Ask Your Question

why does setting target as Opencl slow down DNN?

asked 2018-05-09 01:11:52 -0500

ashr gravatar image

I have used OpenCV DNN module for object detection. It is seen that while setting preferable target of network to Opencl the forward pass slows down by a factor of ~10x( Windows as well as embedded platforms). I can even detect GPU device with OpenCL support (OpenCL C 1.2 ). Is there any reason why this could be happening?

edit retag flag offensive close merge delete


could you add the output of:

opencv_version -opencl

to your question ? (this will show the GPU features)

berak gravatar imageberak ( 2018-05-09 02:43:58 -0500 )edit

hi i have the same problem . i don't the meaning "opencv_version -opencl" . this is a instruction code in makefile????could you explain it more clearly??

yanpeilun01 gravatar imageyanpeilun01 ( 2018-07-04 02:31:27 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2018-05-09 02:38:47 -0500

dkurt gravatar image

@ashr, efficiency measurements is such of things which require accurate description. Here we have just "some object detection network is ten times slower on some GPU with OpenCL 1.2". Please specify the network, vendor of GPU, model of CPU and GPU, version of drivers, forward pass estimations in milliseconds both for CPU and GPU, version of OpenCV.

edit flag offensive delete link more



@dkurt Please find the details as follows: I have used mobilenet-ssd network on Intel core-i5 4590 CPU. It also has GPU device Intel(R) HD Graphics 4600, OpenCL_C_Version: OpenCL C 1.2 , Driver version : OpenCV version : 3.4.1. Detection with default target takes 0.06s whereas with Opencl target it is 0.85 s.

ashr gravatar imageashr ( 2018-05-09 05:11:52 -0500 )edit
Login/Signup to Answer

Question Tools

1 follower


Asked: 2018-05-09 01:11:52 -0500

Seen: 544 times

Last updated: May 09 '18