[FFmpeg-devel] [PATCH] adpcm fixes and improvements

[Andreas Rheinhardt]

Michael Niedermayer:
> On Thu, Aug 24, 2023 at 12:51:28AM +0200, Paul B Mahol wrote:
>> On Thu, Aug 24, 2023 at 12:36 AM Michael Niedermayer <michael at niedermayer.cc>
>> wrote:
>>
>>> On Wed, Aug 23, 2023 at 09:04:35PM +0200, Paul B Mahol wrote:
>>>> On Wed, Aug 23, 2023 at 9:02 PM Michael Niedermayer <
>>> michael at niedermayer.cc>
>>>> wrote:
>>>>
>>>>> On Wed, Aug 16, 2023 at 06:53:42PM +0200, Paul B Mahol wrote:
>>>>>> On Wed, Aug 16, 2023 at 6:38 PM Michael Niedermayer <
>>>>> michael at niedermayer.cc>
>>>>>> wrote:
>>>>>>
>>>>>>> On Tue, Aug 15, 2023 at 04:49:05PM +0200, Paul B Mahol wrote:
>>>>>>>> Attached
>>>>>>>
>>>>>>> [...]
>>>>>>>>  adpcm.c |    2 +-
>>>>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>>> 3305dbe07ca935958fa213f5cadc339ad3cc3592
>>>>>>> 0003-avcodec-adpcm-use-already-existing-pointer-for-4xm-d.patch
>>>>>>>> From c6ad6dc7b8725d897e36399e5c7b8174caeb92e6 Mon Sep 17 00:00:00
>>>>> 2001
>>>>>>>> From: Paul B Mahol <onemda at gmail.com>
>>>>>>>> Date: Tue, 15 Aug 2023 14:18:47 +0200
>>>>>>>> Subject: [PATCH 3/4] avcodec/adpcm: use already existing pointer
>>> for
>>>>> 4xm
>>>>>>>>  decoder
>>>>>>>>
>>>>>>>> Signed-off-by: Paul B Mahol <onemda at gmail.com>
>>>>>>>> ---
>>>>>>>>  libavcodec/adpcm.c | 2 +-
>>>>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> diff --git a/libavcodec/adpcm.c b/libavcodec/adpcm.c
>>>>>>>> index b0c3b91a3b..9993c9e531 100644
>>>>>>>> --- a/libavcodec/adpcm.c
>>>>>>>> +++ b/libavcodec/adpcm.c
>>>>>>>> @@ -1211,7 +1211,7 @@ static int
>>> adpcm_decode_frame(AVCodecContext
>>>>>>> *avctx, AVFrame *frame,
>>>>>>>>
>>>>>>>>          for (int i = 0; i < channels; i++) {
>>>>>>>>              ADPCMChannelStatus *cs = &c->status[i];
>>>>>>>> -            samples = (int16_t *)frame->data[i];
>>>>>>>> +            samples = samples_p[i];
>>>>>>>>              for (int n = nb_samples >> 1; n > 0; n--) {
>>>>>>>>                  int v = bytestream2_get_byteu(&gb);
>>>>>>>>                  *samples++ = adpcm_ima_expand_nibble(cs, v &
>>> 0x0F,
>>>>> 4);
>>>>>>>
>>>>>>> should be ok if tested
>>>>>>>
>>>>>>>
>>>>>>>> --
>>>>>>>> 2.39.1
>>>>>>>>
>>>>>>>
>>>>>>>>  libavcodec/adpcm.c                     |  388
>>>>>>> +++++++++++++++++----------------
>>>>>>>>  tests/ref/fate/adpcm-creative-8-2.6bit |    2
>>>>>>>>  tests/ref/fate/adpcm-creative-8-2bit   |    2
>>>>>>>>  tests/ref/fate/adpcm-creative-8-4bit   |    2
>>>>>>>>  tests/ref/fate/adpcm-ms-mono           |   60 +----
>>>>>>>>  5 files changed, 227 insertions(+), 227 deletions(-)
>>>>>>>> 1760df1de66b4227e71ffe942dedcf7d8a33ad48
>>>>>>> 0004-avcodec-adpcm-consume-all-input-when-decoding.patch
>>>>>>>> From 19789bca53548d672bff30b88a8838edaa876bdb Mon Sep 17 00:00:00
>>>>> 2001
>>>>>>>> From: Paul B Mahol <onemda at gmail.com>
>>>>>>>> Date: Tue, 15 Aug 2023 15:25:22 +0200
>>>>>>>> Subject: [PATCH 4/4] avcodec/adpcm: consume all input when
>>> decoding
>>>>>>>>
>>>>>>>> Stops multiple decoding calls for single packet.
>>>>>>>> Also makes decoding faster.
>>>>>>>
>>>>>>> This increases latency, which can be problem if packets are
>>>>>>> sufficiently large
>>>>>>>
>>>>>>
>>>>>> Then reduce size at demuxer level. there is option for it.
>>>>>
>>>>> if that is so, then please explain exactly which option should be used
>>>>> in the commit message
>>>>>
>>>>>
>>>> ffmpeg -h demuxer=wav
>>>
>>> what about the demuxers that are not named "wav" ?
>>>
>>>
>> I think you are blocking this just for sake of blocking because you have
>> nothing more constructive to do.
>> The delay is always present, unless packet size is exact as block align.
>> And for ADPCM this is irrelevant.
> 
> IIUC, before this patch if a 1gb sized ADPCM packet comes from a demuxer
> that is decoded in small kb sized pieces and returned at that granularity
> to the user.
> after this patch (please correct me if iam wrong)
> you get a failure as the decoded data is not even addressable with int
> but lets assume int is 64bits, you get a 4gb or something like that
> audio frame. It maybe faster to decode that in one go but by how much is that
> faster from lets say decode it in 10-100kb sized chunks ?
> 
> IMHO theres a point where it is too big, and simply returning the data
> to the user with some granularity lower than "unlimited" makes sense.
> Iam not asking for a new feature, rather your patch removes this
> 

If a user wants his audio in small bits, he should not send 1GB packets.

- Andreas