switching to `CL_PERF_TEST_P`, adding `safeFinish()` to flush CL queue

modules/objdetect/perf/opencl/perf_cascades.cpp

 #include "perf_precomp.hpp"
 #include <opencv2/imgproc.hpp>
 
+#include "opencv2/ts/ocl_perf.hpp"
+
 using namespace std;
 using namespace cv;
 using namespace perf;
@@ -10,7 +12,7 @@ using std::tr1::get;
 typedef std::tr1::tuple<std::string, std::string, int> Cascade_Image_MinSize_t;
 typedef perf::TestBaseWithParam<Cascade_Image_MinSize_t> Cascade_Image_MinSize;
 
-PERF_TEST_P( Cascade_Image_MinSize, CascadeClassifier_UMat,
+OCL_PERF_TEST_P(Cascade_Image_MinSize, CascadeClassifier,
                  testing::Combine(
                     testing::Values( string("cv/cascadeandhog/cascades/haarcascade_frontalface_alt.xml"),
                                      string("cv/cascadeandhog/cascades/lbpcascade_frontalface.xml") ),
@@ -42,6 +44,7 @@ PERF_TEST_P( Cascade_Image_MinSize, CascadeClassifier_UMat,
     while(next())
     {
         faces.clear();
+        cvtest::ocl::perf::safeFinish();
 
         startTimer();
         cc.detectMultiScale(uimg, faces, 1.1, 3, 0, minSize);