1.7.0-rc2: StackOverflowError with complex-valued matrix exp

[daviehh]

On macos, using version 1.7.0-rc2, julia just shows StackOverflowError with no other info when taking the matrix exp with a ~ 300x300 complex-valued matrix. Minimum example:

using LinearAlgebra

n = 300
m = rand(ComplexF64, n, n);
mex = exp(m);

Alro ran with --startup-file=no to make sure it's not some clash with other packages.

my versioninfo():

Julia Version 1.7.0-rc2
Commit f23fc0d27a (2021-10-20 12:45 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.5.0)
  CPU: Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)
Environment:
  JULIA_MATHLINK = /Applications/Mathematica.app/Contents/Frameworks/mathlink.framework
  JULIA_MATHKERNEL = /Applications/Mathematica.app/Contents/MacOS/MathKernel

Same code runs fine with Version 1.6.3 (2021-09-23).

Hardware: Macbook pro (intel) with 16 GB ram, activity monitor shows low memory usage.

In addition, the same code sometimes gives ERROR: LoadError: ReadOnlyMemoryError(), not sure how to reproduce that one...

Thanks!

[vtjnash]

I can confirm (and seems fixed on master). Looks like an openblas issue:

(lldb) bt
* thread JuliaLang/julia#1, queue = 'com.apple.main-thread'
  * frame #0: 0x000000011b9e23a8 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 24
    frame JuliaLang/julia#1: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#2: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#3: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#4: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#5: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#6: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#7: 0x000000011b9e2533 libopenblas64_.0.3.13.dylib`zgetrf_parallel + 419
    frame JuliaLang/julia#8: 0x000000011b7e0372 libopenblas64_.0.3.13.dylib`zgesv_64_ + 402
    frame JuliaLang/julia#9: 0x00000001100fc1a4
    frame JuliaLang/julia#10: 0x00000001100fead8
    frame JuliaLang/julia#11: 0x0000000110100ea4
    frame JuliaLang/julia#12: 0x00000001063385f0 libjulia-internal.1.7.dylib`do_call + 208

[daviehh]

Looks like it, current master uses libopenblas64_.0.3.17.dylib, and by running

BLAS.lbt_forward("/path/to/libopenblas64_.0.3.17.dylib"; clear=true)

in 1.7.0-rc2 the issue is resolved,

so maybe just bump openblas_jll for 1.7?

[carstenbauer]

https://discourse.julialang.org/t/inv-causes-stack-overflow-on-julia-1-7-0-and-mac-os/72411/12 seems to be very similar. Perhaps the same cause?

[stevengj]

I think it's the same cause as the abovementioned discourse thread.

I'm getting the same issue with exp as above on 1.7.0 on macOS (x86_64). It boils down to a LAPACK call: I get StackOverflowError from

using LinearAlgebra
n = 300
A = rand(ComplexF64,n,n)
B = copy(A)
LAPACK.gesv!(A,B)

or even if I directly ccall to LAPACK:

import LinearAlgebra.BLAS: @blasfunc, libblastrampoline, BlasInt
ipiv = similar(A, BlasInt, n)
info = Ref{BlasInt}()
ccall((@blasfunc(cgesv_), libblastrampoline), Cvoid,
    (Ref{BlasInt}, Ref{BlasInt}, Ptr{ComplexF32}, Ref{BlasInt}, Ptr{BlasInt},
    Ptr{ComplexF32}, Ref{BlasInt}, Ptr{BlasInt}),
    n, size(B,2), A, max(1,stride(A,2)), ipiv, B, max(1,stride(B,2)), info)

Similarly, the StackOverflowError in the discourse thread for inv(A) boils down to a ccall((@blasfunc(sgetrf_), libblastrampoline), ...).

[carstenbauer]

Mentioned by @KristofferC on Slack: JuliaPackaging/Yggdrasil#3996

[gbaraldi]

It might be that macos is more susceptible to these stackoverflows because, unless I understood incorrectly, the default pthread stack is 512kb on macos and it's larger on other OSs. Linux seems to be 2Mb and windows 1Mb.