From 313dd2932f540cbbdcec93ed97a0cc004293962c Mon Sep 17 00:00:00 2001 From: rogerman Date: Fri, 4 Mar 2016 19:32:06 +0000 Subject: [PATCH] Video Filter: - Upgrade xBRZ to v1.4. - Add 6xBRZ pixel scaler. --- desmume/src/filter/filter.h | 1 + desmume/src/filter/videofilter.cpp | 3 +- desmume/src/filter/videofilter.h | 1 + desmume/src/filter/xbrz.cpp | 877 ++++++++++++++--------------- desmume/src/filter/xbrz.h | 20 +- 5 files changed, 451 insertions(+), 451 deletions(-) diff --git a/desmume/src/filter/filter.h b/desmume/src/filter/filter.h index d09ff8a27..03be9e93e 100644 --- a/desmume/src/filter/filter.h +++ b/desmume/src/filter/filter.h @@ -51,3 +51,4 @@ void Render2xBRZ(SSurface Src, SSurface Dst); void Render3xBRZ(SSurface Src, SSurface Dst); void Render4xBRZ(SSurface Src, SSurface Dst); void Render5xBRZ(SSurface Src, SSurface Dst); +void Render6xBRZ(SSurface Src, SSurface Dst); diff --git a/desmume/src/filter/videofilter.cpp b/desmume/src/filter/videofilter.cpp index 12ab1ada4..d140be566 100644 --- a/desmume/src/filter/videofilter.cpp +++ b/desmume/src/filter/videofilter.cpp @@ -50,7 +50,8 @@ const VideoFilterAttributes VideoFilterAttributesList[] = { {VideoFilterTypeID_4xBRZ, "4xBRZ", &Render4xBRZ, 4, 1, 0}, {VideoFilterTypeID_5xBRZ, "5xBRZ", &Render5xBRZ, 5, 1, 0}, {VideoFilterTypeID_HQ3X, "HQ3x", &RenderHQ3X, 3, 1, 0}, - {VideoFilterTypeID_HQ3XS, "HQ3xS", &RenderHQ3XS, 3, 1, 0} }; + {VideoFilterTypeID_HQ3XS, "HQ3xS", &RenderHQ3XS, 3, 1, 0}, + {VideoFilterTypeID_6xBRZ, "6xBRZ", &Render6xBRZ, 6, 1, 0} }; // Parameters for Scanline filter int scanline_filter_a = 0; diff --git a/desmume/src/filter/videofilter.h b/desmume/src/filter/videofilter.h index 1c349f48f..d4b8656db 100644 --- a/desmume/src/filter/videofilter.h +++ b/desmume/src/filter/videofilter.h @@ -72,6 +72,7 @@ enum VideoFilterTypeID VideoFilterTypeID_5xBRZ, VideoFilterTypeID_HQ3X, VideoFilterTypeID_HQ3XS, + VideoFilterTypeID_6xBRZ, VideoFilterTypeIDCount // Make sure this one is always last }; diff --git a/desmume/src/filter/xbrz.cpp b/desmume/src/filter/xbrz.cpp index 162be0c3a..f983bd2e7 100644 --- a/desmume/src/filter/xbrz.cpp +++ b/desmume/src/filter/xbrz.cpp @@ -1,6 +1,6 @@ // **************************************************************************** // * This file is part of the HqMAME project. It is distributed under * -// * GNU General Public License: http://www.gnu.org/licenses/gpl.html * +// * GNU General Public License: http://www.gnu.org/licenses/gpl-3.0 * // * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved * // * * // * Additionally and as a special exception, the author gives permission * @@ -13,6 +13,8 @@ // * do so, delete this exception statement from your version. * // **************************************************************************** +// 2016-03-04 (rogerman): Update to XBRZ 1.4. +// // 2014-11-18 (rogerman): Update to XBRZ 1.1. // // 2014-02-06 (rogerman): Modified for use in DeSmuME by removing C++11 code. @@ -23,6 +25,15 @@ #include #include #include +#include + +#ifdef _MSC_VER + #define FORCE_INLINE __forceinline +#elif defined __GNUC__ + #define FORCE_INLINE __attribute__((always_inline)) inline +#else + #define FORCE_INLINE inline +#endif #define COLOR_MASK_A 0xFF000000 #define COLOR_MASK_R 0x00FF0000 @@ -34,30 +45,57 @@ namespace template inline unsigned char getByte(uint32_t val) { return static_cast((val >> (8 * N)) & 0xff); } -inline unsigned char getAlpha(uint32_t val) { return getByte<3>(val); } -inline unsigned char getRed (uint32_t val) { return getByte<2>(val); } -inline unsigned char getGreen(uint32_t val) { return getByte<1>(val); } -inline unsigned char getBlue (uint32_t val) { return getByte<0>(val); } +inline unsigned char getAlpha(uint32_t pix) { return getByte<3>(pix); } +inline unsigned char getRed (uint32_t pix) { return getByte<2>(pix); } +inline unsigned char getGreen(uint32_t pix) { return getByte<1>(pix); } +inline unsigned char getBlue (uint32_t pix) { return getByte<0>(pix); } +inline uint32_t makePixel( unsigned char r, unsigned char g, unsigned char b) { return (r << 16) | (g << 8) | b; } +inline uint32_t makePixel(unsigned char a, unsigned char r, unsigned char g, unsigned char b) { return (a << 24) | (r << 16) | (g << 8) | b; } -template inline -T abs(T value) +template FORCE_INLINE +unsigned char gradientRGB_calcColor(unsigned char colFront, unsigned char colBack) { - //static_assert(std::is_signed::value, "abs() requires signed types"); - return value < 0 ? -value : value; + return (colFront * M + colBack * (N - M)) / N; +} + +FORCE_INLINE unsigned char gradientARGB_calcColor(unsigned char colFront, unsigned char colBack, unsigned int weightFront, unsigned int weightBack, unsigned int weightSum) +{ + return static_cast((colFront * weightFront + colBack * weightBack) / weightSum); +}; + +template inline +uint32_t gradientRGB(uint32_t pixFront, uint32_t pixBack) //blend front color with opacity M / N over opaque background: http://en.wikipedia.org/wiki/Alpha_compositing#Alpha_blending +{ + //static_assert(0 < M && M < N && N <= 1000, ""); + //auto calcBlend = [](unsigned char colFront, unsigned char colBack) -> unsigned char { return (colFront * M + colBack * (N - M)) / N; }; + + return makePixel(gradientRGB_calcColor(getRed (pixFront), getRed (pixBack)), + gradientRGB_calcColor(getGreen(pixFront), getGreen(pixBack)), + gradientRGB_calcColor(getBlue (pixFront), getBlue (pixBack))); } template inline -void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with opacity M / N +uint32_t