Merge pull request #7716 from unknownbrackets/clut

Avoid accidental sign ext for > 24 bit clut shift, clut wrapping

GPU/Common/DepalettizeShaderCommon.cpp

@@ -88,7 +88,7 @@ void GenerateDepalShader300(char *buffer, GEBufferFormat pixelFormat) {
 		texturePixels = 512;
 
 	if (shift) {
-		WRITE(p, "  index = ((index >> %i) & 0x%02x)", shift, mask);
+		WRITE(p, "  index = (int(uint(index) >> %i) & 0x%02x)", shift, mask);
 	} else {
 		WRITE(p, "  index = (index & 0x%02x)", mask);
 	}
@@ -213,6 +213,7 @@ void GenerateDepalShaderFloat(char *buffer, GEBufferFormat pixelFormat, ShaderLa
 	}
 
 	// Offset by half a texel (plus clutBase) to turn NEAREST filtering into FLOOR.
+	// Technically, the clutBase should be |'d, not added, but that's hard with floats.
 	float texel_offset = ((float)clutBase + 0.5f) / texturePixels;
 	sprintf(offset, " + %f", texel_offset);
 

GPU/Directx9/DepalettizeShaderDX9.cpp

@@ -109,7 +109,7 @@ LPDIRECT3DTEXTURE9 DepalShaderCacheDX9::GetClutTexture(const u32 clutID, u32 *ra
 	memcpy(rect.pBits, rawClut, 1024);
 	tex->texture->UnlockRect(0);
 
-	pD3Ddevice->SetSamplerState(1, D3DSAMP_ADDRESSU, D3DTADDRESS_CLAMP);
+	pD3Ddevice->SetSamplerState(1, D3DSAMP_ADDRESSU, D3DTADDRESS_WRAP);
 	pD3Ddevice->SetSamplerState(1, D3DSAMP_ADDRESSV, D3DTADDRESS_CLAMP);
 	pD3Ddevice->SetSamplerState(1, D3DSAMP_MINFILTER, D3DTEXF_POINT);
 	pD3Ddevice->SetSamplerState(1, D3DSAMP_MAGFILTER, D3DTEXF_POINT);

GPU/Directx9/TextureCacheDX9.cpp

@@ -70,14 +70,14 @@ TextureCacheDX9::TextureCacheDX9() : cacheSizeEstimate_(0), secondCacheSizeEstim
 	tmpTexBuf16.resize(1024 * 512);  // 1MB
 	tmpTexBufRearrange.resize(1024 * 512);   // 2MB
 
-	// Aren't these way too big?
-	clutBufConverted_ = (u32 *)AllocateAlignedMemory(4096 * sizeof(u32), 16);  // 16KB
-	clutBufRaw_ = (u32 *)AllocateAlignedMemory(4096 * sizeof(u32), 16);  // 16KB
+	// TODO: Clamp down to 256/1KB?  Need to check mipmapShareClut and clamp loadclut.
+	clutBufConverted_ = (u32 *)AllocateAlignedMemory(1024 * sizeof(u32), 16);  // 4KB
+	clutBufRaw_ = (u32 *)AllocateAlignedMemory(1024 * sizeof(u32), 16);  // 4KB
 
 	// Zap these so that reads from uninitialized parts of the CLUT look the same in
 	// release and debug
-	memset(clutBufConverted_, 0, 4096 * sizeof(u32));
-	memset(clutBufRaw_, 0, 4096 * sizeof(u32));
+	memset(clutBufConverted_, 0, 1024 * sizeof(u32));
+	memset(clutBufRaw_, 0, 1024 * sizeof(u32));
 
