libavcodec/opus_celt.c
b70d7a4a
 /*
  * Copyright (c) 2012 Andrew D'Addesio
  * Copyright (c) 2013-2014 Mozilla Corporation
07b78340
  * Copyright (c) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
b70d7a4a
  *
2c7d3ecf
  * This file is part of FFmpeg.
b70d7a4a
  *
2c7d3ecf
  * FFmpeg is free software; you can redistribute it and/or
b70d7a4a
  * modify it under the terms of the GNU Lesser General Public
  * License as published by the Free Software Foundation; either
  * version 2.1 of the License, or (at your option) any later version.
  *
2c7d3ecf
  * FFmpeg is distributed in the hope that it will be useful,
b70d7a4a
  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  * Lesser General Public License for more details.
  *
  * You should have received a copy of the GNU Lesser General Public
2c7d3ecf
  * License along with FFmpeg; if not, write to the Free Software
b70d7a4a
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
 /**
  * @file
  * Opus CELT decoder
  */
 
e538108c
 #include "opus_celt.h"
0660a09d
 #include "opustab.h"
e538108c
 #include "opus_pvq.h"
b70d7a4a
 
79450adf
 /* Use the 2D z-transform to apply prediction in both the time domain (alpha)
  * and the frequency domain (beta) */
07b78340
 static void celt_decode_coarse_energy(CeltFrame *f, OpusRangeCoder *rc)
b70d7a4a
 {
     int i, j;
79450adf
     float prev[2] = { 0 };
     float alpha = ff_celt_alpha_coef[f->size];
     float beta  = ff_celt_beta_coef[f->size];
     const uint8_t *model = ff_celt_coarse_energy_dist[f->size][0];
b70d7a4a
 
79450adf
     /* intra frame */
     if (opus_rc_tell(rc) + 3 <= f->framebits && ff_opus_rc_dec_log(rc, 3)) {
         alpha = 0.0f;
         beta  = 1.0f - (4915.0f/32768.0f);
07b78340
         model = ff_celt_coarse_energy_dist[f->size][1];
b70d7a4a
     }
 
     for (i = 0; i < CELT_MAX_BANDS; i++) {
07b78340
         for (j = 0; j < f->channels; j++) {
             CeltBlock *block = &f->block[j];
b70d7a4a
             float value;
             int available;
 
07b78340
             if (i < f->start_band || i >= f->end_band) {
                 block->energy[i] = 0.0;
b70d7a4a
                 continue;
             }
 
07b78340
             available = f->framebits - opus_rc_tell(rc);
b70d7a4a
             if (available >= 15) {
                 /* decode using a Laplace distribution */
                 int k = FFMIN(i, 20) << 1;
317be31e
                 value = ff_opus_rc_dec_laplace(rc, model[k] << 7, model[k+1] << 6);
b70d7a4a
             } else if (available >= 2) {
317be31e
                 int x = ff_opus_rc_dec_cdf(rc, ff_celt_model_energy_small);
b70d7a4a
                 value = (x>>1) ^ -(x&1);
             } else if (available >= 1) {
317be31e
                 value = -(float)ff_opus_rc_dec_log(rc, 1);
b70d7a4a
             } else value = -1;
 
07b78340
             block->energy[i] = FFMAX(-9.0f, block->energy[i]) * alpha + prev[j] + value;
b70d7a4a
             prev[j] += beta * value;
         }
     }
 }
 
07b78340
 static void celt_decode_fine_energy(CeltFrame *f, OpusRangeCoder *rc)
b70d7a4a
 {
     int i;
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
b70d7a4a
         int j;
07b78340
         if (!f->fine_bits[i])
b70d7a4a
             continue;
 
07b78340
         for (j = 0; j < f->channels; j++) {
             CeltBlock *block = &f->block[j];
b70d7a4a
             int q2;
             float offset;
07b78340
             q2 = ff_opus_rc_get_raw(rc, f->fine_bits[i]);
             offset = (q2 + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f - 0.5f;
             block->energy[i] += offset;
b70d7a4a
         }
     }
 }
 
07b78340
 static void celt_decode_final_energy(CeltFrame *f, OpusRangeCoder *rc)
b70d7a4a
 {
     int priority, i, j;
07b78340
     int bits_left = f->framebits - opus_rc_tell(rc);
b70d7a4a
 
     for (priority = 0; priority < 2; priority++) {
07b78340
         for (i = f->start_band; i < f->end_band && bits_left >= f->channels; i++) {
             if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
b70d7a4a
                 continue;
 
07b78340
             for (j = 0; j < f->channels; j++) {
b70d7a4a
                 int q2;
                 float offset;
317be31e
                 q2 = ff_opus_rc_get_raw(rc, 1);
07b78340
                 offset = (q2 - 0.5f) * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
                 f->block[j].energy[i] += offset;
b70d7a4a
                 bits_left--;
             }
         }
     }
 }
 
07b78340
 static void celt_decode_tf_changes(CeltFrame *f, OpusRangeCoder *rc)
b70d7a4a
 {
     int i, diff = 0, tf_select = 0, tf_changed = 0, tf_select_bit;
07b78340
     int consumed, bits = f->transient ? 2 : 4;
b70d7a4a
 
     consumed = opus_rc_tell(rc);
07b78340
     tf_select_bit = (f->size != 0 && consumed+bits+1 <= f->framebits);
b70d7a4a
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
         if (consumed+bits+tf_select_bit <= f->framebits) {
317be31e
             diff ^= ff_opus_rc_dec_log(rc, bits);
b70d7a4a
             consumed = opus_rc_tell(rc);
             tf_changed |= diff;
         }
07b78340
         f->tf_change[i] = diff;
         bits = f->transient ? 4 : 5;
b70d7a4a
     }
 
