Debug a CI crash?

[mcabbott]

Re-enabling some things skipped in #57

Closes #73

[chriselrod]

As an aside, it may be worth adding the option to use LoopVectorization's threads.
But it won't necessarily be faster until I get around to adding cache-level blocking, e.g. Tullio was ahead above 1000x1000 or so for matrix multiplication on my computer:

(If LV is involved in any crashes, I'd take a look.)

[mcabbott]

Would indeed be good to have such an option, impressive stuff!

Re errors, after last commit allowing LV 0.12, there are some in https://github.com/mcabbott/Tullio.jl/pull/79/checks?check_run_id=2158318303 from test/parsing.jl. Would try to boil some down but at present my only computer is an M1 mac on which LV doesn't compile.

[chriselrod]

Would try to boil some down but at present my only computer is an M1 mac on which LV doesn't compile.

I'd appreciate that, especially if you have a workflow for it. Tullio's output isn't the easiest to navigate (neither is LV's, hard problem to avoid with metaprogramming).

Any idea why that is? Both with native Julia and under Rosetta?
I was considering buying a Mac Mini to setup with buildkite for running CI jobs, but I'd prefer to wait to buy an ARM CPU until I can get one with SVE.

[mcabbott]

Oh good point, I was seeing how far I could get without Rosetta, but with it, I can indeed run things.

The extent of my workflow is to look at what @tullio <expr> verbose=2 grad=false prints out, and trim that. Perhaps it could be more organised somehow!

With native Julia, I think the first hurdle is that HWloc.jl needs a library via BinaryBuilder. Have been meaning to learn a little how all that works, but haven't yet. I can download and build hwloc natively, outside of Julia, so I think it's just a plumbing problem of some sort.

[chriselrod]

Ah. Do any BinaryBuilder packages work yet? Looking at the ARM issue, the most recent binary builder comment seems to imply they don't (i.e., they haven't been built for Apple ARM yet).

[mcabbott]

None that I have tried work, at the moment. I don't understand whether they can be fixed piecemeal, or must wait for a global re-build.

[mcabbott]

Merged but not quite fixed? E.g.

https://github.com/mcabbott/Tullio.jl/pull/105/checks?check_run_id=2457882209

1h 17m 29s

  gradients.jl:335
229
  gradients.jl:349
230
ERROR: Package Tullio errored during testing (received signal: KILL)