gradientARGB(uint32_t pixFront, uint32_t pixBack) //find intermediate color between two colors with alpha channels (=> NO alpha blending!!!) { - //static_assert(0 < M && M < N && N <= 256, "possible overflow of (col & byte1Mask) * M + (dst & byte1Mask) * (N - M)"); - - dst = (COLOR_MASK_B & (((col & COLOR_MASK_B) * M + (dst & COLOR_MASK_B) * (N - M)) / N)) | // - (COLOR_MASK_G & (((col & COLOR_MASK_G) * M + (dst & COLOR_MASK_G) * (N - M)) / N)) | //this works because next higher 8 bits are free - (COLOR_MASK_R & (((col & COLOR_MASK_R) * M + (dst & COLOR_MASK_R) * (N - M)) / N)) | // - (COLOR_MASK_A & (((((col & COLOR_MASK_A) >> 8) * M + ((dst & COLOR_MASK_A) >> 8) * (N - M)) / N) << 8)); //next 8 bits are not free, so shift - //the last row operating on a potential alpha channel costs only ~1% perf => negligible! + //static_assert(0 < M && M < N && N <= 1000, ""); + + const unsigned int weightFront = getAlpha(pixFront) * M; + const unsigned int weightBack = getAlpha(pixBack) * (N - M); + const unsigned int weightSum = weightFront + weightBack; + if (weightSum == 0) + return 0; + /* + auto calcColor = [=](unsigned char colFront, unsigned char colBack) + { + return static_cast((colFront * weightFront + colBack * weightBack) / weightSum); + }; + */ + return makePixel(static_cast(weightSum / N), + gradientARGB_calcColor(getRed (pixFront), getRed (pixBack), weightFront, weightBack, weightSum), + gradientARGB_calcColor(getGreen(pixFront), getGreen(pixBack), weightFront, weightBack, weightSum), + gradientARGB_calcColor(getBlue (pixFront), getBlue (pixBack), weightFront, weightBack, weightSum)); } @@ -73,15 +111,6 @@ void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with // -//inline -//uint32_t alphaBlend2(uint32_t pix1, uint32_t pix2, double alpha) -//{ -// return (redMask & static_cast((pix1 & redMask ) * alpha + (pix2 & redMask ) * (1 - alpha))) | -// (greenMask & static_cast((pix1 & greenMask) * alpha + (pix2 & greenMask) * (1 - alpha))) | -// (blueMask & static_cast((pix1 & blueMask ) * alpha + (pix2 & blueMask ) * (1 - alpha))); -//} - - uint32_t* byteAdvance( uint32_t* ptr, int bytes) { return reinterpret_cast< uint32_t*>(reinterpret_cast< char*>(ptr) + bytes); } const uint32_t* byteAdvance(const uint32_t* ptr, int bytes) { return reinterpret_cast(reinterpret_cast(ptr) + bytes); } @@ -101,16 +130,6 @@ void fillBlock(uint32_t* trg, int pitch, uint32_t col, int blockWidth, int block inline void fillBlock(uint32_t* trg, int pitch, uint32_t col, int n) { fillBlock(trg, pitch, col, n, n); } - -#ifdef _MSC_VER -#define FORCE_INLINE __forceinline -#elif defined __GNUC__ -#define FORCE_INLINE __attribute__((always_inline)) inline -#else -#define FORCE_INLINE inline -#endif - - enum RotationDegree //clock-wise { ROT_0, @@ -164,199 +183,6 @@ template inline T square(T value) { return value * value; } -/* -inline -void rgbtoLuv(uint32_t c, double& L, double& u, double& v) -{ - //http://www.easyrgb.com/index.php?X=MATH&H=02#text2 - double r = getRed (c) / 255.0; - double g = getGreen(c) / 255.0; - double b = getBlue (c) / 255.0; - - if ( r > 0.04045 ) - r = std::pow(( ( r + 0.055 ) / 1.055 ) , 2.4); - else - r /= 12.92; - if ( g > 0.04045 ) - g = std::pow(( ( g + 0.055 ) / 1.055 ) , 2.4); - else - g /= 12.92; - if ( b > 0.04045 ) - b = std::pow(( ( b + 0.055 ) / 1.055 ) , 2.4); - else - b /= 12.92; - - r *= 100; - g *= 100; - b *= 100; - - double x = 0.4124564 * r + 0.3575761 * g + 0.1804375 * b; - double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b; - double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b; - //--------------------- - double var_U = 4 * x / ( x + 15 * y + 3 * z ); - double var_V = 9 * y / ( x + 15 * y + 3 * z ); - double var_Y = y / 100; - - if ( var_Y > 0.008856 ) var_Y = std::pow(var_Y , 1.0/3 ); - else var_Y = 7.787 * var_Y + 16.0 / 116; - - const double ref_X = 95.047; //Observer= 2°, Illuminant= D65 - const double ref_Y = 100.000; - const double ref_Z = 108.883; - - const double ref_U = ( 4 * ref_X ) / ( ref_X + ( 15 * ref_Y ) + ( 3 * ref_Z ) ); - const double ref_V = ( 9 * ref_Y ) / ( ref_X + ( 15 * ref_Y ) + ( 3 * ref_Z ) ); - - L = ( 116 * var_Y ) - 16; - u = 13 * L * ( var_U - ref_U ); - v = 13 * L * ( var_V - ref_V ); -} -*/ - -inline -void rgbtoLab(uint32_t c, unsigned char& L, signed char& A, signed char& B) -{ - //code: http://www.easyrgb.com/index.php?X=MATH - //test: http://www.workwithcolor.com/color-converter-01.htm - //------RGB to XYZ------ - double r = getRed (c) / 255.0; - double g = getGreen(c) / 255.0; - double b = getBlue (c) / 255.0; - - r = r > 0.04045 ? std::pow(( r + 0.055 ) / 1.055, 2.4) : r / 12.92; - r = g > 0.04045 ? std::pow(( g + 0.055 ) / 1.055, 2.4) : g / 12.92; - r = b > 0.04045 ? std::pow(( b + 0.055 ) / 1.055, 2.4) : b / 12.92; - - r *= 100; - g *= 100; - b *= 100; - - double x = 0.4124564 * r + 0.3575761 * g + 0.1804375 * b; - double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b; - double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b; - //------XYZ to Lab------ - const double refX = 95.047; // - const double refY = 100.000; //Observer= 2°, Illuminant= D65 - const double refZ = 108.883; // - double var_X = x / refX; - double var_Y = y / refY; - double var_Z = z / refZ; - - var_X = var_X > 0.008856 ? std::pow(var_X, 1.0 / 3) : 7.787 * var_X + 4.0 / 29; - var_Y = var_Y > 0.008856 ? std::pow(var_Y, 1.0 / 3) : 7.787 * var_Y + 4.0 / 29; - var_Z = var_Z > 0.008856 ? std::pow(var_Z, 1.0 / 3) : 7.787 * var_Z + 4.0 / 29; - - L = static_cast(116 * var_Y - 