[FFmpeg-devel] Parallelizing MPEG encoding with MPI

[Måns Rullgård]

Roman Shaposhnik <rvs at sun.com> writes:

> On Dec 6, 2008, at 8:08 PM, M?ns Rullg?rd wrote:
>> "Jason Garrett-Glaser" <darkshikari at gmail.com> writes:
>>
>>> On Sat, Dec 6, 2008 at 12:20 PM, Andreas Simbuerger <simi at osaft.eu>  
>>> wrote:
>>>> Greetings everybody!
>>>>
>>>> I'm currently working on a programming exercise for my diploma at  
>>>> the
>>>> University of Passau. We try to enhance MPEG encoding by introducing
>>>> parallelism on GOP encoding with the MPI libraries.
>>>
>>> Isn't MPI rather unnecessary for video encoding parallelism?  It is
>>> rather easy to develop a model (see
>>> http://akuvian.org/src/x264/sliceless_threads.txt) which does not
>>> require communication between threads except to signal the maximum
>>> motion vector search range, which can be done using an atomic write  
>>> to
>>> a struct variable.
>>
>> MPI works on networks of machines with no shared memory.
>
> Not quite true. MPI implementations are capable of utilizing
> shared memory for message passing if they notice that entities
> are running on the same machine. That's one way of utilizing
> those crazy multicore CPUs *and* clusters at the same time -- just
> write MPI and let the runtime figure out the most efficient way
> of passing messages.

What I said is true, but not the entire truth.

-- 
M?ns Rullg?rd
mans at mansr.com