opencv - Initial cv::gpu::GpuMat Call Takes Forever -

all, know cuda takes time initialize described in opencv documentation, length of initial call obscenely long. feel have done wrong unsure.

i initialize cuda manually cv::gpu::setdevice(0); call. takes 1-3 seconds expected. however, when enter first test function , execute:

cv::gpu::mat input_img(input->rows, input->cols, input->type()); 

or

cv::gpu::mat input_img; //this instant input_img.create(input->rows, input->cols, input->type(); 

these take on 45 seconds execute.

now have hand written test cuda applications , not take long transfer image over, makes me believe have either compiled opencv incorrectly or programs use opencv incorrectly. have advice?

edit: see i'm confused because allocating device memory doesn't generate ptx, i'm pretty sure jit compiler isn't running on create call...