Save States: Properly rescale and convert the internal 3D framebuffer to and from a standard format -- 32-bit RGBA8888. (Related to commit c1eafc5 and completes all the work in PR #190 by @SuuperW.)

desmume/src/GPU.cpp

@@ -718,8 +718,8 @@ static FORCEINLINE void CopyLineExpand(void *__restrict dst, const void *__restr
 }
 
 template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
-static void CopyLineExpandHinted(const void *__restrict srcBuffer, const size_t srcLineIndex,
+void CopyLineExpandHinted(const void *__restrict srcBuffer, const size_t srcLineIndex,
-								 void *__restrict dstBuffer, const size_t dstLineIndex, const size_t dstLineWidth, const size_t dstLineCount)
+						  void *__restrict dstBuffer, const size_t dstLineIndex, const size_t dstLineWidth, const size_t dstLineCount)
 {
 	switch (INTEGERSCALEHINT)
 	{
@@ -784,7 +784,7 @@ static void CopyLineExpandHinted(const void *__restrict srcBuffer, const size_t
 }
 
 template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
-static void CopyLineExpandHinted(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer)
+void CopyLineExpandHinted(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer)
 {
 	CopyLineExpandHinted<INTEGERSCALEHINT, USELINEINDEX, NEEDENDIANSWAP, ELEMENTSIZE>(srcBuffer, lineInfo.indexNative,
 																					  dstBuffer, lineInfo.indexCustom, lineInfo.widthCustom, lineInfo.renderCount);
@@ -1184,8 +1184,8 @@ static FORCEINLINE void CopyLineReduce(void *__restrict dst, const void *__restr
 }
 
 template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
-static void CopyLineReduceHinted(const void *__restrict srcBuffer, const size_t srcLineIndex, const size_t srcLineWidth,
+void CopyLineReduceHinted(const void *__restrict srcBuffer, const size_t srcLineIndex, const size_t srcLineWidth,
-								 void *__restrict dstBuffer, const size_t dstLineIndex)
+						  void *__restrict dstBuffer, const size_t dstLineIndex)
 {
 	switch (INTEGERSCALEHINT)
 	{
@@ -1250,7 +1250,7 @@ static void CopyLineReduceHinted(const void *__restrict srcBuffer, const size_t
 }
 
 template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
-static void CopyLineReduceHinted(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer)
+void CopyLineReduceHinted(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer)
 {
 	CopyLineReduceHinted<INTEGERSCALEHINT, USELINEINDEX, NEEDENDIANSWAP, ELEMENTSIZE>(srcBuffer, lineInfo.indexCustom, lineInfo.widthCustom,
 																					  dstBuffer, lineInfo.indexNative);
@@ -6393,7 +6393,6 @@ GPUEngineA::GPUEngineA()
 	_captureWorkingB16 = (u16 *)malloc_alignedPage(GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(u16));
 	_captureWorkingA32 = (FragmentColor *)malloc_alignedPage(GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(FragmentColor));
 	_captureWorkingB32 = (FragmentColor *)malloc_alignedPage(GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(FragmentColor));
-	gfx3d_Update3DFramebuffers(_3DFramebufferMain, _3DFramebuffer16);
 }
 
 GPUEngineA::~GPUEngineA()
@@ -6404,7 +6403,6 @@ GPUEngineA::~GPUEngineA()
 	free_aligned(this->_captureWorkingB16);
 	free_aligned(this->_captureWorkingA32);
 	free_aligned(this->_captureWorkingB32);
-	gfx3d_Update3DFramebuffers(NULL, NULL);
 }
 
 GPUEngineA* GPUEngineA::Allocate()
@@ -6552,7 +6550,6 @@ void GPUEngineA::SetCustomFramebufferSize(size_t w, size_t h)
 	this->_captureWorkingB16 = newCaptureWorkingB16;
 	this->_captureWorkingA32 = newCaptureWorkingA32;
 	this->_captureWorkingB32 = newCaptureWorkingB32;
-	gfx3d_Update3DFramebuffers(this->_3DFramebufferMain, this->_3DFramebuffer16);
 	
 	const NDSDisplayInfo &dispInfo = GPU->GetDisplayInfo();
 	
@@ -8133,6 +8130,19 @@ GPUSubsystem::GPUSubsystem()
 	
 	gfx3d_init();
 	
