A small analysis of vector composition results on different architectures

[Dmitry Kazakov]

Hi!

First, thanks everyone who helped me with running the composition benchmark
on their machines! Now I have some important information about what we
should do next with these optimizations!

In the first attached document you can find a table, comparing the results
on various architectures.

Here are conclusions I came to:

1) On some architectures, e.g. non-Intel, the usage of Vc may degrade the
performance, so Krita should have some automatic benchmarking capabilities
(in settings dialog, for example), which would select the best available
implementation of the composite op. Such benchmark might be executed on the
first run of Krita and its result might be written to a Krita config file.

2) Some old processors (Nehalem, Merom) have some problems when one of the
pixels (src or dst) have alpha channel set to 255. I can't explain right
now why this happens. Probably, there is some branch misprediction... The
rest of the cases on these CPUs are handled properly.

So in general, I  think these two problems should be solved and after that
the branch can be merged into master.


-- 
Dmitry Kazakov
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kimageshop/attachments/20121104/ef23d884/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Krita_Vc_Arch_Comparison.pdf
Type: application/pdf
Size: 33477 bytes
Desc: not available
URL: <http://mail.kde.org/pipermail/kimageshop/attachments/20121104/ef23d884/attachment-0001.pdf>