Cocoa Port: Copying the contents of the display window via Edit > Copy now copies the actual contents of the display window instead of directly copying the NDS screens with a generic vertical layout.

desmume/src/frontend/cocoa/ClientDisplayView.cpp

@@ -1017,7 +1017,12 @@ void ClientDisplayView::FlushView()
 	this->_viewNeedsFlush = false;
 }
 
-void ClientDisplayView::FinishFrameAtIndex(const u8 bufferIndex)
+void ClientDisplayView::CopyFrameToBuffer(uint32_t *dstBuffer)
+{
+	// Do nothing. This is implementation dependent.
+}
+
+void ClientDisplayView::FinishFrameAtIndex(const uint8_t bufferIndex)
 {
 	// Do nothing. This is implementation dependent.
 }

desmume/src/frontend/cocoa/ClientDisplayView.h

@@ -320,7 +320,9 @@ public:
 	virtual void ProcessDisplays();
 	virtual void UpdateView();
 	virtual void FlushView();
-	virtual void FinishFrameAtIndex(const u8 bufferIndex);
+	virtual void FinishFrameAtIndex(const uint8_t bufferIndex);
+	
+	virtual void CopyFrameToBuffer(uint32_t *dstBuffer);
 	
 	// Emulator interface
 	const NDSDisplayInfo& GetEmuDisplayInfo() const;

desmume/src/frontend/cocoa/OGLDisplayOutput.cpp

@@ -34,6 +34,7 @@ static const char *HUDOutputVertShader_100 = {"\
 	uniform vec2 viewSize; \n\
 	uniform float scalar; \n\
 	uniform float angleDegrees; \n\
+	uniform bool renderFlipped; \n\
 	\n\
 	VARYING vec4 vtxColor; \n\
 	VARYING vec2 texCoord[1]; \n\
@@ -54,6 +55,11 @@ static const char *HUDOutputVertShader_100 = {"\
 		vtxColor = inColor; \n\
 		texCoord[0] = inTexCoord0; \n\
 		gl_Position = vec4(projection * rotation * scale * inPosition, 0.0, 1.0);\n\
+		\n\
+		if (renderFlipped)\n\
+		{\n\
+			gl_Position.y *= -1.0;\n\
+		}\n\
 	} \n\
 "};
 
@@ -77,6 +83,7 @@ static const char *Sample1x1OutputVertShader_100 = {"\
 	uniform vec2 viewSize; \n\
 	uniform float scalar; \n\
 	uniform float angleDegrees; \n\
+	uniform bool renderFlipped; \n\
 	\n\
 	VARYING vec2 texCoord[1]; \n\
 	\n\
@@ -95,6 +102,11 @@ static const char *Sample1x1OutputVertShader_100 = {"\
 		\n\
 		texCoord[0] = inTexCoord0; \n\
 		gl_Position = vec4(projection * rotation * scale * inPosition, 0.0, 1.0);\n\
+		\n\
+		if (renderFlipped)\n\
+		{\n\
+			gl_Position.y *= -1.0;\n\
+		}\n\
 	} \n\
 "};
 
@@ -111,6 +123,7 @@ static const char *BicubicSample4x4Output_VertShader_110 = {"\
 	uniform vec2 viewSize; \n\
 	uniform float scalar; \n\
 	uniform float angleDegrees; \n\
+	uniform bool renderFlipped; \n\
 	\n\
 	VARYING vec2 texCoord[16];\n\
 	\n\
@@ -150,6 +163,11 @@ static const char *BicubicSample4x4Output_VertShader_110 = {"\
 		texCoord[12] = xystart + vec2( 2.0, 2.0);\n\
 		\n\
 		gl_Position = vec4(projection * rotation * scale * inPosition, 0.0, 1.0);\n\
+		\n\
+		if (renderFlipped)\n\
+		{\n\
+			gl_Position.y *= -1.0;\n\
+		}\n\
 	}\n\
 "};
 
@@ -167,6 +185,7 @@ static const char *BicubicSample5x5Output_VertShader_110 = {"\
 	uniform vec2 viewSize; \n\
 	uniform float scalar; \n\
 	uniform float angleDegrees; \n\
+	uniform bool renderFlipped; \n\
 	\n\
 	VARYING vec2 texCoord[25];\n\
 	\n\
@@ -216,6 +235,11 @@ static const char *BicubicSample5x5Output_VertShader_110 = {"\
 		texCoord[12] = xystart + vec2( 2.0, 2.0);\n\
 		\n\
 		gl_Position = vec4(projection * rotation * scale * inPosition, 0.0, 1.0);\n\
+		\n\
+		if (renderFlipped)\n\
+		{\n\
+			gl_Position.y *= -1.0;\n\
+		}\n\
 	}\n\
 "};
 
@@ -234,6 +258,7 @@ static const char *BicubicSample6x6Output_VertShader_110 = {"\
 	uniform vec2 viewSize; \n\
 	uniform float scalar; \n\
 	uniform float angleDegrees; \n\
+	uniform bool renderFlipped; \n\
 	\n\
 	VARYING vec2 texCoord[36];\n\
 	\n\
@@ -295,6 +320,11 @@ static const char *BicubicSample6x6Output_VertShader_110 = {"\
 		texCoord[30] = xystart + vec2( 3.0, 3.0);\n\
 		\n\
 		gl_Position = vec4(projection * rotation * scale * inPosition, 0.0, 1.0);\n\
+		\n\
+		if (renderFlipped)\n\
+		{\n\
+			gl_Position.y *= -1.0;\n\
+		}\n\
 	}\n\
 "};
 
