[SOLVED] CUDA Examples segfault or are unstable

[mbarrow]

Hi, I want to benchmark the CUDA TLED performance for a Neo-Hookean cube.

I don’t have confidence in the simulator because many of the examples under src/applications/plugins/SofaCUDA/examples/ seem to be unstable or simply cause runSofa to segfault.

I’ve tried version 16.12 and 15.03 with much the same results (15.03 seems worse however), so I’m hoping for some help.

Does anyone have a simple .scn file known to work with ver 16.10 that uses the CUDA TLED FEM to simulate a block of Neo-Hookean solid?

[Hugo]

Dear mbarrow,

I recompiled the SofaCUDA plugin, there’s indeed some scene not working. About the CudaTetrahedronTLEDForceField.scn, it seems that there is some missing files.

I asked about it, and some devs told me that is a pretty old (2009!) piece of code. Would you feel confortable-enough with the CUDA API to look at the implementation?

Best,

Hugo

[mbarrow]

Hi, I would like to but I’m confused about one thing. There seem to be some advanced simulators showcased on the site (Eye surgery sim, tissue tearing etc). Do these not use the CUDA TLED FEM? I would like to focus on the most state of the art implementation as the code seems complex and we want to benchmark the state of the art configuration.

Cheers!

[Hugo]

Hi mbarrow,

In SOFA, the TLED approach was developed during a PhD thesis that ended in 2009. Since then, the other GPU-based simulation were implemented using the SofaCUDA plugin. These implementations do not only rely on explicit schemes. Most of them were done using implicit or semi-implicit schemes.

Cheers,

Hugo

[mbarrow]

Thank you Hugo. Sorry I’m a beginner. Does this mean I’ve been looking at the wrong GPU-classes if I want to benchmark state of the art SOFA GPU performance or is the Implicit / Semi-Implicit code not available for general users?

Cheers,

Michael

[Hugo]

Hi mbarrrow,

No worries. The TLED would not be my start in SOFA. You could rather have a look into the plugin SofaCUDA. There, you will find the FEM codes in their GPU version. Best,

Hugo

[mbarrow]

Thank you, that is very helpful!

[Hugo]

I therefore set this topic as solved.
Do not hesitate to post further question later on, mbarrow.

Cheers,

Hugo

[Author]

Posts