A64FX processor in Isambard 2 not recognized by archspec

[boegel]

I'm seeing this on the Isambard 2 A64FX partition

$ archspec --version
archspec, version 0.1.2
$ archspec cpu
aarch64

Detailed CPU info:

$ head -8 proc/cpuinfo
processor	: 0
BogoMIPS	: 200.00
Features	: fp asimd evtstrm sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm fcma dcpop sve
CPU implementer	: 0x46
CPU architecture: 8
CPU variant	: 0x1
CPU part	: 0x001
CPU revision	: 0

The following hack fixes the problem:

$ diff -U 7 cpu/microarchitectures.json.orig cpu/microarchitectures.json
--- cpu/microarchitectures.json.orig	2021-02-27 18:00:11.000000000 +0000
+++ cpu/microarchitectures.json	2021-02-27 18:05:58.000000000 +0000
@@ -1411,16 +1411,14 @@
     "a64fx": {
       "from": ["aarch64"],
       "vendor": "Fujitsu",
       "features": [
         "fp",
         "asimd",
         "evtstrm",
-        "aes",
-        "pmull",
         "sha1",
         "sha2",
         "crc32",
         "atomics",
         "cpuid",
         "asimdrdm",
         "fphp",

Would it make sense to just drop aes and pmull from the a64fx entry?
I'm not sure why these features are not being reported on Isambard...

Perhaps related to OS?

$ cat /etc/redhat-release
Red Hat Enterprise Linux release 8.2 (Ootpa)

@christopheredsall Any input on this?

[h-murai]

FYI, this is the result on Fugaku (A64FX).

$ head -8 /proc/cpuinfo
processor	: 0
BogoMIPS	: 200.00
Features	: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm fcma dcpop sve
CPU implementer	: 0x46
CPU architecture: 8
CPU variant	: 0x1
CPU part	: 0x001
CPU revision	: 0

There are both aes and pmull in the Features line.

[boegel]

Could this difference be due to Fugaku being an FX1000 system, while Isambard 2 is an FX700?
Cfr. https://www.fujitsu.com/global/products/computing/servers/supercomputer/specifications/

[h-murai]

And, this is the result for FX700 on our site. It's completely the same as that for Fugaku.

$ head -8 /proc/cpuinfo
processor	: 0
BogoMIPS	: 200.00
Features	: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm fcma dcpop sve
CPU implementer	: 0x46
CPU architecture: 8
CPU variant	: 0x1
CPU part	: 0x001
CPU revision	: 0

[christopheredsall]

Might be a difference between Fujitsu FX700 and HPE Apollo 80

[tomgreen66]

I help support Isambard. Have seen some references there are optional in the ARM spec. Will try and get definitive answer.

[giordano]

I also noticed the same difference between Isambard and Fugaku when checking features detection in Julia: JuliaLang/julia#44194. For what is worth, LLVM doesn't seems to enable AES for A64FX: https://github.com/llvm/llvm-project/blob/5c116d50e42f93ef4fa9a9719121378ec6b99b69/llvm/lib/Target/AArch64/AArch64.td#L997-L999

[giordano]

Uhm, on Isambard I'm still seeing

$ archspec cpu
aarch64
$ python3 -c "import archspec; print(archspec.__version__)"
0.1.4

😕

[boegel]

@alalazo Any idea why this wasn't resolved via #44?

[giordano]

@alalazo Any idea why this wasn't resolved via #44?

I wonder if it is because I didn't remove pmull. For the record, I have the same problem also on Ookami, which also has the FX700 system HPE Apollo 80. Feels like only FX1000 has aes+pmull.

[tomgreen66]

Sounds like the has finally been fixed. Can look at any Spack issue you experience on Isambard, particular modules are expected for Cray support. I have been using a local copy with removal of aes and pmull. Now looking towards next iteration of Isambard next year.

[giordano]

@tomgreen66 as mentioned in #59 (comment), the problem with archspec misdetecting the CPU seems to be solved by removing pmull, but Spack is still getting it wrong. Does your spack detect the system as a64fx or generic aarch64?