/src/aom/aom_dsp/entdec.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved. |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 2 Clause License and |
5 | | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
6 | | * was not distributed with this source code in the LICENSE file, you can |
7 | | * obtain it at www.aomedia.org/license/software. If the Alliance for Open |
8 | | * Media Patent License 1.0 was not distributed with this source code in the |
9 | | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
10 | | */ |
11 | | |
12 | | #include <assert.h> |
13 | | #include "aom_dsp/entdec.h" |
14 | | #include "aom_dsp/prob.h" |
15 | | |
16 | | /*A range decoder. |
17 | | This is an entropy decoder based upon \cite{Mar79}, which is itself a |
18 | | rediscovery of the FIFO arithmetic code introduced by \cite{Pas76}. |
19 | | It is very similar to arithmetic encoding, except that encoding is done with |
20 | | digits in any base, instead of with bits, and so it is faster when using |
21 | | larger bases (i.e.: a byte). |
22 | | The author claims an average waste of $\frac{1}{2}\log_b(2b)$ bits, where $b$ |
23 | | is the base, longer than the theoretical optimum, but to my knowledge there |
24 | | is no published justification for this claim. |
25 | | This only seems true when using near-infinite precision arithmetic so that |
26 | | the process is carried out with no rounding errors. |
27 | | |
28 | | An excellent description of implementation details is available at |
29 | | http://www.arturocampos.com/ac_range.html |
30 | | A recent work \cite{MNW98} which proposes several changes to arithmetic |
31 | | encoding for efficiency actually re-discovers many of the principles |
32 | | behind range encoding, and presents a good theoretical analysis of them. |
33 | | |
34 | | End of stream is handled by writing out the smallest number of bits that |
35 | | ensures that the stream will be correctly decoded regardless of the value of |
36 | | any subsequent bits. |
37 | | od_ec_dec_tell() can be used to determine how many bits were needed to decode |
38 | | all the symbols thus far; other data can be packed in the remaining bits of |
39 | | the input buffer. |
40 | | @PHDTHESIS{Pas76, |
41 | | author="Richard Clark Pasco", |
42 | | title="Source coding algorithms for fast data compression", |
43 | | school="Dept. of Electrical Engineering, Stanford University", |
44 | | address="Stanford, CA", |
45 | | month=May, |
46 | | year=1976, |
47 | | URL="http://www.richpasco.org/scaffdc.pdf" |
48 | | } |
49 | | @INPROCEEDINGS{Mar79, |
50 | | author="Martin, G.N.N.", |
51 | | title="Range encoding: an algorithm for removing redundancy from a digitised |
52 | | message", |
53 | | booktitle="Video & Data Recording Conference", |
54 | | year=1979, |
55 | | address="Southampton", |
56 | | month=Jul, |
57 | | URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz" |
58 | | } |
59 | | @ARTICLE{MNW98, |
60 | | author="Alistair Moffat and Radford Neal and Ian H. Witten", |
61 | | title="Arithmetic Coding Revisited", |
62 | | journal="{ACM} Transactions on Information Systems", |
63 | | year=1998, |
64 | | volume=16, |
65 | | number=3, |
66 | | pages="256--294", |
67 | | month=Jul, |
68 | | URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf" |
69 | | }*/ |
70 | | |
71 | | /*This is meant to be a large, positive constant that can still be efficiently |
72 | | loaded as an immediate (on platforms like ARM, for example). |
73 | | Even relatively modest values like 100 would work fine.*/ |
74 | 302k | #define OD_EC_LOTS_OF_BITS (0x4000) |
75 | | |
76 | | /*The return value of od_ec_dec_tell does not change across an od_ec_dec_refill |
77 | | call.*/ |
78 | 23.2M | static void od_ec_dec_refill(od_ec_dec *dec) { |
79 | 23.2M | int s; |
80 | 23.2M | od_ec_window dif; |
81 | 23.2M | int16_t cnt; |
82 | 23.2M | const unsigned char *bptr; |
