[Bug 747216] applemedia: implement GstAppleCoreVideoMemory

[GStreamer (GNOME Bugzilla)]

https://bugzilla.gnome.org/show_bug.cgi?id=747216

--- Comment #41 from Alessandro Decina <alessandro.d at gmail.com> ---
(In reply to Nicolas Dufresne (stormer) from comment #39)
> Ok, does this still apply cleanly after Alessandro ? Is Alessandro
> interested in discussing his work or will he always push stuff at random
> point in time without further discussion with the others ?

Uh, what's with the polemical tone? :)

Texture sharing in vtdec was disabled since September
(f02425c4afcd85260a1b387aeddf863774257917). Turned out that the fix involved
switching to IOSurface (8ae003326157438c12c45589e050c5f446723f61), which is
something that has actually been discussed since the IOSurface framework was
introduced (years - in the meantime safari, chrome and firefox all switched to
it). I finally did it. I was aware of this patchset of course but it has
nothing to do with what I did so I didn't say anything here.

Now as for this patchset, I'm personally +0 on landing it _as it is_ (and the
reason why I haven't landed it) for the following reasons:

- performance: while the code is well written (although I think it's too
complex in places, like the locking logic), it doesn't _practically_ improve
anything performance wise. Its purpose is to delay mapping pixel buffers, but
for all practical purposes pixel buffers get _always_ mapped except that in the
avfvideosrc ! vtenc case, so it doesn't really improve anything. avfvideosrc ! 
vtenc as it is today in master already avoids mapping pixel buffers using the
CoreMedia meta (which I don't think is ideal, see below);

- simplicity: we all agree with the GstMemory approach, but this code also
keeps the metas around. I wrote the metas in my 1-day port to 1.x back in the
day.  In retrospect that was a poor choice and the metas should likely be
removed now. We've been talking about _at least_ unifying the CoreVideo and
CoreMedia metas for a while. In fact the whole coremedia/corevideo buffer stuff
should probably just be one thing;

- mainteinance: say we merge this as is (since it's got tests and it's probably
good enough to work), it looks to me like we're adding code without improving
performance and without reducing debt. Since Ilya is not interested, I will
have to be the one to clean up the metas/buffer implementations anyway and if
we land this I will have to update even more code

-- 
You are receiving this mail because:
You are the QA Contact for the bug.