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 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
27 * G.722 ADPCM audio encoder
33 #define FREEZE_INTERVAL 128
35 /* This is an arbitrary value. Allowing insanely large values leads to strange
36 problems, so we limit it to a reasonable value */
37 #define MAX_FRAME_SIZE 32768
39 /* We clip the value of avctx->trellis to prevent data type overflows and
40 undefined behavior. Using larger values is insanely slow anyway. */
42 #define MAX_TRELLIS 16
44 static av_cold int g722_encode_close(AVCodecContext *avctx)
46 G722Context *c = avctx->priv_data;
48 for (i = 0; i < 2; i++) {
49 av_freep(&c->paths[i]);
50 av_freep(&c->node_buf[i]);
51 av_freep(&c->nodep_buf[i]);
56 static av_cold int g722_encode_init(AVCodecContext * avctx)
58 G722Context *c = avctx->priv_data;
61 if (avctx->channels != 1) {
62 av_log(avctx, AV_LOG_ERROR, "Only mono tracks are allowed.\n");
63 return AVERROR_INVALIDDATA;
66 c->band[0].scale_factor = 8;
67 c->band[1].scale_factor = 2;
68 c->prev_samples_pos = 22;
71 int frontier = 1 << avctx->trellis;
72 int max_paths = frontier * FREEZE_INTERVAL;
74 for (i = 0; i < 2; i++) {
75 c->paths[i] = av_mallocz(max_paths * sizeof(**c->paths));
76 c->node_buf[i] = av_mallocz(2 * frontier * sizeof(**c->node_buf));
77 c->nodep_buf[i] = av_mallocz(2 * frontier * sizeof(**c->nodep_buf));
78 if (!c->paths[i] || !c->node_buf[i] || !c->nodep_buf[i]) {
79 ret = AVERROR(ENOMEM);
85 if (avctx->frame_size) {
86 /* validate frame size */
87 if (avctx->frame_size & 1 || avctx->frame_size > MAX_FRAME_SIZE) {
90 if (avctx->frame_size == 1)
92 else if (avctx->frame_size > MAX_FRAME_SIZE)
93 new_frame_size = MAX_FRAME_SIZE;
95 new_frame_size = avctx->frame_size - 1;
97 av_log(avctx, AV_LOG_WARNING, "Requested frame size is not "
98 "allowed. Using %d instead of %d\n", new_frame_size,
100 avctx->frame_size = new_frame_size;
103 /* This is arbitrary. We use 320 because it's 20ms @ 16kHz, which is
104 a common packet size for VoIP applications */
105 avctx->frame_size = 320;
108 if (avctx->trellis) {
109 /* validate trellis */
110 if (avctx->trellis < MIN_TRELLIS || avctx->trellis > MAX_TRELLIS) {
111 int new_trellis = av_clip(avctx->trellis, MIN_TRELLIS, MAX_TRELLIS);
112 av_log(avctx, AV_LOG_WARNING, "Requested trellis value is not "
113 "allowed. Using %d instead of %d\n", new_trellis,
115 avctx->trellis = new_trellis;
121 g722_encode_close(avctx);
125 static const int16_t low_quant[33] = {
126 35, 72, 110, 150, 190, 233, 276, 323,
127 370, 422, 473, 530, 587, 650, 714, 786,
128 858, 940, 1023, 1121, 1219, 1339, 1458, 1612,
129 1765, 1980, 2195, 2557, 2919
132 static inline void filter_samples(G722Context *c, const int16_t *samples,
133 int *xlow, int *xhigh)
136 c->prev_samples[c->prev_samples_pos++] = samples[0];
137 c->prev_samples[c->prev_samples_pos++] = samples[1];
138 ff_g722_apply_qmf(c->prev_samples + c->prev_samples_pos - 24, &xout1, &xout2);
139 *xlow = xout1 + xout2 >> 13;
140 *xhigh = xout1 - xout2 >> 13;
141 if (c->prev_samples_pos >= PREV_SAMPLES_BUF_SIZE) {
142 memmove(c->prev_samples,
143 c->prev_samples + c->prev_samples_pos - 22,
144 22 * sizeof(c->prev_samples[0]));
145 c->prev_samples_pos = 22;
149 static inline int encode_high(const struct G722Band *state, int xhigh)
151 int diff = av_clip_int16(xhigh - state->s_predictor);
152 int pred = 141 * state->scale_factor >> 8;
153 /* = diff >= 0 ? (diff < pred) + 2 : diff >= -pred */
