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 FFmpeg.
10 * FFmpeg 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 * FFmpeg 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 FFmpeg; 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"
26 #include "libavutil/opt.h"
34 * G.726 Standard uses rather odd 11bit floating point arithmentic for
35 * numerous occasions. It's a mistery to me why they did it this way
36 * instead of simply using 32bit integer arithmetic.
38 typedef struct Float11 {
39 uint8_t sign; /**< 1bit sign */
40 uint8_t exp; /**< 4bit exponent */
41 uint8_t mant; /**< 6bit mantissa */
44 static inline Float11* i2f(int i, Float11* f)
49 f->exp = av_log2_16bit(i) + !!i;
50 f->mant = i? (i<<6) >> f->exp : 1<<5;
54 static inline int16_t mult(Float11* f1, Float11* f2)
58 exp = f1->exp + f2->exp;
59 res = (((f1->mant * f2->mant) + 0x30) >> 4);
60 res = exp > 19 ? res << (exp - 19) : res >> (19 - exp);
61 return (f1->sign ^ f2->sign) ? -res : res;
64 static inline int sgn(int value)
66 return (value < 0) ? -1 : 1;
69 typedef struct G726Tables {
70 const int* quant; /**< quantization table */
71 const int16_t* iquant; /**< inverse quantization table */
72 const int16_t* W; /**< special table #1 ;-) */
73 const uint8_t* F; /**< special table #2 */
76 typedef struct G726Context {
79 G726Tables tbls; /**< static tables needed for computation */
81 Float11 sr[2]; /**< prev. reconstructed samples */
82 Float11 dq[6]; /**< prev. difference */
83 int a[2]; /**< second order predictor coeffs */
84 int b[6]; /**< sixth order predictor coeffs */
85 int pk[2]; /**< signs of prev. 2 sez + dq */
87 int ap; /**< scale factor control */
88 int yu; /**< fast scale factor */
89 int yl; /**< slow scale factor */
90 int dms; /**< short average magnitude of F[i] */
91 int dml; /**< long average magnitude of F[i] */
92 int td; /**< tone detect */
94 int se; /**< estimated signal for the next iteration */
95 int sez; /**< estimated second order prediction */
96 int y; /**< quantizer scaling factor for the next iteration */
100 static const int quant_tbl16[] = /**< 16kbit/s 2bits per sample */
102 static const int16_t iquant_tbl16[] =
103 { 116, 365, 365, 116 };
104 static const int16_t W_tbl16[] =
105 { -22, 439, 439, -22 };
106 static const uint8_t F_tbl16[] =
109 static const int quant_tbl24[] = /**< 24kbit/s 3bits per sample */
110 { 7, 217, 330, INT_MAX };
111 static const int16_t iquant_tbl24[] =
112 { INT16_MIN, 135, 273, 373, 373, 273, 135, INT16_MIN };
113 static const int16_t W_tbl24[] =
114 { -4, 30, 137, 582, 582, 137, 30, -4 };
115 static const uint8_t F_tbl24[] =
116 { 0, 1, 2, 7, 7, 2, 1, 0 };
118 static const int quant_tbl32[] = /**< 32kbit/s 4bits per sample */
119 { -125, 79, 177, 245, 299, 348, 399, INT_MAX };
120 static const int16_t iquant_tbl32[] =
121 { INT16_MIN, 4, 135, 213, 273, 323, 373, 425,
122 425, 373, 323, 273, 213, 135, 4, INT16_MIN };
123 static const int16_t W_tbl32[] =
124 { -12, 18, 41, 64, 112, 198, 355, 1122,
125 1122, 355, 198, 112, 64, 41, 18, -12};
126 static const uint8_t F_tbl32[] =
127 { 0, 0, 0, 1, 1, 1, 3, 7, 7, 3, 1, 1, 1, 0, 0, 0 };
129 static const int quant_tbl40[] = /**< 40kbit/s 5bits per sample */
130 { -122, -16, 67, 138, 197, 249, 297, 338,
131 377, 412, 444, 474, 501, 527, 552, INT_MAX };
132 static const int16_t iquant_tbl40[] =
133 { INT16_MIN, -66, 28, 104, 169, 224, 274, 318,
134 358, 395, 429, 459, 488, 514, 539, 566,
135 566, 539, 514, 488, 459, 429, 395, 358,
136 318, 274, 224, 169, 104, 28, -66, INT16_MIN };
137 static const int16_t W_tbl40[] =
138 { 14, 14, 24, 39, 40, 41, 58, 100,
139 141, 179, 219, 280, 358, 440, 529, 696,
140 696, 529, 440, 358, 280, 219, 179, 141,
141 100, 58, 41, 40, 39, 24, 14, 14 };
142 static const uint8_t F_tbl40[] =
143 { 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 3, 4, 5, 6, 6,
144 6, 6, 5, 4, 3, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
146 static const G726Tables G726Tables_pool[] =
147 {{ quant_tbl16, iquant_tbl16, W_tbl16, F_tbl16 },
148 { quant_tbl24, iquant_tbl24, W_tbl24, F_tbl24 },
149 { quant_tbl32, iquant_tbl32, W_tbl32, F_tbl32 },
150 { quant_tbl40, iquant_tbl40, W_tbl40, F_tbl40 }};
154 * Para 4.2.2 page 18: Adaptive quantizer.
