Problems with openblas in IJulia/vscode/Juno in 1.6.4

[ctkelley]

This line of code

n=512; L=rand(n,n); z=rand(n); b=L\z;

runs in the REPL and in IJUlia with 1.6.3. I get a StackOverflowError: in IJulia under 1.6.4. Running Monterey. Others report the same problem in vscode and Juno.

Please see
https://discourse.julialang.org/t/downgrade-to-1-6-3/72129/2

[metab0t]

Reproduced on Windows 10 with Julia 1.6.4 and IJulia 1.23.2.
It is strange that it works when n <= 65 and fails when n > 65.

[KristofferC]

I wonder if it has something to do with https://github.com/JuliaLang/julia/issues/43008

[daviehh]

With https://github.com/JuliaLang/julia/issues/43008, I can somewhat get around the stackoverflow with setting blas thread to 1 with BLAS.set_num_threads(1), does it work for you?

[metab0t]

With #886, I can somewhat get around the stackoverflow with setting blas thread to 1 with BLAS.set_num_threads(1), does it work for you?

Yes, it works for me and solves the problem.

[KristofferC]

I guess JuliaLang/julia@b8e63b9 was the change on 1.6.4 that caused this? @ViralBShah, is the JuliaPackaging/Yggdrasil#3732 change to change the number of threads also now getting applied when new builds of e.g. OpenBLAS 0.3.10 are created? They all seem to include this common file https://github.com/JuliaPackaging/Yggdrasil/blob/b7f1359fb814e02d67a1e6ea80767b2d1d9188d5/O/OpenBLAS/OpenBLAS%400.3.10/build_tarballs.jl#L3 so it seems so?

[ctkelley]

It's a workaround, but not a solution. I need all the blas threads I can get, so I'm sticking with 1.6.3.

[ViralBShah]

AFAICT, you have to explicitly bump the BB_REL version to get the new openblas builds on 1.6.

I believe other patches were required which needed a new BB_REL version to be built and that would have picked up other changes from common.jl.

[ctkelley]

Still a problem in 1.7.0 on Macbook Pro running M1.

[ViralBShah]

Unfortunately I have not tried M1 - but this may be an M1 specific problem with openblas.

[ctkelley]

No joy on intel Macs either. Same problem.

[ViralBShah]

This works fine on Intel macs for me using the official binaries.

julia> versioninfo()
Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.5.0)
  CPU: Intel(R) Core(TM) i5-7360U CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)

julia> a = randn(66,66); b = rand(66); norm(b-a*(a\b))
4.2670948674214197e-13

julia> a = randn(1666,1666); b = rand(1666); norm(b-a*(a\b))
1.6409946780787127e-11

[BenjaminBorn]

It works for me in a standard REPL. But in the VS Code REPL, it only works if I manually set the number of BLAS threads to 1 (see below), even setting it to 2 threads results in an error.

julia> versioninfo()
Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.5.0)
  CPU: Intel(R) Core(TM) i7-7567U CPU @ 3.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)

julia> a = randn(66,66); b = rand(66); norm(b-a*(a\b))
ERROR: ReadOnlyMemoryError()

julia> BLAS.set_num_threads(1)

julia> a = randn(66,66); b = rand(66); norm(b-a*(a\b))
1.9695810178621183e-13

[ctkelley]

It always worked for me in the REPL, and fails in IJulia in 1.6.4 and 1.7.0 on both M1 macs and Intel Macs.

[ViralBShah]

I believe this is the same as #890. Closing this as a duplicate, since that one already discusses the fix.