[FFmpeg-devel] PATCH - libmad MP3 decoding support

[David Fletcher]

On Fri May 6 21:29:51 EEST 2022 Lynne dev at lynne.ee wrote:

> I disagree with this patch. The native decoder should be fixed instead,
> or if it's not able to be fixed, users should use specialized libraries.
> We're not adding a yet another decoding library to fix an issue on
> very old ARM chips.

> I suspect the problem is with the MDCT. We already have a better and
> more precise fixed-point MDCT in libavutil/tx. The plan was to replace
> the MDCT in mp3 once lavu/tx has more assembly, but if it helps,
> should be fine to replace it now.

Hi Lynne,

I don't have a strong feeling either way. The patch (an earlier version)
is already doing what I needed in the application with the Reciva
radios. I posted it here to contribute back to FFmpeg, but if there's a
better way to fix the problem that's great.

If there's a better version of MDCT integrated into the native mp3 fixed
point decoder I can test that on the ARMv4 hardware to see if the
decoding bug is fixed. Certainly these are old chips but still lots of
them around - a big part of them being useful is the assembly
optimisations otherwise everything just slows down too much.

Best regards, David.