[ This file is way out of date ] Performance: - MMX ycrcb conversion to rgb - combine multiplies across weight/quant/idct prescale (?) - parse: - re-arrange data so that coeff blocks are all one big array (alignment + one big memset (mmx!) at beginning of segment) - more efficient bookkeeping (vs current brute force mark and sweep) in second and third passes of parse - still optimize vlc: - combine lookup tables that use the same index: - first level of classes, class_index_mask, class_index_rshift (all indexed by maxbits) - vlc_lookups, vlc_index_mask, vlc_index_rshift (all indexed by class) - sign_mask, sign_rshift (indexed by vlc len) - think about optimizing vlc/getbits interface based on a few observations: - there are three lookups in vlc of the form ((bits & mask) >> shift) are really doing this: bitstream_show_skip(bs,skip,len) // show len bits, beginning skip bits from current position - if we add that interface, and then mmx getbits, this could free registers for better tuning the rest of the vlc lookup code. - note that start and len are bounded to the range 0-16, it might pay to ensure that after flush, show can always count on at least 16 bits remaining in bs->current_word - (there are multiple shows for each flush - eliminates branch in show) - since we parse a whole video segment before we do idcts, we can reserve mmx registers for getbits state for the entire duration of parsing a video segment - note that bitstream state is re-initialized everytime we start a new video segment - mmx version of 248 idct - tune cache footprint: access input and output withouth polluting L1 - get everything working in Windows and use VTune to analyze and improve x86 performance. Documentation: - there is none! - the contents of this file has/will move to the project task list on sourceforge.

Simpan