Source/Core/Common/Src/Thread.h b/Source/Core/Common/Src/Thread.h
index 4df20da844..f667cc170c 100644
--- a/Source/Core/Common/Src/Thread.h
+++ b/Source/Core/Common/Src/Thread.h
@@ -137,7 +137,31 @@ private:
 #endif
 };
 
-
+#ifdef _WIN32
+// Event(WaitForSingleObject) is too expensive
+// as it always enters Ring0 regardless of the state of lock
+// This EventEx will try to stay in Ring3 as much as possible
+// If the lock can be obtained in the first time, Ring0 won't be entered at all
+class EventEx
+{
+public:
+	EventEx();
+	void Init();
+	void Shutdown();
+	void Set();
+	// Infinite wait
+	void Spin();
+	// Infinite wait with sleep
+	void Wait();
+	// Wait with message processing and sleep
+	bool MsgWait();
+private:
+	volatile long m_Lock;
+};
+#else
+// TODO: implement for Linux
+#define EventEx	Event
+#endif
 
 class Event
 {
@@ -182,9 +206,10 @@ private:
 void InitThreading();
 void SleepCurrentThread(int ms);
 
-// YieldCPU: Use this function during a spin-wait to make the current thread
-// relax while another thread is working. This may be more efficient than using
-// events because event functions use kernel calls.
+// YieldCPU: This function is only effective on HyperThreading CPU
+// Use this function during a spin-wait to make the current thread
+// relax while another thread is working. This may be more efficient
+// than using events because event functions use kernel calls.
 inline void YieldCPU()
 {
 #ifdef _WIN32
diff --git a/Source/Core/Core/Src/Core.cpp b/Source/Core/Core/Src/Core.cpp
index 17c8ba38c9..a5fc0a91dd 100644
--- a/Source/Core/Core/Src/Core.cpp
+++ b/Source/Core/Core/Src/Core.cpp
@@ -548,13 +548,14 @@ void ScreenShot()
 // This should only be called from VI
 void VideoThrottle()
 {
-	u32 TargetVPS = (SConfig::GetInstance().m_Framelimit > 1) ? SConfig::GetInstance().m_Framelimit * 10
-		: VideoInterface::TargetRefreshRate;
+	u32 TargetVPS = (SConfig::GetInstance().m_Framelimit > 1) ?
+		SConfig::GetInstance().m_Framelimit * 10 : VideoInterface::TargetRefreshRate;
 
 	// When frame limit is NOT off
 	if (SConfig::GetInstance().m_Framelimit)
 	{
-		u32 frametime = DrawnVideo * 1000 / TargetVPS;
+		// Make the limiter a bit loose
+		u32 frametime = DrawnVideo * 1000 / ++TargetVPS;
 		while ((u32)Timer.GetTimeDifference() < frametime)
 			Common::YieldCPU();
 			//Common::SleepCurrentThread(1);
@@ -567,7 +568,7 @@ void VideoThrottle()
 		SCoreStartupParameter& _CoreParameter = SConfig::GetInstance().m_LocalCoreStartupParameter;
 
 		u32 FPS = Common::AtomicLoad(DrawnFrame) * 1000 / ElapseTime;
-		u32 VPS = DrawnVideo * 1000 / ElapseTime;
+		u32 VPS = --DrawnVideo * 1000 / ElapseTime;
 		u32 Speed = VPS * 100 / VideoInterface::TargetRefreshRate;
 		
 		// Settings are shown the same for both extended and summary info
@@ -611,7 +612,7 @@ void VideoThrottle()
 					SystemTimers::GetTicksPerSecond() / 1000000,
 					_CoreParameter.bSkipIdle ? "~" : "",
 					TicksPercentage);
-		
+
 		#else	// Summary information
 		std::string SFPS = StringFromFormat("FPS: %u - VPS: %u - SPEED: %u%%", FPS, VPS, Speed);
 		#endif
@@ -630,7 +631,7 @@ void VideoThrottle()
 		Common::AtomicStore(DrawnFrame, 0);
 		DrawnVideo = 0;
 	}
-	
+
 	DrawnVideo++;
 }
 
diff --git a/Source/Core/Core/Src/HW/SystemTimers.cpp b/Source/Core/Core/Src/HW/SystemTimers.cpp
index 8661bd7408..d284e2ef0c 100644
--- a/Source/Core/Core/Src/HW/SystemTimers.cpp
+++ b/Source/Core/Core/Src/HW/SystemTimers.cpp
@@ -132,13 +132,7 @@ int
 
