Update to xBRZ 1.6

filter/xbrz.cpp

@@ -1,13 +1,14 @@
 // ****************************************************************************
-// * This file is part of the HqMAME project. It is distributed under         *
+// * This file is part of the xBRZ project. It is distributed under           *
-// * GNU General Public License: http://www.gnu.org/licenses/gpl-3.0          *
+// * GNU General Public License: https://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     *
-// * to link the code of this program with the MAME library (or with modified *
+// * to link the code of this program with the following libraries            *
-// * versions of MAME that use the same license as MAME), and distribute      *
+// * (or with modified versions that use the same licenses), and distribute   *
-// * linked combinations including the two. You must obey the GNU General     *
+// * linked combinations including the two: MAME, FreeFileSync, Snes9x        *
-// * Public License in all respects for all of the code used other than MAME. *
+// * You must obey the GNU General Public License in all respects for all of  *
+// * the code used other than MAME, FreeFileSync, Snes9x.                     *
 // * If you modify this file, you may extend this exception to your version   *
 // * of the file, but you are not obligated to do so. If you do not wish to   *
 // * do so, delete this exception statement from your version.                *
@@ -15,27 +16,16 @@
 
 #include "xbrz.h"
 #include <cassert>
-#include <algorithm>
 #include <vector>
+#include <algorithm>
+#include <cmath> //std::sqrt
+#include "xbrz_tools.h"
+
+using namespace xbrz;
 
-#ifndef WIN32
-#include <cmath>
-#endif
 
 namespace
 {
-template <uint32_t N> inline
-unsigned char getByte(uint32_t val) { return static_cast<unsigned char>((val >> (8 * N)) & 0xff); }
-
-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 <unsigned int M, unsigned int N> 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
 {
@@ -84,26 +74,6 @@ uint32_t gradientARGB(uint32_t pixFront, uint32_t pixBack) //find intermediate c
 //
 
 
-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<const uint32_t*>(reinterpret_cast<const char*>(ptr) + bytes); }
-
-
-//fill block  with the given color
-inline
-void fillBlock(uint32_t* trg, int pitch, uint32_t col, int blockWidth, int blockHeight)
-{
-    //for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
-    //    std::fill(trg, trg + blockWidth, col);
-
-    for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
-        for (int x = 0; x < blockWidth; ++x)
-            trg[x] = col;
-}
-
-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__
@@ -166,7 +136,7 @@ template <class T> inline
 T square(T value) { return value * value; }
 
 
-
+#if 0
 inline
 double distRGB(uint32_t pix1, uint32_t pix2)
 {
@@ -177,6 +147,7 @@ double distRGB(uint32_t pix1, uint32_t pix2)
     //euklidean RGB distance
     return std::sqrt(square(r_diff) + square(g_diff) + square(b_diff));
 }
+#endif
 
 
 inline
@@ -206,26 +177,20 @@ double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight)
 }
 
 
-struct DistYCbCrBuffer //30% perf boost compared to distYCbCr()!
+inline
+double distYCbCrBuffered(uint32_t pix1, uint32_t pix2)
 {
-public:
+    //30% perf boost compared to plain distYCbCr()!
-    static double dist(uint32_t pix1, uint32_t pix2)
+    //consumes 64 MB memory; using double is only 2% faster, but takes 128 MB
+    static const std::vector<float> diffToDist = []
     {
-#if defined _MSC_VER && _MSC_VER < 1900
+        std::vector<float> tmp;
-#error function scope static initialization is not yet thread-safe!
-#endif
-        static const DistYCbCrBuffer inst;
-        return inst.distImpl(pix1, pix2);
-    }
 
-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 r_diff = getByte<2>(i) * 2 - 0xFF;
-            const int g_diff = getByte<1>(i) * 2 - 255;
+            const int g_diff = getByte<1>(i) * 2 - 0xFF;
-            const int b_diff = getByte<0>(i) * 2 - 255;
+            const int b_diff = getByte<0>(i) * 2 - 0xFF;
 
             const double k_b = 0.0593; //ITU-R BT.2020 conversion
             const double k_r = 0.2627; //
@@ -238,28 +203,31 @@ private:
             const double c_b = scale_b * (b_diff - y);
             const double c_r = scale_r * (r_diff - y);
 
-            buffer[i] = static_cast<float>(std::sqrt(square(y) + square(c_b) + square(c_r)));
+            tmp.push_back(static_cast<float>(std::sqrt(square(y) + square(c_b) + square(c_r))));
-        }
         }
+        return tmp;
+    }();
 
-    double distImpl(uint32_t pix1, uint32_t pix2) const
-    {
     //if (pix1 == pix2) -> 8% perf degradation!
     //    return 0;
-        //if (pix1 > pix2)
+    //if (pix1 < pix2)
     //    std::swap(pix1, pix2); -> 30% perf degradation!!!
-
+#if 1
     const int r_diff = static_cast<int>(getRed  (pix1)) - getRed  (pix2);
     const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
     const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
 