+	for (size_t line = 0; line < GPU_VRAM_BLOCK_LINES + 1; line++)
+	{
+		GPUEngineLineInfo &lineInfo = this->_lineInfo[line];
+		
+		lineInfo.indexNative = line;
+		lineInfo.indexCustom = lineInfo.indexNative;
+		lineInfo.widthCustom = GPU_FRAMEBUFFER_NATIVE_WIDTH;
+		lineInfo.renderCount = 1;
+		lineInfo.pixelCount = GPU_FRAMEBUFFER_NATIVE_WIDTH;
+		lineInfo.blockOffsetNative = lineInfo.indexNative * GPU_FRAMEBUFFER_NATIVE_WIDTH;
+		lineInfo.blockOffsetCustom = lineInfo.indexCustom * GPU_FRAMEBUFFER_NATIVE_WIDTH;
+	}
+	
 	_engineMain = GPUEngineA::Allocate();
 	_engineSub = GPUEngineB::Allocate();
 	
@@ -8443,6 +8453,11 @@ const NDSDisplayInfo& GPUSubsystem::GetDisplayInfo()
 	return this->_displayInfo;
 }
 
+const GPUEngineLineInfo& GPUSubsystem::GetLineInfoAtIndex(size_t l)
+{
+	return this->_lineInfo[l];
+}
+
 u32 GPUSubsystem::GetFPSRender3D() const
 {
 	return this->_render3DFrameCount;
@@ -8593,6 +8608,19 @@ void GPUSubsystem::SetCustomFramebufferSize(size_t w, size_t h)
 	_gpuDstToSrcSSSE3_u16_8e = newGpuDstToSrcSSSE3_u16_8e;
 	_gpuDstToSrcSSSE3_u32_4e = newGpuDstToSrcSSSE3_u32_4e;
 	
+	for (size_t line = 0; line < GPU_VRAM_BLOCK_LINES + 1; line++)
+	{
+		GPUEngineLineInfo &lineInfo = this->_lineInfo[line];
+		
+		lineInfo.indexNative = line;
+		lineInfo.indexCustom = _gpuCaptureLineIndex[lineInfo.indexNative];
+		lineInfo.widthCustom = w;
+		lineInfo.renderCount = _gpuCaptureLineCount[lineInfo.indexNative];
+		lineInfo.pixelCount = lineInfo.widthCustom * lineInfo.renderCount;
+		lineInfo.blockOffsetNative = lineInfo.indexNative * GPU_FRAMEBUFFER_NATIVE_WIDTH;
+		lineInfo.blockOffsetCustom = lineInfo.indexCustom * lineInfo.widthCustom;
+	}
+	
 	CurrentRenderer->RenderFinish();
 	CurrentRenderer->SetRenderNeedsFinish(false);
 	
@@ -9153,8 +9181,8 @@ u8* GPUSubsystem::_DownscaleAndConvertForSavestate(const NDSDisplayID displayID,
 					isIntermediateBufferMissing = (intermediateBuffer == NULL);
 					if (!isIntermediateBufferMissing)
 					{
-						const u32 *src = (u32 *)this->_displayInfo.customBuffer[displayID];
+						const u32 *__restrict src = (u32 *__restrict)this->_displayInfo.customBuffer[displayID];
-						u32 *dst = (u32 *)intermediateBuffer;
+						u32 *__restrict dst = (u32 *__restrict)intermediateBuffer;
 						
 						for (size_t l = 0; l < GPU_FRAMEBUFFER_NATIVE_HEIGHT; l++)
 						{
@@ -9650,3 +9678,7 @@ template void GPUEngineBase::ParseReg_BGnY<GPULayerID_BG3>();
 template void GPUSubsystem::RenderLine<NDSColorFormat_BGR555_Rev>(const size_t l);
 template void GPUSubsystem::RenderLine<NDSColorFormat_BGR666_Rev>(const size_t l);
 template void GPUSubsystem::RenderLine<NDSColorFormat_BGR888_Rev>(const size_t l);
+
+// These functions are used in gfx3d.cpp
+template void CopyLineExpandHinted<0xFFFF, false, true, 4>(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer);
+template void CopyLineReduceHinted<0xFFFF, false, true, 4>(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer);

desmume/src/GPU.h

@@ -1763,6 +1763,7 @@ private:
 	GPUEngineB *_engineSub;
 	NDSDisplay *_display[2];
 	float _backlightIntensityTotal[2];
+	GPUEngineLineInfo _lineInfo[GPU_VRAM_BLOCK_LINES + 1];
 	
 	int _pending3DRendererID;
 	bool _needChange3DRenderer;
@@ -1797,6 +1798,7 @@ public:
 	void ForceFrameStop();
 	
 	const NDSDisplayInfo& GetDisplayInfo(); // Frontends need to call this whenever they need to read the video buffers from the emulator core
+	const GPUEngineLineInfo& GetLineInfoAtIndex(size_t l);
 	u32 GetFPSRender3D() const;
 	
 	GPUEngineA* GetEngineMain();
@@ -1898,6 +1900,20 @@ public:
 	void SetClientData(void *clientData);
 };
 
