ipcm little endian misidentified as big endian

[Dyami Caliri]

Summary of the bug: When loading an mp4 file with ipcm audio, little endian, 24 bit, libavformat thinks it is big-endian. The resulting decoded samples are then 'static' due to the byte swapping.

We are using libavformat/libavcodec to decode media files. I will upload a sample, 'ipcm_le.mp4'. QuickTime player on macOS correctly identifies this is as little endian.

Dumping the format after loading the audio shows this format:

Audio: pcm_s24be (ipcm / 0x6D637069), 48000 Hz, 1 channels, s32 (24 bit), 1152 kb/s

My own inspection into these files suggests that there is a 'pcmC' sub-atom with a flag for endianness. I am not familiar enough with ffmpeg code to try to submit a patch.

[Dyami Caliri]

I couldn't get your upload ftp server to work, so I am providing the sample here:
​https://www.dragonframe.com/download/ipcm_le_ticket_9763.mp4

[Aleksoid1978]

Can you describe the format of an atom 'pcmC' ?

[Balling]

Did you try forcing the byte endiness? Is this due to M1?

[Balling]

can you verify ffplay.exe -acodec pcm_s24le ipcm_le_ticket_9763.mp4 works on larger samples?

[Dyami Caliri]

Replying to Balling:

can you verify ffplay.exe -acodec pcm_s24le ipcm_le_ticket_9763.mp4 works on larger samples?

The sample I provided (ipcm_le_ticket_9763.mp4) just has a couple of small sounds effects in it. It works with the command you have here, as do larger samples.

[Dyami Caliri]

Replying to Balling:

Did you try forcing the byte endiness? Is this due to M1?

libavformat detects it as pcm_s24be. If I override that and tell it to use pcm_s24le, the sound will be correct. It is not due to M1, as it also happens on x86_64.

It is an issue with avformat assuming that it is big-endian when it is not.

[Dyami Caliri]

Replying to Aleksoid1978:

Can you describe the format of an atom 'pcmC' ?

I couldn't find a formal description anywhere, but I looked at several files, and here is what I see:

0E 70636D43 00000000 01180000 00

0x0E = Length 14 bytes
'pcmC' = identifier
0x00000000 = no idea

0x01 = 0x01 -> little endian, 0x00 -> big endian
0x18 = 24 bits per sample?

I don't know the rest.

[IvanB]

Proposed patch here: ​https://patchwork.ffmpeg.org/project/ffmpeg/patch/20220607082618.64814-1-4rusxg@gmail.com/

[Marton Balint]

Thanks, I have factorized out the duplicated code from your patch and simplified it a bit:

​https://patchwork.ffmpeg.org/project/ffmpeg/list/?series=6849

[Balling]

Fixed in cbe216d3a62fe22b16a1ba40b9b3f13e5b4ac9a9.