-        return buffer[(((r_diff + 255) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte
+    return diffToDist[(((r_diff + 0xFF) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte
-                      (((g_diff + 255) / 2) <<  8) |
+                                        (((g_diff + 0xFF) / 2) <<  8) |
-                      (( b_diff + 255) / 2)];
+                                        (( b_diff + 0xFF) / 2)];
-    }
+#else //not noticeably faster:
+    const int r_diff_tmp = ((pix1 & 0xFF0000) + 0xFF0000 - (pix2 & 0xFF0000)) / 2;
+    const int g_diff_tmp = ((pix1 & 0x00FF00) + 0x00FF00 - (pix2 & 0x00FF00)) / 2; //slightly reduce precision (division by 2) to squeeze value into single byte
+    const int b_diff_tmp = ((pix1 & 0x0000FF) + 0x0000FF - (pix2 & 0x0000FF)) / 2;
 
-    std::vector<float> buffer; //consumes 64 MB memory; using double is only 2% faster, but takes 128 MB
+    return diffToDist[(r_diff_tmp & 0xFF0000) | (g_diff_tmp & 0x00FF00) | (b_diff_tmp & 0x0000FF)];
-};
+#endif
+}
 
 
 enum BlendType
@@ -373,7 +341,7 @@ DEF_GETTER(g, a) DEF_GETTER(h, d) DEF_GETTER(i,	g)
 
 
 //compress four blend types into a single byte
-inline BlendType getTopL   (unsigned char b) { return static_cast<BlendType>(0x3 & b); }
+//inline BlendType getTopL   (unsigned char b) { return static_cast<BlendType>(0x3 & b); }
 inline BlendType getTopR   (unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 2)); }
 inline BlendType getBottomR(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 4)); }
 inline BlendType getBottomL(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 6)); }
@@ -446,7 +414,7 @@ void blendPixel(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;
@@ -515,7 +483,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
     //"sizeof(uint32_t) * srcWidth * (yLast - yFirst)" bytes without risk of accidental overwriting before accessing
     const int bufferSize = srcWidth;
     unsigned char* preProcBuffer = reinterpret_cast<unsigned char*>(trg + yLast * Scaler::scale * trgWidth) - bufferSize;
-    std::fill(preProcBuffer, preProcBuffer + bufferSize, 0);
+    std::fill(preProcBuffer, preProcBuffer + bufferSize, '\0');
     static_assert(BLEND_NONE == 0, "");
 
     //initialize preprocessing buffer for first row of current stripe: detect upper left and right corner blending
@@ -644,7 +612,8 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
             }
 
             //fill block of size scale * scale with the given color
-            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!
+            fillBlock(out, trgWidth * sizeof(uint32_t), ker4.f, Scaler::scale, 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 5% perf-improvement
@@ -1039,7 +1008,7 @@ struct ColorDistanceRGB
 {
     static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
     {
-        return DistYCbCrBuffer::dist(pix1, pix2);
+        return distYCbCrBuffered(pix1, pix2);
 
         //if (pix1 == pix2) //about 4% perf boost
         //    return 0;
@@ -1061,15 +1030,31 @@ struct ColorDistanceARGB
             3. if a1 = 1,  ??? maybe: 255 * (1 - a2) + a2 * distYCbCr()
         */
 
-        //return std::min(a1, a2) * DistYCbCrBuffer::dist(pix1, pix2) + 255 * abs(a1 - a2);
+        //return std::min(a1, a2) * distYCbCrBuffered(pix1, pix2) + 255 * abs(a1 - a2);
         //=> following code is 15% faster:
-        const double d = DistYCbCrBuffer::dist(pix1, pix2);
+        const double d = distYCbCrBuffered(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)));
+        //alternative? return std::sqrt(a1 * a2 * square(distYCbCrBuffered(pix1, pix2)) + square(255 * (a1 - a2)));
+    }
+};
+
+
+struct ColorDistanceUnbufferedARGB
+{
+    static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
+    {
+        const double a1 = getAlpha(pix1) / 255.0 ;
+        const double a2 = getAlpha(pix2) / 255.0 ;
+
+        const double d = distYCbCr(pix1, pix2, luminanceWeight);
+        if (a1 < a2)
+            return a1 * d + 255 * (a2 - a1);
+        else
+            return a2 * d + 255 * (a1 - a2);
     }
 };
 
@@ -1096,8 +1081,25 @@ struct ColorGradientARGB
 
 void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, ColorFormat colFmt, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
 {
+    static_assert(SCALE_FACTOR_MAX == 6, "");
     switch (colFmt)
     {
+        case ColorFormat::RGB:
+            switch (factor)
+            {
+                case 2:
+                    return scaleImage<Scaler2x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                case 3:
+                    return scaleImage<Scaler3x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                case 4:
+                    return scaleImage<Scaler4x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                case 5:
+                    return scaleImage<Scaler5x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                case 6:
+                    return scaleImage<Scaler6x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+            }
+            break;
+
         case ColorFormat::ARGB:
             switch (factor)
             {
@@ -1114,19 +1116,19 @@ void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth
             }
             break;
 
