Add Kodi-specific patch

Kodi 15.0 contains an updated version of libsquish:
https://github.com/xbmc/xbmc/tree/master/tools/depends/native/libsquish-native

The OpenElec project provides a separate tarball including the Kodi-
specific patches:
http://sources.openelec.tv/devel/libsquish-1.10-openelec.tar.gz

This patch contains the relevant diff between upstream libsquish 1.13
and the OpenElec tarball.

Signed-off-by: Bernd Kuhls <bernd.kuhls@t-online.de>

diff -uwNr 1.13/squish.cpp libsquish-1.10-openelec/squish.cpp
--- 1.13/squish.cpp	2015-04-30 12:48:49.000000000 +0200
+++ libsquish-1.10-openelec/squish.cpp	2015-01-09 10:58:43.000000000 +0100
@@ -23,6 +23,7 @@
 
    -------------------------------------------------------------------------- */
 
+#include <string.h>
 #include <squish.h>
 #include "colourset.h"
 #include "maths.h"
@@ -39,7 +40,7 @@
     // grab the flag bits
     int method = flags & ( kDxt1 | kDxt3 | kDxt5 );
     int fit = flags & ( kColourIterativeClusterFit | kColourClusterFit | kColourRangeFit );
-    int extra = flags & kWeightColourByAlpha;
+	int extra = flags & ( kWeightColourByAlpha | kSourceBGRA );
 
     // set defaults
     if( method != kDxt3 && method != kDxt5 )
@@ -124,8 +125,30 @@
     return blockcount*blocksize;
 }
 
+void CopyRGBA( u8 const* source, u8* dest, int flags )
+{
+	if (flags & kSourceBGRA)
+	{
+		// convert from bgra to rgba
+		dest[0] = source[2];
+		dest[1] = source[1];
+		dest[2] = source[0];
+		dest[3] = source[3];
+	}
+	else
+	{
+		for( int i = 0; i < 4; ++i )
+			*dest++ = *source++;
+	}
+}
+
 void CompressImage( u8 const* rgba, int width, int height, void* blocks, int flags, float* metric )
 {
+	CompressImage(rgba, width, height, width*4, blocks, flags, metric);
+}
+  
+void CompressImage( u8 const* rgba, int width, int height, int pitch, void* blocks, int flags, float* metric )
+{
     // fix any bad flags
     flags = FixFlags( flags );
 
@@ -154,20 +177,14 @@
                     if( sx < width && sy < height )
                     {
                         // copy the rgba value
-                        u8 const* sourcePixel = rgba + 4*( width*sy + sx );
-                        for( int i = 0; i < 4; ++i )
-                            *targetPixel++ = *sourcePixel++;
-
+						u8 const* sourcePixel = rgba + pitch*sy + 4*sx;
+						CopyRGBA(sourcePixel, targetPixel, flags);
                         // enable this pixel
                         mask |= ( 1 << ( 4*py + px ) );
                     }
-                    else
-                    {
-                        // skip this pixel as its outside the image
                         targetPixel += 4;
                     }
                 }
-            }
 
             // compress it into the output
             CompressMasked( sourceRgba, mask, targetBlock, flags, metric );
@@ -180,6 +197,11 @@
 
 void DecompressImage( u8* rgba, int width, int height, void const* blocks, int flags )
 {
+	DecompressImage( rgba, width, height, width*4, blocks, flags );
+}
+
+void DecompressImage( u8* rgba, int width, int height, int pitch, void const* blocks, int flags )
+{
     // fix any bad flags
     flags = FixFlags( flags );
 
@@ -207,24 +229,132 @@
                     int sy = y + py;
                     if( sx < width && sy < height )
                     {
-                        u8* targetPixel = rgba + 4*( width*sy + sx );
+						u8* targetPixel = rgba + pitch*sy + 4*sx;
 
