scuffle_transmuxer/

lib.rs

//! A crate for transmuxing video streams.
#![cfg_attr(feature = "docs", doc = "\n\nSee the [changelog][changelog] for a full release history.")]
#![cfg_attr(feature = "docs", doc = "## Feature flags")]
#![cfg_attr(feature = "docs", doc = document_features::document_features!())]
//! ## License
//!
//! This project is licensed under the MIT or Apache-2.0 license.
//! You can choose between one of them if you use this work.
//!
//! `SPDX-License-Identifier: MIT OR Apache-2.0`
#![allow(clippy::single_match)]
// #![deny(missing_docs)]
#![deny(unsafe_code)]
#![deny(unreachable_pub)]

use std::collections::VecDeque;
use std::fmt::Debug;
use std::io;

use byteorder::{BigEndian, ReadBytesExt};
use bytes::{Buf, Bytes};
use isobmff::boxes::{
    Brand, ChunkOffsetBox, FileTypeBox, HandlerBox, HandlerType, MediaBox, MediaDataBox, MediaHeaderBox,
    MediaInformationBox, MovieBox, MovieExtendsBox, MovieFragmentBox, MovieFragmentHeaderBox, MovieHeaderBox,
    SampleDescriptionBox, SampleSizeBox, SampleTableBox, SampleToChunkBox, SoundMediaHeaderBox, TimeToSampleBox, TrackBox,
    TrackExtendsBox, TrackFragmentBaseMediaDecodeTimeBox, TrackFragmentBox, TrackFragmentHeaderBox, TrackHeaderBox,
    TrackRunBox, VideoMediaHeaderBox,
};
use isobmff::{IsoSized, UnknownBox};
use scuffle_bytes_util::zero_copy::Serialize;
use scuffle_flv::audio::AudioData;
use scuffle_flv::audio::body::AudioTagBody;
use scuffle_flv::audio::body::legacy::LegacyAudioTagBody;
use scuffle_flv::audio::body::legacy::aac::AacAudioData;
use scuffle_flv::audio::header::AudioTagHeader;
use scuffle_flv::audio::header::legacy::{LegacyAudioTagHeader, SoundType};
use scuffle_flv::script::{OnMetaData, ScriptData};
use scuffle_flv::tag::{FlvTag, FlvTagData};
use scuffle_flv::video::VideoData;
use scuffle_flv::video::body::VideoTagBody;
use scuffle_flv::video::body::enhanced::{ExVideoTagBody, VideoPacket, VideoPacketCodedFrames, VideoPacketSequenceStart};
use scuffle_flv::video::body::legacy::LegacyVideoTagBody;
use scuffle_flv::video::header::enhanced::VideoFourCc;
use scuffle_flv::video::header::legacy::{LegacyVideoTagHeader, LegacyVideoTagHeaderAvcPacket};
use scuffle_flv::video::header::{VideoFrameType, VideoTagHeader, VideoTagHeaderData};
use scuffle_h264::Sps;

mod codecs;
mod define;
mod errors;

pub use define::*;
pub use errors::TransmuxError;

struct Tags<'a> {
    video_sequence_header: Option<VideoSequenceHeader<'a>>,
    audio_sequence_header: Option<AudioSequenceHeader>,
    scriptdata_tag: Option<OnMetaData<'a>>,
}

#[derive(Debug, Clone)]
pub struct Transmuxer<'a> {
    // These durations are measured in timescales
    /// sample_freq * 1000
    audio_duration: u64,
    /// fps * 1000
    video_duration: u64,
    sequence_number: u32,
    last_video_timestamp: u32,
    settings: Option<(VideoSettings, AudioSettings)>,
    tags: VecDeque<FlvTag<'a>>,
}

impl Default for Transmuxer<'_> {
    fn default() -> Self {
        Self::new()
    }
}

impl<'a> Transmuxer<'a> {
    pub fn new() -> Self {
        Self {
            sequence_number: 1,
            tags: VecDeque::new(),
            audio_duration: 0,
            video_duration: 0,
            last_video_timestamp: 0,
            settings: None,
        }
    }

