* Initial implementation of the G.726 ADPCM audio codec.
[libav.git] / libavcodec / g726.c
CommitLineData
e5966052
RS
1/*
2 * G.726 ADPCM audio codec
3 * Copyright (c) 2004 Roman Shaposhnik.
4 *
5 * This is a very straightforward rendition of the G.726
6 * Section 4 "Computational Details".
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 */
22#include <limits.h>
23#include "avcodec.h"
24#include "common.h"
25
26/*
27 * G.726 Standard uses rather odd 11bit floating point arithmentic for
28 * numerous occasions. It's a mistery to me why they did it this way
29 * instead of simply using 32bit integer arithmetic.
30 */
31typedef struct Float11 {
32 int sign; /* 1bit sign */
33 int exp; /* 4bit exponent */
34 int mant; /* 6bit mantissa */
35} Float11;
36
37static inline Float11* i2f(int16_t i, Float11* f)
38{
39 f->sign = (i < 0);
40 if (f->sign)
41 i = -i;
42 f->exp = av_log2_16bit(i) + !!i;
43 f->mant = i? (i<<6) >> f->exp :
44 1<<5;
45 return f;
46}
47
48static inline int16_t mult(Float11* f1, Float11* f2)
49{
50 int res, exp;
51
52 exp = f1->exp + f2->exp;
53 res = (((f1->mant * f2->mant) + 0x30) >> 4) << 7;
54 res = exp > 26 ? res << (exp - 26) : res >> (26 - exp);
55 return (f1->sign ^ f2->sign) ? -res : res;
56}
57
58static inline int clamp(int value, int min, int max)
59{
60 if (value < min)
61 return min;
62 else if (value > max)
63 return max;
64 else
65 return value;
66}
67
68static inline int sgn(int value)
69{
70 return (value < 0) ? -1 : 1;
71}
72
73typedef struct G726Tables {
74 int bits; /* bits per sample */
75 int* quant; /* quantization table */
76 int* iquant; /* inverse quantization table */
77 int* W; /* special table #1 ;-) */
78 int* F; /* special table #2 */
79} G726Tables;
80
81typedef struct G726Context {
82 G726Tables* tbls; /* static tables needed for computation */
83
84 Float11 sr[2]; /* prev. reconstructed samples */
85 Float11 dq[6]; /* prev. difference */
86 int a[2]; /* second order predictor coeffs */
87 int b[6]; /* sixth order predictor coeffs */
88 int pk[2]; /* signs of prev. 2 sez + dq */
89
90 int ap; /* scale factor control */
91 int yu; /* fast scale factor */
92 int yl; /* slow scale factor */
93 int dms; /* short average magnitude of F[i] */
94 int dml; /* long average magnitude of F[i] */
95 int td; /* tone detect */
96
97 int se; /* estimated signal for the next iteration */
98 int sez; /* estimated second order prediction */
99 int y; /* quantizer scaling factor for the next iteration */
100} G726Context;
101
102static int quant_tbl16[] = /* 16kbit/s 2bits per sample */
