You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

406 lines
15KB

  1. /*
  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
  7. *
  8. * This file is part of Libav.
  9. *
  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.
  14. *
  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.
  19. *
  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
  23. */
  24. /**
  25. * @file
  26. * G.722 ADPCM audio encoder
  27. */
  28. #include "libavutil/avassert.h"
  29. #include "avcodec.h"
  30. #include "internal.h"
  31. #include "g722.h"
  32. #include "libavutil/common.h"
  33. #define FREEZE_INTERVAL 128
  34. /* This is an arbitrary value. Allowing insanely large values leads to strange
  35. problems, so we limit it to a reasonable value */
  36. #define MAX_FRAME_SIZE 32768
  37. /* We clip the value of avctx->trellis to prevent data type overflows and
  38. undefined behavior. Using larger values is insanely slow anyway. */
  39. #define MIN_TRELLIS 0
  40. #define MAX_TRELLIS 16
  41. static av_cold int g722_encode_close(AVCodecContext *avctx)
  42. {
  43. G722Context *c = avctx->priv_data;
  44. int i;
  45. for (i = 0; i < 2; i++) {
  46. av_freep(&c->paths[i]);
  47. av_freep(&c->node_buf[i]);
  48. av_freep(&c->nodep_buf[i]);
  49. }
  50. #if FF_API_OLD_ENCODE_AUDIO
  51. av_freep(&avctx->coded_frame);
  52. #endif
  53. return 0;
  54. }
  55. static av_cold int g722_encode_init(AVCodecContext * avctx)
  56. {
  57. G722Context *c = avctx->priv_data;
  58. int ret;
  59. if (avctx->channels != 1) {
  60. av_log(avctx, AV_LOG_ERROR, "Only mono tracks are allowed.\n");
  61. return AVERROR_INVALIDDATA;
  62. }
  63. c->band[0].scale_factor = 8;
  64. c->band[1].scale_factor = 2;
  65. c->prev_samples_pos = 22;
  66. if (avctx->trellis) {
  67. int frontier = 1 << avctx->trellis;
  68. int max_paths = frontier * FREEZE_INTERVAL;
  69. int i;
  70. for (i = 0; i < 2; i++) {
  71. c->paths[i] = av_mallocz(max_paths * sizeof(**c->paths));
  72. c->node_buf[i] = av_mallocz(2 * frontier * sizeof(**c->node_buf));
  73. c->nodep_buf[i] = av_mallocz(2 * frontier * sizeof(**c->nodep_buf));
  74. if (!c->paths[i] || !c->node_buf[i] || !c->nodep_buf[i]) {
  75. ret = AVERROR(ENOMEM);
  76. goto error;
  77. }
  78. }
  79. }
  80. if (avctx->frame_size) {
  81. /* validate frame size */
  82. if (avctx->frame_size & 1 || avctx->frame_size > MAX_FRAME_SIZE) {
  83. int new_frame_size;
  84. if (avctx->frame_size == 1)
  85. new_frame_size = 2;
  86. else if (avctx->frame_size > MAX_FRAME_SIZE)
  87. new_frame_size = MAX_FRAME_SIZE;
  88. else
  89. new_frame_size = avctx->frame_size - 1;
  90. av_log(avctx, AV_LOG_WARNING, "Requested frame size is not "
  91. "allowed. Using %d instead of %d\n", new_frame_size,
  92. avctx->frame_size);
  93. avctx->frame_size = new_frame_size;
  94. }
  95. } else {
  96. /* This is arbitrary. We use 320 because it's 20ms @ 16kHz, which is
  97. a common packet size for VoIP applications */
  98. avctx->frame_size = 320;
  99. }
  100. avctx->delay = 22;
  101. if (avctx->trellis) {
  102. /* validate trellis */
  103. if (avctx->trellis < MIN_TRELLIS || avctx->trellis > MAX_TRELLIS) {
  104. int new_trellis = av_clip(avctx->trellis, MIN_TRELLIS, MAX_TRELLIS);