156 static inline uint8_t quant(G726Context* c, int d)
158 int sign, exp, i, dln;
165 exp = av_log2_16bit(d);
166 dln = ((exp<<7) + (((d<<7)>>exp)&0x7f)) - (c->y>>2);
168 while (c->tbls.quant[i] < INT_MAX && c->tbls.quant[i] < dln)
173 if (c->code_size != 2 && i == 0) /* I'm not sure this is a good idea */
180 * Para 4.2.3 page 22: Inverse adaptive quantizer.
182 static inline int16_t inverse_quant(G726Context* c, int i)
186 dql = c->tbls.iquant[i] + (c->y >> 2);
187 dex = (dql>>7) & 0xf; /* 4bit exponent */
188 dqt = (1<<7) + (dql & 0x7f); /* log2 -> linear */
189 return (dql < 0) ? 0 : ((dqt<<dex) >> 7);
192 static int16_t g726_decode(G726Context* c, int I)
194 int dq, re_signal, pk0, fa1, i, tr, ylint, ylfrac, thr2, al, dq0;
196 int I_sig= I >> (c->code_size - 1);
198 dq = inverse_quant(c, I);
200 /* Transition detect */
201 ylint = (c->yl >> 15);
202 ylfrac = (c->yl >> 10) & 0x1f;
203 thr2 = (ylint > 9) ? 0x1f << 10 : (0x20 + ylfrac) << ylint;
204 tr= (c->td == 1 && dq > ((3*thr2)>>2));
206 if (I_sig) /* get the sign */
208 re_signal = c->se + dq;
210 /* Update second order predictor coefficient A2 and A1 */
211 pk0 = (c->sez + dq) ? sgn(c->sez + dq) : 0;
212 dq0 = dq ? sgn(dq) : 0;
219 /* This is a bit crazy, but it really is +255 not +256 */
220 fa1 = av_clip((-c->a[0]*c->pk[0]*pk0)>>5, -256, 255);
222 c->a[1] += 128*pk0*c->pk[1] + fa1 - (c->a[1]>>7);
223 c->a[1] = av_clip(c->a[1], -12288, 12288);
224 c->a[0] += 64*3*pk0*c->pk[0] - (c->a[0] >> 8);
225 c->a[0] = av_clip(c->a[0], -(15360 - c->a[1]), 15360 - c->a[1]);
228 c->b[i] += 128*dq0*sgn(-c->dq[i].sign) - (c->b[i]>>8);
231 /* Update Dq and Sr and Pk */
233 c->pk[0] = pk0 ? pk0 : 1;
235 i2f(re_signal, &c->sr[0]);
237 c->dq[i] = c->dq[i-1];
239 c->dq[0].sign = I_sig; /* Isn't it crazy ?!?! */
241 c->td = c->a[1] < -11776;
244 c->dms += (c->tbls.F[I]<<4) + ((- c->dms) >> 5);
245 c->dml += (c->tbls.F[I]<<4) + ((- c->dml) >> 7);
249 c->ap += (-c->ap) >> 4;
250 if (c->y <= 1535 || c->td || abs((c->dms << 2) - c->dml) >= (c->dml >> 3))
254 /* Update Yu and Yl */
255 c->yu = av_clip(c->y + c->tbls.W[I] + ((-c->y)>>5), 544, 5120);
256 c->yl += c->yu + ((-c->yl)>>6);
258 /* Next iteration for Y */
259 al = (c->ap >= 256) ? 1<<6 : c->ap >> 2;
260 c->y = (c->yl + (c->yu - (c->yl>>6))*al) >> 6;
262 /* Next iteration for SE and SEZ */
265 c->se += mult(i2f(c->b[i] >> 2, &f), &c->dq[i]);
268 c->se += mult(i2f(c->a[i] >> 2, &f), &c->sr[i]);
271 return av_clip(re_signal << 2, -0xffff, 0xffff);
274 static av_cold int g726_reset(G726Context *c)
278 c->tbls = G726Tables_pool[c->code_size - 2];
279 for (i=0; i<2; i++) {
280 c->sr[i].mant = 1<<5;
283 for (i=0; i<6; i++) {
284 c->dq[i].mant = 1<<5;
294 #if CONFIG_ADPCM_G726_ENCODER
295 static int16_t g726_encode(G726Context* c, int16_t sig)
299 i = quant(c, sig/4 - c->se) & ((1<<c->code_size) - 1);
304 /* Interfacing to the libavcodec */
306 static av_cold int g726_encode_init(AVCodecContext *avctx)
308 G726Context* c = avctx->priv_data;
310 if (avctx->strict_std_compliance > FF_COMPLIANCE_UNOFFICIAL &&
311 avctx->sample_rate != 8000) {
312 av_log(avctx, AV_LOG_ERROR, "Sample rates other than 8kHz are not "
313 "allowed when the compliance level is higher than unofficial. "
314 "Resample or reduce the compliance level.\n");
315 return AVERROR(EINVAL);
317 av_assert0(avctx->sample_rate > 0);
319 if(avctx->channels != 1){
