diff --git a/GPU/Common/TextureCacheCommon.cpp b/GPU/Common/TextureCacheCommon.cpp index c8d4bc01d4..e0c2d4667b 100644 --- a/GPU/Common/TextureCacheCommon.cpp +++ b/GPU/Common/TextureCacheCommon.cpp @@ -1556,15 +1556,16 @@ static CheckAlphaResult DecodeDXTBlocks(uint8_t *out, int outPitch, uint32_t tex u32 blockIndex = (y / 4) * (bufw / 4); int blockHeight = std::min(h - y, 4); for (int x = 0; x < minw; x += 4) { + int blockWidth = std::min(minw - x, 4); switch (n) { case 1: - DecodeDXT1Block(dst + outPitch32 * y + x, (const DXT1Block *)src + blockIndex, outPitch32, blockHeight, &alphaSum); + DecodeDXT1Block(dst + outPitch32 * y + x, (const DXT1Block *)src + blockIndex, outPitch32, blockWidth, blockHeight, &alphaSum); break; case 3: - DecodeDXT3Block(dst + outPitch32 * y + x, (const DXT3Block *)src + blockIndex, outPitch32, blockHeight); + DecodeDXT3Block(dst + outPitch32 * y + x, (const DXT3Block *)src + blockIndex, outPitch32, blockWidth, blockHeight); break; case 5: - DecodeDXT5Block(dst + outPitch32 * y + x, (const DXT5Block *)src + blockIndex, outPitch32, blockHeight); + DecodeDXT5Block(dst + outPitch32 * y + x, (const DXT5Block *)src + blockIndex, outPitch32, blockWidth, blockHeight); break; } blockIndex++; diff --git a/GPU/Common/TextureDecoder.cpp b/GPU/Common/TextureDecoder.cpp index 22e45b151b..a12be66cb9 100644 --- a/GPU/Common/TextureDecoder.cpp +++ b/GPU/Common/TextureDecoder.cpp @@ -421,9 +421,9 @@ class DXTDecoder { public: inline void DecodeColors(const DXT1Block *src, bool ignore1bitAlpha); inline void DecodeAlphaDXT5(const DXT5Block *src); - inline void WriteColorsDXT1(u32 *dst, const DXT1Block *src, int pitch, int height); - inline void WriteColorsDXT3(u32 *dst, const DXT3Block *src, int pitch, int height); - inline void WriteColorsDXT5(u32 *dst, const DXT5Block *src, int pitch, int height); + inline void WriteColorsDXT1(u32 *dst, const DXT1Block *src, int pitch, int width, int height); + inline void WriteColorsDXT3(u32 *dst, const DXT3Block *src, int pitch, int width, int height); + inline void WriteColorsDXT5(u32 *dst, const DXT5Block *src, int pitch, int width, int height); bool AnyNonFullAlpha() const { return anyNonFullAlpha_; } @@ -507,11 +507,11 @@ void DXTDecoder::DecodeAlphaDXT5(const DXT5Block *src) { } } -void DXTDecoder::WriteColorsDXT1(u32 *dst, const DXT1Block *src, int pitch, int height) { +void DXTDecoder::WriteColorsDXT1(u32 *dst, const DXT1Block *src, int pitch, int width, int height) { bool anyColor3 = false; for (int y = 0; y < height; y++) { int colordata = src->lines[y]; - for (int x = 0; x < 4; x++) { + for (int x = 0; x < width; x++) { int col = colordata & 3; if (col == 3) { anyColor3 = true; @@ -527,11 +527,11 @@ void DXTDecoder::WriteColorsDXT1(u32 *dst, const DXT1Block *src, int pitch, int } } -void DXTDecoder::WriteColorsDXT3(u32 *dst, const DXT3Block *src, int pitch, int height) { +void DXTDecoder::WriteColorsDXT3(u32 *dst, const DXT3Block *src, int pitch, int width, int height) { for (int y = 0; y < height; y++) { int