  105. av_log(avctx, AV_LOG_WARNING, "Requested trellis value is not "
  106. "allowed. Using %d instead of %d\n", new_trellis,
  107. avctx->trellis);
  108. avctx->trellis = new_trellis;
  109. }
  110. }
  111. #if FF_API_OLD_ENCODE_AUDIO
  112. avctx->coded_frame = avcodec_alloc_frame();
  113. if (!avctx->coded_frame) {
  114. ret = AVERROR(ENOMEM);
  115. goto error;
  116. }
  117. #endif
  118. return 0;
  119. error:
  120. g722_encode_close(avctx);
  121. return ret;
  122. }
  123. static const int16_t low_quant[33] = {
  124. 35, 72, 110, 150, 190, 233, 276, 323,
  125. 370, 422, 473, 530, 587, 650, 714, 786,
  126. 858, 940, 1023, 1121, 1219, 1339, 1458, 1612,
  127. 1765, 1980, 2195, 2557, 2919
  128. };
  129. static inline void filter_samples(G722Context *c, const int16_t *samples,
  130. int *xlow, int *xhigh)
  131. {
  132. int xout1, xout2;
  133. c->prev_samples[c->prev_samples_pos++] = samples[0];
  134. c->prev_samples[c->prev_samples_pos++] = samples[1];
  135. ff_g722_apply_qmf(c->prev_samples + c->prev_samples_pos - 24, &xout1, &xout2);
  136. *xlow = xout1 + xout2 >> 14;
  137. *xhigh = xout1 - xout2 >> 14;
  138. if (c->prev_samples_pos >= PREV_SAMPLES_BUF_SIZE) {
  139. memmove(c->prev_samples,
  140. c->prev_samples + c->prev_samples_pos - 22,
  141. 22 * sizeof(c->prev_samples[0]));
  142. c->prev_samples_pos = 22;
  143. }
  144. }
  145. static inline int encode_high(const struct G722Band *state, int xhigh)
  146. {
  147. int diff = av_clip_int16(xhigh - state->s_predictor);
  148. int pred = 141 * state->scale_factor >> 8;
  149. /* = diff >= 0 ? (diff < pred) + 2 : diff >= -pred */
  150. return ((diff ^ (diff >> (sizeof(diff)*8-1))) < pred) + 2*(diff >= 0);
  151. }
  152. static inline int encode_low(const struct G722Band* state, int xlow)
  153. {
  154. int diff = av_clip_int16(xlow - state->s_predictor);
  155. /* = diff >= 0 ? diff : -(diff + 1) */
  156. int limit = diff ^ (diff >> (sizeof(diff)*8-1));
  157. int i = 0;
  158. limit = limit + 1 << 10;
  159. if (limit > low_quant[8] * state->scale_factor)
  160. i = 9;
  161. while (i < 29 && limit > low_quant[i] * state->scale_factor)
  162. i++;
  163. return (diff < 0 ? (i < 2 ? 63 : 33) : 61) - i;
  164. }
  165. static void g722_encode_trellis(G722Context *c, int trellis,
  166. uint8_t *dst, int nb_samples,
  167. const int16_t *samples)
  168. {
  169. int i, j, k;
  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];
  175. for (i = 0; i < 2; i++) {
  176. nodes[i] = c->nodep_buf[i];
  177. nodes_next[i] = c->nodep_buf[i] + frontier;
  178. memset(c->nodep_buf[i], 0, 2 * frontier * sizeof(*c->nodep_buf[i]));
  179. nodes[i][0] = c->node_buf[i] + frontier;
  180. nodes[i][0]->ssd = 0;
  181. nodes[i][0]->path = 0;
  182. nodes[i][0]->state = c->band[i];
  183. }
  184. for (i = 0; i < nb_samples >> 1; i++) {
  185. int xlow, xhigh;
  186. struct TrellisNode *next[2];
  187. int heap_pos[2] = {0, 0};
  188. for (j = 0; j < 2; j++) {
  189. next[j] = c->node_buf[j] + frontier*(i & 1);
  190. memset(nodes_next[j], 0, frontier * sizeof(**nodes_next));
  191. }
  192. filter_samples(c, &samples[2*i], &xlow, &xhigh);
  193. for (j = 0; j < frontier && nodes[0][j]; j++) {
  194. /* Only k >> 2 affects the future adaptive state, therefore testing
