[OpenBLAS_jll] Upgrade to new build optimised for PowerPC

[giordano]

This only difference compared to previous build is that this one enables use of dynamic architecture detection also for the PowerPC architecture.

Ref: https://discourse.julialang.org/t/building-openblas-for-powerpc-defaults-to-power8/93972.

[vchuravy]

Can you also update the source recipe here?

[giordano]

Uhm, where exactly? In https://github.com/JuliaLang/julia/blob/80dc13595c0aa29062705547debab679b5735c08/deps/openblas.mk I don't see anything PowerPC-specific (apart from a patch). Maybe I should just delete

julia/Make.inc

Lines 834 to 835 in e536c77

	OPENBLAS_DYNAMIC_ARCH:=0
	OPENBLAS_TARGET_ARCH:=POWER8

to allow optimising for the current architecture as we do on Intel?

[ViralBShah]

That seems about right.

[vchuravy]

Yeah or just set dynamic arch to true

[ViralBShah]

Not relevant for this PR, but it would be nice for all the logic to set the various OPENBLAS build options to live in openblas.mk.

[ViralBShah]

If you just don't set OPENBLAS_TARGET_ARCH explicitly, we do default to DYNAMIC_ARCH=1 (higher up in Make.inc). May be best to just delete those two lines.

[giordano]

Ok, done! To be honest I was a bit lost among those options 😅

[ViralBShah]

Yeah, it is quite a maze. it would be nice to refactor those and put them all in openblas.mk.

[vchuravy]

Running make -C deps compile-openblas I see tuning for power9 and https://github.com/xianyi/OpenBLAS/blob/10be02c89605334b3095c0ac76f4aa328ae54f72/cmake/arch.cmake#L57-L58 does confirm that this is correct.

Nice!

Needs closer look

[vchuravy]

I spoke to soon.

It looks like we only built Power9...

/bin/sh: line 1: 609325 Segmentation fault      (core dumped) OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 ./test_sbgemm > SBBLAT3.SUMM

[vchuravy]

Hm still getting a seqfault on Power8, but this time I actually saw it build many Power6, Power8, and Power9...

/bin/sh: line 1: 714674 Segmentation fault      (core dumped) OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 ./test_sbgemm > SBBLAT3.SUMM
make[2]: *** [Makefile:167: level3] Error 139
make[2]: Leaving directory '/nobackup/users/vchuravy/dev/julia/deps/scratch/openblas-b89fb708caa5a5a32de8f4306c4ff132e0228e9a/test'
make[1]: *** [Makefile:149: tests] Error 2
make[1]: Leaving directory '/nobackup/users/vchuravy/dev/julia/deps/scratch/openblas-b89fb708caa5a5a32de8f4306c4ff132e0228e9a'

[RajalakshmiSR]

What is the compiler version and make command used to build openblas? Are you trying it on POWER8 or POWER9?

[vchuravy]

This was with:

make USE_BINARYBUILDER_OPENBLAS=0 -C deps compile-openblas

[vchuravy@satori-login-001 ~]$ gcc --version
gcc (GCC) 8.3.1 20191121 (Red Hat 8.3.1-5)
Copyright (C) 2018 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

[vchuravy@satori-login-001 ~]$ cat /proc/cpuinfo 
processor	: 0
cpu		: POWER9, altivec supported
clock		: 3800.000000MHz
revision	: 2.2 (pvr 004e 1202)

Slightly surprised thought this was a Power8 cluster xD

[RajalakshmiSR]

This was with:

make USE_BINARYBUILDER_OPENBLAS=0 -C deps compile-openblas

[vchuravy@satori-login-001 ~]$ gcc --version
gcc (GCC) 8.3.1 20191121 (Red Hat 8.3.1-5)
Copyright (C) 2018 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

[vchuravy@satori-login-001 ~]$ cat /proc/cpuinfo 
processor	: 0
cpu		: POWER9, altivec supported
clock		: 3800.000000MHz
revision	: 2.2 (pvr 004e 1202)

Slightly surprised thought this was a Power8 cluster xD

Thanks, I can recreate it. Will debug this.

[RajalakshmiSR]

This is fixed here OpenMathLib/OpenBLAS#3718

[RajalakshmiSR]

openblas 0.3.22 will have the fix included.

[KristofferC]

Good to go now?

[RajalakshmiSR]

Thanks. looks good.