Enabling NEON instructions for fftw

Enabling NEON instructions for fftw

[Philip Balister]

So, recent versions of fftw have NEON support for the single precision
build. In the poast, I just passed --enable-neon to configure or the
armv7a case. Now, this is not entirely correct, since some armv7a's lack
a NEON coprocessor.

Does anyone have a suggestion for how to handle this better? Personally,
I think if you want to run fftwf without NEON, you should have your head
examined, but I would like to avoid generating packages that SIGILL on
people.

Philip

Re: Enabling NEON instructions for fftw

[Mark Hatle]

On 1/3/13 12:42 PM, Philip Balister wrote:
> So, recent versions of fftw have NEON support for the single precision
> build. In the poast, I just passed --enable-neon to configure or the
> armv7a case. Now, this is not entirely correct, since some armv7a's lack
> a NEON coprocessor.
>
> Does anyone have a suggestion for how to handle this better? Personally,
> I think if you want to run fftwf without NEON, you should have your head
> examined, but I would like to avoid generating packages that SIGILL on
> people.

This says to me that either you need a machine specific package, or use an 
alternative architecture for that package.. (that is compatible).. then the 
package can switch neon on/off depending on arch of tuning flags.

There is an armv7a-neon tuning defined.  This enabled the TUNE_FEATURES of 
'neon'.  So you should be able to check for that in the fftw recipe, and enable 
the --enable-neon when it's set.  (Using PACKAGECONFIG is likely best...)

Then you can change the DEFAULTTUNE_<recipe> = "armv7-neon" in your local.conf.

--Mark

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

Re: Enabling NEON instructions for fftw

[Philip Balister]

On 01/03/2013 02:21 PM, Mark Hatle wrote:
> On 1/3/13 12:42 PM, Philip Balister wrote:
>> So, recent versions of fftw have NEON support for the single precision
>> build. In the poast, I just passed --enable-neon to configure or the
>> armv7a case. Now, this is not entirely correct, since some armv7a's lack
>> a NEON coprocessor.
>>
>> Does anyone have a suggestion for how to handle this better? Personally,
>> I think if you want to run fftwf without NEON, you should have your head
>> examined, but I would like to avoid generating packages that SIGILL on
>> people.
> 
> This says to me that either you need a machine specific package, or use
> an alternative architecture for that package.. (that is compatible)..
> then the package can switch neon on/off depending on arch of tuning flags.
> 
> There is an armv7a-neon tuning defined.  This enabled the TUNE_FEATURES
> of 'neon'.  So you should be able to check for that in the fftw recipe,
> and enable the --enable-neon when it's set.  (Using PACKAGECONFIG is
> likely best...)
> 
> Then you can change the DEFAULTTUNE_<recipe> = "armv7-neon" in your
> local.conf.

I really do not want to depend on users putting entries in local.conf to
obtain proper operation on a specific machine. It looks like the neon
tune is already selected for the machine I am building for.

In the end, this boils down to, do we care about supporting machines
with/without NEON from one set of packages by making ones with NEON
machine specific.

For now, I will purse the route of fftwf building with --enable if the
neon tune is available for the machine.

Philip

Philip

Philip

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

Re: Enabling NEON instructions for fftw

[Mark Hatle]

On 1/4/13 9:36 AM, Philip Balister wrote:
> On 01/03/2013 02:21 PM, Mark Hatle wrote:
>> On 1/3/13 12:42 PM, Philip Balister wrote:
>>> So, recent versions of fftw have NEON support for the single precision
>>> build. In the poast, I just passed --enable-neon to configure or the
>>> armv7a case. Now, this is not entirely correct, since some armv7a's lack
>>> a NEON coprocessor.
>>>
>>> Does anyone have a suggestion for how to handle this better? Personally,
>>> I think if you want to run fftwf without NEON, you should have your head
>>> examined, but I would like to avoid generating packages that SIGILL on
>>> people.
>>
>> This says to me that either you need a machine specific package, or use
>> an alternative architecture for that package.. (that is compatible)..
>> then the package can switch neon on/off depending on arch of tuning flags.
>>
>> There is an armv7a-neon tuning defined.  This enabled the TUNE_FEATURES
>> of 'neon'.  So you should be able to check for that in the fftw recipe,
>> and enable the --enable-neon when it's set.  (Using PACKAGECONFIG is
>> likely best...)
>>
>> Then you can change the DEFAULTTUNE_<recipe> = "armv7-neon" in your
>> local.conf.
>
> I really do not want to depend on users putting entries in local.conf to
> obtain proper operation on a specific machine. It looks like the neon
> tune is already selected for the machine I am building for.
>
> In the end, this boils down to, do we care about supporting machines
> with/without NEON from one set of packages by making ones with NEON
> machine specific.

Machine specific is one answer, but my preference is to have a neon specific 
package arch selected.  Looking at the definitions, my expectation is that once 
the -neon version of armv7 is selected the package arch should similarly be 
modified to end in "-neon".

So as long as the BSP/machine uses the neon tune, then everything should be 
compiled and labeled properly.

> For now, I will purse the route of fftwf building with --enable if the
> neon tune is available for the machine.

Definitely the right way to do it, no matter how the package's arch is named.

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