Home | History | Annotate | Download | only in celt
      1 /* Copyright (c) 2007-2008 CSIRO
      2    Copyright (c) 2007-2010 Xiph.Org Foundation
      3    Copyright (c) 2008 Gregory Maxwell
      4    Written by Jean-Marc Valin and Gregory Maxwell */
      5 /*
      6    Redistribution and use in source and binary forms, with or without
      7    modification, are permitted provided that the following conditions
      8    are met:
      9 
     10    - Redistributions of source code must retain the above copyright
     11    notice, this list of conditions and the following disclaimer.
     12 
     13    - Redistributions in binary form must reproduce the above copyright
     14    notice, this list of conditions and the following disclaimer in the
     15    documentation and/or other materials provided with the distribution.
     16 
     17    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
     18    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
     19    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
     20    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
     21    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
     22    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
     23    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
     24    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
     25    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
     26    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
     27    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     28 */
     29 
     30 #ifdef HAVE_CONFIG_H
     31 #include "config.h"
     32 #endif
     33 
     34 #define CELT_ENCODER_C
     35 
     36 #include "cpu_support.h"
     37 #include "os_support.h"
     38 #include "mdct.h"
     39 #include <math.h>
     40 #include "celt.h"
     41 #include "pitch.h"
     42 #include "bands.h"
     43 #include "modes.h"
     44 #include "entcode.h"
     45 #include "quant_bands.h"
     46 #include "rate.h"
     47 #include "stack_alloc.h"
     48 #include "mathops.h"
     49 #include "float_cast.h"
     50 #include <stdarg.h>
     51 #include "celt_lpc.h"
     52 #include "vq.h"
     53 
     54 
     55 /** Encoder state
     56  @brief Encoder state
     57  */
     58 struct OpusCustomEncoder {
     59    const OpusCustomMode *mode;     /**< Mode used by the encoder */
     60    int channels;
     61    int stream_channels;
     62 
     63    int force_intra;
     64    int clip;
     65    int disable_pf;
     66    int complexity;
     67    int upsample;
     68    int start, end;
     69 
     70    opus_int32 bitrate;
     71    int vbr;
     72    int signalling;
     73    int constrained_vbr;      /* If zero, VBR can do whatever it likes with the rate */
     74    int loss_rate;
     75    int lsb_depth;
     76    int lfe;
     77    int disable_inv;
     78    int arch;
     79 
     80    /* Everything beyond this point gets cleared on a reset */
     81 #define ENCODER_RESET_START rng
     82 
     83    opus_uint32 rng;
     84    int spread_decision;
     85    opus_val32 delayedIntra;
     86    int tonal_average;
     87    int lastCodedBands;
     88    int hf_average;
     89    int tapset_decision;
     90 
     91    int prefilter_period;
     92    opus_val16 prefilter_gain;
     93    int prefilter_tapset;
     94 #ifdef RESYNTH
     95    int prefilter_period_old;
     96    opus_val16 prefilter_gain_old;
     97    int prefilter_tapset_old;
     98 #endif
     99    int consec_transient;
    100    AnalysisInfo analysis;
    101    SILKInfo silk_info;
    102 
    103    opus_val32 preemph_memE[2];
    104    opus_val32 preemph_memD[2];
    105 
    106    /* VBR-related parameters */
    107    opus_int32 vbr_reservoir;
    108    opus_int32 vbr_drift;
    109    opus_int32 vbr_offset;
    110    opus_int32 vbr_count;
    111    opus_val32 overlap_max;
    112    opus_val16 stereo_saving;
    113    int intensity;
    114    opus_val16 *energy_mask;
    115    opus_val16 spec_avg;
    116 
    117 #ifdef RESYNTH
    118    /* +MAX_PERIOD/2 to make space for overlap */
    119    celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2];
    120 #endif
    121 
    122    celt_sig in_mem[1]; /* Size = channels*mode->overlap */
    123    /* celt_sig prefilter_mem[],  Size = channels*COMBFILTER_MAXPERIOD */