                         // copy the rgba value
+						CopyRGBA(sourcePixel, targetPixel, flags);
+					}
+					sourcePixel += 4;
+				}
+			}
+			
+			// advance
+			sourceBlock += bytesPerBlock;
+		}
+	}
+}
+
+static double ErrorSq(double x, double y)
+{
+	return (x - y) * (x - y);
+}
+
+static void ComputeBlockWMSE(u8 const *original, u8 const *compressed, unsigned int w, unsigned int h, double &cmse, double &amse)
+{
+	// Computes the MSE for the block and weights it by the variance of the original block.
+	// If the variance of the original block is less than 4 (i.e. a standard deviation of 1 per channel)
+	// then the block is close to being a single colour. Quantisation errors in single colour blocks
+	// are easier to see than similar errors in blocks that contain more colours, particularly when there
+	// are many such blocks in a large area (eg a blue sky background) as they cause banding.  Given that
+	// banding is easier to see than small errors in "complex" blocks, we weight the errors by a factor
+	// of 5. This implies that images with large, single colour areas will have a higher potential WMSE
+	// than images with lots of detail.
+
+	cmse = amse = 0;
+	unsigned int sum_p[4];  // per channel sum of pixels
+	unsigned int sum_p2[4]; // per channel sum of pixels squared
+	memset(sum_p, 0, sizeof(sum_p));
+	memset(sum_p2, 0, sizeof(sum_p2));
+	for( unsigned int py = 0; py < 4; ++py )
+	{
+		for( unsigned int px = 0; px < 4; ++px )
+		{
+			if( px < w && py < h )
+			{
+				double pixelCMSE = 0;
+				for( int i = 0; i < 3; ++i )
+				{
+					pixelCMSE += ErrorSq(original[i], compressed[i]);
+					sum_p[i] += original[i];
+					sum_p2[i] += (unsigned int)original[i]*original[i];
+				}
+				if( original[3] == 0 && compressed[3] == 0 )
+					pixelCMSE = 0; // transparent in both, so colour is inconsequential
+				amse += ErrorSq(original[3], compressed[3]);
+				cmse += pixelCMSE;
+				sum_p[3] += original[3];
+				sum_p2[3] += (unsigned int)original[3]*original[3];
+			}
+			original += 4;
+			compressed += 4;
+		}
+	}
+	unsigned int variance = 0;
                         for( int i = 0; i < 4; ++i )
-                            *targetPixel++ = *sourcePixel++;
+		variance += w*h*sum_p2[i] - sum_p[i]*sum_p[i];
+	if( variance < 4 * w * w * h * h )
+	{
+		amse *= 5;
+		cmse *= 5;
                     }
-                    else
+}
+  
+void ComputeMSE( u8 const *rgba, int width, int height, u8 const *dxt, int flags, double &colourMSE, double &alphaMSE )
                     {
-                        // skip this pixel as its outside the image
-                        sourcePixel += 4;
+	ComputeMSE(rgba, width, height, width*4, dxt, flags, colourMSE, alphaMSE);
+}
+                
+void ComputeMSE( u8 const *rgba, int width, int height, int pitch, u8 const *dxt, int flags, double &colourMSE, double &alphaMSE )
+{
+	// fix any bad flags
+	flags = FixFlags( flags );
+	colourMSE = alphaMSE = 0;
+
+	// initialise the block input
+	squish::u8 const* sourceBlock = dxt;
+	int bytesPerBlock = ( ( flags & squish::kDxt1 ) != 0 ) ? 8 : 16;
+
+	// loop over blocks
+	for( int y = 0; y < height; y += 4 )
+	{
+		for( int x = 0; x < width; x += 4 )
+		{
+			// decompress the block
+			u8 targetRgba[4*16];
+			Decompress( targetRgba, sourceBlock, flags );
+			u8 const* sourcePixel = targetRgba;
+
+			// copy across to a similar pixel block
+			u8 originalRgba[4*16];
+			u8* originalPixel = originalRgba;
+
+			for( int py = 0; py < 4; ++py )
+			{
+				for( int px = 0; px < 4; ++px )
+				{
+					int sx = x + px;
+					int sy = y + py;
+					if( sx < width && sy < height )
+					{
+						u8 const* targetPixel = rgba + pitch*sy + 4*sx;
+						CopyRGBA(targetPixel, originalPixel, flags);
                     }
+					sourcePixel += 4;
+					originalPixel += 4;
                 }
             }
 
