Cocoa Port: Fix a possible crash when performing Copy-Framebuffer-To-Clipboard operations on Metal display views.

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

@@ -102,7 +102,7 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 	id<MTLComputePipelineState> _fetch666ConvertOnlyPipeline;
 	id<MTLComputePipelineState> _fetch888ConvertOnlyPipeline;
 	id<MTLComputePipelineState> deposterizePipeline;
-	id<MTLRenderPipelineState> hudPipeline;
+	id<MTLRenderPipelineState> hudBGRAPipeline;
 	id<MTLRenderPipelineState> hudRGBAPipeline;
 	
 	id<MTLSamplerState> samplerHUDBox;
@@ -151,7 +151,7 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 @property (readonly, nonatomic) id<MTLLibrary> defaultLibrary;
 
 @property (readonly, nonatomic) id<MTLComputePipelineState> deposterizePipeline;
-@property (readonly, nonatomic) id<MTLRenderPipelineState> hudPipeline;
+@property (readonly, nonatomic) id<MTLRenderPipelineState> hudBGRAPipeline;
 @property (readonly, nonatomic) id<MTLRenderPipelineState> hudRGBAPipeline;
 @property (readonly, nonatomic) id<MTLSamplerState> samplerHUDBox;
 @property (readonly, nonatomic) id<MTLSamplerState> samplerHUDText;
@@ -188,7 +188,6 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 	MTLRenderPassDescriptor *_outputRenderPassDesc;
 	MTLRenderPassColorAttachmentDescriptor *colorAttachment0Desc;
 	id<MTLComputePipelineState> pixelScalePipeline;
-	id<MTLRenderPipelineState> outputRGBAPipeline;
 	id<MTLRenderPipelineState> outputDrawablePipeline;
 	MTLPixelFormat drawableFormat;
 	
@@ -229,7 +228,6 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 @property (assign, nonatomic) MetalDisplayViewSharedData *sharedData;
 @property (readonly, nonatomic) MTLRenderPassColorAttachmentDescriptor *colorAttachment0Desc;
 @property (retain) id<MTLComputePipelineState> pixelScalePipeline;
-@property (retain) id<MTLRenderPipelineState> outputRGBAPipeline;
 @property (retain) id<MTLRenderPipelineState> outputDrawablePipeline;
 @property (assign) MTLPixelFormat drawableFormat;
 @property (retain) id<MTLBuffer> bufCPUFilterDstMain;
@@ -257,7 +255,8 @@ typedef DisplayViewShaderProperties DisplayViewShaderProperties;
 			   hudPipelineState:(id<MTLRenderPipelineState>)hudPipelineState
 					texDisplays:(MetalTexturePair)texDisplay
 						   mrfi:(MetalRenderFrameInfo)mrfi
-						doYFlip:(BOOL)willFlip;
+						doYFlip:(BOOL)willFlip
+					   doSwapRB:(BOOL)willSwapRB;
 - (void) renderStartAtIndex:(uint8_t)index;
 - (void) renderFinishAtIndex:(uint8_t)index;
 - (ClientDisplayBufferState) renderBufferStateAtIndex:(uint8_t)index;

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

@@ -33,7 +33,7 @@
 @synthesize defaultLibrary;
 
 @synthesize deposterizePipeline;
-@synthesize hudPipeline;
+@synthesize hudBGRAPipeline;
 @synthesize hudRGBAPipeline;
 @synthesize samplerHUDBox;
 @synthesize samplerHUDText;
@@ -184,13 +184,14 @@
 		[[[hudPipelineDesc vertexBuffers] objectAtIndexedSubscript:3] setMutability:MTLMutabilityImmutable];
 		[[[hudPipelineDesc vertexBuffers] objectAtIndexedSubscript:4] setMutability:MTLMutabilityImmutable];
 		[[[hudPipelineDesc vertexBuffers] objectAtIndexedSubscript:5] setMutability:MTLMutabilityImmutable];
+		[[[hudPipelineDesc fragmentBuffers] objectAtIndexedSubscript:0] setMutability:MTLMutabilityImmutable];
 	}
 #endif
 	
 	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatBGRA8Unorm];
-	hudPipeline = [[device newRenderPipelineStateWithDescriptor:hudPipelineDesc error:nil] retain];
+	hudBGRAPipeline = [[device newRenderPipelineStateWithDescriptor:hudPipelineDesc error:nil] retain];
 	
