2 * G.726 ADPCM audio codec
3 * Copyright (c) 2004 Roman Shaposhnik.
5 * This is a very straightforward rendition of the G.726
6 * Section 4 "Computational Details".
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.
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.
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
28 * G.726 Standard uses rather odd 11bit floating point arithmentic for
29 * numerous occasions. It's a mistery to me why they did it this way
30 * instead of simply using 32bit integer arithmetic.
32 typedef struct Float11 {
33 int sign; /**< 1bit sign */
34 int exp; /**< 4bit exponent */
35 int mant; /**< 6bit mantissa */
38 static inline Float11* i2f(int16_t i, Float11* f)
43 f->exp = av_log2_16bit(i) + !!i;
44 f->mant = i? (i<<6) >> f->exp :
49 static inline int16_t mult(Float11* f1, Float11* f2)
53 exp = f1->exp + f2->exp;
54 res = (((f1->mant * f2->mant) + 0x30) >> 4) << 7;
55 res = exp > 26 ? res << (exp - 26) : res >> (26 - exp);
56 return (f1->sign ^ f2->sign) ? -res : res;
59 static inline int sgn(int value)
61 return (value < 0) ? -1 : 1;
64 typedef struct G726Tables {
65 int bits; /**< bits per sample */
66 int* quant; /**< quantization table */
67 int* iquant; /**< inverse quantization table */
68 int* W; /**< special table #1 ;-) */
69 int* F; /**< special table #2 */
72 typedef struct G726Context {
73 G726Tables* tbls; /**< static tables needed for computation */
75 Float11 sr[2]; /**< prev. reconstructed samples */
76 Float11 dq[6]; /**< prev. difference */
77 int a[2]; /**< second order predictor coeffs */
78 int b[6]; /**< sixth order predictor coeffs */
79 int pk[2]; /**< signs of prev. 2 sez + dq */
81 int ap; /**< scale factor control */
82 int yu; /**< fast scale factor */
83 int yl; /**< slow scale factor */
84 int dms; /**< short average magnitude of F[i] */
85 int dml; /**< long average magnitude of F[i] */
86 int td; /**< tone detect */
88 int se; /**< estimated signal for the next iteration */
89 int sez; /**< estimated second order prediction */
90 int y; /**< quantizer scaling factor for the next iteration */
93 static int quant_tbl16[] = /**< 16kbit/s 2bits per sample */
95 static int iquant_tbl16[] =
96 { 116, 365, 365, 116 };
97 static int W_tbl16[] =
98 { -22, 439, 439, -22 };
99 static int F_tbl16[] =
102 static int quant_tbl24[] = /**< 24kbit/s 3bits per sample */
103 { 7, 217, 330, INT_MAX };
104 static int iquant_tbl24[] =
105 { INT_MIN, 135, 273, 373, 373, 273, 135, INT_MIN };
106 static int W_tbl24[] =
107 { -4, 30, 137, 582, 582, 137, 30, -4 };
108 static int F_tbl24[] =
109 { 0, 1, 2, 7, 7, 2, 1, 0 };
111 static int quant_tbl32[] = /**< 32kbit/s 4bits per sample */
112 { -125, 79, 177, 245, 299, 348, 399, INT_MAX };
113 static int iquant_tbl32[] =
114 { INT_MIN, 4, 135, 213, 273, 323, 373, 425,
115 425, 373, 323, 273, 213, 135, 4, INT_MIN };
116 static int W_tbl32[] =
117 { -12, 18, 41, 64, 112, 198, 355, 1122,
118 1122, 355, 198, 112, 64, 41, 18, -12};
119 static int F_tbl32[] =
120 { 0, 0, 0, 1, 1, 1, 3, 7, 7, 3, 1, 1, 1, 0, 0, 0 };
122 static int quant_tbl40[] = /**< 40kbit/s 5bits per sample */
123 { -122, -16, 67, 138, 197, 249, 297, 338,
124 377, 412, 444, 474, 501, 527, 552, INT_MAX };
125 static int iquant_tbl40[] =
126 { INT_MIN, -66, 28, 104, 169, 224, 274, 318,
127 358, 395, 429, 459, 488, 514, 539, 566,
128 566, 539, 514, 488, 459, 429, 395, 358,
129 318, 274, 224, 169, 104, 28, -66, INT_MIN };
130 static int W_tbl40[] =
131 { 14, 14, 24, 39, 40, 41, 58, 100,
132 141, 179, 219, 280, 358, 440, 529, 696,
133 696, 529, 440, 358, 280, 219, 179, 141,
134 100, 58, 41, 40, 39, 24, 14, 14 };
135 static int F_tbl40[] =
136 { 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 3, 4, 5, 6, 6,
137 6, 6, 5, 4, 3, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
139 static G726Tables G726Tables_pool[] =
140 {{ 2, quant_tbl16, iquant_tbl16, W_tbl16, F_tbl16 },
141 { 3, quant_tbl24, iquant_tbl24, W_tbl24, F_tbl24 },
142 { 4, quant_tbl32, iquant_tbl32, W_tbl32, F_tbl32 },
143 { 5, quant_tbl40, iquant_tbl40, W_tbl40, F_tbl40 }};
147 * Para 4.2.2 page 18: Adaptive quantizer.