@@ -4996,6 +5026,8 @@ OGLVideoOutput::OGLVideoOutput()
 	
 	_texCPUFilterDstID[NDSDisplayID_Main] = 0;
 	_texCPUFilterDstID[NDSDisplayID_Touch] = 0;
+	
+	_fboFrameCopyID = 0;
 }
 
 OGLVideoOutput::~OGLVideoOutput()
@@ -5011,6 +5043,7 @@ OGLVideoOutput::~OGLVideoOutput()
 		this->_layerList = NULL;
 	}
 	
+	glDeleteFramebuffersEXT(1, &this->_fboFrameCopyID);
 	glDeleteTextures(2, this->_texCPUFilterDstID);
 }
 
@@ -5118,7 +5151,7 @@ void OGLVideoOutput::Init()
 	glDisable(GL_DITHER);
 	glDisable(GL_STENCIL_TEST);
 	
-	glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
+	glBlendFuncSeparate(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_ZERO, GL_ONE);
 	
 	// Set up fixed-function pipeline render states.
 	if (!this->_contextInfo->IsShaderSupported())
@@ -5133,10 +5166,9 @@ void OGLVideoOutput::Init()
 	glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 	
 	// Set up textures
-	glGenTextures(2, this->_texCPUFilterDstID);
-	
 	glTextureRangeAPPLE(GL_TEXTURE_RECTANGLE_ARB, this->_vfMasterDstBufferSize, this->_vfMasterDstBuffer);
 	
+	glGenTextures(2, this->_texCPUFilterDstID);
 	glBindTexture(GL_TEXTURE_RECTANGLE_ARB, this->_texCPUFilterDstID[NDSDisplayID_Main]);
 	glTexParameteri(GL_TEXTURE_RECTANGLE_ARB, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 	glTexParameteri(GL_TEXTURE_RECTANGLE_ARB, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
@@ -5154,6 +5186,8 @@ void OGLVideoOutput::Init()
 	glTexImage2D(GL_TEXTURE_RECTANGLE_ARB, 0, GL_RGBA, this->_vf[NDSDisplayID_Touch]->GetDstWidth(), this->_vf[NDSDisplayID_Touch]->GetDstHeight(), 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, this->_vf[NDSDisplayID_Touch]->GetDstBufferPtr());
 	
 	glBindTexture(GL_TEXTURE_RECTANGLE_ARB, 0);
+	
+	glGenFramebuffersEXT(1, &this->_fboFrameCopyID);
 }
 
 void OGLVideoOutput::SetOutputFilter(const OutputFilterTypeID filterID)
@@ -5222,7 +5256,7 @@ void OGLVideoOutput::ProcessDisplays()
 	}
 }
 
-void OGLVideoOutput::FinishFrameAtIndex(const u8 bufferIndex)
+void OGLVideoOutput::FinishFrameAtIndex(const uint8_t bufferIndex)
 {
 	for (size_t i = 0; i < _layerList->size(); i++)
 	{
@@ -5235,7 +5269,42 @@ void OGLVideoOutput::FinishFrameAtIndex(const u8 bufferIndex)
 	}
 }
 
-void OGLVideoOutput::RenderViewOGL()
+void OGLVideoOutput::CopyFrameToBuffer(uint32_t *dstBuffer)
+{
+	for (size_t i = 0; i < _layerList->size(); i++)
+	{
+		OGLVideoLayer *theLayer = (*_layerList)[i];
+		theLayer->SetNeedsUpdateViewport();
+	}
+	
+	GLuint texFrameCopyID = 0;
+	
+	glGenTextures(1, &texFrameCopyID);
+	glBindTexture(GL_TEXTURE_RECTANGLE_ARB, texFrameCopyID);
+	glTexParameteri(GL_TEXTURE_RECTANGLE_ARB, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+	glTexParameteri(GL_TEXTURE_RECTANGLE_ARB, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+	glTexParameteri(GL_TEXTURE_RECTANGLE_ARB, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+	glTexParameteri(GL_TEXTURE_RECTANGLE_ARB, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+	glTexImage2D(GL_TEXTURE_RECTANGLE_ARB, 0, GL_RGBA, this->_viewportWidth, this->_viewportHeight, 0, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV, dstBuffer);
+	
+	glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, this->_fboFrameCopyID);
+	glFramebufferTexture2DEXT(GL_FRAMEBUFFER_EXT, GL_COLOR_ATTACHMENT0_EXT, GL_TEXTURE_RECTANGLE_ARB, texFrameCopyID, 0);
+	
+	this->RenderFrameOGL(true);
+	glReadPixels(0, 0, this->_renderProperty.clientWidth, this->_renderProperty.clientHeight, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV, dstBuffer);
+	
+	glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0);
+	glBindTexture(GL_TEXTURE_RECTANGLE_ARB, 0);
+	glDeleteTextures(1, &texFrameCopyID);
+	
+	for (size_t i = 0; i < _layerList->size(); i++)
+	{
+		OGLVideoLayer *theLayer = (*_layerList)[i];
+		theLayer->SetNeedsUpdateViewport();
+	}
+}
+
+void OGLVideoOutput::RenderFrameOGL(bool isRenderingFlipped)
 {
 	if (this->_needUpdateViewport)
 	{
@@ -5251,7 +5320,7 @@ void OGLVideoOutput::RenderViewOGL()
 		
