diff --git a/modules/gpu/include/opencv2/gpu/gpu.hpp b/modules/gpu/include/opencv2/gpu/gpu.hpp
index 949acba690..db5aa43763 100644
--- a/modules/gpu/include/opencv2/gpu/gpu.hpp
+++ b/modules/gpu/include/opencv2/gpu/gpu.hpp
@@ -395,7 +395,7 @@ CV_EXPORTS void GaussianBlur(const GpuMat& src, GpuMat& dst, Size ksize, GpuMat&
 
 //! applies Laplacian operator to the image
 //! supports only ksize = 1 and ksize = 3
-CV_EXPORTS void Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize = 1, double scale = 1, Stream& stream = Stream::Null());
+CV_EXPORTS void Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize = 1, double scale = 1, int borderType = BORDER_DEFAULT, Stream& stream = Stream::Null());
 
 
 ////////////////////////////// Arithmetics ///////////////////////////////////
@@ -1094,8 +1094,8 @@ public:
 
     bool use_local_init_data_cost;
 private:
-	GpuMat messages_buffers;
-        
+    GpuMat messages_buffers;
+
     GpuMat temp;
     GpuMat out;
 };
diff --git a/modules/gpu/src/filtering.cpp b/modules/gpu/src/filtering.cpp
index 2349857ee7..ebf990001a 100644
--- a/modules/gpu/src/filtering.cpp
+++ b/modules/gpu/src/filtering.cpp
@@ -87,7 +87,7 @@ void cv::gpu::Scharr(const GpuMat&, GpuMat&, int, int, int, double, int, int) {
 void cv::gpu::Scharr(const GpuMat&, GpuMat&, int, int, int, GpuMat&, double, int, int, Stream&) { throw_nogpu(); }
 void cv::gpu::GaussianBlur(const GpuMat&, GpuMat&, Size, double, double, int, int) { throw_nogpu(); }
 void cv::gpu::GaussianBlur(const GpuMat&, GpuMat&, Size, GpuMat&, double, double, int, int, Stream&) { throw_nogpu(); }
-void cv::gpu::Laplacian(const GpuMat&, GpuMat&, int, int, double, Stream&) { throw_nogpu(); }
+void cv::gpu::Laplacian(const GpuMat&, GpuMat&, int, int, double, int, Stream&) { throw_nogpu(); }
 
 #else
 
@@ -664,8 +664,8 @@ namespace cv { namespace gpu { namespace device
     namespace imgproc
     {
         template <typename T, typename D>
-        void filter2D_gpu(DevMem2Db srcWhole, int ofsX, int ofsY, DevMem2Db dst, 
-                          int kWidth, int kHeight, int anchorX, int anchorY, const float* kernel, 
+        void filter2D_gpu(DevMem2Db srcWhole, int ofsX, int ofsY, DevMem2Db dst,
+                          int kWidth, int kHeight, int anchorX, int anchorY, const float* kernel,
                           int borderMode, const float* borderValue, cudaStream_t stream);
     }
 }}}
@@ -708,14 +708,14 @@ namespace
         nppFilter2D_t func;
     };
 
-    typedef void (*gpuFilter2D_t)(DevMem2Db srcWhole, int ofsX, int ofsY, DevMem2Db dst, 
-                                   int kWidth, int kHeight, int anchorX, int anchorY, const float* kernel, 
+    typedef void (*gpuFilter2D_t)(DevMem2Db srcWhole, int ofsX, int ofsY, DevMem2Db dst,
+                                   int kWidth, int kHeight, int anchorX, int anchorY, const float* kernel,
                                    int borderMode, const float* borderValue, cudaStream_t stream);
 
     struct GpuFilter2D : public BaseFilter_GPU
     {
         GpuFilter2D(Size ksize_, Point anchor_, gpuFilter2D_t func_, const GpuMat& kernel_, int brd_type_) :
-            BaseFilter_GPU(ksize_, anchor_), func(func_), kernel(kernel_), brd_type(brd_type_) 
+            BaseFilter_GPU(ksize_, anchor_), func(func_), kernel(kernel_), brd_type(brd_type_)
         {
         }
 
