ORC support for arm64, and compositor performance

[Nicolas Dufresne]

Le jeudi 10 février 2022 à 00:18 +0200, Mart Raudsepp via gstreamer-devel a
écrit :
> Ühel kenal päeval, T, 08.02.2022 kell 09:20, kirjutas David Ing via
> gstreamer-devel:
> > Mart,
> > 
> > I am curious:  What is the highest version of gstreamer that you
> > observed this crash with?
> 
> It was back in March 2021 with GStreamer 1.18.4 and ORC git master at
> the time.
> There have been not many changes to ORC since then, but there is a
> loadoff{b,w,l} commit that was done for performance reasons, so who
> knows, maybe it helps.
> 
> > Also:  It sounds like you verified it was definitely ORC by disabling
> > those optimizations, and running without a crash.  Is that correct?
> 
> The crash was there only after ORC was updated from 0.4.32 to git from
> March 2021 state, otherwise it seemed to work. But due to being orders
> of magnitude slower, it wasn't able to keep up in performance with the
> live pipeline, so triggered other issues.
> 
> The crash backtrace pointed squarely at the new ORC code really.
> It wasn't an instant crash on the first frame, as far as I recall, but
> something that happened quite some time into the pipelines runtime.
> 
> I feel quite bad for not having saved the backtrace and following up on
> it, but I thought I will with a more thorough analysis, but ended up
> having to quickly move back to x86_64 instead and it was left on the
> sidelines.
> 
> So really, just take it as a warning to test thoroughly for your use
> cases and that's all.

I do recall that before some more arm64 opcode was added (somewhere last year),
arm64 support was pretty much always falling back C code, so quite slow. At that
time, the unit test was indeed noticed to crash. But I remember that was worked
on and fixed.

> 
> 
> Mart