From ea1245d19121d74b1f8398c4c8a0e66d2bb2f5c7 Mon Sep 17 00:00:00 2001 From: "Jasper St. Pierre" Date: Sun, 10 Aug 2014 15:28:42 -0400 Subject: TextureDecoder: Pass the TLUT address straight into the texture decoder This removes the requirement for the TextureDecoder to have access to global texture memory. --- Source/Core/VideoCommon/TextureDecoder_x64.cpp | 56 +++++++++++++------------- 1 file changed, 28 insertions(+), 28 deletions(-) (limited to 'Source/Core/VideoCommon/TextureDecoder_x64.cpp') diff --git a/Source/Core/VideoCommon/TextureDecoder_x64.cpp b/Source/Core/VideoCommon/TextureDecoder_x64.cpp index 867566032b..9603ec1695 100644 --- a/Source/Core/VideoCommon/TextureDecoder_x64.cpp +++ b/Source/Core/VideoCommon/TextureDecoder_x64.cpp @@ -83,9 +83,9 @@ struct DXTBlock u8 lines[4]; }; -static inline void decodebytesC4_5A3_To_rgba32(u32 *dst, const u8 *src, int tlutaddr) +static inline void decodebytesC4_5A3_To_rgba32(u32 *dst, const u8 *src, const u8* tlut_) { - u16 *tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 4; x++) { u8 val = src[x]; @@ -94,9 +94,9 @@ static inline void decodebytesC4_5A3_To_rgba32(u32 *dst, const u8 *src, int tlut } } -static inline void decodebytesC4IA8_To_RGBA(u32* dst, const u8* src, int tlutaddr) +static inline void decodebytesC4IA8_To_RGBA(u32* dst, const u8* src, const u8* tlut_) { - u16* tlut = (u16*)(texMem+tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 4; x++) { u8 val = src[x]; @@ -105,9 +105,9 @@ static inline void decodebytesC4IA8_To_RGBA(u32* dst, const u8* src, int tlutadd } } -static inline void decodebytesC4RGB565_To_RGBA(u32* dst, const u8* src, int tlutaddr) +static inline void decodebytesC4RGB565_To_RGBA(u32* dst, const u8* src, const u8* tlut_) { - u16* tlut = (u16*)(texMem+tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 4; x++) { u8 val = src[x]; @@ -116,9 +116,9 @@ static inline void decodebytesC4RGB565_To_RGBA(u32* dst, const u8* src, int tlut } } -static inline void decodebytesC8_5A3_To_RGBA32(u32 *dst, const u8 *src, int tlutaddr) +static inline void decodebytesC8_5A3_To_RGBA32(u32 *dst, const u8 *src, const u8* tlut_) { - u16 *tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 8; x++) { u8 val = src[x]; @@ -126,27 +126,27 @@ static inline void decodebytesC8_5A3_To_RGBA32(u32 *dst, const u8 *src, int tlut } } -static inline void decodebytesC8IA8_To_RGBA(u32* dst, const u8* src, int tlutaddr) +static inline void decodebytesC8IA8_To_RGBA(u32* dst, const u8* src, const u8* tlut_) { - u16* tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 8; x++) { *dst++ = decodeIA8Swapped(tlut[src[x]]); } } -static inline void decodebytesC8RGB565_To_RGBA(u32* dst, const u8* src, int tlutaddr) +static inline void decodebytesC8RGB565_To_RGBA(u32* dst, const u8* src, const u8* tlut_) { - u16* tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 8; x++) { *dst++ = decode565RGBA(Common::swap16(tlut[src[x]])); } } -static inline void decodebytesC14X2_5A3_To_RGBA(u32 *dst, const u16 *src, int tlutaddr) +static inline void decodebytesC14X2_5A3_To_RGBA(u32 *dst, const u16 *src, const u8* tlut_) { - u16 *tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 4; x++) { u16 val = Common::swap16(src[x]); @@ -154,9 +154,9 @@ static inline void decodebytesC14X2_5A3_To_RGBA(u32 *dst, const u16 *src, int tl } } -static inline void decodebytesC14X2IA8_To_RGBA(u32* dst, const u16* src, int tlutaddr) +static inline void decodebytesC14X2IA8_To_RGBA(u32* dst, const u16* src, const u8* tlut_) { - u16* tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 4; x++) { u16 val = Common::swap16(src[x]); @@ -164,9 +164,9 @@ static inline void decodebytesC14X2IA8_To_RGBA(u32* dst, const u16* src, int tlu } } -static inline void decodebytesC14X2rgb565_To_RGBA(u32* dst, const u16* src, int tlutaddr) +static inline void decodebytesC14X2rgb565_To_RGBA(u32* dst, const u16* src, const u8* tlut_) { - u16* tlut = (u16*)(texMem + tlutaddr); + const u16* tlut = (u16*) tlut_; for (int x = 0; x < 4; x++) { u16 val = Common::swap16(src[x]); @@ -287,7 +287,7 @@ static inline void SetOpenMPThreadCount(int width, int height) // TODO: complete SSE2 