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
26 #include "libavutil/channel_layout.h"
27 #include "libavutil/opt.h"
35 * G.726 Standard uses rather odd 11-bit floating point arithmetic for
36 * numerous occasions. It's a mystery to me why they did it this way
37 * instead of simply using 32-bit integer arithmetic.
39 typedef struct Float11 {
40 uint8_t sign; /**< 1 bit sign */
41 uint8_t exp; /**< 4 bits exponent */
42 uint8_t mant; /**< 6 bits mantissa */
45 static inline Float11* i2f(int i, Float11* f)
50 f->exp = av_log2_16bit(i) + !!i;
51 f->mant = i? (i<<6) >> f->exp : 1<<5;
55 static inline int16_t mult(Float11* f1, Float11* f2)
59 exp = f1->exp + f2->exp;
60 res = (((f1->mant * f2->mant) + 0x30) >> 4);
61 res = exp > 19 ? res << (exp - 19) : res >> (19 - exp);
62 return (f1->sign ^ f2->sign) ? -res : res;
65 static inline int sgn(int value)
67 return (value < 0) ? -1 : 1;
70 typedef struct G726Tables {
71 const int* quant; /**< quantization table */
72 const int16_t* iquant; /**< inverse quantization table */
73 const int16_t* W; /**< special table #1 ;-) */
74 const uint8_t* F; /**< special table #2 */
77 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 */
98 int little_endian; /**< little-endian bitstream as used in aiff and Sun AU */
101 static const int quant_tbl16[] = /**< 16kbit/s 2 bits per sample */
103 static const int16_t iquant_tbl16[] =
104 { 116, 365, 365, 116 };
105 static const int16_t W_tbl16[] =
106 { -22, 439, 439, -22 };
107 static const uint8_t F_tbl16[] =
110 static const int quant_tbl24[] = /**< 24kbit/s 3 bits per sample */
111 { 7, 217, 330, INT_MAX };
112 static const int16_t iquant_tbl24[] =
113 { INT16_MIN, 135, 273, 373, 373, 273, 135, INT16_MIN };
114 static const int16_t W_tbl24[] =
115 { -4, 30, 137, 582, 582, 137, 30, -4 };
116 static const uint8_t F_tbl24[] =
117 { 0, 1, 2, 7, 7, 2, 1, 0 };
119 static const int quant_tbl32[] = /**< 32kbit/s 4 bits per sample */
120 { -125, 79, 177, 245, 299, 348, 399, INT_MAX };
121 static const int16_t iquant_tbl32[] =
122 { INT16_MIN, 4, 135, 213, 273, 323, 373, 425,
123 425, 373, 323, 273, 213, 135, 4, INT16_MIN };
124 static const int16_t W_tbl32[] =
125 { -12, 18, 41, 64, 112, 198, 355, 1122,
126 1122, 355, 198, 112, 64, 41, 18, -12};
127 static const uint8_t F_tbl32[] =
128 { 0, 0, 0, 1, 1, 1, 3, 7, 7, 3, 1, 1, 1, 0, 0, 0 };
130 static const int quant_tbl40[] = /**< 40kbit/s 5 bits per sample */
131 { -122, -16, 67, 138, 197, 249, 297, 338,
132 377, 412, 444, 474, 501, 527, 552, INT_MAX };
133 static const int16_t iquant_tbl40[] =
134 { INT16_MIN, -66, 28, 104, 169, 224, 274, 318,
135 358, 395, 429, 459, 488, 514, 539, 566,
136 566, 539, 514, 488, 459, 429, 395, 358,
137 318, 274, 224, 169, 104, 28, -66, INT16_MIN };
138 static const int16_t W_tbl40[] =
139 { 14, 14, 24, 39, 40, 41, 58, 100,
140 141, 179, 219, 280, 358, 440, 529, 696,
141 696, 529, 440, 358, 280, 219, 179, 141,
142 100, 58, 41, 40, 39, 24, 14, 14 };
143 static const uint8_t F_tbl40[] =
144 { 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 3, 4, 5, 6, 6,
145 6, 6, 5, 4, 3, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
147 static const G726Tables G726Tables_pool[] =
148 {{ quant_tbl16, iquant_tbl16, W_tbl16, F_tbl16 },
149 { quant_tbl24, iquant_tbl24, W_tbl24, F_tbl24 },
150 { quant_tbl32, iquant_tbl32, W_tbl32, F_tbl32 },
151 { quant_tbl40, iquant_tbl40, W_tbl40, F_tbl40 }};
155 * Paragraph 4.2.2 page 18: Adaptive quantizer.
