Merge pull request #867 from comex/gpu-determinism-pr1

Some small preparatory changes to VideoCommon for GPU thread determinism mode.

Source/Core/Common/ChunkFile.h

@@ -25,6 +25,7 @@
 
 #include "Common/Common.h"
 #include "Common/FileUtil.h"
+#include "Common/Flag.h"
 
 // ewww
 #if _LIBCPP_VERSION
@@ -165,6 +166,14 @@ public:
 			Do(x[i]);
 	}
 
+	void Do(Common::Flag& flag)
+	{
+		bool s = flag.IsSet();
+		Do(s);
+		if (mode == MODE_READ)
+			flag.Set(s);
+	}
+
 	template <typename T>
 	void Do(T& x)
 	{

Source/Core/VideoBackends/D3D/main.cpp

@@ -160,11 +160,6 @@ bool VideoBackend::Initialize(void *window_handle)
 
 void VideoBackend::Video_Prepare()
 {
-	// Better be safe...
-	s_efbAccessRequested = FALSE;
-	s_FifoShuttingDown = FALSE;
-	s_swapRequested = FALSE;
-
 	// internal interfaces
 	g_renderer = new Renderer(m_window_handle);
 	g_texture_cache = new TextureCache;
@@ -196,10 +191,6 @@ void VideoBackend::Shutdown()
 	// TODO: should be in Video_Cleanup
 	if (g_renderer)
 	{
-		s_efbAccessRequested = FALSE;
-		s_FifoShuttingDown = FALSE;
-		s_swapRequested = FALSE;
-
 		// VideoCommon
 		Fifo_Shutdown();
 		CommandProcessor::Shutdown();

Source/Core/VideoBackends/OGL/main.cpp

@@ -189,10 +189,6 @@ void VideoBackend::Video_Prepare()
 
 	g_renderer = new Renderer;
 
-	s_efbAccessRequested = false;
-	s_FifoShuttingDown = false;
-	s_swapRequested = false;
-
 	CommandProcessor::Init();
 	PixelEngine::Init();
 
@@ -230,9 +226,6 @@ void VideoBackend::Video_Cleanup()
 {
 	if (g_renderer)
 	{
-		s_efbAccessRequested = false;
-		s_FifoShuttingDown = false;
-		s_swapRequested = false;
 		Fifo_Shutdown();
 
 		// The following calls are NOT Thread Safe

Source/Core/VideoCommon/CommandProcessor.cpp

@@ -317,7 +317,7 @@ void STACKALIGN GatherPipeBursted()
 	}
 
 	if (IsOnThread())
-		SetCpStatus(true);
+		SetCPStatusFromCPU();
 
 	// update the fifo pointer
 	if (fifo.CPWritePointer >= fifo.CPEnd)
@@ -361,30 +361,17 @@ void UpdateInterruptsFromVideoBackend(u64 userdata)
 	CoreTiming::ScheduleEvent_Threadsafe(0, et_UpdateInterrupts, userdata);
 }
 
