[FFmpeg-user] Zeranoe Windows builds spiking CPU to 100% randomly

[Jan Ehrhardt]

Roger Pack in gmane.comp.video.ffmpeg.user (Wed, 11 Jul 2012 15:13:26
-0600):
>> What still strikes me: on my most troublesome file (50i PAFF interlaced
>> Canon HF M306 recording) the ancient Mencoder-33883 is as quick as or
>> even faster than the fastest FFmpeg. See output below.
>
>That seems odd since I'd guess that mencoder just "calls out" to
>ffmpeg underneath, anyway (or libavcodec or what not).  You could
>compare it to some "ancient" versions of ffmpeg I suppose :)

I compiled the latest Mencoder (r35020) with underneath FFmpeg
ffmpeg-HEAD-d3abbb1 (pthreaded) and ffmpeg-HEAD-d3abbb1 with the same
configs. The speed difference is still there:

https://sessiondatabase.net/jwplayer/dubbel5s.gif

The first encode is a one pass with 2 times the same input file, knitted
together to a 8:08 minute movie. However, the latest Mencoder has
troubles with 2 files as input files: the audio in the second file is
out-of sync. Moreover, this MEncoder has troubles with encoding 50i PAFF
interlaced files to WMV files. So, the second encode does both files
after each other with '-mc 1 -vf scale=448:256,harddup,telecine' and
afterwards I knit them together with asfbin (the 'pause' between the
encodes). The telecine filter was the only way to get a viewable and
hearable result.

But it is far from ideal:

https://sessiondatabase.net/jwplayer/telecine.gif

The still shows a person passing the camera at a short distance. The
whole encoded video shows those stripes. I will report this issue in the
MEncoder users list.

The third encode is an ancient MEncoder. Core from r33986, Stream &
demux_lavc from r34118 (which supports >2GB input files), FFmpeg 0.8.1
from Sherpya's r33883. As you can see, it is almost as fast as the
latest MEncoder. And the transcode is without any issues. No troubles
with more input files, no extra filters needed.

The last encode is ffmpeg-HEAD-d3abbb1 on its own. Sloooow. I know I can
go faster, but used this one for a fair comparison. Andy Sheen's version
N-42432-gb27c7d7 (w32threaded) goes up to 111/112 fps, but even that is
still slower than the ancient MEncoder 33986/34188/33883.

It looks like MEncoder has some internal mechanism for extra threading.

Jan

PS. No, I do not have hyper threaded cores ;-)