Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

CUDA platform hardcoded in geometry engine

See original GitHub issue

When playing with different tests and platforms I realized that perses currently requires a CUDA platform to work. Specifically, FFAllAngleGeometryEngine._log_propose method has the platform hardcoded in https://github.com/choderalab/perses/blob/87bc1d00c6c087b0dc13b38f83f5e2c4d13e012d/perses/rjmc/geometry.py#L453

My guess is that this is a bug and we actually want to allow different platforms to be used in this case. Uless there’s a real reason why CUDA is required for this method.

Issue Analytics

State:
Created a year ago
Comments:9 (9 by maintainers)

Top GitHub Comments

1reaction

zhang-ivycommented, Aug 19, 2022

Ah I see that three contexts are being created in _logp_propose, and all three of them are being cleaned up here: https://github.com/choderalab/perses/blob/6aedcd0e1fa91e7039b4482db37def1bc20bf55c/perses/rjmc/geometry.py#L693

There is another context created in _corrected_reduced_potential that is never deleted, so I wonder if this is what’s causing the slowdown

0reactions

jchoderacommented, Aug 19, 2022

For example, in _logp_propose, I count that we create three Context objects but only clean up one of them.

Top Results From Across the Web

CUDA C++ Best Practices Guide

CUDA C++ Best Practices Guide. The programming guide to using the CUDA Toolkit to obtain the best performance from NVIDIA GPUs.

geometryEngine | ArcGIS Maps SDK for JavaScript 4.25

A client-side geometry engine for testing, measuring, and analyzing the spatial relationship between two or more 2D geometries. If more than one geometry...

Writing a Game Engine from Scratch - Part 4: Graphics Library

Now obviously we hard-coded a lot of things. For example, OpenGL allows you to specify your own Data Layout for each Vertex. We...

First experience with portable high-performance geometry ...

A next generation geometry library must be adaptable to scalar, vector and accelerator/GPU platforms. To avoid the large potential effort going.

Optimized CUDA function pipeline. Rectangles, circles and ...

Rectangles, circles and arrows denote CUDA functions, data arrays and ... SPHASE was hard-coded for the specific geometry of the tank in question...