With the new code, the tests in meas_extensions_trailedSources run about 5% faster than on main. (that's a "real world" number, not an iteration bench mark).

Using the HEAD^ commit seems to get a minor speed up of 3 seconds out of 413 (main takes 431 seconds) but that might just be normal variation.

Looks like Jenkins passed without needing to patch meas_deblender.

Ok, I'll just put this in review (probably should have done so earlier), and we'll get it in as-is.

I don't know how Eli Rykoff got that 7x speedup with his original one; it's not substantially different from my first attempt. It did lack some of the lifetime management of later steps (hence the lack of safety), but I'm skeptical that could make such a big difference, unless it's that thread-safe atomics (used in shared_ptr reference counts) are really unusually expensive on M1 Macs.

Eli Rykoff do you have some example code you are running to test the loop efficiency so Jim and I can try it?

Matthias Wittgen, mind taking a look at this pybind11 iteration optimization?

Not that this went though a couple of iterations, so it makes sense to just look at all changes in the PR at once, rather than commit-by-commit.

https://github.com/lsst/afw/pull/640/files