157 static inline uint8_t quant(G726Context* c, int d)
159 int sign, exp, i, dln;
166 exp = av_log2_16bit(d);
167 dln = ((exp<<7) + (((d<<7)>>exp)&0x7f)) - (c->y>>2);
169 while (c->tbls.quant[i] < INT_MAX && c->tbls.quant[i] < dln)
174 if (c->code_size != 2 && i == 0) /* I'm not sure this is a good idea */
181 * Paragraph 4.2.3 page 22: Inverse adaptive quantizer.
183 static inline int16_t inverse_quant(G726Context* c, int i)
187 dql = c->tbls.iquant[i] + (c->y >> 2);
188 dex = (dql>>7) & 0xf; /* 4-bit exponent */
189 dqt = (1<<7) + (dql & 0x7f); /* log2 -> linear */
190 return (dql < 0) ? 0 : ((dqt<<dex) >> 7);
193 static int16_t g726_decode(G726Context* c, int I)
195 int dq, re_signal, pk0, fa1, i, tr, ylint, ylfrac, thr2, al, dq0;
197 int I_sig= I >> (c->code_size - 1);
199 dq = inverse_quant(c, I);
201 /* Transition detect */
202 ylint = (c->yl >> 15);
203 ylfrac = (c->yl >> 10) & 0x1f;
204 thr2 = (ylint > 9) ? 0x1f << 10 : (0x20 + ylfrac) << ylint;
205 tr= (c->td == 1 && dq > ((3*thr2)>>2));
207 if (I_sig) /* get the sign */
209 re_signal = (int16_t)(c->se + dq);
211 /* Update second order predictor coefficient A2 and A1 */
212 pk0 = (c->sez + dq) ? sgn(c->sez + dq) : 0;
213 dq0 = dq ? sgn(dq) : 0;
220 /* This is a bit crazy, but it really is +255 not +256 */
221 fa1 = av_clip_intp2((-c->a[0]*c->pk[0]*pk0)>>5, 8);
223 c->a[1] += 128*pk0*c->pk[1] + fa1 - (c->a[1]>>7);
224 c->a[1] = av_clip(c->a[1], -12288, 12288);
225 c->a[0] += 64*3*pk0*c->pk[0] - (c->a[0] >> 8);
226 c->a[0] = av_clip(c->a[0], -(15360 - c->a[1]), 15360 - c->a[1]);
229 c->b[i] += 128*dq0*sgn(-c->dq[i].sign) - (c->b[i]>>8);
232 /* Update Dq and Sr and Pk */
234 c->pk[0] = pk0 ? pk0 : 1;
236 i2f(re_signal, &c->sr[0]);
238 c->dq[i] = c->dq[i-1];
240 c->dq[0].sign = I_sig; /* Isn't it crazy ?!?! */
242 c->td = c->a[1] < -11776;
245 c->dms += (c->tbls.F[I]<<4) + ((- c->dms) >> 5);
246 c->dml += (c->tbls.F[I]<<4) + ((- c->dml) >> 7);
250 c->ap += (-c->ap) >> 4;
251 if (c->y <= 1535 || c->td || abs((c->dms << 2) - c->dml) >= (c->dml >> 3))
255 /* Update Yu and Yl */
256 c->yu = av_clip(c->y + c->tbls.W[I] + ((-c->y)>>5), 544, 5120);
257 c->yl += c->yu + ((-c->yl)>>6);
259 /* Next iteration for Y */
260 al = (c->ap >= 256) ? 1<<6 : c->ap >> 2;
261 c->y = (c->yl + (c->yu - (c->yl>>6))*al) >> 6;
263 /* Next iteration for SE and SEZ */
266 c->se += mult(i2f(c->b[i] >> 2, &f), &c->dq[i]);
269 c->se += mult(i2f(c->a[i] >> 2, &f), &c->sr[i]);
272 return av_clip(re_signal * 4, -0xffff, 0xffff);
275 static av_cold int g726_reset(G726Context *c)
279 c->tbls = G726Tables_pool[c->code_size - 2];
280 for (i=0; i<2; i++) {
281 c->sr[i].mant = 1<<5;
284 for (i=0; i<6; i++) {
285 c->dq[i].mant = 1<<5;
295 #if CONFIG_ADPCM_G726_ENCODER || CONFIG_ADPCM_G726LE_ENCODER
296 static int16_t g726_encode(G726Context* c, int16_t sig)
300 i = av_mod_uintp2(quant(c, sig/4 - c->se), c->code_size);
305 /* Interfacing to the libavcodec */
