DeformableMirror performance issues

[mperrin]

Reported by @sdwill: The get_opd function can be very slow for the DeformableMirror class, in some circumstances with large numbers of actuators or high sampling.

In particular, a low-hanging fruit for efficiency is that the evaluation of the OPD should be done for just inside the DM aperture, not including all the zero padding in the wavefront.

@sdwill has volunteered to work on this.

[mperrin]

@sdwill can you comment on this one? I believe the issue here is specifically with the DM class option for evaluating the surface as a sum of many gaussians, not via convolution. And the convolution path is much faster. Can you provide any more specific instance of a case in which you were encountering performance issues?

[sdwill]

Just ran some tests of ContinuousDeformableMirror on my machine with 48x48 actuators.

If no influence function is specified via the influence_function keyword, then get_opd() becomes very slow, because _get_surface_via_gaussian_influence_functions() performs a for-loop over all actuators. If an influence function is passed in during initialization, the performance appears to be much better, as expected.

It seems to me that _get_surface_via_gaussian_influence_functions() should be combined with _get_surface_via_conv() to improve efficiency when using a Gaussian influence function.

[ermaier]

I was wondering if there are any updates on this enhancement? Currently experiencing a lot of this issue in trying to implement Electric Field Conjugation with POPPY. In order to build the complex response matrix for EFC, I have to poke each actuator in my 34x34 DM and record the final complex wavefront. The first time I do this calculation, when the rest of the DM is flat, this takes ~90 seconds (using a full 28 core node on our high performance computing cluster). When I later recalculate the matrix, after there is already a shape on the DM, it takes 40-50 minutes. @douglase and I profiled my response matrix builder function and the culprit is definitely _get_surface_via_gaussian_influence_functions re-interpolating the surface all the time...

Currently trying to work around this by assigning the final derived mirror shape to the OPD of an optic that is not a DM.

[mperrin]

Ah. The trick is not to use the Gaussian influence function at all. There is another path that uses a convolution kernel with a measured influence function and that is many orders of magnitude faster. Apologies, I don’t think I realized you were using the basic Gaussian version actively and it was slowing you down. I should provide you with example code for how we’re modeling the kiloDMs in HiCAT. I’ll try to email you code tomorrow, and please remind me if I don’t.

[ermaier]

It's no problem, I didn't even realize it was an issue until last week until I started trying to recalculate the response matrix with a shape on the mirror. Good to know there's an alternative...here's a reminder :P

[mperrin]

We should update the docs for ContinuousDeformableMirror to add some brief remarks on performance, and recommendation to use the convolution pathway rather than the Gaussian.