optimization of less often used texture formats. // TODO: refactor algorithms using _mm_loadl_epi64 unaligned loads to prefer 128-bit aligned loads. -PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int height, int texformat, int tlutaddr, TlutFormat tlutfmt) +PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int height, int texformat, const u8* tlut, TlutFormat tlutfmt) { SetOpenMPThreadCount(width, height); @@ -304,7 +304,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 8) for (int x = 0, yStep = (y / 8) * Wsteps8; x < width; x += 8,yStep++) for (int iy = 0, xStep = 8 * yStep; iy < 8; iy++,xStep++) - decodebytesC4_5A3_To_rgba32(dst + (y + iy) * width + x, src + 4 * xStep, tlutaddr); + decodebytesC4_5A3_To_rgba32(dst + (y + iy) * width + x, src + 4 * xStep, tlut); } else if (tlutfmt == GX_TL_IA8) { @@ -312,7 +312,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 8) for (int x = 0, yStep = (y / 8) * Wsteps8; x < width; x += 8,yStep++) for (int iy = 0, xStep = 8 * yStep; iy < 8; iy++,xStep++) - decodebytesC4IA8_To_RGBA(dst + (y + iy) * width + x, src + 4 * xStep, tlutaddr); + decodebytesC4IA8_To_RGBA(dst + (y + iy) * width + x, src + 4 * xStep, tlut); } else if (tlutfmt == GX_TL_RGB565) @@ -321,7 +321,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 8) for (int x = 0, yStep = (y / 8) * Wsteps8; x < width; x += 8,yStep++) for (int iy = 0, xStep = 8 * yStep; iy < 8; iy++,xStep++) - decodebytesC4RGB565_To_RGBA(dst + (y + iy) * width + x, src + 4 * xStep, tlutaddr); + decodebytesC4RGB565_To_RGBA(dst + (y + iy) * width + x, src + 4 * xStep, tlut); } break; case GX_TF_I4: @@ -562,7 +562,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 4) for (int x = 0, yStep = (y / 4) * Wsteps8; x < width; x += 8, yStep++) for (int iy = 0, xStep = 4 * yStep; iy < 4; iy++, xStep++) - decodebytesC8_5A3_To_RGBA32((u32*)dst + (y + iy) * width + x, src + 8 * xStep, tlutaddr); + decodebytesC8_5A3_To_RGBA32((u32*)dst + (y + iy) * width + x, src + 8 * xStep, tlut); } else if (tlutfmt == GX_TL_IA8) { @@ -570,7 +570,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 4) for (int x = 0, yStep = (y / 4) * Wsteps8; x < width; x += 8, yStep++) for (int iy = 0, xStep = 4 * yStep; iy < 4; iy++, xStep++) - decodebytesC8IA8_To_RGBA(dst + (y + iy) * width + x, src + 8 * xStep, tlutaddr); + decodebytesC8IA8_To_RGBA(dst + (y + iy) * width + x, src + 8 * xStep, tlut); } else if (tlutfmt == GX_TL_RGB565) @@ -579,7 +579,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 4) for (int x = 0, yStep = (y / 4) * Wsteps8; x < width; x += 8, yStep++) for (int iy = 0, xStep = 4 * yStep; iy < 4; iy++, xStep++) - decodebytesC8RGB565_To_RGBA(dst + (y + iy) * width + x, src + 8 * xStep, tlutaddr); + decodebytesC8RGB565_To_RGBA(dst + (y + iy) * width + x, src + 8 * xStep, tlut); } break; @@ -675,7 +675,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 4) for (int x = 0, yStep = (y / 4) * Wsteps4; x < width; x += 4, yStep++) for (int iy = 0, xStep = 4 * yStep; iy < 4; iy++, xStep++) - decodebytesC14X2_5A3_To_RGBA(dst + (y + iy) * width + x, (u16*)(src + 8 * xStep), tlutaddr); + decodebytesC14X2_5A3_To_RGBA(dst + (y + iy) * width + x, (u16*)(src + 8 * xStep), tlut); } else if (tlutfmt == GX_TL_IA8) { @@ -683,7 +683,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 4) for (int x = 0, yStep = (y / 4) * Wsteps4; x < width; x += 4, yStep++) for (int iy = 0, xStep = 4 * yStep; iy < 4; iy++, xStep++) - decodebytesC14X2IA8_To_RGBA(dst + (y + iy) * width + x, (u16*)(src + 8 * xStep), tlutaddr); + decodebytesC14X2IA8_To_RGBA(dst + (y + iy) * width + x, (u16*)(src + 8 * xStep), tlut); } else if (tlutfmt == GX_TL_RGB565) { @@ -691,7 +691,7 @@ PC_TexFormat _TexDecoder_DecodeImpl(u32 * dst, const u8 * src, int width, int he for (int y = 0; y < height; y += 4) for (int x = 0, yStep = (y / 4) * Wsteps4; x < width; x += 4, yStep++) for (int iy = 0, xStep = 4 * yStep; iy < 4; iy++, xStep++) - decodebytesC14X2rgb565_To_RGBA(dst + (y + iy) * width + x, (u16*)(src + 8 * xStep), tlutaddr); + decodebytesC14X2rgb565_To_RGBA(dst + (y + iy) * width + x, (u16*)(src + 8 * xStep), tlut); } break; case GX_TF_RGB565: -- cgit v1.2.3