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 file is part of Libav.
10 * Libav is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
15 * Libav is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with Libav; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 #include "libavutil/avassert.h"
32 * G.726 Standard uses rather odd 11bit floating point arithmentic for
33 * numerous occasions. It's a mistery to me why they did it this way
34 * instead of simply using 32bit integer arithmetic.
36 typedef struct Float11 {
37 uint8_t sign; /**< 1bit sign */
38 uint8_t exp; /**< 4bit exponent */
39 uint8_t mant; /**< 6bit mantissa */
42 static inline Float11* i2f(int i, Float11* f)
47 f->exp = av_log2_16bit(i) + !!i;
48 f->mant = i? (i<<6) >> f->exp : 1<<5;
52 static inline int16_t mult(Float11* f1, Float11* f2)
56 exp = f1->exp + f2->exp;
57 res = (((f1->mant * f2->mant) + 0x30) >> 4);
58 res = exp > 19 ? res << (exp - 19) : res >> (19 - exp);
59 return (f1->sign ^ f2->sign) ? -res : res;
62 static inline int sgn(int value)
64 return (value < 0) ? -1 : 1;
67 typedef struct G726Tables {
68 const int* quant; /**< quantization table */
69 const int16_t* iquant; /**< inverse quantization table */
70 const int16_t* W; /**< special table #1 ;-) */
71 const uint8_t* F; /**< special table #2 */
74 typedef struct G726Context {
75 G726Tables tbls; /**< static tables needed for computation */
77 Float11 sr[2]; /**< prev. reconstructed samples */
78 Float11 dq[6]; /**< prev. difference */
79 int a[2]; /**< second order predictor coeffs */
80 int b[6]; /**< sixth order predictor coeffs */
81 int pk[2]; /**< signs of prev. 2 sez + dq */
83 int ap; /**< scale factor control */
84 int yu; /**< fast scale factor */
85 int yl; /**< slow scale factor */
86 int dms; /**< short average magnitude of F[i] */
87 int dml; /**< long average magnitude of F[i] */
88 int td; /**< tone detect */
90 int se; /**< estimated signal for the next iteration */
91 int sez; /**< estimated second order prediction */
92 int y; /**< quantizer scaling factor for the next iteration */
96 static const int quant_tbl16[] = /**< 16kbit/s 2bits per sample */
98 static const int16_t iquant_tbl16[] =
99 { 116, 365, 365, 116 };
100 static const int16_t W_tbl16[] =
101 { -22, 439, 439, -22 };
102 static const uint8_t F_tbl16[] =
105 static const int quant_tbl24[] = /**< 24kbit/s 3bits per sample */
106 { 7, 217, 330, INT_MAX };
107 static const int16_t iquant_tbl24[] =
108 { INT16_MIN, 135, 273, 373, 373, 273, 135, INT16_MIN };
109 static const int16_t W_tbl24[] =
110 { -4, 30, 137, 582, 582, 137, 30, -4 };
111 static const uint8_t F_tbl24[] =
112 { 0, 1, 2, 7, 7, 2, 1, 0 };
114 static const int quant_tbl32[] = /**< 32kbit/s 4bits per sample */
115 { -125, 79, 177, 245, 299, 348, 399, INT_MAX };
116 static const int16_t iquant_tbl32[] =
117 { INT16_MIN, 4, 135, 213, 273, 323, 373, 425,
118 425, 373, 323, 273, 213, 135, 4, INT16_MIN };
119 static const int16_t W_tbl32[] =
120 { -12, 18, 41, 64, 112, 198, 355, 1122,
121 1122, 355, 198, 112, 64, 41, 18, -12};
122 static const uint8_t F_tbl32[] =
123 { 0, 0, 0, 1, 1, 1, 3, 7, 7, 3, 1, 1, 1, 0, 0, 0 };
125 static const int quant_tbl40[] = /**< 40kbit/s 5bits per sample */
126 { -122, -16, 67, 138, 197, 249, 297, 338,
127 377, 412, 444, 474, 501, 527, 552, INT_MAX };
128 static const int16_t iquant_tbl40[] =
129 { INT16_MIN, -66, 28, 104, 169, 224, 274, 318,
130 358, 395, 429, 459, 488, 514, 539, 566,
131 566, 539, 514, 488, 459, 429, 395, 358,
132 318, 274, 224, 169, 104, 28, -66, INT16_MIN };
133 static const int16_t W_tbl40[] =
134 { 14, 14, 24, 39, 40, 41, 58, 100,
135 141, 179, 219, 280, 358, 440, 529, 696,
136 696, 529, 440, 358, 280, 219, 179, 141,
137 100, 58, 41, 40, 39, 24, 14, 14 };
138 static const uint8_t F_tbl40[] =
139 { 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 3, 4, 5, 6, 6,
140 6, 6, 5, 4, 3, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
142 static const G726Tables G726Tables_pool[] =
143 {{ quant_tbl16, iquant_tbl16, W_tbl16, F_tbl16 },
144 { quant_tbl24, iquant_tbl24, W_tbl24, F_tbl24 },
145 { quant_tbl32, iquant_tbl32, W_tbl32, F_tbl32 },
146 { quant_tbl40, iquant_tbl40, W_tbl40, F_tbl40 }};
150 * Para 4.2.2 page 18: Adaptive quantizer.
