fix output matrix allocation in cv::subtract(cherry picked from commit 629461c836)

modules/core/src/arithm.cpp

@@ -1553,43 +1553,58 @@ void cv::add( InputArray src1, InputArray src2, OutputArray dst,
     arithm_op(src1, src2, dst, mask, dtype, getAddTab() );
 }
 
-void cv::subtract( InputArray src1, InputArray src2, OutputArray dst,
+void cv::subtract( InputArray _src1, InputArray _src2, OutputArray _dst,
                InputArray mask, int dtype )
 {
 #ifdef HAVE_TEGRA_OPTIMIZATION
-    if (mask.empty() && src1.depth() == CV_8U && src2.depth() == CV_8U)
+    int kind1 = _src1.kind(), kind2 = _src2.kind();
+    Mat src1 = _src1.getMat(), src2 = _src2.getMat();
+    bool src1Scalar = checkScalar(src1, _src2.type(), kind1, kind2);
+    bool src2Scalar = checkScalar(src2, _src1.type(), kind2, kind1);
+
+    if (!src1Scalar && !src2Scalar && mask.empty() &&
+        src1.depth() == CV_8U && src2.depth() == CV_8U)
+    {
+        if (dtype == -1)
         {
-        if (dtype == -1 && dst.fixedType())
-            dtype = dst.depth();
+            if (_dst.fixedType())
+            {
+                dtype = _dst.depth();
+            }
+            else
+            {
+                dtype = src1.depth();
+            }
+        }
 
-        dtype = CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src1.channels());
+        dtype = CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), _src1.channels());
 
-        if (!dst.fixedType() || dtype == dst.type())
+        if (dtype == _dst.type())
         {
-            dst.create(src1.size(), dtype);
+            _dst.create(_src1.size(), dtype);
 
             if (dtype == CV_16S)
             {
-                Mat _dst = dst.getMat();
-                if(tegra::subtract_8u8u16s(src1.getMat(), src2.getMat(), _dst))
+                Mat dst = _dst.getMat();
+                if(tegra::subtract_8u8u16s(src1, src2, dst))
                     return;
             }
             else if (dtype == CV_32F)
             {
-                Mat _dst = dst.getMat();
-                if(tegra::subtract_8u8u32f(src1.getMat(), src2.getMat(), _dst))
+                Mat dst = _dst.getMat();
+                if(tegra::subtract_8u8u32f(src1, src2, dst))
                     return;
             }
             else if (dtype == CV_8S)
             {
-                Mat _dst = dst.getMat();
-                if(tegra::subtract_8u8u8s(src1.getMat(), src2.getMat(), _dst))
+                Mat dst = _dst.getMat();
+                if(tegra::subtract_8u8u8s(src1, src2, dst))
                     return;
             }
         }
     }
 #endif
-    arithm_op(src1, src2, dst, mask, dtype, getSubTab() );
+    arithm_op(_src1, _src2, _dst, mask, dtype, getSubTab() );
 }
 
 void cv::absdiff( InputArray src1, InputArray src2, OutputArray dst )

modules/core/test/test_arithm.cpp

@@ -1580,12 +1580,21 @@ TEST_P(Mul1, One)
 
 INSTANTIATE_TEST_CASE_P(Arithm, Mul1, testing::Values(Size(2, 2), Size(1, 1)));
 
-TEST(Subtract8u8u16s, EmptyOutputMat)
+TEST(Subtract, EmptyOutputMat)
 {
     cv::Mat src1 = cv::Mat::zeros(16, 16, CV_8UC1);
     cv::Mat src2 = cv::Mat::zeros(16, 16, CV_8UC1);
-    cv::Mat dst;
-    cv::subtract(src1, src2, dst, cv::noArray(), CV_16S);
-    ASSERT_FALSE(dst.empty());
-    ASSERT_EQ(0, cv::countNonZero(dst));
+    cv::Mat dst1, dst2, dst3;
+
+    cv::subtract(src1, src2, dst1, cv::noArray(), CV_16S);
+    cv::subtract(src1, src2, dst2);
+    cv::subtract(src1, cv::Scalar::all(0), dst3, cv::noArray(), CV_16S);
+
+    ASSERT_FALSE(dst1.empty());
+    ASSERT_FALSE(dst2.empty());
+    ASSERT_FALSE(dst3.empty());
+
+    ASSERT_EQ(0, cv::countNonZero(dst1));
+    ASSERT_EQ(0, cv::countNonZero(dst2));
+    ASSERT_EQ(0, cv::countNonZero(dst3));
 }