-        case ColorFormat::RGB:
+        case ColorFormat::ARGB_UNBUFFERED:
             switch (factor)
             {
                 case 2:
-                    return scaleImage<Scaler2x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                    return scaleImage<Scaler2x<ColorGradientARGB>, ColorDistanceUnbufferedARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
                 case 3:
-                    return scaleImage<Scaler3x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                    return scaleImage<Scaler3x<ColorGradientARGB>, ColorDistanceUnbufferedARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
                 case 4:
-                    return scaleImage<Scaler4x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                    return scaleImage<Scaler4x<ColorGradientARGB>, ColorDistanceUnbufferedARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
                 case 5:
-                    return scaleImage<Scaler5x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                    return scaleImage<Scaler5x<ColorGradientARGB>, ColorDistanceUnbufferedARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
                 case 6:
-                    return scaleImage<Scaler6x<ColorGradientRGB>, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
+                    return scaleImage<Scaler6x<ColorGradientARGB>, ColorDistanceUnbufferedARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
             }
             break;
     }
@@ -1138,84 +1140,133 @@ bool xbrz::equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, doub
 {
     switch (colFmt)
     {
-        case ColorFormat::ARGB:
-            return ColorDistanceARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
-
         case ColorFormat::RGB:
             return ColorDistanceRGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
+        case ColorFormat::ARGB:
+            return ColorDistanceARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
+        case ColorFormat::ARGB_UNBUFFERED:
+            return ColorDistanceUnbufferedARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
     }
     assert(false);
     return false;
 }
 
 
-void xbrz::nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight, int srcPitch,
+void xbrz::bilinearScale(const uint32_t* src, int srcWidth, int srcHeight,
-                                uint32_t* trg, int trgWidth, int trgHeight, int trgPitch,
+                         /**/  uint32_t* trg, int trgWidth, int trgHeight)
-                                SliceType st, int yFirst, int yLast)
 {
-    if (srcPitch < srcWidth * static_cast<int>(sizeof(uint32_t))  ||
+    bilinearScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
-        trgPitch < trgWidth * static_cast<int>(sizeof(uint32_t)))
+                  trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
-    {
+    0, trgHeight, [](uint32_t pix) { return pix; });
-        assert(false);
-        return;
 }
 
-    switch (st)
-    {
-        case NN_SCALE_SLICE_SOURCE:
-            //nearest-neighbor (going over source image - fast for upscaling, since source is read only once
-            yFirst = std::max(yFirst, 0);
-            yLast  = std::min(yLast, srcHeight);
-            if (yFirst >= yLast || trgWidth <= 0 || trgHeight <= 0) return;
 
-            for (int y = yFirst; y < yLast; ++y)
+void xbrz::nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight,
+                                /**/  uint32_t* trg, int trgWidth, int trgHeight)
 {
-                //mathematically: ySrc = floor(srcHeight * yTrg / trgHeight)
+    nearestNeighborScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
-                // => search for integers in: [ySrc, ySrc + 1) * trgHeight / srcHeight
+                         trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
+    0, trgHeight, [](uint32_t pix) { return pix; });
+}
 
-                //keep within for loop to support MT input slices!
-                const int yTrg_first = ( y      * trgHeight + srcHeight - 1) / srcHeight; //=ceil(y * trgHeight / srcHeight)
-                const int yTrg_last  = ((y + 1) * trgHeight + srcHeight - 1) / srcHeight; //=ceil(((y + 1) * trgHeight) / srcHeight)
-                const int blockHeight = yTrg_last - yTrg_first;
 
-                if (blockHeight > 0)
+#if 0
+//#include <ppl.h>
+void bilinearScaleCpu(const uint32_t* src, int srcWidth, int srcHeight,
+                      /**/  uint32_t* trg, int trgWidth, int trgHeight)
 {
-                    const uint32_t* srcLine = byteAdvance(src, y * srcPitch);
+    const int TASK_GRANULARITY = 16;
-                    uint32_t* trgLine  = byteAdvance(trg, yTrg_first * trgPitch);
-                    int xTrg_first = 0;
 
-                    for (int x = 0; x < srcWidth; ++x)
+    concurrency::task_group tg;
-                    {
-                        int xTrg_last = ((x + 1) * trgWidth + srcWidth - 1) / srcWidth;
-                        const int blockWidth = xTrg_last - xTrg_first;
-                        if (blockWidth > 0)
-                        {
-                            xTrg_first = xTrg_last;
-                            fillBlock(trgLine, trgPitch, srcLine[x], blockWidth, blockHeight);
-                            trgLine += blockWidth;
-                        }
-                    }
-                }
-            }
-            break;
 