07b78340
     if (tf_select_bit && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
                          ff_celt_tf_select[f->size][f->transient][1][tf_changed])
317be31e
         tf_select = ff_opus_rc_dec_log(rc, 1);
b70d7a4a
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
         f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
b70d7a4a
     }
 }
 
07b78340
 static void celt_decode_allocation(CeltFrame *f, OpusRangeCoder *rc)
b70d7a4a
 {
     // approx. maximum bit allocation for each band before boost/trim
     int cap[CELT_MAX_BANDS];
     int boost[CELT_MAX_BANDS];
     int threshold[CELT_MAX_BANDS];
     int bits1[CELT_MAX_BANDS];
     int bits2[CELT_MAX_BANDS];
     int trim_offset[CELT_MAX_BANDS];
 
07b78340
     int skip_start_band = f->start_band;
b70d7a4a
     int dynalloc       = 6;
     int alloctrim      = 5;
     int extrabits      = 0;
 
07b78340
     int skip_bit             = 0;
     int intensity_stereo_bit = 0;
     int dual_stereo_bit      = 0;
b70d7a4a
 
     int remaining, bandbits;
     int low, high, total, done;
     int totalbits;
     int consumed;
     int i, j;
 
     consumed = opus_rc_tell(rc);
 
     /* obtain spread flag */
07b78340
     f->spread = CELT_SPREAD_NORMAL;
     if (consumed + 4 <= f->framebits)
         f->spread = ff_opus_rc_dec_cdf(rc, ff_celt_model_spread);
b70d7a4a
 
     /* generate static allocation caps */
     for (i = 0; i < CELT_MAX_BANDS; i++) {
07b78340
         cap[i] = (ff_celt_static_caps[f->size][f->channels - 1][i] + 64)
                  * ff_celt_freq_range[i] << (f->channels - 1) << f->size >> 2;
b70d7a4a
     }
 
     /* obtain band boost */
07b78340
     totalbits = f->framebits << 3; // convert to 1/8 bits
b70d7a4a
     consumed = opus_rc_tell_frac(rc);
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
b70d7a4a
         int quanta, band_dynalloc;
 
         boost[i] = 0;
 
07b78340
         quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size;
b70d7a4a
         quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta));
         band_dynalloc = dynalloc;
         while (consumed + (band_dynalloc<<3) < totalbits && boost[i] < cap[i]) {
317be31e
             int add = ff_opus_rc_dec_log(rc, band_dynalloc);
b70d7a4a
             consumed = opus_rc_tell_frac(rc);
             if (!add)
                 break;
 
             boost[i]     += quanta;
             totalbits    -= quanta;
             band_dynalloc = 1;
         }
         /* dynalloc is more likely to occur if it's already been used for earlier bands */
         if (boost[i])
             dynalloc = FFMAX(2, dynalloc - 1);
     }
 
     /* obtain allocation trim */
     if (consumed + (6 << 3) <= totalbits)
317be31e
         alloctrim = ff_opus_rc_dec_cdf(rc, ff_celt_model_alloc_trim);
b70d7a4a
 
     /* anti-collapse bit reservation */
07b78340
     totalbits = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1;
     f->anticollapse_needed = 0;
     if (f->blocks > 1 && f->size >= 2 &&
         totalbits >= ((f->size + 2) << 3))
         f->anticollapse_needed = 1 << 3;
     totalbits -= f->anticollapse_needed;
b70d7a4a
 
     /* band skip bit reservation */
     if (totalbits >= 1 << 3)
         skip_bit = 1 << 3;
     totalbits -= skip_bit;
 
     /* intensity/dual stereo bit reservation */
07b78340
     if (f->channels == 2) {
         intensity_stereo_bit = ff_celt_log2_frac[f->end_band - f->start_band];
         if (intensity_stereo_bit <= totalbits) {
             totalbits -= intensity_stereo_bit;
b70d7a4a
             if (totalbits >= 1 << 3) {
07b78340
                 dual_stereo_bit = 1 << 3;
b70d7a4a
                 totalbits -= 1 << 3;
             }
         } else
07b78340
             intensity_stereo_bit = 0;
b70d7a4a
     }
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
         int trim     = alloctrim - 5 - f->size;
         int band     = ff_celt_freq_range[i] * (f->end_band - i - 1);
         int duration = f->size + 3;
         int scale    = duration + f->channels - 1;
b70d7a4a
 
         /* PVQ minimum allocation threshold, below this value the band is
          * skipped */
0660a09d
         threshold[i] = FFMAX(3 * ff_celt_freq_range[i] << duration >> 4,
07b78340
                              f->channels << 3);
b70d7a4a
 
         trim_offset[i] = trim * (band << scale) >> 6;
 
07b78340
         if (ff_celt_freq_range[i] << f->size == 1)
             trim_offset[i] -= f->channels << 3;
b70d7a4a
     }
 
     /* bisection */
     low  = 1;
     high = CELT_VECTORS - 1;
     while (low <= high) {
         int center = (low + high) >> 1;
         done = total = 0;
 
