Module mp4ff implements MP4 media file parsing and writing for AVC and HEVC video, AAC and AC-3 audio, stpp and wvtt subtitles, and timed metadata tracks. It is focused on fragmented files as used for streaming in MPEG-DASH, MSS and HLS fMP4, but can also decode and encode all boxes needed for progressive MP4 files.

Some useful command line tools are available in cmd directory.

1. mp4ff-info prints a tree of the box hierarchy of a mp4 file with information about the boxes.
2. mp4ff-pslister extracts and displays SPS and PPS for AVC or HEVC in a mp4 or a bytestream (Annex B) file. Partial information is printed for HEVC.
3. mp4ff-nallister lists NALUs and picture types for video in progressive or fragmented file
4. mp4ff-subslister lists details of wvtt or stpp (WebVTT or TTML in ISOBMFF) subtitle samples
5. mp4ff-crop crops a progressive mp4 file to a specified duration
6. mp4ff-encrypt encrypts a fragmented file using cenc or cbcs Common Encryption scheme
7. mp4ff-decrypt decrypts a fragmented file encrypted using cenc or cbcs Common Encryption scheme

You can install these tools by going to their respective directory and run go install . or directly from the repo with

go install github.com/Eyevinn/mp4ff/cmd/mp4ff-info@latest
go install github.com/Eyevinn/mp4ff/cmd/mp4ff-encrypt@latest
...

for each individual tool.

Example code for some common use cases is available in the examples directory. The examples and their functions are:

1. initcreator creates typical init segments (ftyp + moov) for different video and audio codecs
2. resegmenter reads a segmented file (CMAF track) and resegments it with other segment durations using FullSample
3. segmenter takes a progressive mp4 file and creates init and media segments from it. This tool has been extended to support generation of segments with multiple tracks as well as reading and writing mdat in lazy mode
4. multitrack parses a fragmented file with multiple tracks
5. combine-segs combines single-track init and media segments into multi-track segments
6. add-sidx adds a top-level sidx box describing the segments of a fragmented files.

The top-level packages in the mp4ff module are

1. mp4 provides support for for parsing (called Decode) and writing (Encode) a plethor of mp4 boxes. It also contains helper functions for extracting, encrypting, dectrypting samples and a lot more.
2. avc deals with AVC (aka H.264) video in the mp4ff/avc package including parsing of SPS and PPS, and finding start-codes in Annex B byte streams.
3. hevc provides structures and functions for dealing with HEVC video and its packaging
4. sei provides support for handling Supplementary Enhancement Information (SEI) such as timestamps for AVC and HEVC video.
5. av1 provides basic support for AV1 video packaging
6. aac provides support for AAC audio. This includes handling ADTS headers which is common for AAC inside MPEG-2 TS streams.
7. bits provides bit-wise and byte-wise readers and writers used by the other packages.

The top level structure for both non-fragmented and fragmented mp4 files is mp4.File.

In a progressive (non-fragmented) mp4.File, the top-level attributes Ftyp, Moov, and Mdat point to the corresponding boxes.

A fragmented mp4.File can be more or less complete, like a single init segment, one or more media segments, or a combination of both, like a CMAF track which renders into a playable one-track asset. It can also have multiple tracks. For fragmented files, the following high-level attributes are used:

• Init contains a ftyp and a moov box and provides the general metadata for a fragmented file. It corresponds to a CMAF header. It can also contain one or more sidx boxes.
• Segments is a slice of MediaSegment which start with an optional styp box, possibly one or more sidx boxes and then one or moreFragments.
• Fragment is a mp4 fragment with exactly one moof box followed by a mdat box where the latter contains the media data. It can have one or more trun boxes containing the metadata for the samples. The fragment can start with one or more emsg boxes.

It should be noted that it is sometimes hard to decide what should belong to a Segment or Fragment.

All child boxes of container boxes such as MoovBox are listed in the Children attribute, but the most prominent child boxes have direct links with names which makes it possible to write a path such as

fragment.Moof.Traf.Trun

to access the (only) trun box in a fragment with only one traf box, or

fragment.Moof.Trafs[1].Trun[1]

to get the second trun of the second traf box (provided that they exist). Care must be taken to assert that none of the intermediate pointers are nil to avoid panic.

A typical use case is to generate a fragmented file consisting of an init segment followed by a series of media segments.

