ref: a6e2476e4d56cde5249ad0525afdda85f38f27da
dir: /common/mp4v2/INTERNALS/
January 7, 2002 MP4V2 LIBRARY INTERNALS ======================= This document provides an overview of the internals of the mp4v2 library to aid those who wish to modify and extend it. Before reading this document, I recommend familiarizing yourself with the MP4 (or Quicktime) file format standard and the mp4v2 library API. The API is described in a set of man pages in mpeg4ip/doc/mp4v2, or if you prefer by looking at mp4.h. All the library code is written in C++, however the library API follows uses C calling conventions hence is linkable by both C and C++ programs. The library has been compiled and used on Linux, BSD, Windows, and Mac OS X. Other than libc, the library has no external dependencies, and hence can be used independently of the mpeg4ip package if desired. The library is used for both real-time recording and playback in mpeg4ip, and its runtime performance is up to those tasks. On the IA32 architecture compiled with gcc, the stripped library is approximately 600 KB code and initialized data. It is useful to think of the mp4v2 library as consisting of four layers: infrastructure, file format, generic tracks, and type specific track helpers. A description of each layer follows, from the fundamental to the optional. Infrastructure ============== The infrastructure layer provides basic file I/O, memory allocation, error handling, string utilities, and protected arrays. The source files for this layer are mp4file_io, mp4util, and mp4array. Note that the array classes uses preprocessor macros instead of C++ templates. The rationale for this is to increase portability given the sometimes incomplete support by some compilers for templates. File Format =========== The file format layer provides the translation from the on-disk MP4 file format to in-memory C++ structures and back to disk. It is intended to exactly match the MP4 specification in syntax and semantics. It represents the majority of the code. There are three key structures at the file format layer: atoms, properties, and descriptors. Atoms are the primary containers within an mp4 file. They can contain any combination of properties, other atoms, or descriptors. The mp4atom files contain the base class for all the atoms, and provide generic functions that cover most cases. Most atoms are covered in atom_standard.cpp. Atoms that have a special read, generation or write needs are contained in their subclass contained in file atom_<name>.cpp, where <name> is the four letter name of the atom defined in the MP4 specification. Atoms that only specifies the properties of the atom or the possible child atoms in the case of a container atom are located in atom_standard.cpp. In more specialized cases the atom specific file provides routines to initialize, read, or write the atom. Properties are the atomic pieces of information. The basic types of properties are integers, floats, strings, and byte arrays. For integers and floats there are subclasses that represent the different storage sizes, e.g. 8, 16, 24, 32, and 64 bit integers. For strings, there is 1 property class with a number of options regarding exact storage details, e.g. null terminated, fixed length, counted. For implementation reasons, there are also two special properties, table and descriptor, that are actually containers for groups of properties. I.e by making these containers provide a property interface much code can be written in a generic fashion. The mp4property files contain all the property related classes. Descriptors are containers that derive from the MPEG conventions and use different encoding rules than the atoms derived from the QuickTime file format. This means more use of bitfields and conditional existence with an emphasis on bit efficiency at the cost of encoding/decoding complexity. Descriptors can contain other descriptors and/or properties. The mp4descriptor files contain the generic base class for descriptors. Also the mp4property files have a descriptor wrapper class that allows a descriptor to behave as if it were a property. The specific descriptors are implemented as subclasses of the base class descriptor in manner similar to that of atoms. The descriptors, ocidescriptors, and qosqualifiers files contain these implementations. Each atom/property/descriptor has a name closely related to that in the MP4 specification. The difference being that the mp4v2 library doesn't use '-' or '_' in property names and capitalizes the first letter of each word, e.g. "thisIsAPropertyName". A complete name specifies the complete container path. The names follow the C/C++ syntax for elements and array indices. Examples are: "moov.mvhd.duration" "moov.trak[2].tkhd.duration" "moov.trak[3].minf.mdia.stbl.stsz[101].sampleSize" Note "*" can be used as a wildcard for an atom name (only). This is most useful when dealing with the stsd atom which contains child atoms with various names, but shared property names. Note that internally when performance matters the code looks up a property by name once, and then stores the returned pointer