+template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
+void CopyLineExpandHinted(const void *__restrict srcBuffer, const size_t srcLineIndex,
+						  void *__restrict dstBuffer, const size_t dstLineIndex, const size_t dstLineWidth, const size_t dstLineCount);
+
+template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
+void CopyLineExpandHinted(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer);
+
+template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
+void CopyLineReduceHinted(const void *__restrict srcBuffer, const size_t srcLineIndex, const size_t srcLineWidth,
+						  void *__restrict dstBuffer, const size_t dstLineIndex);
+
+template <s32 INTEGERSCALEHINT, bool USELINEINDEX, bool NEEDENDIANSWAP, size_t ELEMENTSIZE>
+void CopyLineReduceHinted(const GPUEngineLineInfo &lineInfo, const void *__restrict srcBuffer, void *__restrict dstBuffer);
+
 extern GPUSubsystem *GPU;
 extern MMU_struct MMU;
 

desmume/src/gfx3d.cpp

@@ -283,8 +283,7 @@ static float normalTable[1024];
 #define fix10_2float(v) (((float)((s32)(v))) / (float)(1<<9))
 
 // Color buffer that is filled by the 3D renderer and is read by the GPU engine.
-static FragmentColor *_gfx3d_colorMain = NULL;
+static CACHE_ALIGN FragmentColor _gfx3d_savestateBuffer[GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT];
-static u16 *_gfx3d_color16 = NULL;
 
 // Matrix stack handling
 //TODO: decouple stack pointers from matrix stack type
@@ -609,6 +608,7 @@ void gfx3d_reset()
 	memset(gxPIPE.param, 0, sizeof(gxPIPE.param));
 	memset(colorRGB, 0, sizeof(colorRGB));
 	memset(&tempVertInfo, 0, sizeof(tempVertInfo));
+	memset(_gfx3d_savestateBuffer, 0, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT * sizeof(u32));
 
 	MatrixInit(mtxCurrent[MATRIXMODE_PROJECTION]);
 	MatrixInit(mtxCurrent[MATRIXMODE_POSITION]);
@@ -2717,24 +2717,54 @@ SFORMAT SF_GFX3D[]={
 	{ "GTVC", 4, 1, &tempVertInfo.count},
 	{ "GTVM", 4, 4, tempVertInfo.map},
 	{ "GTVF", 4, 1, &tempVertInfo.first},
-	{ "G3CX", 1, 4*GPU_FRAMEBUFFER_NATIVE_WIDTH*GPU_FRAMEBUFFER_NATIVE_HEIGHT, _gfx3d_colorMain},
+	{ "G3CX", 1, 4*GPU_FRAMEBUFFER_NATIVE_WIDTH*GPU_FRAMEBUFFER_NATIVE_HEIGHT, _gfx3d_savestateBuffer},
 	{ 0 }
 };
 
-void gfx3d_Update3DFramebuffers(FragmentColor *framebufferMain, u16 *framebuffer16)
-{
-	_gfx3d_colorMain = framebufferMain;
-	_gfx3d_color16 = framebuffer16;
-}
-
 //-------------savestate
-void gfx3d_savestate(EMUFILE &os)
+void gfx3d_PrepareSaveStateBufferWrite()
 {
 	if (CurrentRenderer->GetRenderNeedsFinish())
 	{
 		GPU->ForceRender3DFinishAndFlush(true);
 	}
 	
+	const size_t w = CurrentRenderer->GetFramebufferWidth();
+	const size_t h = CurrentRenderer->GetFramebufferHeight();
+	
+	if ( (w == GPU_FRAMEBUFFER_NATIVE_WIDTH) && (h == GPU_FRAMEBUFFER_NATIVE_HEIGHT) ) // Framebuffer is at the native size
+	{
+		if (CurrentRenderer->GetColorFormat() == NDSColorFormat_BGR666_Rev)
+		{
+			ColorspaceConvertBuffer6665To8888<false, false>((u32 *)CurrentRenderer->GetFramebuffer(), (u32 *)_gfx3d_savestateBuffer, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
+		}
+		else
+		{
+			ColorspaceCopyBuffer32<false, false>((u32 *)CurrentRenderer->GetFramebuffer(), (u32 *)_gfx3d_savestateBuffer, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
+		}
+	}
+	else // Framebuffer is at a custom size
+	{
+		const FragmentColor *__restrict src = CurrentRenderer->GetFramebuffer();
+		FragmentColor *__restrict dst = _gfx3d_savestateBuffer;
+		
+		for (size_t l = 0; l < GPU_FRAMEBUFFER_NATIVE_HEIGHT; l++)
+		{
+			const GPUEngineLineInfo &lineInfo = GPU->GetLineInfoAtIndex(l);
+			CopyLineReduceHinted<0xFFFF, false, true, 4>(lineInfo, src, dst);
+			src += lineInfo.pixelCount;
+			dst += GPU_FRAMEBUFFER_NATIVE_WIDTH;
+		}
+		
+		if (CurrentRenderer->GetColorFormat() == NDSColorFormat_BGR666_Rev)
+		{
+			ColorspaceConvertBuffer6665To8888<false, false>((u32 *)_gfx3d_savestateBuffer, (u32 *)_gfx3d_savestateBuffer, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
+		}
+	}
+}
+
+void gfx3d_savestate(EMUFILE &os)
+{
 	//version
 	os.write_32LE(4);
 
