neka-nat/cupoch

cluster_dbscan over voxel_down_sample function becomes slower in cupoch

Closed this issue · 4 comments

AlexandrineRibeiro commented

I test cluster_dbscan and voxel_down_sample functions in both open3d and cupoch. I apply cluster_dbscan in the resulting point cloud from voxel_down_sample and cupoch is slower than open3d.

I've the following processing times:

For 196133 points:

geometry::PointCloud with 196133 points.
transform (CPU) [sec]: 0.0009989738464355469
transform (GPU) [sec]: 0.0009970664978027344
voxel_down_sample (CPU) [sec]: 0.011995553970336914
voxel_down_sample (GPU) [sec]: 0.01075434684753418
cluster_dbscan (CPU) [sec]: 0.05086064338684082
cluster_dbscan (GPU) [sec]: 0.16341233253479004

For 1440 points:

geometry::PointCloud with 1440 points.
transform (CPU) [sec]: 0.0
transform (GPU) [sec]: 0.0
voxel_down_sample (CPU) [sec]: 0.0
voxel_down_sample (GPU) [sec]: 0.001996755599975586
cluster_dbscan (CPU) [sec]: 0.002991914749145508
cluster_dbscan (GPU) [sec]: 0.9901258945465088

Is there anyway gpu could make the calculation faster than cpu?

neka-nat commented

Hi,
Thank you for the question!
I have questions about your runtime environment.

  • Is it a Windows environment?
  • What is the number of OpenMP parallelism in Open3D?
AlexandrineRibeiro commented

The environment tested on has the following specs:

  • Intel Core i7-8750H CPU
  • Nvidia GTX1070 GPU
  • OMP_NUM_THREAD=1
  • Python 3.7.9
  • Windows 10
neka-nat commented

If you can use the ubuntu version, it may be faster.
The Windows version does not use a library called RMM, which uses a memory pool.
Also, these functions are not very fast, as you can see from the README results.

AlexandrineRibeiro commented

Hi @neka-nat,

Thank you for your answer.

Feel free to close this issue.

  • 👍1