16); - A = static_cast< signed char>(500 * (var_X - var_Y)); - B = static_cast< signed char>(200 * (var_Y - var_Z)); -}; - - -inline -double distLAB(uint32_t pix1, uint32_t pix2) -{ - unsigned char L1 = 0; //[0, 100] - signed char a1 = 0; //[-128, 127] - signed char b1 = 0; //[-128, 127] - rgbtoLab(pix1, L1, a1, b1); - - unsigned char L2 = 0; - signed char a2 = 0; - signed char b2 = 0; - rgbtoLab(pix2, L2, a2, b2); - - //----------------------------- - //http://www.easyrgb.com/index.php?X=DELT - - //Delta E/CIE76 - return std::sqrt(square(1.0 * L1 - L2) + - square(1.0 * a1 - a2) + - square(1.0 * b1 - b2)); -} - - -/* -inline -void rgbtoHsl(uint32_t c, double& h, double& s, double& l) -{ - //http://www.easyrgb.com/index.php?X=MATH&H=18#text18 - const int r = getRed (c); - const int g = getGreen(c); - const int b = getBlue (c); - - const int varMin = numeric::min(r, g, b); - const int varMax = numeric::max(r, g, b); - const int delMax = varMax - varMin; - - l = (varMax + varMin) / 2.0 / 255.0; - - if (delMax == 0) //gray, no chroma... - { - h = 0; - s = 0; - } - else - { - s = l < 0.5 ? - delMax / (1.0 * varMax + varMin) : - delMax / (2.0 * 255 - varMax - varMin); - - double delR = ((varMax - r) / 6.0 + delMax / 2.0) / delMax; - double delG = ((varMax - g) / 6.0 + delMax / 2.0) / delMax; - double delB = ((varMax - b) / 6.0 + delMax / 2.0) / delMax; - - if (r == varMax) - h = delB - delG; - else if (g == varMax) - h = 1 / 3.0 + delR - delB; - else if (b == varMax) - h = 2 / 3.0 + delG - delR; - - if (h < 0) - h += 1; - if (h > 1) - h -= 1; - } -} - -inline -double distHSL(uint32_t pix1, uint32_t pix2, double lightningWeight) -{ - double h1 = 0; - double s1 = 0; - double l1 = 0; - rgbtoHsl(pix1, h1, s1, l1); - double h2 = 0; - double s2 = 0; - double l2 = 0; - rgbtoHsl(pix2, h2, s2, l2); - - //HSL is in cylindric coordinatates where L represents height, S radius, H angle, - //however we interpret the cylinder as a bi-conic solid with top/bottom radius 0, middle radius 1 - assert(0 <= h1 && h1 <= 1); - assert(0 <= h2 && h2 <= 1); - - double r1 = l1 < 0.5 ? - l1 * 2 : - 2 - l1 * 2; - - double x1 = r1 * s1 * std::cos(h1 * 2 * numeric::pi); - double y1 = r1 * s1 * std::sin(h1 * 2 * numeric::pi); - double z1 = l1; - - double r2 = l2 < 0.5 ? - l2 * 2 : - 2 - l2 * 2; - - double x2 = r2 * s2 * std::cos(h2 * 2 * numeric::pi); - double y2 = r2 * s2 * std::sin(h2 * 2 * numeric::pi); - double z2 = l2; - - return 255 * std::sqrt(square(x1 - x2) + square(y1 - y2) + square(lightningWeight * (z1 - z2))); -} -*/ - inline double distRGB(uint32_t pix1, uint32_t pix2) @@ -370,19 +196,6 @@ double distRGB(uint32_t pix1, uint32_t pix2) } -inline -double distNonLinearRGB(uint32_t pix1, uint32_t pix2) -{ - //non-linear rgb: http://www.compuphase.com/cmetric.htm - const double r_diff = static_cast(getRed (pix1)) - getRed (pix2); - const double g_diff = static_cast(getGreen(pix1)) - getGreen(pix2); - const double b_diff = static_cast(getBlue (pix1)) - getBlue (pix2); - - const double r_avg = (static_cast(getRed(pix1)) + getRed(pix2)) / 2; - return std::sqrt((2 + r_avg / 255) * square(r_diff) + 4 * square(g_diff) + (2 + (255 - r_avg) / 255) * square(b_diff)); -} - - inline double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight) { @@ -410,56 +223,60 @@ double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight) } -inline -double distYCbCrAlpha(uint32_t pix1, uint32_t pix2, double lumaWeight) +struct DistYCbCrBuffer //30% perf boost compared to distYCbCr()! { - const double a1 = getAlpha(pix1) / 255.0 ; - const double a2 = getAlpha(pix2) / 255.0 ; - - /* - Requirements for a color distance handling alpha channel: with a1, a2 in [0, 1] - - 1. if a1 = a2, distance should be: a1 * distYCbCr() - 2. if a1 = 0, distance should be: a2 * distYCbCr(black, white) = a2 * 255 - 3. if a1 = 1, distance should be: 255 * (1 - a2) + a2 * distYCbCr() - */ - return std::min(a1, a2) * distYCbCr(pix1, pix2, lumaWeight) + 255 * abs(a1 - a2); -} - - -inline -double distYUV(uint32_t pix1, uint32_t pix2, double luminanceWeight) -{ - //perf: it's not worthwhile to buffer the YUV-conversion, the direct code is faster by ~ 6% - //since RGB -> YUV conversion is essentially a matrix multiplication, we can calculate the RGB diff before the conversion (distributive property) - const double r_diff = static_cast(getRed (pix1)) - getRed (pix2); - const double g_diff = static_cast(getGreen(pix1)) - getGreen(pix2); - const double b_diff = static_cast(getBlue (pix1)) - getBlue (pix2); - - //http://en.wikipedia.org/wiki/YUV#Conversion_to.2Ffrom_RGB - const double w_b = 0.114; - const double w_r = 0.299; - const double w_g = 1 - w_r - w_b; - - const double u_max = 0.436; - const double v_max = 0.615; - - const double scale_u = u_max / (1 - w_b); - const double scale_v = v_max / (1 - w_r); - - double y = w_r * r_diff + w_g * g_diff + w_b * b_diff;//value range: 255 * [-1, 1] - double u = scale_u * (b_diff - y); //value range: 255 * 2 * u_max * [-1, 1] - double v = scale_v * (r_diff - y); //value range: 255 * 2 * v_max * [-1, 1] - -#ifndef NDEBUG - const double eps = 0.5; +public: + static double dist(uint32_t pix1, uint32_t pix2) + { +#if defined _MSC_VER && _MSC_VER < 1900 +#error function scope static initialization is not yet thread-safe! #endif - assert(abs(y) <= 255 + eps); - assert(abs(u) <= 255 * 2 * u_max + eps); - assert(abs(v) <= 255 * 2 * v_max + eps); + static const DistYCbCrBuffer inst; + return inst.distImpl(pix1, pix2); + } - return std::sqrt(square(luminanceWeight * y) + square(u) + square(v)); -} +private: + DistYCbCrBuffer() : buffer(256 * 256 * 256) + { + for (uint32_t i = 0; i < 256 * 256 * 256; ++i) //startup time: 114 ms on Intel Core i5 (four cores) + { + const int r_diff = getByte<2>(i) * 2 - 255; + const int g_diff = getByte<1>(i) * 2 - 255; + const int b_diff = getByte<0>(i) * 2 - 255; + + const double k_b = 0.0593; //ITU-R BT.2020 conversion + const double k_r = 0.2627; // + const double k_g = 1.0 - k_b - k_r; + + const double scale_b = 0.5 / (1.0 - k_b); + const double scale_r = 0.5 / (1.0 - k_r); + + const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr! + const double c_b = scale_b * (b_diff - y); + const double c_r = scale_r * (r_diff - y); + + buffer[i] = static_cast(std::sqrt(square(y) + square(c_b) + square(c_r))); + } + } + + double distImpl(uint32_t pix1, uint32_t pix2) const + { + //if (pix1 == pix2) -> 8% perf degradation! + // return 0; + //if (pix1 > pix2) + // std::swap(pix1, pix2); -> 30% perf degradation!!! + + const int r_diff = static_cast(getRed (pix1)) - getRed (pix2); + const int g_diff = static_cast(getGreen(pix1)) - getGreen(pix2); + const int b_diff = static_cast(getBlue (pix1)) - getBlue (pix2); + + return buffer[(((r_diff + 255) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte + (((g_diff + 255) / 2) << 8) | + (( b_diff + 255) / 2)]; + } + + std::vector buffer; //consumes 64 MB memory; using double is only 2% faster, but takes 128 MB +}; enum BlendType @@ -492,7 +309,7 @@ input kernel area naming convention: ----------------- | A | B | C | D | ----|---|---|---| -| E | F | G | H | //evalute the four corners between F, G, J, K +| E | F | G | H | //evaluate the four corners between F, G, J, K ----|---|---|---| //input pixel is at position F | I | J | K | L | ----|---|---|---| @@ -511,11 +328,11 @@ BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg) ker.g == ker.k)) return result; - //auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_); }; + //auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight); }; const int weight = 4; - double jg = ColorDistance::dist(ker.i, ker.f, cfg.luminanceWeight_) + ColorDistance::dist(ker.f, ker.c, cfg.luminanceWeight_) + ColorDistance::dist(ker.n, ker.k, cfg.luminanceWeight_) + ColorDistance::dist(ker.k, ker.h, cfg.luminanceWeight_) + weight * ColorDistance::dist(ker.j, ker.g, cfg.luminanceWeight_); - double fk = ColorDistance::dist(ker.e, ker.j, cfg.luminanceWeight_) + ColorDistance::dist(ker.j, ker.o, cfg.luminanceWeight_) + ColorDistance::dist(ker.b, ker.g, cfg.luminanceWeight_) + ColorDistance::dist(ker.g, ker.l, cfg.luminanceWeight_) + weight * ColorDistance::dist(ker.f, ker.k, cfg.luminanceWeight_); + double jg = ColorDistance::dist(ker.i, ker.f, cfg.luminanceWeight) + ColorDistance::dist(ker.f, ker.c, cfg.luminanceWeight) + ColorDistance::dist(ker.n, ker.k, cfg.luminanceWeight) + ColorDistance::dist(ker.k, ker.h, cfg.luminanceWeight) + weight * ColorDistance::dist(ker.j, ker.g, cfg.luminanceWeight); + double fk = ColorDistance::dist(ker.e, ker.j, cfg.luminanceWeight) + ColorDistance::dist(ker.j, ker.o, cfg.luminanceWeight) + ColorDistance::dist(ker.b, ker.g, cfg.luminanceWeight) + ColorDistance::dist(ker.g, ker.l, cfg.luminanceWeight) + weight * ColorDistance::dist(ker.f, ker.k, cfg.luminanceWeight); if (jg < fk) //test sample: 70% of values max(jg, fk) / min(jg, fk) are between 1.1 and 3.7 with median being 1.8 { @@ -591,18 +408,19 @@ template <> inline unsigned char rotateBlendInfo(unsigned char b) { ret template <> inline unsigned char rotateBlendInfo(unsigned char b) { return ((b << 4) | (b >> 4)) & 0xff; } template <> inline unsigned char rotateBlendInfo(unsigned char b) { return ((b << 6) | (b >> 2)) & 0xff; } - +/* #ifndef NDEBUG -int debugPixelX = -1; -int debugPixelY = 84; -bool breakIntoDebugger = false; + int debugPixelX = -1; + int debugPixelY = 12; + __declspec(thread) bool breakIntoDebugger = false; #endif +*/ template FORCE_INLINE double sPixEQ(uint32_t col1, uint32_t col2, const xbrz::ScalerCfg& cfg) { - return ColorDistance::dist(col1, col2, cfg.luminanceWeight_) < cfg.equalColorTolerance_; + return ColorDistance::dist(col1, col2, cfg.luminanceWeight) < cfg.equalColorTolerance; } template @@ -623,7 +441,7 @@ const bool sPixDoLineBlend(const Kernel_3x3& ker, const char blend, const xbrz:: return true; //make sure there is no second blending in an adjacent rotation for this pixel: handles insular pixels, mario eyes - if (getTopR(blend) != BLEND_NONE && !sPixEQ(e, g, cfg)) //but support double-blending for 90° corners + if (getTopR(blend) != BLEND_NONE && !sPixEQ(e, g, cfg)) //but support double-blending for 90¡ corners return false; if (getBottomL(blend) != BLEND_NONE && !sPixEQ(e, c, cfg)) return false; @@ -648,7 +466,7 @@ const bool sPixDoLineBlend(const Kernel_3x3& ker, const char blend, const xbrz:: return true; } - + /* input kernel area naming convention: ------------- @@ -661,7 +479,7 @@ input kernel area naming convention: */ template FORCE_INLINE //perf: quite worth it! -void scalePixel(const Kernel_3x3& ker, +void blendPixel(const Kernel_3x3& ker, uint32_t* target, int trgWidth, unsigned char blendInfo, //result of preprocessing all four corners of pixel "e" const xbrz::ScalerCfg& cfg) @@ -684,9 +502,10 @@ void scalePixel(const Kernel_3x3& ker, const unsigned char blend = rotateBlendInfo(blendInfo); if (getBottomR(blend) >= BLEND_NORMAL) - {/* - auto eq = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_) < cfg.equalColorTolerance_; }; - auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_); }; + { +/* + auto