152 static inline uint8_t quant(G726Context* c, int d)
154 int sign, exp, i, dln;
161 exp = av_log2_16bit(d);
162 dln = ((exp<<7) + (((d<<7)>>exp)&0x7f)) - (c->y>>2);
164 while (c->tbls.quant[i] < INT_MAX && c->tbls.quant[i] < dln)
169 if (c->code_size != 2 && i == 0) /* I'm not sure this is a good idea */
176 * Para 4.2.3 page 22: Inverse adaptive quantizer.
178 static inline int16_t inverse_quant(G726Context* c, int i)
182 dql = c->tbls.iquant[i] + (c->y >> 2);
183 dex = (dql>>7) & 0xf; /* 4bit exponent */
184 dqt = (1<<7) + (dql & 0x7f); /* log2 -> linear */
185 return (dql < 0) ? 0 : ((dqt<<dex) >> 7);
188 static int16_t g726_decode(G726Context* c, int I)
190 int dq, re_signal, pk0, fa1, i, tr, ylint, ylfrac, thr2, al, dq0;
192 int I_sig= I >> (c->code_size - 1);
194 dq = inverse_quant(c, I);
196 /* Transition detect */
197 ylint = (c->yl >> 15);
198 ylfrac = (c->yl >> 10) & 0x1f;
199 thr2 = (ylint > 9) ? 0x1f << 10 : (0x20 + ylfrac) << ylint;
200 tr= (c->td == 1 && dq > ((3*thr2)>>2));
202 if (I_sig) /* get the sign */
204 re_signal = c->se + dq;
206 /* Update second order predictor coefficient A2 and A1 */
207 pk0 = (c->sez + dq) ? sgn(c->sez + dq) : 0;
208 dq0 = dq ? sgn(dq) : 0;
215 /* This is a bit crazy, but it really is +255 not +256 */
216 fa1 = av_clip((-c->a[0]*c->pk[0]*pk0)>>5, -256, 255);
218 c->a[1] += 128*pk0*c->pk[1] + fa1 - (c->a[1]>>7);
219 c->a[1] = av_clip(c->a[1], -12288, 12288);
220 c->a[0] += 64*3*pk0*c->pk[0] - (c->a[0] >> 8);
221 c->a[0] = av_clip(c->a[0], -(15360 - c->a[1]), 15360 - c->a[1]);
224 c->b[i] += 128*dq0*sgn(-c->dq[i].sign) - (c->b[i]>>8);
227 /* Update Dq and Sr and Pk */
229 c->pk[0] = pk0 ? pk0 : 1;
231 i2f(re_signal, &c->sr[0]);
233 c->dq[i] = c->dq[i-1];
235 c->dq[0].sign = I_sig; /* Isn't it crazy ?!?! */
237 c->td = c->a[1] < -11776;
240 c->dms += (c->tbls.F[I]<<4) + ((- c->dms) >> 5);
241 c->dml += (c->tbls.F[I]<<4) + ((- c->dml) >> 7);
245 c->ap += (-c->ap) >> 4;
246 if (c->y <= 1535 || c->td || abs((c->dms << 2) - c->dml) >= (c->dml >> 3))
250 /* Update Yu and Yl */
251 c->yu = av_clip(c->y + c->tbls.W[I] + ((-c->y)>>5), 544, 5120);
252 c->yl += c->yu + ((-c->yl)>>6);
254 /* Next iteration for Y */
255 al = (c->ap >= 256) ? 1<<6 : c->ap >> 2;
256 c->y = (c->yl + (c->yu - (c->yl>>6))*al) >> 6;
258 /* Next iteration for SE and SEZ */
261 c->se += mult(i2f(c->b[i] >> 2, &f), &c->dq[i]);
264 c->se += mult(i2f(c->a[i] >> 2, &f), &c->sr[i]);
267 return av_clip(re_signal << 2, -0xffff, 0xffff);
270 static av_cold int g726_reset(G726Context* c, int index)
274 c->tbls = G726Tables_pool[index];
275 for (i=0; i<2; i++) {
276 c->sr[i].mant = 1<<5;
279 for (i=0; i<6; i++) {
280 c->dq[i].mant = 1<<5;
290 #if CONFIG_ADPCM_G726_ENCODER
291 static int16_t g726_encode(G726Context* c, int16_t sig)
295 i = quant(c, sig/4 - c->se) & ((1<<c->code_size) - 1);
300 /* Interfacing to the libavcodec */
302 static av_cold int g726_encode_init(AVCodecContext *avctx)
304 G726Context* c = avctx->priv_data;
306 if (avctx->strict_std_compliance > FF_COMPLIANCE_UNOFFICIAL &&
307 avctx->sample_rate != 8000) {
308 av_log(avctx, AV_LOG_ERROR, "Sample rates other than 8kHz are not "
309 "allowed when the compliance level is higher than unofficial. "
310 "Resample or reduce the compliance level.\n");
311 return AVERROR(EINVAL);
313 av_assert0(avctx->sample_rate > 0);
315 if(avctx->channels != 1){
316 av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
320 if (avctx->bit_rate % avctx->sample_rate) {
