GFX3D:

- Remove NDS_3D_GetLineData(), as it wasn't appropriately named or used for its intended purpose.
- Add NDS_3D_RenderFinish(), which is what should have been used in the first place. (Purpose: Blocks the thread until 3D rendering is finished.)
- Optimize gpu3DNull so that it doesn't have to clear the 3D layer every frame.

task.h:
- Fix multiple #include compiling bug.

SoftRasterizer:
- Improve the performance of rendering in multithreaded mode.
- Improve the stability of reset and shutdown in multithreaded mode.
- Do some minor code cleanup.

OpenGL Renderer:
- Improve the stability of reset and shutdown.
- Do some minor code cleanup.

desmume/src/OGLRender.cpp

@@ -117,8 +117,7 @@ enum OGLTextureUnitID
 
 // Multithreading States
 static bool enableMultithreading = false;
-static bool isReadPixelsWorking = false;
+static Task oglReadPixelsTask[2];
-static Task oglReadPixelsTask;
 
 // Polygon Info
 static GLfloat polyAlpha = 1.0f;
@@ -498,12 +497,31 @@ static void OGLInitShaders(const char *oglExtensionString)
 
 static void OGLReset()
 {
+	gpuScreen3DHasNewData[0] = false;
+	gpuScreen3DHasNewData[1] = false;
+	
+	if (enableMultithreading)
+	{
+		for (unsigned int i = 0; i < 2; i++)
+		{
+			oglReadPixelsTask[i].finish();
+		}
+	}
+	
+	if(!BEGINGL())
+		return;
+	
+	glFinish();
+	
+	for (unsigned int i = 0; i < 2; i++)
+	{
+		memset(GPU_screen3D[i], 0, sizeof(GPU_screen3D[i]));
+	}
+	
 	if(isShaderSupported)
 	{
 		hasTexture = false;
 		
-		if(BEGINGL())
-		{
 		glUniform1i(uniformPolyID, 0);
 		glUniform1f(uniformPolyAlpha, 1.0f);
 		glUniform2f(uniformTexScale, 1.0f, 1.0f);
@@ -513,27 +531,18 @@ static void OGLReset()
 		glUniform1i(uniformWBuffer, 0);
 		glUniform1i(uniformEnableAlphaTest, GL_TRUE);
 		glUniform1f(uniformAlphaTestRef, 0.0f);
-			
-			ENDGL();
-		}
 	}
 	else
 	{
 		memset(color4fBuffer, 0, VERTLIST_SIZE * 4 * sizeof(GLfloat));
 	}
 	
-	TexCache_Reset();
+	ENDGL();
-	if (currTexture) 
-		delete currTexture;
-	currTexture = NULL;
-	
-	for (unsigned int i = 0; i < 2; i++)
-	{
-		memset(GPU_screen3D[i], 0, sizeof(GPU_screen3D[i]));
-		gpuScreen3DHasNewData[i] = false;
-	}
 	
 	memset(vertIndexBuffer, 0, VERT_INDEX_BUFFER_SIZE * sizeof(GLushort));
+	currTexture = NULL;
+	
+	Default3D_Reset();
 }
 
 //static class OGLTexCacheUser : public ITexCacheUser
@@ -636,13 +645,25 @@ static bool OGLIsMinimumVersionSupported(const char *oglVersionString)
 
 static char OGLInit(void)
 {
+	char result = 0;
+	
 	if(!oglrender_init)
-		return 0;
+		return result;
 	if(!oglrender_init())
-		return 0;
+		return result;
+	
+	result = Default3D_Init();
+	if (result == 0)
+	{
+		return result;
+	}
 
 	if(!BEGINGL())
-		return 0;
+	{
+		INFO("OpenGL: Could not initialize -- BEGINGL() failed.");
+		result = 0;
+		return result;
+	}
 	
 	// Get OpenGL info
 	const char *oglVendorString = (const char *)glGetString(GL_VENDOR);
@@ -655,13 +676,10 @@ static char OGLInit(void)
 			 OGL_MINIMUM_GPU_VERSION_REQUIRED_MAJOR, OGL_MINIMUM_GPU_VERSION_REQUIRED_MINOR, OGL_MINIMUM_GPU_VERSION_REQUIRED_REVISION,
 			 oglVersionString, oglVendorString, oglRendererString);
 		
-		return 0;
+		result = 0;
+		return result;
 	}
 	
-	glViewport(0, 0, 256, 192);
-	if (glGetError() != GL_NO_ERROR)
-		return 0;
-	
 	const char *oglExtensionString = (const char *)glGetString(GL_EXTENSIONS);
 
 	for (u8 i = 0; i < 255; i++)
@@ -936,8 +954,6 @@ static char OGLInit(void)
 	ENDGL();
 	
