Can findOverlappingTris run on a GPU? How can I improve the performance of my Python version?

[jiushi-ai]

Hello!

findOverlappingTris runs in parallel across all mesh faces:

MeshLib/source/MRMesh/MROverlappingTris.cpp

Line 14 in a7e75b8

if ( BitSetParallelFor( mp.mesh.topology.getFaceIds( mp.region ), [&]( FaceId f )

Not sure how such parallelism could be implemented with python. Most likely this is main reason behind slowdown. And also general python efficiency loss.

Originally posted by @Grantim in #5567

[Grantim]

Hello!

As far as this function is trivially parallelized over all region faces, it can actually be implemented on GPU. But it seems to me that it would be overcomplication. I think it is better to update c++ implementation to be more useful and fit your needs rather then implementing it in python using GPU (which I believe will not be much faster then CPU c++ version).

Could you please explain a bit more, what changes do you want to add in this function?

[jiushi-ai]

Hello!

I've sent my code files to your email address support@meshinspector.com, including the test STL file.

The find_overlapping_tris_parallel function is implemented using a multi-process concurrent approach. It's the program's entry point; you only need to look at its code interface.

The functionality adjustments are in the process_close_triangles function, where I added my own necessary checks.

isOverlappingFlag = isOverlapping(t, leaf_triangle,FACE_NORMALS[f],FACE_NORMALS[int(face)])
This is the actual overlapping logic I need, handling faces that are occluded.

My local CPU has 8 cores, so it starts 12 processes to detect this test STL, taking about 5 minutes to complete.

[Grantim]

Thanks, we have received the code and mesh. And now I see what checks to you want to perform.

Anyway it does not seem to be effective to write such low-level code in python. Here is small PR that will allow to pass user-defined predicate to detect overlaps.
(Please note that using python callable like this will still be locked with GIL but I believe that it will be faster anyway)

[jiushi-ai]

Thank you. I've seen the latest C++ adjustments. I look forward to your new Python version so I can verify the execution efficiency.

Also, I don't understand: when the areas of two faces differ significantly, why isn't it considered an overlap?

if f_dbl_area * settings.minAreaFraction > f1_dbl_area:

return mm.ProcessOneResult.ContinueProcessing

[Grantim]

I don't understand: when the areas of two faces differ significantly, why isn't it considered an overlap?

This check was useful in our cases to exclude degenerated faces and spikes (which lead to false positive undercut checks), I think it can be disabled by setting minAreaFraction to negative value

[jiushi-ai]

Hi, I understand your explanation; it excludes degenerate surfaces from overlap detection access.

I also have a question about terminology:
What do ST and MT stand for in Meshlib (ST & MT)? Are they versions? What are the functional differences?

What version of meshlib is currently being installed using pip install meshlib?

[Grantim]

In tums of WASM version of MeshLib ST & MT could mean singlethread and multithread. But it is not applicable to python version. Could you please link where have you found these abbreviations ("ST" and "MT")?

[jiushi-ai]

Hello, on this page, I saw this description in a list comparing Meshlib with other software:

https://meshlib.io/blog/comparing-3d-simplification-libraries/

[jiushi-ai]

Hello. Do you have any code examples with Python version enabling GPU?

[Grantim]

Hello, on this page, I saw this description in a list comparing Meshlib with other software:

https://meshlib.io/blog/comparing-3d-simplification-libraries/

I see, we have singlethread and parallel versions of simplification algorithm, so ST and MT there means singlethread and multithread

---

Hello. Do you have any code examples with Python version enabling GPU?

Some algorithms have GPU (CUDA) implementations that are present in meshlib.mrcudapy for example auto rebuild:

MeshLib/python_scripts/mesh_rebuild.py

Lines 1 to 10 in 33023f4

	from meshlib import mrmeshpy as mm
	from meshlib import mrcudapy as mc
	def rebuild( mesh : mm.Mesh, voxelSize : float, decimate : bool = True )->mm.Mesh:
	params = mm.RebuildMeshSettings()
	params.voxelSize = voxelSize # mm.suggestVoxelSize(mesh,5e7) # suggest amount of voxels to use
	if mc.isCudaAvailable():
	params.fwn = mc.FastWindingNumber(mesh) # only if cuda is available, for speedup
	params.decimate = decimate
	return mm.rebuildMesh(mesh,params) # it can leave some small components

[jiushi-ai]

If mc.isCudaAvailable() is called, will it automatically find that GPU? Because I have two GPUs, and generally, isn't it necessary to specify the GPU ID?

[Grantim]

Yes, it will take first available CUDA capable GPU