07b78340
         for (i = f->end_band - 1; i >= f->start_band; i--) {
0660a09d
             bandbits = ff_celt_freq_range[i] * ff_celt_static_alloc[center][i]
07b78340
                        << (f->channels - 1) << f->size >> 2;
b70d7a4a
 
             if (bandbits)
                 bandbits = FFMAX(0, bandbits + trim_offset[i]);
             bandbits += boost[i];
 
             if (bandbits >= threshold[i] || done) {
                 done = 1;
                 total += FFMIN(bandbits, cap[i]);
07b78340
             } else if (bandbits >= f->channels << 3)
                 total += f->channels << 3;
b70d7a4a
         }
 
         if (total > totalbits)
             high = center - 1;
         else
             low = center + 1;
     }
     high = low--;
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
0660a09d
         bits1[i] = ff_celt_freq_range[i] * ff_celt_static_alloc[low][i]
07b78340
                    << (f->channels - 1) << f->size >> 2;
b70d7a4a
         bits2[i] = high >= CELT_VECTORS ? cap[i] :
0660a09d
                    ff_celt_freq_range[i] * ff_celt_static_alloc[high][i]
07b78340
                    << (f->channels - 1) << f->size >> 2;
b70d7a4a
 
         if (bits1[i])
             bits1[i] = FFMAX(0, bits1[i] + trim_offset[i]);
         if (bits2[i])
             bits2[i] = FFMAX(0, bits2[i] + trim_offset[i]);
         if (low)
             bits1[i] += boost[i];
         bits2[i] += boost[i];
 
         if (boost[i])
07b78340
             skip_start_band = i;
b70d7a4a
         bits2[i] = FFMAX(0, bits2[i] - bits1[i]);
     }
 
     /* bisection */
     low  = 0;
     high = 1 << CELT_ALLOC_STEPS;
     for (i = 0; i < CELT_ALLOC_STEPS; i++) {
         int center = (low + high) >> 1;
         done = total = 0;
 
07b78340
         for (j = f->end_band - 1; j >= f->start_band; j--) {
b70d7a4a
             bandbits = bits1[j] + (center * bits2[j] >> CELT_ALLOC_STEPS);
 
             if (bandbits >= threshold[j] || done) {
                 done = 1;
                 total += FFMIN(bandbits, cap[j]);
07b78340
             } else if (bandbits >= f->channels << 3)
                 total += f->channels << 3;
b70d7a4a
         }
         if (total > totalbits)
             high = center;
         else
             low = center;
     }
 
     done = total = 0;
07b78340
     for (i = f->end_band - 1; i >= f->start_band; i--) {
b70d7a4a
         bandbits = bits1[i] + (low * bits2[i] >> CELT_ALLOC_STEPS);
 
         if (bandbits >= threshold[i] || done)
             done = 1;
         else
07b78340
             bandbits = (bandbits >= f->channels << 3) ?
                        f->channels << 3 : 0;
b70d7a4a
 
         bandbits     = FFMIN(bandbits, cap[i]);
07b78340
         f->pulses[i] = bandbits;
b70d7a4a
         total      += bandbits;
     }
 
     /* band skipping */
07b78340
     for (f->coded_bands = f->end_band; ; f->coded_bands--) {
b70d7a4a
         int allocation;
07b78340
         j = f->coded_bands - 1;
b70d7a4a
 
07b78340
         if (j == skip_start_band) {
b70d7a4a
             /* all remaining bands are not skipped */
             totalbits += skip_bit;
             break;
         }
 
         /* determine the number of bits available for coding "do not skip" markers */
         remaining   = totalbits - total;
07b78340
         bandbits    = remaining / (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]);
         remaining  -= bandbits  * (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]);
         allocation  = f->pulses[j] + bandbits * ff_celt_freq_range[j]
                       + FFMAX(0, remaining - (ff_celt_freq_bands[j] - ff_celt_freq_bands[f->start_band]));
b70d7a4a
 
         /* a "do not skip" marker is only coded if the allocation is
            above the chosen threshold */
07b78340
         if (allocation >= FFMAX(threshold[j], (f->channels + 1) <<3 )) {
317be31e
             if (ff_opus_rc_dec_log(rc, 1))
b70d7a4a
                 break;
 
             total      += 1 << 3;
             allocation -= 1 << 3;
         }
 
         /* the band is skipped, so reclaim its bits */
07b78340
         total -= f->pulses[j];
         if (intensity_stereo_bit) {
             total -= intensity_stereo_bit;
             intensity_stereo_bit = ff_celt_log2_frac[j - f->start_band];
             total += intensity_stereo_bit;
b70d7a4a
         }
 
07b78340
         total += f->pulses[j] = (allocation >= f->channels << 3) ?
                               f->channels << 3 : 0;
b70d7a4a
     }
 
     /* obtain stereo flags */
07b78340
     f->intensity_stereo = 0;
     f->dual_stereo      = 0;
     if (intensity_stereo_bit)
         f->intensity_stereo = f->start_band +
                           ff_opus_rc_dec_uint(rc, f->coded_bands + 1 - f->start_band);
     if (f->intensity_stereo <= f->start_band)
         totalbits += dual_stereo_bit; /* no intensity stereo means no dual stereo */
     else if (dual_stereo_bit)
         f->dual_stereo = ff_opus_rc_dec_log(rc, 1);
b70d7a4a
 
     /* supply the remaining bits in this frame to lower bands */
     remaining = totalbits - total;
07b78340
     bandbits  = remaining / (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]);
     remaining -= bandbits * (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]);
     for (i = f->start_band; i < f->coded_bands; i++) {
0660a09d
         int bits = FFMIN(remaining, ff_celt_freq_range[i]);
b70d7a4a
 
