[FFmpeg-devel] [PATCH] mpegaudio float default
Tue May 18 22:18:46 CEST 2010
On 18/05/10 20:51, Michael Niedermayer wrote:
> On Mon, May 17, 2010 at 11:47:08AM -0400, Ronald S. Bultje wrote:
>> On Mon, May 17, 2010 at 11:18 AM, Michael Niedermayer<michaelni at gmx.at> wrote:
>>> On Thu, May 13, 2010 at 06:23:46PM -0400, Ronald S. Bultje wrote:
>>>> On Thu, May 13, 2010 at 6:12 PM, Michael Niedermayer<michaelni at gmx.at> wrote:
>>>>> On Thu, May 13, 2010 at 10:56:36PM +0100, M?ns Rullg?rd wrote:
>>>>>> Michael Niedermayer<michaelni at gmx.at> writes:
>>>>>> 3x slower on Cortex-A8. We need to discuss betters way to choose
>>>>>> implementation. Ideally we should have something that extends well to
>>>>>> other codecs too. There are fixed-point versions of at least wma
>>>>>> floating about, and IMO we should be working towards merging them.
>>>>> someone is working on fixp wma ? :)
>>>> Rockbox (?) already has a fxp wma1/2 decoder, based on the FFmpeg one
>>>> from several years ago, that they'd like to backport into our tree if
>>>> we're interested.
>>> ive been pointed to
>>> http://www.mail-archive.com/rockbox-dev at cool.haxx.se/msg07388.html
>>> http://www.mail-archive.com/rockbox-dev at cool.haxx.se/msg07413.html
>>> This thread does not sound like the rockbox devels seriously intent to
>>> integrate fixed point support back into the ffmpeg codecs.
>>> But iam not on their ML so its easy for me to misinterrpret this.
>> I think you're right, and re-reading my original reply I was too
>> positive. Just alone looking at their reference to tremor, I don't
>> think they have any idea of how to do this in a good way (= code
>> sharing between fixp/float impls where possible, etc.), they simply
>> want to plumb a new codec in our tree called "wmafixed", which decodes
>> streams that are syntactically similar to wma1/2 but which shares 0%
>> code with any of the rest of the FFmpeg tree.
>> I don't think that's useful.
> so back to my original question
> do we have someone who wants to work on adding fixp support to out wma decoder?
> mans? ben?
> if no i might (emphasis is "might") eventually try to do it myself,
> but mans would be the more ideal person to do this because he is actually
> interrested in it and has fpu-less hardware.
> Its not particularely hard work, just converting one function at a time to
> fixp and verifying that it still works. One just has to try to maximize
> precission while preventing overflows and redusing rescaling steps
The thing is that rockbox has good mdct code. We should try to get that
merged first. After that step is done we can merge the cook patch that
was posted years ago. But currently I don't have time to do it, I'll
talk to some Rockbox guys tomorrow though and see if they can do it.
More information about the ffmpeg-devel