render improvements

[Lars Knoll]

On Tuesday 14 June 2005 12:23, David Reveman wrote:
> On Mon, 2005-06-13 at 10:12 +0200, Lars Knoll wrote:
> > On Wednesday 08 June 2005 22:01, David Reveman wrote:
> > > On Wed, 2005-06-08 at 18:00 +0200, Lars Knoll wrote:
> > > > On Wednesday 08 June 2005 14:44, David Reveman wrote:
> > > > > Hi,
> > > > >
> > > > > I just realized that the render improvements you committed to
> > > > > xserver (fbcompose.c) on the 27th of May, broke, or rather removed,
> > > > > the software convolution filter code I had in there since some time
> > > > > ago. Do you think you can fix that?
> > > >
> > > > That should not be too hard. It just got forgotten. I'll try to add
> > > > it in again over the weekend (as well as finish the gradients).
> > >
> > > Good, I could do it myself but as you wrote the new code I thought it'd
> > > be easier for you to do it. I realized the missing convolution code as
> > > I was trying to merge my transparent/pad/reflect picture extend code,
> > > which need some more work but I'll send a patch to the xorg list when
> > > this is done so you can use it for your gradients.
> >
> > Ok, I just submitted it to xserver. I'll submit it to xorg a bit later,
> > once I got it built and tested.
>
> Thanks for spending some time on getting this working again.
>
> I saw that you changed so that the ConvolutionFilter is added by
> default, I'm not sure that we want to do that without verifying that all
> DDX's are aware of it. But I think kdrive's ATI driver is the only one
> that accelerate transformations except Xgl so we could probably just
> make sure that it doesn't try to accelerate it as Nereast or Bilinear
> filters and it should be fine.

Not adding it by default means it will not be usable. The driver should check 
that he supports the filter that is set and fall back to software otherwise. 
Everything else is a bug in the driver IMO and should be fixed there.

> I tried the ConvolutionFilter but it doesn't seem to work... while
> trying this I also realized that there's seems to be issues with
> transformations in general and the new fbcompose.c implementation.

Strange. I tried it both with Xephyr and Xorg CVS and it worked fine for me. 
Do you have an example that fails?

> The rendertest program available from cairo CVS is exposing a lot of
> problems. You'll need XCB for building the x protocol version of
> rendertest and you can easily verify the output by building the
> libpixman version of it as well.

Ok, I'll check it out tomorrow.

> I haven't had the time to get familiar with the new fbcompose.c
> implementation yet so I can't easily see where the problems is right
> now.

It should be easy to fix. All the tests I had seemed to work fine. 

The problem is that we currently lack a good test suite for Render, that 
covers the complete specs (or at least most of it). We should try to put 
something together there as it's also needed for everyone that wants to 
implement drivers that accelerate Render.

Cheers,
Lars