/src/avm/avm_dsp/entdec.c
Line | Count | Source |
1 | | /* |
2 | | * Copyright (c) 2021, Alliance for Open Media. All rights reserved |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 3-Clause Clear License |
5 | | * and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear |
6 | | * License was not distributed with this source code in the LICENSE file, you |
7 | | * can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. If the |
8 | | * Alliance for Open Media Patent License 1.0 was not distributed with this |
9 | | * source code in the PATENTS file, you can obtain it at |
10 | | * aomedia.org/license/patent-license/. |
11 | | */ |
12 | | |
13 | | #include <assert.h> |
14 | | #include "avm_dsp/entdec.h" |
15 | | #include "avm_dsp/prob.h" |
16 | | |
17 | | /*A range decoder. |
18 | | This is an entropy decoder based upon \cite{Mar79}, which is itself a |
19 | | rediscovery of the FIFO arithmetic code introduced by \cite{Pas76}. |
20 | | It is very similar to arithmetic encoding, except that encoding is done with |
21 | | digits in any base, instead of with bits, and so it is faster when using |
22 | | larger bases (i.e.: a byte). |
23 | | The author claims an average waste of $\frac{1}{2}\log_b(2b)$ bits, where $b$ |
24 | | is the base, longer than the theoretical optimum, but to my knowledge there |
25 | | is no published justification for this claim. |
26 | | This only seems true when using near-infinite precision arithmetic so that |
27 | | the process is carried out with no rounding errors. |
28 | | |
29 | | An excellent description of implementation details is available at |
30 | | http://www.arturocampos.com/ac_range.html |
31 | | A recent work \cite{MNW98} which proposes several changes to arithmetic |
32 | | encoding for efficiency actually re-discovers many of the principles |
33 | | behind range encoding, and presents a good theoretical analysis of them. |
34 | | |
35 | | End of stream is handled by writing out the smallest number of bits that |
36 | | ensures that the stream will be correctly decoded regardless of the value of |
37 | | any subsequent bits. |
38 | | avm_od_ec_dec_tell() can be used to determine how many bits were needed to |
39 | | decode all the symbols thus far; other data can be packed in the remaining |
40 | | bits of the input buffer. |
41 | | @PHDTHESIS{Pas76, |
42 | | author="Richard Clark Pasco", |
43 | | title="Source coding algorithms for fast data compression", |
44 | | school="Dept. of Electrical Engineering, Stanford University", |
45 | | address="Stanford, CA", |
46 | | month=May, |
47 | | year=1976, |
48 | | URL="http://www.richpasco.org/scaffdc.pdf" |
49 | | } |
50 | | @INPROCEEDINGS{Mar79, |
51 | | author="Martin, G.N.N.", |
52 | | title="Range encoding: an algorithm for removing redundancy from a digitised |
53 | | message", |
54 | | booktitle="Video & Data Recording Conference", |
55 | | year=1979, |
56 | | address="Southampton", |
57 | | month=Jul, |
58 | | URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz" |
59 | | } |
60 | | @ARTICLE{MNW98, |
61 | | author="Alistair Moffat and Radford Neal and Ian H. Witten", |
62 | | title="Arithmetic Coding Revisited", |
63 | | journal="{ACM} Transactions on Information Systems", |
64 | | year=1998, |
65 | | volume=16, |
66 | | number=3, |
67 | | pages="256--294", |
68 | | month=Jul, |
69 | | URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf" |
70 | | }*/ |
71 | | |
72 | | /*Initializes the decoder. |
73 | | buf: The input buffer to use. |