 		if (theLayer->IsVisible())
 		{
-			theLayer->RenderOGL();
+			theLayer->RenderOGL(isRenderingFlipped);
 		}
 	}
 }
@@ -5632,6 +5701,7 @@ OGLImage::OGLImage(OGLContextInfo *contextInfo, GLsizei imageWidth, GLsizei imag
 		_uniformAngleDegrees = glGetUniformLocation(finalOutputProgramID, "angleDegrees");
 		_uniformScalar = glGetUniformLocation(finalOutputProgramID, "scalar");
 		_uniformViewSize = glGetUniformLocation(finalOutputProgramID, "viewSize");
+		_uniformRenderFlipped = glGetUniformLocation(finalOutputProgramID, "renderFlipped");
 		glUseProgram(0);
 	}
 	else
@@ -5771,6 +5841,7 @@ void OGLImage::UploadTransformationOGL()
 		glUniform2f(this->_uniformViewSize, w, h);
 		glUniform1f(this->_uniformAngleDegrees, 0.0f);
 		glUniform1f(this->_uniformScalar, s);
+		glUniform1i(this->_uniformRenderFlipped, GL_FALSE);
 	}
 	else
 	{
@@ -6228,6 +6299,7 @@ OGLHUDLayer::OGLHUDLayer(OGLVideoOutput *oglVO)
 		_uniformAngleDegrees = glGetUniformLocation(_program->GetProgramID(), "angleDegrees");
 		_uniformScalar = glGetUniformLocation(_program->GetProgramID(), "scalar");
 		_uniformViewSize = glGetUniformLocation(_program->GetProgramID(), "viewSize");
+		_uniformRenderFlipped = glGetUniformLocation(_program->GetProgramID(), "renderFlipped");
 		glUseProgram(0);
 	}
 	else
@@ -6438,7 +6510,7 @@ void OGLHUDLayer::_UpdateVerticesOGL()
 	this->_output->ClearHUDNeedsUpdate();
 }
 
-void OGLHUDLayer::RenderOGL()
+void OGLHUDLayer::RenderOGL(bool isRenderingFlipped)
 {
 	size_t hudLength = this->_output->GetHUDString().length();
 	size_t hudTouchLineLength = 0;
@@ -6489,6 +6561,7 @@ void OGLHUDLayer::RenderOGL()
 		if (this->_needUpdateViewport)
 		{
 			glUniform2f(this->_uniformViewSize, this->_output->GetViewProperties().clientWidth, this->_output->GetViewProperties().clientHeight);
+			glUniform1i(this->_uniformRenderFlipped, (isRenderingFlipped) ? GL_TRUE : GL_FALSE);
 			this->_needUpdateViewport = false;
 		}
 	}
@@ -6645,6 +6718,7 @@ OGLDisplayLayer::OGLDisplayLayer(OGLVideoOutput *oglVO)
 		_uniformAngleDegrees = glGetUniformLocation(finalOutputProgramID, "angleDegrees");
 		_uniformScalar = glGetUniformLocation(finalOutputProgramID, "scalar");
 		_uniformViewSize = glGetUniformLocation(finalOutputProgramID, "viewSize");
+		_uniformRenderFlipped = glGetUniformLocation(finalOutputProgramID, "renderFlipped");
 		glUseProgram(0);
 	}
 	else
@@ -7134,7 +7208,7 @@ void OGLDisplayLayer::ProcessOGL()
 	glViewport(0, 0, this->_output->GetViewportWidth(), this->_output->GetViewportHeight());
 }
 
-void OGLDisplayLayer::RenderOGL()
+void OGLDisplayLayer::RenderOGL(bool isRenderingFlipped)
 {
 	const NDSDisplayInfo &emuDisplayInfo = this->_output->GetEmuDisplayInfo();
 	
@@ -7145,6 +7219,7 @@ void OGLDisplayLayer::RenderOGL()
 		if (this->_needUpdateViewport)
 		{
 			glUniform2f(this->_uniformViewSize, this->_output->GetViewportWidth(), this->_output->GetViewportHeight());
+			glUniform1i(this->_uniformRenderFlipped, (isRenderingFlipped) ? GL_TRUE : GL_FALSE);
 			this->_needUpdateViewport = false;
 		}
 	}
@@ -7156,7 +7231,15 @@ void OGLDisplayLayer::RenderOGL()
 		
 		glMatrixMode(GL_PROJECTION);
 		glLoadIdentity();
+		
+		if (isRenderingFlipped)
+		{
+			glOrtho(-w/2.0, -w/2.0 + w, -h/2.0 + h, -h/2.0, -1.0, 1.0);
+		}
+		else
+		{
 			glOrtho(-w/2.0, -w/2.0 + w, -h/2.0, -h/2.0 + h, -1.0, 1.0);
+		}
 		
 		this->_needUpdateViewport = false;
 	}

desmume/src/frontend/cocoa/OGLDisplayOutput.h

@@ -212,6 +212,7 @@ protected:
 	GLint _uniformAngleDegrees;
 	GLint _uniformScalar;
 	GLint _uniformViewSize;
+	GLint _uniformRenderFlipped;
 	
 	void UploadVerticesOGL();
 	void UploadTexCoordsOGL();
@@ -257,6 +258,7 @@ protected:
 	GLint _uniformAngleDegrees;
 	GLint _uniformScalar;
 	GLint _uniformViewSize;
+	GLint _uniformRenderFlipped;
 	
 public:
 	virtual ~OGLVideoLayer() {};
@@ -268,7 +270,7 @@ public:
 	virtual bool IsVisible();
 	virtual void SetVisibility(const bool visibleState);
 	
