flac dev - Oct 2024 - C API: How to get a seektable for very long files?

Op zo 13 okt 2024 om 22:33 schreef Stefan Oltmanns <stefan-oltmanns at
gmx.net>:
>
> Is the seektable written at the beginning of the file in the metadata
> block or can there also be a second metadata block at the end?
>
Only at the start of the file.
>
> If it's at the beginning, would it possible to reserve space for N seek
> points and during encoding remember a seek point after X samples,
> resulting in M seek points when encoding is finished. If M <= N all seek
> points are written, otherwise only every 2nd, 3rd etc.
> Is it possible to do that? The functions all expect a a total_samples
> argument, which is not known at the beginning.
>
No, you can only provide a seek table template with specific sample
numbers. There is no way to ask the encoder to add a seek point every
so many samples.
>
> The signal is the FM-modulated video signal of video tapes (like VHS).
> The idea is to capture the signal directly from the video head amplifier
> in the VCR and later demodulate/decode it in software, providing higher
> quality than traditional capture of analog video. See this project:
> https://github.com/oyvindln/vhs-decode/
> I started to design a capture device, as there is no 40 MHz continuous
> sampling hardware available at consumer prices:
> https://github.com/Stefan-Olt/MISRC
>
I've seen similar uses before. Maybe this one can serve as some
inspiration: https://www.youtube.com/watch?v=ZrEFU22C8l8 According to
that guy, he used cheap hardware.

Op ma 14 okt 2024 om 00:09 schreef Stefan Oltmanns <stefan-oltmanns at
gmx.net>:
>
> I think there is another major issue for me: In
> METADATA_BLOCK_STREAMINFO the field for the length is only 36 bit,
> that's not even half an hour at 40 MHz sample rate, resulting in that
> the encoder sets it to 0 for longer captures. In the seekpoint the
> sample number is 64 bit, which is more than enough.
> But how does the decoder handle the seektable when the total number of
> samples is unknown? Or does the seektable override the info from
> METADATA_BLOCK_STREAMINFO?
When a suitable seektable is found, it overrides the information from
streaminfo, yes.
>
> I used this functions now to add seekpoints, but all remain placeholders
> according to metaflac:
>
> FLAC__metadata_object_new
> FLAC__metadata_object_seektable_template_append_placeholders
> FLAC__stream_encoder_set_metadata
> (encoder init & loop)
> FLAC__metadata_object_seektable_template_sort
>
Yes, that is correct, because you asked for placeholder points. You
should ask for spaced points. I just tested what happens if you make a
seek table template with a total_samples that is bigger than the
eventual total_samples that is encoded, and found a bug in the
encoder. It works, but the resulting seek table isn't valid. You could
try to use that approach anyway in the meantime, perhaps it works just
fine?

I'd say, have your implementation prepare a seek table template for 5
hours of recording (I assume that is above the upper bound of such
captures?), the stream encoder will fill in those seek points when it
reaches them, and leaves the unused ones unfilled. I will work on a
fix, so the stream encoder converts those unused points to placeholder
points in the future.

The stream encoder works pretty simple: you give it metadata to add to
the start of the stream, it adds those verbatim. After encoding is
finished, it will update streaminfo and seektable. There is no
metadata at the end of the stream, and the metadata blocks should not
be changed during encoding.

I see the difficulty here now by the way: metaflac also refuses to
write a seektable when the streaminfo metadata block specifies 0 total
samples, which is unavoidable in your case.