-void SetCpStatus(bool isCPUThread)
+void SetCPStatusFromGPU()
 {
-	// overflow & underflow check
-	fifo.bFF_HiWatermark = (fifo.CPReadWriteDistance > fifo.CPHiWatermark);
-	fifo.bFF_LoWatermark = (fifo.CPReadWriteDistance < fifo.CPLoWatermark);
-
 	// breakpoint
-	if (!isCPUThread)
+	if (fifo.bFF_BPEnable)
 	{
-		if (fifo.bFF_BPEnable)
+		if (fifo.CPBreakpoint == fifo.CPReadPointer)
 		{
-			if (fifo.CPBreakpoint == fifo.CPReadPointer)
+			if (!fifo.bFF_Breakpoint)
 			{
-				if (!fifo.bFF_Breakpoint)
+				INFO_LOG(COMMANDPROCESSOR, "Hit breakpoint at %i", fifo.CPReadPointer);
-				{
+				fifo.bFF_Breakpoint = true;
-					INFO_LOG(COMMANDPROCESSOR, "Hit breakpoint at %i", fifo.CPReadPointer);
-					fifo.bFF_Breakpoint = true;
-				}
-			}
-			else
-			{
-				if (fifo.bFF_Breakpoint)
-					INFO_LOG(COMMANDPROCESSOR, "Cleared breakpoint at %i", fifo.CPReadPointer);
-				fifo.bFF_Breakpoint = false;
 			}
 		}
 		else
@@ -394,6 +381,20 @@ void SetCpStatus(bool isCPUThread)
 			fifo.bFF_Breakpoint = false;
 		}
 	}
+	else
+	{
+		if (fifo.bFF_Breakpoint)
+			INFO_LOG(COMMANDPROCESSOR, "Cleared breakpoint at %i", fifo.CPReadPointer);
+		fifo.bFF_Breakpoint = false;
+	}
+	SetCPStatusFromCPU();
+}
+
+void SetCPStatusFromCPU()
+{
+	// overflow & underflow check
+	fifo.bFF_HiWatermark = (fifo.CPReadWriteDistance > fifo.CPHiWatermark);
+	fifo.bFF_LoWatermark = (fifo.CPReadWriteDistance < fifo.CPLoWatermark);
 
 	bool bpInt = fifo.bFF_Breakpoint && fifo.bFF_BPInt;
 	bool ovfInt = fifo.bFF_HiWatermark && fifo.bFF_HiWatermarkInt;
@@ -408,15 +409,14 @@ void SetCpStatus(bool isCPUThread)
 		{
 			if (!interrupt || bpInt || undfInt || ovfInt)
 			{
-				if (!isCPUThread)
+				if (Core::IsGPUThread())
 				{
-					// GPU thread:
+					// Schedule the interrupt asynchronously
 					interruptWaiting = true;
 					CommandProcessor::UpdateInterruptsFromVideoBackend(userdata);
 				}
 				else
 				{
-					// CPU thread:
 					interruptSet = interrupt;
 					INFO_LOG(COMMANDPROCESSOR,"Interrupt set");
 					ProcessorInterface::SetInterrupt(INT_CAUSE_CP, interrupt);

Source/Core/VideoCommon/CommandProcessor.h

@@ -135,7 +135,8 @@ void DoState(PointerWrap &p);
 
 void RegisterMMIO(MMIO::Mapping* mmio, u32 base);
 
-void SetCpStatus(bool isCPUThread = false);
+void SetCPStatusFromGPU();
+void SetCPStatusFromCPU();
 void GatherPipeBursted();
 void UpdateInterrupts(u64 userdata);
 void UpdateInterruptsFromVideoBackend(u64 userdata);

Source/Core/VideoCommon/Fifo.cpp

@@ -148,7 +148,7 @@ void RunGpuLoop()
 
 		VideoFifo_CheckAsyncRequest();
 
-		CommandProcessor::SetCpStatus();
+		CommandProcessor::SetCPStatusFromGPU();
 
 		Common::AtomicStore(CommandProcessor::VITicks, CommandProcessor::m_cpClockOrigin);
 
@@ -184,7 +184,7 @@ void RunGpuLoop()
 					Common::AtomicStore(fifo.SafeCPReadPointer, fifo.CPReadPointer);
 			}
 
-			CommandProcessor::SetCpStatus();
+			CommandProcessor::SetCPStatusFromGPU();
 
 			// This call is pretty important in DualCore mode and must be called in the FIFO Loop.
 			// If we don't, s_swapRequested or s_efbAccessRequested won't be set to false
@@ -247,5 +247,5 @@ void RunGpu()
 
 		fifo.CPReadWriteDistance -= 32;
 	}
-	CommandProcessor::SetCpStatus();
+	CommandProcessor::SetCPStatusFromGPU();
 }

Source/Core/VideoCommon/MainBase.cpp

@@ -1,5 +1,4 @@
-
+#include "Common/Event.h"
-#include "Common/Atomic.h"
 #include "Core/ConfigManager.h"
 