07b78340
         f->pulses[i] += bits + bandbits * ff_celt_freq_range[i];
b70d7a4a
         remaining    -= bits;
     }
 
07b78340
     for (i = f->start_band; i < f->coded_bands; i++) {
         int N = ff_celt_freq_range[i] << f->size;
b70d7a4a
         int prev_extra = extrabits;
07b78340
         f->pulses[i] += extrabits;
b70d7a4a
 
         if (N > 1) {
             int dof;        // degrees of freedom
             int temp;       // dof * channels * log(dof)
             int offset;     // fine energy quantization offset, i.e.
                             // extra bits assigned over the standard
                             // totalbits/dof
             int fine_bits, max_bits;
 
07b78340
             extrabits = FFMAX(0, f->pulses[i] - cap[i]);
             f->pulses[i] -= extrabits;
b70d7a4a
 
             /* intensity stereo makes use of an extra degree of freedom */
07b78340
             dof = N * f->channels
                   + (f->channels == 2 && N > 2 && !f->dual_stereo && i < f->intensity_stereo);
             temp = dof * (ff_celt_log_freq_range[i] + (f->size<<3));
b70d7a4a
             offset = (temp >> 1) - dof * CELT_FINE_OFFSET;
             if (N == 2) /* dof=2 is the only case that doesn't fit the model */
                 offset += dof<<1;
 
             /* grant an additional bias for the first and second pulses */
07b78340
             if (f->pulses[i] + offset < 2 * (dof << 3))
b70d7a4a
                 offset += temp >> 2;
07b78340
             else if (f->pulses[i] + offset < 3 * (dof << 3))
b70d7a4a
                 offset += temp >> 3;
 
07b78340
             fine_bits = (f->pulses[i] + offset + (dof << 2)) / (dof << 3);
             max_bits  = FFMIN((f->pulses[i]>>3) >> (f->channels - 1),
b70d7a4a
                               CELT_MAX_FINE_BITS);
 
             max_bits  = FFMAX(max_bits, 0);
 
07b78340
             f->fine_bits[i] = av_clip(fine_bits, 0, max_bits);
b70d7a4a
 
             /* if fine_bits was rounded down or capped,
                give priority for the final fine energy pass */
07b78340
             f->fine_priority[i] = (f->fine_bits[i] * (dof<<3) >= f->pulses[i] + offset);
b70d7a4a
 
             /* the remaining bits are assigned to PVQ */
07b78340
             f->pulses[i] -= f->fine_bits[i] << (f->channels - 1) << 3;
b70d7a4a
         } else {
             /* all bits go to fine energy except for the sign bit */
07b78340
             extrabits = FFMAX(0, f->pulses[i] - (f->channels << 3));
             f->pulses[i] -= extrabits;
             f->fine_bits[i] = 0;
             f->fine_priority[i] = 1;
b70d7a4a
         }
 
         /* hand back a limited number of extra fine energy bits to this band */
         if (extrabits > 0) {
07b78340
             int fineextra = FFMIN(extrabits >> (f->channels + 2),
                                   CELT_MAX_FINE_BITS - f->fine_bits[i]);
             f->fine_bits[i] += fineextra;
b70d7a4a
 
07b78340
             fineextra <<= f->channels + 2;
             f->fine_priority[i] = (fineextra >= extrabits - prev_extra);
b70d7a4a
             extrabits -= fineextra;
         }
     }
07b78340
     f->remaining = extrabits;
b70d7a4a
 
     /* skipped bands dedicate all of their bits for fine energy */
07b78340
     for (; i < f->end_band; i++) {
         f->fine_bits[i]     = f->pulses[i] >> (f->channels - 1) >> 3;
         f->pulses[i]        = 0;
         f->fine_priority[i] = f->fine_bits[i] < 1;
b70d7a4a
     }
 }
 
07b78340
 static void celt_denormalize(CeltFrame *f, CeltBlock *block, float *data)
b70d7a4a
 {
     int i, j;
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
         float *dst = data + (ff_celt_freq_bands[i] << f->size);
7e98da9c
         float norm = exp2f(block->energy[i] + ff_celt_mean_energy[i]);
b70d7a4a
 
07b78340
         for (j = 0; j < ff_celt_freq_range[i] << f->size; j++)
b70d7a4a
             dst[j] *= norm;
     }
 }
 
07b78340
 static void celt_postfilter_apply_transition(CeltBlock *block, float *data)
b70d7a4a
 {
07b78340
     const int T0 = block->pf_period_old;
     const int T1 = block->pf_period;
b70d7a4a
 
     float g00, g01, g02;
     float g10, g11, g12;
 
     float x0, x1, x2, x3, x4;
 
     int i;
 
07b78340
     if (block->pf_gains[0]     == 0.0 &&
         block->pf_gains_old[0] == 0.0)
b70d7a4a
         return;
 
