1.7.0: `inv` Causes Stack Overflow

[SebastianAment]

Following up on this Discourse post. According to @stevengj, this issue might be solved by the resolution this Yggdrasil issue.

I am running into an odd error when using Julia 1.7.0 on my 2017 Mac Book Pro (Intel i7) on Mac OS 12.01. In particular, executing

n = 1000;
A = randn(Float32, n, n);
inv(A)
   ERROR: StackOverflowError:

With no further stack trace printed afterwards.

• The same error occurs with ComplexF32 and ComplexF64 types, but not Float64.
• According to @carstenbauer, the issue is also present in 1.6.4, 1.7.0.rc-2, 1.7.0.rc-3, 1.7.0.
• 1.6.3 and the current master branch do not exhibit the issue.

More detailed info: On my machine, inv works for all element types up to a certain size. For Float32, the failure size is 514 by 514 and for the complex types, it’s 258 by 258. Below these cutoffs, everything works as usual.

[ViralBShah]

Does that still happen if you set the number of blas threads to 32?

[SebastianAment]

Also, #886 is probably related.

[giordano]

This is fixed by JuliaLang/julia#43300, as shown in JuliaPackaging/Yggdrasil#3999 (comment)

[giordano]

According to @carstenbauer, the issue is also present in 1.6.4, 1.7.0.rc-2, 1.7.0.rc-3, 1.7.0.

Also, I can't reproduce the issue anywhere in v1.6.4. I have a PR already open for it, but with no way to test it

[SebastianAment]

Does that still happen if you set the number of blas threads to 32?

I had

julia> BLAS.get_num_threads()
2

It only works if I set BLAS.set_num_threads(1).

[SebastianAment]

According to @carstenbauer, the issue is also present in 1.6.4, 1.7.0.rc-2, 1.7.0.rc-3, 1.7.0.

Also, I can't reproduce the issue anywhere in v1.6.4. I have a PR already open for it, but with no way to test it

This CI test of one of my packages runs 1.6-MacOS-latest (i.e. 1.6.4 to my knowledge) and fails with the same error.

[giordano]

I believe Viral was asking about building OpenBLAS with 32 threads instead of 512, and the answer is always in JuliaPackaging/Yggdrasil#3999 (comment)

[carstenbauer]

According to @carstenbauer, the issue is also present in 1.6.4, 1.7.0.rc-2, 1.7.0.rc-3, 1.7.0.

Also, I can't reproduce the issue anywhere in v1.6.4. I have a PR already open for it, but with no way to test it

[ViralBShah]

The bizarre thing is that it fails with all data types except Float64, and I generally only test for Float64 manually. We should add tests for lu in LinearAlgebra tests in Julia for size 2000 for all the BLAS data types.

[stevengj]

Basically this issue is a duplicate of #886.