Am 14.10.24 um 09:11 schrieb Martijn van Beurden:
> Op zo 13 okt 2024 om 22:33 schreef Stefan Oltmanns <stefan-oltmanns at
gmx.net>:
>>
>
>>
>> The signal is the FM-modulated video signal of video tapes (like VHS).
>> The idea is to capture the signal directly from the video head
amplifier
>> in the VCR and later demodulate/decode it in software, providing higher
>> quality than traditional capture of analog video. See this project:
>> https://github.com/oyvindln/vhs-decode/
>> I started to design a capture device, as there is no 40 MHz continuous
>> sampling hardware available at consumer prices:
>> https://github.com/Stefan-Olt/MISRC
>>
>
> I've seen similar uses before. Maybe this one can serve as some
> inspiration: https://www.youtube.com/watch?v=ZrEFU22C8l8 According to
> that guy, he used cheap hardware.
That's indeed an interesting device.
>
> Op ma 14 okt 2024 om 00:09 schreef Stefan Oltmanns <stefan-oltmanns at
gmx.net>:
>>
>> I think there is another major issue for me: In
>> METADATA_BLOCK_STREAMINFO the field for the length is only 36 bit,
>> that's not even half an hour at 40 MHz sample rate, resulting in
that
>> the encoder sets it to 0 for longer captures. In the seekpoint the
>> sample number is 64 bit, which is more than enough.
>> But how does the decoder handle the seektable when the total number of
>> samples is unknown? Or does the seektable override the info from
>> METADATA_BLOCK_STREAMINFO?
>
> When a suitable seektable is found, it overrides the information from
> streaminfo, yes.
Unfortunately that doesn't seem to be the case. I just made a capture
that is > 30 Minutes with total samples set to 0 and a seek table: All
players I tried cannot seek in the file and cannot determine it's
length: VLC, Celluloid and DeaDBeef

I wondered why I can seek files I manually compressed with flac from
raw: The total samples field is just wrong, the seek table is correct,
having sample_numer values a lot higher than total samples. The players
display the wrong length. Seems to be a bug in flac.
>
>>
>> I used this functions now to add seekpoints, but all remain
placeholders
>> according to metaflac:
>>
>> FLAC__metadata_object_new
>> FLAC__metadata_object_seektable_template_append_placeholders
>> FLAC__stream_encoder_set_metadata
>> (encoder init & loop)
>> FLAC__metadata_object_seektable_template_sort
>>
>
> Yes, that is correct, because you asked for placeholder points. You
> should ask for spaced points. I just tested what happens if you make a
> seek table template with a total_samples that is bigger than the
> eventual total_samples that is encoded, and found a bug in the
> encoder. It works, but the resulting seek table isn't valid. You could
> try to use that approach anyway in the meantime, perhaps it works just
> fine?
I fixed that manually after calling
FLAC__metadata_object_seektable_template_sort, according to metaflac
everything is correct:

     for(int i = seektable->data.seek_table.num_points-1; i>=0; i--) {
         if (seektable->data.seek_table.points[i].stream_offset != 0) break;
         seektable->data.seek_table.points[i].sample_number
0xFFFFFFFFFFFFFFFF;
     }

>
> I'd say, have your implementation prepare a seek table template for 5
> hours of recording (I assume that is above the upper bound of such
> captures?), the stream encoder will fill in those seek points when it
> reaches them, and leaves the unused ones unfilled. I will work on a
> fix, so the stream encoder converts those unused points to placeholder
> points in the future.
I have done that now, 2^18 seekpoints for a maximum length of 2^41
samples, that sould be enough.
>
> The stream encoder works pretty simple: you give it metadata to add to
> the start of the stream, it adds those verbatim. After encoding is
> finished, it will update streaminfo and seektable. There is no
> metadata at the end of the stream, and the metadata blocks should not
> be changed during encoding.
>
> I see the difficulty here now by the way: metaflac also refuses to
> write a seektable when the streaminfo metadata block specifies 0 total
> samples, which is unavoidable in your case.
Just an idea: Of course the size of the total samples field cannot be
changed, and a seektable currently doesn't seem to work with total
samples = 0.

Is it possible to extend the specification with full backward
compatibility in a way like this?

If total samples is 0, the last seek point in the seek table that is not
a place holder can indicate the total number of samples:
In case frame_samples is 0, stream_offset is pointing to the last frame
and sample_number is the total number of samples.

FLAC__stream_decoder_get_total_samples would check that and return the
correct value.

Not sure if that breaks any old versions / 3rd party decoder. Another
option would be to use the first place holder, set stream_offset to the
total samples and some magic value in frame_samples. That should
definitly be backward compatible as these values are undefined for
placeholders.

Another option would of course be to add a new type of
metadata, like METADATA_BLOCK_STREAMINFO2, but I don't know how
backwards compatible that is.

Best regards
Stefan