1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
|
[ This file is way out of date ]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.
|