    124    /* opus_val16 oldBandE[],     Size = channels*mode->nbEBands */
    125    /* opus_val16 oldLogE[],      Size = channels*mode->nbEBands */
    126    /* opus_val16 oldLogE2[],     Size = channels*mode->nbEBands */
    127    /* opus_val16 energyError[],  Size = channels*mode->nbEBands */
    128 };
    129 
    130 int celt_encoder_get_size(int channels)
    131 {
    132    CELTMode *mode = opus_custom_mode_create(48000, 960, NULL);
    133    return opus_custom_encoder_get_size(mode, channels);
    134 }
    135 
    136 OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int channels)
    137 {
    138    int size = sizeof(struct CELTEncoder)
    139          + (channels*mode->overlap-1)*sizeof(celt_sig)    /* celt_sig in_mem[channels*mode->overlap]; */
    140          + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */
    141          + 4*channels*mode->nbEBands*sizeof(opus_val16);  /* opus_val16 oldBandE[channels*mode->nbEBands]; */
    142                                                           /* opus_val16 oldLogE[channels*mode->nbEBands]; */
    143                                                           /* opus_val16 oldLogE2[channels*mode->nbEBands]; */
    144                                                           /* opus_val16 energyError[channels*mode->nbEBands]; */
    145    return size;
    146 }
    147 
    148 #ifdef CUSTOM_MODES
    149 CELTEncoder *opus_custom_encoder_create(const CELTMode *mode, int channels, int *error)
    150 {
    151    int ret;
    152    CELTEncoder *st = (CELTEncoder *)opus_alloc(opus_custom_encoder_get_size(mode, channels));
    153    /* init will handle the NULL case */
    154    ret = opus_custom_encoder_init(st, mode, channels);
    155    if (ret != OPUS_OK)
    156    {
    157       opus_custom_encoder_destroy(st);
    158       st = NULL;
    159    }
    160    if (error)
    161       *error = ret;
    162    return st;
    163 }
    164 #endif /* CUSTOM_MODES */
    165 
    166 static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode,
    167                                          int channels, int arch)
    168 {
    169    if (channels < 0 || channels > 2)
    170       return OPUS_BAD_ARG;
    171 
    172    if (st==NULL || mode==NULL)
    173       return OPUS_ALLOC_FAIL;
    174 
    175    OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels));
    176 
    177    st->mode = mode;
    178    st->stream_channels = st->channels = channels;
    179 
    180    st->upsample = 1;
    181    st->start = 0;
    182    st->end = st->mode->effEBands;
    183    st->signalling = 1;
    184    st->arch = arch;
    185 
    186    st->constrained_vbr = 1;
    187    st->clip = 1;
    188 
    189    st->bitrate = OPUS_BITRATE_MAX;
    190    st->vbr = 0;
    191    st->force_intra  = 0;
    192    st->complexity = 5;
    193    st->lsb_depth=24;
    194 
    195    opus_custom_encoder_ctl(st, OPUS_RESET_STATE);
    196 
    197    return OPUS_OK;
    198 }
    199 
    200 #ifdef CUSTOM_MODES
    201 int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels)
    202 {
    203    return opus_custom_encoder_init_arch(st, mode, channels, opus_select_arch());
    204 }
    205 #endif
    206 
    207 int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels,
    208                       int arch)
    209 {
    210    int ret;
    211    ret = opus_custom_encoder_init_arch(st,
    212            opus_custom_mode_create(48000, 960, NULL), channels, arch);
    213    if (ret != OPUS_OK)
    214       return ret;
    215    st->upsample = resampling_factor(sampling_rate);
    216    return OPUS_OK;
    217 }
    218 
    219 #ifdef CUSTOM_MODES
    220 void opus_custom_encoder_destroy(CELTEncoder *st)
    221 {
    222    opus_free(st);
    223 }
    224 #endif /* CUSTOM_MODES */
    225 
    226 
    227 static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C,
    228                               opus_val16 *tf_estimate, int *tf_chan, int allow_weak_transients,
    229                               int *weak_transient)
    230 {
    231    int i;
    232    VARDECL(opus_val16, tmp);
    233    opus_val32 mem0,mem1;
    234    int is_transient = 0;
    235    opus_int32 mask_metric = 0;
    236    int c;
    237    opus_val16 tf_max;
    238    int len2;