321 av_log(avctx, AV_LOG_ERROR, "Bitrate - Samplerate combination is invalid\n");
322 return AVERROR(EINVAL);
324 c->code_size = (avctx->bit_rate + avctx->sample_rate/2) / avctx->sample_rate;
325 if (c->code_size < 2 || c->code_size > 5) {
326 av_log(avctx, AV_LOG_ERROR, "Invalid number of bits %d\n", c->code_size);
327 return AVERROR(EINVAL);
329 avctx->bits_per_coded_sample = c->code_size;
331 g726_reset(c, c->code_size - 2);
333 avctx->coded_frame = avcodec_alloc_frame();
334 if (!avctx->coded_frame)
335 return AVERROR(ENOMEM);
336 avctx->coded_frame->key_frame = 1;
338 /* select a frame size that will end on a byte boundary and have a size of
339 approximately 1024 bytes */
340 avctx->frame_size = ((int[]){ 4096, 2736, 2048, 1640 })[c->code_size - 2];
345 static av_cold int g726_encode_close(AVCodecContext *avctx)
347 av_freep(&avctx->coded_frame);
351 static int g726_encode_frame(AVCodecContext *avctx,
352 uint8_t *dst, int buf_size, void *data)
354 G726Context *c = avctx->priv_data;
355 const int16_t *samples = data;
359 init_put_bits(&pb, dst, 1024*1024);
361 for (i = 0; i < avctx->frame_size; i++)
362 put_bits(&pb, c->code_size, g726_encode(c, *samples++));
366 return put_bits_count(&pb)>>3;
370 static av_cold int g726_decode_init(AVCodecContext *avctx)
372 G726Context* c = avctx->priv_data;
374 if (avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
375 avctx->sample_rate != 8000) {
376 av_log(avctx, AV_LOG_ERROR, "Only 8kHz sample rate is allowed when "
377 "the compliance level is strict. Reduce the compliance level "
378 "if you wish to decode the stream anyway.\n");
379 return AVERROR(EINVAL);
382 if(avctx->channels != 1){
383 av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
387 c->code_size = avctx->bits_per_coded_sample;
388 if (c->code_size < 2 || c->code_size > 5) {
389 av_log(avctx, AV_LOG_ERROR, "Invalid number of bits %d\n", c->code_size);
390 return AVERROR(EINVAL);
392 g726_reset(c, c->code_size - 2);
394 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
399 static int g726_decode_frame(AVCodecContext *avctx,
400 void *data, int *data_size,
403 const uint8_t *buf = avpkt->data;
404 int buf_size = avpkt->size;
405 G726Context *c = avctx->priv_data;
406 int16_t *samples = data;
408 int out_samples, out_size;
410 out_samples = buf_size * 8 / c->code_size;
411 out_size = out_samples * av_get_bytes_per_sample(avctx->sample_fmt);
412 if (*data_size < out_size) {
413 av_log(avctx, AV_LOG_ERROR, "Output buffer is too small\n");
414 return AVERROR(EINVAL);
417 init_get_bits(&gb, buf, buf_size * 8);
419 while (out_samples--)
420 *samples++ = g726_decode(c, get_bits(&gb, c->code_size));
422 if (get_bits_left(&gb) > 0)
423 av_log(avctx, AV_LOG_ERROR, "Frame invalidly split, missing parser?\n");
425 *data_size = out_size;
429 #if CONFIG_ADPCM_G726_ENCODER
430 AVCodec ff_adpcm_g726_encoder = {
432 .type = AVMEDIA_TYPE_AUDIO,
433 .id = CODEC_ID_ADPCM_G726,
434 .priv_data_size = sizeof(G726Context),
435 .init = g726_encode_init,
436 .encode = g726_encode_frame,
437 .close = g726_encode_close,
438 .capabilities = CODEC_CAP_SMALL_LAST_FRAME,
439 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},
440 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),
444 AVCodec ff_adpcm_g726_decoder = {
446 .type = AVMEDIA_TYPE_AUDIO,
447 .id = CODEC_ID_ADPCM_G726,
448 .priv_data_size = sizeof(G726Context),
449 .init = g726_decode_init,
450 .decode = g726_decode_frame,
451 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),