[FFmpeg-devel] [PATCH] frame_thread_encoder: make task indexing deterministic.

[Michael Niedermayer]

On Sat, Apr 01, 2017 at 06:34:50PM -0400, Ronald S. Bultje wrote:
> Hi,
> 
> On Sat, Apr 1, 2017 at 5:44 PM, Michael Niedermayer <michael at niedermayer.cc>
> wrote:
> 
> > On Sat, Apr 01, 2017 at 08:50:37AM -0400, Ronald S. Bultje wrote:
> > > Hi,
> > >
> > > On Sat, Apr 1, 2017 at 6:15 AM, Michael Niedermayer
> > <michael at niedermayer.cc>
> > > wrote:
> > > [.]
> > > >
> > > > about the patch itself, the changes done to indexing seem not to
> > > > change anything, it makes it possible for the variables to overflow
> > > > though.
> > > >
> > >
> > > That's intentional, they're both unsigned and the one is guaranteed to be
> > > larger than the other so the difference after subtract would still be
> > > guaranteed to be <= n_threads. So although the indices overflow, the
> > > subtraction will always be correct.
> > >
> > >
> > > > IIUC the only change your patch does is to remove the outdata check
> > > > from the quoted warning
> > > >
> > >
> > > Yes, because if you change the delay to be fixed (and equal to
> > n_threads),
> > > you no longer need the check. If the difference between indices is
> > smaller
> > > than n_threads, you don't return a packet. If it's equal, you do return a
> > > packet.
> > >
> >
> > > its a while ago that i worked n this code but isnt this just "missing"
> > > > a finished_task_mutex lock over the access ?
> > >
> > >
> > > I don't know, it depends on what you're trying to accomplish. If you
> > want a
> > > fixed-frame delay, you don't need the access and so no mutex is needed. I
> > > haven't thought much about what you need to get a variable-frame delay.
> > To
> > > me, a fixed-frame delay for intra-only codecs (which is what this API
> > > currently accomplishes in practice) is exactly what you would expect,
> > > variable-frame delay implies you're not actually using all threads. But
> > > maybe I misunderstand.
> >
> > Lower delay is better, so is using fewer threads if theres no need for
> > more
> > it means less delay for any form of real time commuication
> > less memory used, less L2 cache used and consequently better use of
> > the CPU.
> > If decoding with 1 thread is fast enough for the rate at which packets
> > become available using 12 threads will result in more cpu cycles per
> > frame, more memory used, more watts, a higher energy bill, ...
> 
> 

> We already do automatic thread detection, and it's a user-settable option
> also.

automatic thread detection detects the maximum that might be usefull
on the hardware, not the optimal

Allowing the user to set the number of threads means the user can
set the number but the user does not know the best number for each
file, the number also changes at runtime.
But even if it didnt change and the user knew the best it would be
very inconvenient for the user to set it by hand
Its better if it just uses the optimal automatically


> 
> Can we please review the patch?

This comment feels offensive
I reviewed your patch, or at least tried to.

"... the changes done to indexing seem not to change anything, ..."
If the changes are not needed, please remove them from the patch or
if they are neeed please explain why

having an index count to UINT_MAX is confusing, i prefer it to stay
within 0..BUFFER_SIZE.


'...isnt this just "missing" a finished_task_mutex lock over the access ?'

If it just is missing a lock, adding this lock is the fix i would
prefer.
Of course you can as well use an atomic access or somethig else that
keeps tsan from complaining.

Supporting changing delay and number of threads at runtime was a
big factor in the design of the frame thread encoder. Its something
i cared about, so i am not to positive towards patches removing that.

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Opposition brings concord. Out of discord comes the fairest harmony.
-- Heraclitus
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20170402/329bf750/attachment.sig>