yading@10: /*
yading@10:  * Copyright (c) CMU 1993 Computer Science, Speech Group
yading@10:  *                        Chengxiang Lu and Alex Hauptmann
yading@10:  * Copyright (c) 2005 Steve Underwood <steveu at coppice.org>
yading@10:  * Copyright (c) 2009 Kenan Gillet
yading@10:  * Copyright (c) 2010 Martin Storsjo
yading@10:  *
yading@10:  * This file is part of Libav.
yading@10:  *
yading@10:  * Libav is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * Libav is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with Libav; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: /**
yading@10:  * @file
yading@10:  * G.722 ADPCM audio encoder
yading@10:  */
yading@10: 
yading@10: #include "libavutil/avassert.h"
yading@10: #include "avcodec.h"
yading@10: #include "internal.h"
yading@10: #include "g722.h"
yading@10: #include "libavutil/common.h"
yading@10: 
yading@10: #define FREEZE_INTERVAL 128
yading@10: 
yading@10: /* This is an arbitrary value. Allowing insanely large values leads to strange
yading@10:    problems, so we limit it to a reasonable value */
yading@10: #define MAX_FRAME_SIZE 32768
yading@10: 
yading@10: /* We clip the value of avctx->trellis to prevent data type overflows and
yading@10:    undefined behavior. Using larger values is insanely slow anyway. */
yading@10: #define MIN_TRELLIS 0
yading@10: #define MAX_TRELLIS 16
yading@10: 
yading@10: static av_cold int g722_encode_close(AVCodecContext *avctx)
yading@10: {
yading@10:     G722Context *c = avctx->priv_data;
yading@10:     int i;
yading@10:     for (i = 0; i < 2; i++) {
yading@10:         av_freep(&c->paths[i]);
yading@10:         av_freep(&c->node_buf[i]);
yading@10:         av_freep(&c->nodep_buf[i]);
yading@10:     }
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static av_cold int g722_encode_init(AVCodecContext * avctx)
yading@10: {
yading@10:     G722Context *c = avctx->priv_data;
yading@10:     int ret;
yading@10: 
yading@10:     if (avctx->channels != 1) {
yading@10:         av_log(avctx, AV_LOG_ERROR, "Only mono tracks are allowed.\n");
yading@10:         return AVERROR_INVALIDDATA;
yading@10:     }
yading@10: 
yading@10:     c->band[0].scale_factor = 8;
yading@10:     c->band[1].scale_factor = 2;
yading@10:     c->prev_samples_pos = 22;
yading@10: 
yading@10:     if (avctx->trellis) {
yading@10:         int frontier = 1 << avctx->trellis;
yading@10:         int max_paths = frontier * FREEZE_INTERVAL;
yading@10:         int i;
yading@10:         for (i = 0; i < 2; i++) {
yading@10:             c->paths[i] = av_mallocz(max_paths * sizeof(**c->paths));
yading@10:             c->node_buf[i] = av_mallocz(2 * frontier * sizeof(**c->node_buf));
yading@10:             c->nodep_buf[i] = av_mallocz(2 * frontier * sizeof(**c->nodep_buf));
yading@10:             if (!c->paths[i] || !c->node_buf[i] || !c->nodep_buf[i]) {
yading@10:                 ret = AVERROR(ENOMEM);
yading@10:                 goto error;
yading@10:             }
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     if (avctx->frame_size) {
yading@10:         /* validate frame size */
yading@10:         if (avctx->frame_size & 1 || avctx->frame_size > MAX_FRAME_SIZE) {
yading@10:             int new_frame_size;
yading@10: 
yading@10:             if (avctx->frame_size == 1)
yading@10:                 new_frame_size = 2;
yading@10:             else if (avctx->frame_size > MAX_FRAME_SIZE)
yading@10:                 new_frame_size = MAX_FRAME_SIZE;
yading@10:             else
yading@10:                 new_frame_size = avctx->frame_size - 1;
yading@10: 
yading@10:             av_log(avctx, AV_LOG_WARNING, "Requested frame size is not "