 	// Multithreading Setup
-	isReadPixelsWorking = false;
-
 	if (CommonSettings.num_cores > 1)
 	{
 #ifdef _WINDOWS
@@ -947,7 +963,11 @@ static char OGLInit(void)
 		enableMultithreading = false;
 #else
 		enableMultithreading = true;
-		oglReadPixelsTask.start(false);
+		
+		for (unsigned int i = 0; i < 2; i++)
+		{
+			oglReadPixelsTask[i].start(false);
+		}
 #endif
 	}
 	else
@@ -960,27 +980,28 @@ static char OGLInit(void)
 	INFO("OpenGL: Initialized successfully.\n[GPU Info - Version: %s, Vendor: %s, Renderer: %s]\n",
 		 oglVersionString, oglVendorString, oglRendererString);
 
-	return 1;
+	return result;
 }
 
 static void OGLClose()
 {
-	if (enableMultithreading)
-	{
-		oglReadPixelsTask.finish();
-		oglReadPixelsTask.shutdown();
-		isReadPixelsWorking = false;
-	}
-	
 	gpuScreen3DHasNewData[0] = false;
 	gpuScreen3DHasNewData[1] = false;
 	
-	delete [] vertIndexBuffer;
+	if (enableMultithreading)
-	vertIndexBuffer = NULL;
+	{
+		for (unsigned int i = 0; i < 2; i++)
+		{
+			oglReadPixelsTask[i].finish();
+			oglReadPixelsTask[i].shutdown();
+		}
+	}
 	
 	if(!BEGINGL())
 		return;
 	
+	glFinish();
+	
 	if(isShaderSupported)
 	{
 		glUseProgram(0);
@@ -992,6 +1013,8 @@ static void OGLClose()
 		glDeleteShader(vertexShaderID);
 		glDeleteShader(fragmentShaderID);
 		
+		glActiveTexture(GL_TEXTURE0 + OGLTextureUnitID_ToonTable);
+		glBindTexture(GL_TEXTURE_1D, 0);
 		glDeleteTextures(1, &texToonTableID);
 
 		isShaderSupported = false;
@@ -1004,27 +1027,20 @@ static void OGLClose()
 	
 	if (isVAOSupported)
 	{
+		glBindVertexArray(0);
 		glDeleteVertexArrays(1, &vaoMainStatesID);
 		isVAOSupported = false;
 	}
 
-	//kill the tex cache to free all the texture ids
-	TexCache_Reset();
-
-	while(!freeTextureIds.empty())
-	{
-		GLuint temp = freeTextureIds.front();
-		freeTextureIds.pop();
-		glDeleteTextures(1,&temp);
-	}
-
 	if (isVBOSupported)
 	{
+		glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);
 		glDeleteBuffersARB(1, &vboVertexID);
 	}
 	
 	if (isPBOSupported)
 	{
+		glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0);
 		glDeleteBuffersARB(2, pboRenderDataID);
 		pboRenderBuffer[0] = NULL;
 		pboRenderBuffer[1] = NULL;
@@ -1033,6 +1049,8 @@ static void OGLClose()
 	// FBO
 	if (isFBOSupported)
 	{
+		glActiveTexture(GL_TEXTURE0 + OGLTextureUnitID_ClearImage);
+		glBindTexture(GL_TEXTURE_2D, 0);
 		glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0);
 		
 		glDeleteFramebuffersEXT(1, &fboClearImageID);
@@ -1040,7 +1058,27 @@ static void OGLClose()
 		glDeleteTextures(1, &texClearImageDepthStencilID);
 	}
 	
+	//kill the tex cache to free all the texture ids
+	TexCache_Reset();
+	
+	glActiveTexture(GL_TEXTURE0);
+	glBindTexture(GL_TEXTURE_2D, 0);
+	
+	while(!freeTextureIds.empty())
+	{
+		GLuint temp = freeTextureIds.front();
+		freeTextureIds.pop();
+		glDeleteTextures(1,&temp);
+	}
+	
+	glFinish();
+	
 	ENDGL();
+	
+	delete [] vertIndexBuffer;
+	vertIndexBuffer = NULL;
+	
+	Default3D_Close();
 }
 
 static void texDeleteCallback(TexCacheItem* item)
@@ -1318,15 +1356,14 @@ static void GL_ReadFramebuffer()
 {
 	static unsigned int bufferIndex = 0;
 	
