[cairo] [cairo-announce] [ANNOUNCE] pixman release 0.15.20 (Release Candidate) now available

[Siarhei Siamashka]

On Wednesday 12 August 2009, Soeren Sandmann wrote:
> Siarhei Siamashka <siarhei.siamashka at gmail.com> writes:
> > > Btw., I would appreciate if you could try blitter-test and see if it
> > > passes on ARM.
> >
> > It does not. Function 'neon_composite_over_n_8_0565' is not completely
> > identical to C path, we discussed it on irc.
>
> Is this with the blitters-test in the 0.15.20 release?

yes

> There are a bunch of fixes to the C path in 0.15.20, and it's entirely
> possible that the NEON version was right and the C path wrong. That was the
> case for the SSE2 and MMX versions of that function.
>
> > But just disabling this function results in quite a measurable
> > performance regression on text operations.
>
> How much of a difference is it? Just one-bit differences or more than
> that?

The difference is quite serious. I did not look deep into the code, but my
wild guess is that just some color components might be swapped or messed up.
This is the only explanation why it can't be seen easily in real applications
(which mostly use black/white/grayscale fonts).

The attached patch makes the test pass (by using alternative intrinsics part
of code).

-- 
Best regards,
Siarhei Siamashka
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-ARM-a-fix-to-pass-blitters-test-for-neon_composite.patch
Type: text/x-diff
Size: 969 bytes
Desc: not available
Url : http://lists.cairographics.org/archives/cairo/attachments/20090812/438a2f15/attachment-0001.patch