149 static inline uint8_t quant(G726Context* c, int d)
151 int sign, exp, i, dln;
158 exp = av_log2_16bit(d);
159 dln = ((exp<<7) + (((d<<7)>>exp)&0x7f)) - (c->y>>2);
161 while (c->tbls->quant[i] < INT_MAX && c->tbls->quant[i] < dln)
166 if (c->tbls->bits != 2 && i == 0) /* I'm not sure this is a good idea */
173 * Para 4.2.3 page 22: Inverse adaptive quantizer.
175 static inline int16_t inverse_quant(G726Context* c, int i)
179 dql = c->tbls->iquant[i] + (c->y >> 2);
180 dex = (dql>>7) & 0xf; /* 4bit exponent */
181 dqt = (1<<7) + (dql & 0x7f); /* log2 -> linear */
182 return (dql < 0) ? 0 : ((dqt<<7) >> (14-dex));
185 static inline int16_t g726_iterate(G726Context* c, int16_t I)
187 int dq, re_signal, pk0, fa1, i, tr, ylint, ylfrac, thr2, al, dq0;
190 dq = inverse_quant(c, I);
191 if (I >> (c->tbls->bits - 1)) /* get the sign */
193 re_signal = c->se + dq;
195 /* Transition detect */
196 ylint = (c->yl >> 15);
197 ylfrac = (c->yl >> 10) & 0x1f;
198 thr2 = (ylint > 9) ? 0x1f << 10 : (0x20 + ylfrac) << ylint;
199 if (c->td == 1 && abs(dq) > ((thr2+(thr2>>1))>>1))
204 /* Update second order predictor coefficient A2 and A1 */
205 pk0 = (c->sez + dq) ? sgn(c->sez + dq) : 0;
206 dq0 = dq ? sgn(dq) : 0;
213 /* This is a bit crazy, but it really is +255 not +256 */
214 fa1 = clip((-c->a[0]*c->pk[0]*pk0)>>5, -256, 255);
216 c->a[1] += 128*pk0*c->pk[1] + fa1 - (c->a[1]>>7);
217 c->a[1] = clip(c->a[1], -12288, 12288);
218 c->a[0] += 64*3*pk0*c->pk[0] - (c->a[0] >> 8);
219 c->a[0] = clip(c->a[0], -(15360 - c->a[1]), 15360 - c->a[1]);
222 c->b[i] += 128*dq0*sgn(-c->dq[i].sign) - (c->b[i]>>8);
225 /* Update Dq and Sr and Pk */
227 c->pk[0] = pk0 ? pk0 : 1;
229 i2f(re_signal, &c->sr[0]);
231 c->dq[i] = c->dq[i-1];
233 c->dq[0].sign = I >> (c->tbls->bits - 1); /* Isn't it crazy ?!?! */
235 /* Update tone detect [I'm not sure 'tr == 0' is really needed] */
236 c->td = (tr == 0 && c->a[1] < -11776);
239 c->dms += ((c->tbls->F[I]<<9) - c->dms) >> 5;
240 c->dml += ((c->tbls->F[I]<<11) - c->dml) >> 7;
243 else if (c->y > 1535 && !c->td && (abs((c->dms << 2) - c->dml) < (c->dml >> 3)))
244 c->ap += (-c->ap) >> 4;
246 c->ap += (0x200 - c->ap) >> 4;
248 /* Update Yu and Yl */
249 c->yu = clip(c->y + (((c->tbls->W[I] << 5) - c->y) >> 5), 544, 5120);
250 c->yl += c->yu + ((-c->yl)>>6);
252 /* Next iteration for Y */
253 al = (c->ap >= 256) ? 1<<6 : c->ap >> 2;
254 c->y = (c->yl + (c->yu - (c->yl>>6))*al) >> 6;
256 /* Next iteration for SE and SEZ */
259 c->se += mult(i2f(c->b[i] >> 2, &f), &c->dq[i]);
