[FFmpeg-devel] MPEG-2 Acceleration Refactor
Sun Jun 17 12:28:16 CEST 2007
I guess I may as well reply on-list then...
On Sat, 16 Jun 2007 at 23:46 -0700, Greg Hulands wrote:
> Hi Michael,
> On 16/06/2007, at 6:52 PM, Michael Niedermayer wrote:
>> On Sat, Jun 16, 2007 at 06:44:21PM -0700, Greg Hulands wrote:
>>> Based on the START/STOP_TIMER stuff showing the patch is actually
>>> faster I have tidied the patch a little more and removed the function
>>> prototypes to the _fast functions at the top of the file.
>>> If there are any other problems from stopping this getting committed,
>>> please let me know.
>> yes, why is it faster now and was slower before?
> John Dalgliesh, the guy who originally wrote the patch on an older svn
I'll just add that my patch was in an untested and unbenchmarked state;
the first try at this factorisation. I didn't then (and still dont) have
time to complete it.
> I clean built ffmpeg with -s on gcc and then did a diff of the symbols dumped
> from the mpeg12.o files.
Ahh, that should have been -S not -s.
> I have attached these so you can look at them. I am
> not sure if that explains why it is faster or not. I know I have said this
> before, but I am out of my depth with things at the moment, so if you are
> able to hand hold me through this it would be greatly appreciated. If this
> isn't what is needed to determine why it is faster, could you outline what is
> required and how to do it please.
I don't think too much can be gleaned from those lists of symbols.
As for the patch, it is now a lot better wrt cosmetics, there are a couple
left 'tho and a few indentation irregularities. But that doesn't matter if
it doesn't function as intended. (Except for allowing swift rejection :)
If you send me the output of the two -S runs via private mail then I'll
try to find time to see what's it's doing and if it's good. Hmmm.. maybe
it's faster even if not inlined from the cache benefit plus predicability
of those added conditions? I'd be surprised 'tho...
More information about the ffmpeg-devel