2 * Copyright (c) CMU 1993 Computer Science, Speech Group
3 * Chengxiang Lu and Alex Hauptmann
4 * Copyright (c) 2005 Steve Underwood <steveu at coppice.org>
5 * Copyright (c) 2009 Kenan Gillet
6 * Copyright (c) 2010 Martin Storsjo
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
27 * G.722 ADPCM audio encoder
30 #include "libavutil/avassert.h"
34 #include "libavutil/common.h"
36 #define FREEZE_INTERVAL 128
38 /* This is an arbitrary value. Allowing insanely large values leads to strange
39 problems, so we limit it to a reasonable value */
40 #define MAX_FRAME_SIZE 32768
42 /* We clip the value of avctx->trellis to prevent data type overflows and
43 undefined behavior. Using larger values is insanely slow anyway. */
45 #define MAX_TRELLIS 16
47 static av_cold int g722_encode_close(AVCodecContext *avctx)
49 G722Context *c = avctx->priv_data;
51 for (i = 0; i < 2; i++) {
52 av_freep(&c->paths[i]);
53 av_freep(&c->node_buf[i]);
54 av_freep(&c->nodep_buf[i]);
59 static av_cold int g722_encode_init(AVCodecContext * avctx)
61 G722Context *c = avctx->priv_data;
63 c->band[0].scale_factor = 8;
64 c->band[1].scale_factor = 2;
65 c->prev_samples_pos = 22;
67 if (avctx->frame_size) {
68 /* validate frame size */
69 if (avctx->frame_size & 1 || avctx->frame_size > MAX_FRAME_SIZE) {
72 if (avctx->frame_size == 1)
74 else if (avctx->frame_size > MAX_FRAME_SIZE)
75 new_frame_size = MAX_FRAME_SIZE;
77 new_frame_size = avctx->frame_size - 1;
79 av_log(avctx, AV_LOG_WARNING, "Requested frame size is not "
80 "allowed. Using %d instead of %d\n", new_frame_size,
82 avctx->frame_size = new_frame_size;
85 /* This is arbitrary. We use 320 because it's 20ms @ 16kHz, which is
86 a common packet size for VoIP applications */
87 avctx->frame_size = 320;
89 avctx->initial_padding = 22;
92 /* validate trellis */
93 if (avctx->trellis < MIN_TRELLIS || avctx->trellis > MAX_TRELLIS) {
94 int new_trellis = av_clip(avctx->trellis, MIN_TRELLIS, MAX_TRELLIS);
95 av_log(avctx, AV_LOG_WARNING, "Requested trellis value is not "
96 "allowed. Using %d instead of %d\n", new_trellis,
98 avctx->trellis = new_trellis;
100 if (avctx->trellis) {
101 int frontier = 1 << avctx->trellis;
102 int max_paths = frontier * FREEZE_INTERVAL;
104 for (int i = 0; i < 2; i++) {
105 c->paths[i] = av_calloc(max_paths, sizeof(**c->paths));
106 c->node_buf[i] = av_calloc(frontier, 2 * sizeof(**c->node_buf));
107 c->nodep_buf[i] = av_calloc(frontier, 2 * sizeof(**c->nodep_buf));
108 if (!c->paths[i] || !c->node_buf[i] || !c->nodep_buf[i])
109 return AVERROR(ENOMEM);
114 ff_g722dsp_init(&c->dsp);
119 static const int16_t low_quant[33] = {
120 35, 72, 110, 150, 190, 233, 276, 323,
121 370, 422, 473, 530, 587, 650, 714, 786,
122 858, 940, 1023, 1121, 1219, 1339, 1458, 1612,
123 1765, 1980, 2195, 2557, 2919
126 static inline void filter_samples(G722Context *c, const int16_t *samples,
127 int *xlow, int *xhigh)
130 c->prev_samples[c->prev_samples_pos++] = samples[0];
131 c->prev_samples[c->prev_samples_pos++] = samples[1];
132 c->dsp.apply_qmf(c->prev_samples + c->prev_samples_pos - 24, xout);
133 *xlow = xout[0] + xout[1] >> 14;
134 *xhigh = xout[0] - xout[1] >> 14;
135 if (c->prev_samples_pos >= PREV_SAMPLES_BUF_SIZE) {
136 memmove(c->prev_samples,
137 c->prev_samples + c->prev_samples_pos - 22,
138 22 * sizeof(c->prev_samples[0]));
139 c->prev_samples_pos = 22;
143 static inline int encode_high(const struct G722Band *state, int xhigh)
145 int diff = av_clip_int16(xhigh - state->s_predictor);
146 int pred = 141 * state->scale_factor >> 8;
147 /* = diff >= 0 ? (diff < pred) + 2 : diff >= -pred */