    239    /* Forward masking: 6.7 dB/ms. */
    240 #ifdef FIXED_POINT
    241    int forward_shift = 4;
    242 #else
    243    opus_val16 forward_decay = QCONST16(.0625f,15);
    244 #endif
    245    /* Table of 6*64/x, trained on real data to minimize the average error */
    246    static const unsigned char inv_table[128] = {
    247          255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25,
    248           23, 22, 21, 20, 19, 18, 17, 16, 16, 15, 15, 14, 13, 13, 12, 12,
    249           12, 12, 11, 11, 11, 10, 10, 10,  9,  9,  9,  9,  9,  9,  8,  8,
    250            8,  8,  8,  7,  7,  7,  7,  7,  7,  6,  6,  6,  6,  6,  6,  6,
    251            6,  6,  6,  6,  6,  6,  6,  6,  6,  5,  5,  5,  5,  5,  5,  5,
    252            5,  5,  5,  5,  5,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,
    253            4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  3,  3,
    254            3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  2,
    255    };
    256    SAVE_STACK;
    257    ALLOC(tmp, len, opus_val16);
    258 
    259    *weak_transient = 0;
    260    /* For lower bitrates, let's be more conservative and have a forward masking
    261       decay of 3.3 dB/ms. This avoids having to code transients at very low
    262       bitrate (mostly for hybrid), which can result in unstable energy and/or
    263       partial collapse. */
    264    if (allow_weak_transients)
    265    {
    266 #ifdef FIXED_POINT
    267       forward_shift = 5;
    268 #else
    269       forward_decay = QCONST16(.03125f,15);
    270 #endif
    271    }
    272    len2=len/2;
    273    for (c=0;c<C;c++)
    274    {
    275       opus_val32 mean;
    276       opus_int32 unmask=0;
    277       opus_val32 norm;
    278       opus_val16 maxE;
    279       mem0=0;
    280       mem1=0;
    281       /* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */
    282       for (i=0;i<len;i++)
    283       {
    284          opus_val32 x,y;
    285          x = SHR32(in[i+c*len],SIG_SHIFT);
    286          y = ADD32(mem0, x);
    287 #ifdef FIXED_POINT
    288          mem0 = mem1 + y - SHL32(x,1);
    289          mem1 = x - SHR32(y,1);
    290 #else
    291          mem0 = mem1 + y - 2*x;
    292          mem1 = x - .5f*y;
    293 #endif
    294          tmp[i] = SROUND16(y, 2);
    295          /*printf("%f ", tmp[i]);*/
    296       }
    297       /*printf("\n");*/
    298       /* First few samples are bad because we don't propagate the memory */
    299       OPUS_CLEAR(tmp, 12);
    300 
    301 #ifdef FIXED_POINT
    302       /* Normalize tmp to max range */
    303       {
    304          int shift=0;
    305          shift = 14-celt_ilog2(MAX16(1, celt_maxabs16(tmp, len)));
    306          if (shift!=0)
    307          {
    308             for (i=0;i<len;i++)
    309                tmp[i] = SHL16(tmp[i], shift);
    310          }
    311       }
    312 #endif
    313 
    314       mean=0;
    315       mem0=0;
    316       /* Grouping by two to reduce complexity */
    317       /* Forward pass to compute the post-echo threshold*/
    318       for (i=0;i<len2;i++)
    319       {
    320          opus_val16 x2 = PSHR32(MULT16_16(tmp[2*i],tmp[2*i]) + MULT16_16(tmp[2*i+1],tmp[2*i+1]),16);
    321          mean += x2;
    322 #ifdef FIXED_POINT
    323          /* FIXME: Use PSHR16() instead */
    324          tmp[i] = mem0 + PSHR32(x2-mem0,forward_shift);
    325 #else
    326          tmp[i] = mem0 + MULT16_16_P15(forward_decay,x2-mem0);
    327 #endif
    328          mem0 = tmp[i];
    329       }
    330 
    331       mem0=0;
    332       maxE=0;
    333       /* Backward pass to compute the pre-echo threshold */
    334       for (i=len2-1;i>=0;i--)
    335       {
    336          /* Backward masking: 13.9 dB/ms. */
    337 #ifdef FIXED_POINT
    338          /* FIXME: Use PSHR16() instead */
    339          tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3);
    340 #else
    341          tmp[i] = mem0 + MULT16_16_P15(QCONST16(0.125f,15),tmp[i]-mem0);
    342 #endif
    343          mem0 = tmp[i];
    344          maxE = MAX16(maxE, mem0);
    345       }
    346       /*for (i=0;i<len2;i++)printf("%f ", tmp[i]/mean);printf("\n");*/
    347 
    348       /* Compute the ratio of the "frame energy" over the harmonic mean of the energy.
