VideoCommon: implement swap requests in the full async way

6 files changed, 42 insertions, 51 deletions

diff --git a/Source/Core/VideoCommon/AsyncRequests.cpp b/Source/Core/VideoCommon/AsyncRequests.cpp
index dce56c0d7b..9430d390fa 100644
--- a/Source/Core/VideoCommon/AsyncRequests.cpp
+++ b/Source/Core/VideoCommon/AsyncRequests.cpp
@@ -4,7 +4,7 @@
 AsyncRequests AsyncRequests::s_singleton;
 
 AsyncRequests::AsyncRequests()
-: m_enable(false)
+: m_enable(false), m_passthrough(true)
 {
 }
 
@@ -34,6 +34,13 @@ void AsyncRequests::PullEventsInternal()
 void AsyncRequests::PushEvent(const AsyncRequests::Event& event, bool blocking)
 {
 	std::unique_lock<std::mutex> lock(m_mutex);
+
+	if (m_passthrough)
+	{
+		HandleEvent(event);
+		return;
+	}
+
 	m_empty.store(false);
 	m_wake_me_up_again |= blocking;
 
@@ -65,6 +72,7 @@ void AsyncRequests::SetEnable(bool enable)
 
 void AsyncRequests::HandleEvent(const AsyncRequests::Event& e)
 {
+	EFBRectangle rc;
 	switch (e.type)
 	{
 		case Event::EFB_POKE_COLOR:
@@ -83,6 +91,16 @@ void AsyncRequests::HandleEvent(const AsyncRequests::Event& e)
 			*e.efb_peek.data = g_renderer->AccessEFB(PEEK_Z, e.efb_peek.x, e.efb_peek.y, 0);
 			break;
 
+		case Event::SWAP_EVENT:
+			Renderer::Swap(e.swap_event.xfbAddr, e.swap_event.fbWidth, e.swap_event.fbStride, e.swap_event.fbHeight, rc);
+			break;
+
 	}
 }
 
+void AsyncRequests::SetPassthrough(bool enable)
+{
+	std::unique_lock<std::mutex> lock(m_mutex);
+	m_passthrough = enable;
+}
+
diff --git a/Source/Core/VideoCommon/AsyncRequests.h b/Source/Core/VideoCommon/AsyncRequests.h
index d3cfece7eb..fe3d5c3234 100644
--- a/Source/Core/VideoCommon/AsyncRequests.h
+++ b/Source/Core/VideoCommon/AsyncRequests.h
@@ -22,6 +22,7 @@ public:
 			EFB_POKE_Z,
 			EFB_PEEK_COLOR,
 			EFB_PEEK_Z,
+			SWAP_EVENT,
 		} type;
 		u64 time;
 
@@ -40,6 +41,14 @@ public:
 				u16 y;
 				u32* data;
 			} efb_peek;
+
+			struct
+			{
+				u32 xfbAddr;
+				u32 fbWidth;
+				u32 fbStride;
+				u32 fbHeight;
+			} swap_event;
 		};
 	};
 
@@ -52,6 +61,7 @@ public:
 	}
 	void PushEvent(const Event& event, bool blocking = false);
 	void SetEnable(bool enable);
+	void SetPassthrough(bool enable);
 
 	static AsyncRequests* GetInstance() { return &s_singleton; }
 
@@ -68,4 +78,5 @@ private:
 
 	bool m_wake_me_up_again;
 	bool m_enable;
+	bool m_passthrough;
 };
diff --git a/Source/Core/VideoCommon/Fifo.cpp b/Source/Core/VideoCommon/Fifo.cpp
index b8f4002d4d..74470a9fcd 100644
--- a/Source/Core/VideoCommon/Fifo.cpp
+++ b/Source/Core/VideoCommon/Fifo.cpp
@@ -284,6 +284,7 @@ void RunGpuLoop()
 	bool yield_cpu = cpu_info.num_cores <= 2;
 
 	AsyncRequests::GetInstance()->SetEnable(true);
+	AsyncRequests::GetInstance()->SetPassthrough(false);
 
 	while (GpuRunningState)
 	{
@@ -383,6 +384,7 @@ void RunGpuLoop()
 	// wake up SyncGPU if we were interrupted
 	s_video_buffer_cond.notify_all();
 	AsyncRequests::GetInstance()->SetEnable(false);
+	AsyncRequests::GetInstance()->SetPassthrough(true);
 }
 
 
diff --git a/Source/Core/VideoCommon/MainBase.cpp b/Source/Core/VideoCommon/MainBase.cpp
index 47eab7d941..25741ea7d8 100644
--- a/Source/Core/VideoCommon/MainBase.cpp
+++ b/Source/Core/VideoCommon/MainBase.cpp
@@ -19,7 +19,6 @@
 
 bool s_BackendInitialized = false;
 
-Common::Flag s_swapRequested;
 static Common::Flag s_FifoShuttingDown;
 
 static Common::Flag s_perfQueryRequested;
@@ -61,45 +60,11 @@ void VideoBackendHardware::Video_SetRendering(bool bEnabled)
 	Fifo_SetRendering(bEnabled);
 }
 
