fixed ocl::meanStddev and created accuracy test for this function

modules/ocl/src/arithm.cpp

@@ -454,23 +454,15 @@ Scalar cv::ocl::sqrSum(const oclMat &src)
 
 void cv::ocl::meanStdDev(const oclMat &src, Scalar &mean, Scalar &stddev)
 {
-    CV_Assert(src.depth() <= CV_32S);
-    cv::Size sz(1, 1);
-    int channels = src.oclchannels();
-    Mat m1(sz, CV_MAKETYPE(CV_32S, channels), cv::Scalar::all(0)),
-        m2(sz, CV_MAKETYPE(CV_32S, channels), cv::Scalar::all(0));
-    oclMat dst1(m1), dst2(m2);
-
-//    arithmetic_sum_run(src, dst1, "arithm_op_sum");
-//    arithmetic_sum_run(src, dst2, "arithm_op_squares_sum");
-
-    m1 = (Mat)dst1;
-    m2 = (Mat)dst2;
-    int i = 0, *p = (int *)m1.data, *q = (int *)m2.data;
-    for (; i < channels; i++)
+    double total = 1.0 / src.size().area();
+
+    mean = sum(src);
+    stddev = sqrSum(src);
+
+    for (int i = 0; i < 4; ++i)
     {
-        mean.val[i] = (double)p[i] / (src.cols * src.rows);
-        stddev.val[i] = std::sqrt(std::max((double) q[i] / (src.cols * src.rows) - mean.val[i] * mean.val[i] , 0.));
+        mean[i] *= total;
+        stddev[i] = std::sqrt(std::max(stddev[i] * total - mean.val[i] * mean.val[i] , 0.));
     }
 }
 

modules/ocl/test/test_arithm.cpp

@@ -1440,6 +1440,30 @@ TEST_P(SetIdentity, Mat)
     }
 }
 
+//////////////////////////////// setIdentity /////////////////////////////////////////////////
+
+typedef ArithmTestBase MeanStdDev;
+
+TEST_P(MeanStdDev, Mat)
+{
+    for (int j = 0; j < LOOP_TIMES; j++)
+    {
+        random_roi();
+
+        Scalar cpu_mean, cpu_stddev;
+        Scalar gpu_mean, gpu_stddev;
+
+        cv::meanStdDev(src1_roi, cpu_mean, cpu_stddev);
+        cv::ocl::meanStdDev(gsrc1, gpu_mean, gpu_stddev);
+
+        for (int i = 0; i < 4; ++i)
+        {
+            EXPECT_NEAR(cpu_mean[i], gpu_mean[i], 1e-5);
+            EXPECT_NEAR(cpu_stddev[i], gpu_stddev[i], 0.1);
+        }
+    }
+}
+
 //////////////////////////////////////// Instantiation /////////////////////////////////////////
 
 INSTANTIATE_TEST_CASE_P(Arithm, Lut, Combine(testing::Range(CV_8U, CV_USRTYPE1), testing::Range(1, 5), Bool(), Bool()));
@@ -1470,5 +1494,6 @@ INSTANTIATE_TEST_CASE_P(Arithm, Compare, Combine(testing::Range(CV_8U, CV_USRTYP
 INSTANTIATE_TEST_CASE_P(Arithm, Pow, Combine(Values(CV_32F, CV_64F), testing::Range(1, 5), Bool()));
 INSTANTIATE_TEST_CASE_P(Arithm, AddWeighted, Combine(testing::Range(CV_8U, CV_USRTYPE1), testing::Range(1, 5), Bool()));
 INSTANTIATE_TEST_CASE_P(Arithm, SetIdentity, Combine(testing::Range(CV_8U, CV_USRTYPE1), testing::Range(1, 5), Bool()));
+INSTANTIATE_TEST_CASE_P(Arithm, MeanStdDev, Combine(testing::Range(CV_8U, CV_USRTYPE1), testing::Range(1, 5), Bool()));
 
 #endif // HAVE_OPENCL