CUDA build common lines

[garrettwrong]

Adds a CUDA kernel for building common lines. Approx order of magnitude faster before tuning/optimization. Matches Python clmatrix. Have some concerns whether Python matches MATLAB. TBD

Currently has changes from other branches in review.

[garrettwrong]

Implemented a better PF memory layout for a pretty good speedup. This takes the 89px JSB2017 problem from ~10hours for a common matrix line build to about 10 minutes.

I'll need to address the unit tests etc next.

[garrettwrong]

Happy to report that for JSB 80S 179px class averages I am achieving mean aligned angular distance of 0.36 degrees with the CUDA implementations
at commit a79f8c4 as compared with the published MATLAB code. This is up to transposing the image data, and without S weighting or J weighting on. That is, just the base Sync3N algorithm which include building CL matrix, voting procedure, building S, and the global handedness sync.

These kernels can definitely be further optimized. I plan to revisit the S weighting soon. Couple other things coming up...

[garrettwrong]

This will be in 13.1

[codecov]

Codecov Report

Attention: Patch coverage is 29.24528% with 75 lines in your changes missing coverage. Please review.

Project coverage is 86.92%. Comparing base (c8a51b8) to head (c47b626).
Report is 45 commits behind head on develop.

Files with missing lines	Patch %	Lines
src/aspire/abinitio/commonline_base.py	37.70%	38 Missing ⚠️
src/aspire/abinitio/commonline_sync3n.py	13.95%	37 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1167      +/-   ##
===========================================
- Coverage    87.37%   86.92%   -0.46%     
===========================================
  Files          132      132              
  Lines        13639    13735      +96     
===========================================
+ Hits         11917    11939      +22     
- Misses        1722     1796      +74     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[garrettwrong]

First round of self review today. Want to look at it one more time and rerun larger manual tests. Then will open it up.

[garrettwrong]

Total (unweighted) Sync3N algorithm is now right around 30 minutes on caf (A100 GPU) for 3000 179x179 single precision images. Doubles around 45m. (80S JSB class averages).

[j-c-c]

This looks great! Just a couple comments/questions.

[garrettwrong]

Looks good. Just a few things.

Also, codecov seems to complain that GPU paths are not tested? I assume this is due to the auth problems you were talking about last week.

Yeah I did see that with the last push. I was hoping 1199 resolved but only fixed some of the problems there.

The GPU tests are running on caf/decaf.

That auth issue impacted all uploads and is why we had no codecov reports at all for a while. However, it isn't why the caf/decaf report isn't there.

In this case, the caf/decaf codecov upload is failing (says report is not found). I will look into it and try to patch in another tiny PR. I think it is probably relating to not using default directories. (defaults probably assumed by codecov). I just enabled ampere reports in 1199 with the hope to reports for the GPU code we've been adding (its not coverage reporting we previously had).