Run CI on more architectures and more operating systems

[DilumAluthge]

Before this pull request, we run CI on the following OS+arch combinations:

• Linux x64
• Windows x64

After this pull request, we run CI on the following OS+arch combinations:

• Linux x64
• Windows x64
• macOS x64
• Linux x86
• Windows x86

[DilumAluthge]

@chriselrod Of the following, are there any OS+arch combinations that LoopVectorization does NOT support? If so, let me know and I'll remove them from the CI matrix.

• Linux x64
• Windows x64
• macOS x64
• Linux x86
• Windows x86

[chriselrod]

I supported all of them before 0.9, but with all the VectorizationBase changes, I'll probably have to do some bug fixing.

It'd be good to be able to test on ARM as well.

[DilumAluthge]

I supported all of them before 0.9, but with all the VectorizationBase changes, I'll probably have to do some bug fixing.

It'd be good to be able to test on ARM as well.

Unfortunately, GitHub Actions currently does not provide any ARM runners 😢

[DilumAluthge]

with all the VectorizationBase changes

That reminds me - I was planning on going through all of the repositories in the LoopVectorization ecosystem to tweak their CI workflows in the same way that we tweaked the workflows in this repository.

If you can provide me with a full list of those package names (LoopVectorization.jl, VectorizationBase.jl, etc.), I'll work on tweaking each repo.

[chriselrod]

LoopVectorization, VectorizationBase, and SLEEFPirates are the core packages related to SIMD.
But ArrayInterface is also important.

VectorizationBase.jl also needs a lot more tests. It has <50% coverage at the moment.

[codecov-io]

Codecov Report

Merging #163 (b217c2c) into master (fded771) will decrease coverage by 1.09%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #163      +/-   ##
==========================================
- Coverage   92.00%   90.91%   -1.10%     
==========================================
  Files          33       33              
  Lines        4529     4214     -315     
==========================================
- Hits         4167     3831     -336     
- Misses        362      383      +21     

Impacted Files	Coverage Δ
src/mapreduce.jl	77.14% <0.00%> (-9.70%)	⬇️
src/reconstruct_loopset.jl	85.71% <0.00%> (-3.97%)	⬇️
src/broadcast.jl	84.11% <0.00%> (-2.04%)	⬇️
src/lower_constant.jl	83.15% <0.00%> (-2.00%)	⬇️
src/determinestrategy.jl	95.75% <0.00%> (-1.70%)	⬇️
src/graphs.jl	85.85% <0.00%> (-1.15%)	⬇️
src/vectorizationbase_compat/contract_pass.jl	77.04% <0.00%> (-0.91%)	⬇️
src/costs.jl	84.74% <0.00%> (-0.74%)	⬇️
src/lower_store.jl	92.53% <0.00%> (-0.47%)	⬇️
src/filter.jl	88.88% <0.00%> (-0.40%)	⬇️
... and 19 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update fded771...b217c2c. Read the comment docs.

[DilumAluthge]

It's going to be a lot of noise if we add a bunch of CI jobs that always fail.

I think the best approach will be to add these jobs one by one as the bugs get fixed.

[chriselrod]

Okay. Was there a report of the failures?
I just installed Julia 1.5.3 for i686 to debug x86.
Hopefully it's only one set of problems.

[DilumAluthge]

Yeah you can see all the logs.

Go to https://github.com/chriselrod/LoopVectorization.jl/runs/1490884724

And in the left hand side you'll see all the jobs - click a job to open the log.

[DilumAluthge]

But ArrayInterface is also important.

@chriselrod Which ArrayInterface.jl should I make PRs to?

• https://github.com/chriselrod/ArrayInterface.jl
• https://github.com/SciML/ArrayInterface.jl

[chriselrod]

SciML

[DilumAluthge]

I've made several more PRs (to the packages you listed above) to tweak CI. Some tweaks include:

1. Put the docs build and the doctests in two separate jobs. That way, we can still build and deploy the docs even if the doctests are failing.
2. Move the Julia nightly CI into a separate file. That way, we can have two separate badges in the README: CI and CI (Julia nightly). And the CI badge will still be green even if tests are failing on Julia nightly.

[DilumAluthge]

LoopVectorization, VectorizationBase, and SLEEFPirates are the core packages related to SIMD.
But ArrayInterface is also important.

VectorizationBase.jl also needs a lot more tests. It has <50% coverage at the moment.

I think I got all the packages you specified. Let me know if there are other packages that you'd like me to take a look at.

[DilumAluthge]

I also tried to go through and update the CI workflows for the packages related to ProbabilityModels.jl. Let me know if I missed any packages.