-        case NN_SCALE_SLICE_TARGET:
+    for (int i = 0; i < trgHeight; i += TASK_GRANULARITY)
-            //nearest-neighbor (going over target image - slow for upscaling, since source is read multiple times missing out on cache! Fast for similar image sizes!)
+        tg.run([=]
-            yFirst = std::max(yFirst, 0);
+    {
-            yLast  = std::min(yLast, trgHeight);
+        const int iLast = std::min(i + TASK_GRANULARITY, trgHeight);
-            if (yFirst >= yLast || srcHeight <= 0 || srcWidth <= 0) return;
+        xbrz::bilinearScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
+                            trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
+        i, iLast, [](uint32_t pix) { return pix; });
+    });
+    tg.wait();
+}
 
-            for (int y = yFirst; y < yLast; ++y)
+
+//Perf: AMP vs CPU: merely ~10% shorter runtime (scaling 1280x800 -> 1920x1080)
+//#include <amp.h>
+void bilinearScaleAmp(const uint32_t* src, int srcWidth, int srcHeight, //throw concurrency::runtime_exception
+                      /**/  uint32_t* trg, int trgWidth, int trgHeight)
 {
-                uint32_t* trgLine = byteAdvance(trg, y * trgPitch);
+    //C++ AMP reference:       https://msdn.microsoft.com/en-us/library/hh289390.aspx
-                const int ySrc = srcHeight * y / trgHeight;
+    //introduction to C++ AMP: https://msdn.microsoft.com/en-us/magazine/hh882446.aspx
-                const uint32_t* srcLine = byteAdvance(src, ySrc * srcPitch);
+    using namespace concurrency;
-                for (int x = 0; x < trgWidth; ++x)
+    //TODO: pitch
+
+    if (srcHeight <= 0 || srcWidth <= 0) return;
+
+    const float scaleX = static_cast<float>(trgWidth ) / srcWidth;
+    const float scaleY = static_cast<float>(trgHeight) / srcHeight;
+
+    array_view<const uint32_t, 2> srcView(srcHeight, srcWidth, src);
+    array_view<      uint32_t, 2> trgView(trgHeight, trgWidth, trg);
+    trgView.discard_data();
+
+    parallel_for_each(trgView.extent, [=](index<2> idx) restrict(amp) //throw ?
     {
-                    const int xSrc = srcWidth * x / trgWidth;
+        const int y = idx[0];
-                    trgLine[x] = srcLine[xSrc];
+        const int x = idx[1];
-                }
+        //Perf notes:
-            }
+        //    -> float-based calculation is (almost 2x) faster than double!
-            break;
+        //    -> no noticeable improvement via tiling: https://msdn.microsoft.com/en-us/magazine/hh882447.aspx
-    }
+        //    -> no noticeable improvement with restrict(amp,cpu)
+        //    -> iterating over y-axis only is significantly slower!
+        //    -> pre-calculating x,y-dependent variables in a buffer + array_view<> is ~ 20 % slower!
+        const int y1 = srcHeight * y / trgHeight;
+        int y2 = y1 + 1;
+        if (y2 == srcHeight) --y2;
+
+        const float yy1 = y / scaleY - y1;
+        const float y2y = 1 - yy1;
+        //-------------------------------------
+        const int x1 = srcWidth * x / trgWidth;
+        int x2 = x1 + 1;
+        if (x2 == srcWidth) --x2;
+
+        const float xx1 = x / scaleX - x1;
+        const float x2x = 1 - xx1;
+        //-------------------------------------
+        const float x2xy2y = x2x * y2y;
+        const float xx1y2y = xx1 * y2y;
+        const float x2xyy1 = x2x * yy1;
+        const float xx1yy1 = xx1 * yy1;
+
+        auto interpolate = [=](int offset)
+        {
+            /*
+                https://en.wikipedia.org/wiki/Bilinear_interpolation
+                (c11(x2 - x) + c21(x - x1)) * (y2 - y ) +
+                (c12(x2 - x) + c22(x - x1)) * (y  - y1)
+            */
+            const auto c11 = (srcView(y1, x1) >> (8 * offset)) & 0xff;
+            const auto c21 = (srcView(y1, x2) >> (8 * offset)) & 0xff;
+            const auto c12 = (srcView(y2, x1) >> (8 * offset)) & 0xff;
+            const auto c22 = (srcView(y2, x2) >> (8 * offset)) & 0xff;
+
+            return c11 * x2xy2y + c21 * xx1y2y +
+                   c12 * x2xyy1 + c22 * xx1yy1;
+        };
+
+        const float bi = interpolate(0);
+        const float gi = interpolate(1);
+        const float ri = interpolate(2);
+        const float ai = interpolate(3);
+
+        const auto b = static_cast<uint32_t>(bi + 0.5f);
+        const auto g = static_cast<uint32_t>(gi + 0.5f);
+        const auto r = static_cast<uint32_t>(ri + 0.5f);
+        const auto a = static_cast<uint32_t>(ai + 0.5f);
+
+        trgView(y, x) = (a << 24) | (r << 16) | (g << 8) | b;
+    });
+    trgView.synchronize(); //throw ?
 }
+#endif