-	if (enableMultithreading && isReadPixelsWorking)
-	{
-		oglReadPixelsTask.finish();
-		isReadPixelsWorking = false;
-	}
-	
 	bufferIndex = (bufferIndex + 1) & 0x01;
 	gpuScreen3DBufferIndex = bufferIndex;
 	
+	if (enableMultithreading)
+	{
+		oglReadPixelsTask[bufferIndex].finish();
+	}
+	
 	if (isPBOSupported)
 	{
 		if(!BEGINGL()) return;
@@ -1353,14 +1390,12 @@ static void GL_ReadFramebuffer()
 	// H-Blank, and let that logic determine whether a pixel read is needed or not.
 	// This can save us some time in cases where games don't require the 3D layer
 	// for this particular frame.
+	
+	gpuScreen3DHasNewData[bufferIndex] = true;
+	
 	if (enableMultithreading)
 	{
-		isReadPixelsWorking = true;
+		oglReadPixelsTask[bufferIndex].execute(&execReadPixelsTask, &bufferIndex);
-		oglReadPixelsTask.execute(&execReadPixelsTask, &bufferIndex);
-	}
-	else
-	{
-		gpuScreen3DHasNewData[bufferIndex] = true;
 	}
 }
 
@@ -1738,27 +1773,26 @@ static void OGLRender()
 
 static void OGLVramReconfigureSignal()
 {
-	TexCache_Invalidate();
+	Default3D_VramReconfigureSignal();
 }
 
-static u8* OGLGetLineData(u8 lineNumber)
+static void OGLRenderFinish()
 {
 	// If OpenGL is still reading back pixels on a separate thread, wait for it to finish.
-	if (isReadPixelsWorking)
+	// Otherwise, just do the pixel read now.
-	{
-		oglReadPixelsTask.finish();
-		isReadPixelsWorking = false;
-	}
-	
-	// If we're doing a pixel read on this thread and we have new rendered data,
-	// then do the pixel read now.
 	if (gpuScreen3DHasNewData[gpuScreen3DBufferIndex])
+	{
+		if (enableMultithreading)
+		{
+			oglReadPixelsTask[gpuScreen3DBufferIndex].finish();
+		}
+		else
 		{
 			execReadPixelsTask(&gpuScreen3DBufferIndex);
-		gpuScreen3DHasNewData[gpuScreen3DBufferIndex] = false;
 		}
 		
-	return ( gfx3d_convertedScreen + (lineNumber << (8+2)) );
+		gpuScreen3DHasNewData[gpuScreen3DBufferIndex] = false;
+	}
 }
 
 GPU3DInterface gpu3Dgl = {
@@ -1767,6 +1801,6 @@ GPU3DInterface gpu3Dgl = {
 	OGLReset,
 	OGLClose,
 	OGLRender,
-	OGLVramReconfigureSignal,
+	OGLRenderFinish,
-	OGLGetLineData
+	OGLVramReconfigureSignal
 };

desmume/src/gfx3d.cpp

@@ -489,6 +489,8 @@ void gfx3d_init()
 
 void gfx3d_reset()
 {
+	gpu3D->NDS_3D_RenderFinish();
+	
 #ifdef _SHOW_VTX_COUNTERS
 	max_polys = max_verts = 0;
 #endif
@@ -2183,8 +2185,7 @@ void gfx3d_VBlankEndSignal(bool skipFrame)
 
 	drawPending = FALSE;
 
-	//if the null 3d core is chosen, then we need to clear out the 3d buffers to keep old data from being rendered
+	if(!CommonSettings.showGpu.main)
-	if(gpu3D == &gpu3DNull || !CommonSettings.showGpu.main)
 	{
 		memset(gfx3d_convertedScreen,0,sizeof(gfx3d_convertedScreen));
 		return;
@@ -2301,15 +2302,9 @@ void gfx3d_glGetLightColor(unsigned int index, unsigned int* dest)
 
 void gfx3d_GetLineData(int line, u8** dst)
 {
-	if (gpu3D->NDS_3D_GetLineData == NULL)
+	gpu3D->NDS_3D_RenderFinish();
-	{
 	*dst = gfx3d_convertedScreen+((line)<<(8+2));
 }
-	else
-	{
-		*dst = gpu3D->NDS_3D_GetLineData(line);
-	}
-}
 
 void gfx3d_GetLineData15bpp(int line, u16** dst)
 {

desmume/src/rasterize.cpp

@@ -67,7 +67,7 @@ static u8 decal_table[32][64][64];
 static u8 index_lookup_table[65];
 static u8 index_start_table[8];
 
-
+static bool softRastHasNewData = false;
 