to the property class. To add an atom, first you should see if an existing atom exists that can be used. If not, you need to decide if special read/write or generate properties need to be established; for example a property in the atom changes other properties (adds, or subtracts). If there are no special cases, add the atom properties to atom_standard.cpp. If there are special properties, add a new file, add a new class to atoms.h, and add the class to MP4Atom::CreateAtom in mp4atom.cpp. Generic Tracks ============== The two entities at this level are the mp4 file as a whole and the tracks which are contained with it. The mp4file and mp4track files contain the implementation. The critical work done by this layer is to map the collection of atoms, properties, and descriptors that represent a media track into a useful, and consistent set of operations. For example, reading or writing a media sample of a track is a relatively simple operation from the library API perspective. However there are numerous pieces of information in the mp4 file that need to be properly used and updated to do this. This layer handles all those details. Given familiarity with the mp4 spec, the code should be straight-forward. What may not be immediately obvious are the functions to handle chunks of media samples. These exist to allow optimization of the mp4 file layout by reordering the chunks on disk to interleave the media sample chunks of multiple tracks in time order. (See MP4Optimize API doc). Type Specific Track Helpers =========================== This specialized code goes beyond the meta-information about tracks in the mp4 file to understanding and manipulating the information in the track samples. There are currently two helpers in the library: the MPEG-4 Systems Helper, and the RTP Hint Track Helper. The MPEG-4 Systems Helper is currently limited to creating the OD, BIFS, and SDP information about a minimal audio/video scene consistent with the Internet Streaming Media Alliance (ISMA) specifications. We will be evaluating how best to generalize the library's helper functions for MPEG-4 Systems without overburdening the implementation. The code for this helper is found in the isma and odcommands files. The RTP Hint Track Helper is more extensive in its support. The hint tracks contain the track packetization information needed to build RTP packets for streaming. The library can construct RTP packets based on the hint track making RTP based servers significantly easier to write. All code related to rtp hint tracks is in the rtphint files. It would also be useful to look at test/mp4broadcaster and mpeg4ip/server/mp4creator for examples of how this part of the library API can be used. Library API =========== The library API is defined and implemented in the mp4 files. The API uses C linkage conventions, and the mp4.h file adapts itself according to whether C or C++ is the compilation mode. All API calls are implemented in mp4.cpp and basically pass thru's to the MP4File member functions. This ensures that the library has internal access to the same functions as available via the API. All the calls in mp4.cpp use C++ try/catch blocks to protect against any runtime errors in the library. Upon error the library will print a diagnostic message if the verbostiy level has MP4_DETAILS_ERROR set, and return a distinguished error value, typically 0 or -1. The test and util subdirectories contain useful examples of how to use the library. Also the mp4creator and mp4live programs within mpeg4ip demonstrate more complete usage of the library API. Debugging ========= Since mp4 files are fairly complicated, extensive debugging support is built into the library. Multi-level diagnostic messages are available under the control of a verbosity bitmask described in the API. Also the library provides the MP4Dump() call which provides an ASCII version of the mp4 file meta-information. The mp4dump utilitity is a wrapper executable around this function. The mp4extract program is also provided in the utilities directory which is useful for extracting a track from an mp4file and putting the media data back into it's own file. It can also extract each sample of a track into its own file it that is desired. When all else fails, mp4 files are amenable to debugging by direct examination. Since the atom names are four letter ASCII codes finding reference points in a hex dump is feasible. On UNIX, the od command is your friend: "od -t x1z -A x [-j 0xXXXXXX] foo.mp4" will print a hex and ASCII dump, with hex addresses, starting optionally from a specified offset. The library diagnostic messages can provide information on where the library is reading or writing. General caveats =============== The coding convention is to use the C++ throw operator whenever an unrecoverable error occurs. This throw is caught at the API layer in mp4.cpp and translated into an error value. Be careful about indices. Internally, we follow the C/C++ convention to use zero-based indices. However the MP4 spec uses one-based indices for things like samples and hence the library API uses this convention.