filter/xbrz.h

@@ -1,13 +1,14 @@
 // ****************************************************************************
-// * This file is part of the HqMAME project. It is distributed under         *
+// * This file is part of the xBRZ project. It is distributed under           *
-// * GNU General Public License: http://www.gnu.org/licenses/gpl-3.0          *
+// * GNU General Public License: https://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     *
-// * to link the code of this program with the MAME library (or with modified *
+// * to link the code of this program with the following libraries            *
-// * versions of MAME that use the same license as MAME), and distribute      *
+// * (or with modified versions that use the same licenses), and distribute   *
-// * linked combinations including the two. You must obey the GNU General     *
+// * linked combinations including the two: MAME, FreeFileSync, Snes9x        *
-// * Public License in all respects for all of the code used other than MAME. *
+// * You must obey the GNU General Public License in all respects for all of  *
+// * the code used other than MAME, FreeFileSync, Snes9x.                     *
 // * If you modify this file, you may extend this exception to your version   *
 // * of the file, but you are not obligated to do so. If you do not wish to   *
 // * do so, delete this exception statement from your version.                *
@@ -20,7 +21,7 @@
 //#include <cstddef> //size_t
 //#include <cstdint> //uint32_t
 #include <limits>
-#include "xbrz-config.h"
+#include "xbrz_config.h"
 
 namespace xbrz
 {
@@ -43,53 +44,37 @@ enum class ColorFormat //from high bits -> low bits, 8 bit per channel
 {
     RGB,  //8 bit for each red, green, blue, upper 8 bits unused
     ARGB, //including alpha channel, BGRA byte order on little-endian machines
+    ARGB_UNBUFFERED, //like ARGB, but without the one-time buffer creation overhead (ca. 100 - 300 ms) at the expense of a slightly slower scaling time
 };
 
+const int SCALE_FACTOR_MAX = 6;
+
 /*
 -> map source (srcWidth * srcHeight) to target (scale * width x scale * height) image, optionally processing a half-open slice of rows [yFirst, yLast) only
 -> support for source/target pitch in bytes!
 -> if your emulator changes only a few image slices during each cycle (e.g. DOSBox) then there's no need to run xBRZ on the complete image:
    Just make sure you enlarge the source image slice by 2 rows on top and 2 on bottom (this is the additional range the xBRZ algorithm is using during analysis)
-   Caveat: If there are multiple changed slices, make sure they do not overlap after adding these additional rows in order to avoid a memory race condition
+   CAVEAT: If there are multiple changed slices, make sure they do not overlap after adding these additional rows in order to avoid a memory race condition
    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; suggestion: process 8-16 rows at least
+               - there is a minor inefficiency for the first row of a slice, so avoid processing single rows only; suggestion: process at least 8-16 rows
 */
-void scale(size_t factor, //valid range: 2 - 6
+void scale(size_t factor, //valid range: 2 - SCALE_FACTOR_MAX
            const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
            ColorFormat colFmt,
            const ScalerCfg& cfg = ScalerCfg(),
            int yFirst = 0, int yLast = std::numeric_limits<int>::max()); //slice of source image
 
-void nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight,
+void bilinearScale(const uint32_t* src, int srcWidth, int srcHeight,
-                          uint32_t* trg, int trgWidth, int trgHeight);
+                   /**/  uint32_t* trg, int trgWidth, int trgHeight);
+
+void nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight,
+                          /**/  uint32_t* trg, int trgWidth, int trgHeight);
 
-enum SliceType
-{
-    NN_SCALE_SLICE_SOURCE,
-    NN_SCALE_SLICE_TARGET,
-};
-void nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight, int srcPitch, //pitch in bytes!
-                          uint32_t* trg, int trgWidth, int trgHeight, int trgPitch,
-                          SliceType st, int yFirst, int yLast);
 
 //parameter tuning
 bool equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance);
-
-
-
-
-
-//########################### implementation ###########################
-inline
-void nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight,
-                          uint32_t* trg, int trgWidth, int trgHeight)
-{
-    nearestNeighborScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
-                         trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
-                         NN_SCALE_SLICE_TARGET, 0, trgHeight);
-}
 }
 