07b78340
     g00 = block->pf_gains_old[0];
     g01 = block->pf_gains_old[1];
     g02 = block->pf_gains_old[2];
     g10 = block->pf_gains[0];
     g11 = block->pf_gains[1];
     g12 = block->pf_gains[2];
b70d7a4a
 
     x1 = data[-T1 + 1];
     x2 = data[-T1];
     x3 = data[-T1 - 1];
     x4 = data[-T1 - 2];
 
     for (i = 0; i < CELT_OVERLAP; i++) {
         float w = ff_celt_window2[i];
         x0 = data[i - T1 + 2];
 
         data[i] +=  (1.0 - w) * g00 * data[i - T0]                          +
                     (1.0 - w) * g01 * (data[i - T0 - 1] + data[i - T0 + 1]) +
                     (1.0 - w) * g02 * (data[i - T0 - 2] + data[i - T0 + 2]) +
                     w         * g10 * x2                                    +
                     w         * g11 * (x1 + x3)                             +
                     w         * g12 * (x0 + x4);
         x4 = x3;
         x3 = x2;
         x2 = x1;
         x1 = x0;
     }
 }
 
07b78340
 static void celt_postfilter_apply(CeltBlock *block, float *data, int len)
b70d7a4a
 {
07b78340
     const int T = block->pf_period;
b70d7a4a
     float g0, g1, g2;
     float x0, x1, x2, x3, x4;
     int i;
 
07b78340
     if (block->pf_gains[0] == 0.0 || len <= 0)
b70d7a4a
         return;
 
07b78340
     g0 = block->pf_gains[0];
     g1 = block->pf_gains[1];
     g2 = block->pf_gains[2];
b70d7a4a
 
     x4 = data[-T - 2];
     x3 = data[-T - 1];
     x2 = data[-T];
     x1 = data[-T + 1];
 
     for (i = 0; i < len; i++) {
         x0 = data[i - T + 2];
         data[i] += g0 * x2        +
                    g1 * (x1 + x3) +
                    g2 * (x0 + x4);
         x4 = x3;
         x3 = x2;
         x2 = x1;
         x1 = x0;
     }
 }
 
07b78340
 static void celt_postfilter(CeltFrame *f, CeltBlock *block)
b70d7a4a
 {
07b78340
     int len = f->blocksize * f->blocks;
b70d7a4a
 
07b78340
     celt_postfilter_apply_transition(block, block->buf + 1024);
b70d7a4a
 
07b78340
     block->pf_period_old = block->pf_period;
     memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
b70d7a4a
 
07b78340
     block->pf_period = block->pf_period_new;
     memcpy(block->pf_gains, block->pf_gains_new, sizeof(block->pf_gains));
b70d7a4a
 
     if (len > CELT_OVERLAP) {
07b78340
         celt_postfilter_apply_transition(block, block->buf + 1024 + CELT_OVERLAP);
         celt_postfilter_apply(block, block->buf + 1024 + 2 * CELT_OVERLAP,
b70d7a4a
                               len - 2 * CELT_OVERLAP);
 
07b78340
         block->pf_period_old = block->pf_period;
         memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
b70d7a4a
     }
 
07b78340
     memmove(block->buf, block->buf + len, (1024 + CELT_OVERLAP / 2) * sizeof(float));
b70d7a4a
 }
 
07b78340
 static int parse_postfilter(CeltFrame *f, OpusRangeCoder *rc, int consumed)
b70d7a4a
 {
     int i;
 
07b78340
     memset(f->block[0].pf_gains_new, 0, sizeof(f->block[0].pf_gains_new));
     memset(f->block[1].pf_gains_new, 0, sizeof(f->block[1].pf_gains_new));
b70d7a4a
 
07b78340
     if (f->start_band == 0 && consumed + 16 <= f->framebits) {
317be31e
         int has_postfilter = ff_opus_rc_dec_log(rc, 1);
b70d7a4a
         if (has_postfilter) {
             float gain;
             int tapset, octave, period;
 
317be31e
             octave = ff_opus_rc_dec_uint(rc, 6);
             period = (16 << octave) + ff_opus_rc_get_raw(rc, 4 + octave) - 1;
             gain   = 0.09375f * (ff_opus_rc_get_raw(rc, 3) + 1);
07b78340
             tapset = (opus_rc_tell(rc) + 2 <= f->framebits) ?
317be31e
                      ff_opus_rc_dec_cdf(rc, ff_celt_model_tapset) : 0;
b70d7a4a
 
             for (i = 0; i < 2; i++) {
07b78340
                 CeltBlock *block = &f->block[i];
b70d7a4a
 
07b78340
                 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
ba67c234
                 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
                 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
                 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
b70d7a4a
             }
         }
 
         consumed = opus_rc_tell(rc);
     }
 
     return consumed;
 }
 
07b78340
 static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X)
b70d7a4a
 {
     int i, j, k;
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
b70d7a4a
         int renormalize = 0;
         float *xptr;
         float prev[2];
         float Ediff, r;
         float thresh, sqrt_1;
         int depth;
 
         /* depth in 1/8 bits */
07b78340
         depth = (1 + f->pulses[i]) / (ff_celt_freq_range[i] << f->size);
cea529dd
         thresh = exp2f(-1.0 - 0.125f * depth);
07b78340
         sqrt_1 = 1.0f / sqrtf(ff_celt_freq_range[i] << f->size);
b70d7a4a
 
07b78340
         xptr = X + (ff_celt_freq_bands[i] << f->size);
b70d7a4a
 
07b78340
         prev[0] = block->prev_energy[0][i];
         prev[1] = block->prev_energy[1][i];
         if (f->channels == 1) {
             CeltBlock *block1 = &f->block[1];
b70d7a4a
 