307 static av_cold int g726_encode_init(AVCodecContext *avctx)
309 G726Context* c = avctx->priv_data;
311 c->little_endian = !strcmp(avctx->codec->name, "g726le");
313 if (avctx->strict_std_compliance > FF_COMPLIANCE_UNOFFICIAL &&
314 avctx->sample_rate != 8000) {
315 av_log(avctx, AV_LOG_ERROR, "Sample rates other than 8kHz are not "
316 "allowed when the compliance level is higher than unofficial. "
317 "Resample or reduce the compliance level.\n");
318 return AVERROR(EINVAL);
320 if (avctx->sample_rate <= 0) {
321 av_log(avctx, AV_LOG_ERROR, "Invalid sample rate %d\n",
323 return AVERROR(EINVAL);
326 if(avctx->channels != 1){
327 av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
328 return AVERROR(EINVAL);
332 c->code_size = (avctx->bit_rate + avctx->sample_rate/2) / avctx->sample_rate;
334 c->code_size = av_clip(c->code_size, 2, 5);
335 avctx->bit_rate = c->code_size * avctx->sample_rate;
336 avctx->bits_per_coded_sample = c->code_size;
340 /* select a frame size that will end on a byte boundary and have a size of
341 approximately 1024 bytes */
342 avctx->frame_size = ((int[]){ 4096, 2736, 2048, 1640 })[c->code_size - 2];
347 static int g726_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
348 const AVFrame *frame, int *got_packet_ptr)
350 G726Context *c = avctx->priv_data;
351 const int16_t *samples = (const int16_t *)frame->data[0];
353 int i, ret, out_size;
355 out_size = (frame->nb_samples * c->code_size + 7) / 8;
356 if ((ret = ff_alloc_packet2(avctx, avpkt, out_size, 0)) < 0)
358 init_put_bits(&pb, avpkt->data, avpkt->size);
360 for (i = 0; i < frame->nb_samples; i++)
361 if (c->little_endian) {
362 put_bits_le(&pb, c->code_size, g726_encode(c, *samples++));
364 put_bits(&pb, c->code_size, g726_encode(c, *samples++));
367 if (c->little_endian) {
368 flush_put_bits_le(&pb);
373 avpkt->size = out_size;
378 #define OFFSET(x) offsetof(G726Context, x)
379 #define AE AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
380 static const AVOption options[] = {
381 { "code_size", "Bits per code", OFFSET(code_size), AV_OPT_TYPE_INT, { .i64 = 4 }, 2, 5, AE },
385 static const AVCodecDefault defaults[] = {
391 #if CONFIG_ADPCM_G726_ENCODER
392 static const AVClass g726_class = {
393 .class_name = "g726",
394 .item_name = av_default_item_name,
396 .version = LIBAVUTIL_VERSION_INT,
399 AVCodec ff_adpcm_g726_encoder = {
401 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),
402 .type = AVMEDIA_TYPE_AUDIO,
403 .id = AV_CODEC_ID_ADPCM_G726,
404 .priv_data_size = sizeof(G726Context),
405 .init = g726_encode_init,
406 .encode2 = g726_encode_frame,
407 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME,
408 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
409 AV_SAMPLE_FMT_NONE },
410 .priv_class = &g726_class,
411 .defaults = defaults,
415 #if CONFIG_ADPCM_G726LE_ENCODER
416 static const AVClass g726le_class = {
417 .class_name = "g726le",
