switch native flag from -march=native to -mcpu=native for LLVM compilers on Arm 64-bit (aarch64)#5139
Conversation
…tive for aarch64
It was observed that `-march=native` does not actually provide all the CPU
features on e.g. GH200, missing bf16 when building OpenBLAS 0.3.31 with
LLVM 21.1.8. This causes build failures like:
../kernel/arm64/sbgemv_n_neon.c:146:14: error: '__builtin_neon_vld1q_bf16' needs target feature bf16,neon
146 | a3 = vld1q_bf16(a_ptr3);
| ^
Switching to -mtune=native -mcpu=native solves this issue. Therefore, apply
this option generally.
Signed-off-by: Jan André Reuter <[email protected]>
|
Failure was encountered in easybuilders/easybuild-easyconfigs#25420 (comment). Trying a rebuild with this PR. |
|
Test report for PR which caused the issue: easybuilders/easybuild-easyconfigs#25420 (comment) Will run a few more tests to make sure we don't hit any regression because of the change. |
|
OpenBLAS v0.3.30: https://gist.github.com/Thyre/3c3d148fd9ac873ed4c76570f88e7e05 |
|
Can we ask OpenBLAS maintainers for input on this? |
Sure. I'd guess that the changes described in OpenMathLib/OpenBLAS#5643 (comment) may have something to do with the new error, but generally, |
|
Just came across this issue - https://developer.arm.com/community/arm-community-blogs/b/tools-software-ides-blog/posts/compiler-flags-across-architectures-march-mtune-and-mcpu is usually a good reference when it comes with march/mtune/mcpu. I think (and happy to be corrected) for that instruction vld1q_bf16 is optional in armv9 - so march would have to disable it to make it executable on all armv9 microarchitectures. I usually refer to https://developer.arm.com/documentation/109697/2025_12/Feature-descriptions/The-Armv9-2-architecture-extension?lang=en to see what is optional vs mandatory but never easy to understand. I previously hit similar issues and asked on Arm HPC User Group (AHUG) support channels. |
This is a very good reference, thanks a lot! [reuter1@jrc0900 ~]$ diff -Naur mcpu.txt march.txt
--- mcpu.txt 2026-02-25 21:03:44.998971916 +0100
+++ march.txt 2026-02-25 21:03:50.711072338 +0100
@@ -1,39 +1,26 @@
--
-"+aes"
-"+bf16"
+"+bti"
"+ccidx"
"+complxnum"
"+crc"
+"+dit"
"+dotprod"
-"+ete"
+"+flagm"
"-fmv"
-"+fp16fml"
-"+fpac"
"+fp-armv8"
"+fullfp16"
-"+i8mm"
"+jsconv"
"+lse"
-"+mte"
"+neon"
"+outline-atomics"
"+pauth"
-"+perfmon"
-"-rand"
+"+predres"
"+ras"
"+rcpc"
"+rdm"
-"+sha2"
-"+sha3"
-"+sm4"
-"+spe"
+"+sb"
"+ssbs"
"+sve"
"+sve2"
-"+sve-aes"
-"+sve-bitperm"
-"+sve-sha3"
-"+sve-sm4"
"-target-feature"
-"+trbe"
"+v9a"Some of the features are reported as present when checking For now, I'd just go with the TL;DR:
|
Signed-off-by: Jan André Reuter <[email protected]>
-march=native to -mtune=native -mcpu=native for aarch64-march=native to -mcpu=native for aarch64
-march=native to -mcpu=native for aarch64-march=native to -mcpu=native for LLVM compilers on Arm 64-bit (aarch64)
boegel
left a comment
There was a problem hiding this comment.
This problem doesn't occur with the existing OpenBLAS-0.3.30-llvm-compilers-20.1.8.eb, but does occur with a copy that's updated to install OpenBLAS v0.3.31 instead.
I've confirmed that using -mcpu=native as proposed here fixes the problem, and indeed https://developer.arm.com/community/arm-community-blogs/b/tools-software-ides-blog/posts/compiler-flags-across-architectures-march-mtune-and-mcpu makes it pretty clear that this is what we should be doing (and we're already doing so with GCC already, see also https://github.com/easybuilders/easybuild-framework/pull/1974/changes).
So, lgtm!
It was observed that
-march=nativedoes not actually provide all the CPU features on e.g. GH200, missing bf16 when building OpenBLAS 0.3.31 with LLVM 21.1.8. This causes build failures like:Switching to
-mtune=native -mcpu=nativesolves this issue. Therefore, apply this option generally.