 ////optimized float floor useful in limited cases
 ////from http://www.stereopsis.com/FPU.html#convert
@@ -1073,7 +1073,7 @@ static SoftRasterizerEngine mainSoftRasterizer;
 static Task rasterizerUnitTask[_MAX_CORES];
 static RasterizerUnit<true> rasterizerUnit[_MAX_CORES];
 static RasterizerUnit<false> _HACK_viewer_rasterizerUnit;
-static int rasterizerCores;
+static unsigned int rasterizerCores;
 static bool rasterizerUnitTasksInited = false;
 
 static void* execRasterizerUnit(void* arg)
@@ -1085,6 +1085,12 @@ static void* execRasterizerUnit(void* arg)
 
 static char SoftRastInit(void)
 {
+	char result = Default3D_Init();
+	if (result == 0)
+	{
+		return result;
+	}
+	
 	if(!rasterizerUnitTasksInited)
 	{
 		rasterizerUnitTasksInited = true;
@@ -1095,7 +1101,7 @@ static char SoftRastInit(void)
 		rasterizerCores = CommonSettings.num_cores;
 		if (rasterizerCores > _MAX_CORES) 
 			rasterizerCores = _MAX_CORES;
-		if(CommonSettings.num_cores == 1)
+		if(CommonSettings.num_cores <= 1)
 		{
 			rasterizerCores = 1;
 			rasterizerUnit[0].SLI_MASK = 0;
@@ -1146,22 +1152,44 @@ static char SoftRastInit(void)
 	TexCache_Reset();
 
 	printf("SoftRast Initialized with cores=%d\n",rasterizerCores);
-	return 1;
+	return result;
 }
 
-static void SoftRastReset() {
+static void SoftRastReset()
-	TexCache_Reset();
+{
+	if (rasterizerCores > 1)
+	{
+		for(unsigned int i = 0; i < rasterizerCores; i++)
+		{
+			rasterizerUnitTask[i].finish();
+		}
+	}
+	
+	softRastHasNewData = false;
+	
+	Default3D_Reset();
 }
 
 static void SoftRastClose()
 {
-	for(int i=0; i<_MAX_CORES; i++)
+	if (rasterizerCores > 1)
+	{
+		for(unsigned int i = 0; i < rasterizerCores; i++)
+		{
+			rasterizerUnitTask[i].finish();
 			rasterizerUnitTask[i].shutdown();
-	rasterizerUnitTasksInited = false;
+		}
 	}
 	
-static void SoftRastVramReconfigureSignal() {
+	rasterizerUnitTasksInited = false;
-	TexCache_Invalidate();
+	softRastHasNewData = false;
+	
+	Default3D_Close();
+}
+
+static void SoftRastVramReconfigureSignal()
+{
+	Default3D_VramReconfigureSignal();
 }
 
 static void SoftRastConvertFramebuffer()
@@ -1622,24 +1650,44 @@ static void SoftRastRender()
 	mainSoftRasterizer.performCoordAdjustment(true);
 	mainSoftRasterizer.setupTextures(true);
 
+	softRastHasNewData = true;
 	
-	if(rasterizerCores==1)
+	if (rasterizerCores > 1)
 	{
-		rasterizerUnit[0].mainLoop<false>(&mainSoftRasterizer);
+		for(unsigned int i = 0; i < rasterizerCores; i++)
+		{
+			rasterizerUnitTask[i].execute(&execRasterizerUnit, (void *)i);
+		}
 	}
 	else
 	{
-		for(int i=0;i<rasterizerCores;i++) rasterizerUnitTask[i].execute(execRasterizerUnit,(void*)i);
+		rasterizerUnit[0].mainLoop<false>(&mainSoftRasterizer);
-		for(int i=0;i<rasterizerCores;i++) rasterizerUnitTask[i].finish();
+	}
+}
+
+static void SoftRastRenderFinish()
+{
+	if (!softRastHasNewData)
+	{
+		return;
+	}
+	
+	if (rasterizerCores > 1)
+	{
+		for(unsigned int i = 0; i < rasterizerCores; i++)
+		{
+			rasterizerUnitTask[i].finish();
+		}
 	}
 	
 	TexCache_EvictFrame();
 	
-
 	mainSoftRasterizer.framebufferProcess();
 	
 	//	printf("rendered %d of %d polys after backface culling\n",gfx3d.polylist->count-culled,gfx3d.polylist->count);
 	SoftRastConvertFramebuffer();
+	
+	softRastHasNewData = false;
 }
 