 #include "VideoCommon/BPStructs.h"
@@ -18,13 +17,13 @@
 
 bool s_BackendInitialized = false;
 
-volatile u32 s_swapRequested = false;
+Common::Flag s_swapRequested;
-u32 s_efbAccessRequested = false;
+static Common::Flag s_FifoShuttingDown;
-volatile u32 s_FifoShuttingDown = false;
+static Common::Flag s_efbAccessRequested;
+static Common::Event s_efbAccessReadyEvent;
 
-static std::condition_variable s_perf_query_cond;
+static Common::Flag s_perfQueryRequested;
-static std::mutex s_perf_query_lock;
+static Common::Event s_perfQueryReadyEvent;
-static volatile bool s_perf_query_requested;
 
 static volatile struct
 {
@@ -57,7 +56,9 @@ void VideoBackendHardware::Video_EnterLoop()
 void VideoBackendHardware::Video_ExitLoop()
 {
 	ExitGpuLoop();
-	s_FifoShuttingDown = true;
+	s_FifoShuttingDown.Set();
+	s_efbAccessReadyEvent.Set();
+	s_perfQueryReadyEvent.Set();
 }
 
 void VideoBackendHardware::Video_SetRendering(bool bEnabled)
@@ -70,11 +71,11 @@ static void VideoFifo_CheckSwapRequest()
 {
 	if (g_ActiveConfig.bUseXFB)
 	{
-		if (Common::AtomicLoadAcquire(s_swapRequested))
+		if (s_swapRequested.IsSet())
 		{
 			EFBRectangle rc;
 			Renderer::Swap(s_beginFieldArgs.xfbAddr, s_beginFieldArgs.fbWidth, s_beginFieldArgs.fbHeight,rc);
-			Common::AtomicStoreRelease(s_swapRequested, false);
+			s_swapRequested.Clear();
 		}
 	}
 }
@@ -84,7 +85,7 @@ void VideoFifo_CheckSwapRequestAt(u32 xfbAddr, u32 fbWidth, u32 fbHeight)
 {
 	if (g_ActiveConfig.bUseXFB)
 	{
-		if (Common::AtomicLoadAcquire(s_swapRequested))
+		if (s_swapRequested.IsSet())
 		{
 			u32 aLower = xfbAddr;
 			u32 aUpper = xfbAddr + 2 * fbWidth * fbHeight;
@@ -115,7 +116,7 @@ void VideoBackendHardware::Video_EndField()
 {
 	if (s_BackendInitialized)
 	{
-		Common::AtomicStoreRelease(s_swapRequested, true);
+		s_swapRequested.Set();
 	}
 }
 
@@ -138,11 +139,11 @@ bool VideoBackendHardware::Video_Screenshot(const std::string& filename)
 
 void VideoFifo_CheckEFBAccess()
 {
-	if (Common::AtomicLoadAcquire(s_efbAccessRequested))
+	if (s_efbAccessRequested.IsSet())
 	{
 		s_AccessEFBResult = g_renderer->AccessEFB(s_accessEFBArgs.type, s_accessEFBArgs.x, s_accessEFBArgs.y, s_accessEFBArgs.Data);
-
+		s_efbAccessRequested.Clear();
-		Common::AtomicStoreRelease(s_efbAccessRequested, false);
+		s_efbAccessReadyEvent.Set();
 	}
 }
 
@@ -155,13 +156,15 @@ u32 VideoBackendHardware::Video_AccessEFB(EFBAccessType type, u32 x, u32 y, u32
 		s_accessEFBArgs.y = y;
 		s_accessEFBArgs.Data = InputData;
 
-		Common::AtomicStoreRelease(s_efbAccessRequested, true);
+		s_efbAccessRequested.Set();
 
 		if (SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread)
 		{
-			while (Common::AtomicLoadAcquire(s_efbAccessRequested) && !s_FifoShuttingDown)
+			s_efbAccessReadyEvent.Reset();
-				//Common::SleepCurrentThread(1);
+			if (s_FifoShuttingDown.IsSet())
-				Common::YieldCPU();
+				return 0;
+			s_efbAccessRequested.Set();
+			s_efbAccessReadyEvent.Wait();
 		}
 		else
 			VideoFifo_CheckEFBAccess();
@@ -172,23 +175,13 @@ u32 VideoBackendHardware::Video_AccessEFB(EFBAccessType type, u32 x, u32 y, u32
 	return 0;
 }
 
