On Aarch64 you do not need to check for NEON since it is a mandatory part. (The 'asimd' flag is always present) Because of that a number of defines in gcc are no longer present. I had the same issue with the LPCNet code on an 64bit raspberry pi. You need to compile with '-mtune=native -mcpu=native' to enable the correct optimizations. And in places where the code checks for __ARM_NEON__ you need to modify it to also allow __ARM_NEON (with trailing underscores)
Regards, Jeroen On 04/20/2020 04:35 PM, Richard Shaw wrote: > So I'm working on packaging LPCNet for Fedora now and I'm having an > interesting time figuring out the base optimizations each arch can > provide and I currently have them forced as shown below since we can't > assume the host system is the same as the build system. > > x86_64 -> AVX2 > i686 -> AVX > armv7hl -> NEON > s390x -> NONE > ppc64le -> forced -mvsx (more on that later) > aarch64 -> NO CLUE > > So armv7hl has NEON but aarch64 doesn't seem to support NEON, AVX, or > AVX2. > > For ppc64le I'm manually forcing -mvsx since it doesn't support AVX or > AVX2 on Fedora (which is interesting compared to OSX). > > So would VSX allow test_vec to function? Right now because AVX/AVX2 > isn't being used the build is skipping it. > > Thanks, > Richard > KF5OIM > > > > > _______________________________________________ > Freetel-codec2 mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/freetel-codec2
_______________________________________________ Freetel-codec2 mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/freetel-codec2