154 return ((diff ^ (diff >> (sizeof(diff)*8-1))) < pred) + 2*(diff >= 0);
157 static inline int encode_low(const struct G722Band* state, int xlow)
159 int diff = av_clip_int16(xlow - state->s_predictor);
160 /* = diff >= 0 ? diff : -(diff + 1) */
161 int limit = diff ^ (diff >> (sizeof(diff)*8-1));
163 limit = limit + 1 << 10;
164 if (limit > low_quant[8] * state->scale_factor)
166 while (i < 29 && limit > low_quant[i] * state->scale_factor)
168 return (diff < 0 ? (i < 2 ? 63 : 33) : 61) - i;
171 static void g722_encode_trellis(G722Context *c, int trellis,
172 uint8_t *dst, int nb_samples,
173 const int16_t *samples)
176 int frontier = 1 << trellis;
177 struct TrellisNode **nodes[2];
178 struct TrellisNode **nodes_next[2];
179 int pathn[2] = {0, 0}, froze = -1;
180 struct TrellisPath *p[2];
182 for (i = 0; i < 2; i++) {
183 nodes[i] = c->nodep_buf[i];
184 nodes_next[i] = c->nodep_buf[i] + frontier;
185 memset(c->nodep_buf[i], 0, 2 * frontier * sizeof(*c->nodep_buf));
186 nodes[i][0] = c->node_buf[i] + frontier;
187 nodes[i][0]->ssd = 0;
188 nodes[i][0]->path = 0;
189 nodes[i][0]->state = c->band[i];
192 for (i = 0; i < nb_samples >> 1; i++) {
194 struct TrellisNode *next[2];
195 int heap_pos[2] = {0, 0};
197 for (j = 0; j < 2; j++) {
198 next[j] = c->node_buf[j] + frontier*(i & 1);
199 memset(nodes_next[j], 0, frontier * sizeof(**nodes_next));
202 filter_samples(c, &samples[2*i], &xlow, &xhigh);
204 for (j = 0; j < frontier && nodes[0][j]; j++) {
205 /* Only k >> 2 affects the future adaptive state, therefore testing
206 * small steps that don't change k >> 2 is useless, the original
207 * value from encode_low is better than them. Since we step k
208 * in steps of 4, make sure range is a multiple of 4, so that
209 * we don't miss the original value from encode_low. */
210 int range = j < frontier/2 ? 4 : 0;
211 struct TrellisNode *cur_node = nodes[0][j];
213 int ilow = encode_low(&cur_node->state, xlow);
215 for (k = ilow - range; k <= ilow + range && k <= 63; k += 4) {
216 int decoded, dec_diff, pos;
218 struct TrellisNode* node;
223 decoded = av_clip((cur_node->state.scale_factor *
224 ff_g722_low_inv_quant6[k] >> 10)
225 + cur_node->state.s_predictor, -16384, 16383);
226 dec_diff = xlow - decoded;
228 #define STORE_NODE(index, UPDATE, VALUE)\
229 ssd = cur_node->ssd + dec_diff*dec_diff;\
230 /* Check for wraparound. Using 64 bit ssd counters would \
231 * be simpler, but is slower on x86 32 bit. */\
232 if (ssd < cur_node->ssd)\
234 if (heap_pos[index] < frontier) {\
235 pos = heap_pos[index]++;\
236 assert(pathn[index] < FREEZE_INTERVAL * frontier);\
237 node = nodes_next[index][pos] = next[index]++;\
238 node->path = pathn[index]++;\
240 /* Try to replace one of the leaf nodes with the new \
241 * one, but not always testing the same leaf position */\
242 pos = (frontier>>1) + (heap_pos[index] & ((frontier>>1) - 1));\
243 if (ssd >= nodes_next[index][pos]->ssd)\
246 node = nodes_next[index][pos];\
249 node->state = cur_node->state;\
251 c->paths[index][node->path].value = VALUE;\
252 c->paths[index][node->path].prev = cur_node->path;\
253 /* Sift the newly inserted node up in the heap to restore \
254 * the heap property */\
256 int parent = (pos - 1) >> 1;\
257 if (nodes_next[index][parent]->ssd <= ssd)\
259 FFSWAP(struct TrellisNode*, nodes_next[index][parent],\
260 nodes_next[index][pos]);\
263 STORE_NODE(0, ff_g722_update_low_predictor(&node->state, k >> 2), k);