-	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatBGRA8Unorm];
+	[[[hudPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatRGBA8Unorm];
 	hudRGBAPipeline = [[device newRenderPipelineStateWithDescriptor:hudPipelineDesc error:nil] retain];
 	
 	[hudPipelineDesc release];
@@ -318,7 +319,7 @@
 	[_fetch666ConvertOnlyPipeline release];
 	[_fetch888ConvertOnlyPipeline release];
 	[deposterizePipeline release];
-	[hudPipeline release];
+	[hudBGRAPipeline release];
 	[hudRGBAPipeline release];
 	[hudIndexBuffer release];
 	
@@ -812,7 +813,6 @@
 @synthesize sharedData;
 @synthesize colorAttachment0Desc;
 @synthesize pixelScalePipeline;
-@synthesize outputRGBAPipeline;
 @synthesize outputDrawablePipeline;
 @synthesize drawableFormat;
 @synthesize bufCPUFilterDstMain;
@@ -855,9 +855,8 @@
 	[colorAttachment0Desc setTexture:nil];
 	
 	pixelScalePipeline = nil;
-	outputRGBAPipeline = nil;
 	outputDrawablePipeline = nil;
-	drawableFormat = MTLPixelFormatInvalid;
+	drawableFormat = MTLPixelFormatRGBA8Unorm;
 	
 	_texDisplaySrcDeposterize[NDSDisplayID_Main][0]  = nil;
 	_texDisplaySrcDeposterize[NDSDisplayID_Touch][0] = nil;
@@ -951,7 +950,6 @@
 	[texPairProcess.touch release];
 	
 	[self setPixelScalePipeline:nil];
-	[self setOutputRGBAPipeline:nil];
 	[self setOutputDrawablePipeline:nil];
 	[self setTexHUDCharMap:nil];
 	
@@ -1180,14 +1178,8 @@
 			break;
 	}
 	
-	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatBGRA8Unorm];
-	[self setOutputRGBAPipeline:[[sharedData device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil]];
-	
-	if ([self drawableFormat] != MTLPixelFormatInvalid)
-	{
 	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:[self drawableFormat]];
 	[self setOutputDrawablePipeline:[[sharedData device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil]];
-	}
 	
 #if HAVE_OSAVAILABLE && defined(MAC_OS_X_VERSION_10_13) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_13)
 	if (@available(macOS 10.13, *))
@@ -1197,6 +1189,7 @@
 		[[[outputPipelineDesc vertexBuffers] objectAtIndexedSubscript:2] setMutability:MTLMutabilityImmutable];
 		[[[outputPipelineDesc vertexBuffers] objectAtIndexedSubscript:3] setMutability:MTLMutabilityImmutable];
 		[[[outputPipelineDesc fragmentBuffers] objectAtIndexedSubscript:0] setMutability:MTLMutabilityImmutable];
+		[[[outputPipelineDesc fragmentBuffers] objectAtIndexedSubscript:1] setMutability:MTLMutabilityImmutable];
 	}
 #endif
 	
@@ -1223,18 +1216,12 @@
 		[[[outputPipelineDesc vertexBuffers] objectAtIndexedSubscript:2] setMutability:MTLMutabilityImmutable];
 		[[[outputPipelineDesc vertexBuffers] objectAtIndexedSubscript:3] setMutability:MTLMutabilityImmutable];
 		[[[outputPipelineDesc fragmentBuffers] objectAtIndexedSubscript:0] setMutability:MTLMutabilityImmutable];
+		[[[outputPipelineDesc fragmentBuffers] objectAtIndexedSubscript:1] setMutability:MTLMutabilityImmutable];
 	}
 #endif
 	
-	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:MTLPixelFormatBGRA8Unorm];
-	outputRGBAPipeline = [[[sharedData device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil] retain];
-	
-	if ([self drawableFormat] != MTLPixelFormatInvalid)
-	{
 	[[[outputPipelineDesc colorAttachments] objectAtIndexedSubscript:0] setPixelFormat:[self drawableFormat]];
 	outputDrawablePipeline = [[[sharedData device] newRenderPipelineStateWithDescriptor:outputPipelineDesc error:nil] retain];
-	}
-	
 	[outputPipelineDesc release];
 	
