diff --git a/samples/gpu/performance/performance.cpp b/samples/gpu/performance/performance.cpp
index 1b5a0bf8c5..03e00654f0 100644
--- a/samples/gpu/performance/performance.cpp
+++ b/samples/gpu/performance/performance.cpp
@@ -33,9 +33,8 @@ void TestSystem::run()
     }
 
     cout << setiosflags(ios_base::fixed | ios_base::left);
-    cout << "\nCPU Total: " << setprecision(3) << cpu_total_ / getTickFrequency() << " sec\n";
-    cout << "GPU Total: " << setprecision(3) << gpu_total_ / getTickFrequency() << " sec (x" 
-        << setprecision(3) << static_cast<double>(cpu_total_) / gpu_total_ << ")\n";
+    cout << "\nAverage GPU Speedup: x" << setprecision(3) 
+        << speedup_total_ / num_subtests_called_ << endl;
     cout << resetiosflags(ios_base::fixed | ios_base::left);
 }
 
@@ -47,7 +46,9 @@ 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;
@@ -72,6 +73,7 @@ void TestSystem::flush()
     cout << resetiosflags(ios_base::fixed | ios_base::left) << endl;
     
     can_flush_ = false;
+    num_subtests_called_++;
 }
 
 
diff --git a/samples/gpu/performance/performance.h b/samples/gpu/performance/performance.h
index 354c54e411..8fbe667f28 100644
--- a/samples/gpu/performance/performance.h
+++ b/samples/gpu/performance/performance.h
@@ -42,7 +42,6 @@ public:
     {
         int64 delta = cv::getTickCount() - cpu_started_;
         cpu_elapsed_ += delta;
-        cpu_total_ += delta; 
         can_flush_ = true;
     }  
 
@@ -52,7 +51,6 @@ public:
     {
         int64 delta = cv::getTickCount() - gpu_started_;
         gpu_elapsed_ += delta;
-        gpu_total_ += delta; 
         can_flush_ = true;
     }
 
@@ -64,9 +62,8 @@ public:
     }
 
 private:
-    TestSystem(): can_flush_(false),
-                  cpu_elapsed_(0), cpu_total_(0), 
-                  gpu_elapsed_(0), gpu_total_(0) {};
+    TestSystem(): can_flush_(false), cpu_elapsed_(0), gpu_elapsed_(0), 
+                  speedup_total_(0.0), num_subtests_called_(0) {};
 
     void flush();
 
@@ -77,8 +74,11 @@ private:
 
     bool can_flush_;
 
-    int64 cpu_started_, cpu_elapsed_, cpu_total_;
-    int64 gpu_started_, gpu_elapsed_, gpu_total_;
+    int64 cpu_started_, cpu_elapsed_;
+    int64 gpu_started_, gpu_elapsed_;
+
+    double speedup_total_;
+    int num_subtests_called_;
 };
 
 
diff --git a/samples/gpu/performance/tests.cpp b/samples/gpu/performance/tests.cpp
index bd4957fd9b..db31237c22 100644
--- a/samples/gpu/performance/tests.cpp
+++ b/samples/gpu/performance/tests.cpp
@@ -7,24 +7,52 @@ using namespace cv;
 
 TEST(matchTemplate)
 {
+    Mat image, templ, result;
+    gen(image, 3000, 3000, CV_8U);
+
+    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";
 
-        Mat image; gen(image, 3000, 3000, CV_8U);
-        Mat templ; gen(templ, templ_size, templ_size, CV_8U);
-        Mat result;
+        gen(templ, templ_size, templ_size, CV_8U);
 
         CPU_ON;
         matchTemplate(image, templ, result, CV_TM_SQDIFF);
         CPU_OFF;
 
-        gpu::GpuMat d_image(image);
-        gpu::GpuMat d_templ(templ);
-        gpu::GpuMat d_result;
+        d_templ = templ;
 
         GPU_ON;
         gpu::matchTemplate(d_image, d_templ, d_result, CV_TM_SQDIFF);
         GPU_OFF;
     }
 }
+
+
+TEST(minMaxLoc) 
+{
+    Mat src;
+    gpu::GpuMat d_src;
+
+    double min_val, max_val;
+    Point min_loc, max_loc;
+
+    for (int size = 2000; size <= 8000; size *= 2)
+    {
+        SUBTEST << "img " << size << ", 32F, no mask";
+
+        gen(src, size, size, CV_32F);
+
+        CPU_ON;
+        minMaxLoc(src, &min_val, &max_val, &min_loc, &max_loc);
+        CPU_OFF;
+
+        d_src = src;
+
+        GPU_ON;
+        gpu::minMaxLoc(d_src, &min_val, &max_val, &min_loc, &max_loc);
+        GPU_OFF;
+    }
+}
\ No newline at end of file