core: fix an old shortcoming in texcache where games that burned through sprites memory with lots of tiny sprites could use gigabytes of texcache buffers (metal slug)

desmume/ChangeLog

@@ -7,6 +7,7 @@ General/Core:
  bug: fix div and sqrt busy flag bug
  bug: fix vectest
  bug: fix lid savestate desync
+ bug: fix texcache memory GB explosion when games use tons of tiny 3d sprites
  enh: support devkitpro argv
  enh: add gbaslot-rom commandline
 
@@ -24,8 +25,10 @@ Windows:
  
 Linux/OSX:
  bug: fix building for nosse2 systems
+ bug: fix --num-cores
  enh: add --nojoy=1 to fix laptops with accelerometers
  
+ 
 0.9.4 -> 0.9.5 (r2437-r3075)
 
 0.9.5 introduces an entirely rewritten main emulation loop

desmume/src/driver.cpp

@@ -84,7 +84,7 @@ public:
 
 		engine.updateFogTable();
 	
-		engine.initFramebuffer(kViewportWidth,kViewportHeight,gfx3d.state.enableClearImage);
+		engine.initFramebuffer(kViewportWidth,kViewportHeight,gfx3d.state.enableClearImage?true:false);
 		engine.updateToonTable();
 		engine.updateFloatColors();
 		engine.performClipping(checkMaterialInterpolate->IsChecked());

desmume/src/rasterize.cpp

@@ -1437,7 +1437,7 @@ static void SoftRastRender()
 	if(gfx3d.state.enableFog && CommonSettings.GFX3D_Fog)
 		mainSoftRasterizer.updateFogTable();
 	
-	mainSoftRasterizer.initFramebuffer(256,192,gfx3d.state.enableClearImage);
+	mainSoftRasterizer.initFramebuffer(256,192,gfx3d.state.enableClearImage?true:false);
 	mainSoftRasterizer.updateToonTable();
 	mainSoftRasterizer.updateFloatColors();
 	mainSoftRasterizer.performClipping(CommonSettings.GFX3D_HighResolutionInterpolateColor);

desmume/src/texcache.cpp

@@ -193,7 +193,11 @@ public:
 	TTexCacheItemMultimap index;
 
 	//this ought to be enough for anyone
-	static const u32 kMaxCacheSize = 64*1024*1024; 
+	//static const u32 kMaxCacheSize = 64*1024*1024; 
+	//changed by zeromus on 15-dec. I couldnt find any games that were getting anywhere NEAR 64
+	static const u32 kMaxCacheSize = 16*1024*1024; 
+	//metal slug burns through sprites so fast, it can test it pretty quickly though
+
 	//this is not really precise, it is off by a constant factor
 	u32 cache_size;
 
@@ -307,7 +311,7 @@ public:
 			if(mspal.size != 0 && memcmp(curr->dump.palette,pal,mspal.size)) goto REJECT;
 
 			//when the texture data doesn't match
-			if(ms.memcmp(curr->dump.texture,sizeof(curr->dump.texture))) goto REJECT;
+			if(ms.memcmp(&curr->dump.texture[0],sizeof(curr->dump.texture))) goto REJECT;
 
 			//if the texture is 4x4 then the index data must match
 			if(textureMode == TEXMODE_4X4)
@@ -342,7 +346,6 @@ public:
 		newitem->sizeY=sizeY;
 		newitem->invSizeX=1.0f/((float)(sizeX));
 		newitem->invSizeY=1.0f/((float)(sizeY));
-		newitem->dump.textureSize = ms.dump(newitem->dump.texture,sizeof(newitem->dump.texture));
 		newitem->decode_len = sizeX*sizeY*4;
 		newitem->mode = textureMode;
 		newitem->decoded = new u8[newitem->decode_len];
@@ -356,13 +359,15 @@ public:
 		{
 			memcpy(newitem->dump.palette, pal, palSize*2);
 		}
-		//dump 4x4 index data for cache keying
-		newitem->dump.indexSize = 0;
+
+		//dump texture and 4x4 index data for cache keying
+		const int texsize = newitem->dump.textureSize = ms.size;
+		const int indexsize = newitem->dump.indexSize = msIndex.size;
+		newitem->dump.texture = new u8[texsize+indexsize];
+		ms.dump(&newitem->dump.texture[0],newitem->dump.maxTextureSize); //dump texture
 		if(textureMode == TEXMODE_4X4)
-		{
-			newitem->dump.indexSize = min(msIndex.size,(int)sizeof(newitem->dump.texture) - newitem->dump.textureSize);
-			msIndex.dump(newitem->dump.texture+newitem->dump.textureSize,newitem->dump.indexSize);
-		}
+			msIndex.dump(newitem->dump.texture+newitem->dump.textureSize,newitem->dump.indexSize); //dump 4x4
+
 
 		//============================================================================ 
 		//Texture conversion
@@ -637,6 +642,7 @@ public:
 
 	void evict(u32 target = kMaxCacheSize)
 	{
+		printf("%d %d/%d\n",index.size(),cache_size/1024,target/1024);
 		//dont do anything unless we're over the target
 		if(cache_size<target) return;
 

desmume/src/texcache.h

@@ -44,11 +44,14 @@ public:
 
 	TexCache_TexFormat cacheFormat;
 
-	//TODO - this is a little wasteful
-	struct {
-		int					textureSize, indexSize;
-		u8					texture[128*1024]; // 128Kb texture slot
-		u8					palette[256*2];
+	struct Dump {
+		~Dump() {
+			delete[] texture;
+		}
+		int textureSize, indexSize;
+		static const int maxTextureSize=128*1024;
+		u8* texture;
+		u8 palette[256*2];
 	} dump;
 };