[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Softfloat on ARM?

On Mon, 2007-02-12 at 16:42 +0000, Wookey wrote:
> On 2007-02-12 15:51 +0100, David Fokkema wrote:
> > Hi group,
> > 
> > Regarding floating point operations in ImageMagick / NetPBM there were
> > suggestions to try out the EABI port. Right now, that seems a bit
> > painful. However, on the Debian EABI port page in the wiki it says that
> > the current way of working (oldabi) is about ten times slower than a
> > program compiled with softfloat, if I understand correctly. I tried
> > simply to use an -msoft-float compiler switch. That issued a 'missing
> > -lfloat' error and that seems to be a very old library removed from
> > debian. IEEE softfloat is software released a long time ago which only
> > creates an object file and doesn't have an ARM template???
> > 
> > I'm really confused.
> > 
> > Is it possible for me to just compile ImageMagick / NetPBM with some
> > sort of softfloat support which doesn't generate hard illegal
> > instruction faults and how could I do that?
> In a word: No.
> You can't (easily) mix soft-float compiled binaries with non-soft-float
> compiled binaries on existing arm. (that's one major reason why
> EABI/armel will imporve people's lives - this sort of thing becomes
> possible).

Yes, I gathered something like that from the wiki. I just kinda hoped
that for just one application it was possible to link it to something
like -lsoftm instead of -lm.

> The problem is that you can't link soft-float and hard-float binaries. 
> To built it with soft-float requires all the libraries to also be
> soft-float so that implies a lot of rebuilding and static-linking. 

For complex applications or an entire OS, I figured that. Pity it
doesn't work for less complex binaries.

> Fixing netpbm to resize without doing FP (on arm) might not be a huge
> job - I have no idea. It would certainly be appreciated by other
> people using arm boxes in the way you are - soft-float is still very
> slow in comparison to integer algorithms. 

That's an idea. The idea behind netpbm is to keep everything very
simple. Maybe I can write a new pnmscale or modify it in some way.
Actually, that sounds like a great idea! (hoping that it is doable in
sane amounts of time, ;-)

> FP is often used quite gratuitously by developers unfamiliar with the
> idea that it might not be avilable.

True. I don't ever give it a thought. Thanks!


Reply to: