Re: armhf: neon instruction ?
On Mon, Jan 10, 2022 at 3:49 PM Mathieu Malaterre <malat@debian.org> wrote:
>
> Dear arm porters,
>
> Could someone please clarify if I can build a package using gcc flags such as:
>
> -march=armv7-a -mfpu=neon-vfpv4 -mfloat-abi=hard -mfp16-format=ieee
>
> Ref:
>
> * https://salsa.debian.org/debian-phototools-team/highway/-/blob/master/CMakeLists.txt#L182-185
A package built with those options is likely to not work on currently supported
processors that only have VFPv3 rather than VFPv4 (All Cortex-A8, A9 and more).
The highest supported FPU implemetnation at the moment is "vfpv3-d16", so
no VFPv4 extensions, NEON or extra double-precision registers, see also [1].
A future Debian release might drop support for implementations with less than 32
double-precision registers or with no NEON, but would likely still need runtime
detection for vfpv4 features that are not in vfpv3 (fused
multiply-accumulate and
half precision floats).
Arnd
[1] https://wiki.debian.org/ArmHardFloatPort/VfpComparison
Reply to: