[x265] Questions about parallelism

[Nicolas Morey-Chaisemartin]

On 02/02/2015 11:43 AM, Deepthi Nandakumar wrote:
>
>
> We ran into this exact issue doing performance analyses on 40-core+ machines, where we found slicetypeDecide being a bottleneck. Please checkout this commit.
>
> Changeset:
>
> 	
>
> 9033 (d36211d0190f) slicetype: allow queue to fill past full to prevent bottlenecks …
>
> User:
>
> 	
>
> Steve Borho <steve at borho.org <mailto:steve at borho.org>>
>
> Date:
>
> 	
>
> 2015-01-06 15:38:58 +0530 (3 weeks)
>
>
> This patch essentially lets lookahead run slightly ahead of the Frame Encoders, so the output queue always has decided frames available for encoding.
>
I was looking how my quick patch would conflicts against master and found out this one. Which is pretty much what I did.

> We're very close to tagging 1.5 - which will contain this and other improvements.
This is good. I've been waiting to update for a while but we had to release some code last week. Hopefully it won't be too much of a hassle to integrate.

>
>     The idea would be something like this:
>     Input -> Input Thread <-> LookAhead queue <-> LookAhead Thread <-> input_queue <-> Main thread <-> FrameEncoders <-> Main Thread -> Output stream
>
> We could try this out, but on generic x86 multicore, it may not benefit much. It would be good to see if the current tip fixes the performance issues on your accelerator pipeline as well.

I'll give a try at merging the code and see where we land.

Thanks for the feedback.

Nicolas

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150202/3d645ed3/attachment.html>