418 .item_name = av_default_item_name,
420 .version = LIBAVUTIL_VERSION_INT,
423 AVCodec ff_adpcm_g726le_encoder = {
425 .long_name = NULL_IF_CONFIG_SMALL("G.726 little endian ADPCM (\"right-justified\")"),
426 .type = AVMEDIA_TYPE_AUDIO,
427 .id = AV_CODEC_ID_ADPCM_G726LE,
428 .priv_data_size = sizeof(G726Context),
429 .init = g726_encode_init,
430 .encode2 = g726_encode_frame,
431 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME,
432 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
433 AV_SAMPLE_FMT_NONE },
434 .priv_class = &g726le_class,
435 .defaults = defaults,
439 #if CONFIG_ADPCM_G726_DECODER || CONFIG_ADPCM_G726LE_DECODER
440 static av_cold int g726_decode_init(AVCodecContext *avctx)
442 G726Context* c = avctx->priv_data;
444 if(avctx->channels > 1){
445 avpriv_request_sample(avctx, "Decoding more than one channel");
446 return AVERROR_PATCHWELCOME;
449 avctx->channel_layout = AV_CH_LAYOUT_MONO;
451 c->little_endian = !strcmp(avctx->codec->name, "g726le");
453 c->code_size = avctx->bits_per_coded_sample;
454 if (c->code_size < 2 || c->code_size > 5) {
455 av_log(avctx, AV_LOG_ERROR, "Invalid number of bits %d\n", c->code_size);
456 return AVERROR(EINVAL);
460 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
465 static int g726_decode_frame(AVCodecContext *avctx, void *data,
466 int *got_frame_ptr, AVPacket *avpkt)
468 AVFrame *frame = data;
469 const uint8_t *buf = avpkt->data;
470 int buf_size = avpkt->size;
471 G726Context *c = avctx->priv_data;
474 int out_samples, ret;
476 out_samples = buf_size * 8 / c->code_size;
478 /* get output buffer */
479 frame->nb_samples = out_samples;
480 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
482 samples = (int16_t *)frame->data[0];
484 init_get_bits(&gb, buf, buf_size * 8);
486 while (out_samples--)
487 *samples++ = g726_decode(c, c->little_endian ?
488 get_bits_le(&gb, c->code_size) :
489 get_bits(&gb, c->code_size));
491 if (get_bits_left(&gb) > 0)
492 av_log(avctx, AV_LOG_ERROR, "Frame invalidly split, missing parser?\n");
499 static void g726_decode_flush(AVCodecContext *avctx)
501 G726Context *c = avctx->priv_data;
506 #if CONFIG_ADPCM_G726_DECODER
507 AVCodec ff_adpcm_g726_decoder = {
509 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),
510 .type = AVMEDIA_TYPE_AUDIO,
511 .id = AV_CODEC_ID_ADPCM_G726,
512 .priv_data_size = sizeof(G726Context),
513 .init = g726_decode_init,
514 .decode = g726_decode_frame,
515 .flush = g726_decode_flush,
516 .capabilities = AV_CODEC_CAP_DR1,
520 #if CONFIG_ADPCM_G726LE_DECODER
521 AVCodec ff_adpcm_g726le_decoder = {
523 .type = AVMEDIA_TYPE_AUDIO,
524 .id = AV_CODEC_ID_ADPCM_G726LE,
525 .priv_data_size = sizeof(G726Context),
526 .init = g726_decode_init,
527 .decode = g726_decode_frame,
528 .flush = g726_decode_flush,
529 .capabilities = AV_CODEC_CAP_DR1,
530 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM little-endian"),