armhf support in OpenEmbedded?

armhf support in OpenEmbedded?

[Leon Woestenberg]

[-- Attachment #1: Type: text/plain, Size: 342 bytes --]

Hello,

do we already have support for the ARM "armhf" ABI in OpenEmbedded?  (A
quick search on the ML didn't get me hits).

Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a
CPUs.


http://wiki.debian.org/ArmHardFloatPort
http://www.powerdeveloper.org/forums/viewtopic.php?p=13645#13645

Regards,

Leon.

[-- Attachment #2: Type: text/html, Size: 529 bytes --]

Re: armhf support in OpenEmbedded?

[Mark Hatle]

In oe-core, there are definitions for various tunings that use the ARM EABI w/ 
HF extension.

The following tunings support that mode:  (And armhf is avsolutely not required 
to be for the armv7-a CPU.. it compatible with any CPU that supports VFP 
instructions... note there are armv7-a systems that do -not- have VFP support, 
so these ABI extensions will not work there!)

armv5ehf-vfp armv5tehf-vfp
armv5ehfb-vfp armv5tehfb-vfp
armv5hf-vfp armv5thf-vfp
armv5hfb-vfp armv5thfb-vfp
armv6hf armv6thf
armv6hfb armv6thfb
armv7hf armv7thf armv7hf-neon armv7thf-neon
armv7ahfb armv7athfb armv7ahfb-neon armv7athfb-neon
armv7hf armv7thf armv7hf-neon armv7thf-neon
armv7hfb armv7thfb armv7hfb-neon armv7thfb-neon

See conf/machine/include/arm/*

--Mark

On 2/6/12 3:01 PM, Leon Woestenberg wrote:
> Hello,
>
> do we already have support for the ARM "armhf" ABI in OpenEmbedded? (A quick
> search on the ML didn't get me hits).
>
> Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a CPUs.
>
>
> http://wiki.debian.org/ArmHardFloatPort
> http://www.powerdeveloper.org/forums/viewtopic.php?p=13645#13645
>
> Regards,
>
> Leon.
>
>
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core

Re: armhf support in OpenEmbedded?

[Koen Kooi]

Op 6 feb. 2012, om 22:01 heeft Leon Woestenberg het volgende geschreven:

> Hello,
> 
> do we already have support for the ARM "armhf" ABI in OpenEmbedded?  (A quick search on the ML didn't get me hits).
> 
> Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a CPUs.

Yes, we've had it for ages in OE-classic and Marks tune overhaul made it easy to use in OE-core as well. Also note that armhf has practically no adavantage over softfp in real world applications. The only things getting a decent speedup in proper benchmarks was povray.

regards,

Koen

Re: armhf support in OpenEmbedded?

[Mark Hatle]

On 2/6/12 3:17 PM, Koen Kooi wrote:
>
> Op 6 feb. 2012, om 22:01 heeft Leon Woestenberg het volgende geschreven:
>
>> Hello,
>>
>> do we already have support for the ARM "armhf" ABI in OpenEmbedded?  (A quick search on the ML didn't get me hits).
>>
>> Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a CPUs.
>
> Yes, we've had it for ages in OE-classic and Marks tune overhaul made it easy to use in OE-core as well. Also note that armhf has practically no adavantage over softfp in real world applications. The only things getting a decent speedup in proper benchmarks was povray.

I strongly discourage people from using it.  If you don't care about ABI 
compatibility it might be a few cycles faster.. but it does limit your ability 
to draw from existing EABI software.  (Be it commercial or open source.)

(And yes, I realize this is a minor concern for most embedded systems.. but it 
is an issue.)

--Mark

> regards,
>
> Koen
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core

Re: armhf support in OpenEmbedded?

