FFmpeg audio API

This page is for discussion regarding the reworking of the FFmpeg audio API to accommodate the requirements needed for today's audio codecs.

Features needed

• Generalized channel mixing (SIMD optimized) - users should be able to set their own channel mixing coefficients.
• Codec alterable channel mixing coefficients - the codec should be able to set and update the channel mixing coefficients during runtime (DCA supports this feature, maybe AC-3 also).
• Output channel request function - specify the number of output channels, default should be >2 channels mapped to 2 channels
• Channel reordering - currently there are different orders depending on the codec.
• SIMD optimized interleaving
• Allow planar output - don't duplicate the interleaving code in every codec
• Support bit depths other than 16-bit - 8-bit/24-bit/32-bit/float
• Channel selection - ability to access one channel from a multichannel stream

Feature wish list

• Dolby Pro Logic Surround Sound decoding.
• Add a better FFT routine. (Would the KISS implementation be a good candidate?)
• Fixed point MDCT/FFT implementations
• Custom audio filter support. (Basing it on the video filter API ideas?)
• Proper API for enabling SIMD optimized code.

Current ideas

Threads with previous discussions in the subject:

• http://thread.gmane.org/gmane.comp.video.ffmpeg.devel/47485/focus=48097
  • The thread has several ideas already implemented.
• http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/2007-November/038323.html
  • Discussion of general ideas and requirements for the new API.

Downmixing templates

/** Initialization routine for the libavcodec multichannel audio mixer
 *
 * The multichannel mixer does not know the "position" of the speakers and it doesn't need to either. But
 * depending on the mixing matrix it will unknowingly reorder channels to the native order.
 *
 * @param mix
 * This is the actual mixing context. It will hold the all the information needed to perform mixing.
 * If the passed argument is NULL it will allocate a context. If not null it will reinit the passed
 * context. The mix context is of fixed size and will be large enough to support a MAX_MIX_CHANNELS
 * amount of channels.
 *
 * @param inchannels
 * Number of inchannels, this is set by the input stream. This value will be stored in the mixing context.
 *
 * @param outchannels 
 * Number of outchannels, this is set by the user. This value will be stored in the mixing context.
 *
 * @param mixing_coeff_index_matrix
 * Table with inchannels*outchannels index elements, a negative index means that the mixing coeffs should be negated.
 * For example (simplified) [1,2] would mean coeff[1]+coeff[2] while [1,-2] would mean coeff[1]-coeff[2].
 *
 * @param mixing_coeffs_table
 * Table with mixing coeffs, it is this table the mixing_coeff_index_matrix will refer too. It is declared as void* to
 * make it possible for a future addition of fixed point mixing.
 */
void ff_mix_init(DMIXContext* mix, unsigned int inchannels, unsigned int outchannels, int8_t* mixing_coeff_index_matrix, void* mixing_coeffs_table);