[FFmpeg-devel] [PATCH v23 04/10] avformat/evc_demuxer: Added demuxer to handle reading EVC video files

[James Almer]

On 6/7/2023 10:36 AM, Dawid Kozinski/Multimedia (PLT) /SRPOL/Staff 
Engineer/Samsung Electronics wrote:
> 
> 
> 
>> -----Original Message-----
>> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of James
>> Almer
>> Sent: poniedziałek, 29 maja 2023 15:08
>> To: ffmpeg-devel at ffmpeg.org
>> Subject: Re: [FFmpeg-devel] [PATCH v23 04/10] avformat/evc_demuxer: Added
>> demuxer to handle reading EVC video files
>>
>> On 5/26/2023 7:31 AM, Dawid Kozinski wrote:
>>> +static int get_temporal_id(const uint8_t *bits, int bits_size) {
>>> +    int temporal_id = 0;
>>> +    short t = 0;
>>> +
>>> +    if (bits_size >= EVC_NALU_HEADER_SIZE) {
>>> +        unsigned char *p = (unsigned char *)bits;
>>> +        // forbidden_zero_bit
>>> +        if ((p[0] & 0x80) != 0)
>>> +            return -1;
>>> +
>>> +        for (int i = 0; i < EVC_NALU_HEADER_SIZE; i++)
>>> +            t = (t << 8) | p[i];
>>
>> I think this can be replaced with AV_RB16.
>>
>>> +
>>> +        temporal_id = (t >> 6) & 0x0007;
>>> +    }
>>> +
>>> +    return temporal_id;
>>> +}
>>
>> [...]
>>
>>> +static uint32_t read_nal_unit_length(const uint8_t *bits, int
>>> +bits_size) {
>>> +    uint32_t nalu_len = 0;
>>> +
>>> +    if (bits_size >= EVC_NALU_LENGTH_PREFIX_SIZE) {
>>> +
>>> +        int t = 0;
>>> +        unsigned char *p = (unsigned char *)bits;
>>> +
>>> +        for (int i = 0; i < EVC_NALU_LENGTH_PREFIX_SIZE; i++)
>>> +            t = (t << 8) | p[i];
>>
>> AV_RB32.
>>
>>> +
>>> +        nalu_len = t;
>>> +        if (nalu_len == 0)   // Invalid bitstream size
>>> +            return 0;
>>> +    }
>>> +
>>> +    return nalu_len;
>>> +}
>>
>> [...]
>>
>>> +static int evc_frame_merge_filter(AVBSFContext *bsf, AVPacket *out) {
>>> +    EVCMergeContext *ctx = bsf->priv_data;
>>> +    AVPacket *in = ctx->in;
>>> +
>>> +    int free_space = 0;
>>> +    size_t  nalu_size = 0;
>>> +    uint8_t *nalu = NULL;
>>> +    int au_end_found = 0;
>>> +    int err;
>>> +
>>> +    err = ff_bsf_get_packet_ref(bsf, in);
>>> +    if (err < 0)
>>> +        return err;
>>> +
>>> +    nalu_size = read_nal_unit_length(in->data,
>> EVC_NALU_LENGTH_PREFIX_SIZE);
>>> +    if(nalu_size <= 0) {
>>> +        av_packet_unref(in);
>>> +        return AVERROR_INVALIDDATA;
>>> +    }
>>> +
>>> +    nalu = in->data + EVC_NALU_LENGTH_PREFIX_SIZE;
>>> +    nalu_size = in->size - EVC_NALU_LENGTH_PREFIX_SIZE;
>>> +
>>> +    // NAL unit parsing needed to determine if end of AU was found
>>> +    err = parse_nal_unit(nalu, nalu_size, bsf);
>>> +    if (err < 0) {
>>> +        av_log(bsf, AV_LOG_ERROR, "NAL Unit parsing error\n");
>>> +        av_packet_unref(in);
>>> +
>>> +        return err;
>>> +    }
>>> +
>>> +    au_end_found = end_of_access_unit_found(bsf);
>>> +
>>> +    free_space = ctx->au_buffer.capacity - ctx->au_buffer.data_size;
>>> +    while( free_space < in->size ) {
>>> +        ctx->au_buffer.capacity *= 2;
>>> +        free_space = ctx->au_buffer.capacity -
>>> + ctx->au_buffer.data_size;
>>> +
>>> +        if(free_space >= in->size) {
>>> +            ctx->au_buffer.data = av_realloc(ctx->au_buffer.data, ctx-
>>> au_buffer.capacity);
>>> +        }
>>> +    }
>>> +
>>> +    memcpy(ctx->au_buffer.data + ctx->au_buffer.data_size, in->data,
>>> + in->size);
>>
>> This is including the Annex-B's nal_unit_lenght value into the assembled
> packet. I
>> assume libxevd expects this? Also, is it defined as part of the ISOBMFF
>> encapsulation? Patch 07/10 does not seem to strip it.
> 
> Could you expand on this topic?
> What do you mean by "Patch 07/10 does not seem to strip it"?
> Patch 7/10 expands the implementation of the MOV Muxer to support the EVC
> stream.
> The function ff_isom_write_evcc parses the received data, extracts NAL units
> of types SPS, PPS, APS, SEI, and inserts their contents into the
> corresponding elements of the arrays array, which is a component of the
> EVCDecoderConfigurationRecord structure. Additionally, the
> ff_isom_write_evcc function populates the relevant fields of the
> EVCDecoderConfigurationRecord structure based on the parsed SPS NAL unit
> data, and then calls the evcc_write function, which writes the
> EVCDecoderConfigurationRecord to the stream.
> 
> It seems that I don't understand what's wrong with our implementation ( I
> mean MOV muxer for EVC) and it seems I need your help in this matter. I need
> an explanation of what's wrong and what should be done to fix the issue.

I mean the packets that go into ISOBMFF Sample, not the evcc atom. Is 
the encapsulation supposed to include the nal_unit_length from Annex-B?