103 { 260, INT_MAX };
104static int iquant_tbl16[] =
105 { 116, 365, 365, 116 };
106static int W_tbl16[] =
107 { -22, 439, 439, -22 };
108static int F_tbl16[] =
109 { 0, 7, 7, 0 };
110
111static int quant_tbl24[] = /* 24kbit/s 3bits per sample */
112 { 7, 217, 330, INT_MAX };
113static int iquant_tbl24[] =
114 { INT_MIN, 135, 273, 373, 373, 273, 135, INT_MIN };
115static int W_tbl24[] =
116 { -4, 30, 137, 582, 582, 137, 30, -4 };
117static int F_tbl24[] =
118 { 0, 1, 2, 7, 7, 2, 1, 0 };
119
120static int quant_tbl32[] = /* 32kbit/s 4bits per sample */
121 { -125, 79, 177, 245, 299, 348, 399, INT_MAX };
122static int iquant_tbl32[] =
123 { INT_MIN, 4, 135, 213, 273, 323, 373, 425,
124 425, 373, 323, 273, 213, 135, 4, INT_MIN };
125static int W_tbl32[] =
126 { -12, 18, 41, 64, 112, 198, 355, 1122,
127 1122, 355, 198, 112, 64, 41, 18, -12};
128static int F_tbl32[] =
129 { 0, 0, 0, 1, 1, 1, 3, 7, 7, 3, 1, 1, 1, 0, 0, 0 };
130
131static int quant_tbl40[] = /* 40kbit/s 5bits per sample */
132 { -122, -16, 67, 138, 197, 249, 297, 338,
133 377, 412, 444, 474, 501, 527, 552, INT_MAX };
134static int iquant_tbl40[] =
135 { INT_MIN, -66, 28, 104, 169, 224, 274, 318,
136 358, 395, 429, 459, 488, 514, 539, 566,
137 566, 539, 514, 488, 459, 429, 395, 358,
138 318, 274, 224, 169, 104, 28, -66, INT_MIN };
139static int W_tbl40[] =
140 { 14, 14, 24, 39, 40, 41, 58, 100,
141 141, 179, 219, 280, 358, 440, 529, 696,
142 696, 529, 440, 358, 280, 219, 179, 141,
143 100, 58, 41, 40, 39, 24, 14, 14 };
144static int F_tbl40[] =
145 { 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 3, 4, 5, 6, 6,
146 6, 6, 5, 4, 3, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
147
148static G726Tables G726Tables_pool[] =
149 {{ 2, quant_tbl16, iquant_tbl16, W_tbl16, F_tbl16 },
150 { 3, quant_tbl24, iquant_tbl24, W_tbl24, F_tbl24 },
151 { 4, quant_tbl32, iquant_tbl32, W_tbl32, F_tbl32 },
152 { 5, quant_tbl40, iquant_tbl40, W_tbl40, F_tbl40 }};
153
154
155/*
156 * Para 4.2.2 page 18: Adaptive quantizer.
157 */
158static inline uint8_t quant(G726Context* c, int d)
159{
160 int sign, exp, i, dln;
161
162 sign = i = 0;
163 if (d < 0) {
164 sign = 1;
165 d = -d;
166 }
167 exp = av_log2_16bit(d);
168 dln = ((exp<<7) + (((d<<7)>>exp)&0x7f)) - (c->y>>2);
169
170 while (c->tbls->quant[i] < INT_MAX && c->tbls->quant[i] < dln)
171 ++i;
172
173 if (sign)
174 i = ~i;
175 if (c->tbls->bits != 2 && i == 0) /* I'm not sure this is a good idea */
176 i = 0xff;
177
178 return i;
179}
180
181/*
182 * Para 4.2.3 page 22: Inverse adaptive quantizer.