74 | | storage: The size in bytes of the input buffer.*/ |
75 | | void avm_od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf, |
76 | 17.4k | uint32_t storage) { |
77 | 17.4k | dec->buf = buf; |
78 | 17.4k | dec->end = buf + storage; |
79 | 17.4k | dec->bptr = buf; |
80 | 17.4k | dec->dif = ((od_ec_window)1 << (OD_EC_WINDOW_SIZE - 1)) - 1; |
81 | 17.4k | dec->rng = 0x8000; |
82 | 17.4k | dec->cnt = -15; |
83 | 17.4k | dec->tell_offs = dec->cnt + 1; |
84 | 17.4k | od_ec_dec_refill(dec); |
85 | 17.4k | } |
86 | | |
87 | | /*Decode a single binary value. |
88 | | f: The probability that the bit is one, scaled by 32768. |
89 | | Return: The value decoded (0 or 1).*/ |
90 | 0 | int avm_od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f) { |
91 | 0 | od_ec_window dif; |
92 | 0 | od_ec_window vw; |
93 | 0 | unsigned r; |
94 | 0 | unsigned r_new; |
95 | 0 | unsigned v; |
96 | 0 | int ret; |
97 | 0 | assert(0 < f); |
98 | 0 | assert(f < 32768U); |
99 | 0 | dif = dec->dif; |
100 | 0 | r = dec->rng; |
101 | 0 | assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); |
102 | 0 | assert(32768U <= r); |
103 | 0 | v = od_ec_prob_scale(f, r, 0, 2); |
104 | 0 | vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16); |
105 | 0 | ret = 1; |
106 | 0 | r_new = v; |
107 | 0 | if (dif >= vw) { |
108 | 0 | r_new = r - v; |
109 | 0 | dif -= vw; |
110 | 0 | ret = 0; |
111 | 0 | } |
112 | 0 | return od_ec_dec_normalize(dec, dif, r_new, ret); |
113 | 0 | } |
114 | | |
115 | | /*Decode a single binary value, with 50/50 probability. |
116 | | Return: The value decoded (0 or 1).*/ |
117 | 0 | int od_ec_decode_bool_bypass(od_ec_dec *dec) { |
118 | 0 | return od_ec_decode_literal_bypass(dec, 1); |
119 | 0 | } |
120 | | |
121 | | /*Decode a literal of n_bits |
122 | | n_bits: Number of bits to decode 1..8 |
123 | | Return: The value decoded (0..2^n_bits-1).*/ |
124 | 4.47M | int od_ec_decode_literal_bypass(od_ec_dec *dec, int n_bits) { |
125 | 4.47M | od_ec_window dif; |
126 | 4.47M | od_ec_window vw; |
127 | 4.47M | unsigned r; |
128 | 4.47M | int ret; |
129 | 4.47M | dif = dec->dif; |
130 | 4.47M | r = dec->rng; |
131 | 4.47M | assert((r & 1) == 0); |
132 | 4.47M | assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); |
133 | 4.47M | assert(32768U <= r); |
134 | 4.47M | assert(0 < n_bits && n_bits <= 32); |
135 | 4.47M | vw = (od_ec_window)r << (OD_EC_WINDOW_SIZE - 16); |
136 | 4.47M | ret = 0; |
137 | 9.69M | for (int bit = 0; bit < n_bits; bit++) { |
138 | 5.21M | vw >>= 1; |
139 | 5.21M | ret <<= 1; |
140 | 5.21M | if (dif >= vw) { |
141 | 2.68M | dif -= vw; |
142 | 2.68M | } else { |
143 | 2.53M | ret |= 1; |
144 | 2.53M | } |
145 | 5.21M | } |
146 | 4.47M | return od_ec_dec_bypass_normalize(dec, dif, n_bits, ret); |
147 | 4.47M | } |
148 | | |
149 | | /*Decode unary-coded symbol. |
150 | | max_bits: Max number of decoded bits. |
151 | | Return: The value decoded (0..2^n_bits-1).*/ |
152 | | OD_WARN_UNUSED_RESULT int od_ec_decode_unary_bypass(od_ec_dec *dec, |
153 | | int max_bits) |
154 | | OD_ARG_NONNULL(1); |
155 | 196k | int od_ec_decode_unary_bypass(od_ec_dec *dec, int max_bits) { |
156 | 196k | if (dec->cnt < max_bits - 1) od_ec_dec_refill(dec); |
157 | 196k | od_ec_window dif; |
158 | 196k | od_ec_window vw; |
159 | 196k | unsigned r; |
160 | 196k | int ret; |
161 | 196k | dif = dec->dif; |
162 | 196k | r = dec->rng; |
163 | 196k | assert((r & 1) == 0); |
164 | 196k | assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); |
165 | 196k | assert(32768U <= r); |