[Leon Woestenberg]

[-- Attachment #1: Type: text/plain, Size: 1107 bytes --]

Hello,

On Mon, Feb 6, 2012 at 10:22 PM, Mark Hatle <mark.hatle@windriver.com>wrote:

> On 2/6/12 3:17 PM, Koen Kooi wrote:
>
>>
>> Op 6 feb. 2012, om 22:01 heeft Leon Woestenberg het volgende geschreven:
>>
>>  do we already have support for the ARM "armhf" ABI in OpenEmbedded?  (A
>>> quick search on the ML didn't get me hits).
>>> Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a
>>> CPUs.
>>>
>>
>> Yes, we've had it for ages in OE-classic and Marks tune overhaul made it
>> easy to use in OE-core as well. Also note that armhf has practically no
>> adavantage over softfp in real world applications. The only things getting
>> a decent speedup in proper benchmarks was povray.
>>
>
> I strongly discourage people from using it.  If you don't care about ABI
> compatibility it might be a few cycles faster.. but it does limit your
> ability to draw from existing EABI software.  (Be it commercial or open
> source.)
>

Can someone please lift that large rock I have been under for ages?

Thanks!

Leon


/me climbing back into the OE trenches

[-- Attachment #2: Type: text/html, Size: 1631 bytes --]

Re: armhf support in OpenEmbedded?

[Koen Kooi]

Op 6 feb. 2012, om 22:22 heeft Mark Hatle het volgende geschreven:

> On 2/6/12 3:17 PM, Koen Kooi wrote:
>> 
>> Op 6 feb. 2012, om 22:01 heeft Leon Woestenberg het volgende geschreven:
>> 
>>> Hello,
>>> 
>>> do we already have support for the ARM "armhf" ABI in OpenEmbedded?  (A quick search on the ML didn't get me hits).
>>> 
>>> Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a CPUs.
>> 
>> Yes, we've had it for ages in OE-classic and Marks tune overhaul made it easy to use in OE-core as well. Also note that armhf has practically no adavantage over softfp in real world applications. The only things getting a decent speedup in proper benchmarks was povray.
> 
> I strongly discourage people from using it.  If you don't care about ABI compatibility it might be a few cycles faster..

The issue with hf isn't technical, it's that projects like debian and ubuntu are hyping it. Debian tells everyone how much faster hf is, neglicting to say that their previous tune is armv4t. So from softfp-armv4t to hardfp-armv7a...

> but it does limit your ability to draw from existing EABI software.  (Be it commercial or open source.)
> 
> (And yes, I realize this is a minor concern for most embedded systems.. but it is an issue.)

I discussed that at length during my job interview at TI about 3 years ago, we're finally seeing the problem in real life :)


regards,

Koen

Re: armhf support in OpenEmbedded?

[Philip Balister]

On 02/06/2012 04:17 PM, Koen Kooi wrote:
> 
> Op 6 feb. 2012, om 22:01 heeft Leon Woestenberg het volgende geschreven:
> 
>> Hello,
>>
>> do we already have support for the ARM "armhf" ABI in OpenEmbedded?  (A quick search on the ML didn't get me hits).
>>
>> Basically,"armhf" is built with -mfloat-abi=hard, and tuned for armv7-a CPUs.
> 
> Yes, we've had it for ages in OE-classic and Marks tune overhaul made it easy to use in OE-core as well. Also note that armhf has practically no adavantage over softfp in real world applications. The only things getting a decent speedup in proper benchmarks was povray.

From what I saw, the big win for hf is for functions that return floats.
I try to avoid writing functions that return floats and use pointers to
the variable :)

Philip

Re: armhf support in OpenEmbedded?

[Phil Blundell]

On Mon, 2012-02-06 at 22:01 +0100, Leon Woestenberg wrote:
> do we already have support for the ARM "armhf" ABI in OpenEmbedded?
> (A quick search on the ML didn't get me hits).

Yes, oe-core supports hard-float tunings, see the various TUNE_FEATUREs
defined in meta/conf/machine/include/tune*.inc.  It's never been
entirely obvious whether this is actually very much of a win in practice
though, unless your workload is extraordinarily FP-intensive.

p.