Debian Codegraph @

mpeg2enc/NOTES

raw
1 SOME NOTES IN AND AROUND MPEG ENCODING
2
3 (1) The importance of precision...
4
5 Its not just quantizing "about right" that matters.
6
7 However, mp1e has a very very good idea: why not precompile a
8 combined quantisation / fdct matrix with the multipliers set up
9 "just right" to include the necessary frig factors. If nothing
10 else it saves accumulating numeric errors in the intermediate
11 results of the coefficients!
12
13 That mp1e guy was *smart*...
14
15 (2)
16
17 If you want to accurately compute an integer division using
18 multiplication...
19
20
21 Notation: [x](k) means x floor-ed to nearest k. [x] is [x] to the nearest
22 unit.
23
24
25 [x / d](1) = [Kx/d](K) = [ [K/d]x +(K%d)x/d ](K) / K
26
27 Observe that for large K >> d [K/d]x >> (K%d)x/d
28 so that the (K%d)x/d is in effect a rounding correction
29 and hence need not itself be desperately accurate. Thus we
30 expand another step to obtain...
31
32 = [[K/d]x + [(K%d)[K/d]x + d](K)/K ](K) /K
33
34 Where d is a very small correction factor. We now observe that since
35 we're flooring to K on the outside we're really only interested in
36 whether the rounding factor generates a "carry" so it doesn't need to
37 be 100% accurate.
38
39 = [[K/d]x ](K)/K + [[K/d]x]%K + (K%d)[K/d]x + d](K)/K
40
41 For K twice word size the first term is simply the high word of the
42 product [K/d]x
43
44 The second term is simply the low word of the product [K/d]x plus the
45 term (K%d)[K/d]x + d which for large K and small d will fit into a
46 single word. Thus the carry can at most be a single bit.
47
48 We can compute it "exactly" adding a small fixed round up d (32
49 appears to work nicely for K=2^16) and d < 256). However, for MPEG
50 applications we have fairly limited range of d and x. For these we
51 can get "almost" right results with a (coarse approximation). We know
52 0 <= K%d < d. We also know the factor gets very small for larger d.
53 Thus we really only need a correction that "works" for small d.
54 Taking a wild guess to try x ~= (K%d)[K/d]x for smallish d.
55
56
57 Miracle of miracles it works just great! No too-lows, and only very
58 rare too highs for large x.
59
60 Thus [x/d] ~= hiword([K/d]x) + (loword([K/d]x)+x) >= K )
61
62 Clever eh. The actual computation we quantize in MPEG is
63
64 [coefficient[i]*32 + quantisation/2) / quantisation*2*coeffquant[i] ]
65
66 So, combining with Mr mp1e's sneakiness what we need is a table of
67 [K/quantisation*2*coeffquant[i]] for each quantisation. Ooooerrr...
68
69 If we want to get *really* clever we combine the quantisation with the
70 DCT coefficients to go yet faster....
71
72
73 (3) Predicting motion vectors....
74
75
76 You'd think you could use the motion vectors from the P frame as an
77 upper bound for the those needed in the following (in encoding order)
78 B-frames. Wrrrrrooooong. Doesn't work. Well, actually it works, but:
79
80 (a) The gains ain't impressive (10% maybe).
81 (b) The is some quality loss.
82 (c) It is well night impossible to get *any* gains in the MPEG-2
83 interlaced (field MC case) which is where you'd really want the speed gains.
84
85 (4) The wonderful DCT quantisation matrix
86
87 Experiments have revealed that if you're compressing on the margins of
88 acceptable quality you do *far* better if you squeeze the higher
89 frequency DCT coefficients using the quantisation matrix rather than
90 simply allow the encoder to jack up the global quantisation factor.
91
92 Why? The latter starts run an unacceptable risk of producing
93 noticeable mosaic-ing due to big DC offssets. The latter just
94 starts blurring the results by killing off h.f. components.
95
96
97 (5) Spares and Sub-sampled matching.
98
99 You get *good* results searching for sub-sampled matches at
100 sub-sampling resolution.
101
102 However, if you go below the sampling resolution you *will* miss
103 the odd good match.
104
105 Furthermore, the sad of sub-sampled match leading (eventually) to
106 an optimal 1-pel match can be surprisingly much higher than you'd
107 expect. 6*sad scaled for sampling density can occasionally be
108 observed. I was amazed...
109
110 Andrew
111
112
113
114
115