-	virtual void RenderOGL() = 0;
+	virtual void RenderOGL(bool isRenderingFlipped) = 0;
 	virtual void FinishOGL(const u8 bufferIndex) {};
 };
 
@@ -292,7 +294,7 @@ public:
 	
 	void CopyHUDFont(const FT_Face &fontFace, const size_t glyphSize, const size_t glyphTileSize, GlyphInfo *glyphInfo);
 	
-	virtual void RenderOGL();
+	virtual void RenderOGL(bool isRenderingFlipped);
 };
 
 class OGLDisplayLayer : public OGLVideoLayer
@@ -324,7 +326,7 @@ public:
 	
 	void LoadNativeDisplayByID_OGL(const NDSDisplayID displayID);
 	void ProcessOGL();
-	virtual void RenderOGL();
+	virtual void RenderOGL(bool isRenderingFlipped);
 	virtual void FinishOGL(const u8 bufferIndex);
 };
 
@@ -381,6 +383,7 @@ protected:
 	bool _hasOGLPixelScaler;
 	std::vector<OGLVideoLayer *> *_layerList;
 	GLuint _texCPUFilterDstID[2];
+	GLuint _fboFrameCopyID;
 	
 	void _UpdateViewport();
 	
@@ -419,8 +422,9 @@ public:
 	
 	// Client view interface
 	virtual void ProcessDisplays();
-	virtual void FinishFrameAtIndex(const u8 bufferIndex);
+	virtual void FinishFrameAtIndex(const uint8_t bufferIndex);
-	virtual void RenderViewOGL();
+	virtual void CopyFrameToBuffer(uint32_t *dstBuffer);;
+	virtual void RenderFrameOGL(bool isRenderingFlipped);
 	virtual void LockDisplayTextures();
 	virtual void UnlockDisplayTextures();
 };

desmume/src/frontend/cocoa/cocoa_output.h

@@ -170,6 +170,7 @@
 
 - (void) setScaleFactor:(float)theScaleFactor;
 - (void) hudUpdate;
+- (NSImage *) copyImageFromView;
 - (NSImage *) image;
 - (NSBitmapImageRep *) bitmapImageRep;
 

desmume/src/frontend/cocoa/cocoa_output.mm

@@ -1140,7 +1140,7 @@
 
 - (void) handleCopyToPasteboard
 {
-	NSImage *screenshot = [self image];
+	NSImage *screenshot = [self copyImageFromView];
 	if (screenshot == nil)
 	{
 		return;
@@ -1188,6 +1188,44 @@
 	OSSpinLockUnlock(&spinlockNDSFrameInfo);
 }
 
+- (NSImage *) copyImageFromView
+{
+	NSSize viewSize = NSMakeSize(_cdv->GetViewProperties().clientWidth, _cdv->GetViewProperties().clientHeight);
+	NSUInteger w = viewSize.width;
+	NSUInteger h = viewSize.height;
+	
+	NSImage *newImage = [[NSImage alloc] initWithSize:viewSize];
+	if (newImage == nil)
+	{
+		return newImage;
+	}
+	
+	// Render the frame in an NSBitmapImageRep
+	NSBitmapImageRep *newImageRep = [[[NSBitmapImageRep alloc] initWithBitmapDataPlanes:NULL
+																			 pixelsWide:w
+																			 pixelsHigh:h
+																		  bitsPerSample:8
+																		samplesPerPixel:4
+																			   hasAlpha:YES
+																			   isPlanar:NO
+																		 colorSpaceName:NSCalibratedRGBColorSpace
+																			bytesPerRow:w * 4
+																		   bitsPerPixel:32] autorelease];
+	if (newImageRep == nil)
+	{
+		[newImage release];
+		newImage = nil;
+		return newImage;
+	}
+	
+	_cdv->CopyFrameToBuffer((uint32_t *)[newImageRep bitmapData]);
+	
+	// Attach the rendered frame to the NSImageRep
+	[newImage addRepresentation:newImageRep];
+	
+	return [newImage autorelease];
+}
+
 - (NSImage *) image
 {
 	pthread_rwlock_rdlock(self.rwlockProducer);

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

@@ -55,6 +55,7 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 	id<MTLComputePipelineState> _fetch888Pipeline;
 	id<MTLComputePipelineState> deposterizePipeline;
 	id<MTLRenderPipelineState> hudPipeline;
+	id<MTLRenderPipelineState> hudRGBAPipeline;
 	
 	id<MTLBlitCommandEncoder> _fetchEncoder;
 	
@@ -109,6 +110,7 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 
 @property (readonly, nonatomic) id<MTLComputePipelineState> deposterizePipeline;
 @property (readonly, nonatomic) id<MTLRenderPipelineState> hudPipeline;
+@property (readonly, nonatomic) id<MTLRenderPipelineState> hudRGBAPipeline;
 @property (readonly, nonatomic) id<MTLSamplerState> samplerHUDBox;
 @property (readonly, nonatomic) id<MTLSamplerState> samplerHUDText;
 