148 return ((diff ^ (diff >> (sizeof(diff)*8-1))) < pred) + 2*(diff >= 0);
151 static inline int encode_low(const struct G722Band* state, int xlow)
153 int diff = av_clip_int16(xlow - state->s_predictor);
154 /* = diff >= 0 ? diff : -(diff + 1) */
155 int limit = diff ^ (diff >> (sizeof(diff)*8-1));
157 limit = limit + 1 << 10;
158 if (limit > low_quant[8] * state->scale_factor)
160 while (i < 29 && limit > low_quant[i] * state->scale_factor)
162 return (diff < 0 ? (i < 2 ? 63 : 33) : 61) - i;
165 static void g722_encode_trellis(G722Context *c, int trellis,
166 uint8_t *dst, int nb_samples,
167 const int16_t *samples)
170 int frontier = 1 << trellis;
171 struct TrellisNode **nodes[2];
172 struct TrellisNode **nodes_next[2];
173 int pathn[2] = {0, 0}, froze = -1;
174 struct TrellisPath *p[2];
176 for (i = 0; i < 2; i++) {
177 nodes[i] = c->nodep_buf[i];
178 nodes_next[i] = c->nodep_buf[i] + frontier;
179 memset(c->nodep_buf[i], 0, 2 * frontier * sizeof(*c->nodep_buf[i]));
180 nodes[i][0] = c->node_buf[i] + frontier;
181 nodes[i][0]->ssd = 0;
182 nodes[i][0]->path = 0;
183 nodes[i][0]->state = c->band[i];
186 for (i = 0; i < nb_samples >> 1; i++) {
188 struct TrellisNode *next[2];
189 int heap_pos[2] = {0, 0};
191 for (j = 0; j < 2; j++) {
192 next[j] = c->node_buf[j] + frontier*(i & 1);
193 memset(nodes_next[j], 0, frontier * sizeof(**nodes_next));
196 filter_samples(c, &samples[2*i], &xlow, &xhigh);
198 for (j = 0; j < frontier && nodes[0][j]; j++) {
199 /* Only k >> 2 affects the future adaptive state, therefore testing
200 * small steps that don't change k >> 2 is useless, the original
201 * value from encode_low is better than them. Since we step k
202 * in steps of 4, make sure range is a multiple of 4, so that
203 * we don't miss the original value from encode_low. */
204 int range = j < frontier/2 ? 4 : 0;
205 struct TrellisNode *cur_node = nodes[0][j];
207 int ilow = encode_low(&cur_node->state, xlow);
209 for (k = ilow - range; k <= ilow + range && k <= 63; k += 4) {
210 int decoded, dec_diff, pos;
212 struct TrellisNode* node;
217 decoded = av_clip_intp2((cur_node->state.scale_factor *
218 ff_g722_low_inv_quant6[k] >> 10)
219 + cur_node->state.s_predictor, 14);
220 dec_diff = xlow - decoded;
222 #define STORE_NODE(index, UPDATE, VALUE)\
223 ssd = cur_node->ssd + dec_diff*dec_diff;\
224 /* Check for wraparound. Using 64 bit ssd counters would \
225 * be simpler, but is slower on x86 32 bit. */\
226 if (ssd < cur_node->ssd)\
228 if (heap_pos[index] < frontier) {\
229 pos = heap_pos[index]++;\
230 av_assert2(pathn[index] < FREEZE_INTERVAL * frontier);\
231 node = nodes_next[index][pos] = next[index]++;\
232 node->path = pathn[index]++;\
234 /* Try to replace one of the leaf nodes with the new \
235 * one, but not always testing the same leaf position */\
236 pos = (frontier>>1) + (heap_pos[index] & ((frontier>>1) - 1));\
237 if (ssd >= nodes_next[index][pos]->ssd)\
240 node = nodes_next[index][pos];\
243 node->state = cur_node->state;\
245 c->paths[index][node->path].value = VALUE;\
246 c->paths[index][node->path].prev = cur_node->path;\
247 /* Sift the newly inserted node up in the heap to restore \
248 * the heap property */\
250 int parent = (pos - 1) >> 1;\
251 if (nodes_next[index][parent]->ssd <= ssd)\
253 FFSWAP(struct TrellisNode*, nodes_next[index][parent],\
254 nodes_next[index][pos]);\
257 STORE_NODE(0, ff_g722_update_low_predictor(&node->state, k >> 2), k);
261 for (j = 0; j < frontier && nodes[1][j]; j++) {
263 struct TrellisNode *cur_node = nodes[1][j];
265 /* We don't try to get any initial guess for ihigh via
266 * encode_high - since there's only 4 possible values, test