 GPU3DInterface gpu3DRasterize = {
@@ -1648,7 +1696,7 @@ GPU3DInterface gpu3DRasterize = {
 	SoftRastReset,
 	SoftRastClose,
 	SoftRastRender,
-	SoftRastVramReconfigureSignal,
+	SoftRastRenderFinish,
-	NULL
+	SoftRastVramReconfigureSignal
 };
 

desmume/src/render3D.cpp

@@ -17,23 +17,62 @@
 */
 
 #include "render3D.h"
+#include "gfx3d.h"
+#include "texcache.h"
 
 int cur3DCore = GPU3D_NULL;
 
-static void NDS_nullFunc1		(void){}
-static char NDS_nullFunc2		(void){ return 1; }
-
 GPU3DInterface gpu3DNull = { 
 	"None",
-	NDS_nullFunc2, //NDS_3D_Init
+	Default3D_Init,
-	NDS_nullFunc1, //NDS_3D_Reset
+	Default3D_Reset,
-	NDS_nullFunc1, //NDS_3D_Close
+	Default3D_Close,
-	NDS_nullFunc1, //NDS_3D_Render
+	Default3D_Render,
-	NDS_nullFunc1, //NDS_3D_VramReconfigureSignal
+	Default3D_RenderFinish,
-	0
+	Default3D_VramReconfigureSignal
 };
 
 GPU3DInterface *gpu3D = &gpu3DNull;
+static bool default3DAlreadyClearedLayer = false;
+
+char Default3D_Init()
+{
+	default3DAlreadyClearedLayer = false;
+	
+	return 1;
+}
+
+void Default3D_Reset()
+{
+	default3DAlreadyClearedLayer = false;
+	
+	TexCache_Reset();
+}
+
+void Default3D_Close()
+{
+	memset(gfx3d_convertedScreen, 0, sizeof(gfx3d_convertedScreen));
+	default3DAlreadyClearedLayer = false;
+}
+
+void Default3D_Render()
+{
+	if (!default3DAlreadyClearedLayer)
+	{
+		memset(gfx3d_convertedScreen, 0, sizeof(gfx3d_convertedScreen));
+		default3DAlreadyClearedLayer = true;
+	}
+}
+
+void Default3D_RenderFinish()
+{
+	// Do nothing
+}
+
+void Default3D_VramReconfigureSignal()
+{
+	TexCache_Invalidate();
+}
 
 void NDS_3D_SetDriver (int core3DIndex)
 {

desmume/src/render3D.h

@@ -30,23 +30,25 @@ typedef struct Render3DInterface
 	const char * name;
 
 	//called once when the plugin starts up
-	char (CALL_CONVENTION*  NDS_3D_Init)					(void);
+	char (CALL_CONVENTION*  NDS_3D_Init)					();
 	
 	//called when the emulator resets (is this necessary?)
-	void (CALL_CONVENTION*  NDS_3D_Reset)					(void);
+	void (CALL_CONVENTION*  NDS_3D_Reset)					();
 	
 	//called when the plugin shuts down
-	void (CALL_CONVENTION*  NDS_3D_Close)					(void);
+	void (CALL_CONVENTION*  NDS_3D_Close)					();
 	
 	//called when the renderer should do its job and render the current display lists
-	void (CALL_CONVENTION*  NDS_3D_Render)					(void);
+	void (CALL_CONVENTION*  NDS_3D_Render)					();
+	
+	// Called whenever 3D rendering needs to finish. This function should block the calling thread
+	// and only release the block when 3D rendering is finished. (Before reading the 3D layer, be
+	// sure to always call this function.)
+	void (CALL_CONVENTION*	NDS_3D_RenderFinish)			();
 
 	//called when the emulator reconfigures its vram. you may need to invalidate your texture cache.
 	void (CALL_CONVENTION*  NDS_3D_VramReconfigureSignal)	();
 
-	//called when the emulator requests rendered graphics data
-	u8* (CALL_CONVENTION*	NDS_3D_GetLineData)				(u8 lineNumber);
-
 } GPU3DInterface;
 
 extern int cur3DCore;
@@ -61,6 +63,13 @@ extern GPU3DInterface gpu3DNull;
 // Extern pointer
 extern GPU3DInterface *gpu3D;
 
+char Default3D_Init();
+void Default3D_Reset();
+void Default3D_Close();
+void Default3D_Render();
+void Default3D_RenderFinish();
+void Default3D_VramReconfigureSignal();
+
 void NDS_3D_SetDriver (int core3DIndex);
 bool NDS_3D_ChangeCore(int newCore);
 

desmume/src/utils/task.h

@@ -16,6 +16,7 @@
 */
 
 #ifndef _TASK_H_
+#define _TASK_H_
 
 //Sort of like a single-thread thread pool.
 //You hand it a worker function and then call finish() to synch with its completion