 	// Set up processing textures.
@@ -1988,6 +1975,7 @@
 					texDisplays:(MetalTexturePair)texDisplay
 						   mrfi:(MetalRenderFrameInfo)mrfi
 						doYFlip:(BOOL)willFlip
+					   doSwapRB:(BOOL)willSwapRB
 {
 	// Generate the command encoder.
 	id<MTLRenderCommandEncoder> rce = [cb renderCommandEncoderWithDescriptor:_outputRenderPassDesc];
@@ -2011,6 +1999,7 @@
 	[rce setVertexBytes:_texCoordBuffer length:sizeof(_texCoordBuffer) atIndex:1];
 	[rce setVertexBytes:&_cdvPropertiesBuffer length:sizeof(_cdvPropertiesBuffer) atIndex:2];
 	[rce setVertexBytes:&doYFlip length:sizeof(uint8_t) atIndex:3];
+	[rce setFragmentBytes:&willSwapRB length:sizeof(uint8_t) atIndex:1];
 	
 	switch (cdp->GetPresenterProperties().mode)
 	{
@@ -2093,6 +2082,7 @@
 		[rce setVertexBytes:&_cdvPropertiesBuffer length:sizeof(_cdvPropertiesBuffer) atIndex:3];
 		[rce setVertexBytes:&doYFlip length:sizeof(uint8_t) atIndex:4];
 		[rce setFragmentTexture:[self texHUDCharMap] atIndex:0];
+		[rce setFragmentBytes:&willSwapRB length:sizeof(uint8_t) atIndex:0];
 		
 		// First, draw the inputs.
 		if (mrfi.willDrawHUDInput)
@@ -2176,7 +2166,12 @@
 	
 	@autoreleasepool
 	{
-		MTLTextureDescriptor *texRenderDesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
+		// Note that this method should ALWAYS return 32-bit RGBA format, not BGRA format.
+		MTLPixelFormat viewDrawableFormat = [self drawableFormat];
+		BOOL willSwapRB = (viewDrawableFormat == MTLPixelFormatBGRA8Unorm) ? YES : NO;
+		id<MTLRenderPipelineState> hudPipelineState = (willSwapRB) ? [sharedData hudBGRAPipeline] : [sharedData hudRGBAPipeline];
+		
+		MTLTextureDescriptor *texRenderDesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:viewDrawableFormat
 																								 width:clientWidth
 																								height:clientHeight
 																							 mipmapped:NO];
@@ -2195,11 +2190,12 @@
 		const MetalRenderFrameInfo mrfi = [self renderFrameInfo];
 		
 		[self renderForCommandBuffer:cb
-				 outputPipelineState:[self outputRGBAPipeline]
+				 outputPipelineState:[self outputDrawablePipeline]
-					hudPipelineState:[sharedData hudRGBAPipeline]
+					hudPipelineState:hudPipelineState
 						 texDisplays:texProcess
 								mrfi:mrfi
-							 doYFlip:NO];
+							 doYFlip:NO
+							doSwapRB:willSwapRB];
 		
 		id<MTLBlitCommandEncoder> bce = [cb blitCommandEncoder];
 		
@@ -2348,13 +2344,15 @@
 	[oldTexTouch release];
 	
 	const MetalRenderFrameInfo mrfi = [presenterObject renderFrameInfo];
+	id<MTLRenderPipelineState> hudPipelineState = ([presenterObject drawableFormat] == MTLPixelFormatBGRA8Unorm) ? [[presenterObject sharedData] hudBGRAPipeline] : [[presenterObject sharedData] hudRGBAPipeline];
 	
 	[presenterObject renderForCommandBuffer:cb
 						outputPipelineState:[presenterObject outputDrawablePipeline]
-						   hudPipelineState:[[presenterObject sharedData] hudPipeline]
+						   hudPipelineState:hudPipelineState
 								texDisplays:_displayTexturePair
 									   mrfi:mrfi
-									doYFlip:NO];
+									doYFlip:NO
+								   doSwapRB:NO];
 	