@@ -162,6 +164,7 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 	MTLRenderPassColorAttachmentDescriptor *colorAttachment0Desc;
 	id<MTLComputePipelineState> pixelScalePipeline;
 	id<MTLRenderPipelineState> displayOutputPipeline;
+	id<MTLRenderPipelineState> displayRGBAOutputPipeline;
 	
 	id<MTLBuffer> _cdvPropertiesBuffer;
 	id<MTLBuffer> _displayVtxPositionBuffer;
@@ -203,6 +206,7 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 @property (readonly, nonatomic) MTLRenderPassColorAttachmentDescriptor *colorAttachment0Desc;
 @property (retain) id<MTLComputePipelineState> pixelScalePipeline;
 @property (retain) id<MTLRenderPipelineState> displayOutputPipeline;
+@property (retain) id<MTLRenderPipelineState> displayRGBAOutputPipeline;
 @property (retain) id<MTLBuffer> bufCPUFilterSrcMain;
 @property (retain) id<MTLBuffer> bufCPUFilterSrcTouch;
 @property (retain) id<MTLBuffer> bufCPUFilterDstMain;
@@ -223,6 +227,10 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 - (void) copyHUDFontUsingFace:(const FT_Face &)fontFace size:(const size_t)glyphSize tileSize:(const size_t)glyphTileSize info:(GlyphInfo *)glyphInfo;
 - (void) processDisplays;
 - (void) updateRenderBuffers;
+- (void) renderForCommandBuffer:(id<MTLCommandBuffer>)cb
+		   displayPipelineState:(id<MTLRenderPipelineState>)displayPipelineState
+			   hudPipelineState:(id<MTLRenderPipelineState>)hudPipelineState;
+- (void) renderAndDownloadToBuffer:(uint32_t *)dstBuffer;
 - (void) renderAndFlushDrawable;
 
 @end
@@ -287,6 +295,8 @@ public:
 	virtual void ProcessDisplays();
 	virtual void UpdateView();
 	virtual void FlushView();
+	
+	virtual void CopyFrameToBuffer(uint32_t *dstBuffer);
 };
 
 #pragma mark -

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

@@ -27,6 +27,7 @@
 
 @synthesize deposterizePipeline;
 @synthesize hudPipeline;
+@synthesize hudRGBAPipeline;
 @synthesize samplerHUDBox;
 @synthesize samplerHUDText;
 
@@ -109,20 +110,24 @@
 	deposterizeThreadGroupsPerGrid = fetchThreadGroupsPerGridNative;
 		
 	MTLRenderPipelineDescriptor *hudPipelineDesc = [[MTLRenderPipelineDescriptor alloc] init];
-	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatBGRA8Unorm];
 	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setBlendingEnabled:YES];
 	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setRgbBlendOperation:MTLBlendOperationAdd];
 	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setAlphaBlendOperation:MTLBlendOperationAdd];
 	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setSourceRGBBlendFactor:MTLBlendFactorSourceAlpha];
-	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setSourceAlphaBlendFactor:MTLBlendFactorSourceAlpha];
+	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setSourceAlphaBlendFactor:MTLBlendFactorZero];
 	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setDestinationRGBBlendFactor:MTLBlendFactorOneMinusSourceAlpha];
-	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setDestinationAlphaBlendFactor:MTLBlendFactorOneMinusSourceAlpha];
+	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setDestinationAlphaBlendFactor:MTLBlendFactorOne];
 	
 	id<MTLFunction> hudFragmentFunction = [defaultLibrary newFunctionWithName:@"hud_fragment"];
 	[hudPipelineDesc setVertexFunction:[defaultLibrary newFunctionWithName:@"hud_vertex"]];
 	[hudPipelineDesc setFragmentFunction:hudFragmentFunction];
+	
+	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatBGRA8Unorm];
 	hudPipeline = [[device newRenderPipelineStateWithDescriptor:hudPipelineDesc error:nil] retain];
 	
+	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatRGBA8Unorm];
+	hudRGBAPipeline = [[device newRenderPipelineStateWithDescriptor:hudPipelineDesc error:nil] retain];
+	
 	[hudPipelineDesc release];
 	
 	hudIndexBuffer = [[device newBufferWithLength:(sizeof(uint16_t) * HUD_TOTAL_ELEMENTS * 6) options:MTLResourceStorageModeManaged] retain];
@@ -274,6 +279,7 @@
 	[_fetch888Pipeline release];
 	[deposterizePipeline release];
 	[hudPipeline release];
+	[hudRGBAPipeline release];
 	[hudIndexBuffer release];
 	
 	[_bufMasterBrightMode[NDSDisplayID_Main] release];
@@ -728,6 +734,7 @@
 @synthesize colorAttachment0Desc;
 @synthesize pixelScalePipeline;
 @synthesize displayOutputPipeline;
+@synthesize displayRGBAOutputPipeline;
 @synthesize bufCPUFilterSrcMain;
 @synthesize bufCPUFilterSrcTouch;
 @synthesize bufCPUFilterDstMain;
@@ -764,6 +771,7 @@
 	
 	pixelScalePipeline = nil;
 	displayOutputPipeline = nil;
+	displayRGBAOutputPipeline = nil;
 	
 	_cdvPropertiesBuffer = nil;
 	_displayVtxPositionBuffer = nil;
@@ -828,6 +836,7 @@
 	[[self colorAttachment0Desc] setTexture:nil];
 	[self setPixelScalePipeline:nil];
 	[self setDisplayOutputPipeline:nil];
+	[self setDisplayRGBAOutputPipeline:nil];
 	[self setTexHUDCharMap:nil];
 	
