diff --git a/autotests/read/psd/laba_16bit.png b/autotests/read/psd/laba_16bit.png index d0db793..d3d317c 100644 Binary files a/autotests/read/psd/laba_16bit.png and b/autotests/read/psd/laba_16bit.png differ diff --git a/autotests/read/psd/laba_8bit.png b/autotests/read/psd/laba_8bit.png index 1fd7e71..c691745 100644 Binary files a/autotests/read/psd/laba_8bit.png and b/autotests/read/psd/laba_8bit.png differ diff --git a/src/imageformats/fastmath_p.h b/src/imageformats/fastmath_p.h new file mode 100644 index 0000000..17bd2dd --- /dev/null +++ b/src/imageformats/fastmath_p.h @@ -0,0 +1,35 @@ +/* + Approximated math functions used into conversions. + + SPDX-FileCopyrightText: Edward Kmett + SPDX-FileCopyrightText: 2023 Mirco Miranda + + SPDX-License-Identifier: BSD-3-Clause +*/ +#ifndef FASTMATH_P_H +#define FASTMATH_P_H + +#include + +/*! + * \brief fastPow + * Based on Edward Kmett code released into the public domain. + * See also: https://github.com/ekmett/approximate + */ +inline double fastPow(double x, double y) +{ + union { + double d; + qint32 i[2]; + } u = {x}; +#if Q_BYTE_ORDER == Q_LITTLE_ENDIAN + u.i[1] = qint32(y * (u.i[1] - 1072632447) + 1072632447); + u.i[0] = 0; +#else // never tested + u.i[0] = qint32(y * (u.i[0] - 1072632447) + 1072632447); + u.i[1] = 0; +#endif + return u.d; +} + +#endif // FASTMATH_P_H diff --git a/src/imageformats/psd.cpp b/src/imageformats/psd.cpp index 942a300..d1a2484 100644 --- a/src/imageformats/psd.cpp +++ b/src/imageformats/psd.cpp @@ -3,7 +3,7 @@ SPDX-FileCopyrightText: 2003 Ignacio Castaño SPDX-FileCopyrightText: 2015 Alex Merry - SPDX-FileCopyrightText: 2022 Mirco Miranda + SPDX-FileCopyrightText: 2022-2023 Mirco Miranda SPDX-License-Identifier: LGPL-2.0-or-later */ @@ -21,7 +21,6 @@ /* * Limitations of the current code: * - 32-bit float image are converted to 16-bit integer image. - * NOTE: Qt 6.2 allow 32-bit float images (RGB only) * - Other color spaces cannot directly be read due to lack of QImage support for * color spaces other than RGB (and Grayscale). Where possible, a conversion * to RGB is done: @@ -33,6 +32,7 @@ * color management engine (e.g. LittleCMS). */ +#include "fastmath_p.h" #include "psd_p.h" #include "util_p.h" @@ -51,7 +51,7 @@ typedef quint8 uchar; * This should not be a problem because the Qt's QColorSpace supports the linear * sRgb colorspace. * - * Using linear conversion, the loading speed is improved by 4x. Anyway, if you are using + * Using linear conversion, the loading speed is slightly improved. Anyway, if you are using * an software that discard color info, you should comment it. * * At the time I'm writing (07/2022), Gwenview and Krita supports linear sRgb but KDE @@ -845,6 +845,7 @@ inline void cmykToRgb(uchar *target, qint32 targetChannels, const char *source, auto s = reinterpret_cast(source); auto t = reinterpret_cast(target); auto max = double(std::numeric_limits::max()); + auto invmax = 1.0 / max; // speed improvements by ~10% if (sourceChannels < 4) { qDebug() << "cmykToRgb: image is not a valid CMYK!"; @@ -853,10 +854,10 @@ inline void cmykToRgb(uchar *target, qint32 targetChannels, const char *source, for (qint32 w = 0; w < width; ++w) { auto ps = s + sourceChannels * w; - auto C = 1 - *(ps + 0) / max; - auto M = 1 - *(ps + 1) / max; - auto Y = 1 - *(ps + 2) / max; - auto K = 1 - *(ps + 3) / max; + auto C = 1 - *(ps + 0) * invmax; + auto M = 1 - *(ps + 1) * invmax; + auto Y = 1 - *(ps + 2) * invmax; + auto K = 1 - *(ps + 3) * invmax; auto pt = t + targetChannels * w; *(pt + 0) = T(std::min(max - (C * (1 - K) + K) * max + 0.5, max)); @@ -881,8 +882,9 @@ inline double gammaCorrection(double linear) #ifdef PSD_FAST_LAB_CONVERSION return linear; #else - // NOTE: pow() slow down the performance by a 4 factor :( - return (linear > 0.0031308 ? 1.055 * std::pow(linear, 1.0 / 2.4) - 0.055 : 12.92 * linear); + // Replacing fastPow with std::pow the conversion time is 2/3 times longer: using fastPow + // there are minimal differences in the conversion that are not visually noticeable. + return (linear > 0.0031308 ? 1.055 * fastPow(linear, 1.0 / 2.4) - 0.055 : 12.92 * linear); #endif } @@ -892,6 +894,7 @@ inline void labToRgb(uchar *target, qint32 targetChannels, const char *source, q auto s = reinterpret_cast(source); auto t = reinterpret_cast(target); auto max = double(std::numeric_limits::max()); + auto invmax = 1.0 / max; if (sourceChannels < 3) { qDebug() << "labToRgb: image is not a valid LAB!"; @@ -900,14 +903,14 @@ inline void labToRgb(uchar *target, qint32 targetChannels, const char *source, q for (qint32 w = 0; w < width; ++w) { auto ps = s + sourceChannels * w; - auto L = (*(ps + 0) / max) * 100.0; - auto A = (*(ps + 1) / max) * 255.0 - 128.0; - auto B = (*(ps + 2) / max) * 255.0 - 128.0; + auto L = (*(ps + 0) * invmax) * 100.0; + auto A = (*(ps + 1) * invmax) * 255.0 - 128.0; + auto B = (*(ps + 2) * invmax) * 255.0 - 128.0; // converting LAB to XYZ (D65 illuminant) - auto Y = (L + 16.0) / 116.0; - auto X = A / 500.0 + Y; - auto Z = Y - B / 200.0; + auto Y = (L + 16.0) * (1.0 / 116.0); + auto X = A * (1.0 / 500.0) + Y; + auto Z = Y - B * (1.0 / 200.0); // NOTE: use the constants of the illuminant of the target RGB color space X = finv(X) * 0.9504; // D50: * 0.9642