[FFmpeg-devel] [PATCH] simplify dnxhd threading

[Baptiste Coudurier]

On 09/25/2009 11:36 AM, Reimar D?ffinger wrote:
> On Fri, Sep 25, 2009 at 10:51:36AM -0700, Baptiste Coudurier wrote:
>> 68 contexts, even when not using threads ? I'm concerned about memory.
>
> Yes, that probably should be checked. Though I think that giving each thread
> its own whole context is overdoing it anyway.
>
>> I'd like to see extended benchmarks with 8+ threads before
>> considering it, if it's not faster, I don't really see the point.
>
> Well, do you by chance know a commandline where it is even faster with -threads 2?

You have interlaced mode: -flags +ildct, bitrate: 120Mb or 185Mb for 
1080 25fps.

Here -threads 2 seems to be faster on my core 2 duo.
On quadcore -threads 4 achieves real time encoding at 1080, the encoder 
is basically used to capture real time from HD-SDI link.

> There was no difference between -threads 2 and -threads 1 in my tests.

Well, given high bitrates involved here, results tend to vary, indeed.
I usually test with -f null output toi avoid actually writing the data, 
because my hard drive cannot keep up reading and writing.

> Also note that I explicitly pointed out that my patch can speed up things
> even with only 2 threads if there are other tasks running or encoding takes
> very different time for different parts of the image (no idea if that can happen
> with DNxHD).
> That said, feel free to consider this more an RFC than a patch to be applied.

Sure, I kinda like the approach if it's faster.

-- 
Baptiste COUDURIER
Key fingerprint                 8D77134D20CC9220201FC5DB0AC9325C5C1ABAAA
FFmpeg maintainer                                  http://www.ffmpeg.org