@@ -2905,6 +2935,59 @@ bool gfx3d_loadstate(EMUFILE &is, int size)
 	return true;
 }
 
+void gfx3d_FinishLoadStateBufferRead()
+{
+	const Render3DDeviceInfo &deviceInfo = CurrentRenderer->GetDeviceInfo();
+	
+	switch (deviceInfo.renderID)
+	{
+		case RENDERID_NULL:
+			memset(CurrentRenderer->GetFramebuffer(), 0, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT * sizeof(FragmentColor));
+			break;
+			
+		case RENDERID_SOFTRASTERIZER:
+		{
+			const size_t w = CurrentRenderer->GetFramebufferWidth();
+			const size_t h = CurrentRenderer->GetFramebufferHeight();
+			
+			if ( (w == GPU_FRAMEBUFFER_NATIVE_WIDTH) && (h == GPU_FRAMEBUFFER_NATIVE_HEIGHT) ) // Framebuffer is at the native size
+			{
+				if (CurrentRenderer->GetColorFormat() == NDSColorFormat_BGR666_Rev)
+				{
+					ColorspaceConvertBuffer8888To6665<false, false>((u32 *)_gfx3d_savestateBuffer, (u32 *)CurrentRenderer->GetFramebuffer(), GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
+				}
+				else
+				{
+					ColorspaceCopyBuffer32<false, false>((u32 *)_gfx3d_savestateBuffer, (u32 *)CurrentRenderer->GetFramebuffer(), GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
+				}
+			}
+			else // Framebuffer is at a custom size
+			{
+				if (CurrentRenderer->GetColorFormat() == NDSColorFormat_BGR666_Rev)
+				{
+					ColorspaceConvertBuffer8888To6665<false, false>((u32 *)_gfx3d_savestateBuffer, (u32 *)_gfx3d_savestateBuffer, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
+				}
+				
+				const FragmentColor *__restrict src = _gfx3d_savestateBuffer;
+				FragmentColor *__restrict dst = CurrentRenderer->GetFramebuffer();
+				
+				for (size_t l = 0; l < GPU_FRAMEBUFFER_NATIVE_HEIGHT; l++)
+				{
+					const GPUEngineLineInfo &lineInfo = GPU->GetLineInfoAtIndex(l);
+					CopyLineExpandHinted<0xFFFF, false, true, 4>(lineInfo, src, dst);
+					src += GPU_FRAMEBUFFER_NATIVE_WIDTH;
+					dst += lineInfo.pixelCount;
+				}
+			}
+			break;
+		}
+			
+		default:
+			// Do nothing. Loading the 3D framebuffer is unsupported on this 3D renderer.
+			break;
+	}
+}
+
 void gfx3d_parseCurrentDISP3DCNT()
 {
 	const IOREG_DISP3DCNT &DISP3DCNT = gfx3d.state.savedDISP3DCNT;

desmume/src/gfx3d.h

@@ -636,9 +636,10 @@ void gfx3d_glGetLightColor(const size_t index, u32 &dst);
 
 struct SFORMAT;
 extern SFORMAT SF_GFX3D[];
-void gfx3d_Update3DFramebuffers(FragmentColor *framebufferMain, u16 *framebuffer16);
+void gfx3d_PrepareSaveStateBufferWrite();
 void gfx3d_savestate(EMUFILE &os);
 bool gfx3d_loadstate(EMUFILE &is, int size);
+void gfx3d_FinishLoadStateBufferRead();
 
 void gfx3d_ClearStack();
 

desmume/src/saves.cpp

@@ -1061,6 +1061,7 @@ static void writechunks(EMUFILE &os)
 
 	save_time = tm.get_Ticks();
 	
+	gfx3d_PrepareSaveStateBufferWrite();
 	wifiHandler->PrepareSaveStateWrite();
 
 	savestate_WriteChunk(os,1,SF_ARM9);
@@ -1184,7 +1185,13 @@ static bool ReadStateChunks(EMUFILE &is, s32 totalsize)
 	}
 	
 	if (chunkError)
+	{
 		msgbox->warn("There was an error loading the savestate. Your game session is probably corrupt now.");
+	}
+	else
+	{
+		gfx3d_FinishLoadStateBufferRead();
+	}
 	
 	if (haveInfo)
 	{