320 av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
321 return AVERROR(EINVAL);
325 c->code_size = (avctx->bit_rate + avctx->sample_rate/2) / avctx->sample_rate;
327 c->code_size = av_clip(c->code_size, 2, 5);
328 avctx->bit_rate = c->code_size * avctx->sample_rate;
329 avctx->bits_per_coded_sample = c->code_size;
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;
369 #define OFFSET(x) offsetof(G726Context, x)
370 #define AE AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
371 static const AVOption options[] = {
372 { "code_size", "Bits per code", OFFSET(code_size), AV_OPT_TYPE_INT, { 4 }, 2, 5, AE },
376 static const AVClass class = {
377 .class_name = "g726",
378 .item_name = av_default_item_name,
380 .version = LIBAVUTIL_VERSION_INT,
383 static const AVCodecDefault defaults[] = {
388 AVCodec ff_adpcm_g726_encoder = {
390 .type = AVMEDIA_TYPE_AUDIO,
391 .id = CODEC_ID_ADPCM_G726,
392 .priv_data_size = sizeof(G726Context),
393 .init = g726_encode_init,
394 .encode = g726_encode_frame,
395 .close = g726_encode_close,
396 .capabilities = CODEC_CAP_SMALL_LAST_FRAME,
397 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},
398 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),
399 .priv_class = &class,
400 .defaults = defaults,
404 #if CONFIG_ADPCM_G726_DECODER
405 static av_cold int g726_decode_init(AVCodecContext *avctx)
407 G726Context* c = avctx->priv_data;
409 if (avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
410 avctx->sample_rate != 8000) {
411 av_log(avctx, AV_LOG_ERROR, "Only 8kHz sample rate is allowed when "
412 "the compliance level is strict. Reduce the compliance level "
413 "if you wish to decode the stream anyway.\n");
414 return AVERROR(EINVAL);
417 if(avctx->channels != 1){
418 av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
419 return AVERROR(EINVAL);
422 c->code_size = avctx->bits_per_coded_sample;
423 if (c->code_size < 2 || c->code_size > 5) {
424 av_log(avctx, AV_LOG_ERROR, "Invalid number of bits %d\n", c->code_size);
425 return AVERROR(EINVAL);
429 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
431 avcodec_get_frame_defaults(&c->frame);
432 avctx->coded_frame = &c->frame;
437 static int g726_decode_frame(AVCodecContext *avctx, void *data,
438 int *got_frame_ptr, AVPacket *avpkt)
440 const uint8_t *buf = avpkt->data;
441 int buf_size = avpkt->size;
442 G726Context *c = avctx->priv_data;
445 int out_samples, ret;
447 out_samples = buf_size * 8 / c->code_size;
449 /* get output buffer */
450 c->frame.nb_samples = out_samples;
451 if ((ret = avctx->get_buffer(avctx, &c->frame)) < 0) {
452 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
455 samples = (int16_t *)c->frame.data[0];
457 init_get_bits(&gb, buf, buf_size * 8);
459 while (out_samples--)
460 *samples++ = g726_decode(c, get_bits(&gb, c->code_size));
462 if (get_bits_left(&gb) > 0)
463 av_log(avctx, AV_LOG_ERROR, "Frame invalidly split, missing parser?\n");
466 *(AVFrame *)data = c->frame;
471 static void g726_decode_flush(AVCodecContext *avctx)
473 G726Context *c = avctx->priv_data;
477 AVCodec ff_adpcm_g726_decoder = {
479 .type = AVMEDIA_TYPE_AUDIO,
480 .id = CODEC_ID_ADPCM_G726,
481 .priv_data_size = sizeof(G726Context),
482 .init = g726_decode_init,
483 .decode = g726_decode_frame,
484 .flush = g726_decode_flush,
485 .capabilities = CODEC_CAP_DR1,
486 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),