Fix indexing bug with pinthreads()

[JamesWrigley]

When the :random strategy is selected pinthreads() will default to ignoring hyperthreads, but in this case the number of CPU IDs returned will be less than the total number of threads available, and taking a view of the returned CPU IDs will fail since the end index is out of bounds. There are probably other situations that would trigger this.

[carstenbauer]

Thanks for the PR! However, I don't think we should just use min(length(cpuids), nthreads) here because this will subtly only pin a subset of the Julia threads. I think for the case length(cpuids) < nthreads we should throw an explicit error indicating that there aren't enough cores for the number of Julia threads (and perhaps also suggest that, if HT is available, hyperthreads=true could be a solution). This way, the user can be certain that all threads are properly pinned if the function doesn't error. What do you think?

[JamesWrigley]

Thanks for the review :) I promise I haven't forgotten about this, just really busy ATM so I'll probably get back to it in a week or so.

[carstenbauer]

Hey @JamesWrigley, any time to finish this? :)

[JamesWrigley]

cough sorry :flushed: I see that you fixed the original issue in cc75d18.

Now there's only two changes:

• Test for :random.
• I sneaked in 3e9a51c to improve the error message for passing an invalid symbol to pinthreads(). I was passing :firstn not realizing it was removed and got confused by the MethodError.

Also noticed that openblas_pinthreads() still relies on the old _strategy_* methods, but I see that's fixed in #28.

[JamesWrigley]

(I'll rebase when it's ready to merge)