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 static av_cold int g722_encode_init(AVCodecContext * avctx)
37 G722Context *c = avctx->priv_data;
39 if (avctx->channels != 1) {
40 av_log(avctx, AV_LOG_ERROR, "Only mono tracks are allowed.\n");
41 return AVERROR_INVALIDDATA;
44 c->band[0].scale_factor = 8;
45 c->band[1].scale_factor = 2;
46 c->prev_samples_pos = 22;
49 int frontier = 1 << avctx->trellis;
50 int max_paths = frontier * FREEZE_INTERVAL;
52 for (i = 0; i < 2; i++) {
53 c->paths[i] = av_mallocz(max_paths * sizeof(**c->paths));
54 c->node_buf[i] = av_mallocz(2 * frontier * sizeof(**c->node_buf));
55 c->nodep_buf[i] = av_mallocz(2 * frontier * sizeof(**c->nodep_buf));
62 static av_cold int g722_encode_close(AVCodecContext *avctx)
64 G722Context *c = avctx->priv_data;
66 for (i = 0; i < 2; i++) {
67 av_freep(&c->paths[i]);
68 av_freep(&c->node_buf[i]);
69 av_freep(&c->nodep_buf[i]);
74 static const int16_t low_quant[33] = {
75 35, 72, 110, 150, 190, 233, 276, 323,
76 370, 422, 473, 530, 587, 650, 714, 786,
77 858, 940, 1023, 1121, 1219, 1339, 1458, 1612,
78 1765, 1980, 2195, 2557, 2919
81 static inline void filter_samples(G722Context *c, const int16_t *samples,
82 int *xlow, int *xhigh)
85 c->prev_samples[c->prev_samples_pos++] = samples[0];
86 c->prev_samples[c->prev_samples_pos++] = samples[1];
87 ff_g722_apply_qmf(c->prev_samples + c->prev_samples_pos - 24, &xout1, &xout2);
88 *xlow = xout1 + xout2 >> 13;
89 *xhigh = xout1 - xout2 >> 13;
90 if (c->prev_samples_pos >= PREV_SAMPLES_BUF_SIZE) {
91 memmove(c->prev_samples,
92 c->prev_samples + c->prev_samples_pos - 22,
93 22 * sizeof(c->prev_samples[0]));
94 c->prev_samples_pos = 22;
98 static inline int encode_high(const struct G722Band *state, int xhigh)
100 int diff = av_clip_int16(xhigh - state->s_predictor);
101 int pred = 141 * state->scale_factor >> 8;
102 /* = diff >= 0 ? (diff < pred) + 2 : diff >= -pred */
103 return ((diff ^ (diff >> (sizeof(diff)*8-1))) < pred) + 2*(diff >= 0);
106 static inline int encode_low(const struct G722Band* state, int xlow)
108 int diff = av_clip_int16(xlow - state->s_predictor);
109 /* = diff >= 0 ? diff : -(diff + 1) */
110 int limit = diff ^ (diff >> (sizeof(diff)*8-1));
112 limit = limit + 1 << 10;
113 if (limit > low_quant[8] * state->scale_factor)
115 while (i < 29 && limit > low_quant[i] * state->scale_factor)
117 return (diff < 0 ? (i < 2 ? 63 : 33) : 61) - i;
120 static int g722_encode_trellis(AVCodecContext *avctx,
121 uint8_t *dst, int buf_size, void *data)
123 G722Context *c = avctx->priv_data;
124 const int16_t *samples = data;
126 int frontier = 1 << avctx->trellis;
127 struct TrellisNode **nodes[2];
128 struct TrellisNode **nodes_next[2];
129 int pathn[2] = {0, 0}, froze = -1;
130 struct TrellisPath *p[2];
132 for (i = 0; i < 2; i++) {
133 nodes[i] = c->nodep_buf[i];
134 nodes_next[i] = c->nodep_buf[i] + frontier;
135 memset(c->nodep_buf[i], 0, 2 * frontier * sizeof(*c->nodep_buf));
136 nodes[i][0] = c->node_buf[i] + frontier;
137 nodes[i][0]->ssd = 0;
138 nodes[i][0]->path = 0;
139 nodes[i][0]->state = c->band[i];
142 for (i = 0; i < buf_size; i++) {
144 struct TrellisNode *next[2];
145 int heap_pos[2] = {0, 0};
147 for (j = 0; j < 2; j++) {
148 next[j] = c->node_buf[j] + frontier*(i & 1);
149 memset(nodes_next[j], 0, frontier * sizeof(**nodes_next));
152 filter_samples(c, &samples[2*i], &xlow, &xhigh);
154 for (j = 0; j < frontier && nodes[0][j]; j++) {
155 /* Only k >> 2 affects the future adaptive state, therefore testing
156 * small steps that don't change k >> 2 is useless, the orignal
157 * value from encode_low is better than them. Since we step k
158 * in steps of 4, make sure range is a multiple of 4, so that
159 * we don't miss the original value from encode_low. */
160 int range = j < frontier/2 ? 4 : 0;
161 struct TrellisNode *cur_node = nodes[0][j];
163 int ilow = encode_low(&cur_node->state, xlow);
165 for (k = ilow - range; k <= ilow + range && k <= 63; k += 4) {
166 int decoded, dec_diff, pos;
168 struct TrellisNode* node;