262 c->se += mult(i2f(c->a[i] >> 2, &f), &c->sr[i]);
265 return clip(re_signal << 2, -0xffff, 0xffff);
268 static int g726_reset(G726Context* c, int bit_rate)
272 c->tbls = &G726Tables_pool[bit_rate/8000 - 2];
273 for (i=0; i<2; i++) {
278 for (i=0; i<6; i++) {
296 static int16_t g726_decode(G726Context* c, int16_t i)
298 return g726_iterate(c, i);
301 static int16_t g726_encode(G726Context* c, int16_t sig)
305 i = quant(c, sig/4 - c->se) & ((1<<c->tbls->bits) - 1);
310 /* Interfacing to the libavcodec */
312 typedef struct AVG726Context {
319 static int g726_init(AVCodecContext * avctx)
321 AVG726Context* c = (AVG726Context*)avctx->priv_data;
323 if (avctx->channels != 1 ||
324 (avctx->bit_rate != 16000 && avctx->bit_rate != 24000 &&
325 avctx->bit_rate != 32000 && avctx->bit_rate != 40000)) {
326 av_log(avctx, AV_LOG_ERROR, "G726: unsupported audio format\n");
329 if (avctx->sample_rate != 8000 && avctx->strict_std_compliance>=0) {
330 av_log(avctx, AV_LOG_ERROR, "G726: unsupported audio format\n");
333 g726_reset(&c->c, avctx->bit_rate);
334 c->code_size = c->c.tbls->bits;
338 avctx->coded_frame = avcodec_alloc_frame();
339 if (!avctx->coded_frame)
341 avctx->coded_frame->key_frame = 1;
346 static int g726_close(AVCodecContext *avctx)
348 av_freep(&avctx->coded_frame);
352 static int g726_encode_frame(AVCodecContext *avctx,
353 uint8_t *dst, int buf_size, void *data)
355 AVG726Context *c = avctx->priv_data;
356 short *samples = data;
359 init_put_bits(&pb, dst, 1024*1024);
361 for (; buf_size; buf_size--)
362 put_bits(&pb, c->code_size, g726_encode(&c->c, *samples++));
366 return put_bits_count(&pb)>>3;
369 static int g726_decode_frame(AVCodecContext *avctx,
370 void *data, int *data_size,
371 uint8_t *buf, int buf_size)
373 AVG726Context *c = avctx->priv_data;
374 short *samples = data;
382 mask = (1<<c->code_size) - 1;
383 init_get_bits(&gb, buf, buf_size * 8);
385 int s = c->code_size - c->bits_left;;
386 code = (c->bit_buffer << s) | get_bits(&gb, s);
387 *samples++ = g726_decode(&c->c, code & mask);
390 while (get_bits_count(&gb) + c->code_size <= buf_size*8)
391 *samples++ = g726_decode(&c->c, get_bits(&gb, c->code_size) & mask);
393 c->bits_left = buf_size*8 - get_bits_count(&gb);
394 c->bit_buffer = get_bits(&gb, c->bits_left);
397 *data_size = (uint8_t*)samples - (uint8_t*)data;
401 #ifdef CONFIG_ENCODERS
402 AVCodec adpcm_g726_encoder = {
406 sizeof(AVG726Context),
412 #endif //CONFIG_ENCODERS
414 AVCodec adpcm_g726_decoder = {
418 sizeof(AVG726Context),