07b78340
             prev[0] = FFMAX(prev[0], block1->prev_energy[0][i]);
             prev[1] = FFMAX(prev[1], block1->prev_energy[1][i]);
b70d7a4a
         }
07b78340
         Ediff = block->energy[i] - FFMIN(prev[0], prev[1]);
b70d7a4a
         Ediff = FFMAX(0, Ediff);
 
         /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
         short blocks don't have the same energy as long */
7e98da9c
         r = exp2f(1 - Ediff);
07b78340
         if (f->size == 3)
b70d7a4a
             r *= M_SQRT2;
         r = FFMIN(thresh, r) * sqrt_1;
07b78340
         for (k = 0; k < 1 << f->size; k++) {
b70d7a4a
             /* Detect collapse */
07b78340
             if (!(block->collapse_masks[i] & 1 << k)) {
b70d7a4a
                 /* Fill with noise */
0660a09d
                 for (j = 0; j < ff_celt_freq_range[i]; j++)
07b78340
                     xptr[(j << f->size) + k] = (celt_rng(f) & 0x8000) ? r : -r;
b70d7a4a
                 renormalize = 1;
             }
         }
 
         /* We just added some energy, so we need to renormalize */
         if (renormalize)
07b78340
             celt_renormalize_vector(xptr, ff_celt_freq_range[i] << f->size, 1.0f);
b70d7a4a
     }
 }
 
07b78340
 static void celt_decode_bands(CeltFrame *f, OpusRangeCoder *rc)
b70d7a4a
 {
     float lowband_scratch[8 * 22];
     float norm[2 * 8 * 100];
 
07b78340
     int totalbits = (f->framebits << 3) - f->anticollapse_needed;
b70d7a4a
 
     int update_lowband = 1;
     int lowband_offset = 0;
 
     int i, j;
 
07b78340
     memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs));
     memset(f->block[1].coeffs, 0, sizeof(f->block[0].coeffs));
b70d7a4a
 
07b78340
     for (i = f->start_band; i < f->end_band; i++) {
6d0b62a1
         uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 };
07b78340
         int band_offset = ff_celt_freq_bands[i] << f->size;
         int band_size   = ff_celt_freq_range[i] << f->size;
         float *X = f->block[0].coeffs + band_offset;
         float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL;
b70d7a4a
 
         int consumed = opus_rc_tell_frac(rc);
         float *norm2 = norm + 8 * 100;
         int effective_lowband = -1;
6d0b62a1
         int b = 0;
b70d7a4a
 
         /* Compute how many bits we want to allocate to this band */
07b78340
         if (i != f->start_band)
             f->remaining -= consumed;
         f->remaining2 = totalbits - consumed - 1;
         if (i <= f->coded_bands - 1) {
             int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i);
             b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14);
6d0b62a1
         }
b70d7a4a
 
07b78340
         if (ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] &&
b70d7a4a
             (update_lowband || lowband_offset == 0))
             lowband_offset = i;
 
         /* Get a conservative estimate of the collapse_mask's for the bands we're
6d0b62a1
            going to be folding from. */
07b78340
         if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE ||
                                     f->blocks > 1 || f->tf_change[i] < 0)) {
b70d7a4a
             int foldstart, foldend;
 
             /* This ensures we never repeat spectral content within one band */
07b78340
             effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band],
0660a09d
                                       ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]);
b70d7a4a
             foldstart = lowband_offset;
0660a09d
             while (ff_celt_freq_bands[--foldstart] > effective_lowband);
b70d7a4a
             foldend = lowband_offset - 1;
0660a09d
             while (ff_celt_freq_bands[++foldend] < effective_lowband + ff_celt_freq_range[i]);
b70d7a4a
 
             cm[0] = cm[1] = 0;
             for (j = foldstart; j < foldend; j++) {
07b78340
                 cm[0] |= f->block[0].collapse_masks[j];
                 cm[1] |= f->block[f->channels - 1].collapse_masks[j];
b70d7a4a
             }
6d0b62a1
         }
b70d7a4a
 
07b78340
         if (f->dual_stereo && i == f->intensity_stereo) {
b70d7a4a
             /* Switch off dual stereo to do intensity */
07b78340
             f->dual_stereo = 0;
             for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++)
b70d7a4a
                 norm[j] = (norm[j] + norm2[j]) / 2;
         }
 
07b78340
         if (f->dual_stereo) {
8e7e74df
             cm[0] = f->pvq->decode_band(f->pvq, f, rc, i, X, NULL, band_size, b / 2, f->blocks,
07b78340
                                         effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size,
e538108c
                                         norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]);
b70d7a4a
 
8e7e74df
             cm[1] = f->pvq->decode_band(f->pvq, f, rc, i, Y, NULL, band_size, b/2, f->blocks,
07b78340
                                         effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL, f->size,
e538108c
                                         norm2 + band_offset, 0, 1.0f, lowband_scratch, cm[1]);
b70d7a4a
         } else {
8e7e74df
             cm[0] = f->pvq->decode_band(f->pvq, f, rc, i, X, Y, band_size, b, f->blocks,
07b78340
                                         effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size,
e538108c
                                         norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]|cm[1]);
b70d7a4a
             cm[1] = cm[0];
         }
 
07b78340
         f->block[0].collapse_masks[i]               = (uint8_t)cm[0];
         f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1];
         f->remaining += f->pulses[i] + consumed;
b70d7a4a
 
         /* Update the folding position only as long as we have 1 bit/sample depth */
         update_lowband = (b > band_size << 3);
     }
 }
 