 	[cb addScheduledHandler:^(id<MTLCommandBuffer> block) {
 		[presenterObject setRenderBufferState:ClientDisplayBufferState_Reading index:mrfi.renderIndex];

desmume/src/frontend/cocoa/userinterface/MacMetalDisplayViewShaders.metal

@@ -195,9 +195,11 @@ vertex HUDVtx hud_vertex(const device float2 *inPosition [[buffer(0)]],
 
 fragment float4 hud_fragment(const HUDVtx vtx [[stage_in]],
 							 const texture2d<float> tex [[texture(0)]],
-							 const sampler samp [[sampler(0)]])
+							 const sampler samp [[sampler(0)]],
+							 const constant uint8_t &doSwapRB [[buffer(0)]])
 {
-	return tex.sample(samp, vtx.texCoord, (vtx.lowerHUDMipMapLevel) ? level(-0.50f) : level(0.00f)) * vtx.color;
+	const float4 outColor = tex.sample(samp, vtx.texCoord, (vtx.lowerHUDMipMapLevel) ? level(-0.50f) : level(0.00f)) * vtx.color;
+	return (doSwapRB == 0) ? outColor.rgba : outColor.bgra;
 }
 
 #pragma mark Output Filters
@@ -279,9 +281,10 @@ vertex DisplayVtxTex16 display_output_sampletex16_vertex(const device float2 *in
 // Input Pixel Mapping:  00
 fragment float4 output_filter_nearest(const DisplayVtx vtx [[stage_in]],
 									  const texture2d<float> tex [[texture(0)]],
-									  const device float *inBacklightIntensity [[buffer(0)]])
+									  const device float *inBacklightIntensity [[buffer(0)]],
+									  const constant uint8_t &doSwapRB [[buffer(1)]])
 {
-	return float4(tex.sample(genSampler, vtx.texCoord).rgb * *inBacklightIntensity, 1.0f);
+	return (doSwapRB == 0) ? float4(tex.sample(genSampler, vtx.texCoord).rgb * *inBacklightIntensity, 1.0f) : float4(tex.sample(genSampler, vtx.texCoord).bgr * *inBacklightIntensity, 1.0f);
 }
 
 //---------------------------------------
@@ -289,9 +292,10 @@ fragment float4 output_filter_nearest(const DisplayVtx vtx [[stage_in]],
 //                       02|03
 fragment float4 output_filter_bilinear(const DisplayVtx vtx [[stage_in]],
 									   const texture2d<float> tex [[texture(0)]],
-									   const device float *inBacklightIntensity [[buffer(0)]])
+									   const device float *inBacklightIntensity [[buffer(0)]],
+									   const constant uint8_t &doSwapRB [[buffer(1)]])
 {
-	return float4(tex.sample(outputSamplerBilinear, vtx.texCoord).rgb * *inBacklightIntensity, 1.0f);
+	return (doSwapRB == 0) ? float4(tex.sample(outputSamplerBilinear, vtx.texCoord).rgb * *inBacklightIntensity, 1.0f) : float4(tex.sample(outputSamplerBilinear, vtx.texCoord).bgr * *inBacklightIntensity, 1.0f);
 }
 
 //---------------------------------------
@@ -301,7 +305,8 @@ fragment float4 output_filter_bilinear(const DisplayVtx vtx [[stage_in]],
 //                       12|13|14|15
 fragment float4 output_filter_bicubic_bspline(const DisplayVtxTex16 vtx [[stage_in]],
 											  const texture2d<float> tex [[texture(0)]],
-											  const device float *inBacklightIntensity [[buffer(0)]])
+											  const device float *inBacklightIntensity [[buffer(0)]],
+											  const constant uint8_t &doSwapRB [[buffer(1)]])
 {
 	float2 f = fract(vtx.texCoord05);
 	float4 wx = bicubic_weight_bspline(f.x);
@@ -328,7 +333,7 @@ fragment float4 output_filter_bicubic_bspline(const DisplayVtxTex16 vtx [[stage_
 	                         +  tex.sample(genSampler, vtx.texCoord14) * wx.b
 	                         +  tex.sample(genSampler, vtx.texCoord15) * wx.a) * wy.a;
 	
-	return float4(outFragment.rgb * *inBacklightIntensity, 1.0f);
+	return (doSwapRB == 0) ? float4(outFragment.rgb * *inBacklightIntensity, 1.0f) : float4(outFragment.bgr * *inBacklightIntensity, 1.0f);
 }
 
