[PATCH 0/1] beagleboard: enable hard floating point abi

[PATCH 0/1] beagleboard: enable hard floating point abi

[Darren Hart]

There have been requests to enable the hard floating point abi for the
Yocto Project Beagleboard BSP. With the recent cortexa8 tune file changes
I submitted to oe-core, subject "tune: update arm7a and cortexa8 tune files",
it is now possible for the machine config to change DEFAULTTUNE to take
advantage of the newly added "hf" tune variants.

Some binary libraries are incompatible with hardfp, as is 3D support on the
Beagleboard. We don't ship either, so I feel it makes sense to take advantage
of as much of the core CPU features as possible.

This addresses [YOCTO 1203]
http://bugzilla.yoctoproject.org/show_bug.cgi?id=1204

I don't feel very strongly about this however. So if someone would like to
argue for hardfp not being the default, please do so and I can easily
comment out the DEFAULTTUNE change and add a comment in the machine config
describing how to enable hardfp.

The following changes since commit 53bdf3fd520b75730cc802f52121dd394894e130:

  tune: Add hard floating point variants of cortexa8 tunes (2011-08-19 15:02:10 -0700)

are available in the git repository at:
  git://git.pokylinux.org/poky-contrib dvhart/beagleboard-hardfp
  http://git.pokylinux.org/cgit.cgi/poky-contrib/log/?h=dvhart/beagleboard-hardfp

Darren Hart (1):
  beagleboard: enable hard floating point abi

 meta-yocto/conf/machine/beagleboard.conf |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)

-- 
1.7.6

[PATCH 1/1] beagleboard: enable hard floating point abi

[Darren Hart]

Fixes [YOCTO #1203]

Using the hard floating point abi is incompatible with some binary libaries and
3D support for the Beagleboard. As we do not provide these in poky and
meta-yocto, we can take advantage of the hard floating point abi. This can be
easily disabled by anyone building on top of meta-yocto if they need the soft
abi by overriding DEFAULT_TUNE.

Signed-off-by: Darren Hart <dvhart@linux.intel.com>
CC: Jason Kridner <jkridner@beagleboard.org>
CC: Koen Kooi <koen@dominion.thruhere.net>
---
 meta-yocto/conf/machine/beagleboard.conf |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)

diff --git a/meta-yocto/conf/machine/beagleboard.conf b/meta-yocto/conf/machine/beagleboard.conf
index 0b3cebc..8f11732 100644
--- a/meta-yocto/conf/machine/beagleboard.conf
+++ b/meta-yocto/conf/machine/beagleboard.conf
@@ -18,6 +18,11 @@ MACHINE_EXTRA_RRECOMMENDS += "beagleboard-audio"
 # Allow for MMC booting (required by the NAND-less Beagleboard XM)
 EXTRA_IMAGEDEPENDS += "u-boot x-load"
 
+# Use the hard floating point abi. This breaks some binary libraries and
+# 3D. Since poky and meta-yocto do not provide these by default, enable the
+# hard floating point abi. Comment out the following or override it to disable
+# it.
+DEFAULTTUNE ?= "cortexa8hf-neon"
 include conf/machine/include/tune-cortexa8.inc
 
 IMAGE_FSTYPES += "tar.bz2 jffs2"
-- 
1.7.6

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Jason Kridner]

On Fri, Aug 19, 2011 at 6:23 PM, Darren Hart <dvhart@linux.intel.com> wrote:
> Fixes [YOCTO #1203]
>
> Using the hard floating point abi is incompatible with some binary libaries and
> 3D support for the Beagleboard. As we do not provide these in poky and
> meta-yocto, we can take advantage of the hard floating point abi. This can be
> easily disabled by anyone building on top of meta-yocto if they need the soft
> abi by overriding DEFAULT_TUNE.
>
> Signed-off-by: Darren Hart <dvhart@linux.intel.com>
> CC: Jason Kridner <jkridner@beagleboard.org>
> CC: Koen Kooi <koen@dominion.thruhere.net>
> ---
>  meta-yocto/conf/machine/beagleboard.conf |    5 +++++
>  1 files changed, 5 insertions(+), 0 deletions(-)
>
> diff --git a/meta-yocto/conf/machine/beagleboard.conf b/meta-yocto/conf/machine/beagleboard.conf
> index 0b3cebc..8f11732 100644
> --- a/meta-yocto/conf/machine/beagleboard.conf
> +++ b/meta-yocto/conf/machine/beagleboard.conf
> @@ -18,6 +18,11 @@ MACHINE_EXTRA_RRECOMMENDS += "beagleboard-audio"
>  # Allow for MMC booting (required by the NAND-less Beagleboard XM)
>  EXTRA_IMAGEDEPENDS += "u-boot x-load"
>
> +# Use the hard floating point abi. This breaks some binary libraries and
> +# 3D. Since poky and meta-yocto do not provide these by default, enable the
> +# hard floating point abi. Comment out the following or override it to disable
> +# it.
> +DEFAULTTUNE ?= "cortexa8hf-neon"

