GICP vs VGICPCuda Performance Issue

[espackman]

Hi!
I've been experimenting with this library, and I came up with some interesting results I'd like to get some feedback on.
Using the Python implementation, I modified the kitti.py file to perform some kitti specific calibrations and export the poses to a text file. Strangely enough, when testing each of the registration types available (GICP, VGICP, VGICPCuda, NDTCuda), my testing reported that the multithreaded, non-Cuda variants performed the process more quickly than the Cuda variants. I'm more than happy to share my work, but in a general sense was this to be expected? I would have thought it would be the opposite.
The test was performed locally on WSL, with a pretty beefy workstation (16-core Threadripper pro, Nvidia Quadro P4000, 32 Gb RAM). Perhaps its due to a performance gap between the GPU and CPU? Thanks, more than happy to provide more info to help understand this.

[JACKLiuDay]

Hi，I have the same problem as you. But my problem is VGICPcuda and ndtcuda not work on my PC. I tired VGICP without cuda and it works well. Cloud you give me some advice on how to use VGICPCUDA properly.
#116
And I tried other project which has NDT CUDA. And the Nvidia NX board works well but the ndt registration takes only 40ms, and the whole time is 100ms on average. So it looks like some gap between CPU and GPU makes the whole time more.

[espackman]

Hi Jack! Apologies for the delay in response.

I am unsure what is going wrong here as well, and have not found a solution. I reviewed a comparison between VGICP and VGICPCuda using the Kitti dataset, and found that the VGICP is performing roughly 2x as fast as the VGICPCuda method while having about the same accuracy. The NDTCuda method is performing poorly, both in terms of accuracy and speed.

I'm hoping the fast_gicp author @koide3 can take a look at this issue!

[koide3]

Hi guys, sorry for the late response.
Generally, the GPU-based algorithms should be much faster than the CPU-based ones.
It may take a few seconds for the first time to wait for the GPU to be initialized, but then it should work quickly.
Can you do some profiling to see which part of the code is taking a long processing time?

[espackman]

Sure thing! I'll post something in the next few days