Using LibMPEG3 for advanced MPEG decoding

Author: Heroine Virtual Ltd. (Motion picture solutions for Linux without a warranty) Harassment:

libmpeg3 is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2, or (at your option) any later version.

libmpeg3 is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

In addition to the GPL's warranty stipulation, libmpeg3 is distributed WITHOUT GUARANTEED SUPPORT. Support that is not guaranteed includes technical support, compiler troubleshooting, debugging, version matching, updating, among other additional labor which may or may not be required to meet a user's requirements.

Table of contents

Building the library
Using tables of contents for editing

LibMPEG3 decodes several MPEG standards into uncompressed data suitable for editing and playback.

libmpeg3 currently decodes:

MPEG-2 video
MPEG-1 video
mp3 audio
mp2 audio
ac3 audio
MPEG-2 transport streams
MPEG-2 program streams
MPEG-1 program streams
IFO files

The video output can be in many different color models and frame sizes. The audio output can be in twos compliment or floating point. Frame accurate seeking, normally impossible in transport streams, is possible in libmpeg3 through the use of a table of contents. MPEG-2 video in YUV-422 colorspace is decodable. Digital TV broadcasts and DVD's can be edited using libmpeg3. Libmpeg3 takes what is normally a last mile distribution format and makes it editable.

Because of these and other features libmpeg3 is not intended for consumer applications but serves users who are interested in high quality editing and footage acquisition.

Building the library

libmpeg3 depends on the CFLAGS environment variable to get optimization flags. You should set it to

-O3 -march=i686 -fmessage-length=0 -funroll-all-loops -fomit-frame-pointer -malign-loops=2 -malign-jumps=2 -malign-functions=2

You must run make to build the library and should be using Kernel 2.4.9 or later. The makefile automatically determines appropriate parameters and puts the library in i686/libmpeg3.a. Several utilities are also built. Install the utilities by running make install.

Unfortunately libmpeg3 excercizes the system more aggressively than a consumer library and this brings out different bugs in each kernel version.

2.4.9: ext3 filesystem failure
2.4.10: memory management failure when running mpeg3toc
2.4.17: memory management failure after 5 hours of decoding video

As libmpeg3 is not one of the standard MPEG decoding libraries, these utilities are unlike any you've ever seen before. Remember a utility is only as illegal or legal as the guy who runs it.


STEP 1: Verifying file compatibility

Programs using libmpeg3 must #include "libmpeg3.h".

Call mpeg3_check_sig to verify if the file can be read by libmpeg3. This returns a 1 if it is compatible and 0 if it isn't.

STEP 2: Open the file

You need an mpeg3_t* file descriptor:

mpeg3_t* file;

Then you need to open the file:

file = mpeg3_open(char *path);

mpeg3_open returns a NULL if the file couldn't be opened for some reason. Be sure to check this. Everything you do with libmpeg3 requires passing the file pointer.

Another way of opening a file is

mpeg3_open_copy(char *path, mpeg3_t *old_file)

You need to open multiple copies of a file in realtime situations because only one thread can access a mpeg3_t structure at a time. The audio and video can't read simultaneously. The solution is not to repeatedly call mpeg3_open but to call mpeg3_open_copy for every file handle after the first one. This copies tables from the first file to speed up opening.

STEP 3: Set optimization strategies

Call mpeg3_set_cpus(mpeg3_t *file, int cpus) to set how many CPUs should be devoted to video decompression. LibMPEG3 can use any number.

Call mpeg3_set_mmx(mpeg3_t *file, int use_mmx) to set if MMX is used for video. Disabling MMX is mandatory for low bitrate streams since it is very lossy. By the way, lately the compiled MMX output has been producing corrupted video. This is a change in the way modern compilers and CPU's handle MMX from the way it was done 4 years ago but since modern CPU's are so fast, you're better off not using MMX at all.

STEP 4: Query the file.

There are a number of queries for the audio components of the stream:

int mpeg3_has_audio(mpeg3_t *file);
int mpeg3_total_astreams(mpeg3_t *file);             // Number of multiplexed audio streams
int mpeg3_audio_channels(mpeg3_t *file, int stream);
int mpeg3_sample_rate(mpeg3_t *file, int stream);
long mpeg3_audio_samples(mpeg3_t *file, int stream); // Total length
The audio is presented as a number of streams starting at 0 and including mpeg3_total_astreams - 1. Each stream contains a certain number of channels starting at 0 and including mpeg3_audio_channels - 1. The methodology is first determine if the file has audio, then get the number of streams in the file, then for each stream get the number of channels, sample rate, and length.