    349          This essentially corresponds to a bitrate-normalized temporal noise-to-mask
    350          ratio */
    351 
    352       /* As a compromise with the old transient detector, frame energy is the
    353          geometric mean of the energy and half the max */
    354 #ifdef FIXED_POINT
    355       /* Costs two sqrt() to avoid overflows */
    356       mean = MULT16_16(celt_sqrt(mean), celt_sqrt(MULT16_16(maxE,len2>>1)));
    357 #else
    358       mean = celt_sqrt(mean * maxE*.5*len2);
    359 #endif
    360       /* Inverse of the mean energy in Q15+6 */
    361       norm = SHL32(EXTEND32(len2),6+14)/ADD32(EPSILON,SHR32(mean,1));
    362       /* Compute harmonic mean discarding the unreliable boundaries
    363          The data is smooth, so we only take 1/4th of the samples */
    364       unmask=0;
    365       /* We should never see NaNs here. If we find any, then something really bad happened and we better abort
    366          before it does any damage later on. If these asserts are disabled (no hardening), then the table
    367          lookup a few lines below (id = ...) is likely to crash dur to an out-of-bounds read. DO NOT FIX
    368          that crash on NaN since it could result in a worse issue later on. */
    369       celt_assert(!celt_isnan(tmp[0]));
    370       celt_assert(!celt_isnan(norm));
    371       for (i=12;i<len2-5;i+=4)
    372       {
    373          int id;
    374 #ifdef FIXED_POINT
    375          id = MAX32(0,MIN32(127,MULT16_32_Q15(tmp[i]+EPSILON,norm))); /* Do not round to nearest */
    376 #else
    377          id = (int)MAX32(0,MIN32(127,floor(64*norm*(tmp[i]+EPSILON)))); /* Do not round to nearest */
    378 #endif
    379          unmask += inv_table[id];
    380       }
    381       /*printf("%d\n", unmask);*/
    382       /* Normalize, compensate for the 1/4th of the sample and the factor of 6 in the inverse table */
    383       unmask = 64*unmask*4/(6*(len2-17));
    384       if (unmask>mask_metric)
    385       {
    386          *tf_chan = c;
    387          mask_metric = unmask;
    388       }
    389    }
    390    is_transient = mask_metric>200;
    391    /* For low bitrates, define "weak transients" that need to be
    392       handled differently to avoid partial collapse. */
    393    if (allow_weak_transients && is_transient && mask_metric<600) {
    394       is_transient = 0;
    395       *weak_transient = 1;
    396    }
    397    /* Arbitrary metric for VBR boost */
    398    tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42);
    399    /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */
    400    *tf_estimate = celt_sqrt(MAX32(0, SHL32(MULT16_16(QCONST16(0.0069,14),MIN16(163,tf_max)),14)-QCONST32(0.139,28)));
    401    /*printf("%d %f\n", tf_max, mask_metric);*/
    402    RESTORE_STACK;
    403 #ifdef FUZZING
    404    is_transient = rand()&0x1;
    405 #endif
    406    /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/
    407    return is_transient;
    408 }
    409 
    410 /* Looks for sudden increases of energy to decide whether we need to patch
    411    the transient decision */
    412 static int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
    413       int start, int end, int C)
    414 {
    415    int i, c;
    416    opus_val32 mean_diff=0;
    417    opus_val16 spread_old[26];
    418    /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to
    419       avoid false detection caused by irrelevant bands */
    420    if (C==1)
    421    {
    422       spread_old[start] = oldE[start];
    423       for (i=start+1;i<end;i++)
    424          spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]);
    425    } else {
    426       spread_old[start] = MAX16(oldE[start],oldE[start+nbEBands]);
    427       for (i=start+1;i<end;i++)
    428          spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT),
    429                                MAX16(oldE[i],oldE[i+nbEBands]));
