Merge github.com:Itseez/opencv

modules/core/src/ocl.cpp

@@ -4504,8 +4504,8 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
     if (vectorWidths[0] == 1)
     {
         // it's heuristic
-        vectorWidths[CV_8U] = vectorWidths[CV_8S] = 16;
-        vectorWidths[CV_16U] = vectorWidths[CV_16S] = 8;
+        vectorWidths[CV_8U] = vectorWidths[CV_8S] = 4;
+        vectorWidths[CV_16U] = vectorWidths[CV_16S] = 2;
         vectorWidths[CV_32S] = vectorWidths[CV_32F] = vectorWidths[CV_64F] = 1;
     }
 

modules/imgproc/src/opencl/match_template.cl

@@ -161,7 +161,7 @@ __kernel void matchTemplate_Naive_CCORR(__global const uchar * srcptr, int src_s
                 for (int j = 0; j < template_cols; ++j)
                 {
                     T temp = (T)(template[j]);
-                    T src = *(__global const T*)(srcptr + ind + j*(int)sizeof(T1));
+                    T src = vload4(0, (__global const T1*)(srcptr + ind + j*(int)sizeof(T1)));
 
                     sum = mad(convertToWT(src), convertToWT(temp), sum);
 

modules/imgproc/src/pyramids.cpp

@@ -341,12 +341,9 @@ pyrUp_( const Mat& _src, Mat& _dst, int)
     for( int y = 0; y < ssize.height; y++ )
     {
         T* dst0 = _dst.ptr<T>(y*2);
-        T* dst1 = _dst.ptr<T>(y*2+1);
+        T* dst1 = _dst.ptr<T>(std::min(y*2+1, dsize.height-1));
         WT *row0, *row1, *row2;
 
-        if( y*2+1 >= dsize.height )
-            dst1 = dst0;
-
         // fill the ring buffer (horizontal convolution and decimation)
         for( ; sy <= y + 1; sy++ )
         {