 //---------------------------------------
@@ -338,7 +343,8 @@ fragment float4 output_filter_bicubic_bspline(const DisplayVtxTex16 vtx [[stage_
 //                       12|13|14|15
 fragment float4 output_filter_bicubic_mitchell_netravali(const DisplayVtxTex16 vtx [[stage_in]],
 														 const texture2d<float> tex [[texture(0)]],
-														 const device float *inBacklightIntensity [[buffer(0)]])
+														 const device float *inBacklightIntensity [[buffer(0)]],
+														 const constant uint8_t &doSwapRB [[buffer(1)]])
 {
 	float2 f = fract(vtx.texCoord05);
 	float4 wx = bicubic_weight_mitchell_netravali(f.x);
@@ -365,7 +371,7 @@ fragment float4 output_filter_bicubic_mitchell_netravali(const DisplayVtxTex16 v
 	                         +  tex.sample(genSampler, vtx.texCoord14) * wx.b
 	                         +  tex.sample(genSampler, vtx.texCoord15) * wx.a) * wy.a;
 	
-	return float4(outFragment.rgb * *inBacklightIntensity, 1.0f);
+	return (doSwapRB == 0) ? float4(outFragment.rgb * *inBacklightIntensity, 1.0f) : float4(outFragment.bgr * *inBacklightIntensity, 1.0f);
 }
 
 //---------------------------------------
@@ -375,7 +381,8 @@ fragment float4 output_filter_bicubic_mitchell_netravali(const DisplayVtxTex16 v
 //                       12|13|14|15
 fragment float4 output_filter_lanczos2(const DisplayVtxTex16 vtx [[stage_in]],
 									   const texture2d<float> tex [[texture(0)]],
-									   const device float *inBacklightIntensity [[buffer(0)]])
+									   const device float *inBacklightIntensity [[buffer(0)]],
+									   const constant uint8_t &doSwapRB [[buffer(1)]])
 {
 	const float2 f = fract(vtx.texCoord05);
 	float4 wx = bicubic_weight_lanczos2(f.x);
@@ -402,7 +409,7 @@ fragment float4 output_filter_lanczos2(const DisplayVtxTex16 vtx [[stage_in]],
 	                         +  tex.sample(genSampler, vtx.texCoord14) * wx.b
 	                         +  tex.sample(genSampler, vtx.texCoord15) * wx.a) * wy.a;
 	
-	return float4(outFragment.rgb * *inBacklightIntensity, 1.0f);
+	return (doSwapRB == 0) ? float4(outFragment.rgb * *inBacklightIntensity, 1.0f) : float4(outFragment.bgr * *inBacklightIntensity, 1.0f);
 }
 
 //---------------------------------------
@@ -414,7 +421,8 @@ fragment float4 output_filter_lanczos2(const DisplayVtxTex16 vtx [[stage_in]],
 //                      30|31|32|33|34|35
 fragment float4 output_filter_lanczos3(const DisplayVtxTex16 vtx [[stage_in]],
 									   const texture2d<float> tex [[texture(0)]],
-									   const device float *inBacklightIntensity [[buffer(0)]])
+									   const device float *inBacklightIntensity [[buffer(0)]],
+									   const constant uint8_t &doSwapRB [[buffer(1)]])
 {
 	const float2 f = fract(vtx.texCoord05);
 	float3 wx1 = bicubic_weight_lanczos3(0.5f - f.x * 0.5f);
@@ -478,7 +486,7 @@ fragment float4 output_filter_lanczos3(const DisplayVtxTex16 vtx [[stage_in]],
 	                         +  tex.sample(genSampler, vtx.texCoord05 + float2( 2.0, 3.0)) * wx1.b
 	                         +  tex.sample(genSampler, vtx.texCoord05 + float2( 3.0, 3.0)) * wx2.b) * wy2.b;
 	
-	return float4(outFragment.rgb * *inBacklightIntensity, 1.0f);
+	return (doSwapRB == 0) ? float4(outFragment.rgb * *inBacklightIntensity, 1.0f) : float4(outFragment.bgr * *inBacklightIntensity, 1.0f);
 }
 
 #pragma mark NDS Emulation Functions