eq = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight) < cfg.equalColorTolerance; }; + auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight); }; const bool doLineBlend = [&]() -> bool { @@ -694,26 +513,26 @@ void scalePixel(const Kernel_3x3& ker, return true; //make sure there is no second blending in an adjacent rotation for this pixel: handles insular pixels, mario eyes - if (getTopR(blend) != BLEND_NONE && !eq(e, g)) //but support double-blending for 90° corners + if (getTopR(blend) != BLEND_NONE && !eq(e, g)) //but support double-blending for 90° corners return false; if (getBottomL(blend) != BLEND_NONE && !eq(e, c)) return false; //no full blending for L-shapes; blend corner only (handles "mario mushroom eyes") - if (eq(g, h) && eq(h , i) && eq(i, f) && eq(f, c) && !eq(e, i)) + if (!eq(e, i) && eq(g, h) && eq(h , i) && eq(i, f) && eq(f, c)) return false; return true; }(); */ - const uint32_t px = ( ColorDistance::dist(e, f, cfg.luminanceWeight_) <= ColorDistance::dist(e, h, cfg.luminanceWeight_) ) ? f : h; //choose most similar color + const uint32_t px = ColorDistance::dist(e, f, cfg.luminanceWeight) <= ColorDistance::dist(e, h, cfg.luminanceWeight) ? f : h; //choose most similar color OutputMatrix out(target, trgWidth); if (sPixDoLineBlend(ker, blend, cfg)) { - const double fg = ColorDistance::dist(f, g, cfg.luminanceWeight_); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9 - const double hc = ColorDistance::dist(h, c, cfg.luminanceWeight_); // + const double fg = ColorDistance::dist(f, g, cfg.luminanceWeight); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9 + const double hc = ColorDistance::dist(h, c, cfg.luminanceWeight); // const bool haveShallowLine = cfg.steepDirectionThreshold * fg <= hc && e != g && d != g; const bool haveSteepLine = cfg.steepDirectionThreshold * hc <= fg && e != c && b != c; @@ -766,7 +585,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, std::fill(preProcBuffer, preProcBuffer + bufferSize, 0); //static_assert(BLEND_NONE == 0, ""); - //initialize preprocessing buffer for first row: detect upper left and right corner blending + //initialize preprocessing buffer for first row of current stripe: detect upper left and right corner blending //this cannot be optimized for adjacent processing stripes; we must not allow for a memory race condition! if (yFirst > 0) { @@ -815,7 +634,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, */ setTopR(preProcBuffer[x], res.blend_j); - if (x + 1 < srcWidth) + if (x + 1 < bufferSize) setTopL(preProcBuffer[x + 1], res.blend_k); } } @@ -834,39 +653,42 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, for (int x = 0; x < srcWidth; ++x, out += Scaler::scale) { +/* #ifndef NDEBUG breakIntoDebugger = debugPixelX == x && debugPixelY == y; #endif +*/ //all those bounds checks have only insignificant impact on performance! const int x_m1 = std::max(x - 1, 0); //perf: prefer array indexing to additional pointers! const int x_p1 = std::min(x + 1, srcWidth - 1); const int x_p2 = std::min(x + 2, srcWidth - 1); + Kernel_4x4 ker4 = {}; //perf: initialization is negligible + + ker4.a = s_m1[x_m1]; //read sequentially from memory as far as possible + ker4.b = s_m1[x]; + ker4.c = s_m1[x_p1]; + ker4.d = s_m1[x_p2]; + + ker4.e = s_0[x_m1]; + ker4.f = s_0[x]; + ker4.g = s_0[x_p1]; + ker4.h = s_0[x_p2]; + + ker4.i = s_p1[x_m1]; + ker4.j = s_p1[x]; + ker4.k = s_p1[x_p1]; + ker4.l = s_p1[x_p2]; + + ker4.m = s_p2[x_m1]; + ker4.n = s_p2[x]; + ker4.o = s_p2[x_p1]; + ker4.p = s_p2[x_p2]; + //evaluate the four corners on bottom-right of current pixel unsigned char blend_xy = 0; //for current (x, y) position { - Kernel_4x4 ker = {}; //perf: initialization is negligible - ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible - ker.b = s_m1[x]; - ker.c = s_m1[x_p1]; - ker.d = s_m1[x_p2]; - - ker.e = s_0[x_m1]; - ker.f = s_0[x]; - ker.g = s_0[x_p1]; - ker.h = s_0[x_p2]; - - ker.i = s_p1[x_m1]; - ker.j = s_p1[x]; - ker.k = s_p1[x_p1]; - ker.l = s_p1[x_p2]; - - ker.m = s_p2[x_m1]; - ker.n = s_p2[x]; - ker.o = s_p2[x_p1]; - ker.p = s_p2[x_p2]; - - const BlendResult res = preProcessCorners(ker, cfg); + const BlendResult res = preProcessCorners(ker4, cfg); /* preprocessing blend result: --------- @@ -884,34 +706,34 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, blend_xy1 = 0; setTopL(blend_xy1, res.blend_k); //set 1st known corner for (x + 1, y + 1) and buffer for use on next column - if (x + 1 < srcWidth) //set 3rd known corner for (x + 1, y) + if (x + 1 < bufferSize) //set 3rd known corner for (x + 1, y) setBottomL(preProcBuffer[x + 1], res.blend_g); } //fill block of size scale * scale with the given color - fillBlock(out, trgWidth * sizeof(uint32_t), s_0[x], Scaler::scale); //place *after* preprocessing step, to not overwrite the results while processing the the last pixel! + fillBlock(out, trgWidth * sizeof(uint32_t), ker4.f, Scaler::scale); //place *after* preprocessing step, to not overwrite the results while processing the the last pixel! //blend four corners of current pixel - if (blendingNeeded(blend_xy)) //good 20% perf-improvement + if (blendingNeeded(blend_xy)) //good 5% perf-improvement { - Kernel_3x3 ker = {}; //perf: initialization is negligible + Kernel_3x3 ker3 = {}; //perf: initialization is negligible - ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible - ker.b = s_m1[x]; - ker.c = s_m1[x_p1]; + ker3.a = ker4.a; + ker3.b = ker4.b; + ker3.c = ker4.c; - ker.d = s_0[x_m1]; - ker.e = s_0[x]; - ker.f = s_0[x_p1]; + ker3.d = ker4.e; + ker3.e = ker4.f; + ker3.f = ker4.g; - ker.g = s_p1[x_m1]; - ker.h = s_p1[x]; - ker.i = s_p1[x_p1]; + ker3.g = ker4.i; + ker3.h = ker4.j; + ker3.i = ker4.k; - scalePixel(ker, out, trgWidth, blend_xy, cfg); - scalePixel(ker, out, trgWidth, blend_xy, cfg); - scalePixel(ker, out, trgWidth, blend_xy, cfg); - scalePixel(ker, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); } } } @@ -919,112 +741,127 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, //------------------------------------------------------------------------------------ -struct Scaler2x +template +struct Scaler2x : public ColorGradient { static const int scale = 2; + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + template static void blendLineShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); } template static void blendLineSteep(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col); - alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); } template static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<1, 0>(), col); - alphaBlend<1, 4>(out.template ref<0, 1>(), col); - alphaBlend<5, 6>(out.template ref<1, 1>(), col); //[!] fixes 7/8 used in xBR + alphaGrad<1, 4>(out.template ref<1, 0>(), col); + alphaGrad<1, 4>(out.template ref<0, 1>(), col); + alphaGrad<5, 6>(out.template ref<1, 1>(), col); //[!] fixes 7/8 used in xBR } template static void blendLineDiagonal(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 2>(out.template ref<1, 1>(), col); + alphaGrad<1, 2>(out.template ref<1, 1>(), col); } template static void blendCorner(uint32_t col, OutputMatrix& out) { //model a round corner - alphaBlend<21, 100>(out.template ref<1, 1>(), col); //exact: 1 - pi/4 = 0.2146018366 + alphaGrad<21, 100>(out.template ref<1, 1>(), col); //exact: 1 - pi/4 = 0.2146018366 } }; -struct Scaler3x +template +struct Scaler3x : public ColorGradient { static const int scale = 3; + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + template static void blendLineShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); out.template ref() = col; } template static void blendLineSteep(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col); - alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); - alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); out.template ref<2, scale - 1>() = col; } template static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<2, 0>(), col); - alphaBlend<1, 4>(out.template ref<0, 2>(), col); - alphaBlend<3, 4>(out.template ref<2, 1>(), col); - alphaBlend<3, 4>(out.template ref<1, 2>(), col); + alphaGrad<1, 4>(out.template ref<2, 0>(), col); + alphaGrad<1, 4>(out.template ref<0, 2>(), col); + alphaGrad<3, 4>(out.template ref<2, 1>(), col); + alphaGrad<3, 4>(out.template ref<1, 2>(), col); out.template ref<2, 2>() = col; } template static void blendLineDiagonal(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 8>(out.template ref<1, 2>(), col); - alphaBlend<1, 8>(out.template ref<2, 1>(), col); - alphaBlend<7, 8>(out.template ref<2, 2>(), col); + alphaGrad<1, 8>(out.template ref<1, 2>(), col); //conflict with other rotations for this odd scale + alphaGrad<1, 8>(out.template ref<2, 1>(), col); + alphaGrad<7, 8>(out.template ref<2, 2>(), col); // } template static void blendCorner(uint32_t col, OutputMatrix& out) { //model a round corner - alphaBlend<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598 - //alphaBlend<14, 1000>(out.template ref<2, 1>(), col); //0.01413008627 -> negligible - //alphaBlend<14, 1000>(out.template ref<1, 2>(), col); //0.01413008627 + alphaGrad<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598 + //alphaGrad<7, 256>(out.template ref<2, 1>(), col); //0.02826017254 -> negligible + avoid conflicts with other rotations for this odd scale + //alphaGrad<7, 256>(out.template ref<1, 2>(), col); //0.02826017254 } }; -struct Scaler4x +template +struct Scaler4x : public ColorGradient { static const int scale = 4; + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + template static void blendLineShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); out.template ref() = col; out.template ref() = col; @@ -1033,11 +870,11 @@ struct Scaler4x template static void blendLineSteep(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col); - alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); - alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col); - alphaBlend<3, 4>(out.template ref<3, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); out.template ref<2, scale - 1>() = col; out.template ref<3, scale - 1>() = col; @@ -1046,19 +883,23 @@ struct Scaler4x template static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<3, 4>(out.template ref<3, 1>(), col); - alphaBlend<3, 4>(out.template ref<1, 3>(), col); - alphaBlend<1, 4>(out.template ref<3, 0>(), col); - alphaBlend<1, 4>(out.template ref<0, 3>(), col); - alphaBlend<1, 3>(out.template ref<2, 2>(), col); //[!] fixes 1/4 used in xBR - out.template ref<3, 3>() = out.template ref<3, 2>() = out.template ref<2, 3>() = col; + alphaGrad<3, 4>(out.template ref<3, 1>(), col); + alphaGrad<3, 4>(out.template ref<1, 3>(), col); + alphaGrad<1, 4>(out.template ref<3, 0>(), col); + alphaGrad<1, 4>(out.template ref<0, 3>(), col); + + alphaGrad<1, 3>(out.template ref<2, 2>(), col); //[!] fixes 1/4 used in xBR + + out.template ref<3, 3>() = col; + out.template ref<3, 2>() = col; + out.template ref<2, 3>() = col; } template static void blendLineDiagonal(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 