183 */
184static inline int16_t inverse_quant(G726Context* c, int i)
185{
186 int dql, dex, dqt;
187
188 dql = c->tbls->iquant[i] + (c->y >> 2);
189 dex = (dql>>7) & 0xf; /* 4bit exponent */
190 dqt = (1<<7) + (dql & 0x7f); /* log2 -> linear */
191 return (dql < 0) ? 0 : ((dqt<<7) >> (14-dex));
192}
193
194static inline int16_t g726_iterate(G726Context* c, int16_t I)
195{
196 int dq, re_signal, pk0, fa1, i, tr, ylint, ylfrac, thr2, al, dq0;
197 Float11 f;
198
199 dq = inverse_quant(c, I);
200 if (I >> (c->tbls->bits - 1)) /* get the sign */
201 dq = -dq;
202 re_signal = c->se + dq;
203
204 /* Transition detect */
205 ylint = (c->yl >> 15);
206 ylfrac = (c->yl >> 10) & 0x1f;
207 thr2 = (ylint > 9) ? 0x1f << 10 : (0x20 + ylfrac) << ylint;
208 if (c->td == 1 && abs(dq) > ((thr2+(thr2>>1))>>1))
209 tr = 1;
210 else
211 tr = 0;
212
213 /* Update second order predictor coefficient A2 and A1 */
214 pk0 = (c->sez + dq) ? sgn(c->sez + dq) : 0;
215 dq0 = dq ? sgn(dq) : 0;
216 if (tr) {
217 c->a[0] = 0;
218 c->a[1] = 0;
219 for (i=0; i<6; i++)
220 c->b[i] = 0;
221 } else {
222 /* This is a bit crazy, but it really is +255 not +256 */
223 fa1 = clamp((-c->a[0]*c->pk[0]*pk0)>>5, -256, 255);
224
225 c->a[1] += 128*pk0*c->pk[1] + fa1 - (c->a[1]>>7);
226 c->a[1] = clamp(c->a[1], -12288, 12288);
227 c->a[0] += 64*3*pk0*c->pk[0] - (c->a[0] >> 8);
228 c->a[0] = clamp(c->a[0], -(15360 - c->a[1]), 15360 - c->a[1]);
229
230 for (i=0; i<6; i++)
231 c->b[i] += 128*dq0*sgn(-c->dq[i].sign) - (c->b[i]>>8);
232 }
233
234 /* Update Dq and Sr and Pk */
235 c->pk[1] = c->pk[0];
236 c->pk[0] = pk0 ? pk0 : 1;
237 c->sr[1] = c->sr[0];
238 i2f(re_signal, &c->sr[0]);
239 for (i=5; i>0; i--)
240 c->dq[i] = c->dq[i-1];
241 i2f(dq, &c->dq[0]);
242 c->dq[0].sign = I >> (c->tbls->bits - 1); /* Isn't it crazy ?!?! */
243
244 /* Update tone detect [I'm not sure 'tr == 0' is really needed] */
245 c->td = (tr == 0 && c->a[1] < -11776);
246
247 /* Update Ap */
248 c->dms += ((c->tbls->F[I]<<9) - c->dms) >> 5;
249 c->dml += ((c->tbls->F[I]<<11) - c->dml) >> 7;
250 if (tr)
251 c->ap = 256;
252 else if (c->y > 1535 && !c->td && (abs((c->dms << 2) - c->dml) < (c->dml >> 3)))
253 c->ap += (-c->ap) >> 4;
254 else
255 c->ap += (0x200 - c->ap) >> 4;
256
257 /* Update Yu and Yl */
258 c->yu = clamp(c->y + (((c->tbls->W[I] << 5) - c->y) >> 5), 544, 5120);
259 c->yl += c->yu + ((-c->yl)>>6);
260
261 /* Next iteration for Y */
262 al = (c->ap >= 256) ? 1<<6 : c->ap >> 2;
263 c->y = (c->yl + (c->yu - (c->yl>>6))*al) >> 6;
264
265 /* Next iteration for SE and SEZ */
266 c->se = 0;
267 for (i=0; i<6; i++)
268 c->se += mult(i2f(c->b[i] >> 2, &f), &c->dq[i]);
269 c->sez = c->se >> 1;
270 for (i=0; i<2; i++)
271 c->se += mult(i2f(c->a[i] >> 2, &f), &c->sr[i]);
272 c->se >>= 1;
273
274 return clamp(re_signal << 2, -0xffff, 0xffff);
275}
276
277static int g726_reset(G726Context* c, int bit_rate)
278{
279 int i;
280
281 c->tbls = &G726Tables_pool[bit_rate/8000 - 2];
282 for (i=0; i<2; i++) {
283 i2f(0, &c->sr[i]);