 	[self setSharedData:nil];
@@ -996,7 +1005,6 @@
 - (void) setOutputFilter:(OutputFilterTypeID)filterID
 {
 	MTLRenderPipelineDescriptor *outputPipelineDesc = [[MTLRenderPipelineDescriptor alloc] init];
-	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:[self pixelFormat]];
 	[outputPipelineDesc setAlphaToOneEnabled:YES];
 	
 	switch (filterID)
@@ -1033,7 +1041,12 @@
 			break;
 	}
 	
+	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:[self pixelFormat]];
 	[self setDisplayOutputPipeline:[[self device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil]];
+	
+	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatRGBA8Unorm];
+	[self setDisplayRGBAOutputPipeline:[[self device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil]];
+	
 	[outputPipelineDesc release];
 }
 
@@ -1047,12 +1060,16 @@
 	[self setDevice:[sharedData device]];
 	
 	MTLRenderPipelineDescriptor *outputPipelineDesc = [[MTLRenderPipelineDescriptor alloc] init];
-	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:[self pixelFormat]];
 	[outputPipelineDesc setAlphaToOneEnabled:YES];
 	[outputPipelineDesc setVertexFunction:[[sharedData defaultLibrary] newFunctionWithName:@"display_output_vertex"]];
 	[outputPipelineDesc setFragmentFunction:[[sharedData defaultLibrary] newFunctionWithName:@"output_filter_bilinear"]];
 	
+	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:[self pixelFormat]];
 	displayOutputPipeline = [[[self device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil] retain];
+	
+	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatRGBA8Unorm];
+	displayRGBAOutputPipeline = [[[self device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil] retain];
+	
 	[outputPipelineDesc release];
 	
 	_cdvPropertiesBuffer = [[[self device] newBufferWithLength:sizeof(DisplayViewShaderProperties) options:MTLResourceStorageModeManaged] retain];
@@ -1581,31 +1598,24 @@
 	pthread_mutex_unlock(&_mutexBufferUpdate);
 }
 
-- (void) renderAndFlushDrawable
+- (void) renderForCommandBuffer:(id<MTLCommandBuffer>)cb
+		   displayPipelineState:(id<MTLRenderPipelineState>)displayPipelineState
+			   hudPipelineState:(id<MTLRenderPipelineState>)hudPipelineState
 {
-	@autoreleasepool
+	id<MTLRenderCommandEncoder> rce = [cb renderCommandEncoderWithDescriptor:_outputRenderPassDesc];
-	{
-		pthread_mutex_lock(&_mutexDisplayTextureUpdate);
-		pthread_mutex_lock(&_mutexBufferUpdate);
-				
-		// Now that everything is set up, go ahead and draw everything.
-		id<CAMetalDrawable> layerDrawable = [self nextDrawable];
-		[colorAttachment0Desc setTexture:[layerDrawable texture]];
-		id<MTLCommandBuffer> cb = [self newCommandBuffer];
-		id<MTLRenderCommandEncoder> ce = [cb renderCommandEncoderWithDescriptor:_outputRenderPassDesc];
 	
 	if (_needEncodeViewport)
 	{
-			[ce setViewport:_newViewport];
+		[rce setViewport:_newViewport];
 	}
 	
 	// Draw the NDS displays.
 	if (_willDrawDisplays)
 	{
-			[ce setRenderPipelineState:[self displayOutputPipeline]];
+		[rce setRenderPipelineState:displayPipelineState];
-			[ce setVertexBuffer:_displayVtxPositionBuffer offset:0 atIndex:0];
+		[rce setVertexBuffer:_displayVtxPositionBuffer offset:0 atIndex:0];
-			[ce setVertexBuffer:_displayTexCoordBuffer offset:0 atIndex:1];
+		[rce setVertexBuffer:_displayTexCoordBuffer offset:0 atIndex:1];
-			[ce setVertexBuffer:_cdvPropertiesBuffer offset:0 atIndex:2];
+		[rce setVertexBuffer:_cdvPropertiesBuffer offset:0 atIndex:2];
 		
 		switch (_cdv->GetViewProperties().mode)
 		{
@@ -1613,8 +1623,8 @@
 			{
 				if (_cdv->IsSelectedDisplayEnabled(NDSDisplayID_Main))
 				{
-						[ce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Main] atIndex:0];
+					[rce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Main] atIndex:0];
-						[ce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:0 vertexCount:4];
+					[rce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:0 vertexCount:4];
 				}
 				break;
 			}
@@ -1623,8 +1633,8 @@
 			{
 				if (_cdv->IsSelectedDisplayEnabled(NDSDisplayID_Touch))
 				{
-						[ce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Touch] atIndex:0];
+					[rce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Touch] atIndex:0];
-						[ce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:4 vertexCount:4];
+					[rce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:4 vertexCount:4];
 				}
 				break;
 			}
@@ -1642,8 +1652,8 @@
 					{
 						if (_cdv->IsSelectedDisplayEnabled(majorDisplayID))
 						{
-								[ce setFragmentTexture:_texDisplayOutput[majorDisplayID] atIndex:0];
+							[rce setFragmentTexture:_texDisplayOutput[majorDisplayID] atIndex:0];
-								[ce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:majorDisplayVtx vertexCount:4];
+							[rce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:majorDisplayVtx vertexCount:4];
 						}
 						break;
 					}
@@ -1654,14 +1664,14 @@
 				
 				if (_cdv->IsSelectedDisplayEnabled(NDSDisplayID_Main))
 				{
-						[ce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Main] atIndex:0];
+					[rce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Main] atIndex:0];
-						[ce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:0 vertexCount:4];
+					[rce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:0 vertexCount:4];
 				}
 				
 				if (_cdv->IsSelectedDisplayEnabled(NDSDisplayID_Touch))
 				{
-						[ce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Touch] atIndex:0];
+					[rce setFragmentTexture:_texDisplayOutput[NDSDisplayID_Touch] atIndex:0];
-						[ce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:4 vertexCount:4];
+					[rce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:4 vertexCount:4];
 				}
 			}
 				