    430    }
    431    for (i=end-2;i>=start;i--)
    432       spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHIFT));
    433    /* Compute mean increase */
    434    c=0; do {
    435       for (i=IMAX(2,start);i<end-1;i++)
    436       {
    437          opus_val16 x1, x2;
    438          x1 = MAX16(0, newE[i + c*nbEBands]);
    439          x2 = MAX16(0, spread_old[i]);
    440          mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2))));
    441       }
    442    } while (++c<C);
    443    mean_diff = DIV32(mean_diff, C*(end-1-IMAX(2,start)));
    444    /*printf("%f %f %d\n", mean_diff, max_diff, count);*/
    445    return mean_diff > QCONST16(1.f, DB_SHIFT);
    446 }
    447 
    448 /** Apply window and compute the MDCT for all sub-frames and
    449     all channels in a frame */
    450 static void compute_mdcts(const CELTMode *mode, int shortBlocks, celt_sig * OPUS_RESTRICT in,
    451                           celt_sig * OPUS_RESTRICT out, int C, int CC, int LM, int upsample,
    452                           int arch)
    453 {
    454    const int overlap = mode->overlap;
    455    int N;
    456    int B;
    457    int shift;
    458    int i, b, c;
    459    if (shortBlocks)
    460    {
    461       B = shortBlocks;
    462       N = mode->shortMdctSize;
    463       shift = mode->maxLM;
    464    } else {
    465       B = 1;
    466       N = mode->shortMdctSize<<LM;
    467       shift = mode->maxLM-LM;
    468    }
    469    c=0; do {
    470       for (b=0;b<B;b++)
    471       {
    472          /* Interleaving the sub-frames while doing the MDCTs */
    473          clt_mdct_forward(&mode->mdct, in+c*(B*N+overlap)+b*N,
    474                           &out[b+c*N*B], mode->window, overlap, shift, B,
    475                           arch);
    476       }
    477    } while (++c<CC);
    478    if (CC==2&&C==1)
    479    {
    480       for (i=0;i<B*N;i++)
    481          out[i] = ADD32(HALF32(out[i]), HALF32(out[B*N+i]));
    482    }
    483    if (upsample != 1)
    484    {
    485       c=0; do
    486       {
    487          int bound = B*N/upsample;
    488          for (i=0;i<bound;i++)
    489             out[c*B*N+i] *= upsample;
    490          OPUS_CLEAR(&out[c*B*N+bound], B*N-bound);
    491       } while (++c<C);
    492    }
    493 }
    494 
    495 
    496 void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
    497                         int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip)
    498 {
    499    int i;
    500    opus_val16 coef0;
    501    celt_sig m;
    502    int Nu;
    503 
    504    coef0 = coef[0];
    505    m = *mem;
    506 
    507    /* Fast path for the normal 48kHz case and no clipping */
    508    if (coef[1] == 0 && upsample == 1 && !clip)
    509    {
    510       for (i=0;i<N;i++)
    511       {
    512          opus_val16 x;
    513          x = SCALEIN(pcmp[CC*i]);
    514          /* Apply pre-emphasis */
    515          inp[i] = SHL32(x, SIG_SHIFT) - m;
    516          m = SHR32(MULT16_16(coef0, x), 15-SIG_SHIFT);
    517       }
    518       *mem = m;
    519       return;
    520    }
    521 
    522    Nu = N/upsample;
    523    if (upsample!=1)
    524    {
    525       OPUS_CLEAR(inp, N);
    526    }
    527    for (i=0;i<Nu;i++)
    528       inp[i*upsample] = SCALEIN(pcmp[CC*i]);
    529 
    530 #ifndef FIXED_POINT
    531    if (clip)
    532    {
    533       /* Clip input to avoid encoding non-portable files */
    534       for (i=0;i<Nu;i++)
    535          inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample]));
    536    }
    537 #else
    538    (void)clip; /* Avoids a warning about clip being unused. */
    539 #endif
    540 #ifdef CUSTOM_MODES
    541    if (coef[1] != 0)
    542    {
    543       opus_val16 coef1 = coef[1];
    544       opus_val16 coef2 = coef[2];
    545       for (i=0;i<N;i++)
    546       {
    547          celt_sig x, tmp;
