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/avassert.h"
27 #include "libavutil/channel_layout.h"
28 #include "libavutil/opt.h"
36 * G.726 Standard uses rather odd 11bit floating point arithmentic for
37 * numerous occasions. It's a mystery to me why they did it this way
38 * instead of simply using 32bit integer arithmetic.
40 typedef struct Float11 {
41 uint8_t sign; /**< 1bit sign */
42 uint8_t exp; /**< 4bit exponent */
43 uint8_t mant; /**< 6bit mantissa */
46 static inline Float11* i2f(int i, Float11* f)
51 f->exp = av_log2_16bit(i) + !!i;
52 f->mant = i? (i<<6) >> f->exp : 1<<5;
56 static inline int16_t mult(Float11* f1, Float11* f2)
60 exp = f1->exp + f2->exp;
61 res = (((f1->mant * f2->mant) + 0x30) >> 4);
62 res = exp > 19 ? res << (exp - 19) : res >> (19 - exp);
63 return (f1->sign ^ f2->sign) ? -res : res;
66 static inline int sgn(int value)
68 return (value < 0) ? -1 : 1;
71 typedef struct G726Tables {
72 const int* quant; /**< quantization table */
73 const int16_t* iquant; /**< inverse quantization table */
74 const int16_t* W; /**< special table #1 ;-) */
75 const uint8_t* F; /**< special table #2 */
78 typedef struct G726Context {
81 G726Tables tbls; /**< static tables needed for computation */
83 Float11 sr[2]; /**< prev. reconstructed samples */
84 Float11 dq[6]; /**< prev. difference */
85 int a[2]; /**< second order predictor coeffs */
86 int b[6]; /**< sixth order predictor coeffs */
87 int pk[2]; /**< signs of prev. 2 sez + dq */
89 int ap; /**< scale factor control */
90 int yu; /**< fast scale factor */
91 int yl; /**< slow scale factor */
92 int dms; /**< short average magnitude of F[i] */
93 int dml; /**< long average magnitude of F[i] */
94 int td; /**< tone detect */
96 int se; /**< estimated signal for the next iteration */
97 int sez; /**< estimated second order prediction */
98 int y; /**< quantizer scaling factor for the next iteration */
102 static const int quant_tbl16[] = /**< 16kbit/s 2bits per sample */
104 static const int16_t iquant_tbl16[] =
105 { 116, 365, 365, 116 };
106 static const int16_t W_tbl16[] =
107 { -22, 439, 439, -22 };
108 static const uint8_t F_tbl16[] =
111 static const int quant_tbl24[] = /**< 24kbit/s 3bits per sample */
112 { 7, 217, 330, INT_MAX };
113 static const int16_t iquant_tbl24[] =
114 { INT16_MIN, 135, 273, 373, 373, 273, 135, INT16_MIN };
115 static const int16_t W_tbl24[] =
116 { -4, 30, 137, 582, 582, 137, 30, -4 };
117 static const uint8_t F_tbl24[] =
118 { 0, 1, 2, 7, 7, 2, 1, 0 };
120 static const int quant_tbl32[] = /**< 32kbit/s 4bits per sample */
121 { -125, 79, 177, 245, 299, 348, 399, INT_MAX };
122 static const int16_t iquant_tbl32[] =
123 { INT16_MIN, 4, 135, 213, 273, 323, 373, 425,
124 425, 373, 323, 273, 213, 135, 4, INT16_MIN };
125 static const int16_t W_tbl32[] =
126 { -12, 18, 41, 64, 112, 198, 355, 1122,
127 1122, 355, 198, 112, 64, 41, 18, -12};
128 static const uint8_t F_tbl32[] =
129 { 0, 0, 0, 1, 1, 1, 3, 7, 7, 3, 1, 1, 1, 0, 0, 0 };
131 static const 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 };
134 static const int16_t iquant_tbl40[] =
135 { INT16_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, INT16_MIN };
139 static const int16_t 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 };
144 static const uint8_t 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 };
148 static const G726Tables G726Tables_pool[] =
149 {{ quant_tbl16, iquant_tbl16, W_tbl16, F_tbl16 },
150 { quant_tbl24, iquant_tbl24, W_tbl24, F_tbl24 },
151 { quant_tbl32, iquant_tbl32, W_tbl32, F_tbl32 },
152 { quant_tbl40, iquant_tbl40, W_tbl40, F_tbl40 }};
156 * Para 4.2.2 page 18: Adaptive quantizer.
158 static inline uint8_t quant(G726Context* c, int d)
160 int sign, exp, i, dln;
167 exp = av_log2_16bit(d);
168 dln = ((exp<<7) + (((d<<7)>>exp)&0x7f)) - (c->y>>2);
170 while (c->tbls.quant[i] < INT_MAX && c->tbls.quant[i] < dln)
175 if (c->code_size != 2 && i == 0) /* I'm not sure this is a good idea */
182 * Para 4.2.3 page 22: Inverse adaptive quantizer.