@@ -1675,29 +1685,29 @@
 	{
 		uint8_t isScreenOverlay = 0;
 		
-			[ce setRenderPipelineState:[sharedData hudPipeline]];
+		[rce setRenderPipelineState:hudPipelineState];
-			[ce setVertexBuffer:_hudVtxPositionBuffer offset:0 atIndex:0];
+		[rce setVertexBuffer:_hudVtxPositionBuffer offset:0 atIndex:0];
-			[ce setVertexBuffer:_hudVtxColorBuffer offset:0 atIndex:1];
+		[rce setVertexBuffer:_hudVtxColorBuffer offset:0 atIndex:1];
-			[ce setVertexBuffer:_hudTexCoordBuffer offset:0 atIndex:2];
+		[rce setVertexBuffer:_hudTexCoordBuffer offset:0 atIndex:2];
-			[ce setVertexBuffer:_cdvPropertiesBuffer offset:0 atIndex:3];
+		[rce setVertexBuffer:_cdvPropertiesBuffer offset:0 atIndex:3];
-			[ce setFragmentTexture:[self texHUDCharMap] atIndex:0];
+		[rce setFragmentTexture:[self texHUDCharMap] atIndex:0];
 		
 		// First, draw the inputs.
 		if (_willDrawHUDInput)
 		{
 			isScreenOverlay = 1;
-				[ce setVertexBytes:&isScreenOverlay length:sizeof(uint8_t) atIndex:4];
+			[rce setVertexBytes:&isScreenOverlay length:sizeof(uint8_t) atIndex:4];
-				[ce setFragmentSamplerState:[sharedData samplerHUDBox] atIndex:0];
+			[rce setFragmentSamplerState:[sharedData samplerHUDBox] atIndex:0];
-				[ce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
+			[rce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
 							indexCount:_hudTouchLineLength * 6
 							 indexType:MTLIndexTypeUInt16
 						   indexBuffer:[sharedData hudIndexBuffer]
 					 indexBufferOffset:(_hudStringLength + HUD_INPUT_ELEMENT_LENGTH) * 6 * sizeof(uint16_t)];
 			
 			isScreenOverlay = 0;
-				[ce setVertexBytes:&isScreenOverlay length:sizeof(uint8_t) atIndex:4];
+			[rce setVertexBytes:&isScreenOverlay length:sizeof(uint8_t) atIndex:4];
-				[ce setFragmentSamplerState:[sharedData samplerHUDText] atIndex:0];
+			[rce setFragmentSamplerState:[sharedData samplerHUDText] atIndex:0];
-				[ce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
+			[rce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
 							indexCount:HUD_INPUT_ELEMENT_LENGTH * 6
 							 indexType:MTLIndexTypeUInt16
 						   indexBuffer:[sharedData hudIndexBuffer]
@@ -1705,24 +1715,96 @@
 		}
 		
 		// Next, draw the backing text box.
-			[ce setVertexBytes:&isScreenOverlay length:sizeof(uint8_t) atIndex:4];
+		[rce setVertexBytes:&isScreenOverlay length:sizeof(uint8_t) atIndex:4];
-			[ce setFragmentSamplerState:[sharedData samplerHUDBox] atIndex:0];
+		[rce setFragmentSamplerState:[sharedData samplerHUDBox] atIndex:0];
-			[ce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
+		[rce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
 						indexCount:6
 						 indexType:MTLIndexTypeUInt16
 					   indexBuffer:[sharedData hudIndexBuffer]
 				 indexBufferOffset:0];
 		
 		// Finally, draw each character inside the box.
-			[ce setFragmentSamplerState:[sharedData samplerHUDText] atIndex:0];
+		[rce setFragmentSamplerState:[sharedData samplerHUDText] atIndex:0];
-			[ce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
+		[rce drawIndexedPrimitives:MTLPrimitiveTypeTriangle
 						indexCount:(_hudStringLength - 1) * 6
 						 indexType:MTLIndexTypeUInt16
 					   indexBuffer:[sharedData hudIndexBuffer]
 				 indexBufferOffset:6 * sizeof(uint16_t)];
 	}
 	
