Cocoa Port: Try to fix some more synchronization issues in Metal display views.

- Also remove the usage of _rwlockFrame and change it to a simple
pthread_mutex_t, since CocoaDSDisplay objects no longer have a need for
a full pthread_rwlock_t.

desmume/src/frontend/cocoa/cocoa_GPU.h

@@ -103,7 +103,6 @@ typedef std::map<CGDirectDisplayID, int64_t> DisplayLinkFlushTimeLimitMap;
 @property (assign) NSSize gpuDimensions;
 @property (assign) NSUInteger gpuScale;
 @property (assign) NSUInteger gpuColorFormat;
-@property (readonly) pthread_rwlock_t *gpuFrameRWLock;
 
 @property (assign) BOOL layerMainGPU;
 @property (assign) BOOL layerMainBG0;

desmume/src/frontend/cocoa/cocoa_GPU.mm

@@ -52,7 +52,7 @@ class GPUEventHandlerOSX : public GPUEventHandlerDefault
 private:
 	GPUClientFetchObject *_fetchObject;
 	
-	pthread_rwlock_t _rwlockFrame;
+	pthread_mutex_t _mutexFrame;
 	pthread_mutex_t _mutex3DRender;
 	pthread_mutex_t _mutexApplyGPUSettings;
 	pthread_mutex_t _mutexApplyRender3DSettings;
@@ -65,8 +65,7 @@ public:
 	GPUClientFetchObject* GetFetchObject() const;
 	void SetFetchObject(GPUClientFetchObject *fetchObject);
 	
-	void FramebufferLockWrite();
-	void FramebufferLockRead();
+	void FramebufferLock();
 	void FramebufferUnlock();
 	void Render3DLock();
 	void Render3DUnlock();
@@ -75,7 +74,6 @@ public:
 	void ApplyRender3DSettingsLock();
 	void ApplyRender3DSettingsUnlock();
 	
-	pthread_rwlock_t* GetFrameRWLock();
 	bool GetRender3DNeedsFinish();
 	
 	virtual void DidFrameBegin(bool isFrameSkipRequested, const u8 targetBufferIndex, const size_t line);
@@ -94,7 +92,6 @@ public:
 @dynamic gpuDimensions;
 @dynamic gpuScale;
 @dynamic gpuColorFormat;
-@dynamic gpuFrameRWLock;
 
 @dynamic layerMainGPU;
 @dynamic layerMainBG0;
@@ -256,7 +253,7 @@ public:
 - (void) setGpuDimensions:(NSSize)theDimensions
 {
 	gpuEvent->Render3DLock();
-	gpuEvent->FramebufferLockWrite();
+	gpuEvent->FramebufferLock();
 	
 #ifdef ENABLE_SHARED_FETCH_OBJECT
 	pthread_rwlock_wrlock([[self sharedData] rwlockFramebufferAtIndex:0]);
@@ -278,7 +275,7 @@ public:
 - (NSSize) gpuDimensions
 {
 	gpuEvent->Render3DLock();
-	gpuEvent->FramebufferLockRead();
+	gpuEvent->FramebufferLock();
 	const NSSize dimensions = NSMakeSize(GPU->GetCustomFramebufferWidth(), GPU->GetCustomFramebufferHeight());
 	gpuEvent->FramebufferUnlock();
 	gpuEvent->Render3DUnlock();
@@ -313,7 +310,7 @@ public:
 	
 	// Change the color format.
 	gpuEvent->Render3DLock();
-	gpuEvent->FramebufferLockWrite();
+	gpuEvent->FramebufferLock();
 	
 #ifdef ENABLE_SHARED_FETCH_OBJECT
 	pthread_rwlock_wrlock([[self sharedData] rwlockFramebufferAtIndex:0]);
@@ -335,7 +332,7 @@ public:
 - (NSUInteger) gpuColorFormat
 {
 	gpuEvent->Render3DLock();
-	gpuEvent->FramebufferLockRead();
+	gpuEvent->FramebufferLock();
 	const NSUInteger colorFormat = (NSUInteger)GPU->GetDisplayInfo().colorFormat;
 	gpuEvent->FramebufferUnlock();
 	gpuEvent->Render3DUnlock();
@@ -343,11 +340,6 @@ public:
 	return colorFormat;
 }
 
-- (pthread_rwlock_t *) gpuFrameRWLock
-{
-	return gpuEvent->GetFrameRWLock();
-}
-
 #ifdef ENABLE_SHARED_FETCH_OBJECT
 - (void) setOutputList:(NSMutableArray *)theOutputList mutexPtr:(pthread_mutex_t *)theMutex
 {
@@ -868,7 +860,7 @@ public:
 
 - (void) clearWithColor:(const uint16_t)colorBGRA5551
 {
-	gpuEvent->FramebufferLockWrite();
+	gpuEvent->FramebufferLock();
 	
 #ifdef ENABLE_SHARED_FETCH_OBJECT
 	const u8 bufferIndex = GPU->GetDisplayInfo().bufferIndex;
@@ -1239,7 +1231,7 @@ GPUEventHandlerOSX::GPUEventHandlerOSX()
 {
 	_fetchObject = nil;
 	_render3DNeedsFinish = false;
-	pthread_rwlock_init(&_rwlockFrame, NULL);
+	pthread_mutex_init(&_mutexFrame, NULL);
 	pthread_mutex_init(&_mutex3DRender, NULL);
 	pthread_mutex_init(&_mutexApplyGPUSettings, NULL);
 	pthread_mutex_init(&_mutexApplyRender3DSettings, NULL);
@@ -1252,7 +1244,7 @@ GPUEventHandlerOSX::~GPUEventHandlerOSX()
 		pthread_mutex_unlock(&this->_mutex3DRender);
 	}
 	
