Fix CUDA integral.

modules/cudaarithm/test/test_reductions.cpp

@@ -1076,7 +1076,7 @@ CUDA_TEST_P(Integral, Accuracy)
 
 INSTANTIATE_TEST_CASE_P(CUDA_Arithm, Integral, testing::Combine(
     ALL_DEVICES,
-    testing::Values(cv::Size(128, 128), cv::Size(113, 113), cv::Size(768, 1066)),
+    testing::Values(cv::Size(16, 16), cv::Size(128, 128), cv::Size(113, 113), cv::Size(768, 1066)),
     WHOLE_SUBMAT));
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////////

modules/cudev/include/opencv2/cudev/grid/detail/integral.hpp

@@ -598,7 +598,7 @@ namespace integral_detail
     __host__ static void integral(const GlobPtr<uchar>& src, const GlobPtr<uint>& dst, int rows, int cols, cudaStream_t stream)
     {
         if (deviceSupports(FEATURE_SET_COMPUTE_30)
-            && (cols % 16 == 0)
+            && (cols % 64 == 0)
             && reinterpret_cast<intptr_t>(src.data) % 32 == 0
             && reinterpret_cast<intptr_t>(dst.data) % 32 == 0)
         {