diff --git a/samples/gpu/performance/performance.cpp b/samples/gpu/performance/performance.cpp
index 03e00654f0..668b733eea 100644
--- a/samples/gpu/performance/performance.cpp
+++ b/samples/gpu/performance/performance.cpp
@@ -14,6 +14,15 @@ void Test::gen(Mat& mat, int rows, int cols, int type)
 }
 
 
+void Test::gen(Mat& mat, int rows, int cols, int type, double low, double high)
+{
+    mat.create(rows, cols, type);
+
+    RNG rng(0);
+    rng.fill(mat, RNG::UNIFORM, Scalar::all(low), Scalar::all(high));
+}
+
+
 void TestSystem::run()
 {
     cout << setiosflags(ios_base::left);
@@ -46,13 +55,12 @@ void TestSystem::flush()
 
     int cpu_time = static_cast<int>(cpu_elapsed_ / getTickFrequency() * 1000.0);
     int gpu_time = static_cast<int>(gpu_elapsed_ / getTickFrequency() * 1000.0);
-
-    double speedup = static_cast<double>(cpu_time) / gpu_time;
-    speedup_total_ += speedup;
-
     cpu_elapsed_ = 0;
     gpu_elapsed_ = 0;
 
+    double speedup = static_cast<double>(cpu_time) / std::max(1, gpu_time);
+    speedup_total_ += speedup;
+
     cout << "    " << setiosflags(ios_base::fixed | ios_base::left);
 
     stringstream stream;
diff --git a/samples/gpu/performance/performance.h b/samples/gpu/performance/performance.h
index 8fbe667f28..9fbb1f6f2a 100644
--- a/samples/gpu/performance/performance.h
+++ b/samples/gpu/performance/performance.h
@@ -15,6 +15,7 @@ public:
     const std::string& name() const { return name_; }
 
     void gen(cv::Mat& mat, int rows, int cols, int type);
+    void gen(cv::Mat& mat, int rows, int cols, int type, double low, double high);
 
     virtual void run() = 0;
 
@@ -83,7 +84,7 @@ private:
 
 
 #define TEST(name) \
-    struct name##_test: public Test \
+    struct name##_test: Test \
     { \
         name##_test(): Test(#name) { TestSystem::instance()->add(this); } \
         void run(); \
diff --git a/samples/gpu/performance/tests.cpp b/samples/gpu/performance/tests.cpp
index db31237c22..f7753c5876 100644
--- a/samples/gpu/performance/tests.cpp
+++ b/samples/gpu/performance/tests.cpp
@@ -8,30 +8,30 @@ using namespace cv;
 TEST(matchTemplate)
 {
     Mat image, templ, result;
-    gen(image, 3000, 3000, CV_8U);
+    gen(image, 3000, 3000, CV_32F);
 
     gpu::GpuMat d_image(image), d_templ, d_result;
 
     for (int templ_size = 5; templ_size <= 1000; templ_size *= 2)
     {
-        SUBTEST << "img 3000, templ " << templ_size << ", 8U, SQDIFF";
+        SUBTEST << "img " << image.rows << ", templ " << templ_size << ", 32F, CCORR";
 
-        gen(templ, templ_size, templ_size, CV_8U);
+        gen(templ, templ_size, templ_size, CV_32F);
 
         CPU_ON;
-        matchTemplate(image, templ, result, CV_TM_SQDIFF);
+        matchTemplate(image, templ, result, CV_TM_CCORR);
         CPU_OFF;
 
         d_templ = templ;
 
         GPU_ON;
-        gpu::matchTemplate(d_image, d_templ, d_result, CV_TM_SQDIFF);
+        gpu::matchTemplate(d_image, d_templ, d_result, CV_TM_CCORR);
         GPU_OFF;
     }
 }
 
 
-TEST(minMaxLoc) 
+TEST(minMaxLoc)
 {
     Mat src;
     gpu::GpuMat d_src;
@@ -55,4 +55,80 @@ TEST(minMaxLoc)
         gpu::minMaxLoc(d_src, &min_val, &max_val, &min_loc, &max_loc);
         GPU_OFF;
     }
+}
+
+
+TEST(remap)
+{
+    Mat src, dst, xmap, ymap;
+    gpu::GpuMat d_src, d_dst, d_xmap, d_ymap;
+
+    for (int size = 1000; size <= 8000; size *= 2)
+    {
+        SUBTEST << "img " << size << " and 8UC1, 32FC1 maps";
+
+        gen(src, size, size, CV_8UC1);
+        gen(xmap, size, size, CV_32FC1, 0, size);
+        gen(ymap, size, size, CV_32FC1, 0, size);
+
+        CPU_ON;
+        remap(src, dst, xmap, ymap, INTER_LINEAR);
+        CPU_OFF;
+
+        d_src = src;
+        d_xmap = xmap;
+        d_ymap = ymap;
+
+        GPU_ON;
+        gpu::remap(d_src, d_dst, d_xmap, d_ymap);
+        GPU_OFF;
+    }
+}
+
+
+TEST(dft)
+{
+    Mat src, dst;
+    gpu::GpuMat d_src, d_dst;
+
+    for (int size = 1000; size <= 4000; size += 1000)
+    {
+        SUBTEST << "size " << size << ", 32FC2, complex-to-complex";
+
+        gen(src, size, size, CV_32FC2);
+
+        CPU_ON;
+        dft(src, dst);
+        CPU_OFF;
+
+        d_src = src;
+
+        GPU_ON;
+        gpu::dft(d_src, d_dst, Size(size, size));
+        GPU_OFF;
+    }
+}
+
+
+TEST(cornerHarris)
+{
+    Mat src, dst;
+    gpu::GpuMat d_src, d_dst;
+
+    for (int size = 2000; size <= 4000; size *= 2)
+    {
+        SUBTEST << "size " << size << ", 32FC1";
+
+        gen(src, size, size, CV_32FC1);
+
+        CPU_ON;
+        cornerHarris(src, dst, 5, 7, 0.1, BORDER_REFLECT101);
+        CPU_OFF;
+
+        d_src = src;
+
+        GPU_ON;
+        gpu::cornerHarris(d_src, d_dst, 5, 7, 0.1);
+        GPU_OFF;
+    }
 }
\ No newline at end of file