I'm not really fond of what this breaks, but I don't yet have a vision
of what I'd be doing with meta-yocto yet and therefore am open to this
project giving test coverage to new conditions like this.

>  include conf/machine/include/tune-cortexa8.inc
>
>  IMAGE_FSTYPES += "tar.bz2 jffs2"
> --
> 1.7.6
>
>

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Koen Kooi]

Op 20 aug. 2011, om 00:23 heeft Darren Hart het volgende geschreven:

> Fixes [YOCTO #1203]
> 
> Using the hard floating point abi is incompatible with some binary libaries and
> 3D support for the Beagleboard. As we do not provide these in poky and
> meta-yocto, we can take advantage of the hard floating point abi. 

What advantage are you talking about? So far everyone has been unable to provide real-world numbers[1] that show hardfp making a difference compared to a properly configured softfp. The numbers debian and meego are showing are comparing it against completely vfpless builds, which borders on fraud.

Yocto is of course free to do what it pleases, but switching to hardfp isn't going to make beagle support better, only worse. But if that's your goal, go for it, meta-ti is still there to provide the absolute best beagle support.

[1] povray is the only outlier, but I don't think people will build beagle based renderfarms in the near future.

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Darren Hart]

On 08/23/2011 05:38 AM, Koen Kooi wrote:
> 
> Op 20 aug. 2011, om 00:23 heeft Darren Hart het volgende geschreven:
> 
>> Fixes [YOCTO #1203]
>> 
>> Using the hard floating point abi is incompatible with some binary
>> libaries and 3D support for the Beagleboard. As we do not provide
>> these in poky and meta-yocto, we can take advantage of the hard
>> floating point abi.
> 
> What advantage are you talking about? So far everyone has been unable
> to provide real-world numbers[1] that show hardfp making a difference
> compared to a properly configured softfp. The numbers debian and
> meego are showing are comparing it against completely vfpless builds,


That's good reasoning to stick with softfp+neon. Unfortunately I can't
find the mail threads that first got me looking into adding hardfp
support. As I said, I'm not sold on the idea, but it was requested so I
looked into how to address it.

If nobody comes forward saying they would really like to have this, I'm
going to modify the patch series to disable hardfp by default, but leave
the infrastructure in place so people can enable if they like.

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Tom Rini]

On Tue, Aug 23, 2011 at 8:31 AM, Darren Hart <dvhart@linux.intel.com> wrote:
> On 08/23/2011 05:38 AM, Koen Kooi wrote:
>>
>> Op 20 aug. 2011, om 00:23 heeft Darren Hart het volgende geschreven:
>>
>>> Fixes [YOCTO #1203]
>>>
>>> Using the hard floating point abi is incompatible with some binary
>>> libaries and 3D support for the Beagleboard. As we do not provide
>>> these in poky and meta-yocto, we can take advantage of the hard
>>> floating point abi.
>>
>> What advantage are you talking about? So far everyone has been unable
>> to provide real-world numbers[1] that show hardfp making a difference
>> compared to a properly configured softfp. The numbers debian and
>> meego are showing are comparing it against completely vfpless builds,
>
>
> That's good reasoning to stick with softfp+neon. Unfortunately I can't
> find the mail threads that first got me looking into adding hardfp
> support. As I said, I'm not sold on the idea, but it was requested so I
> looked into how to address it.
>
> If nobody comes forward saying they would really like to have this, I'm
> going to modify the patch series to disable hardfp by default, but leave
> the infrastructure in place so people can enable if they like.

Well, are there some softfp related config bits we need on the yocto
side that meta-ti has, in order to bring that performance back in
line?

