On Mon, Jul 21, 2014 at 09:50:41PM +0200, Manuel Bouyer wrote: > [...] > > I tested earmhf vs earmv7hf on the beaglebone black. with glxgears, > the performance improvement is less than 1%. > > Do you have other tests that show a large gain with earmv7hf vs earmhf ? > For 1% I don't think we should bother providing all of > earmhf, earmv6hf and earmv7hf.
I rebuilt a distribution; this time with build.sh -m evbarm -a earmv7hf -V "CPUFLAGS=-mcpu=cortex-a8 -mtune=cortex-a8 -mfpu=neon -funsafe-math-optimizations" Now glxgears shows a real improvement over earmhf; from about 24.8FPS to 25.8FPS (this would be about 4%). But we probably don't want to provide evbarm binaries with such CPU-specific compilation options (-funsafe-math-optimizations may also be a problem for some use case). -- Manuel Bouyer <bou...@antioche.eu.org> NetBSD: 26 ans d'experience feront toujours la difference --