-// Run from the graphics thread (from Fifo.cpp)
-static void VideoFifo_CheckSwapRequest()
-{
-	if (g_ActiveConfig.bUseXFB)
-	{
-		if (s_swapRequested.IsSet())
-		{
-			EFBRectangle rc;
-			Renderer::Swap(s_beginFieldArgs.xfbAddr, s_beginFieldArgs.fbWidth, s_beginFieldArgs.fbStride, s_beginFieldArgs.fbHeight, rc);
-			s_swapRequested.Clear();
-		}
-	}
-}
-
-// Run from the graphics thread (from Fifo.cpp)
-void VideoFifo_CheckSwapRequestAt(u32 xfbAddr, u32 fbWidth, u32 fbHeight)
-{
-	if (g_ActiveConfig.bUseXFB)
-	{
-		if (s_swapRequested.IsSet())
-		{
-			u32 aLower = xfbAddr;
-			u32 aUpper = xfbAddr + 2 * fbWidth * fbHeight;
-			u32 bLower = s_beginFieldArgs.xfbAddr;
-			u32 bUpper = s_beginFieldArgs.xfbAddr + 2 * s_beginFieldArgs.fbStride * s_beginFieldArgs.fbHeight;
-
-			if (AddressRangesOverlap(aLower, aUpper, bLower, bUpper))
-				VideoFifo_CheckSwapRequest();
-		}
-	}
-}
-
 // Run from the CPU thread (from VideoInterface.cpp)
 void VideoBackendHardware::Video_BeginField(u32 xfbAddr, u32 fbWidth, u32 fbStride, u32 fbHeight)
 {
 	if (s_BackendInitialized && g_ActiveConfig.bUseXFB)
 	{
-		if (!SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread)
-			VideoFifo_CheckSwapRequest();
 		s_beginFieldArgs.xfbAddr = xfbAddr;
 		s_beginFieldArgs.fbWidth = fbWidth;
 		s_beginFieldArgs.fbStride = fbStride;
@@ -110,16 +75,19 @@ void VideoBackendHardware::Video_BeginField(u32 xfbAddr, u32 fbWidth, u32 fbStri
 // Run from the CPU thread (from VideoInterface.cpp)
 void VideoBackendHardware::Video_EndField()
 {
-	if (s_BackendInitialized)
+	if (s_BackendInitialized && g_ActiveConfig.bUseXFB && g_renderer)
 	{
 		SyncGPU(SYNC_GPU_SWAP);
 
-		// Wait until the GPU thread has swapped. Prevents FIFO overflows.
-		while (g_ActiveConfig.bUseXFB && SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread && s_swapRequested.IsSet())
-		{
-			Common::YieldCPU();
-		}
-		s_swapRequested.Set();
+		AsyncRequests::Event e;
+		e.time = 0;
+		e.type = AsyncRequests::Event::SWAP_EVENT;
+
+		e.swap_event.xfbAddr = s_beginFieldArgs.xfbAddr;
+		e.swap_event.fbWidth = s_beginFieldArgs.fbWidth;
+		e.swap_event.fbStride = s_beginFieldArgs.fbStride;
+		e.swap_event.fbHeight = s_beginFieldArgs.fbHeight;
+		AsyncRequests::GetInstance()->PushEvent(e, false);
 	}
 }
 
@@ -246,7 +214,6 @@ void VideoBackendHardware::InitializeShared()
 {
 	VideoCommon_Init();
 
-	s_swapRequested.Clear();
 	s_perfQueryRequested.Clear();
 	s_FifoShuttingDown.Clear();
 	memset((void*)&s_beginFieldArgs, 0, sizeof(s_beginFieldArgs));
@@ -268,7 +235,6 @@ void VideoBackendHardware::DoState(PointerWrap& p)
 	VideoCommon_DoState(p);
 	p.DoMarker("VideoCommon");
 
-	p.Do(s_swapRequested);
 	p.Do(s_beginFieldArgs);
 	p.DoMarker("VideoBackendHardware");
 
@@ -308,7 +274,6 @@ void VideoBackendHardware::RunLoop(bool enable)
 
 void VideoFifo_CheckAsyncRequest()
 {
-	VideoFifo_CheckSwapRequest();
 	VideoFifo_CheckPerfQueryRequest();
 	VideoFifo_CheckBBoxRequest();
 }
diff --git a/Source/Core/VideoCommon/MainBase.h b/Source/Core/VideoCommon/MainBase.h
index b8eebf37a0..c9d059cf7a 100644
--- a/Source/Core/VideoCommon/MainBase.h
+++ b/Source/Core/VideoCommon/MainBase.h
@@ -4,6 +4,3 @@
 #include "Common/Flag.h"
 
 extern bool s_BackendInitialized;
-extern Common::Flag s_swapRequested;
-
-void VideoFifo_CheckSwapRequestAt(u32 xfbAddr, u32 fbWidth, u32 fbHeight);
diff --git a/Source/Core/VideoCommon/RenderBase.cpp b/Source/Core/VideoCommon/RenderBase.cpp
index cb9be880e0..ccceb954dd 100644
--- a/Source/Core/VideoCommon/RenderBase.cpp
+++ b/Source/Core/VideoCommon/RenderBase.cpp
@@ -116,7 +116,6 @@ void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidt
 	if (!fbWidth || !fbHeight)
 		return;
 
-	VideoFifo_CheckSwapRequestAt(xfbAddr, fbWidth, fbHeight);
 	XFBWrited = true;
 
 	if (g_ActiveConfig.bUseXFB)
@@ -126,7 +125,6 @@ void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidt
 	else
 	{
 		Swap(xfbAddr, fbWidth, fbWidth, fbHeight, sourceRc, Gamma);
-		s_swapRequested.Clear();
 	}
 }