[1.7.0-b4] segfault on m1

[mryodo]

Hi!
Sorry for that if it is my fault, but I'm getting a segfault on the clean install of 1.7.0-b4 on m1 mac with the eig for the simplest matrices. mwe:

julia> using LinearAlgebra;

julia> A=rand(10, 10);

julia> eigen(A)

signal (11): Segmentation fault: 11
in expression starting at REPL[3]:1
.Ldgemm_tcopy_L8_M8_20 at /Applications/Julia-1.7.app/Contents/Resources/julia/lib/julia/libopenblas64_.dylib (unknown line)
Allocations: 1721957 (Pool: 1721200; Big: 757); GC: 2

Version info:

Julia Version 1.7.0-beta4
Commit d0c90f37ba (2021-08-24 12:35 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin20.5.0)
  CPU: Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, cyclone)

have a suspicion that something happened upon the transition from macOS 12b5 to b6.
Sorry if this is a waste of time!

[jeremiahpslewis]

I can confirm that it segfaults on my machine with the same setup.

Julia Version 1.7.0-beta4
Commit d0c90f37ba (2021-08-24 12:35 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin20.5.0)
  CPU: Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, cyclone)

[mryodo]

the issue is still preset in v1.7.0-rc1's binary
probably should close in favour of #42184

[mryodo]

okay, so #42293 by @jeremiahpslewis (kudos) fixes the build problem with julia on cmake12
unfortunately, OpenBLAS problem is still present even after build; one maybe wants to pull from OpenMathLib/OpenBLAS#3383 (apparently I am not able to figure out how)

[domschl]

Ideally Julia shouldn't use OpenBLAS on Apple M1. By using Apple's proprietary BLAS implementation (via Accelerate), a speed-up of 2x - 4x can be observed, since Apple's implementation uses their undocumented matrix multiplier hardware.
That will probably not be possible to achieve via OpenBLAS in the foreseeable future...

[jeremiahpslewis]

@domschl This is super interesting, but probably worth opening a new issue on?

[KristofferC]

Ideally Julia shouldn't use OpenBLAS on Apple M1

Julia can easily change the BLAS used (on 1.7+) see https://github.com/staticfloat/libblastrampoline and https://discourse.julialang.org/t/does-mac-m1-in-multithreads-is-slower-that-in-single-thread/61114/12?u=kristoffer.carlsson

[mryodo]

Ideally Julia shouldn't use OpenBLAS on Apple M1

Julia can easily change the BLAS used (on 1.7+) see https://github.com/staticfloat/libblastrampoline and https://discourse.julialang.org/t/does-mac-m1-in-multithreads-is-slower-that-in-single-thread/61114/12?u=kristoffer.carlsson

Thanks, it's quite an interesting discussion...
Although the need for the wrapper for the simple * imply that it would be hard to push this through out all the packages, no?
If so, it seems like some kind of a at least temporary workaround for openblas is still a good idea?

[domschl]

I made a new issue #42312 for Apple M1 BLAS HW-acceleration discussions as suggested so that the discussion can continue there...

[mryodo]

Okay, let me close this one: the segfault is gone in macos12-beta7 (guess, they did it again, ***) on the latest dev version.

If somebody still has the same problem, feel free to comment here (also, even before temporary workaround through LBT were working sort of fine)

Have a nice day, check the discussion above!