Re: [ARMedslack] mfloat-abi
Am Friday 28 January 2011 12:39:34 schrieb Stuart Winter: > > - soft: only software emulation > > - softfp + mfpu=xxx: enables the compiler to use FP > > instructions/registers inside functions, but not as function > > parameters. This way, code remains compatible with 'soft' libraries > > - hard: it's a different ABI, where parameters can be passed in FP > > registers, avoiding expensive moves between core and FP registers. But > > such code is not compatible with 'soft' libraries. > > You can also do > > gcc -dumpspecs > to see what the compiler's defaults are. > Hmm, I did not really see what's the default here (soft, softfp, hard)? BTW: A test with Ogg Vorbis decoding using mpd on my Dockstar gave me these results: - using libvorbis (floating point): ~20% CPU usage - using tremor (vorbis ported to integer only): ~2% CPU usage As soft/softfpe is some kind of porting floating point operations to integer I've assumed that Armed Slack uses mfloat-abi=hard, which forwards the floating point instructions to the slow kernel FP emulation. ___ ARMedslack mailing list ARMedslack@lists.armedslack.org http://lists.armedslack.org/mailman/listinfo/armedslack
Re: [ARMedslack] mfloat-abi
> Hmm, I did not really see what's the default here (soft, softfp, hard)? Me neither - I thought that the specs there were what was used by default. I had a further quick read about it and that isn't the case. > As soft/softfpe is some kind of porting floating point operations to integer > I've assumed that Armed Slack uses mfloat-abi=hard, which forwards the > floating point instructions to the slow kernel FP emulation. Yes, it's been ages since I looked at the issue -- I know the oldabi ARMedslack used that method, but I'd forgotten about the EABI version. The floating point used is the same as Debian use for their armel port: http://wiki.debian.org/ArmEabiPort -- Stuart Winter Slackware ARM: www.armedslack.org ___ ARMedslack mailing list ARMedslack@lists.armedslack.org http://lists.armedslack.org/mailman/listinfo/armedslack
Re: [ARMedslack] mfloat-abi
Am Friday 04 February 2011 15:07:34 schrieb Stuart Winter: > > Hmm, I did not really see what's the default here (soft, softfp, hard)? > > Me neither - I thought that the specs there were what was used by default. > I had a further quick read about it and that isn't the case. > > > As soft/softfpe is some kind of porting floating point operations to > > integer I've assumed that Armed Slack uses mfloat-abi=hard, which > > forwards the floating point instructions to the slow kernel FP emulation. > > Yes, it's been ages since I looked at the issue -- I know the oldabi > ARMedslack used that method, but I'd forgotten about the EABI version. > > The floating point used is the same as Debian use for their armel port: > http://wiki.debian.org/ArmEabiPort They say: "Floating point performance, with or without an FPU is very much faster, and mixing soft and hardfloat code is possible". This was not possible with OABI but I've never tested this on EABI. Building libvorbis with soft, softfp & hard may enlighten me ;-) If it is possible to build these three then mixing is allowed and a cpu usage comparison can follow. ___ ARMedslack mailing list ARMedslack@lists.armedslack.org http://lists.armedslack.org/mailman/listinfo/armedslack