diff --git a/modules/core/src/gpumat.cpp b/modules/core/src/gpumat.cpp
index 0baa8645f6..e5d9367390 100644
--- a/modules/core/src/gpumat.cpp
+++ b/modules/core/src/gpumat.cpp
@@ -1037,6 +1037,11 @@ namespace
         }
     };
 
+    template <typename T> static inline bool isAligned(const T* ptr, size_t size)
+    {
+        return reinterpret_cast<size_t>(ptr) % size == 0;
+    }
+
     //////////////////////////////////////////////////////////////////////////
     // CudaFuncTable
 
@@ -1165,6 +1170,13 @@ namespace
                     CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
             }
 
+            bool aligned = isAligned(src.data, 16) && isAligned(dst.data, 16);
+            if (!aligned)
+            {
+                cv::gpu::convertTo(src, dst);
+                return;
+            }
+
             const func_t func = funcs[src.depth()][dst.depth()][src.channels() - 1];
             CV_DbgAssert(func != 0);
 
diff --git a/modules/gpu/test/test_core.cpp b/modules/gpu/test/test_core.cpp
index 06b9fa309c..01c339f2c2 100644
--- a/modules/gpu/test/test_core.cpp
+++ b/modules/gpu/test/test_core.cpp
@@ -850,7 +850,7 @@ TEST_P(Divide_Array, WithScale)
         cv::Mat dst_gold;
         cv::divide(mat1, mat2, dst_gold, scale, depth.second);
 
-        EXPECT_MAT_NEAR(dst_gold, dst, depth.first >= CV_32F || depth.second >= CV_32F ? 1e-4 : 1.0);
+        EXPECT_MAT_NEAR(dst_gold, dst, depth.first >= CV_32F || depth.second >= CV_32F ? 1e-2 : 1.0);
     }
 }
 
diff --git a/modules/gpu/test/test_filters.cpp b/modules/gpu/test/test_filters.cpp
index 99c463b84a..2e441ee97c 100644
--- a/modules/gpu/test/test_filters.cpp
+++ b/modules/gpu/test/test_filters.cpp
@@ -334,7 +334,7 @@ TEST_P(Laplacian, Accuracy)
     cv::Mat dst_gold;
     cv::Laplacian(src, dst_gold, -1, ksize.width);
 
-    EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
+    EXPECT_MAT_NEAR(dst_gold, dst, src.depth() < CV_32F ? 0.0 : 1e-3);
 }
 
 INSTANTIATE_TEST_CASE_P(GPU_Filter, Laplacian, testing::Combine(