267 for (j = 0; j < frontier && nodes[1][j]; j++) {
269 struct TrellisNode *cur_node = nodes[1][j];
271 /* We don't try to get any initial guess for ihigh via
272 * encode_high - since there's only 4 possible values, test
273 * them all. Testing all of these gives a much, much larger
274 * gain than testing a larger range around ilow. */
275 for (ihigh = 0; ihigh < 4; ihigh++) {
276 int dhigh, decoded, dec_diff, pos;
278 struct TrellisNode* node;
280 dhigh = cur_node->state.scale_factor *
281 ff_g722_high_inv_quant[ihigh] >> 10;
282 decoded = av_clip(dhigh + cur_node->state.s_predictor,
284 dec_diff = xhigh - decoded;
286 STORE_NODE(1, ff_g722_update_high_predictor(&node->state, dhigh, ihigh), ihigh);
290 for (j = 0; j < 2; j++) {
291 FFSWAP(struct TrellisNode**, nodes[j], nodes_next[j]);
293 if (nodes[j][0]->ssd > (1 << 16)) {
294 for (k = 1; k < frontier && nodes[j][k]; k++)
295 nodes[j][k]->ssd -= nodes[j][0]->ssd;
296 nodes[j][0]->ssd = 0;
300 if (i == froze + FREEZE_INTERVAL) {
301 p[0] = &c->paths[0][nodes[0][0]->path];
302 p[1] = &c->paths[1][nodes[1][0]->path];
303 for (j = i; j > froze; j--) {
304 dst[j] = p[1]->value << 6 | p[0]->value;
305 p[0] = &c->paths[0][p[0]->prev];
306 p[1] = &c->paths[1][p[1]->prev];
309 pathn[0] = pathn[1] = 0;
310 memset(nodes[0] + 1, 0, (frontier - 1)*sizeof(**nodes));
311 memset(nodes[1] + 1, 0, (frontier - 1)*sizeof(**nodes));
315 p[0] = &c->paths[0][nodes[0][0]->path];
316 p[1] = &c->paths[1][nodes[1][0]->path];
317 for (j = i; j > froze; j--) {
318 dst[j] = p[1]->value << 6 | p[0]->value;
319 p[0] = &c->paths[0][p[0]->prev];
320 p[1] = &c->paths[1][p[1]->prev];
322 c->band[0] = nodes[0][0]->state;
323 c->band[1] = nodes[1][0]->state;
326 static av_always_inline void encode_byte(G722Context *c, uint8_t *dst,
327 const int16_t *samples)
329 int xlow, xhigh, ilow, ihigh;
330 filter_samples(c, samples, &xlow, &xhigh);
331 ihigh = encode_high(&c->band[1], xhigh);
332 ilow = encode_low (&c->band[0], xlow);
333 ff_g722_update_high_predictor(&c->band[1], c->band[1].scale_factor *
334 ff_g722_high_inv_quant[ihigh] >> 10, ihigh);
335 ff_g722_update_low_predictor(&c->band[0], ilow >> 2);
336 *dst = ihigh << 6 | ilow;
339 static void g722_encode_no_trellis(G722Context *c,
340 uint8_t *dst, int nb_samples,
341 const int16_t *samples)
344 for (i = 0; i < nb_samples; i += 2)
345 encode_byte(c, dst++, &samples[i]);
348 static int g722_encode_frame(AVCodecContext *avctx,
349 uint8_t *dst, int buf_size, void *data)
351 G722Context *c = avctx->priv_data;
352 const int16_t *samples = data;
355 nb_samples = avctx->frame_size - (avctx->frame_size & 1);
358 g722_encode_trellis(c, avctx->trellis, dst, nb_samples, samples);
360 g722_encode_no_trellis(c, dst, nb_samples, samples);
362 /* handle last frame with odd frame_size */
363 if (nb_samples < avctx->frame_size) {
364 int16_t last_samples[2] = { samples[nb_samples], samples[nb_samples] };
365 encode_byte(c, &dst[nb_samples >> 1], last_samples);
368 return (avctx->frame_size + 1) >> 1;
371 AVCodec ff_adpcm_g722_encoder = {
373 .type = AVMEDIA_TYPE_AUDIO,
374 .id = CODEC_ID_ADPCM_G722,
375 .priv_data_size = sizeof(G722Context),
376 .init = g722_encode_init,
377 .close = g722_encode_close,
378 .encode = g722_encode_frame,
379 .capabilities = CODEC_CAP_SMALL_LAST_FRAME,
380 .long_name = NULL_IF_CONFIG_SMALL("G.722 ADPCM"),
381 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},