    /// Feed raw FLV data to the transmuxer.
    pub fn demux(&mut self, data: Bytes) -> Result<(), TransmuxError> {
        let mut cursor = io::Cursor::new(data);
        while cursor.has_remaining() {
            cursor.read_u32::<BigEndian>()?; // previous tag size
            if !cursor.has_remaining() {
                break;
            }

            let tag = FlvTag::demux(&mut cursor)?;
            self.tags.push_back(tag);
        }

        Ok(())
    }

    /// Feed a single FLV tag to the transmuxer.
    pub fn add_tag(&mut self, tag: FlvTag<'a>) {
        self.tags.push_back(tag);
    }

    /// Get the next transmuxed packet. This will return `None` if there is not
    /// enough data to create a packet.
    pub fn mux(&mut self) -> Result<Option<TransmuxResult>, TransmuxError> {
        let mut writer = Vec::new();

        let Some((video_settings, _)) = &self.settings else {
            let Some((video_settings, audio_settings)) = self.init_sequence(&mut writer)? else {
                if self.tags.len() > 30 {
                    // We are clearly not getting any sequence headers, so we should just give up
                    return Err(TransmuxError::NoSequenceHeaders);
                }

                // We don't have enough tags to create an init segment yet
                return Ok(None);
            };

            self.settings = Some((video_settings.clone(), audio_settings.clone()));

            return Ok(Some(TransmuxResult::InitSegment {
                data: Bytes::from(writer),
                audio_settings,
                video_settings,
            }));
        };

        loop {
            let Some(tag) = self.tags.pop_front() else {
                return Ok(None);
            };

            let mdat_data;
            let total_duration;
            let trun_sample;
            let mut is_audio = false;
            let mut is_keyframe = false;

            let duration =
                if self.last_video_timestamp == 0 || tag.timestamp_ms == 0 || tag.timestamp_ms < self.last_video_timestamp {
                    1000 // the first frame is always 1000 ticks where the
                // timescale is 1000 * fps.
                } else {
                    // Since the delta is in milliseconds (ie 1/1000 of a second)
                    // Rounding errors happen. Our presision is only 1/1000 of a second.
                    // So if we have a 30fps video the delta should be 33.33ms (1000/30)
                    // But we can only represent this as 33ms or 34ms. So we will get rounding
                    // errors. To fix this we just check if the delta is 1 more or 1 less than the
                    // expected delta. And if it is we just use the expected delta.
                    // The reason we use a timescale which is 1000 * fps is because then we can
                    // always represent the delta as an integer. If we use a timescale of 1000, we
                    // would run into the same rounding errors.
                    let delta = tag.timestamp_ms as f64 - self.last_video_timestamp as f64;
                    let expected_delta = 1000.0 / video_settings.framerate;
                    if (delta - expected_delta).abs() <= 1.0 {
                        1000
                    } else {
                        (delta * video_settings.framerate) as u32
                    }
                };

            match tag.data {
                FlvTagData::Audio(AudioData {
                    body: AudioTagBody::Legacy(LegacyAudioTagBody::Aac(AacAudioData::Raw(data))),
                    ..
                }) => {
                    let (sample, duration) = codecs::aac::trun_sample(&data)?;

                    trun_sample = sample;
                    mdat_data = data;
                    total_duration = duration;
                    is_audio = true;
                }
                FlvTagData::Video(VideoData {
                    header:
                        VideoTagHeader {
                            frame_type,
                            data:
                                VideoTagHeaderData::Legacy(LegacyVideoTagHeader::AvcPacket(
                                    LegacyVideoTagHeaderAvcPacket::Nalu { composition_time_offset },
                                )),
                        },
                    body: VideoTagBody::Legacy(LegacyVideoTagBody::Other { data }),
                    ..
                }) => {
                    let composition_time =
                        ((composition_time_offset as f64 * video_settings.framerate) / 1000.0).floor() * 1000.0;

