[cairo] Pixman ARM Performance

Koen Kooi k.kooi at student.utwente.nl
Fri Jul 24 14:59:29 PDT 2009


On 24-07-09 21:59, Siarhei Siamashka wrote:
> On Friday 24 July 2009, Koen Kooi wrote:
>> On 24-07-09 18:53, Siarhei Siamashka wrote:
>>> Right now we are shooting in the dark and still have
>>> problems (not having -mfloat-abi option is also not good for current
>>> toolchains).
>>
>> I still don't see what the problem is for someone to put their floatabi
>> in CFLAGS.
>
> Because this someone may want to compile all pixman code with -mfloat-abi=soft
> and just 'pixman-arm-neon.c' with -mfloat-abi=softfp.

Does that actually work as intended?

> Why would someone want it? In order to get a binary which can run everywhere
> starting from armv4t, but at the same time could make use of neon
> optimizations where they are supported. Think about binary distributions
> which target lowest common denominator.
>
> I myself also don't need this feature at all. But pixman already has it
> partially working on ARM (and fully working on x86 and PPC). It would be nice
> to either drop cpu features autodetection completely for ARM, or implement it
> in a way which actually works right.
>
>> If setting CFLAGS is too hard for people I wonder why they
>> are trying to build from source anyway.
>
> You just don't need runtime autodetection feature, that's the source of all
> the misunderstanding.

I don't see much use in wasting precious cachelines on stuff that will 
never get executed, I like my binaries nice and tiny on pre-v7a arms :)

regards,

Koen



More information about the cairo mailing list