  195. * small steps that don't change k >> 2 is useless, the original
  196. * value from encode_low is better than them. Since we step k
  197. * in steps of 4, make sure range is a multiple of 4, so that
  198. * we don't miss the original value from encode_low. */
  199. int range = j < frontier/2 ? 4 : 0;
  200. struct TrellisNode *cur_node = nodes[0][j];
  201. int ilow = encode_low(&cur_node->state, xlow);
  202. for (k = ilow - range; k <= ilow + range && k <= 63; k += 4) {
  203. int decoded, dec_diff, pos;
  204. uint32_t ssd;
  205. struct TrellisNode* node;
  206. if (k < 0)
  207. continue;
  208. decoded = av_clip((cur_node->state.scale_factor *
  209. ff_g722_low_inv_quant6[k] >> 10)
  210. + cur_node->state.s_predictor, -16384, 16383);
  211. dec_diff = xlow - decoded;
  212. #define STORE_NODE(index, UPDATE, VALUE)\
  213. ssd = cur_node->ssd + dec_diff*dec_diff;\
  214. /* Check for wraparound. Using 64 bit ssd counters would \
  215. * be simpler, but is slower on x86 32 bit. */\
  216. if (ssd < cur_node->ssd)\
  217. continue;\
  218. if (heap_pos[index] < frontier) {\
  219. pos = heap_pos[index]++;\
  220. av_assert2(pathn[index] < FREEZE_INTERVAL * frontier);\
  221. node = nodes_next[index][pos] = next[index]++;\
  222. node->path = pathn[index]++;\
  223. } else {\
  224. /* Try to replace one of the leaf nodes with the new \
  225. * one, but not always testing the same leaf position */\
  226. pos = (frontier>>1) + (heap_pos[index] & ((frontier>>1) - 1));\
  227. if (ssd >= nodes_next[index][pos]->ssd)\
  228. continue;\
  229. heap_pos[index]++;\
  230. node = nodes_next[index][pos];\
  231. }\
  232. node->ssd = ssd;\
  233. node->state = cur_node->state;\
  234. UPDATE;\
  235. c->paths[index][node->path].value = VALUE;\
  236. c->paths[index][node->path].prev = cur_node->path;\
  237. /* Sift the newly inserted node up in the heap to restore \
  238. * the heap property */\
  239. while (pos > 0) {\
  240. int parent = (pos - 1) >> 1;\
  241. if (nodes_next[index][parent]->ssd <= ssd)\
  242. break;\
  243. FFSWAP(struct TrellisNode*, nodes_next[index][parent],\
  244. nodes_next[index][pos]);\
  245. pos = parent;\
  246. }
  247. STORE_NODE(0, ff_g722_update_low_predictor(&node->state, k >> 2), k);
  248. }
  249. }
  250. for (j = 0; j < frontier && nodes[1][j]; j++) {
  251. int ihigh;
  252. struct TrellisNode *cur_node = nodes[1][j];
  253. /* We don't try to get any initial guess for ihigh via
  254. * encode_high - since there's only 4 possible values, test
  255. * them all. Testing all of these gives a much, much larger
  256. * gain than testing a larger range around ilow. */
  257. for (ihigh = 0; ihigh < 4; ihigh++) {
  258. int dhigh, decoded, dec_diff, pos;
  259. uint32_t ssd;
  260. struct TrellisNode* node;
  261. dhigh = cur_node->state.scale_factor *
  262. ff_g722_high_inv_quant[ihigh] >> 10;
  263. decoded = av_clip(dhigh + cur_node->state.s_predictor,
  264. -16384, 16383);
  265. dec_diff = xhigh - decoded;
  266. STORE_NODE(1, ff_g722_update_high_predictor(&node->state, dhigh, ihigh), ihigh);
  267. }
  268. }
  269. for (j = 0; j < 2; j++) {
  270. FFSWAP(struct TrellisNode**, nodes[j], nodes_next[j]);
  271. if (nodes[j][0]->ssd > (1 << 16)) {
  272. for (k = 1; k < frontier && nodes[j][k]; k++)
