[OpenBLAS] build 64 Aarch without special case

[matbesancon]

Following JuliaLang/julia#39436, first attempt at building blas with 64 bits even on aarch64 without the special condition

[matbesancon]

I'm not sure if something special should be made to trigger the various builds here?

[giordano]

I'm not sure if something special should be made to trigger the various builds here?

Yes, you need to touch all the files O/OpenBLAS/OpenBLAS*@x.y.z/build_tarballs.jl for the version x.y.z you want to rebuild (0.3.13?)

[ViralBShah]

By touch, I assume you mean make at least a whitespace change, right? Would be nice to get openblas to 64-bit on aarch64. Will need changes in various other packages as well - and it will be nice to get it done.

[giordano]

There is also OpenBLAS high count.

Related to aarch64 OpenBLAS there is JuliaLang/julia#39612, too

[ViralBShah]

I think OpenBLAS high count is not crucial at this point (since it cannot be automatically used anyways). I presume this is in preparation of using BINARY=64 on Julia base for aarch64.

[matbesancon]

woops, should not have touched HighCount 3.12?

[giordano]

The version is wrong, but it isn't good it's failing....

[matbesancon]

also it's failing on x86, so independently from the change of the PR :/

[giordano]

Provided that you still need to edit 0.3.13 instead of 0.3.12, I'm very confused by the error: it looks like we're hitting the bug which was supposedly fixed by #2415 😕

[matbesancon]

Is there a direction to investigate in particular? Not sure I can be particularly helpful

[staticfloat]

@giordano didn't we just add some compiler header fixes for an AVX512 bug?

Googling this error message turns up this: https://gcc.gnu.org/legacy-ml/gcc-patches/2018-01/msg01962.html

[giordano]

Yes #2465 (comment)

[staticfloat]

Excellent. In that thread:

Note this patch doesn't apply to GCC 6 as _mm512_abs_pd and _mm512_mask_abs_pd are not defined

[giordano]

But this is failing with libgfortran 4 and 6, so not GCC 6

[giordano]

I mean, one possibility is that I built the wrong GCC..... But JuliaPackaging/BinaryBuilderBase.jl#101 (comment) would indicate that it isn't the case

[staticfloat]

I believe this is an OpenBLAS bug. I can confirm locally that 0.3.12 breaks and 0.3.13 works, so I say we just build 0.3.13 and use that.

[giordano]

Good to go?

[staticfloat]

Note that this version is going to be incompatible with, like, everything because the aarch64 build is now ILP64.