                    let sample = codecs::avc::trun_sample(frame_type, composition_time as u32, duration, &data)?;

                    trun_sample = sample;
                    total_duration = duration;
                    mdat_data = data;

                    is_keyframe = frame_type == VideoFrameType::KeyFrame;
                }
                FlvTagData::Video(VideoData {
                    header: VideoTagHeader { frame_type, .. },
                    body:
                        VideoTagBody::Enhanced(ExVideoTagBody::NoMultitrack {
                            video_four_cc: VideoFourCc::Av1,
                            packet: VideoPacket::CodedFrames(VideoPacketCodedFrames::Other(data)),
                        }),
                    ..
                }) => {
                    let sample = codecs::av1::trun_sample(frame_type, duration, &data)?;

                    trun_sample = sample;
                    total_duration = duration;
                    mdat_data = data;

                    is_keyframe = frame_type == VideoFrameType::KeyFrame;
                }
                FlvTagData::Video(VideoData {
                    header: VideoTagHeader { frame_type, .. },
                    body:
                        VideoTagBody::Enhanced(ExVideoTagBody::NoMultitrack {
                            video_four_cc: VideoFourCc::Hevc,
                            packet,
                        }),
                    ..
                }) => {
                    let (composition_time, data) = match packet {
                        VideoPacket::CodedFrames(VideoPacketCodedFrames::Hevc {
                            composition_time_offset,
                            data,
                        }) => (Some(composition_time_offset), data),
                        VideoPacket::CodedFramesX { data } => (None, data),
                        _ => continue,
                    };

                    let composition_time =
                        ((composition_time.unwrap_or_default() as f64 * video_settings.framerate) / 1000.0).floor() * 1000.0;

                    let sample = codecs::hevc::trun_sample(frame_type, composition_time as i32, duration, &data)?;

                    trun_sample = sample;
                    total_duration = duration;
                    mdat_data = data;

                    is_keyframe = frame_type == VideoFrameType::KeyFrame;
                }
                _ => {
                    // We don't support anything else
                    continue;
                }
            }

            let trafs = {
                let (main_duration, main_id) = if is_audio {
                    (self.audio_duration, 2)
                } else {
                    (self.video_duration, 1)
                };

                let traf = TrackFragmentBox {
                    tfhd: TrackFragmentHeaderBox::new(main_id, None, None, None, None, None),
                    trun: vec![TrackRunBox::new(vec![trun_sample], None)],
                    sbgp: vec![],
                    sgpd: vec![],
                    subs: vec![],
                    saiz: vec![],
                    saio: vec![],
                    tfdt: Some(TrackFragmentBaseMediaDecodeTimeBox::new(main_duration)),
                    meta: None,
                    udta: None,
                };
                // TODO: traf.optimize();

                vec![traf]
            };

            let mut moof = MovieFragmentBox {
                mfhd: MovieFragmentHeaderBox::new(self.sequence_number),
                meta: None,
                traf: trafs,
                udta: None,
            };

            // We need to get the moof size so that we can set the data offsets.
            let moof_size = moof.size();

            // We just created the moof, and therefore we know that the first traf is the
            // video traf and the second traf is the audio traf. So we can just unwrap them
            // and set the data offsets.
            let traf = moof.traf.first_mut().expect("we just created the moof with a traf");

            // Again we know that these exist because we just created it.
            let trun = traf.trun.first_mut().expect("we just created the video traf with a trun");

            // We now define the offsets.
            // So the video offset will be the size of the moof + 8 bytes for the mdat
            // header.
            trun.data_offset = Some(moof_size as i32 + 8);

            // We then write the moof to the writer.
            moof.serialize(&mut writer)?;

            // We create an mdat box and write it to the writer.
            MediaDataBox::new(mdat_data.into()).serialize(&mut writer)?;

