OpenCL not using GPU with HAAR evaluator

edit

As the title says, when I use a HAAR cascade (new or old format) cv::ocl::setUseOpenCL() command has no effect on detection speed or graphic card memory usage. When I change the cascade to an LBP one, switching OpenCL on and off using the aforementioned command yields almost 2x difference in achieved FPS and a significant difference in graphic card memory usage. Another weird thing is that this happens no matter whether I use cv::Mat or cv::UMat, the latter being a little slower due to data copying.

Specs:

General configuration for OpenCV 3.1.0-dev =====================================
  Version control:               3.1.0-117-g35014a9

  Platform:
    Host:                        Linux 3.19.0-20-lowlatency x86_64
    CMake:                       2.8.12.2
    CMake generator:             Unix Makefiles
    CMake build tool:            /usr/bin/make
    Configuration:               Release
  Other third-party libraries:
    Use IPP:                     9.0.1 [9.0.1]
         at:                     /mnt/OpenCV/opencv/3rdparty/ippicv/unpack/ippicv_lnx
    Use IPP Async:               NO
    Use VA:                      NO
    Use Intel VA-API/OpenCL:     NO
    Use Eigen:                   YES (ver 3.2.0)
    Use Cuda:                    YES (ver 6.5)
    Use OpenCL:                  YES
    Use custom HAL:              NO
  OpenCL:                        <dynamic loading="" of="" opencl="" library="">
    Include path:                /mnt/OpenCV/opencv/3rdparty/include/opencl/1.2
    Use AMDFFT:                  NO
    Use AMDBLAS:                 NO
  Graphic card:
    NVidia GTX 760