Page 1 of 1

VLP_16 and HDL_64 very slow and cpu intensive

Posted: 15 Jun 2017, 04:48
by gu4rd1an
Hello,
I'm doing some experiments using these lidars and I find them quite cpu intensive (99% at 1hz on a i7-7820HK, linux) and very slow (even disabling the pointcloud visualization).

Am I missing anything? Tweaking, configuration... or is it the expected behavior?

Re: VLP_16 and HDL_64 very slow and cpu intensive

Posted: 15 Jun 2017, 21:49
by fferri
Hi gu4rd1an,

if you can, upload your scene (or a minimal reproducible testcase) to dropbox or similar: it will be easier to understand what's going on and give you an answer.

Re: VLP_16 and HDL_64 very slow and cpu intensive

Posted: 16 Jun 2017, 05:13
by gu4rd1an
This is really a simple scene: default floor and a bunch of cubes around.
There is a Velodyne in the middle of the scene (spinning at 20hz), I get something like 5fps.
The octree has been disabled to speed up the simulation.

This is sub-optimal at the moment. I would like my simulation to run at (least) 30fps while the lidar spins at 15-20hz in a more complex scene.
CPU: 7820HK (in complex scenes cpu reaches 100%)
GPU: 1070gtx

https://drive.google.com/open?id=0B-eQH ... WlUcVJ6SVU

Re: VLP_16 and HDL_64 very slow and cpu intensive

Posted: 16 Jun 2017, 08:29
by coppelia
Hello,

you are scanning at 20 Hz, which makes things quite slow, because there are many points. You have several possibilities to speed-up things (from what I can see the bottleneck is the call to simExtVision_handleVelodyneHDL64E: 1/3 of the time is used for vision sensor handling (i.e. rendering), 2/3 is used for computing the points from the various vision sensor's depth maps.

So you can try following:
  • reduce the vision sensor resolution. There will be more overlapping points (resolution will be worse)
  • reduce the number of generated points in the vision sensor filter. There will be less points
Above changes should be exact same for all the involved vision sensors, otherwise it will not work anymore.

Cheers