-- 
Tom

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Darren Hart]

On 08/23/2011 08:34 AM, Tom Rini wrote:
> On Tue, Aug 23, 2011 at 8:31 AM, Darren Hart <dvhart@linux.intel.com> wrote:
>> On 08/23/2011 05:38 AM, Koen Kooi wrote:
>>>
>>> Op 20 aug. 2011, om 00:23 heeft Darren Hart het volgende geschreven:
>>>
>>>> Fixes [YOCTO #1203]
>>>>
>>>> Using the hard floating point abi is incompatible with some binary
>>>> libaries and 3D support for the Beagleboard. As we do not provide
>>>> these in poky and meta-yocto, we can take advantage of the hard
>>>> floating point abi.
>>>
>>> What advantage are you talking about? So far everyone has been unable
>>> to provide real-world numbers[1] that show hardfp making a difference
>>> compared to a properly configured softfp. The numbers debian and
>>> meego are showing are comparing it against completely vfpless builds,
>>
>>
>> That's good reasoning to stick with softfp+neon. Unfortunately I can't
>> find the mail threads that first got me looking into adding hardfp
>> support. As I said, I'm not sold on the idea, but it was requested so I
>> looked into how to address it.
>>
>> If nobody comes forward saying they would really like to have this, I'm
>> going to modify the patch series to disable hardfp by default, but leave
>> the infrastructure in place so people can enable if they like.
> 
> Well, are there some softfp related config bits we need on the yocto
> side that meta-ti has, in order to bring that performance back in
> line?

Are you referring to a specific performance measure?

Jason, Koen, are there any such config bits? We are currently using the
cortexa8-neon tune configuration from oe-core.

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Tom Rini]

On Tue, Aug 23, 2011 at 11:55 AM, Darren Hart <dvhart@linux.intel.com> wrote:
> On 08/23/2011 08:34 AM, Tom Rini wrote:
>> On Tue, Aug 23, 2011 at 8:31 AM, Darren Hart <dvhart@linux.intel.com> wrote:
>>> On 08/23/2011 05:38 AM, Koen Kooi wrote:
>>>>
>>>> Op 20 aug. 2011, om 00:23 heeft Darren Hart het volgende geschreven:
>>>>
>>>>> Fixes [YOCTO #1203]
>>>>>
>>>>> Using the hard floating point abi is incompatible with some binary
>>>>> libaries and 3D support for the Beagleboard. As we do not provide
>>>>> these in poky and meta-yocto, we can take advantage of the hard
>>>>> floating point abi.
>>>>
>>>> What advantage are you talking about? So far everyone has been unable
>>>> to provide real-world numbers[1] that show hardfp making a difference
>>>> compared to a properly configured softfp. The numbers debian and
>>>> meego are showing are comparing it against completely vfpless builds,
>>>
>>>
>>> That's good reasoning to stick with softfp+neon. Unfortunately I can't
>>> find the mail threads that first got me looking into adding hardfp
>>> support. As I said, I'm not sold on the idea, but it was requested so I
>>> looked into how to address it.
>>>
>>> If nobody comes forward saying they would really like to have this, I'm
>>> going to modify the patch series to disable hardfp by default, but leave
>>> the infrastructure in place so people can enable if they like.
>>
>> Well, are there some softfp related config bits we need on the yocto
>> side that meta-ti has, in order to bring that performance back in
>> line?
>
> Are you referring to a specific performance measure?

Just the general ones that have caused people to say softfp+neon isn't
enough (outside of povray :)).

-- 
Tom

Re: [PATCH 1/1] beagleboard: enable hard floating point abi

[Philip Balister]

On 08/23/2011 05:38 AM, Koen Kooi wrote:
>
> Op 20 aug. 2011, om 00:23 heeft Darren Hart het volgende geschreven:
>
>> Fixes [YOCTO #1203]
>>
>> Using the hard floating point abi is incompatible with some binary libaries and
>> 3D support for the Beagleboard. As we do not provide these in poky and
>> meta-yocto, we can take advantage of the hard floating point abi.
>
> What advantage are you talking about? So far everyone has been unable to provide real-world numbers[1] that show hardfp making a difference compared to a properly configured softfp. The numbers debian and meego are showing are comparing it against completely vfpless builds, which borders on fraud.
>
> Yocto is of course free to do what it pleases, but switching to hardfp isn't going to make beagle support better, only worse. But if that's your goal, go for it, meta-ti is still there to provide the absolute best beagle support.
>
> [1] povray is the only outlier, but I don't think people will build beagle based renderfarms in the near future.

The big win for hard float ABI (emphasis mine) is functions that return 
floating point values. I personally care about this, but we also know 
returning floats is "bad" so we sort of avoid it.

I agree with Phil Blundell that this is not a beagle specific thing.

Long term I expect hard float to become normal, but I see no need to 
rush there (no matter what other distros are claiming).

Philip