[Ffmpeg-devel] Threading in libavcodec

[Bryan Mayland]

M?ns Rullg?rd wrote:
>>> Fourth, some random speed results (defaults, vhq4, NO bframes. X2
>>> 4200+ CPU).
>>>
>>> 1 thread : 1:48
>>> 2 threads : 1:05
>>> 3 threads : 1:06
>>> 4 threads : 1:08
>>> 16 threads: 1:11
>>>       
>> So, as I said, it scales pretty well.
>>     
>
> What are those numbers?  Minutes:seconds?  Then it's a far cry from what I'd
> say scales well.  A speedup of 1.66 for two threads is decent but hardly
> impressive.  And before you can talk about scalability you'd need to test
> it on something with a few more CPUs.
>   
    Yeah they are minutes:seconds for encoding his test clip.  The 
problem with his algorithm (imo) is that he uses a busy wait loop like 
while (nothingtodo) ;  That means if you start more threads than you 
have cpus, you're going to get a real performance hit.  I did mine with 
semaphores so the kernel handles what the threads do when there's no 
work to do (and coincidently got the same 39% improvement in the 2 
thread case, but stayed the same up to 4 threads).

    In xvid VHQ1 mode, the motion search takes 50% of encoder time and 
80% in VHQ4 mode.  Based on that, the theoretical max improvement would 
obviously be:
2 cpus: 25% - 40%
3 cpus: 33.% - 53.3%
4 cpus: 37.5% - 60%