There are also queries for the video components:

int mpeg3_has_video(mpeg3_t *file);
int mpeg3_total_vstreams(mpeg3_t *file);            // Number of multiplexed video streams
int mpeg3_video_width(mpeg3_t *file, int stream);
int mpeg3_video_height(mpeg3_t *file, int stream);
float mpeg3_frame_rate(mpeg3_t *file, int stream);  // Frames/sec
long mpeg3_video_frames(mpeg3_t *file, int stream); // Total length
int mpeg3_colormodel(mpeg3_t *file, int stream);
The video behavior is the same as with audio, except that video has no subdivision under streams. Frame rate is a floating point number of frames per second.

mpeg3_colormodel returns either MPEG3_YUV420P or MPEG3_YUV422P. MPEG3_YUV422P is only encountered in high quality video not available in any consumer distribution medium.

STEP 5: Seeking to a point in the file

Each audio stream and each video stream has a position in the file independant of each other stream. A variety of methods are available for specifying the position of a stream: byte offset, frame, sample. Which method you use depends on whether you're seeking audio or video and whether you have a table of contents for the stream.

The preferred seeking method if you're writing a player is:

int mpeg3_seek_byte(mpeg3_t *file, int64_t byte);
int64_t mpeg3_tell_byte(mpeg3_t *file);
This seeks all tracks to an absolute byte offset in the file. The byte offset is from 0 to the result of:

mpeg3_get_bytes(mpeg3_t *file)
The alternative to byte seeking is frame or sample seeking. Frame seeking is only possible if a table of contents exists. The mpeg3toc that comes with libmpeg3 creates tables of contents from MPEG 1 & 2 streams. Sample seeking is only possible if the stream is fixed bitrate audio. The audio seeking is handled by:

int mpeg3_set_sample(mpeg3_t *file, long sample, int stream);    // Seek
long mpeg3_get_sample(mpeg3_t *file, int stream);    // Tell current position
and the video seeking is handled by:

int mpeg3_set_frame(mpeg3_t *file, long frame, int stream); // Seek
long mpeg3_get_frame(mpeg3_t *file, int stream);            // Tell current position
You can either perform percentage seeking or absolute byte seeking but not both on the same file handle. Once you perform either method, the file becomes configured for that method.

If you're in byte seeking mode and you want the current time stamp in the file you can't use mpeg3_get_frame or mpeg3_get_sample because you don't know the total length in the desired units. The mpeg3_audio_samples and mpeg3_video_frames commands don't work in percentage seeking either. Instead use

double mpeg3_get_time(mpeg3_t *file);
which gives you the last timecode read in seconds. The MPEG standard specifies timecodes being placed in the streams. Now you know the absolute byte position in the file and the current time stamp, enough to update a progress bar or a text box.

Finally, there is a way to seek to the previous frame of video:

int mpeg3_previous_frame(mpeg3_t *file, int stream);
Because MPEG 1 & 2 are really hairy, the set commands won't do much good for playing backwards. mpeg3_previous_frame does some tricks to seek to the previous frame. Next you have to call a read_frame command to read it.

STEP 6: Read the data

To read audio data use:

int mpeg3_read_audio(mpeg3_t *file, 
		float *output_f,      // Pointer to pre-allocated buffer of floats
		short *output_i,      // Pointer to pre-allocated buffer if int16's
		int channel,          // Channel to decode
		long samples,         // Number of samples to decode
		int stream);          // Stream containing the channel
This decodes a buffer of sequential floats or int16's for a single channel, depending on which *output... parameter has a nonzero argument. To get a floating point buffer pass a pre-allocated buffer to output_f and NULL to output_i. To get an int16 buffer pass NULL to output_f and a pre-allocated buffer to output_i. Alternatively you can pass NULL to both buffer arguments and the decoder won't render anything.

After reading an audio buffer, the current position in the one stream is advanced. Remember that if you're using percentage seeking you can't call mpeg3_set_sample to rewind and read every channel. How then, do you read more than one channel of audio data? Use

mpeg3_reread_audio(mpeg3_t *file, 
		float *output_f,      /* Pointer to pre-allocated buffer of floats */
		short *output_i,      /* Pointer to pre-allocated buffer of int16's */
		int channel,          /* Channel to decode */
		long samples,         /* Number of samples to decode */
		int stream);
to read each remaining channel after the first channel.