 #endif

filter/xbrz_config.h

@@ -1,13 +1,14 @@
 // ****************************************************************************
-// * This file is part of the HqMAME project. It is distributed under         *
+// * This file is part of the xBRZ project. It is distributed under           *
-// * GNU General Public License: http://www.gnu.org/licenses/gpl-3.0          *
+// * GNU General Public License: https://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     *
-// * to link the code of this program with the MAME library (or with modified *
+// * to link the code of this program with the following libraries            *
-// * versions of MAME that use the same license as MAME), and distribute      *
+// * (or with modified versions that use the same licenses), and distribute   *
-// * linked combinations including the two. You must obey the GNU General     *
+// * linked combinations including the two: MAME, FreeFileSync, Snes9x        *
-// * Public License in all respects for all of the code used other than MAME. *
+// * You must obey the GNU General Public License in all respects for all of  *
+// * the code used other than MAME, FreeFileSync, Snes9x.                     *
 // * If you modify this file, you may extend this exception to your version   *
 // * of the file, but you are not obligated to do so. If you do not wish to   *
 // * do so, delete this exception statement from your version.                *

filter/xbrz_tools.h

@@ -0,0 +1,268 @@
+// ****************************************************************************
+// * This file is part of the xBRZ project. It is distributed under           *
+// * GNU General Public License: https://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     *
+// * to link the code of this program with the following libraries            *
+// * (or with modified versions that use the same licenses), and distribute   *
+// * linked combinations including the two: MAME, FreeFileSync, Snes9x        *
+// * You must obey the GNU General Public License in all respects for all of  *
+// * the code used other than MAME, FreeFileSync, Snes9x.                     *
+// * If you modify this file, you may extend this exception to your version   *
+// * of the file, but you are not obligated to do so. If you do not wish to   *
+// * do so, delete this exception statement from your version.                *
+// ****************************************************************************
+
+#ifndef XBRZ_TOOLS_H_825480175091875
+#define XBRZ_TOOLS_H_825480175091875
+
+#include <cassert>
+#include <algorithm>
+#include <type_traits>
+
+
+namespace xbrz
+{
+template <uint32_t N> inline
+unsigned char getByte(uint32_t val) { return static_cast<unsigned char>((val >> (8 * N)) & 0xff); }
+
+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 a, unsigned char r, unsigned char g, unsigned char b) { return (a << 24) | (r << 16) | (g << 8) | b; }
+inline uint32_t makePixel(                 unsigned char r, unsigned char g, unsigned char b) { return             (r << 16) | (g << 8) | b; }
+
+inline uint32_t rgb555to888(uint16_t pix) { return ((pix & 0x7C00) << 9) | ((pix & 0x03E0) << 6) | ((pix & 0x001F) << 3); }
+inline uint32_t rgb565to888(uint16_t pix) { return ((pix & 0xF800) << 8) | ((pix & 0x07E0) << 5) | ((pix & 0x001F) << 3); }
+
+inline uint16_t rgb888to555(uint32_t pix) { return static_cast<uint16_t>(((pix & 0xF80000) >> 9) | ((pix & 0x00F800) >> 6) | ((pix & 0x0000F8) >> 3)); }
+inline uint16_t rgb888to565(uint32_t pix) { return static_cast<uint16_t>(((pix & 0xF80000) >> 8) | ((pix & 0x00FC00) >> 5) | ((pix & 0x0000F8) >> 3)); }
+
+
+template <class Pix> inline
+Pix* byteAdvance(Pix* ptr, int bytes)
+{
+    using PixNonConst = typename std::remove_cv<Pix>::type;
+    using PixByte     = typename std::conditional<std::is_same<Pix, PixNonConst>::value, char, const char>::type;
+
+    static_assert(std::is_integral<PixNonConst>::value, "Pix* is expected to be cast-able to char*");
+
+    return reinterpret_cast<Pix*>(reinterpret_cast<PixByte*>(ptr) + bytes);
+}
+
+
+//fill block  with the given color
+template <class Pix> inline
+void fillBlock(Pix* trg, int pitch, Pix col, int blockWidth, int blockHeight)
+{
+    //for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
+    //    std::fill(trg, trg + blockWidth, col);
+
+    for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
+        for (int x = 0; x < blockWidth; ++x)
+            trg[x] = col;
+}
+
+
+//nearest-neighbor (going over target image - slow for upscaling, since source is read multiple times missing out on cache! Fast for similar image sizes!)
+template <class PixSrc, class PixTrg, class PixConverter>
+void nearestNeighborScale(const PixSrc* src, int srcWidth, int srcHeight, int srcPitch,
+                          /**/  PixTrg* trg, int trgWidth, int trgHeight, int trgPitch,
+                          int yFirst, int yLast, PixConverter pixCvrt /*convert PixSrc to PixTrg*/)
+{
+    static_assert(std::is_integral<PixSrc>::value, "PixSrc* is expected to be cast-able to char*");
+    static_assert(std::is_integral<PixTrg>::value, "PixTrg* is expected to be cast-able to char*");
+
+    static_assert(std::is_same<decltype(pixCvrt(PixSrc())), PixTrg>::value, "PixConverter returning wrong pixel format");
+
+    if (srcPitch < srcWidth * static_cast<int>(sizeof(PixSrc))  ||
+        trgPitch < trgWidth * static_cast<int>(sizeof(PixTrg)))
+    {
+        assert(false);
+        return;
+    }
+
+    yFirst = std::max(yFirst, 0);
+    yLast  = std::min(yLast, trgHeight);
+    if (yFirst >= yLast || srcHeight <= 0 || srcWidth <= 0) return;
+
+    for (int y = yFirst; y < yLast; ++y)
+    {
+        const int ySrc = srcHeight * y / trgHeight;
+        const PixSrc* const srcLine = byteAdvance(src, ySrc * srcPitch);
+        PixTrg*       const trgLine = byteAdvance(trg, y    * trgPitch);
+
+        for (int x = 0; x < trgWidth; ++x)
+        {
+            const int xSrc = srcWidth * x / trgWidth;
+            trgLine[x] = pixCvrt(srcLine[xSrc]);
+        }
+    }
+}
+
+
+//nearest-neighbor (going over source image - fast for upscaling, since source is read only once
+template <class PixSrc, class PixTrg, class PixConverter>
+void nearestNeighborScaleOverSource(const PixSrc* src, int srcWidth, int srcHeight, int srcPitch,