            // Increase our sequence number and duration.
            self.sequence_number += 1;

            if is_audio {
                self.audio_duration += total_duration as u64;
                return Ok(Some(TransmuxResult::MediaSegment(MediaSegment {
                    data: Bytes::from(writer),
                    ty: MediaType::Audio,
                    keyframe: false,
                    timestamp: self.audio_duration - total_duration as u64,
                })));
            } else {
                self.video_duration += total_duration as u64;
                self.last_video_timestamp = tag.timestamp_ms;
                return Ok(Some(TransmuxResult::MediaSegment(MediaSegment {
                    data: Bytes::from(writer),
                    ty: MediaType::Video,
                    keyframe: is_keyframe,
                    timestamp: self.video_duration - total_duration as u64,
                })));
            }
        }
    }

    /// Internal function to find the tags we need to create the init segment.
    fn find_tags(&self) -> Tags<'a> {
        let tags = self.tags.iter();
        let mut video_sequence_header = None;
        let mut audio_sequence_header = None;
        let mut scriptdata_tag = None;

        for tag in tags {
            if video_sequence_header.is_some() && audio_sequence_header.is_some() && scriptdata_tag.is_some() {
                break;
            }

            match &tag.data {
                FlvTagData::Video(VideoData {
                    body: VideoTagBody::Legacy(LegacyVideoTagBody::AvcVideoPacketSeqHdr(data)),
                    ..
                }) => {
                    video_sequence_header = Some(VideoSequenceHeader::Avc(data.clone()));
                }
                FlvTagData::Video(VideoData {
                    body:
                        VideoTagBody::Enhanced(ExVideoTagBody::NoMultitrack {
                            video_four_cc: VideoFourCc::Av1,
                            packet: VideoPacket::SequenceStart(VideoPacketSequenceStart::Av1(config)),
                        }),
                    ..
                }) => {
                    video_sequence_header = Some(VideoSequenceHeader::Av1(config.clone()));
                }
                FlvTagData::Video(VideoData {
                    body:
                        VideoTagBody::Enhanced(ExVideoTagBody::NoMultitrack {
                            video_four_cc: VideoFourCc::Hevc,
                            packet: VideoPacket::SequenceStart(VideoPacketSequenceStart::Hevc(config)),
                        }),
                    ..
                }) => {
                    video_sequence_header = Some(VideoSequenceHeader::Hevc(config.clone()));
                }
                FlvTagData::Audio(AudioData {
                    body: AudioTagBody::Legacy(LegacyAudioTagBody::Aac(AacAudioData::SequenceHeader(data))),
                    header:
                        AudioTagHeader::Legacy(LegacyAudioTagHeader {
                            sound_size, sound_type, ..
                        }),
                    ..
                }) => {
                    audio_sequence_header = Some(AudioSequenceHeader {
                        data: AudioSequenceHeaderData::Aac(data.clone()),
                        sound_size: *sound_size,
                        sound_type: *sound_type,
                    });
                }
                FlvTagData::ScriptData(ScriptData::OnMetaData(metadata)) => {
                    scriptdata_tag = Some(*metadata.clone());
                }
                _ => {}
            }
        }

        Tags {
            video_sequence_header,
            audio_sequence_header,
            scriptdata_tag,
        }
    }

    /// Create the init segment.
    fn init_sequence(
        &mut self,
        writer: &mut impl io::Write,
    ) -> Result<Option<(VideoSettings, AudioSettings)>, TransmuxError> {
        // We need to find the tag that is the video sequence header
        // and the audio sequence header
        let Tags {
            video_sequence_header,
            audio_sequence_header,
            scriptdata_tag,
        } = self.find_tags();

        let Some(video_sequence_header) = video_sequence_header else {
            return Ok(None);
        };
        let Some(audio_sequence_header) = audio_sequence_header else {
            return Ok(None);
        };

        let video_codec;
        let audio_codec;
        let video_width;
        let video_height;
        let audio_channels;
        let audio_sample_rate;
        let mut video_fps = 0.0;

        let mut estimated_video_bitrate = 0;
        let mut estimated_audio_bitrate = 0;