07b78340
 int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc,
                          float **output, int channels, int frame_size,
                          int start_band,  int end_band)
b70d7a4a
 {
aef5f9ab
     int i, j, downmix = 0;
b70d7a4a
     int consumed;           // bits of entropy consumed thus far for this frame
d2119f62
     MDCT15Context *imdct;
b70d7a4a
 
07b78340
     if (channels != 1 && channels != 2) {
         av_log(f->avctx, AV_LOG_ERROR, "Invalid number of coded channels: %d\n",
                channels);
b70d7a4a
         return AVERROR_INVALIDDATA;
     }
07b78340
     if (start_band < 0 || start_band > end_band || end_band > CELT_MAX_BANDS) {
         av_log(f->avctx, AV_LOG_ERROR, "Invalid start/end band: %d %d\n",
                start_band, end_band);
b70d7a4a
         return AVERROR_INVALIDDATA;
     }
 
07b78340
     f->silence        = 0;
     f->transient      = 0;
     f->anticollapse   = 0;
     f->flushed        = 0;
     f->channels       = channels;
     f->start_band     = start_band;
     f->end_band       = end_band;
     f->framebits      = rc->rb.bytes * 8;
 
     f->size = av_log2(frame_size / CELT_SHORT_BLOCKSIZE);
     if (f->size > CELT_MAX_LOG_BLOCKS ||
         frame_size != CELT_SHORT_BLOCKSIZE * (1 << f->size)) {
         av_log(f->avctx, AV_LOG_ERROR, "Invalid CELT frame size: %d\n",
b70d7a4a
                frame_size);
         return AVERROR_INVALIDDATA;
     }
 
07b78340
     if (!f->output_channels)
         f->output_channels = channels;
b70d7a4a
 
07b78340
     memset(f->block[0].collapse_masks, 0, sizeof(f->block[0].collapse_masks));
     memset(f->block[1].collapse_masks, 0, sizeof(f->block[1].collapse_masks));
b70d7a4a
 
     consumed = opus_rc_tell(rc);
 
     /* obtain silence flag */
07b78340
     if (consumed >= f->framebits)
         f->silence = 1;
b70d7a4a
     else if (consumed == 1)
07b78340
         f->silence = ff_opus_rc_dec_log(rc, 15);
b70d7a4a
 
 
07b78340
     if (f->silence) {
         consumed = f->framebits;
         rc->total_bits += f->framebits - opus_rc_tell(rc);
b70d7a4a
     }
 
     /* obtain post-filter options */
07b78340
     consumed = parse_postfilter(f, rc, consumed);
b70d7a4a
 
     /* obtain transient flag */
07b78340
     if (f->size != 0 && consumed+3 <= f->framebits)
         f->transient = ff_opus_rc_dec_log(rc, 3);
b70d7a4a
 
07b78340
     f->blocks    = f->transient ? 1 << f->size : 1;
     f->blocksize = frame_size / f->blocks;
b70d7a4a
 
07b78340
     imdct = f->imdct[f->transient ? 0 : f->size];
b70d7a4a
 
07b78340
     if (channels == 1) {
b70d7a4a
         for (i = 0; i < CELT_MAX_BANDS; i++)
07b78340
             f->block[0].energy[i] = FFMAX(f->block[0].energy[i], f->block[1].energy[i]);
b70d7a4a
     }
 
07b78340
     celt_decode_coarse_energy(f, rc);
     celt_decode_tf_changes   (f, rc);
     celt_decode_allocation   (f, rc);
     celt_decode_fine_energy  (f, rc);
     celt_decode_bands        (f, rc);
b70d7a4a
 
07b78340
     if (f->anticollapse_needed)
         f->anticollapse = ff_opus_rc_get_raw(rc, 1);
b70d7a4a
 
07b78340
     celt_decode_final_energy(f, rc);
b70d7a4a
 
     /* apply anti-collapse processing and denormalization to
      * each coded channel */
07b78340
     for (i = 0; i < f->channels; i++) {
         CeltBlock *block = &f->block[i];
b70d7a4a
 
07b78340
         if (f->anticollapse)
             process_anticollapse(f, block, f->block[i].coeffs);
b70d7a4a
 
07b78340
         celt_denormalize(f, block, f->block[i].coeffs);
b70d7a4a
     }
 
     /* stereo -> mono downmix */
07b78340
     if (f->output_channels < f->channels) {
         f->dsp->vector_fmac_scalar(f->block[0].coeffs, f->block[1].coeffs, 1.0, FFALIGN(frame_size, 16));
aef5f9ab
         downmix = 1;
07b78340
     } else if (f->output_channels > f->channels)
         memcpy(f->block[1].coeffs, f->block[0].coeffs, frame_size * sizeof(float));
b70d7a4a
 
07b78340
     if (f->silence) {
b70d7a4a
         for (i = 0; i < 2; i++) {
07b78340
             CeltBlock *block = &f->block[i];
b70d7a4a
 
07b78340
             for (j = 0; j < FF_ARRAY_ELEMS(block->energy); j++)
                 block->energy[j] = CELT_ENERGY_SILENCE;
b70d7a4a
         }
07b78340
         memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs));
         memset(f->block[1].coeffs, 0, sizeof(f->block[1].coeffs));
b70d7a4a
     }
 
     /* transform and output for each output channel */
07b78340
     for (i = 0; i < f->output_channels; i++) {
         CeltBlock *block = &f->block[i];
         float m = block->emph_coeff;
b70d7a4a
 
