Merge remote-tracking branch 'upstream/3.4' into merge-3.4

modules/core/src/alloc.cpp

@@ -112,6 +112,13 @@ bool isAlignedAllocationEnabled()
     }
     return useMemalign;
 }
+// do not use variable directly, details: https://github.com/opencv/opencv/issues/15691
+static const bool g_force_initialization_memalign_flag
+#if defined __GNUC__
+    __attribute__((unused))
+#endif
+    = isAlignedAllocationEnabled();
+
 #endif
 
 #ifdef OPENCV_ALLOC_ENABLE_STATISTICS

modules/core/src/copy.cpp

@@ -711,6 +711,13 @@ static bool ipp_flip(Mat &src, Mat &dst, int flip_mode)
 #ifdef HAVE_IPP_IW
     CV_INSTRUMENT_REGION_IPP();
 
+    // Details: https://github.com/opencv/opencv/issues/12943
+    if (flip_mode <= 0 /* swap rows */
+        && cv::ipp::getIppTopFeatures() != ippCPUID_SSE42
+        && (int64_t)(src.total()) * src.elemSize() >= CV_BIG_INT(0x80000000)/*2Gb*/
+    )
+        return false;
+
     IppiAxis ippMode;
     if(flip_mode < 0)
         ippMode = ippAxsBoth;

modules/core/src/count_non_zero.simd.hpp

@@ -179,7 +179,25 @@ static int countNonZero32f( const float* src, int len )
 
 static int countNonZero64f( const double* src, int len )
 {
-    return countNonZero_(src, len);
+    int nz = 0, i = 0;
+#if CV_SIMD_64F
+    v_int64 sum1 = vx_setzero_s64();
+    v_int64 sum2 = vx_setzero_s64();
+    v_float64 zero = vx_setzero_f64();
+    int step = v_float64::nlanes * 2;
+    int len0 = len & -step;
+
+    for(i = 0; i < len0; i += step )
+        {
+        sum1 += v_reinterpret_as_s64(vx_load(&src[i]) == zero);
+        sum2 += v_reinterpret_as_s64(vx_load(&src[i + step / 2]) == zero);
+        }
+
+    // N.B the value is incremented by -1 (0xF...F) for each value
+    nz = i + (int)v_reduce_sum(sum1 + sum2);
+    v_cleanup();
+#endif
+    return nz + countNonZero_(src + i, len - i);
 }
 
 CountNonZeroFunc getCountNonZeroTab(int depth)