psd: LAB conversion speed improved

Using an approximated pow function improves the conversion speed of LAB images by more than 50% without an appreciable visual difference.

src/imageformats/fastmath_p.h

@@ -0,0 +1,35 @@
+/*
+    Approximated math functions used into conversions.
+
+    SPDX-FileCopyrightText: Edward Kmett
+    SPDX-FileCopyrightText: 2023 Mirco Miranda <mircomir@outlook.com>
+
+    SPDX-License-Identifier: BSD-3-Clause
+*/
+#ifndef FASTMATH_P_H
+#define FASTMATH_P_H
+
+#include <QtGlobal>
+
+/*!
+ * \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

src/imageformats/psd.cpp

@@ -32,6 +32,7 @@
  *   color management engine (e.g. LittleCMS).
  */
 
+#include "fastmath_p.h"
 #include "psd_p.h"
 #include "util_p.h"
 
@@ -50,7 +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
@@ -898,8 +899,9 @@ 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
 }