    548          x = inp[i];
    549          /* Apply pre-emphasis */
    550          tmp = MULT16_16(coef2, x);
    551          inp[i] = tmp + m;
    552          m = MULT16_32_Q15(coef1, inp[i]) - MULT16_32_Q15(coef0, tmp);
    553       }
    554    } else
    555 #endif
    556    {
    557       for (i=0;i<N;i++)
    558       {
    559          opus_val16 x;
    560          x = inp[i];
    561          /* Apply pre-emphasis */
    562          inp[i] = SHL32(x, SIG_SHIFT) - m;
    563          m = SHR32(MULT16_16(coef0, x), 15-SIG_SHIFT);
    564       }
    565    }
    566    *mem = m;
    567 }
    568 
    569 
    570 
    571 static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias)
    572 {
    573    int i;
    574    opus_val32 L1;
    575    L1 = 0;
    576    for (i=0;i<N;i++)
    577       L1 += EXTEND32(ABS16(tmp[i]));
    578    /* When in doubt, prefer good freq resolution */
    579    L1 = MAC16_32_Q15(L1, LM*bias, L1);
    580    return L1;
    581 
    582 }
    583 
    584 static int tf_analysis(const CELTMode *m, int len, int isTransient,
    585       int *tf_res, int lambda, celt_norm *X, int N0, int LM,
    586       opus_val16 tf_estimate, int tf_chan, int *importance)
    587 {
    588    int i;
    589    VARDECL(int, metric);
    590    int cost0;
    591    int cost1;
    592    VARDECL(int, path0);
    593    VARDECL(int, path1);
    594    VARDECL(celt_norm, tmp);
    595    VARDECL(celt_norm, tmp_1);
    596    int sel;
    597    int selcost[2];
    598    int tf_select=0;
    599    opus_val16 bias;
    600 
    601    SAVE_STACK;
    602    bias = MULT16_16_Q14(QCONST16(.04f,15), MAX16(-QCONST16(.25f,14), QCONST16(.5f,14)-tf_estimate));
    603    /*printf("%f ", bias);*/
    604 
    605    ALLOC(metric, len, int);
    606    ALLOC(tmp, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm);
    607    ALLOC(tmp_1, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm);
    608    ALLOC(path0, len, int);
    609    ALLOC(path1, len, int);
    610 
    611    for (i=0;i<len;i++)
    612    {
    613       int k, N;
    614       int narrow;
    615       opus_val32 L1, best_L1;
    616       int best_level=0;
    617       N = (m->eBands[i+1]-m->eBands[i])<<LM;
    618       /* band is too narrow to be split down to LM=-1 */
    619       narrow = (m->eBands[i+1]-m->eBands[i])==1;
    620       OPUS_COPY(tmp, &X[tf_chan*N0 + (m->eBands[i]<<LM)], N);
    621       /* Just add the right channel if we're in stereo */
    622       /*if (C==2)
    623          for (j=0;j<N;j++)
    624             tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));*/
    625       L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias);
    626       best_L1 = L1;
    627       /* Check the -1 case for transients */
    628       if (isTransient && !narrow)
    629       {
    630          OPUS_COPY(tmp_1, tmp, N);
    631          haar1(tmp_1, N>>LM, 1<<LM);
    632          L1 = l1_metric(tmp_1, N, LM+1, bias);
    633          if (L1<best_L1)
    634          {
    635             best_L1 = L1;
    636             best_level = -1;
    637          }
    638       }
    639       /*printf ("%f ", L1);*/
    640       for (k=0;k<LM+!(isTransient||narrow);k++)
    641       {
    642          int B;
    643 
    644          if (isTransient)
    645             B = (LM-k-1);
    646          else
    647             B = k+1;
    648 
    649          haar1(tmp, N>>k, 1<<k);
    650 
    651          L1 = l1_metric(tmp, N, B, bias);
    652 
    653          if (L1 < best_L1)
    654          {
    655             best_L1 = L1;
    656             best_level = k+1;
    657          }
    658       }
    659       /*printf ("%d ", isTransient ? LM-best_level : best_level);*/
    660       /* metric is in Q1 to be able to select the mid-point (-0.5) for narrower bands */
    661       if (isTransient)
    662          metric[i] = 2*best_level;
    663       else
