[FFmpeg-devel] [RFC] an aes "optimization"

Guillaume POIRIER poirierg
Tue Feb 15 11:13:30 CET 2011


Sorry for responding that late. I have trouve keeping up with the
flood of messages on this list.

2010/7/3 M?ns Rullg?rd <mans at mansr.com>:
> Ramiro Polla <ramiro.polla at gmail.com> writes:
>> On Fri, Jul 2, 2010 at 4:15 PM, Guillaume POIRIER <poirierg at gmail.com> wrote:
>>> On Fri, Jul 2, 2010 at 9:10 PM, Reimar D?ffinger
>>> <Reimar.Doeffinger at gmx.de> wrote:
>>> [..]
>>>> On PPC64 (using gcc 4.3.5) the timer seems seems to be broken, because
>>>> it claims a ridiculously low 213 vs. 216 decicycles (the difference
>>>> is below the precision of measurement)...
>>> The problem is that on PPC, the only CPU-model independent timer
>>> that's available is the "The Time Base Registers", which is increment
>>> by one every four *bus* cycles.
> Are there CPU-specific counters with better accuracy? ?If so, we
> should try to support them.

Yes, they have better accuracy. CPU-version-independent timers measure
different things on different versions.
Some measure CPU cycles, some measure bus cycles. The only advantage
of them over CPU-specific timers is that they are always available,
and always measure something related to time (the CPU-specific timers
can be configured to measure cache events, TLB events, ....).

>>> CF: ffmpeg-svn/doc/ffmpeg_powerpc_performance_evaluation_howto.txt
>> That .txt was removed in r23808 along with --powerpc-perf-enable
> Apologies if I removed something that's still relevant. ?I thought
> that entire file was concerned with the broken code I deleted.

As you know, PPC isn't as relevant as it once was, so I don't think
it's useful to resurrect that file...

Best regards,

Guillaume Poirier
"I am the Jesus Christ of politics. I sacrifice myself for everyone."

?-- Silvio Berlusconi

More information about the ffmpeg-devel mailing list