[Arm-netbook] RedSleeve - New ARM Linux Distro

Gordan Bobic gordan at bobich.net
Fri Apr 13 11:25:16 BST 2012


Vladimir Pantelic wrote:
> Gordan Bobic wrote:
>> Vladimir Pantelic wrote:
>>>  Gordan Bobic wrote:
>>>
>> It's nowhere nearly as clear cut in the case of emulation (soft) vs
>> soft-float (VFP with old ABI), as I've explained here:
>> http://www.raspberrypi.org/forum/distributions/redsleeve-linux-on-raspberry-pi/#p62612
>>
>> In a lot of cases soft-float is actually slower than emulation.
>>
>> Hard-float should be faster in all cases, but it'll only make a
>> difference if the code uses a lot of FP.
> 
> and if it uses a lot of FP, hand coded (neon) assembly for the time
> critical stuff will beat any compiler generated code anyway,
> see e.g. FFmpeg/libav

Maybe on ARM, but only because GCC is unspeakably crap when it comes to 
vectorizing. On x86 ICC does a sterling job of leveraging SIMD extensions:

http://www.altechnative.net/2010/12/31/choice-of-compilers-part-1-x86/

Things may have improved slightly since then, but I very much doubt it.

ARM's own compiler may fare better on ARM, but IIRC there wasn't a 
freebie non-commercial version of that available.

>>>  comparing *only* soft vs hard calling convention yields much more
>>>  reasonable results:
>>>
>>>  http://markmail.org/message/to7jc6dx2h7tt7ak
>> Indeed, on something very FP heavy that doesn't context switch between
>> functions much like povray, soft-float is only 5-10% slower than
>> hard-float. I would expect soft (fully emulated) to be much slower in
>> this case.
>>
>> But all this will still make no difference in applications that don't
>> use FP operations - I wouldn't expect to see any meaningful difference
>> between soft, soft-float and hard-float on those.
> 
> unless the "soft" ones are using armv5 and an old compiler and the
> "hard" ones armv7 and a newer/better compiler

If you are doing soft-only, there isn't much point in targeting past 
armv5tel.

I don't know how much difference a more recent gcc makes all other 
things being equal, but I doubt it is an earth-shattering difference.

I would like to see good empirical evidence to the contrary, though, if 
it exists.

Gordan



More information about the arm-netbook mailing list