 	// This is completely arbitrary. If we find that we need lower latency, we can just
 	// increase this number.
-	IPC_HLE_PERIOD,
-
-	// For DC watchdog hack
-	// Once every 4 frame-period seems to be enough (arbitrary taking 60fps as the ref).
-	// TODO: make it VI output frame rate compliant (30/60 and 25/50)
-	// Assuming game's frame-finish-watchdog wait more than 4 emulated frame-period before starting its mess.
-	FAKE_GP_WATCHDOG_PERIOD;
+	IPC_HLE_PERIOD;
 
 
 
@@ -227,7 +221,7 @@ void AdvanceCallback(int cyclesExecuted)
 void FakeGPWatchdogCallback(u64 userdata, int cyclesLate)
 {
 	CPluginManager::GetInstance().GetVideo()->Video_WaitForFrameFinish();  // lock CPUThread until frame finish
-	CoreTiming::ScheduleEvent(FAKE_GP_WATCHDOG_PERIOD-cyclesLate, et_FakeGPWD);
+	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame() - cyclesLate, et_FakeGPWD);
 }
 
 void PatchEngineCallback(u64 userdata, int cyclesLate)
@@ -235,7 +229,7 @@ void PatchEngineCallback(u64 userdata, int cyclesLate)
 	// Patch mem and run the Action Replay
 	PatchEngine::ApplyFramePatches();
 	PatchEngine::ApplyARPatches();
-	CoreTiming::ScheduleEvent((GetTicksPerSecond() / 5000) - cyclesLate, et_PatchEngine);
+	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame() - cyclesLate, et_PatchEngine);
 }
 
 void Init()
@@ -271,8 +265,6 @@ void Init()
 	if (UsingDSPLLE)
 		DSP_PERIOD = 12000; // TO BE TWEAKED
 
-	FAKE_GP_WATCHDOG_PERIOD = GetTicksPerSecond() / 60;
-
 	// This is the biggest question mark.
 	AI_PERIOD = GetTicksPerSecond() / 80;
 
@@ -297,16 +289,14 @@ void Init()
 	CoreTiming::ScheduleEvent(AI_PERIOD, et_AI);
 	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame(), et_VI);
 	CoreTiming::ScheduleEvent(DSP_PERIOD, et_DSP);
-	CoreTiming::ScheduleEvent(GetTicksPerSecond() / 60, et_SI);
+	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame(), et_SI);
 	CoreTiming::ScheduleEvent(AUDIO_DMA_PERIOD, et_AudioDMA);
 
 	// For DC watchdog hack
 	if (Core::GetStartupParameter().bCPUThread)
-	{
-		CoreTiming::ScheduleEvent(FAKE_GP_WATCHDOG_PERIOD, et_FakeGPWD);
-	}
+		CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame(), et_FakeGPWD);
 
-	CoreTiming::ScheduleEvent(GetTicksPerSecond() / 60, et_PatchEngine);
+	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame(), et_PatchEngine);
 
 	if (Core::GetStartupParameter().bWii)
 		CoreTiming::ScheduleEvent(IPC_HLE_PERIOD, et_IPC_HLE);
diff --git a/Source/Core/DebuggerWX/Src/MemoryWindow.cpp b/Source/Core/DebuggerWX/Src/MemoryWindow.cpp
index 474fac90ef..729c76ecd3 100644
--- a/Source/Core/DebuggerWX/Src/MemoryWindow.cpp
+++ b/Source/Core/DebuggerWX/Src/MemoryWindow.cpp
@@ -391,7 +391,7 @@ void CMemoryWindow::onSearch(wxCommandEvent& event){
 	}
 	if(size){ 
     unsigned char* pnt=&Dest.front();
-    int k=0;
+    unsigned int k=0;
 	//grab 
 
 	wxString txt = addrbox->GetValue();
diff --git a/Source/Core/DolphinWX/Src/LogWindow.cpp b/Source/Core/DolphinWX/Src/LogWindow.cpp
index eecae934c5..ea3bc012ad 100644
--- a/Source/Core/DolphinWX/Src/LogWindow.cpp
+++ b/Source/Core/DolphinWX/Src/LogWindow.cpp
@@ -187,7 +187,7 @@ void CLogWindow::LoadSettings()
 	m_verbosity->SetSelection(verbosity - 1);
 	ini.Get("Options", "Font", &font, 0);
 	m_FontChoice->SetSelection(font);
-	if (m_FontChoice->GetSelection() < Font.size())
+	if (m_FontChoice->GetSelection() < (int)Font.size())
 		m_Log->SetDefaultStyle(wxTextAttr(wxNullColour, wxNullColour, Font.at(m_FontChoice->GetSelection())));
 	ini.Get("Options", "WriteToFile", &m_writeFile, true);
 	m_writeFileCB->SetValue(m_writeFile);
@@ -303,7 +303,7 @@ wxTextCtrl* CLogWindow::CreateTextCtrl(wxPanel* parent, wxWindowID id, long Styl
 	TC->SetBackgroundColour(*wxBLACK);
 	if (m_FontChoice)
 	{
-		if (m_FontChoice->GetSelection() < Font.size())
+		if (m_FontChoice->GetSelection() < (int)Font.size())
 			TC->SetDefaultStyle(wxTextAttr(wxNullColour, wxNullColour, Font.at(m_FontChoice->GetSelection())));
 	}
 	return TC;
diff --git a/Source/Core/VideoCommon/Src/CommandProcessor.cpp b/Source/Core/VideoCommon/Src/CommandProcessor.cpp
index c02fd492c7..ea153180c3 100644
--- a/Source/Core/VideoCommon/Src/CommandProcessor.cpp
+++ b/Source/Core/VideoCommon/Src/CommandProcessor.cpp
@@ -105,7 +105,7 @@ int m_bboxbottom;
 u16 m_tokenReg;

 

 static u32 fake_GPWatchdogLastToken = 0;

-static Common::Event s_fifoIdleEvent;

+static Common::EventEx s_fifoIdleEvent;

 static Common::CriticalSection sFifoCritical;

 

 void FifoCriticalEnter()

@@ -381,7 +381,7 @@ void Write16(const u16 _Value, const u32 _Address)
 			// Touching that game is a no-go so I don't want to take the risk :p

 			while (fifo.bFF_GPReadEnable && ((!fifo.bFF_BPEnable && fifo.CPReadWriteDistance) || (fifo.bFF_BPEnable && !fifo.bFF_Breakpoint)))

 			{

-				s_fifoIdleEvent.MsgWait();

+				s_fifoIdleEvent.Wait();

 			}

 		}

 	}

@@ -573,7 +573,7 @@ void WaitForFrameFinish()
 {

 	while ((fake_GPWatchdogLastToken == fifo.Fake_GPWDToken) && fifo.bFF_GPReadEnable && ((!fifo.bFF_BPEnable && fifo.CPReadWriteDistance) || (fifo.bFF_BPEnable && !fifo.bFF_Breakpoint)));

 	{

-		s_fifoIdleEvent.MsgWait();

+		s_fifoIdleEvent.Wait();

 	}

 

 	fake_GPWatchdogLastToken = fifo.Fake_GPWDToken;

@@ -618,7 +618,7 @@ void STACKALIGN GatherPipeBursted()
 			// Wait for GPU to catch up

 			while (fifo.CPReadWriteDistance > fifo.CPLoWatermark && fifo.bFF_GPReadEnable && (!fifo.bFF_BPEnable || (fifo.bFF_BPEnable && !fifo.bFF_Breakpoint)))

 			{

-				s_fifoIdleEvent.MsgWait();

+				s_fifoIdleEvent.Wait();

 			}

 		}

 		// check if we are in sync

diff --git a/Source/Core/VideoCommon/Src/Fifo.cpp b/Source/Core/VideoCommon/Src/Fifo.cpp
index 172d6095fa..6ca7c3f259 100644
--- a/Source/Core/VideoCommon/Src/Fifo.cpp
+++ b/Source/Core/VideoCommon/Src/Fifo.cpp
@@ -36,7 +36,7 @@ namespace
 static volatile bool fifoStateRun = false;
 static volatile bool EmuRunning = false;
 static u8 *videoBuffer;
-static Common::Event fifo_run_event;
+static Common::EventEx fifo_run_event;
 // STATE_TO_SAVE
 static int size = 0;
 }  // namespace
@@ -170,6 +170,7 @@ void Fifo_EnterLoop(const SVideoInitialize &video_initialize)
 				{
 					Common::AtomicStore(_fifo.bFF_Breakpoint, 1);
 					CommandProcessor::UpdateInterruptsFromVideoPlugin(true);
+					CommandProcessor::FifoCriticalLeave();
 					break;
 				}
 				distToSend = 32;
@@ -208,10 +209,7 @@ void Fifo_EnterLoop(const SVideoInitialize &video_initialize)
 			// leading the CPU thread to wait in Video_BeginField or Video_AccessEFB thus slowing things down.
 			VideoFifo_CheckEFBAccess();
 			VideoFifo_CheckSwapRequest();
-			
-			CommandProcessor::SetFifoIdleFromVideoPlugin();
 		}
-
 		CommandProcessor::SetFifoIdleFromVideoPlugin();
 		if (EmuRunning)
 			Common::YieldCPU();