neka-nat/cupoch

cupoch v0.1.8 rmm and cuda 11.0 on Ubuntu

1939938853 opened this issue · 5 comments

1939938853 commented

Hello,

It seems that cupoch v0.1.8 RMM crash with cuda 11.0 on Ubuntu? So what is your suggestion for RMM with cuda 11.0 on Ubuntu 20.04?

Thank you!

neka-nat commented

Hi,
What kind of crash is occurring?

1939938853 commented

Hi,

I use the master version, that can pass the Cuda 11.0 + RMM and Ubuntu 20.04 build.

But the performance with RMM seems not as my expected so significant. Can you give me any suggestion to improve the performance?

Thank you!

PS:

Here are some test

-- 1.9 million points VoxelDownSample to 60844 points
Open3D+OpenMP 65 ms
Cupoch+RMM 24 ms

-- RemoveRadiusOutliers(100, 20.0), point_size = 60844
Open3D+OpenMP 114 ms
Cupoch+RMM 114 ms

--Estimate Normals(), point_size = 60844
Open3D+OpenMP 59 ms
Cupoch+RMM 28 ms

--Estimate Normals(), point_size = 23733
Open3D+OpenMP 21 ms
Cupoch+RMM 22 ms

neka-nat commented

You may need to calculate more than 100,000 points to get the speed effect.
Also, using pool memory may make it a little faster.

1939938853 commented

How to use pool memory ? Example?

Thank you!

neka-nat commented

In python, you can do the following.
You can also use c++ in the same way.

import cupoch as cph
cph.initialize_allocator(cph.PoolAllocation, 1000000000)