yading@10:                    "allowed. Using %d instead of %d\n", new_frame_size,
yading@10:                    avctx->frame_size);
yading@10:             avctx->frame_size = new_frame_size;
yading@10:         }
yading@10:     } else {
yading@10:         /* This is arbitrary. We use 320 because it's 20ms @ 16kHz, which is
yading@10:            a common packet size for VoIP applications */
yading@10:         avctx->frame_size = 320;
yading@10:     }
yading@10:     avctx->delay = 22;
yading@10: 
yading@10:     if (avctx->trellis) {
yading@10:         /* validate trellis */
yading@10:         if (avctx->trellis < MIN_TRELLIS || avctx->trellis > MAX_TRELLIS) {
yading@10:             int new_trellis = av_clip(avctx->trellis, MIN_TRELLIS, MAX_TRELLIS);
yading@10:             av_log(avctx, AV_LOG_WARNING, "Requested trellis value is not "
yading@10:                    "allowed. Using %d instead of %d\n", new_trellis,
yading@10:                    avctx->trellis);
yading@10:             avctx->trellis = new_trellis;
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     return 0;
yading@10: error:
yading@10:     g722_encode_close(avctx);
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: static const int16_t low_quant[33] = {
yading@10:       35,   72,  110,  150,  190,  233,  276,  323,
yading@10:      370,  422,  473,  530,  587,  650,  714,  786,
yading@10:      858,  940, 1023, 1121, 1219, 1339, 1458, 1612,
yading@10:     1765, 1980, 2195, 2557, 2919
yading@10: };
yading@10: 
yading@10: static inline void filter_samples(G722Context *c, const int16_t *samples,
yading@10:                                   int *xlow, int *xhigh)
yading@10: {
yading@10:     int xout1, xout2;
yading@10:     c->prev_samples[c->prev_samples_pos++] = samples[0];
yading@10:     c->prev_samples[c->prev_samples_pos++] = samples[1];
yading@10:     ff_g722_apply_qmf(c->prev_samples + c->prev_samples_pos - 24, &xout1, &xout2);
yading@10:     *xlow  = xout1 + xout2 >> 14;
yading@10:     *xhigh = xout1 - xout2 >> 14;
yading@10:     if (c->prev_samples_pos >= PREV_SAMPLES_BUF_SIZE) {
yading@10:         memmove(c->prev_samples,
yading@10:                 c->prev_samples + c->prev_samples_pos - 22,
yading@10:                 22 * sizeof(c->prev_samples[0]));
yading@10:         c->prev_samples_pos = 22;
yading@10:     }
yading@10: }
yading@10: 
yading@10: static inline int encode_high(const struct G722Band *state, int xhigh)
yading@10: {
yading@10:     int diff = av_clip_int16(xhigh - state->s_predictor);
yading@10:     int pred = 141 * state->scale_factor >> 8;
yading@10:            /* = diff >= 0 ? (diff < pred) + 2 : diff >= -pred */
yading@10:     return ((diff ^ (diff >> (sizeof(diff)*8-1))) < pred) + 2*(diff >= 0);
yading@10: }
yading@10: 
yading@10: static inline int encode_low(const struct G722Band* state, int xlow)
yading@10: {
yading@10:     int diff  = av_clip_int16(xlow - state->s_predictor);
yading@10:            /* = diff >= 0 ? diff : -(diff + 1) */
yading@10:     int limit = diff ^ (diff >> (sizeof(diff)*8-1));
yading@10:     int i = 0;
yading@10:     limit = limit + 1 << 10;
yading@10:     if (limit > low_quant[8] * state->scale_factor)
yading@10:         i = 9;
yading@10:     while (i < 29 && limit > low_quant[i] * state->scale_factor)
yading@10:         i++;
yading@10:     return (diff < 0 ? (i < 2 ? 63 : 33) : 61) - i;
yading@10: }
yading@10: 
yading@10: static void g722_encode_trellis(G722Context *c, int trellis,
yading@10:                                 uint8_t *dst, int nb_samples,
yading@10:                                 const int16_t *samples)
yading@10: {
yading@10:     int i, j, k;
