[FFmpeg-devel] [PATCH] libavcodec/mpeg12: Remove "fast" mode

[Lynne]

Oct 30, 2023, 01:02 by leo.izen at gmail.com:

> On 10/29/23 11:48, Lynne wrote:
>
>> Oct 29, 2023, 16:43 by kierank at obe.tv:
>>
>>> $subj as discussed at VDD
>>>
>>> Kieran
>>>
>>
>> This isn't Opus or JPEG-XL, there should be a single correct way of decoding something.
>> Looks good to me.
>> _______________________________________________
>>
>
> I'm not sure I understand what you mean - JPEG XL is fully specified. Are you thinking of mjpeg?
>

Anything that does not have a decoder fully defined with fixed-point math
or at least does not define a specific floating point format, with all constants
being given in that float format unambiguously (e.g. regular IEEE 32-bit floats, with all
constant values being given as 32-bit hex), along with a defined and bitexact
mathematical operations (i.e. IEEE). Giving only float constants up to a reasonable
decimal point and defining a maximum error is just not how video codecs
are defined.

Yeah, compared to Opus, JPEG-XL is really ordinary and shouldn't have been there,
and it can be argued that the error margins are small enough to permit bitexactness
up to a bit depth, but I nevertheless would've liked to see fixed-point specifications.