267 * them all. Testing all of these gives a much, much larger
268 * gain than testing a larger range around ilow. */
269 for (ihigh = 0; ihigh < 4; ihigh++) {
270 int dhigh, decoded, dec_diff, pos;
272 struct TrellisNode* node;
274 dhigh = cur_node->state.scale_factor *
275 ff_g722_high_inv_quant[ihigh] >> 10;
276 decoded = av_clip_intp2(dhigh + cur_node->state.s_predictor, 14);
277 dec_diff = xhigh - decoded;
279 STORE_NODE(1, ff_g722_update_high_predictor(&node->state, dhigh, ihigh), ihigh);
283 for (j = 0; j < 2; j++) {
284 FFSWAP(struct TrellisNode**, nodes[j], nodes_next[j]);
286 if (nodes[j][0]->ssd > (1 << 16)) {
287 for (k = 1; k < frontier && nodes[j][k]; k++)
288 nodes[j][k]->ssd -= nodes[j][0]->ssd;
289 nodes[j][0]->ssd = 0;
293 if (i == froze + FREEZE_INTERVAL) {
294 p[0] = &c->paths[0][nodes[0][0]->path];
295 p[1] = &c->paths[1][nodes[1][0]->path];
296 for (j = i; j > froze; j--) {
297 dst[j] = p[1]->value << 6 | p[0]->value;
298 p[0] = &c->paths[0][p[0]->prev];
299 p[1] = &c->paths[1][p[1]->prev];
302 pathn[0] = pathn[1] = 0;
303 memset(nodes[0] + 1, 0, (frontier - 1)*sizeof(**nodes));
304 memset(nodes[1] + 1, 0, (frontier - 1)*sizeof(**nodes));
308 p[0] = &c->paths[0][nodes[0][0]->path];
309 p[1] = &c->paths[1][nodes[1][0]->path];
310 for (j = i; j > froze; j--) {
311 dst[j] = p[1]->value << 6 | p[0]->value;
312 p[0] = &c->paths[0][p[0]->prev];
313 p[1] = &c->paths[1][p[1]->prev];
315 c->band[0] = nodes[0][0]->state;
316 c->band[1] = nodes[1][0]->state;
319 static av_always_inline void encode_byte(G722Context *c, uint8_t *dst,
320 const int16_t *samples)
322 int xlow, xhigh, ilow, ihigh;
323 filter_samples(c, samples, &xlow, &xhigh);
324 ihigh = encode_high(&c->band[1], xhigh);
325 ilow = encode_low (&c->band[0], xlow);
326 ff_g722_update_high_predictor(&c->band[1], c->band[1].scale_factor *
327 ff_g722_high_inv_quant[ihigh] >> 10, ihigh);
328 ff_g722_update_low_predictor(&c->band[0], ilow >> 2);
329 *dst = ihigh << 6 | ilow;
332 static void g722_encode_no_trellis(G722Context *c,
333 uint8_t *dst, int nb_samples,
334 const int16_t *samples)
337 for (i = 0; i < nb_samples; i += 2)
338 encode_byte(c, dst++, &samples[i]);
341 static int g722_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
342 const AVFrame *frame, int *got_packet_ptr)
344 G722Context *c = avctx->priv_data;
345 const int16_t *samples = (const int16_t *)frame->data[0];
346 int nb_samples, out_size, ret;
348 out_size = (frame->nb_samples + 1) / 2;
349 if ((ret = ff_alloc_packet2(avctx, avpkt, out_size, 0)) < 0)
352 nb_samples = frame->nb_samples - (frame->nb_samples & 1);
355 g722_encode_trellis(c, avctx->trellis, avpkt->data, nb_samples, samples);
357 g722_encode_no_trellis(c, avpkt->data, nb_samples, samples);
359 /* handle last frame with odd frame_size */
360 if (nb_samples < frame->nb_samples) {
361 int16_t last_samples[2] = { samples[nb_samples], samples[nb_samples] };
362 encode_byte(c, &avpkt->data[nb_samples >> 1], last_samples);
365 if (frame->pts != AV_NOPTS_VALUE)
366 avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->initial_padding);
371 AVCodec ff_adpcm_g722_encoder = {
373 .long_name = NULL_IF_CONFIG_SMALL("G.722 ADPCM"),
374 .type = AVMEDIA_TYPE_AUDIO,
375 .id = AV_CODEC_ID_ADPCM_G722,
376 .priv_data_size = sizeof(G722Context),
377 .init = g722_encode_init,
378 .close = g722_encode_close,
379 .encode2 = g722_encode_frame,
380 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME,
381 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_NONE },
382 .channel_layouts = (const uint64_t[]){ AV_CH_LAYOUT_MONO, 0 },
383 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,