yading@10:     int frontier = 1 << trellis;
yading@10:     struct TrellisNode **nodes[2];
yading@10:     struct TrellisNode **nodes_next[2];
yading@10:     int pathn[2] = {0, 0}, froze = -1;
yading@10:     struct TrellisPath *p[2];
yading@10: 
yading@10:     for (i = 0; i < 2; i++) {
yading@10:         nodes[i] = c->nodep_buf[i];
yading@10:         nodes_next[i] = c->nodep_buf[i] + frontier;
yading@10:         memset(c->nodep_buf[i], 0, 2 * frontier * sizeof(*c->nodep_buf[i]));
yading@10:         nodes[i][0] = c->node_buf[i] + frontier;
yading@10:         nodes[i][0]->ssd = 0;
yading@10:         nodes[i][0]->path = 0;
yading@10:         nodes[i][0]->state = c->band[i];
yading@10:     }
yading@10: 
yading@10:     for (i = 0; i < nb_samples >> 1; i++) {
yading@10:         int xlow, xhigh;
yading@10:         struct TrellisNode *next[2];
yading@10:         int heap_pos[2] = {0, 0};
yading@10: 
yading@10:         for (j = 0; j < 2; j++) {
yading@10:             next[j] = c->node_buf[j] + frontier*(i & 1);
yading@10:             memset(nodes_next[j], 0, frontier * sizeof(**nodes_next));
yading@10:         }
yading@10: 
yading@10:         filter_samples(c, &samples[2*i], &xlow, &xhigh);
yading@10: 
yading@10:         for (j = 0; j < frontier && nodes[0][j]; j++) {
yading@10:             /* Only k >> 2 affects the future adaptive state, therefore testing
yading@10:              * small steps that don't change k >> 2 is useless, the original
yading@10:              * value from encode_low is better than them. Since we step k
yading@10:              * in steps of 4, make sure range is a multiple of 4, so that
yading@10:              * we don't miss the original value from encode_low. */
yading@10:             int range = j < frontier/2 ? 4 : 0;
yading@10:             struct TrellisNode *cur_node = nodes[0][j];
yading@10: 
yading@10:             int ilow = encode_low(&cur_node->state, xlow);
yading@10: 
yading@10:             for (k = ilow - range; k <= ilow + range && k <= 63; k += 4) {
yading@10:                 int decoded, dec_diff, pos;
yading@10:                 uint32_t ssd;
yading@10:                 struct TrellisNode* node;
yading@10: 
yading@10:                 if (k < 0)
yading@10:                     continue;
yading@10: 
yading@10:                 decoded = av_clip((cur_node->state.scale_factor *
yading@10:                                   ff_g722_low_inv_quant6[k] >> 10)
yading@10:                                 + cur_node->state.s_predictor, -16384, 16383);
yading@10:                 dec_diff = xlow - decoded;
yading@10: 
yading@10: #define STORE_NODE(index, UPDATE, VALUE)\
yading@10:                 ssd = cur_node->ssd + dec_diff*dec_diff;\
yading@10:                 /* Check for wraparound. Using 64 bit ssd counters would \
yading@10:                  * be simpler, but is slower on x86 32 bit. */\
yading@10:                 if (ssd < cur_node->ssd)\
yading@10:                     continue;\
yading@10:                 if (heap_pos[index] < frontier) {\
yading@10:                     pos = heap_pos[index]++;\
yading@10:                     av_assert2(pathn[index] < FREEZE_INTERVAL * frontier);\
yading@10:                     node = nodes_next[index][pos] = next[index]++;\
yading@10:                     node->path = pathn[index]++;\
yading@10:                 } else {\
yading@10:                     /* Try to replace one of the leaf nodes with the new \
yading@10:                      * one, but not always testing the same leaf position */\
yading@10:                     pos = (frontier>>1) + (heap_pos[index] & ((frontier>>1) - 1));\
yading@10:                     if (ssd >= nodes_next[index][pos]->ssd)\