    664          metric[i] = -2*best_level;
    665       /* For bands that can't be split to -1, set the metric to the half-way point to avoid
    666          biasing the decision */
    667       if (narrow && (metric[i]==0 || metric[i]==-2*LM))
    668          metric[i]-=1;
    669       /*printf("%d ", metric[i]/2 + (!isTransient)*LM);*/
    670    }
    671    /*printf("\n");*/
    672    /* Search for the optimal tf resolution, including tf_select */
    673    tf_select = 0;
    674    for (sel=0;sel<2;sel++)
    675    {
    676       cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
    677       cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+1]) + (isTransient ? 0 : lambda);
    678       for (i=1;i<len;i++)
    679       {
    680          int curr0, curr1;
    681          curr0 = IMIN(cost0, cost1 + lambda);
    682          curr1 = IMIN(cost0 + lambda, cost1);
    683          cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
    684          cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]);
    685       }
    686       cost0 = IMIN(cost0, cost1);
    687       selcost[sel]=cost0;
    688    }
    689    /* For now, we're conservative and only allow tf_select=1 for transients.
    690     * If tests confirm it's useful for non-transients, we could allow it. */
    691    if (selcost[1]<selcost[0] && isTransient)
    692       tf_select=1;
    693    cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
    694    cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]) + (isTransient ? 0 : lambda);
    695    /* Viterbi forward pass */
    696    for (i=1;i<len;i++)
    697    {
    698       int curr0, curr1;
    699       int from0, from1;
    700 
    701       from0 = cost0;
    702       from1 = cost1 + lambda;
    703       if (from0 < from1)
    704       {
    705          curr0 = from0;
    706          path0[i]= 0;
    707       } else {
    708          curr0 = from1;
    709          path0[i]= 1;
    710       }
    711 
    712       from0 = cost0 + lambda;
    713       from1 = cost1;
    714       if (from0 < from1)
    715       {
    716          curr1 = from0;
    717          path1[i]= 0;
    718       } else {
    719          curr1 = from1;
    720          path1[i]= 1;
    721       }
    722       cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
    723       cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]);
    724    }
    725    tf_res[len-1] = cost0 < cost1 ? 0 : 1;
    726    /* Viterbi backward pass to check the decisions */
    727    for (i=len-2;i>=0;i--)
    728    {
    729       if (tf_res[i+1] == 1)
    730          tf_res[i] = path1[i+1];
    731       else
    732          tf_res[i] = path0[i+1];
    733    }
    734    /*printf("%d %f\n", *tf_sum, tf_estimate);*/
    735    RESTORE_STACK;
    736 #ifdef FUZZING
    737    tf_select = rand()&0x1;
    738    tf_res[0] = rand()&0x1;
    739    for (i=1;i<len;i++)
    740       tf_res[i] = tf_res[i-1] ^ ((rand()&0xF) == 0);
    741 #endif
    742    return tf_select;
    743 }
    744 
    745 static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM, int tf_select, ec_enc *enc)
    746 {
    747    int curr, i;
    748    int tf_select_rsv;
    749    int tf_changed;
    750    int logp;
    751    opus_uint32 budget;
    752    opus_uint32 tell;
    753    budget = enc->storage*8;
    754    tell = ec_tell(enc);
    755    logp = isTransient ? 2 : 4;
    756    /* Reserve space to code the tf_select decision. */
    757    tf_select_rsv = LM>0 && tell+logp+1 <= budget;
    758    budget -= tf_select_rsv;
    759    curr = tf_changed = 0;
    760    for (i=start;i<end;i++)
    761    {
    762       if (tell+logp<=budget)
    763       {
    764          ec_enc_bit_logp(enc, tf_res[i] ^ curr, logp);
    765          tell = ec_tell(enc);
    766          curr = tf_res[i];
    767          tf_changed |= curr;
    768       }
    769       else
    770          tf_res[i] = curr;