         /* iMDCT and overlap-add */
07b78340
         for (j = 0; j < f->blocks; j++) {
             float *dst  = block->buf + 1024 + j * f->blocksize;
b70d7a4a
 
07b78340
             imdct->imdct_half(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
aef5f9ab
                               f->blocks);
07b78340
             f->dsp->vector_fmul_window(dst, dst, dst + CELT_OVERLAP / 2,
0660a09d
                                        ff_celt_window, CELT_OVERLAP / 2);
b70d7a4a
         }
 
aef5f9ab
         if (downmix)
             f->dsp->vector_fmul_scalar(&block->buf[1024], &block->buf[1024], 0.5f, frame_size);
 
b70d7a4a
         /* postfilter */
07b78340
         celt_postfilter(f, block);
b70d7a4a
 
         /* deemphasis and output scaling */
         for (j = 0; j < frame_size; j++) {
594cd1f3
             const float tmp = block->buf[1024 - frame_size + j] + m;
07b78340
             m = tmp * CELT_EMPH_COEFF;
594cd1f3
             output[i][j] = tmp;
b70d7a4a
         }
594cd1f3
 
07b78340
         block->emph_coeff = m;
b70d7a4a
     }
 
07b78340
     if (channels == 1)
         memcpy(f->block[1].energy, f->block[0].energy, sizeof(f->block[0].energy));
b70d7a4a
 
     for (i = 0; i < 2; i++ ) {
07b78340
         CeltBlock *block = &f->block[i];
b70d7a4a
 
07b78340
         if (!f->transient) {
             memcpy(block->prev_energy[1], block->prev_energy[0], sizeof(block->prev_energy[0]));
             memcpy(block->prev_energy[0], block->energy,         sizeof(block->prev_energy[0]));
b70d7a4a
         } else {
             for (j = 0; j < CELT_MAX_BANDS; j++)
07b78340
                 block->prev_energy[0][j] = FFMIN(block->prev_energy[0][j], block->energy[j]);
b70d7a4a
         }
 
07b78340
         for (j = 0; j < f->start_band; j++) {
             block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
             block->energy[j]         = 0.0;
b70d7a4a
         }
07b78340
         for (j = f->end_band; j < CELT_MAX_BANDS; j++) {
             block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
             block->energy[j]         = 0.0;
b70d7a4a
         }
     }
 
07b78340
     f->seed = rc->range;
b70d7a4a
 
     return 0;
 }
 
07b78340
 void ff_celt_flush(CeltFrame *f)
b70d7a4a
 {
     int i, j;
 
07b78340
     if (f->flushed)
b70d7a4a
         return;
 
     for (i = 0; i < 2; i++) {
07b78340
         CeltBlock *block = &f->block[i];
b70d7a4a
 
         for (j = 0; j < CELT_MAX_BANDS; j++)
07b78340
             block->prev_energy[0][j] = block->prev_energy[1][j] = CELT_ENERGY_SILENCE;
b70d7a4a
 
07b78340
         memset(block->energy, 0, sizeof(block->energy));
         memset(block->buf,    0, sizeof(block->buf));
b70d7a4a
 
07b78340
         memset(block->pf_gains,     0, sizeof(block->pf_gains));
         memset(block->pf_gains_old, 0, sizeof(block->pf_gains_old));
         memset(block->pf_gains_new, 0, sizeof(block->pf_gains_new));
b70d7a4a
 
07b78340
         block->emph_coeff = 0.0;
b70d7a4a
     }
07b78340
     f->seed = 0;
b70d7a4a
 
07b78340
     f->flushed = 1;
b70d7a4a
 }
 
07b78340
 void ff_celt_free(CeltFrame **f)
b70d7a4a
 {
07b78340
     CeltFrame *frm = *f;
b70d7a4a
     int i;
 
07b78340
     if (!frm)
b70d7a4a
         return;
 
07b78340
     for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
         ff_mdct15_uninit(&frm->imdct[i]);
b70d7a4a
 
8e7e74df
     ff_celt_pvq_uninit(&frm->pvq);
 
07b78340
     av_freep(&frm->dsp);
     av_freep(f);
b70d7a4a
 }
 
07b78340
 int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels)
b70d7a4a
 {
07b78340
     CeltFrame *frm;
b70d7a4a
     int i, ret;
 
     if (output_channels != 1 && output_channels != 2) {
         av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
                output_channels);
         return AVERROR(EINVAL);
     }
 
07b78340
     frm = av_mallocz(sizeof(*frm));
     if (!frm)
b70d7a4a
         return AVERROR(ENOMEM);
 
07b78340
     frm->avctx           = avctx;
     frm->output_channels = output_channels;
b70d7a4a
 
8e7e74df
     for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
594cd1f3
         if ((ret = ff_mdct15_init(&frm->imdct[i], 1, i + 3, -1.0f/32768)) < 0)
b70d7a4a
             goto fail;
8e7e74df
 
     if ((ret = ff_celt_pvq_init(&frm->pvq)) < 0)
         goto fail;
b70d7a4a
 
07b78340
     frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
     if (!frm->dsp) {
00d47591
         ret = AVERROR(ENOMEM);
         goto fail;
     }
b70d7a4a
 
07b78340
     ff_celt_flush(frm);
b70d7a4a
 
07b78340
     *f = frm;
b70d7a4a
 
     return 0;
 fail:
07b78340
     ff_celt_free(&frm);
b70d7a4a
     return ret;
 }