yading@10:                         continue;\
yading@10:                     heap_pos[index]++;\
yading@10:                     node = nodes_next[index][pos];\
yading@10:                 }\
yading@10:                 node->ssd = ssd;\
yading@10:                 node->state = cur_node->state;\
yading@10:                 UPDATE;\
yading@10:                 c->paths[index][node->path].value = VALUE;\
yading@10:                 c->paths[index][node->path].prev = cur_node->path;\
yading@10:                 /* Sift the newly inserted node up in the heap to restore \
yading@10:                  * the heap property */\
yading@10:                 while (pos > 0) {\
yading@10:                     int parent = (pos - 1) >> 1;\
yading@10:                     if (nodes_next[index][parent]->ssd <= ssd)\
yading@10:                         break;\
yading@10:                     FFSWAP(struct TrellisNode*, nodes_next[index][parent],\
yading@10:                                                 nodes_next[index][pos]);\
yading@10:                     pos = parent;\
yading@10:                 }
yading@10:                 STORE_NODE(0, ff_g722_update_low_predictor(&node->state, k >> 2), k);
yading@10:             }
yading@10:         }
yading@10: 
yading@10:         for (j = 0; j < frontier && nodes[1][j]; j++) {
yading@10:             int ihigh;
yading@10:             struct TrellisNode *cur_node = nodes[1][j];
yading@10: 
yading@10:             /* We don't try to get any initial guess for ihigh via
yading@10:              * encode_high - since there's only 4 possible values, test
yading@10:              * them all. Testing all of these gives a much, much larger
yading@10:              * gain than testing a larger range around ilow. */
yading@10:             for (ihigh = 0; ihigh < 4; ihigh++) {
yading@10:                 int dhigh, decoded, dec_diff, pos;
yading@10:                 uint32_t ssd;
yading@10:                 struct TrellisNode* node;
yading@10: 
yading@10:                 dhigh = cur_node->state.scale_factor *
yading@10:                         ff_g722_high_inv_quant[ihigh] >> 10;
yading@10:                 decoded = av_clip(dhigh + cur_node->state.s_predictor,
yading@10:                                   -16384, 16383);
yading@10:                 dec_diff = xhigh - decoded;
yading@10: 
yading@10:                 STORE_NODE(1, ff_g722_update_high_predictor(&node->state, dhigh, ihigh), ihigh);
yading@10:             }
yading@10:         }
yading@10: 
yading@10:         for (j = 0; j < 2; j++) {
yading@10:             FFSWAP(struct TrellisNode**, nodes[j], nodes_next[j]);
yading@10: 
yading@10:             if (nodes[j][0]->ssd > (1 << 16)) {
yading@10:                 for (k = 1; k < frontier && nodes[j][k]; k++)
yading@10:                     nodes[j][k]->ssd -= nodes[j][0]->ssd;
yading@10:                 nodes[j][0]->ssd = 0;
yading@10:             }
yading@10:         }
yading@10: 
yading@10:         if (i == froze + FREEZE_INTERVAL) {
yading@10:             p[0] = &c->paths[0][nodes[0][0]->path];
yading@10:             p[1] = &c->paths[1][nodes[1][0]->path];
yading@10:             for (j = i; j > froze; j--) {
yading@10:                 dst[j] = p[1]->value << 6 | p[0]->value;
yading@10:                 p[0] = &c->paths[0][p[0]->prev];
yading@10:                 p[1] = &c->paths[1][p[1]->prev];
yading@10:             }
yading@10:             froze = i;
yading@10:             pathn[0] = pathn[1] = 0;
yading@10:             memset(nodes[0] + 1, 0, (frontier - 1)*sizeof(**nodes));
yading@10:             memset(nodes[1] + 1, 0, (frontier - 1)*sizeof(**nodes));