2>(out.template ref(), col); - alphaBlend<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); out.template ref() = col; } @@ -1066,26 +907,31 @@ struct Scaler4x static void blendCorner(uint32_t col, OutputMatrix& out) { //model a round corner - alphaBlend<68, 100>(out.template ref<3, 3>(), col); //exact: 0.6848532563 - alphaBlend< 9, 100>(out.template ref<3, 2>(), col); //0.08677704501 - alphaBlend< 9, 100>(out.template ref<2, 3>(), col); //0.08677704501 + alphaGrad<68, 100>(out.template ref<3, 3>(), col); //exact: 0.6848532563 + alphaGrad< 9, 100>(out.template ref<3, 2>(), col); //0.08677704501 + alphaGrad< 9, 100>(out.template ref<2, 3>(), col); //0.08677704501 } }; -struct Scaler5x +template +struct Scaler5x : public ColorGradient { static const int scale = 5; + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + template static void blendLineShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); out.template ref() = col; out.template ref() = col; @@ -1096,12 +942,12 @@ struct Scaler5x template static void blendLineSteep(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col); - alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col); - alphaBlend<1, 4>(out.template ref<4, scale - 3>(), col); + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<1, 4>(out.template ref<4, scale - 3>(), col); - alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col); - alphaBlend<3, 4>(out.template ref<3, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); out.template ref<2, scale - 1>() = col; out.template ref<3, scale - 1>() = col; @@ -1112,34 +958,33 @@ struct Scaler5x template static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col); - alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col); - alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<1, 4>(out.template ref(), col); - alphaBlend<3, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + alphaGrad<2, 3>(out.template ref<3, 3>(), col); out.template ref<2, scale - 1>() = col; out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; out.template ref() = col; out.template ref() = col; - - out.template ref<4, scale - 1>() = col; - - alphaBlend<2, 3>(out.template ref<3, 3>(), col); } template static void blendLineDiagonal(uint32_t col, OutputMatrix& out) { - alphaBlend<1, 8>(out.template ref(), col); - alphaBlend<1, 8>(out.template ref(), col); - alphaBlend<1, 8>(out.template ref(), col); + alphaGrad<1, 8>(out.template ref(), col); //conflict with other rotations for this odd scale + alphaGrad<1, 8>(out.template ref(), col); + alphaGrad<1, 8>(out.template ref(), col); // - alphaBlend<7, 8>(out.template ref<4, 3>(), col); - alphaBlend<7, 8>(out.template ref<3, 4>(), col); + alphaGrad<7, 8>(out.template ref<4, 3>(), col); + alphaGrad<7, 8>(out.template ref<3, 4>(), col); out.template ref<4, 4>() = col; } @@ -1148,11 +993,110 @@ struct Scaler5x static void blendCorner(uint32_t col, OutputMatrix& out) { //model a round corner - alphaBlend<86, 100>(out.template ref<4, 4>(), col); //exact: 0.8631434088 - alphaBlend<23, 100>(out.template ref<4, 3>(), col); //0.2306749731 - alphaBlend<23, 100>(out.template ref<3, 4>(), col); //0.2306749731 - //alphaBlend<8, 1000>(out.template ref<4, 2>(), col); //0.008384061834 -> negligible - //alphaBlend<8, 1000>(out.template ref<2, 4>(), col); //0.008384061834 + alphaGrad<86, 100>(out.template ref<4, 4>(), col); //exact: 0.8631434088 + alphaGrad<23, 100>(out.template ref<4, 3>(), col); //0.2306749731 + alphaGrad<23, 100>(out.template ref<3, 4>(), col); //0.2306749731 + //alphaGrad<1, 64>(out.template ref<4, 2>(), col); //0.01676812367 -> negligible + avoid conflicts with other rotations for this odd scale + //alphaGrad<1, 64>(out.template ref<2, 4>(), col); //0.01676812367 + } +}; + + +template +struct Scaler6x : public ColorGradient +{ + static const int scale = 6; + + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + + template + static void blendLineShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineSteep(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<1, 4>(out.template ref<4, scale - 3>(), col); + + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<5, scale - 3>(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; + out.template ref<5, scale - 1>() = col; + + out.template ref<4, scale - 2>() = col; + out.template ref<5, scale - 2>() = col; + } + + template + static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); + + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; + out.template ref<5, scale - 1>() = col; + + out.template ref<4, scale - 2>() = col; + out.template ref<5, scale - 2>() = col; + + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineDiagonal(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); + + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendCorner(uint32_t col, OutputMatrix& out) + { + //model a round corner + alphaGrad<97, 100>(out.template ref<5, 5>(), col); //exact: 0.9711013910 + alphaGrad<42, 100>(out.template ref<4, 5>(), col); //0.4236372243 + alphaGrad<42, 100>(out.template ref<5, 4>(), col); //0.4236372243 + alphaGrad< 6, 100>(out.template ref<5, 3>(), col); //0.05652034508 + alphaGrad< 6, 100>(out.template ref<3, 5>(), col); //0.05652034508 } }; @@ -1162,9 +1106,11 @@ struct ColorDistanceRGB { static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight) { - if (pix1 == pix2) //about 8% perf boost - return 0; - return distYCbCr(pix1, pix2, luminanceWeight); + return DistYCbCrBuffer::dist(pix1, pix2); + + //if (pix1 == pix2) //about 4% perf boost + // return 0; + //return distYCbCr(pix1, pix2, luminanceWeight); } }; @@ -1172,9 +1118,44 @@ struct ColorDistanceARGB { static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight) { - if (pix1 == pix2) - return 0; - return distYCbCrAlpha(pix1, pix2, luminanceWeight); + const double a1 = getAlpha(pix1) / 255.0 ; + const double a2 = getAlpha(pix2) / 255.0 ; + /* + Requirements for a color distance handling alpha channel: with a1, a2 in [0, 1] + + 1. if a1 = a2, distance should be: a1 * distYCbCr() + 2. if a1 = 0, distance should be: a2 * distYCbCr(black, white) = a2 * 255 + 3. if a1 = 1, ??? maybe: 255 * (1 - a2) + a2 * distYCbCr() + */ + + //return std::min(a1, a2) * DistYCbCrBuffer::dist(pix1, pix2) + 255 * abs(a1 - a2); + //=> following code is 15% faster: + const double d = DistYCbCrBuffer::dist(pix1, pix2); + if (a1 < a2) + return a1 * d + 255 * (a2 - a1); + else + return a2 * d + 255 * (a1 - a2); + + //alternative? return std::sqrt(a1 * a2 * square(DistYCbCrBuffer::dist(pix1, pix2)) + square(255 * (a1 - a2))); + } +}; + + +struct ColorGradientRGB +{ + template + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) + { + pixBack = gradientRGB(pixFront, pixBack); + } +}; + +struct ColorGradientARGB +{ + template + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) + { + pixBack = gradientARGB(pixFront, pixBack); } }; } @@ -1188,26 +1169,33 @@ void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth switch (factor) { case 2: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); case 3: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); case 4: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); case 5: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 6: + return scaleImage, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); } + break; + case ColorFormatRGB: switch (factor) { case 2: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); case 3: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); case 4: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); case 5: - return scaleImage(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + return scaleImage, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 6: + return scaleImage, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); } + break; } assert(false); } @@ -1318,3 +1306,8 @@ void Render5xBRZ(SSurface Src, SSurface Dst) { xbrz::scale(5, (const uint32_t *)Src.Surface, (uint32_t *)Dst.Surface, Src.Width, Src.Height, xbrz::ColorFormatRGB); } + +void Render6xBRZ(SSurface Src, SSurface Dst) +{ + xbrz::scale(6, (const uint32_t *)Src.Surface, (uint32_t *)Dst.Surface, Src.Width, Src.Height, xbrz::ColorFormatRGB); +} diff --git a/desmume/src/filter/xbrz.h b/desmume/src/filter/xbrz.h index f9d4e4df6..a92d14dbb 100644 --- a/desmume/src/filter/xbrz.h +++ b/desmume/src/filter/xbrz.h @@ -1,6 +1,6 @@ // **************************************************************************** // * This file is part of the HqMAME project. It is distributed under * -// * GNU General Public License: http://www.gnu.org/licenses/gpl.html * +// * GNU General Public License: http://www.gnu.org/licenses/gpl-3.0 * // * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved * // * * // * Additionally and as a special exception, the author gives permission * @@ -13,6 +13,8 @@ // * do so, delete this exception statement from your version. * // **************************************************************************** +// 2016-03-04 (rogerman): Update to XBRZ 1.4. +// // 2014-11-18 (rogerman): Update to XBRZ 1.1. // // 2014-02-06 (rogerman): Modified for use in DeSmuME by removing C++11 code. @@ -40,16 +42,18 @@ namespace xbrz using a modified approach of xBR: http://board.byuu.org/viewtopic.php?f=10&t=2248 - new rule set preserving small image features +- highly optimized for performance - support alpha channel - support multithreading - support 64-bit architectures - support processing image slices +- support scaling up to 6xBRZ */ enum ColorFormat //from high bits -> low bits, 8 bit per channel { - ColorFormatARGB, //including alpha channel, BGRA byte order on little-endian machines ColorFormatRGB, //8 bit for each red, green, blue, upper 8 bits unused + ColorFormatARGB, //including alpha channel, BGRA byte order on little-endian machines }; /* @@ -61,25 +65,25 @@ enum ColorFormat //from high bits -> low bits, 8 bit per channel in the target image data if you are using multiple threads for processing each enlarged slice! THREAD-SAFETY: - parts of the same image may be scaled by multiple threads as long as the [yFirst, yLast) ranges do not overlap! - - there is a minor inefficiency for the first row of a slice, so avoid processing single rows only + - there is a minor inefficiency for the first row of a slice, so avoid processing single rows only; suggestion: process 8-16 rows at least */ struct ScalerCfg { ScalerCfg() : - luminanceWeight_(1), - equalColorTolerance_(30), + luminanceWeight(1), + equalColorTolerance(30), dominantDirectionThreshold(3.6), steepDirectionThreshold(2.2), newTestAttribute_(0) {} - double luminanceWeight_; - double equalColorTolerance_; + double luminanceWeight; + double equalColorTolerance; double dominantDirectionThreshold; double steepDirectionThreshold; double newTestAttribute_; //unused; test new parameters }; -void scale(size_t factor, //valid range: 2 - 5 +void scale(size_t factor, //valid range: 2 - 6 const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, ColorFormat colFmt, const ScalerCfg& cfg = ScalerCfg(),