173 decoded = av_clip((cur_node->state.scale_factor *
174 ff_g722_low_inv_quant6[k] >> 10)
175 + cur_node->state.s_predictor, -16384, 16383);
176 dec_diff = xlow - decoded;
178 #define STORE_NODE(index, UPDATE, VALUE)\
179 ssd = cur_node->ssd + dec_diff*dec_diff;\
180 /* Check for wraparound. Using 64 bit ssd counters would \
181 * be simpler, but is slower on x86 32 bit. */\
182 if (ssd < cur_node->ssd)\
184 if (heap_pos[index] < frontier) {\
185 pos = heap_pos[index]++;\
186 assert(pathn[index] < FREEZE_INTERVAL * frontier);\
187 node = nodes_next[index][pos] = next[index]++;\
188 node->path = pathn[index]++;\
190 /* Try to replace one of the leaf nodes with the new \
191 * one, but not always testing the same leaf position */\
192 pos = (frontier>>1) + (heap_pos[index] & ((frontier>>1) - 1));\
193 if (ssd >= nodes_next[index][pos]->ssd)\
196 node = nodes_next[index][pos];\
199 node->state = cur_node->state;\
201 c->paths[index][node->path].value = VALUE;\
202 c->paths[index][node->path].prev = cur_node->path;\
203 /* Sift the newly inserted node up in the heap to restore \
204 * the heap property */\
206 int parent = (pos - 1) >> 1;\
207 if (nodes_next[index][parent]->ssd <= ssd)\
209 FFSWAP(struct TrellisNode*, nodes_next[index][parent],\
210 nodes_next[index][pos]);\
213 STORE_NODE(0, ff_g722_update_low_predictor(&node->state, k >> 2), k);
217 for (j = 0; j < frontier && nodes[1][j]; j++) {
219 struct TrellisNode *cur_node = nodes[1][j];
221 /* We don't try to get any initial guess for ihigh via
222 * encode_high - since there's only 4 possible values, test
223 * them all. Testing all of these gives a much, much larger
224 * gain than testing a larger range around ilow. */
225 for (ihigh = 0; ihigh < 4; ihigh++) {
226 int dhigh, decoded, dec_diff, pos;
228 struct TrellisNode* node;
230 dhigh = cur_node->state.scale_factor *
231 ff_g722_high_inv_quant[ihigh] >> 10;
232 decoded = av_clip(dhigh + cur_node->state.s_predictor,
234 dec_diff = xhigh - decoded;
236 STORE_NODE(1, ff_g722_update_high_predictor(&node->state, dhigh, ihigh), ihigh);
240 for (j = 0; j < 2; j++) {
241 FFSWAP(struct TrellisNode**, nodes[j], nodes_next[j]);
243 if (nodes[j][0]->ssd > (1 << 16)) {
244 for (k = 1; k < frontier && nodes[j][k]; k++)
245 nodes[j][k]->ssd -= nodes[j][0]->ssd;
246 nodes[j][0]->ssd = 0;
250 if (i == froze + FREEZE_INTERVAL) {
251 p[0] = &c->paths[0][nodes[0][0]->path];
252 p[1] = &c->paths[1][nodes[1][0]->path];
253 for (j = i; j > froze; j--) {
254 dst[j] = p[1]->value << 6 | p[0]->value;
255 p[0] = &c->paths[0][p[0]->prev];
256 p[1] = &c->paths[1][p[1]->prev];
259 pathn[0] = pathn[1] = 0;
260 memset(nodes[0] + 1, 0, (frontier - 1)*sizeof(**nodes));
261 memset(nodes[1] + 1, 0, (frontier - 1)*sizeof(**nodes));
265 p[0] = &c->paths[0][nodes[0][0]->path];
266 p[1] = &c->paths[1][nodes[1][0]->path];
267 for (j = i; j > froze; j--) {
268 dst[j] = p[1]->value << 6 | p[0]->value;
269 p[0] = &c->paths[0][p[0]->prev];
270 p[1] = &c->paths[1][p[1]->prev];
272 c->band[0] = nodes[0][0]->state;
273 c->band[1] = nodes[1][0]->state;
278 static int g722_encode_frame(AVCodecContext *avctx,
279 uint8_t *dst, int buf_size, void *data)
281 G722Context *c = avctx->priv_data;
282 const int16_t *samples = data;
286 return g722_encode_trellis(avctx, dst, buf_size, data);
288 for (i = 0; i < buf_size; i++) {
289 int xlow, xhigh, ihigh, ilow;
290 filter_samples(c, &samples[2*i], &xlow, &xhigh);
291 ihigh = encode_high(&c->band[1], xhigh);
292 ilow = encode_low(&c->band[0], xlow);
293 ff_g722_update_high_predictor(&c->band[1], c->band[1].scale_factor *
294 ff_g722_high_inv_quant[ihigh] >> 10, ihigh);
295 ff_g722_update_low_predictor(&c->band[0], ilow >> 2);
296 *dst++ = ihigh << 6 | ilow;
301 AVCodec ff_adpcm_g722_encoder = {
303 .type = AVMEDIA_TYPE_AUDIO,
304 .id = CODEC_ID_ADPCM_G722,
305 .priv_data_size = sizeof(G722Context),
306 .init = g722_encode_init,
307 .close = g722_encode_close,
308 .encode = g722_encode_frame,
309 .long_name = NULL_IF_CONFIG_SMALL("G.722 ADPCM"),
310 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},