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     p[0] = &c->paths[0][nodes[0][0]->path];
yading@10:     p[1] = &c->paths[1][nodes[1][0]->path];
yading@10:     for (j = i; j > froze; j--) {
yading@10:         dst[j] = p[1]->value << 6 | p[0]->value;
yading@10:         p[0] = &c->paths[0][p[0]->prev];
yading@10:         p[1] = &c->paths[1][p[1]->prev];
yading@10:     }
yading@10:     c->band[0] = nodes[0][0]->state;
yading@10:     c->band[1] = nodes[1][0]->state;
yading@10: }
yading@10: 
yading@10: static av_always_inline void encode_byte(G722Context *c, uint8_t *dst,
yading@10:                                          const int16_t *samples)
yading@10: {
yading@10:     int xlow, xhigh, ilow, ihigh;
yading@10:     filter_samples(c, samples, &xlow, &xhigh);
yading@10:     ihigh = encode_high(&c->band[1], xhigh);
yading@10:     ilow  = encode_low (&c->band[0], xlow);
yading@10:     ff_g722_update_high_predictor(&c->band[1], c->band[1].scale_factor *
yading@10:                                 ff_g722_high_inv_quant[ihigh] >> 10, ihigh);
yading@10:     ff_g722_update_low_predictor(&c->band[0], ilow >> 2);
yading@10:     *dst = ihigh << 6 | ilow;
yading@10: }
yading@10: 
yading@10: static void g722_encode_no_trellis(G722Context *c,
yading@10:                                    uint8_t *dst, int nb_samples,
yading@10:                                    const int16_t *samples)
yading@10: {
yading@10:     int i;
yading@10:     for (i = 0; i < nb_samples; i += 2)
yading@10:         encode_byte(c, dst++, &samples[i]);
yading@10: }
yading@10: 
yading@10: static int g722_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
yading@10:                              const AVFrame *frame, int *got_packet_ptr)
yading@10: {
yading@10:     G722Context *c = avctx->priv_data;
yading@10:     const int16_t *samples = (const int16_t *)frame->data[0];
yading@10:     int nb_samples, out_size, ret;
yading@10: 
yading@10:     out_size = (frame->nb_samples + 1) / 2;
yading@10:     if ((ret = ff_alloc_packet2(avctx, avpkt, out_size)) < 0)
yading@10:         return ret;
yading@10: 
yading@10:     nb_samples = frame->nb_samples - (frame->nb_samples & 1);
yading@10: 
yading@10:     if (avctx->trellis)
yading@10:         g722_encode_trellis(c, avctx->trellis, avpkt->data, nb_samples, samples);
yading@10:     else
yading@10:         g722_encode_no_trellis(c, avpkt->data, nb_samples, samples);
yading@10: 
yading@10:     /* handle last frame with odd frame_size */
yading@10:     if (nb_samples < frame->nb_samples) {
yading@10:         int16_t last_samples[2] = { samples[nb_samples], samples[nb_samples] };
yading@10:         encode_byte(c, &avpkt->data[nb_samples >> 1], last_samples);
yading@10:     }
yading@10: 
yading@10:     if (frame->pts != AV_NOPTS_VALUE)
yading@10:         avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->delay);
yading@10:     *got_packet_ptr = 1;
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: AVCodec ff_adpcm_g722_encoder = {
yading@10:     .name           = "g722",
yading@10:     .type           = AVMEDIA_TYPE_AUDIO,
yading@10:     .id             = AV_CODEC_ID_ADPCM_G722,
yading@10:     .priv_data_size = sizeof(G722Context),
yading@10:     .init           = g722_encode_init,
yading@10:     .close          = g722_encode_close,
yading@10:     .encode2        = g722_encode_frame,
yading@10:     .capabilities   = CODEC_CAP_SMALL_LAST_FRAME,
yading@10:     .long_name      = NULL_IF_CONFIG_SMALL("G.722 ADPCM"),
yading@10:     .sample_fmts    = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
yading@10:                                                      AV_SAMPLE_FMT_NONE },
yading@10: };