+			// compute the weighted MSE of the block
+			double blockCMSE, blockAMSE;
+			ComputeBlockWMSE(originalRgba, targetRgba, std::min(4, width - x), std::min(4, height - y), blockCMSE, blockAMSE);
+			colourMSE += blockCMSE;
+			alphaMSE += blockAMSE;
             // advance
             sourceBlock += bytesPerBlock;
         }
     }
+	colourMSE /= (width * height * 3);
+	alphaMSE /= (width * height);
 }
 
 } // namespace squish
diff -uwNr 1.13/squish.h libsquish-1.10-openelec/squish.h
--- 1.13/squish.h	2015-04-30 12:55:27.000000000 +0200
+++ libsquish-1.10-openelec/squish.h	2015-01-09 10:58:43.000000000 +0100
@@ -57,7 +57,10 @@
     kColourRangeFit = ( 1 << 4 ),
 
     //! Weight the colour by alpha during cluster fit (disabled by default).
-    kWeightColourByAlpha = ( 1 << 7 )
+	kWeightColourByAlpha = ( 1 << 7 ),
+	
+	//! Source is BGRA rather than RGBA
+	kSourceBGRA = ( 1 << 9 ),
 };
 
 // -----------------------------------------------------------------------------
@@ -194,6 +197,7 @@
     @param rgba   The pixels of the source.
     @param width  The width of the source image.
     @param height The height of the source image.
+	@param pitch	The pitch of the source image.
     @param blocks Storage for the compressed output.
     @param flags  Compression flags.
     @param metric An optional perceptual metric.
@@ -231,6 +235,7 @@
     to allocate for the compressed output.
 */
 void CompressImage( u8 const* rgba, int width, int height, void* blocks, int flags, float* metric = 0 );
+void CompressImage( u8 const* rgba, int width, int height, int pitch, void* blocks, int flags, float* metric = 0 );
 
 // -----------------------------------------------------------------------------
 
@@ -239,6 +244,7 @@
     @param rgba   Storage for the decompressed pixels.
     @param width  The width of the source image.
     @param height The height of the source image.
+	@param pitch    The pitch of the decompressed pixels.
     @param blocks The compressed DXT blocks.
     @param flags  Compression flags.
 
@@ -254,6 +260,32 @@
     Internally this function calls squish::Decompress for each block.
 */
 void DecompressImage( u8* rgba, int width, int height, void const* blocks, int flags );
+void DecompressImage( u8* rgba, int width, int height, int pitch, void const* blocks, int flags );
+
+// -----------------------------------------------------------------------------
+
+/*! @brief Computes MSE of an compressed image in memory.
+
+	@param rgba		The original image pixels.
+	@param width	The width of the source image.
+	@param height	The height of the source image.
+	@param pitch  	The pitch of the source image.
+	@param dxt		The compressed dxt blocks
+	@param flags	Compression flags.
+	@param colourMSE	The MSE of the colour values.
+	@param alphaMSE	The MSE of the alpha values.
+	
+	The colour MSE and alpha MSE are computed across all pixels. The colour MSE is
+	averaged across all rgb values (i.e. colourMSE = sum sum_k ||dxt.k - rgba.k||/3)
+	
+	The flags parameter should specify either kDxt1, kDxt3 or kDxt5 compression, 
+	however, DXT1 will be used by default if none is specified. All other flags 
+	are ignored.
+
+	Internally this function calls squish::Decompress for each block.
+*/
+void ComputeMSE(u8 const *rgba, int width, int height, u8 const *dxt, int flags, double &colourMSE, double &alphaMSE);
+void ComputeMSE(u8 const *rgba, int width, int height, int pitch, u8 const *dxt, int flags, double &colourMSE, double &alphaMSE);
 
 // -----------------------------------------------------------------------------
 
diff -uwNr 1.13/squish.pc.in libsquish-1.10-openelec/squish.pc.in
--- 1.13/squish.pc	1970-01-01 01:00:00.000000000 +0100
+++ libsquish-1.10-openelec/squish.pc	2015-01-09 10:58:43.000000000 +0100
@@ -0,0 +1,13 @@
+prefix=/usr
+exec_prefix=${prefix}
+libdir=${prefix}/lib
+sharedlibdir=${libdir}
+includedir=${prefix}/include
+
+Name: squish
+Description: squish DXT lib
+Version: 1.1.3-kodi
+
+Requires:
+Libs: -L${libdir} -L${sharedlibdir} -lsquish
+Cflags: -I${includedir}