Codecov Report

Merging #47099 into master will decrease coverage by 0.00%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master   #47099      +/-   ##
==========================================
- Coverage   68.85%   68.85%   -0.01%     
==========================================
  Files         436      436              
  Lines       56395    56395              
==========================================
- Hits        38833    38832       -1     
- Misses      17562    17563       +1     

tbh it doesn't look like using any neon intrinsic acceleration for hypot anyway.

Right so hypot just falls back std::hypot. Can we confirm that std::hypot is broken on older gcc for sanity check. Also is this all older versions of gcc <=7.5 or specifically 7.5.

Also can we hold off on this PR. I need to confirm if this has implication elsewhere if we are using gcc toolchain with older version to build stuff.

Right so hypot just falls back std::hypot. Can we confirm that std::hypot is broken on older gcc for sanity check. Also is this all older versions of gcc <=7.5 or specifically 7.5.

Also can we hold off on this PR. I need to confirm if this has implication elsewhere if we are using gcc toolchain with older version to build stuff.

Also since it is not using any intrinsics, if std::hypot itself is broken then I am vary of fixing it here.

I need to confirm if this is a safe change or we might see perf regressions.

I think we are good to go here. Folks building with older gcc for arm64 will not see perf gains unfortunately.

Also looks like without this pr we fallback to Sleef_hypotf8_u05, so std::hypot question still remains.

@kimishpatel , std::hypot works fine, but at::vec256<float>::sqrt() doesn't, which is the reason why I want to disable neon using older compilers.

@kimishpatel , std::hypot works fine, but at::vec256<float>::sqrt() doesn't, which is the reason why I want to disable neon using older compilers.

@malfet, I thought the issue was hypot. But if sqrt is the issue then this intrinsic vsqrtq_f32 is broken. If you have a minimal repro that would be good. Repro is issue #47098 refers to hypot. There are some arm builds which use GCC7 I think: #43265.

@malfet merged this pull request in c86af4a.