83 | 23.2M | const unsigned char *end; |
84 | 23.2M | dif = dec->dif; |
85 | 23.2M | cnt = dec->cnt; |
86 | 23.2M | bptr = dec->bptr; |
87 | 23.2M | end = dec->end; |
88 | 23.2M | s = OD_EC_WINDOW_SIZE - 9 - (cnt + 15); |
89 | 69.8M | for (; s >= 0 && bptr < end; s -= 8, bptr++) { |
90 | | /*Each time a byte is inserted into the window (dif), bptr advances and cnt |
91 | | is incremented by 8, so the total number of consumed bits (the return |
92 | | value of od_ec_dec_tell) does not change.*/ |
93 | 46.6M | assert(s <= OD_EC_WINDOW_SIZE - 8); |
94 | 46.6M | dif ^= (od_ec_window)bptr[0] << s; |
95 | 46.6M | cnt += 8; |
96 | 46.6M | } |
97 | 23.2M | if (bptr >= end) { |
98 | | /*We've reached the end of the buffer. It is perfectly valid for us to need |
99 | | to fill the window with additional bits past the end of the buffer (and |
100 | | this happens in normal operation). These bits should all just be taken |
101 | | as zero. But we cannot increment bptr past 'end' (this is undefined |
102 | | behavior), so we start to increment dec->tell_offs. We also don't want |
103 | | to keep testing bptr against 'end', so we set cnt to OD_EC_LOTS_OF_BITS |
104 | | and adjust dec->tell_offs so that the total number of unconsumed bits in |
105 | | the window (dec->cnt - dec->tell_offs) does not change. This effectively |
106 | | puts lots of zero bits into the window, and means we won't try to refill |
107 | | it from the buffer for a very long time (at which point we'll put lots |
108 | | of zero bits into the window again).*/ |
109 | 151k | dec->tell_offs += OD_EC_LOTS_OF_BITS - cnt; |
110 | 151k | cnt = OD_EC_LOTS_OF_BITS; |
111 | 151k | } |
112 | 23.2M | dec->dif = dif; |
113 | 23.2M | dec->cnt = cnt; |
114 | 23.2M | dec->bptr = bptr; |
115 | 23.2M | } |
116 | | |
117 | | /*Takes updated dif and range values, renormalizes them so that |
118 | | 32768 <= rng < 65536 (reading more bytes from the stream into dif if |
119 | | necessary), and stores them back in the decoder context. |
120 | | dif: The new value of dif. |
121 | | rng: The new value of the range. |
122 | | ret: The value to return. |
123 | | Return: ret. |
124 | | This allows the compiler to jump to this function via a tail-call.*/ |
125 | | static int od_ec_dec_normalize(od_ec_dec *dec, od_ec_window dif, unsigned rng, |
126 | 637M | int ret) { |
127 | 637M | int d; |
128 | 637M | assert(rng <= 65535U); |
129 | | /*The number of leading zeros in the 16-bit binary representation of rng.*/ |
130 | 638M | d = 16 - OD_ILOG_NZ(rng); |
131 | | /*d bits in dec->dif are consumed.*/ |
132 | 638M | dec->cnt -= d; |
133 | | /*This is equivalent to shifting in 1's instead of 0's.*/ |
134 | 638M | dec->dif = ((dif + 1) << d) - 1; |
135 | 638M | dec->rng = rng << d; |
136 | 638M | if (dec->cnt < 0) od_ec_dec_refill(dec); |
137 | 638M | return ret; |
138 | 637M | } |
139 | | |
140 | | /*Initializes the decoder. |
141 | | buf: The input buffer to use. |
142 | | storage: The size in bytes of the input buffer.*/ |
143 | | void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf, |
144 | 162k | uint32_t storage) { |
145 | 162k | dec->buf = buf; |
146 | 162k | dec->tell_offs = 10 - (OD_EC_WINDOW_SIZE - 8); |
147 | 162k | dec->end = buf + storage; |
148 | 162k | dec->bptr = buf; |
149 | 162k | dec->dif = ((od_ec_window)1 << (OD_EC_WINDOW_SIZE - 1)) - 1; |
150 | 162k | dec->rng = 0x8000; |
151 | 162k | dec->cnt = -15; |
152 | 162k | od_ec_dec_refill(dec); |
153 | 162k | } |
154 | | |
155 | | /*Decode a single binary value. |
156 | | f: The probability that the bit is one, scaled by 32768. |
157 | | Return: The value decoded (0 or 1).*/ |
158 | 122M | int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f) { |
159 | 122M | od_ec_window dif; |
160 | 122M | od_ec_window vw; |
161 | 122M | unsigned r; |
162 | 122M | unsigned r_new; |
163 | 122M | unsigned v; |