184 static inline int16_t inverse_quant(G726Context* c, int i)
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<<dex) >> 7);
194 static int16_t g726_decode(G726Context* c, int I)
196 int dq, re_signal, pk0, fa1, i, tr, ylint, ylfrac, thr2, al, dq0;
198 int I_sig= I >> (c->code_size - 1);
200 dq = inverse_quant(c, I);
202 /* Transition detect */
203 ylint = (c->yl >> 15);
204 ylfrac = (c->yl >> 10) & 0x1f;
205 thr2 = (ylint > 9) ? 0x1f << 10 : (0x20 + ylfrac) << ylint;
206 tr= (c->td == 1 && dq > ((3*thr2)>>2));
208 if (I_sig) /* get the sign */
210 re_signal = c->se + dq;
212 /* Update second order predictor coefficient A2 and A1 */
213 pk0 = (c->sez + dq) ? sgn(c->sez + dq) : 0;
214 dq0 = dq ? sgn(dq) : 0;
221 /* This is a bit crazy, but it really is +255 not +256 */
222 fa1 = av_clip((-c->a[0]*c->pk[0]*pk0)>>5, -256, 255);
224 c->a[1] += 128*pk0*c->pk[1] + fa1 - (c->a[1]>>7);
225 c->a[1] = av_clip(c->a[1], -12288, 12288);
226 c->a[0] += 64*3*pk0*c->pk[0] - (c->a[0] >> 8);
227 c->a[0] = av_clip(c->a[0], -(15360 - c->a[1]), 15360 - c->a[1]);
230 c->b[i] += 128*dq0*sgn(-c->dq[i].sign) - (c->b[i]>>8);
233 /* Update Dq and Sr and Pk */
235 c->pk[0] = pk0 ? pk0 : 1;
237 i2f(re_signal, &c->sr[0]);
239 c->dq[i] = c->dq[i-1];
241 c->dq[0].sign = I_sig; /* Isn't it crazy ?!?! */
243 c->td = c->a[1] < -11776;
246 c->dms += (c->tbls.F[I]<<4) + ((- c->dms) >> 5);
247 c->dml += (c->tbls.F[I]<<4) + ((- c->dml) >> 7);
251 c->ap += (-c->ap) >> 4;
252 if (c->y <= 1535 || c->td || abs((c->dms << 2) - c->dml) >= (c->dml >> 3))
256 /* Update Yu and Yl */
257 c->yu = av_clip(c->y + c->tbls.W[I] + ((-c->y)>>5), 544, 5120);
258 c->yl += c->yu + ((-c->yl)>>6);
260 /* Next iteration for Y */
261 al = (c->ap >= 256) ? 1<<6 : c->ap >> 2;
262 c->y = (c->yl + (c->yu - (c->yl>>6))*al) >> 6;
264 /* Next iteration for SE and SEZ */
267 c->se += mult(i2f(c->b[i] >> 2, &f), &c->dq[i]);
270 c->se += mult(i2f(c->a[i] >> 2, &f), &c->sr[i]);
273 return av_clip(re_signal << 2, -0xffff, 0xffff);
276 static av_cold int g726_reset(G726Context *c)
280 c->tbls = G726Tables_pool[c->code_size - 2];
281 for (i=0; i<2; i++) {
282 c->sr[i].mant = 1<<5;
285 for (i=0; i<6; i++) {
286 c->dq[i].mant = 1<<5;
296 #if CONFIG_ADPCM_G726_ENCODER
297 static int16_t g726_encode(G726Context* c, int16_t sig)
301 i = quant(c, sig/4 - c->se) & ((1<<c->code_size) - 1);
306 /* Interfacing to the libavcodec */
308 static av_cold int g726_encode_init(AVCodecContext *avctx)
310 G726Context* c = avctx->priv_data;
312 if (avctx->strict_std_compliance > FF_COMPLIANCE_UNOFFICIAL &&
313 avctx->sample_rate != 8000) {
314 av_log(avctx, AV_LOG_ERROR, "Sample rates other than 8kHz are not "
315 "allowed when the compliance level is higher than unofficial. "
316 "Resample or reduce the compliance level.\n");
317 return AVERROR(EINVAL);
319 av_assert0(avctx->sample_rate > 0);
321 if(avctx->channels != 1){
322 av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
323 return AVERROR(EINVAL);
327 c->code_size = (avctx->bit_rate + avctx->sample_rate/2) / avctx->sample_rate;
329 c->code_size = av_clip(c->code_size, 2, 5);
330 avctx->bit_rate = c->code_size * avctx->sample_rate;
331 avctx->bits_per_coded_sample = c->code_size;
335 #if FF_API_OLD_ENCODE_AUDIO
336 avctx->coded_frame = avcodec_alloc_frame();
337 if (!avctx->coded_frame)
338 return AVERROR(ENOMEM);
339 avctx->coded_frame->key_frame = 1;
342 /* select a frame size that will end on a byte boundary and have a size of
