[Pixman] [cairo] [PATCH/RFC][pixman] More ARM NEON performance updates

Soeren Sandmann sandmann at daimi.au.dk
Fri Feb 19 09:37:51 PST 2010


> > This all seems way too complex to me and implies that there would be an extra
> > overhead introduced on every image creation.
> > 
> > The branch 'fetch-r5g6b5-arm-neon' has much more simple solution and an extra
> > overhead happens just once at setup time. It is not like CPU features are
> > going to change at runtime 
> The main problem I have with it is that it causes ARM stuff to 'leak'
> out outside of the ARM implementations. It introduces an undocumented
> inter-dependency between implementations: they now have to be created
> in a specific order, or they will overwrite each other's fetchers. 

Thinking some more about this, here is another proposal, which is less

Add some new functions to the implementation struct. Something like

        fetch_scanline_t (* get_scanline_fetcher_32) (...)
        fetch_scanline_t (* get_scanline_fetcher_64 (...)
        store_scanline_t (* get_scanline_storer_32) (...)
        store_scanline_t (* get_scanline_storer_64) (...)

By default these just delegate, and the general implementation will
return the fetchers that are now being set up in the various
property_changed() functions. The property_changed() functions are
extended to also take an implementation argument, so that they can
call the new functions to get the fetchers and storers.

This allows CPU specific fetchers for both sampled images and
gradients, while not being excessively complex.

The downside is that some reorganisation of the code is required. 

The existing fetchers would now conceptually be part of the general
implementation instead of being part of the images. Some fetchers
would probably belong in the fast path implementation.

This means that the image files (pixman-linear-gradient.c etc.) would
have to be renamed to pixman-general-linear-gradient.c and their
constructors moved to pixman-image.c. 

There is probably more reorganisation required, but hopefully it would
be mostly a matter of shuffling code around.


More information about the Pixman mailing list