+                                    /**/  PixTrg* trg, int trgWidth, int trgHeight, int trgPitch,
+                                    int yFirst, int yLast, PixConverter pixCvrt /*convert PixSrc to PixTrg*/)
+{
+    static_assert(std::is_integral<PixSrc>::value, "PixSrc* is expected to be cast-able to char*");
+    static_assert(std::is_integral<PixTrg>::value, "PixTrg* is expected to be cast-able to char*");
+
+    static_assert(std::is_same<decltype(pixCvrt(PixSrc())), PixTrg>::value, "PixConverter returning wrong pixel format");
+
+    if (srcPitch < srcWidth * static_cast<int>(sizeof(PixSrc))  ||
+        trgPitch < trgWidth * static_cast<int>(sizeof(PixTrg)))
+    {
+        assert(false);
+        return;
+    }
+
+    yFirst = std::max(yFirst, 0);
+    yLast  = std::min(yLast, srcHeight);
+    if (yFirst >= yLast || trgWidth <= 0 || trgHeight <= 0) return;
+
+    for (int y = yFirst; y < yLast; ++y)
+    {
+        //mathematically: ySrc = floor(srcHeight * yTrg / trgHeight)
+        // => search for integers in: [ySrc, ySrc + 1) * trgHeight / srcHeight
+
+        //keep within for loop to support MT input slices!
+        const int yTrgFirst = ( y      * trgHeight + srcHeight - 1) / srcHeight; //=ceil(y * trgHeight / srcHeight)
+        const int yTrgLast  = ((y + 1) * trgHeight + srcHeight - 1) / srcHeight; //=ceil(((y + 1) * trgHeight) / srcHeight)
+        const int blockHeight = yTrgLast - yTrgFirst;
+
+        if (blockHeight > 0)
+        {
+            const PixSrc* srcLine = byteAdvance(src, y         * srcPitch);
+            /**/  PixTrg* trgLine = byteAdvance(trg, yTrgFirst * trgPitch);
+            int xTrgFirst = 0;
+
+            for (int x = 0; x < srcWidth; ++x)
+            {
+                const int xTrgLast = ((x + 1) * trgWidth + srcWidth - 1) / srcWidth;
+                const int blockWidth = xTrgLast - xTrgFirst;
+                if (blockWidth > 0)
+                {
+                    xTrgFirst = xTrgLast;
+
+                    const auto trgPix = pixCvrt(srcLine[x]);
+                    fillBlock(trgLine, trgPitch, trgPix, blockWidth, blockHeight);
+                    trgLine += blockWidth;
+                }
+            }
+        }
+    }
+}
+
+
+template <class PixTrg, class PixConverter>
+void bilinearScale(const uint32_t* src, int srcWidth, int srcHeight, int srcPitch,
+                   /**/    PixTrg* trg, int trgWidth, int trgHeight, int trgPitch,
+                   int yFirst, int yLast, PixConverter pixCvrt /*convert uint32_t to PixTrg*/)
+{
+    static_assert(std::is_integral<PixTrg>::value,                            "PixTrg* is expected to be cast-able to char*");
+    static_assert(std::is_same<decltype(pixCvrt(uint32_t())), PixTrg>::value, "PixConverter returning wrong pixel format");
+
+    if (srcPitch < srcWidth * static_cast<int>(sizeof(uint32_t)) ||
+        trgPitch < trgWidth * static_cast<int>(sizeof(PixTrg)))
+    {
+        assert(false);
+        return;
+    }
+
+    yFirst = std::max(yFirst, 0);
+    yLast  = std::min(yLast, trgHeight);
+    if (yFirst >= yLast || srcHeight <= 0 || srcWidth <= 0) return;
+
+    const double scaleX = static_cast<double>(trgWidth ) / srcWidth;
+    const double scaleY = static_cast<double>(trgHeight) / srcHeight;
+
+    //perf notes:
+    //    -> double-based calculation is (slightly) faster than float
+    //    -> precalculation gives significant boost; std::vector<> memory allocation is negligible!
+    struct CoeffsX
+    {
+        int     x1 = 0;
+        int     x2 = 0;
+        double xx1 = 0;
+        double x2x = 0;
+    };
+    std::vector<CoeffsX> buf(trgWidth);
+    for (int x = 0; x < trgWidth; ++x)
+    {
+        const int x1 = srcWidth * x / trgWidth;
+        int x2 = x1 + 1;
+        if (x2 == srcWidth) --x2;
+
+        const double xx1 = x / scaleX - x1;
+        const double x2x = 1 - xx1;
+
+        buf[x] = { x1, x2, xx1, x2x };
+    }
+
+    for (int y = yFirst; y < yLast; ++y)
+    {
+        const int y1 = srcHeight * y / trgHeight;
+        int y2 = y1 + 1;
+        if (y2 == srcHeight) --y2;
+
+        const double yy1 = y / scaleY - y1;
+        const double y2y = 1 - yy1;
+
+        const uint32_t* const srcLine     = byteAdvance(src, y1 * srcPitch);
+        const uint32_t* const srcLineNext = byteAdvance(src, y2 * srcPitch);
+        PixTrg*         const trgLine     = byteAdvance(trg, y  * trgPitch);
+
+        for (int x = 0; x < trgWidth; ++x)
+        {
+            //perf: do NOT "simplify" the variable layout without measurement!
+            const int     x1 = buf[x].x1;
+            const int     x2 = buf[x].x2;
+            const double xx1 = buf[x].xx1;
+            const double x2x = buf[x].x2x;
+
+            const double x2xy2y = x2x * y2y;
+            const double xx1y2y = xx1 * y2y;
+            const double x2xyy1 = x2x * yy1;
+            const double xx1yy1 = xx1 * yy1;
+
+            auto interpolate = [=](int offset)
+            {
+                /*
+                    https://en.wikipedia.org/wiki/Bilinear_interpolation
+                    (c11(x2 - x) + c21(x - x1)) * (y2 - y ) +
+                    (c12(x2 - x) + c22(x - x1)) * (y  - y1)
+                */
+                const auto c11 = (srcLine    [x1] >> (8 * offset)) & 0xff;
+                const auto c21 = (srcLine    [x2] >> (8 * offset)) & 0xff;
+                const auto c12 = (srcLineNext[x1] >> (8 * offset)) & 0xff;
+                const auto c22 = (srcLineNext[x2] >> (8 * offset)) & 0xff;
+
+                return c11 * x2xy2y + c21 * xx1y2y +
+                       c12 * x2xyy1 + c22 * xx1yy1;
+            };
+
+            const double bi = interpolate(0);
+            const double gi = interpolate(1);
+            const double ri = interpolate(2);
+            const double ai = interpolate(3);
+
+            const auto b = static_cast<uint32_t>(bi + 0.5);
+            const auto g = static_cast<uint32_t>(gi + 0.5);
+            const auto r = static_cast<uint32_t>(ri + 0.5);
+            const auto a = static_cast<uint32_t>(ai + 0.5);
+
+            const uint32_t trgPix = (a << 24) | (r << 16) | (g << 8) | b;
+
+            trgLine[x] = pixCvrt(trgPix);
+        }
+    }
+}
+}
+
+#endif //XBRZ_TOOLS_H_825480175091875