166 | 196k | assert((0 < max_bits) && (max_bits <= 32)); |
167 | 196k | vw = (od_ec_window)r << (OD_EC_WINDOW_SIZE - 16); |
168 | 196k | ret = 0; |
169 | 196k | int bit; |
170 | 394k | for (bit = 0; bit < max_bits; bit++) { |
171 | 388k | vw >>= 1; |
172 | 388k | if (dif >= vw) { |
173 | 198k | dif -= vw; |
174 | 198k | ret++; |
175 | 198k | } else { |
176 | 189k | bit++; |
177 | 189k | break; |
178 | 189k | } |
179 | 388k | } |
180 | 196k | return od_ec_dec_bypass_normalize(dec, dif, bit, ret); |
181 | 196k | } |
182 | | |
183 | | /*Decodes a symbol given an inverse cumulative distribution function (CDF) |
184 | | table in Q15. |
185 | | icdf: CDF_PROB_TOP minus the CDF, such that symbol s falls in the range |
186 | | [s > 0 ? (CDF_PROB_TOP - icdf[s - 1]) : 0, CDF_PROB_TOP - icdf[s]). |
187 | | The values must be monotonically non-increasing, and icdf[nsyms - 1] |
188 | | must be 0. |
189 | | nsyms: The number of symbols in the alphabet. |
190 | | This should be at most 16. |
191 | | Return: The decoded symbol s.*/ |
192 | | int avm_od_ec_decode_cdf_q15_c(od_ec_dec *dec, const uint16_t *icdf, |
193 | 0 | int nsyms) { |
194 | 0 | od_ec_window dif; |
195 | 0 | unsigned r; |
196 | 0 | unsigned c; |
197 | 0 | unsigned u; |
198 | 0 | unsigned v; |
199 | 0 | int ret; |
200 | 0 | (void)nsyms; |
201 | 0 | dif = dec->dif; |
202 | 0 | r = dec->rng; |
203 | |
|
204 | 0 | assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); |
205 | 0 | assert(icdf[nsyms - 1] == OD_ICDF(CDF_PROB_TOP)); |
206 | 0 | assert(32768U <= r); |
207 | 0 | assert(7 - EC_PROB_SHIFT - CDF_SHIFT >= 0); |
208 | 0 | c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16)); |
209 | 0 | v = r; |
210 | 0 | ret = -1; |
211 | 0 | do { |
212 | 0 | u = v; |
213 | 0 | ret++; |
214 | 0 | v = od_ec_prob_scale(icdf[ret], r, ret, nsyms); |
215 | 0 | } while (c < v); |
216 | 0 | assert(v < u); |
217 | 0 | assert(u <= r); |
218 | 0 | r = u - v; |
219 | 0 | dif -= (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16); |
220 | 0 | return od_ec_dec_normalize(dec, dif, r, ret); |
221 | 0 | } |
222 | | |
223 | | /*Returns the number of bits "used" by the decoded symbols so far. |
224 | | This same number can be computed in either the encoder or the decoder, and is |
225 | | suitable for making coding decisions. |
226 | | Return: The number of bits. |
227 | | This will always be slightly larger than the exact value (e.g., all |
228 | | rounding error is in the positive direction).*/ |
229 | 86.1k | int avm_od_ec_dec_tell(const od_ec_dec *dec) { |
230 | | /*There is a window of bits stored in dec->dif. The difference |
231 | | (dec->bptr - dec->buf) tells us how many bytes have been read into this |
232 | | window. The difference (dec->cnt - dec->tell_offs) tells us how many of |
233 | | the bits in that window remain unconsumed.*/ |
234 | 86.1k | return (int)((dec->bptr - dec->buf) * 8 - dec->cnt + dec->tell_offs); |
235 | 86.1k | } |
236 | | |
237 | | /*Returns the number of bits "used" by the decoded symbols so far. |
238 | | This same number can be computed in either the encoder or the decoder, and is |
239 | | suitable for making coding decisions. |
240 | | Return: The number of bits scaled by 2**OD_BITRES. |
241 | | This will always be slightly larger than the exact value (e.g., all |
242 | | rounding error is in the positive direction).*/ |
243 | 0 | uint64_t avm_od_ec_dec_tell_frac(const od_ec_dec *dec) { |
244 | 0 | return avm_od_ec_tell_frac(avm_od_ec_dec_tell(dec), dec->rng); |
245 | 0 | } |