-	pthread_rwlock_destroy(&this->_rwlockFrame);
+	pthread_mutex_destroy(&this->_mutexFrame);
 	pthread_mutex_destroy(&this->_mutex3DRender);
 	pthread_mutex_destroy(&this->_mutexApplyGPUSettings);
 	pthread_mutex_destroy(&this->_mutexApplyRender3DSettings);
@@ -1270,7 +1262,7 @@ void GPUEventHandlerOSX::SetFetchObject(GPUClientFetchObject *fetchObject)
 
 void GPUEventHandlerOSX::DidFrameBegin(bool isFrameSkipRequested, const u8 targetBufferIndex, const size_t line)
 {
-	this->FramebufferLockWrite();
+	this->FramebufferLock();
 	
 #ifdef ENABLE_SHARED_FETCH_OBJECT
 	if (!isFrameSkipRequested)
@@ -1334,19 +1326,14 @@ void GPUEventHandlerOSX::DidApplyRender3DSettingsEnd()
 	this->ApplyRender3DSettingsUnlock();
 }
 
-void GPUEventHandlerOSX::FramebufferLockWrite()
+void GPUEventHandlerOSX::FramebufferLock()
 {
-	pthread_rwlock_wrlock(&this->_rwlockFrame);
-}
-
-void GPUEventHandlerOSX::FramebufferLockRead()
-{
-	pthread_rwlock_rdlock(&this->_rwlockFrame);
+	pthread_mutex_lock(&this->_mutexFrame);
 }
 
 void GPUEventHandlerOSX::FramebufferUnlock()
 {
-	pthread_rwlock_unlock(&this->_rwlockFrame);
+	pthread_mutex_unlock(&this->_mutexFrame);
 }
 
 void GPUEventHandlerOSX::Render3DLock()
@@ -1384,11 +1371,6 @@ bool GPUEventHandlerOSX::GetRender3DNeedsFinish()
 	return this->_render3DNeedsFinish;
 }
 
-pthread_rwlock_t* GPUEventHandlerOSX::GetFrameRWLock()
-{
-	return &this->_rwlockFrame;
-}
-
 #pragma mark -
 
 CGLContextObj OSXOpenGLRendererContext = NULL;

desmume/src/frontend/cocoa/cocoa_core.mm

@@ -901,7 +901,7 @@ volatile bool execute = true;
 	
 	if ([theOutput isKindOfClass:[CocoaDSDisplay class]])
 	{
-		[theOutput setRwlockProducer:[[self cdsGPU] gpuFrameRWLock]];
+		[theOutput setRwlockProducer:NULL];
 	}
 	else
 	{

desmume/src/frontend/cocoa/cocoa_output.mm

@@ -57,6 +57,8 @@
 	mutexConsume = (pthread_mutex_t *)malloc(sizeof(pthread_mutex_t));
 	pthread_mutex_init(mutexConsume, NULL);
 	
+	rwlockProducer = NULL;
+	
 	return self;
 }
 

desmume/src/frontend/cocoa/userinterface/MacMetalDisplayView.mm

@@ -1299,16 +1299,16 @@
 		id<MTLCommandBuffer> cb = [self newCommandBuffer];
 		id<MTLComputeCommandEncoder> cce = [cb computeCommandEncoder];
 		
-		// Run the video source filters and the pixel scalers
-		if (useDeposterize)
-		{
-			[cce setComputePipelineState:[sharedData deposterizePipeline]];
-			
 		if (needsFetchBuffersLock)
 		{
 			pthread_rwlock_rdlock([sharedData rwlockFramebufferAtIndex:bufferIndex]);
 		}
 		
+		// Run the video source filters and the pixel scalers
+		if (useDeposterize)
+		{
+			[cce setComputePipelineState:[sharedData deposterizePipeline]];
+			
 			if (shouldProcessDisplay[NDSDisplayID_Main])
 			{
 				[cce setTexture:texMain atIndex:0];
@@ -1348,16 +1348,19 @@
 			
 			if (needsFetchBuffersLock)
 			{
+				needsFetchBuffersLock = !isDisplayProcessedMain || !isDisplayProcessedTouch;
+				
 				[cce endEncoding];
 				[cb addCompletedHandler:^(id<MTLCommandBuffer> block) {
+					if (!needsFetchBuffersLock)
+					{
 						pthread_rwlock_unlock([sharedData rwlockFramebufferAtIndex:bufferIndex]);
+					}
 				}];
 				[cb commit];
 				
 				cb = [self newCommandBuffer];
 				cce = [cb computeCommandEncoder];
-				
-				needsFetchBuffersLock = !isDisplayProcessedMain || !isDisplayProcessedTouch;
 			}
 		}
 		
@@ -1366,11 +1369,6 @@
 		{
 			[cce setComputePipelineState:[self pixelScalePipeline]];
 			
-			if (needsFetchBuffersLock)
-			{
-				pthread_rwlock_rdlock([sharedData rwlockFramebufferAtIndex:bufferIndex]);
-			}
-			
 			if (shouldProcessDisplay[NDSDisplayID_Main])
 			{
 				[cce setTexture:texMain atIndex:0];
@@ -1400,6 +1398,8 @@
 			
 			if (needsFetchBuffersLock)
 			{
+				needsFetchBuffersLock = false;
+				
 				[cce endEncoding];
 				[cb addCompletedHandler:^(id<MTLCommandBuffer> block) {
 					pthread_rwlock_unlock([sharedData rwlockFramebufferAtIndex:bufferIndex]);