-static bool QueryResultIsReady()
-{
-	return !s_perf_query_requested || s_FifoShuttingDown;
-}
-
 static void VideoFifo_CheckPerfQueryRequest()
 {
-	if (s_perf_query_requested)
+	if (s_perfQueryRequested.IsSet())
 	{
 		g_perf_query->FlushResults();
-
+		s_perfQueryRequested.Clear();
-		{
+		s_perfQueryReadyEvent.Set();
-		std::lock_guard<std::mutex> lk(s_perf_query_lock);
-		s_perf_query_requested = false;
-		}
-
-		s_perf_query_cond.notify_one();
 	}
 }
 
@@ -204,9 +197,11 @@ u32 VideoBackendHardware::Video_GetQueryResult(PerfQueryType type)
 	{
 		if (SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread)
 		{
-			s_perf_query_requested = true;
+			s_perfQueryReadyEvent.Reset();
-			std::unique_lock<std::mutex> lk(s_perf_query_lock);
+			if (s_FifoShuttingDown.IsSet())
-			s_perf_query_cond.wait(lk, QueryResultIsReady);
+				return 0;
+			s_perfQueryRequested.Set();
+			s_perfQueryReadyEvent.Wait();
 		}
 		else
 			g_perf_query->FlushResults();
@@ -219,10 +214,10 @@ void VideoBackendHardware::InitializeShared()
 {
 	VideoCommon_Init();
 
-	s_swapRequested = 0;
+	s_swapRequested.Clear();
-	s_efbAccessRequested = 0;
+	s_efbAccessRequested.Clear();
-	s_perf_query_requested = false;
+	s_perfQueryRequested.Clear();
-	s_FifoShuttingDown = 0;
+	s_FifoShuttingDown.Clear();
 	memset((void*)&s_beginFieldArgs, 0, sizeof(s_beginFieldArgs));
 	memset(&s_accessEFBArgs, 0, sizeof(s_accessEFBArgs));
 	s_AccessEFBResult = 0;

Source/Core/VideoCommon/MainBase.h

@@ -1,11 +1,10 @@
 #pragma once
 
 #include "Common/CommonTypes.h"
+#include "Common/Flag.h"
 
 extern bool s_BackendInitialized;
-extern u32 s_efbAccessRequested;
+extern Common::Flag s_swapRequested;
-extern volatile u32 s_FifoShuttingDown;
-extern volatile u32 s_swapRequested;
 
 void VideoFifo_CheckEFBAccess();
 void VideoFifo_CheckSwapRequestAt(u32 xfbAddr, u32 fbWidth, u32 fbHeight);

Source/Core/VideoCommon/OpcodeDecoding.cpp

@@ -469,12 +469,13 @@ void OpcodeDecoder_Shutdown()
 u32 OpcodeDecoder_Run(bool skipped_frame)
 {
 	u32 totalCycles = 0;
-	u32 cycles = FifoCommandRunnable();
+	while (true)
-	while (cycles > 0)
 	{
+		u32 cycles = FifoCommandRunnable();
+		if (cycles == 0)
+			break;
 		skipped_frame ? DecodeSemiNop() : Decode();
 		totalCycles += cycles;
-		cycles = FifoCommandRunnable();
 	}
 	return totalCycles;
 }

Source/Core/VideoCommon/RenderBase.cpp

@@ -126,7 +126,7 @@ void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidt
 	else
 	{
 		Swap(xfbAddr, fbWidth, fbHeight,sourceRc,Gamma);
-		Common::AtomicStoreRelease(s_swapRequested, false);
+		s_swapRequested.Clear();
 	}
 }