The first step is to create the init segment. This is done in three steps as can be seen in examples/initcreator:

init := mp4.CreateEmptyInit()
init.AddEmptyTrack(timescale, mediatype, language)
init.Moov.Trak.SetHEVCDescriptor("hvc1", vpsNALUs, spsNALUs, ppsNALUs)

Here the third step fills in codec-specific parameters into the sample descriptor of the single track. Multiple tracks are also available via the slice attribute Traks instead of Trak.

The second step is to start producing media segments. They should use the timescale that was set when creating the init segment. Generally, that timescale should be chosen so that the sample durations have exact values without rounding errors, e.g. 48000 for 48kHz audio.

A media segment contains one or more fragments, where each fragment has a moof and a mdat box. If all samples are available before the segment is created, one can use a single fragment in each segment. Example code for this can be found in examples/segmenter. For low-latency MPEG-DASH generation, short-duration fragments are added to the segment as the corresponding media samples become available.

A simple, but not optimal, way of creating a media segment is to first create a slice of FullSample with the data needed. The definition of mp4.FullSample is

mp4.FullSample{
 Sample: mp4.Sample{
  Flags uint32 // Flag sync sample etc
  Dur   uint32 // Sample duration in mdhd timescale
  Size  uint32 // Size of sample data
  Cto   int32  // Signed composition time offset
 },
 DecodeTime uint64 // Absolute decode time (offset + accumulated sample Dur)
 Data       []byte // Sample data
}

The mp4.Sample part is what will be written into the trun box. DecodeTime is the media timeline accumulated time. The DecodeTime value of the first sample of a fragment, will be set as the BaseMediaDecodeTime in the tfdt box.

Once a number of such full samples are available, they can be added to a media segment like

seg := mp4.NewMediaSegment()
frag := mp4.CreateFragment(uint32(segNr), mp4.DefaultTrakID)
seg.AddFragment(frag)
for _, sample := range samples {
 frag.AddFullSample(sample)
}

This segment can finally be output to a w io.Writer as

err := seg.Encode(w)

or to a sw bits.SliceWriter as

err := seg.EncodeSW(sw)

For multi-track segments, the code is a bit more involved. Please have a look at examples/segmenter to see how it is done. A more optimal way of handling media sample is to handle them lazily, or using intervals, as explained next.

For video and audio, the dominating part of a mp4 file is the media data which is stored in one or more mdat boxes. In some cases, for example when segmenting large progressive files, it is much more memory efficient to just read the movie or fragment metadata from the moov or moof box and defer the reading of the media data from the mdat box to later.

For decoding, this is supported by running mp4.DecodeFile() in lazy mode as

parsedMp4, err = mp4.DecodeFile(ifd, mp4.WithDecodeMode(mp4.DecModeLazyMdat))

In this case, the media data of the mdat box will not be read, but only its size is being saved. To read or copy the actual data corresponding to a sample, one must calculate the corresponding byte range and either call

func (m *MdatBox) ReadData(start, size int64, rs io.ReadSeeker) ([]byte, error)

or

func (m *MdatBox) CopyData(start, size int64, rs io.ReadSeeker, w io.Writer) (nrWritten int64, err error)

Example code for this, including lazy writing of mdat, can be found in examples/segmenter with the lazy mode set.

The use of the interfaces io.Reader and io.Writer for reading and writing boxes gives a lot of flexibility, but is not optimal when it comes to memory allocation. In particular, the Read(p []byte) method needs a slice p of the proper size to read data, which leads to a lot of allocations and copying of data. In order to achieve better performance, it is advantageous to read the full top level boxes into one, or a few, slices and decode these.

To enable that mode, version 0.27 of the code introduced Decode<X>SR(sr bits.SliceReader) methods to every box <X> where mp4ff.bits.SliceReader is an interface. For example, the TrunBox gets the method DecodeTrunSR(sr bits.SliceReader) in addition to its old DecodeTrun(r io.Reader) method. The bits.SliceReader interface provides methods to read all kinds of data structures from an underlying slice of bytes. It has an implementation bits.FixedSliceReader which uses a fixed-size slice as underlying slice, but one could consider implementing a growing version which would get its data from some external source.

The memory allocation and speed improvements achieved by this may vary, but should be substantial, especially compared to versions before 0.27 which used an extra io.LimitReader layer.

Fur further reduction of memory allocation, use a buffered top-level reader, especially when when reading the mdat box of a progressive file.