 	D3DCAPS9 pCaps;
 	ZeroMemory(&pCaps, sizeof(pCaps));
@@ -865,7 +865,12 @@ void TextureCacheDX9::UpdateCurrentClut() {
 	const u32 clutBaseBytes = clutBase * (clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16));
 	// Technically, these extra bytes weren't loaded, but hopefully it was loaded earlier.
 	// If not, we're going to hash random data, which hopefully doesn't cause a performance issue.
-	const u32 clutExtendedBytes = clutTotalBytes_ + clutBaseBytes;
+	//
+	// TODO: Actually, this seems like a hack.  The game can upload part of a CLUT and reference other data.
+	// clutTotalBytes_ is the last amount uploaded.  We should hash clutMaxBytes_, but this will often hash
+	// unrelated old entries for small palettes.
+	// Adding clutBaseBytes may just be mitigating this for some usage patterns.
+	const u32 clutExtendedBytes = std::min(clutTotalBytes_ + clutBaseBytes, clutMaxBytes_);
 
 	clutHash_ = DoReliableHash32((const char *)clutBufRaw_, clutExtendedBytes, 0xC0108888);
 	clutBuf_ = clutBufRaw_;
@@ -973,11 +978,10 @@ void TextureCacheDX9::SetTextureFramebuffer(TexCacheEntry *entry, VirtualFramebu
 			dxstate.viewport.restore();
 
 			const GEPaletteFormat clutFormat = gstate.getClutPaletteFormat();
-			const u32 clutBase = gstate.getClutIndexStartPos();
 			const u32 bytesPerColor = clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16);
-			const u32 clutExtendedColors = (clutTotalBytes_ / bytesPerColor) + clutBase;
+			const u32 clutTotalColors = clutMaxBytes_ / bytesPerColor;
 
-			TexCacheEntry::Status alphaStatus = CheckAlpha(clutBuf_, getClutDestFormat(gstate.getClutPaletteFormat()), clutExtendedColors, clutExtendedColors, 1);
+			TexCacheEntry::Status alphaStatus = CheckAlpha(clutBuf_, getClutDestFormat(gstate.getClutPaletteFormat()), clutTotalColors, clutTotalColors, 1);
 			gstate_c.textureFullAlpha = alphaStatus == TexCacheEntry::STATUS_ALPHA_FULL;
 			gstate_c.textureSimpleAlpha = alphaStatus == TexCacheEntry::STATUS_ALPHA_SIMPLE;
 		} else {

GPU/GLES/DepalettizeShader.cpp

@@ -152,7 +152,7 @@ GLuint DepalShaderCache::GetClutTexture(const u32 clutID, u32 *rawClut) {
 
 	glTexImage2D(GL_TEXTURE_2D, 0, components, texturePixels, 1, 0, components2, dstFmt, (void *)rawClut);
 
-	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);

GPU/GLES/TextureCache.cpp

@@ -75,14 +75,14 @@ TextureCache::TextureCache() : cacheSizeEstimate_(0), secondCacheSizeEstimate_(0
 	tmpTexBuf16.resize(1024 * 512);  // 1MB
 	tmpTexBufRearrange.resize(1024 * 512);   // 2MB
 
-	// Aren't these way too big?
-	clutBufConverted_ = (u32 *)AllocateAlignedMemory(4096 * sizeof(u32), 16);  // 16KB
-	clutBufRaw_ = (u32 *)AllocateAlignedMemory(4096 * sizeof(u32), 16);  // 16KB
+	// TODO: Clamp down to 256/1KB?  Need to check mipmapShareClut and clamp loadclut.
+	clutBufConverted_ = (u32 *)AllocateAlignedMemory(1024 * sizeof(u32), 16);  // 4KB
+	clutBufRaw_ = (u32 *)AllocateAlignedMemory(1024 * sizeof(u32), 16);  // 4KB
 
