keenon / nimblephysics

Nimble: Physics Engine for Biomechanics and Deep Learning
http://www.nimblephysics.org
Other
403 stars 44 forks source link

Adding multithreading to MarkerFitter::fitTrajectory() #113

Closed keenon closed 1 year ago

keenon commented 1 year ago

This seems like a huge processing time win on my laptop, regardless of the number of trials you're using. This uses a strategy where we process 32 timesteps at a time in parallel while doing "single threaded" IK, where all 32 timesteps use the same initialization from at most 32 timesteps ago. I don't know what kind of quality implications this will end up having.

This defaults to disabled, but you can call markerFitter.setParallelIKWarps(True) to enable it.

nickbianco commented 1 year ago

@keenon up to you if you want to hang on to the branch or delete.