  273. nodes[j][k]->ssd -= nodes[j][0]->ssd;
  274. nodes[j][0]->ssd = 0;
  275. }
  276. }
  277. if (i == froze + FREEZE_INTERVAL) {
  278. p[0] = &c->paths[0][nodes[0][0]->path];
  279. p[1] = &c->paths[1][nodes[1][0]->path];
  280. for (j = i; j > froze; j--) {
  281. dst[j] = p[1]->value << 6 | p[0]->value;
  282. p[0] = &c->paths[0][p[0]->prev];
  283. p[1] = &c->paths[1][p[1]->prev];
  284. }
  285. froze = i;
  286. pathn[0] = pathn[1] = 0;
  287. memset(nodes[0] + 1, 0, (frontier - 1)*sizeof(**nodes));
  288. memset(nodes[1] + 1, 0, (frontier - 1)*sizeof(**nodes));
  289. }
  290. }
  291. p[0] = &c->paths[0][nodes[0][0]->path];
  292. p[1] = &c->paths[1][nodes[1][0]->path];
  293. for (j = i; j > froze; j--) {
  294. dst[j] = p[1]->value << 6 | p[0]->value;
  295. p[0] = &c->paths[0][p[0]->prev];
  296. p[1] = &c->paths[1][p[1]->prev];
  297. }
  298. c->band[0] = nodes[0][0]->state;
  299. c->band[1] = nodes[1][0]->state;
  300. }
  301. static av_always_inline void encode_byte(G722Context *c, uint8_t *dst,
  302. const int16_t *samples)
  303. {
  304. int xlow, xhigh, ilow, ihigh;
  305. filter_samples(c, samples, &xlow, &xhigh);
  306. ihigh = encode_high(&c->band[1], xhigh);
  307. ilow = encode_low (&c->band[0], xlow);
  308. ff_g722_update_high_predictor(&c->band[1], c->band[1].scale_factor *
  309. ff_g722_high_inv_quant[ihigh] >> 10, ihigh);
  310. ff_g722_update_low_predictor(&c->band[0], ilow >> 2);
  311. *dst = ihigh << 6 | ilow;
  312. }
  313. static void g722_encode_no_trellis(G722Context *c,
  314. uint8_t *dst, int nb_samples,
  315. const int16_t *samples)
  316. {
  317. int i;
  318. for (i = 0; i < nb_samples; i += 2)
  319. encode_byte(c, dst++, &samples[i]);
  320. }
  321. static int g722_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
  322. const AVFrame *frame, int *got_packet_ptr)
  323. {
  324. G722Context *c = avctx->priv_data;
  325. const int16_t *samples = (const int16_t *)frame->data[0];
  326. int nb_samples, out_size, ret;
  327. out_size = (frame->nb_samples + 1) / 2;
  328. if ((ret = ff_alloc_packet2(avctx, avpkt, out_size)))
  329. return ret;
  330. nb_samples = frame->nb_samples - (frame->nb_samples & 1);
  331. if (avctx->trellis)
  332. g722_encode_trellis(c, avctx->trellis, avpkt->data, nb_samples, samples);
  333. else
  334. g722_encode_no_trellis(c, avpkt->data, nb_samples, samples);
  335. /* handle last frame with odd frame_size */
  336. if (nb_samples < frame->nb_samples) {
  337. int16_t last_samples[2] = { samples[nb_samples], samples[nb_samples] };
  338. encode_byte(c, &avpkt->data[nb_samples >> 1], last_samples);
  339. }
  340. if (frame->pts != AV_NOPTS_VALUE)
  341. avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->delay);
  342. *got_packet_ptr = 1;
  343. return 0;
  344. }
  345. AVCodec ff_adpcm_g722_encoder = {
  346. .name = "g722",
  347. .type = AVMEDIA_TYPE_AUDIO,
  348. .id = AV_CODEC_ID_ADPCM_G722,
  349. .priv_data_size = sizeof(G722Context),
  350. .init = g722_encode_init,
  351. .close = g722_encode_close,
  352. .encode2 = g722_encode_frame,
  353. .capabilities = CODEC_CAP_SMALL_LAST_FRAME,
  354. .long_name = NULL_IF_CONFIG_SMALL("G.722 ADPCM"),
  355. .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
  356. AV_SAMPLE_FMT_NONE },
  357. };