 	// Zap these so that reads from uninitialized parts of the CLUT look the same in
 	// release and debug
-	memset(clutBufConverted_, 0, 4096 * sizeof(u32));
-	memset(clutBufRaw_, 0, 4096 * sizeof(u32));
+	memset(clutBufConverted_, 0, 1024 * sizeof(u32));
+	memset(clutBufRaw_, 0, 1024 * sizeof(u32));
 
 	glGetFloatv(GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT, &maxAnisotropyLevel);
 	SetupTextureDecoder();
@@ -992,13 +992,18 @@ void TextureCache::UpdateCurrentClut() {
 	const u32 clutBaseBytes = clutBase * (clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16));
 	// Technically, these extra bytes weren't loaded, but hopefully it was loaded earlier.
 	// If not, we're going to hash random data, which hopefully doesn't cause a performance issue.
-	const u32 clutExtendedBytes = clutTotalBytes_ + clutBaseBytes;
+	//
+	// TODO: Actually, this seems like a hack.  The game can upload part of a CLUT and reference other data.
+	// clutTotalBytes_ is the last amount uploaded.  We should hash clutMaxBytes_, but this will often hash
+	// unrelated old entries for small palettes.
+	// Adding clutBaseBytes may just be mitigating this for some usage patterns.
+	const u32 clutExtendedBytes = std::min(clutTotalBytes_ + clutBaseBytes, clutMaxBytes_);
 
 	clutHash_ = DoReliableHash32((const char *)clutBufRaw_, clutExtendedBytes, 0xC0108888);
 
 	// Avoid a copy when we don't need to convert colors.
 	if (UseBGRA8888() || clutFormat != GE_CMODE_32BIT_ABGR8888) {
-		const int numColors = (clutMaxBytes_ + clutBaseBytes) / (clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16));
+		const int numColors = clutMaxBytes_ / (clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16));
 		ConvertColors(clutBufConverted_, clutBufRaw_, getClutDestFormat(clutFormat), numColors);
 		clutBuf_ = clutBufConverted_;
 	} else {
@@ -1147,11 +1152,10 @@ void TextureCache::SetTextureFramebuffer(TexCacheEntry *entry, VirtualFramebuffe
 			framebufferManager_->RebindFramebuffer();
 
 			const GEPaletteFormat clutFormat = gstate.getClutPaletteFormat();
-			const u32 clutBase = gstate.getClutIndexStartPos();
 			const u32 bytesPerColor = clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16);
-			const u32 clutExtendedColors = (clutTotalBytes_ / bytesPerColor) + clutBase;
+			const u32 clutTotalColors = clutMaxBytes_ / bytesPerColor;
 
-			TexCacheEntry::Status alphaStatus = CheckAlpha(clutBuf_, getClutDestFormat(gstate.getClutPaletteFormat()), clutExtendedColors, clutExtendedColors, 1);
+			TexCacheEntry::Status alphaStatus = CheckAlpha(clutBuf_, getClutDestFormat(gstate.getClutPaletteFormat()), clutTotalColors, clutTotalColors, 1);
 			gstate_c.textureFullAlpha = alphaStatus == TexCacheEntry::STATUS_ALPHA_FULL;
 			gstate_c.textureSimpleAlpha = alphaStatus == TexCacheEntry::STATUS_ALPHA_SIMPLE;
 		} else {

GPU/GPUState.h

@@ -309,7 +309,11 @@ struct GPUgstate
 	int getClutIndexShift() const { return (clutformat >> 2) & 0x1F; }
 	int getClutIndexMask() const { return (clutformat >> 8) & 0xFF; }
 	int getClutIndexStartPos() const { return ((clutformat >> 16) & 0x1F) << 4; }
-	int transformClutIndex(int index) const { return ((index >> getClutIndexShift()) & getClutIndexMask()) | getClutIndexStartPos(); }
+	u32 transformClutIndex(u32 index) const {
+		// We need to wrap any entries beyond the first 1024 bytes.
+		u32 mask = getClutPaletteFormat() == GE_CMODE_32BIT_ABGR8888 ? 0xFF : 0x1FF;
+		return ((index >> getClutIndexShift()) & getClutIndexMask()) | (getClutIndexStartPos() & mask);
+	}
 	bool isClutIndexSimple() const { return (clutformat & ~3) == 0xC500FF00; } // Meaning, no special mask, shift, or start pos.
 	bool isTextureSwizzled() const { return texmode & 1; }
 	bool isClutSharedForMipmaps() const { return (texmode & 0x100) == 0; }