Make ab initio faster

[Dimitry Tegunov]

Running ab initio 3D refinement, I've noticed it makes reconstructions using the full box size in every iteration. This results in 1 second being spent on SGD, and 55 on maximization. I think a lot of time could be saved by only using a box size matching the resolution + small margin. Extracting downsampled particles just to run ab initio faster seems very inconvenient.

I was also wondering if ab initio works with sub-tomograms now? That would be awesome!

[Takanori Nakane]

Hi Dimitry,

Thank you very much for suggestion. The inefficiency of the M step is on our issue list but it will take time. It is too dangerous to change the core code right now. Hopefully the target of the next version?

I don't know about sub-tomo. I hope Sjors can answer it.

Takanori Nakane

[Dimitry Tegunov]

Ok, I understand. Alternatively: I remember FFTW being forced into single-threaded mode in the earlier versions, starting back in 1.x. However, I haven't had any issues in 1.4 with multithreading enabled. If it's still single-threaded, perhaps you could include a hidden switch to multithread it at our own risk? That would make reconstruction a lot faster in many cases.

[Takanori Nakane]

FFT alone is not the bottleneck. The refactoring of the M step needs careful profiling and implementation, which is impossible by version 3.0.

[Takanori Nakane]

(not for 3.0)