        if let Some(scriptdata_tag) = scriptdata_tag {
            video_fps = scriptdata_tag.framerate.unwrap_or(0.0);
            estimated_video_bitrate = scriptdata_tag.videodatarate.map(|v| (v * 1024.0) as u32).unwrap_or(0);
            estimated_audio_bitrate = scriptdata_tag.audiodatarate.map(|v| (v * 1024.0) as u32).unwrap_or(0);
        }

        let mut compatible_brands = vec![Brand::Iso5, Brand::Iso6];

        let video_stsd_entry = match video_sequence_header {
            VideoSequenceHeader::Avc(config) => {
                compatible_brands.push(Brand::Avc1);
                video_codec = VideoCodec::Avc {
                    constraint_set: config.profile_compatibility,
                    level: config.level_indication,
                    profile: config.profile_indication,
                };

                let sps = Sps::parse_with_emulation_prevention(io::Cursor::new(&config.sps[0]))
                    .map_err(|_| TransmuxError::InvalidAVCDecoderConfigurationRecord)?;
                video_width = sps.width() as u32;
                video_height = sps.height() as u32;

                let frame_rate = sps.frame_rate();
                if let Some(frame_rate) = frame_rate {
                    video_fps = frame_rate;
                }

                UnknownBox::try_from_box(codecs::avc::stsd_entry(config, &sps)?)?
            }
            VideoSequenceHeader::Av1(config) => {
                compatible_brands.push(Brand(*b"av01"));
                let (entry, seq_obu) = codecs::av1::stsd_entry(config)?;

                video_height = seq_obu.max_frame_height as u32;
                video_width = seq_obu.max_frame_width as u32;

                let op_point = &seq_obu.operating_points[0];

                video_codec = VideoCodec::Av1 {
                    profile: seq_obu.seq_profile,
                    level: op_point.seq_level_idx,
                    tier: op_point.seq_tier,
                    depth: seq_obu.color_config.bit_depth as u8,
                    monochrome: seq_obu.color_config.mono_chrome,
                    sub_sampling_x: seq_obu.color_config.subsampling_x,
                    sub_sampling_y: seq_obu.color_config.subsampling_y,
                    color_primaries: seq_obu.color_config.color_primaries,
                    transfer_characteristics: seq_obu.color_config.transfer_characteristics,
                    matrix_coefficients: seq_obu.color_config.matrix_coefficients,
                    full_range_flag: seq_obu.color_config.full_color_range,
                };

                UnknownBox::try_from_box(entry)?
            }
            VideoSequenceHeader::Hevc(config) => {
                compatible_brands.push(Brand(*b"hev1"));
                video_codec = VideoCodec::Hevc {
                    constraint_indicator: config.general_constraint_indicator_flags,
                    level: config.general_level_idc,
                    profile: config.general_profile_idc,
                    profile_compatibility: config.general_profile_compatibility_flags,
                    tier: config.general_tier_flag,
                    general_profile_space: config.general_profile_space,
                };

                let (entry, sps) = codecs::hevc::stsd_entry(config)?;
                if let Some(info) = sps.vui_parameters.as_ref().and_then(|p| p.vui_timing_info.as_ref()) {
                    video_fps = info.time_scale.get() as f64 / info.num_units_in_tick.get() as f64;
                }

                video_width = sps.cropped_width() as u32;
                video_height = sps.cropped_height() as u32;

                UnknownBox::try_from_box(entry)?
            }
        };

        let audio_stsd_entry = match audio_sequence_header.data {
            AudioSequenceHeaderData::Aac(data) => {
                compatible_brands.push(Brand::Mp41);
                let (entry, config) =
                    codecs::aac::stsd_entry(audio_sequence_header.sound_size, audio_sequence_header.sound_type, data)?;

                audio_sample_rate = config.sampling_frequency;

                audio_codec = AudioCodec::Aac {
                    object_type: config.audio_object_type,
                };
                audio_channels = match audio_sequence_header.sound_type {
                    SoundType::Mono => 1,
                    SoundType::Stereo => 2,
                    _ => return Err(TransmuxError::InvalidAudioChannels),
                };

                entry
            }
        };

        if video_fps == 0.0 {
            return Err(TransmuxError::InvalidVideoFrameRate);
        }

        if video_width == 0 || video_height == 0 {
            return Err(TransmuxError::InvalidVideoDimensions);
        }

        if audio_sample_rate == 0 {
            return Err(TransmuxError::InvalidAudioSampleRate);
        }