@@ -1193,7 +1193,7 @@ void cv::gpu::Scharr(const GpuMat& src, GpuMat& dst, int ddepth, int dx, int dy,
     sepFilter2D(src, dst, ddepth, kx, ky, buf, Point(-1,-1), rowBorderType, columnBorderType, stream);
 }
 
-void cv::gpu::Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize, double scale, Stream& stream)
+void cv::gpu::Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize, double scale, int borderType, Stream& stream)
 {
     CV_Assert(ksize == 1 || ksize == 3);
 
@@ -1206,7 +1206,7 @@ void cv::gpu::Laplacian(const GpuMat& src, GpuMat& dst, int ddepth, int ksize, d
     if (scale != 1)
         kernel *= scale;
 
-    filter2D(src, dst, ddepth, kernel, Point(-1,-1), stream);
+    filter2D(src, dst, ddepth, kernel, Point(-1,-1), borderType, stream);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/modules/gpu/test/test_filters.cpp b/modules/gpu/test/test_filters.cpp
index 6f7b8d9113..99c463b84a 100644
--- a/modules/gpu/test/test_filters.cpp
+++ b/modules/gpu/test/test_filters.cpp
@@ -334,10 +334,7 @@ TEST_P(Laplacian, Accuracy)
     cv::Mat dst_gold;
     cv::Laplacian(src, dst_gold, -1, ksize.width);
 
-    if (type == CV_32FC1)
-        EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
-    else
-        EXPECT_MAT_NEAR(getInnerROI(dst_gold, cv::Size(3, 3)), getInnerROI(dst, cv::Size(3, 3)), 0.0);
+    EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
 }
 
 INSTANTIATE_TEST_CASE_P(GPU_Filter, Laplacian, testing::Combine(
diff --git a/samples/gpu/performance/tests.cpp b/samples/gpu/performance/tests.cpp
index a49ba2a778..e025d7dd12 100644
--- a/samples/gpu/performance/tests.cpp
+++ b/samples/gpu/performance/tests.cpp
@@ -273,7 +273,7 @@ TEST(SURF)
 
     SURF surf;
     vector<KeyPoint> keypoints;
-    vector<float> descriptors;
+    Mat descriptors;
 
     surf(src, Mat(), keypoints, descriptors);
 
@@ -899,7 +899,6 @@ TEST(solvePnPRansac)
     }
 }
 
-
 TEST(GaussianBlur)
 {
     for (int size = 1000; size <= 4000; size += 1000)
@@ -928,6 +927,39 @@ TEST(GaussianBlur)
     }
 }
 
+TEST(filter2D)
+{
+    for (int size = 512; size <= 2048; size *= 2)
+    {
+        Mat src;
+        gen(src, size, size, CV_8UC4, 0, 256);
+                
+        for (int ksize = 3; ksize <= 16; ksize += 2)
+        {        
+            SUBTEST << "ksize = " << ksize << ", " << size << 'x' << size << ", 8UC4";
+            
+            Mat kernel;
+            gen(kernel, ksize, ksize, CV_32FC1, 0.0, 1.0);
+
+            Mat dst;
+            cv::filter2D(src, dst, -1, kernel);
+
+            CPU_ON;
+            cv::filter2D(src, dst, -1, kernel);
+            CPU_OFF;
+
+            gpu::GpuMat d_src(src);
+            gpu::GpuMat d_dst;
+
+            gpu::filter2D(d_src, d_dst, -1, kernel);
+
+            GPU_ON;
+            gpu::filter2D(d_src, d_dst, -1, kernel);
+            GPU_OFF;
+        }
+    }
+}
+
 TEST(pyrDown)
 {
     for (int size = 4000; size >= 1000; size -= 1000)