upgrade VectorizationBase

[GiggleLiu]

@chriselrod I am trying to upgrade VectorizationBase to v0.21. In order to see the place that it errors, I removed several lines about stridedpointer first. It magically not show any error.

Can you please help me check whether I am doing it correctly?

[GiggleLiu]

Wait... The tests take more than 3 hours and is still running, I have to kill it:

Why is it so slow? I didn't see the same issue on my local host.

[chriselrod]

Why is it so slow? I didn't see the same issue on my local host.

Did tests pass locally?
If so, I'd let it keep going. Maybe reduce the number of tests.

CI takes many times longer when it's getting code coverage info.

[GiggleLiu]

Why is it so slow? I didn't see the same issue on my local host.

Did tests pass locally?
If so, I'd let it keep going. Maybe reduce the number of tests.

CI takes many times longer when it's getting code coverage info.

Yes. It is not about the number of tests, it is mainly about the first several tests. The compiling is very slow. The test on my local host takes less than 3min. I do not think taking 3 hour is normal.

[chriselrod]

Yes. It is not about the number of tests, it is mainly about the first several tests. The compiling is very slow. The test on my local host takes less than 3min. I do not think taking 3 hour is normal.

Same basic story with Octavian's and LoopVectorization's tests.
I think the slow down you describe is a bit extreme, even compared to those libraries, but sadly it is not exactly unexpected.

[GiggleLiu]

Just removed the coverage from CI. It works! thanks for your hint~

[chriselrod]

You could try something like Octavian does:
https://github.com/JuliaLinearAlgebra/Octavian.jl/blob/ead8fbb53d55fa726d135a7b83af099acec57093/.github/workflows/ci.yml#L14-L58
It has a "test" and "coverage" job.

Then it checks if code coverage is enabled before running the slowest tests. Something like that may help.

It'd let you test a bunch of different sizes (if that's a concern), while still getting a code coverage (under)estimate from some smaller / faster subset of tests.