Re: arm software floating point support going forward

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

 



On Fri, 25 Jan 2013, Josh Boyer wrote:

> On Fri, Jan 25, 2013 at 1:10 PM, Nicolas Pitre <nico@xxxxxxxxxxx> wrote:
> > On Fri, 25 Jan 2013, Josh Boyer wrote:
> >
> >> On Fri, Jan 25, 2013 at 11:22 AM, Dennis Gilmore <dennis@xxxxxxxx> wrote:
> >> > Hi all,
> >> >
> >> > I wanted to kick off a discussion, I think that with the work that
> >> > Seneca is doing for armv6hl to support the Raspberry Pi most of the
> >> > need for building sfp has gone away. I would like us to drop support
> >> > for sfp in F19 that means that anyone running a kirkwood based system
> >> > would get supported software updates for approximately 13 months from
> >> > now. with cubie boards and other devices coming around that are cheap
> >> > and more powerful and similar options I think there is little benefit
> >> > to continuing to support sfp.
> >>
> >> I'm not overly familiar with arm, but from a kernel standpoint you might
> >> be able to enable floating point emulation. That would let you run the
> >> hardfp binaries on the boards without an FPU.
> >
> > No, you can't.
> 
> OK.
> 
> > The only FP emulation the ARM kernel provide is for the antique FPA
> > instruction set that almost never got implemented in hardware, except
> > for a few exceptions that Linux never supported anyway.
> 
> Ah.  See, that would be where I point back to me knowning almost nothing
> about ARM ;).
> 
> > All modern EABI compliant binaries are using the VFP instruction set,
> > and the only thing the kernel implements is the processing of
> > exceptions for them.  To have a full VFP emulation support in the
> > kernel, significant development effort would be needed.
> 
> I find this slightly interesting to be honest.  In the embedded powerpc
> world, math emulation is used as a last ditch effort but it does exist.
> I would have thought the proliferation of ARM devices would generate some
> amount of interest in getting similar support in-kernel, but apparently
> not.

Thing is: a soft float library in user space is 8 to 25 times faster 
than any kernel emulation of machine FP instructions.  I know that 
as I wrote that soft-float library myself.  ;-)

Before EABI, it was common to have user space binaryes compiled for FPA 
and those instructions were always emulated by the kernel, since, as I 
said, there were no actual FPA hardware that Linux supported.

When EABI was introduced, it was decided to go with the soft-float 
library given the ABI was different anyway and performances were much 
better than any emulation.

That was before VFP became a reality.  With VFP hardware you then had 
the ability to still use the same procedure call convention but the 
ability to implement the intra procedure code using actual VFP 
instructions.  That is what most ARMv6+ binary distributions did prior 
the HF variant to remain compatible with the soft-float ABI.

But given that hardware VFP is now prevalent, going with full FP even at 
the function call level was implemented and deployed.



Nicolas
_______________________________________________
arm mailing list
arm@xxxxxxxxxxxxxxxxxxxxxxx
https://admin.fedoraproject.org/mailman/listinfo/arm



[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Index of Archives]     [Linux ARM (Vger)]     [Linux ARM]     [ARM Kernel]     [Fedora User Discussion]     [Older Fedora Users Discussion]     [Fedora Advisory Board]     [Fedora Security]     [Fedora Maintainers]     [Fedora Devel Java]     [Fedora Legacy]     [Fedora Desktop]     [ATA RAID]     [Fedora Marketing]     [Fedora Mentors]     [Fedora Package Announce]     [Fedora Package Review]     [Fedora Music]     [Fedora Packaging]     [Centos]     [Fedora SELinux]     [Coolkey]     [Yum Users]     [Tux]     [Yosemite News]     [Linux Apps]     [KDE Users]     [Fedora Tools]     [Fedora Art]     [Fedora Docs]     [Asterisk PBX]

Powered by Linux