164 | 122M | int ret; |
165 | 122M | assert(0 < f); |
166 | 122M | assert(f < 32768U); |
167 | 122M | dif = dec->dif; |
168 | 122M | r = dec->rng; |
169 | 122M | assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); |
170 | 122M | assert(32768U <= r); |
171 | 122M | v = ((r >> 8) * (uint32_t)(f >> EC_PROB_SHIFT) >> (7 - EC_PROB_SHIFT)); |
172 | 122M | v += EC_MIN_PROB; |
173 | 122M | vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16); |
174 | 122M | ret = 1; |
175 | 122M | r_new = v; |
176 | 122M | if (dif >= vw) { |
177 | 62.7M | r_new = r - v; |
178 | 62.7M | dif -= vw; |
179 | 62.7M | ret = 0; |
180 | 62.7M | } |
181 | 122M | return od_ec_dec_normalize(dec, dif, r_new, ret); |
182 | 122M | } |
183 | | |
184 | | /*Decodes a symbol given an inverse cumulative distribution function (CDF) |
185 | | table in Q15. |
186 | | icdf: CDF_PROB_TOP minus the CDF, such that symbol s falls in the range |
187 | | [s > 0 ? (CDF_PROB_TOP - icdf[s - 1]) : 0, CDF_PROB_TOP - icdf[s]). |
188 | | The values must be monotonically non-increasing, and icdf[nsyms - 1] |
189 | | must be 0. |
190 | | nsyms: The number of symbols in the alphabet. |
191 | | This should be at most 16. |
192 | | Return: The decoded symbol s.*/ |
193 | 520M | int od_ec_decode_cdf_q15(od_ec_dec *dec, const uint16_t *icdf, int nsyms) { |
194 | 520M | od_ec_window dif; |
195 | 520M | unsigned r; |
196 | 520M | unsigned c; |
197 | 520M | unsigned u; |
198 | 520M | unsigned v; |
199 | 520M | int ret; |
200 | 520M | (void)nsyms; |
201 | 520M | dif = dec->dif; |
202 | 520M | r = dec->rng; |
203 | 520M | const int N = nsyms - 1; |
204 | | |
205 | 520M | assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); |
206 | 520M | assert(icdf[nsyms - 1] == OD_ICDF(CDF_PROB_TOP)); |
207 | 521M | assert(32768U <= r); |
208 | 520M | assert(7 - EC_PROB_SHIFT >= 0); |
209 | 520M | c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16)); |
210 | 520M | v = r; |
211 | 520M | ret = -1; |
212 | 1.00G | do { |
213 | 1.00G | u = v; |
214 | 1.00G | v = ((r >> 8) * (uint32_t)(icdf[++ret] >> EC_PROB_SHIFT) >> |
215 | 1.00G | (7 - EC_PROB_SHIFT)); |
216 | 1.00G | v += EC_MIN_PROB * (N - ret); |
217 | 1.00G | } while (c < v); |
218 | 520M | assert(v < u); |
219 | 521M | assert(u <= r); |
220 | 521M | r = u - v; |
221 | 521M | dif -= (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16); |
222 | 521M | return od_ec_dec_normalize(dec, dif, r, ret); |
223 | 521M | } |
224 | | |
225 | | /*Returns the number of bits "used" by the decoded symbols so far. |
226 | | This same number can be computed in either the encoder or the decoder, and is |
227 | | suitable for making coding decisions. |
228 | | Return: The number of bits. |
229 | | This will always be slightly larger than the exact value (e.g., all |
230 | | rounding error is in the positive direction).*/ |
231 | 1.36M | int od_ec_dec_tell(const od_ec_dec *dec) { |
232 | | /*There is a window of bits stored in dec->dif. The difference |
233 | | (dec->bptr - dec->buf) tells us how many bytes have been read into this |
234 | | window. The difference (dec->cnt - dec->tell_offs) tells us how many of |
235 | | the bits in that window remain unconsumed.*/ |
236 | 1.36M | return (int)((dec->bptr - dec->buf) * 8 - dec->cnt + dec->tell_offs); |
237 | 1.36M | } |
238 | | |
239 | | /*Returns the number of bits "used" by the decoded symbols so far. |
240 | | This same number can be computed in either the encoder or the decoder, and is |
241 | | suitable for making coding decisions. |
242 | | Return: The number of bits scaled by 2**OD_BITRES. |
243 | | This will always be slightly larger than the exact value (e.g., all |
244 | | rounding error is in the positive direction).*/ |
245 | 0 | uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec) { |
246 | 0 | return od_ec_tell_frac(od_ec_dec_tell(dec), dec->rng); |
247 | 0 | } |