284 c->a[i] = 0;
285 c->pk[i] = 1;
286 }
287 for (i=0; i<6; i++) {
288 i2f(0, &c->dq[i]);
289 c->b[i] = 0;
290 }
291 c->ap = 0;
292 c->dms = 0;
293 c->dml = 0;
294 c->yu = 544;
295 c->yl = 34816;
296 c->td = 0;
297
298 c->se = 0;
299 c->sez = 0;
300 c->y = 544;
301
302 return 0;
303}
304
305static int16_t g726_decode(G726Context* c, int16_t i)
306{
307 return g726_iterate(c, i);
308}
309
310static int16_t g726_encode(G726Context* c, int16_t sig)
311{
312 uint8_t i;
313
314 i = quant(c, sig/4 - c->se) & ((1<<c->tbls->bits) - 1);
315 g726_iterate(c, i);
316 return i;
317}
318
319/* Interfacing to the libavcodec */
320
321typedef struct AVG726Context {
322 G726Context c;
323 int bits_left;
324 int bit_buffer;
325 int code_size;
326} AVG726Context;
327
328static int g726_init(AVCodecContext * avctx)
329{
330 AVG726Context* c = (AVG726Context*)avctx->priv_data;
331
332 if (avctx->sample_rate != 8000 || avctx->channels != 1 ||
333 (avctx->bit_rate != 16000 && avctx->bit_rate != 24000 &&
334 avctx->bit_rate != 32000 && avctx->bit_rate != 40000)) {
335 av_log(avctx, AV_LOG_ERROR, "G726: unsupported audio format\n");
336 return -1;
337 }
338 g726_reset(&c->c, avctx->bit_rate);
339 c->code_size = c->c.tbls->bits;
340 c->bit_buffer = 0;
341 c->bits_left = 0;
342
343 return 0;
344}
345
346static int g726_encode_frame(AVCodecContext *avctx,
347 uint8_t *dst, int buf_size, void *data)
348{
349 AVG726Context *c = avctx->priv_data;
350 short *samples = data;
351 PutBitContext pb;
352
353 init_put_bits(&pb, dst, 1024*1024);
354
355 for (; buf_size; buf_size--)
356 put_bits(&pb, c->code_size, g726_encode(&c->c, *samples++));
357
358 flush_put_bits(&pb);
359
360 return put_bits_count(&pb)>>3;
361}
362
363static int g726_decode_frame(AVCodecContext *avctx,
364 void *data, int *data_size,
365 uint8_t *buf, int buf_size)
366{
367 AVG726Context *c = avctx->priv_data;
368 short *samples = data;
369 uint8_t code;
370 uint8_t mask;
371 GetBitContext gb;
372
373 if (!buf_size)
374 goto out;
375
376 mask = (1<<c->code_size) - 1;
377 init_get_bits(&gb, buf, buf_size * 8);
378 if (c->bits_left) {
379 int s = c->code_size - c->bits_left;;
380 code = (c->bit_buffer << s) | get_bits(&gb, s);
381 *samples++ = g726_decode(&c->c, code & mask);
382 }
383
384 while (get_bits_count(&gb) + c->code_size <= buf_size*8)
385 *samples++ = g726_decode(&c->c, get_bits(&gb, c->code_size) & mask);
386
387 c->bits_left = buf_size*8 - get_bits_count(&gb);
388 c->bit_buffer = get_bits(&gb, c->bits_left);
389
390out:
391 *data_size = (uint8_t*)samples - (uint8_t*)data;
392 return buf_size;
393}
394
395#ifdef CONFIG_ENCODERS
396AVCodec adpcm_g726_encoder = {
397 "g726",
398 CODEC_TYPE_AUDIO,
399 CODEC_ID_ADPCM_G726,
400 sizeof(AVG726Context),
401 g726_init,
402 g726_encode_frame,
403 NULL,
404 NULL,
405};
406#endif //CONFIG_ENCODERS
407
408AVCodec adpcm_g726_decoder = {
409 "g726",
410 CODEC_TYPE_AUDIO,
411 CODEC_ID_ADPCM_G726,
412 sizeof(AVG726Context),
413 g726_init,
414 NULL,
415 NULL,
416 g726_decode_frame,
417};