To investigate the efficiency of the new SliceReader and SliceWriter methods, benchmarks have been done. The benchmarks are defined in the file mp4/benchmarks_test.go and mp4/benchmarks_srw_test.go. For DecodeFile, one can see a big improvement by going from version 0.26 to version 0.27 which both use the io.Reader interface but another big increase by using the SliceReader source. The latter benchmarks are called BenchmarkDecodeFileSR but have here been given the same name, for easy comparison. Note that the allocations here refers to the heap allocations that are done inside the benchmark loop. Outside that loop, a slice is allocated to keep the input data.

For EncodeFile, one can see that v0.27 is actually worse than v0.26 when used with the io.Writer interface. That is because the code was restructured so that all writes go via the SliceWriter layer in order to reduce code duplication. However, if instead using the SliceWriter methods directly, there is a big relative gain in allocations as can be seen in the last column.

The mp4ff.mp4 contains a lot of box implementations.

Most boxes have their own file named after the box, but in some cases, there may be multiple boxes that have the same content, and the code file then has a generic name like mp4/visualsampleentry.go.

There is an interface for boxes: Box specificied in mp4.box.go,

The interfaces define common Box methods including encode (writing), but not the decode (parsing) methods which have distinct names for each box type and are dispatched from the parsed box name.

That dispatch based on box name is defined by the tables mp4.decodersSR and mp4.decoders for the functions mp4.DecodeBoxSR() and mp4.DecodeBox(), respectively. The SR variant should normally be used for better performance. If a box name is unkonwn, it will result in an UnknownBox being created.

To implement a new box fooo, the following is needed.

Create a file fooo.go and create a struct type FoooBox.

FoooBox must implement the Box interface methods:

Type()
Size()
Encode(w io.Writer)
EncodeSW(sw bits.SliceWriter)
Info()

It also needs its own decode methods DecodeFoooSR and DecodeFooo, which must be added in the decodersSR map and decoders map, respectively For a simple example, look at the PrftBox in prft.go.

A test file fooo_test.go should also have a test using the method boxDiffAfterEncodeAndDecode to check that the box information is equal after encoding and decoding.

Many attributes are public and can therefore be changed in freely. The advantage of this is that it is possible to write code that can manipulate boxes in many different ways, but one must be cautious to avoid breaking links to sub boxes or create inconsistent states in the boxes.

As an example, container boxes such as TrafBox have a method AddChild which adds a box to Children, its slice of children boxes, but also sets a specific member reference such as Tfdt to point to that box. If Children is manipulated directly, that link may no longer be valid.

For fragmented files, one can choose to either encode all boxes in a mp4.File, or only code the ones which are included in the init and media segments. The attribute that controls that is called FragEncMode. Another attribute EncOptimize controls possible optimizations of the file encoding process. Currently, there is only one possible optimization called OptimizeTrun. It can reduce the size of the TrunBox by finding and writing default values in the TfhdBox and omitting the corresponding values from the TrunBox. Note that this may change the size of all ancestor boxes of trun.

Following the ISOBMFF standard, sample numbers and other numbers start at 1 (one-based). This applies to arguments of functions and methods. The actual storage in slices is zero-based, so sample nr 1 has index 0 in the corresponding slice.

The APIs should be fairly stable, but minor non-backwards-compatible changes may happen until version 1.

The main specification for the MP4 file format is the ISO Base Media File Format (ISOBMFF) standard ISO/IEC 14496-12 7th edition 2021. Some boxes are specified in other standards, as should be commented in the code.

MIT, see LICENSE.

Some code in pkg/mp4, comes from or is based on https://github.com/jfbus/mp4 which has Copyright (c) 2015 Jean-François Bustarret.

Some code in pkg/bits comes from or is based on https://github.com/tcnksm/go-casper/tree/master/internal/bits Copyright (c) 2017 Taichi Nakashima.

See CHANGELOG.md.

Join our community on Slack where you can post any questions regarding any of our open source projects. Eyevinn's consulting business can also offer you:

• Further development of this component
• Customization and integration of this component into your platform
• Support and maintenance agreement

Contact [email protected] if you are interested.

Eyevinn Technology is an independent consultant firm specialized in video and streaming. Independent in a way that we are not commercially tied to any platform or technology vendor. As our way to innovate and push the industry forward we develop proof-of-concepts and tools. The things we learn and the code we write we share with the industry in blogs and by open sourcing the code we have written.

Want to know more about Eyevinn and how it is to work here. Contact us at [email protected]!