win32/snes9xw.vcxproj

@@ -321,13 +321,14 @@
     <ClInclude Include="..\filter\blit.h" />
     <ClInclude Include="..\filter\epx.h" />
     <CustomBuild Include="..\filter\hq2x.h" />
-    <ClInclude Include="..\filter\xbrz-config.h" />
     <ClInclude Include="..\filter\xbrz.h" />
     <CustomBuild Include="..\font.h" />
     <CustomBuild Include="..\fxemu.h" />
     <CustomBuild Include="..\fxinst.h" />
     <CustomBuild Include="..\getset.h" />
     <CustomBuild Include="..\gfx.h" />
+    <ClInclude Include="..\filter\xbrz_config.h" />
+    <ClInclude Include="..\filter\xbrz_tools.h" />
     <ClInclude Include="..\jma\7z.h" />
     <ClInclude Include="..\jma\aribitcd.h" />
     <ClInclude Include="..\jma\ariconst.h" />

win32/snes9xw.vcxproj.filters

@@ -231,9 +231,6 @@
     <ClInclude Include="snes_ntsc_impl.h">
       <Filter>Filter</Filter>
     </ClInclude>
-    <ClInclude Include="..\filter\xbrz-config.h">
-      <Filter>Filter</Filter>
-    </ClInclude>
     <ClInclude Include="..\filter\xbrz.h">
       <Filter>Filter</Filter>
     </ClInclude>
@@ -246,6 +243,12 @@
     <ClInclude Include="cgMini.h">
       <Filter>GUI\VideoDriver</Filter>
     </ClInclude>
+    <ClInclude Include="..\filter\xbrz_config.h">
+      <Filter>Filter</Filter>
+    </ClInclude>
+    <ClInclude Include="..\filter\xbrz_tools.h">
+      <Filter>Filter</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="..\bsx.cpp">