343 approximately 1024 bytes */
344 avctx->frame_size = ((int[]){ 4096, 2736, 2048, 1640 })[c->code_size - 2];
349 #if FF_API_OLD_ENCODE_AUDIO
350 static av_cold int g726_encode_close(AVCodecContext *avctx)
352 av_freep(&avctx->coded_frame);
357 static int g726_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
358 const AVFrame *frame, int *got_packet_ptr)
360 G726Context *c = avctx->priv_data;
361 const int16_t *samples = (const int16_t *)frame->data[0];
363 int i, ret, out_size;
365 out_size = (frame->nb_samples * c->code_size + 7) / 8;
366 if ((ret = ff_alloc_packet2(avctx, avpkt, out_size)))
368 init_put_bits(&pb, avpkt->data, avpkt->size);
370 for (i = 0; i < frame->nb_samples; i++)
371 put_bits(&pb, c->code_size, g726_encode(c, *samples++));
375 avpkt->size = out_size;
380 #define OFFSET(x) offsetof(G726Context, x)
381 #define AE AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
382 static const AVOption options[] = {
383 { "code_size", "Bits per code", OFFSET(code_size), AV_OPT_TYPE_INT, { .i64 = 4 }, 2, 5, AE },
387 static const AVClass class = {
388 .class_name = "g726",
389 .item_name = av_default_item_name,
391 .version = LIBAVUTIL_VERSION_INT,
394 static const AVCodecDefault defaults[] = {
399 AVCodec ff_adpcm_g726_encoder = {
401 .type = AVMEDIA_TYPE_AUDIO,
402 .id = AV_CODEC_ID_ADPCM_G726,
403 .priv_data_size = sizeof(G726Context),
404 .init = g726_encode_init,
405 .encode2 = g726_encode_frame,
406 #if FF_API_OLD_ENCODE_AUDIO
407 .close = g726_encode_close,
409 .capabilities = CODEC_CAP_SMALL_LAST_FRAME,
410 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
411 AV_SAMPLE_FMT_NONE },
412 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),
413 .priv_class = &class,
414 .defaults = defaults,
418 #if CONFIG_ADPCM_G726_DECODER
419 static av_cold int g726_decode_init(AVCodecContext *avctx)
421 G726Context* c = avctx->priv_data;
424 avctx->channel_layout = AV_CH_LAYOUT_MONO;
426 c->code_size = avctx->bits_per_coded_sample;
427 if (c->code_size < 2 || c->code_size > 5) {
428 av_log(avctx, AV_LOG_ERROR, "Invalid number of bits %d\n", c->code_size);
429 return AVERROR(EINVAL);
433 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
435 avcodec_get_frame_defaults(&c->frame);
436 avctx->coded_frame = &c->frame;
441 static int g726_decode_frame(AVCodecContext *avctx, void *data,
442 int *got_frame_ptr, AVPacket *avpkt)
444 const uint8_t *buf = avpkt->data;
445 int buf_size = avpkt->size;
446 G726Context *c = avctx->priv_data;
449 int out_samples, ret;
451 out_samples = buf_size * 8 / c->code_size;
453 /* get output buffer */
454 c->frame.nb_samples = out_samples;
455 if ((ret = avctx->get_buffer(avctx, &c->frame)) < 0) {
456 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
459 samples = (int16_t *)c->frame.data[0];
461 init_get_bits(&gb, buf, buf_size * 8);
463 while (out_samples--)
464 *samples++ = g726_decode(c, get_bits(&gb, c->code_size));
466 if (get_bits_left(&gb) > 0)
467 av_log(avctx, AV_LOG_ERROR, "Frame invalidly split, missing parser?\n");
470 *(AVFrame *)data = c->frame;
475 static void g726_decode_flush(AVCodecContext *avctx)
477 G726Context *c = avctx->priv_data;
481 AVCodec ff_adpcm_g726_decoder = {
483 .type = AVMEDIA_TYPE_AUDIO,
484 .id = AV_CODEC_ID_ADPCM_G726,
485 .priv_data_size = sizeof(G726Context),
486 .init = g726_decode_init,
487 .decode = g726_decode_frame,
488 .flush = g726_decode_flush,
489 .capabilities = CODEC_CAP_DR1,
490 .long_name = NULL_IF_CONFIG_SMALL("G.726 ADPCM"),