To read video data there are two methods. RGB frames or YUV frames. To get an RGB frame use:

int mpeg3_read_frame(mpeg3_t *file, 
		unsigned char **output_rows, // Array of pointers to the start of each output row
		int in_x,                    // Location in input frame to take picture
		int in_y, 
		int in_w, 
		int in_h, 
		int out_w,                   // Dimensions of output_rows
		int out_h, 
		int color_model,             // One of the color model #defines given above.
		int stream);
The video decoding works like a camcorder taking copies of a movie screen. The decoder "sees" a region of the movie screen defined by in_x, in_y, in_w, in_h and transfers it to the frame buffer defined by **output_rows. The input values must be within the boundaries given by mpeg3_video_width and mpeg3_video_height. The size of the frame buffer is defined by out_w, out_h. Although the input dimensions are constrained, the frame buffer can be any size.

color_model defines which RGB color model the picture should be decoded to and the possible values are given in libmpeg3.h. The frame buffer pointed to by output_rows must have enough memory allocated to store the color model you select.

You must allocate 4 extra bytes in the last output_row. This is scratch area for the MMX routines.

mpeg3_read_frame advances the position in the one stream by 1 frame.

To read YUV frames use one of two methods:

int mpeg3_read_yuvframe(mpeg3_t *file,
		char *y_output,
		char *u_output,
		char *v_output,
		int in_x,
		int in_y,
		int in_w,
		int in_h,
		int stream);
The behavior of in_x, in_y, in_w, in_h is identical to mpeg3_read_frame except here you have no control over the output frame size. You must allocate in_w * in_h for the y_output, and in_w * in_h / 4 for the u_output and v_output. While mpeg3_read_yuvframe allows cropping of letterbox it still requires one memcpy. A faster alternative is:

int mpeg3_read_yuvframe_ptr(mpeg3_t *file,
		char **y_output,
		char **u_output,
		char **v_output,
		int stream);
This redirects a *y_output, *u_output, and *v_output pointer to the scratch buffer that decoding took place in. Since MPEG is temporal compression, there is always a buffer containing the last decoder output.

For professional use the library can decode YUV 4:2:2 video in addition to YUV 4:2:0. This variable is determined at encoding time, won't affect the usage of mpeg3_read_frame but you do need an extra function call in order to use mpeg3_read_yuvframe. To determine the encoding of the video stream use

mpeg3_colormodel(mpeg3_t *file, int stream)
This returns either MPEG3_YUV420P or MPEG3_YUV422P. The output buffers and the YUV to RGB conversion for mpeg3_read_yuvframe must be adjusted for the higher sampling of MPEG3_YUV422P. When using mpeg3_read_yuvframe_ptr you don't need to adjust any output buffers.

Synchronizing video with audio

To synchronize video with audio in realtime you need to sometimes delay the video and sometimes drop frames. It's easy to calculate the number of frames to drop but if you're using percentage seeking you can't calculate the exact percentage to seek forward by. Instead call

mpeg3_drop_frames(mpeg3_t *file, long frames, int stream);

This skips frames frames from the current position whether in percentage seeking or absolute seeking.

STEP 7: Close the file

Be sure to close the file with mpeg3_close(mpeg3_t *file) when you're done with it.

Using tables of contents for editing

In 1985 everyone watched Smurfs but one guy watched Robotech. In 1990 everyone watched Teenage Mutant Ninja Turtles but one guy watched Transformers. In 1995 everyone watched Pokemon but one guy watched Behind the Scenes. Now everyone wants handheld organizers but one guy wants MPEG editors. For the wierdos who always looked at the camera rig instead of the celebrity, libmpeg3 supports a way of seeking to an exact frame or sample in any kind of MPEG encapsulation format for editing.

A table of contents must be built for any footage to be edited with libmpeg3. Run mpeg3toc <mpeg stream> <output table of contents>

For editing DVD footage, the mpeg stream argument should be the ifo file belonging to the title set to be edited. This utility reads through every file comprising the mpeg stream and records the offset of every 65536th sample and every keyframe so it can be pretty slow.

The resulting table of contents file should be passed to mpeg3_open and mpeg3_open_copy just like a normal file. The only difference is frame seeking of video is available.