Platform s390x missing in TargetList.txt / Instructions for Make

[claudefalbriard]

I'd like to build the OpenBLAS package customized for s390x architecture on platform system z (LinuxOne), in order to meet the dependency in the build of Julia language.

Is this CPU architecture supported under the current release and what are the instructions to configure the make step?

[martin-frbg]

I believe the build system will autodetect z13 or generic s390x targets, but you will need a checkout of the current "develop" tree for this, not the 0.2.19 release. (Easy enough with the green "clone or download" button which will send you a zip archive if you do not want to run "git clone" locally)

[brada4]

You can install Ubuntu-packaged atlas (libatlas-base-dev) while waiting for openblas release and then that to be built into apt (their julia build uses any BLAS)

[edelsohn]

Julia language build really prefers OpenBLAS. Also, the Julia language front-end has not been updated for the Linux on z Systems ABI, so there are other pieces missing for the Julia build.

[tmh1999]

@claudefalbriard I believe the right build step is $ make TARGET=zarch.

@martin-frbg : regarding #1118, does OpenBLAS team have a plan to release anytime soon? Alpine Linux 3.6 is about to be released in early May. We would love to have OpenBLAS s390x in it (with a period to run some tests)

[edelsohn]

The s390x port is not complete. BLAS 3 support is present, but BLAS 1/2 support still is in progress. More pieces will be available soon, but I don't know if it will be complete and stable by early May.

[claudefalbriard]

Thanks for all the answers and position about ongoing work with OpenBLAS and s390x architecture support. My Julia build is an experimental one, so we can work on this task. I will try the different build recipes and give you a position.

[claudefalbriard]

Either of the two build recipes described above throw an identical error at the Julia "make" step:
#error "This arch/CPU is not supported by OpenBLAS."

1. Compiling OpenBLAS developer version with the make TARGET=zarch (executed successfully) adding the PREFIX path /opt/OpenBLAS to the Linux system path
2. Installing libatlas-base-dev from the Ubuntu repository
   How do I tell Julia build to address the matching OpenBLAS version for make, make install ?

[claudefalbriard]

I got one step further, by typing:
make USE_SYSTEM_BLAS=1

and new receive an error at the end of the Julia make step:

checking for ELF helper width... configure: error: Unknown ELF target: s390x
/data/shared/julia/downloads/julia-0.5.1/deps/unwind.mk:15: recipe for target 'build/libunwind-1.1-julia2/config.status' failed
make[1]: *** [build/libunwind-1.1-julia2/config.status] Error 1
Makefile:81: recipe for target 'julia-deps' failed
make: *** [julia-deps] Error 2

[brada4]

I can only recommend 'apt install julia'
Followed by https://github.com/xianyi/OpenBLAS/wiki/faq#debianlts (and updating frequently based on what @edelsohn said)

You can close this issue as there is no problem with OpenBLAS anymore.

[edelsohn]

make NO_LAPACK=1

[martin-frbg]

I wonder what if anything would be left from OpenBLAS with both USE_SYSTEM_BLAS=1 and NO_LAPACK=1 ?? Maybe @sharkcz 's patch for zArch CPU detection (which I just merged) will help a bit.

[brada4]

Ubuntu build of julia (USE_SYSTEM_*=1) should be reasonably usable, and wishing to complicate things you may want to copy their build flags to get further with your build.
Can you build OpenBLAS (make ...) to the degree that builtin tests pass?

[martin-frbg]

Does anybody here happen to know if the Hercules emulator would be good enough to allow debugging such build problems on more mundane architectures ?

[edelsohn]

Maybe. But why use Hercules when one can obtain a z13 VM through the LinuxONE Community Cloud

https://developer.ibm.com/linuxone/

[brada4]

It is a problem with julia build configuration. Ubuntu sort of got past it and no need to rerereinvent the wheel.

[martin-frbg]

Related Julia ticket appears to be JuliaLang/julia#21474

[martin-frbg]

Closing for now as - I believe - the original OpenBLAS issue has been fixed, and the related Julia thread appears to have petered out.