        // The reason we multiply the FPS by 1000 is to avoid rounding errors
        // Consider If we had a video with a framerate of 30fps. That would imply each
        // frame is 33.333333ms So we are limited to a u32 and therefore we could only
        // represent 33.333333ms as 33ms. So this value is 30 * 1000 = 30000 timescale
        // units per second, making each frame 1000 units long instead of 33ms long.
        let video_timescale = (1000.0 * video_fps) as u32;

        // ftyp
        FileTypeBox {
            major_brand: Brand::Iso5,
            minor_version: 512,
            compatible_brands,
        }
        .serialize(&mut *writer)?;

        // moov
        MovieBox {
            mvhd: MovieHeaderBox::new(0, 0, 1000, 0, 1),
            meta: None,
            trak: vec![
                TrackBox::new(
                    TrackHeaderBox::new(0, 0, 1, 0, Some((video_width, video_height))), // tkhd
                    None,                                                               // edts
                    // mdia
                    MediaBox::new(
                        MediaHeaderBox::new(0, 0, video_timescale, 0),                          // mdhd
                        HandlerBox::new(HandlerType::Video, "VideoHandler".to_string().into()), // hdlr
                        // minf
                        MediaInformationBox::new(
                            // stbl
                            SampleTableBox::new(
                                SampleDescriptionBox::new(vec![video_stsd_entry]), // stsd
                                TimeToSampleBox::default(),                        // stts
                                SampleToChunkBox::default(),                       // stsc
                                Some(SampleSizeBox::default()),                    // stsz
                                ChunkOffsetBox::default(),                         // stco
                            ),
                            Some(VideoMediaHeaderBox::default()), // vmhd
                            None,                                 // smhd
                        ),
                    ),
                ),
                TrackBox::new(
                    TrackHeaderBox::new(0, 0, 2, 0, None), // tkhd
                    None,                                  // edts
                    // mdia
                    MediaBox::new(
                        MediaHeaderBox::new(0, 0, audio_sample_rate, 0),                        // mdhd
                        HandlerBox::new(HandlerType::Audio, "SoundHandler".to_string().into()), // hdlr
                        // minf
                        MediaInformationBox::new(
                            // stbl
                            SampleTableBox::new(
                                SampleDescriptionBox::new(vec![UnknownBox::try_from_box(audio_stsd_entry)?]), // stsd
                                TimeToSampleBox::default(),                                                   // stts
                                SampleToChunkBox::default(),                                                  // stsc
                                Some(SampleSizeBox::default()),                                               // stsz
                                ChunkOffsetBox::default(),                                                    // stco
                            ),
                            None,                                 // vmhd
                            Some(SoundMediaHeaderBox::default()), // smhd
                        ),
                    ),
                ),
            ],
            mvex: Some(MovieExtendsBox {
                mehd: None,
                trex: vec![TrackExtendsBox::new(1), TrackExtendsBox::new(2)],
                leva: None,
            }),
            unknown_boxes: vec![],
            udta: None,
        }
        .serialize(writer)?;

        Ok(Some((
            VideoSettings {
                width: video_width,
                height: video_height,
                framerate: video_fps,
                codec: video_codec,
                bitrate: estimated_video_bitrate,
                timescale: video_timescale,
            },
            AudioSettings {
                codec: audio_codec,
                sample_rate: audio_sample_rate,
                channels: audio_channels,
                bitrate: estimated_audio_bitrate,
                timescale: audio_sample_rate,
            },
        )))
    }
}

/// Changelogs generated by [scuffle_changelog]
#[cfg(feature = "docs")]
#[scuffle_changelog::changelog]
pub mod changelog {}

#[cfg(test)]
mod tests;