colordata = src->color.lines[y]; u32 alphadata = src->alphaLines[y]; - for (int x = 0; x < 4; x++) { + for (int x = 0; x < width; x++) { dst[x] = colors_[colordata & 3] | (alphadata << 28); colordata >>= 2; alphadata >>= 4; @@ -540,13 +540,13 @@ void DXTDecoder::WriteColorsDXT3(u32 *dst, const DXT3Block *src, int pitch, int } } -void DXTDecoder::WriteColorsDXT5(u32 *dst, const DXT5Block *src, int pitch, int height) { +void DXTDecoder::WriteColorsDXT5(u32 *dst, const DXT5Block *src, int pitch, int width, int height) { // 48 bits, 3 bit index per pixel, 12 bits per line. u64 alphadata = ((u64)(u16)src->alphadata1 << 32) | (u32)src->alphadata2; for (int y = 0; y < height; y++) { int colordata = src->color.lines[y]; - for (int x = 0; x < 4; x++) { + for (int x = 0; x < width; x++) { dst[x] = colors_[colordata & 3] | (alpha_[alphadata & 7] << 24); colordata >>= 2; alphadata >>= 3; @@ -619,24 +619,24 @@ uint32_t GetDXT5Texel(const DXT5Block *src, int x, int y) { } // This could probably be done faster by decoding two or four blocks at a time with SSE/NEON. -void DecodeDXT1Block(u32 *dst, const DXT1Block *src, int pitch, int height, u32 *alpha) { +void DecodeDXT1Block(u32 *dst, const DXT1Block *src, int pitch, int width, int height, u32 *alpha) { DXTDecoder dxt; dxt.DecodeColors(src, false); - dxt.WriteColorsDXT1(dst, src, pitch, height); + dxt.WriteColorsDXT1(dst, src, pitch, width, height); *alpha &= dxt.AnyNonFullAlpha() ? 0 : 1; } -void DecodeDXT3Block(u32 *dst, const DXT3Block *src, int pitch, int height) { +void DecodeDXT3Block(u32 *dst, const DXT3Block *src, int pitch, int width, int height) { DXTDecoder dxt; dxt.DecodeColors(&src->color, true); - dxt.WriteColorsDXT3(dst, src, pitch, height); + dxt.WriteColorsDXT3(dst, src, pitch, width, height); } -void DecodeDXT5Block(u32 *dst, const DXT5Block *src, int pitch, int height) { +void DecodeDXT5Block(u32 *dst, const DXT5Block *src, int pitch, int width, int height) { DXTDecoder dxt; dxt.DecodeColors(&src->color, true); dxt.DecodeAlphaDXT5(src); - dxt.WriteColorsDXT5(dst, src, pitch, height); + dxt.WriteColorsDXT5(dst, src, pitch, width, height); } #ifdef _M_SSE diff --git a/GPU/Common/TextureDecoder.h b/GPU/Common/TextureDecoder.h index 3df27b4a3b..ad54815ff7 100644 --- a/GPU/Common/TextureDecoder.h +++ b/GPU/Common/TextureDecoder.h @@ -65,9 +65,9 @@ struct DXT5Block { u8 alpha1; u8 alpha2; }; -void DecodeDXT1Block(u32 *dst, const DXT1Block *src, int pitch, int height, u32 *alpha); -void DecodeDXT3Block(u32 *dst, const DXT3Block *src, int pitch, int height); -void DecodeDXT5Block(u32 *dst, const DXT5Block *src, int pitch, int height); +void DecodeDXT1Block(u32 *dst, const DXT1Block *src, int pitch, int width, int height, u32 *alpha); +void DecodeDXT3Block(u32 *dst, const DXT3Block *src, int pitch, int width, int height); +void DecodeDXT5Block(u32 *dst, const DXT5Block *src, int pitch, int width, int height); uint32_t GetDXT1Texel(const DXT1Block *src, int x, int y); uint32_t GetDXT3Texel(const DXT3Block *src, int x, int y);