-		[ce endEncoding];
+	[rce endEncoding];
+}
+
+- (void) renderAndDownloadToBuffer:(uint32_t *)dstBuffer
+{
+	const size_t clientWidth  = _cdv->GetViewProperties().clientWidth;
+	const size_t clientHeight = _cdv->GetViewProperties().clientHeight;
+	
+	@autoreleasepool
+	{
+		MTLTextureDescriptor *texRenderDesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
+																								 width:clientWidth
+																								height:clientHeight
+																							 mipmapped:NO];
+		[texRenderDesc setResourceOptions:MTLResourceStorageModePrivate];
+		[texRenderDesc setStorageMode:MTLStorageModePrivate];
+		[texRenderDesc setUsage:MTLTextureUsageRenderTarget];
+		
+		id<MTLTexture> texRender = [[[self device] newTextureWithDescriptor:texRenderDesc] retain];
+		id<MTLBuffer> dstMTLBuffer = [[[self device] newBufferWithLength:clientWidth * clientHeight * sizeof(uint32_t) options:MTLResourceStorageModeManaged] retain];
+		
+		pthread_mutex_lock(&_mutexDisplayTextureUpdate);
+		pthread_mutex_lock(&_mutexBufferUpdate);
+		
+		// Now that everything is set up, go ahead and draw everything.
+		[colorAttachment0Desc setTexture:texRender];
+		id<MTLCommandBuffer> cb = [self newCommandBuffer];
+		
+		[self renderForCommandBuffer:cb displayPipelineState:[self displayRGBAOutputPipeline] hudPipelineState:[sharedData hudRGBAPipeline]];
+		
+		id<MTLBlitCommandEncoder> bce = [cb blitCommandEncoder];
+		
+		[bce copyFromTexture:texRender
+				 sourceSlice:0
+				 sourceLevel:0
+				sourceOrigin:MTLOriginMake(0, 0, 0)
+				  sourceSize:MTLSizeMake(clientWidth, clientHeight, 1)
+					toBuffer:dstMTLBuffer
+		   destinationOffset:0
+	  destinationBytesPerRow:clientWidth * sizeof(uint32_t)
+	destinationBytesPerImage:clientWidth * clientHeight * sizeof(uint32_t)];
+		
+		[bce synchronizeResource:dstMTLBuffer];
+		[bce endEncoding];
+		
+		[cb addCompletedHandler:^(id<MTLCommandBuffer> block) {
+			pthread_mutex_unlock(&_mutexBufferUpdate);
+			pthread_mutex_unlock(&_mutexDisplayTextureUpdate);
+			
+			memcpy(dstBuffer, [dstMTLBuffer contents], clientWidth * clientHeight * sizeof(uint32_t));
+			
+			[texRender release];
+			[dstMTLBuffer release];
+		}];
+		
+		[cb commit];
+		[cb waitUntilCompleted];
+	}
+}
+
+- (void) renderAndFlushDrawable
+{
+	@autoreleasepool
+	{
+		pthread_mutex_lock(&_mutexDisplayTextureUpdate);
+		pthread_mutex_lock(&_mutexBufferUpdate);
+				
+		// Now that everything is set up, go ahead and draw everything.
+		id<CAMetalDrawable> layerDrawable = [self nextDrawable];
+		[colorAttachment0Desc setTexture:[layerDrawable texture]];
+		id<MTLCommandBuffer> cb = [self newCommandBuffer];
+		
+		[self renderForCommandBuffer:cb displayPipelineState:[self displayOutputPipeline] hudPipelineState:[sharedData hudPipeline]];
 		
 		[cb presentDrawable:layerDrawable];
 		[cb addCompletedHandler:^(id<MTLCommandBuffer> block) {
@@ -1999,6 +2081,11 @@ void MacMetalDisplayView::FlushView()
 	[(DisplayViewMetalLayer *)this->GetFrontendLayer() renderAndFlushDrawable];
 }
 
+void MacMetalDisplayView::CopyFrameToBuffer(uint32_t *dstBuffer)
+{
+	[(DisplayViewMetalLayer *)this->GetFrontendLayer() renderAndDownloadToBuffer:dstBuffer];
+}
+
 #pragma mark -
 void SetupHQnxLUTs_Metal(id<MTLDevice> &device, id<MTLTexture> &texLQ2xLUT, id<MTLTexture> &texHQ2xLUT, id<MTLTexture> &texHQ3xLUT, id<MTLTexture> &texHQ4xLUT)
 {

desmume/src/frontend/cocoa/userinterface/MacOGLDisplayView.h

@@ -100,7 +100,8 @@ public:
 	virtual void ProcessDisplays();
 	virtual void UpdateView();
 	virtual void FlushView();
-	virtual void FinishFrameAtIndex(const u8 bufferIndex);
+	virtual void CopyFrameToBuffer(uint32_t *dstBuffer);
+	virtual void FinishFrameAtIndex(const uint8_t bufferIndex);
 	virtual void LockDisplayTextures();
 	virtual void UnlockDisplayTextures();
 };

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

@@ -75,7 +75,7 @@
 {
 	CGLSetCurrentContext(glContext);
 	CGLLockContext(glContext);
-	_cdv->RenderViewOGL();
+	_cdv->RenderFrameOGL(false);
 	[super drawInCGLContext:glContext pixelFormat:pixelFormat forLayerTime:timeInterval displayTime:timeStamp];
 	CGLUnlockContext(glContext);
 }
@@ -447,12 +447,20 @@ void MacOGLDisplayView::FlushView()
 	
 	CGLLockContext(this->_context);
 	CGLSetCurrentContext(this->_context);
-	this->RenderViewOGL();
+	this->RenderFrameOGL(false);
 	CGLFlushDrawable(this->_context);
 	CGLUnlockContext(this->_context);
 }
 
-void MacOGLDisplayView::FinishFrameAtIndex(const u8 bufferIndex)
+void MacOGLDisplayView::CopyFrameToBuffer(uint32_t *dstBuffer)
+{
+	CGLLockContext(this->_context);
+	CGLSetCurrentContext(this->_context);
+	this->OGLVideoOutput::CopyFrameToBuffer(dstBuffer);
+	CGLUnlockContext(this->_context);
+}
+
+void MacOGLDisplayView::FinishFrameAtIndex(const uint8_t bufferIndex)
 {
 	CGLLockContext(this->_context);
 	CGLSetCurrentContext(this->_context);