    771       logp = isTransient ? 4 : 5;
    772    }
    773    /* Only code tf_select if it would actually make a difference. */
    774    if (tf_select_rsv &&
    775          tf_select_table[LM][4*isTransient+0+tf_changed]!=
    776          tf_select_table[LM][4*isTransient+2+tf_changed])
    777       ec_enc_bit_logp(enc, tf_select, 1);
    778    else
    779       tf_select = 0;
    780    for (i=start;i<end;i++)
    781       tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]];
    782    /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);printf("\n");*/
    783 }
    784 
    785 
    786 static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X,
    787       const opus_val16 *bandLogE, int end, int LM, int C, int N0,
    788       AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate,
    789       int intensity, opus_val16 surround_trim, opus_int32 equiv_rate, int arch)
    790 {
    791    int i;
    792    opus_val32 diff=0;
    793    int c;
    794    int trim_index;
    795    opus_val16 trim = QCONST16(5.f, 8);
    796    opus_val16 logXC, logXC2;
    797    /* At low bitrate, reducing the trim seems to help. At higher bitrates, it's less
    798       clear what's best, so we're keeping it as it was before, at least for now. */
    799    if (equiv_rate < 64000) {
    800       trim = QCONST16(4.f, 8);
    801    } else if (equiv_rate < 80000) {
    802       opus_int32 frac = (equiv_rate-64000) >> 10;
    803       trim = QCONST16(4.f, 8) + QCONST16(1.f/16.f, 8)*frac;
    804    }
    805    if (C==2)
    806    {
    807       opus_val16 sum = 0; /* Q10 */
    808       opus_val16 minXC; /* Q10 */
    809       /* Compute inter-channel correlation for low frequencies */
    810       for (i=0;i<8;i++)
    811       {
    812          opus_val32 partial;
    813          partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)],
    814                (m->eBands[i+1]-m->eBands[i])<<LM, arch);
    815          sum = ADD16(sum, EXTRACT16(SHR32(partial, 18)));
    816       }
    817       sum = MULT16_16_Q15(QCONST16(1.f/8, 15), sum);
    818       sum = MIN16(QCONST16(1.f, 10), ABS16(sum));
    819       minXC = sum;
    820       for (i=8;i<intensity;i++)
    821       {
    822          opus_val32 partial;
    823          partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)],
    824                (m->eBands[i+1]-m->eBands[i])<<LM, arch);
    825          minXC = MIN16(minXC, ABS16(EXTRACT16(SHR32(partial, 18))));
    826       }
    827       minXC = MIN16(QCONST16(1.f, 10), ABS16(minXC));
    828       /*printf ("%f\n", sum);*/
    829       /* mid-side savings estimations based on the LF average*/
    830       logXC = celt_log2(QCONST32(1.001f, 20)-MULT16_16(sum, sum));
    831       /* mid-side savings estimations based on min correlation */
    832       logXC2 = MAX16(HALF16(logXC), celt_log2(QCONST32(1.001f, 20)-MULT16_16(minXC, minXC)));
    833 #ifdef FIXED_POINT
    834       /* Compensate for Q20 vs Q14 input and convert output to Q8 */
    835       logXC = PSHR32(logXC-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8);
    836       logXC2 = PSHR32(logXC2-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8);
    837 #endif
    838 
    839       trim += MAX16(-QCONST16(4.f, 8), MULT16_16_Q15(QCONST16(.75f,15),logXC));
    840       *stereo_saving = MIN16(*stereo_saving + QCONST16(0.25f, 8), -HALF16(logXC2));
    841    }
    842 
    843    /* Estimate spectral tilt */
    844    c=0; do {
    845       for (i=0;i<end-1;i++)
    846       {
    847          diff += bandLogE[i+c*m->nbEBands]*(opus_int32)(2+2*i-end);
    848       }
    849    } while (++c<C);
    850    diff /= C*(end-1);
    851    /*printf("%f\n", diff);*/
    852    trim -= MAX32(-QCONST16(2.f, 8), MIN32(QCONST16(2.f, 8), SHR32(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 ));
    853    trim -= SHR16(surround_trim, DB_SHIFT-8);
    854    trim -= 2*SHR16(tf_estimate, 14-8);
    855 #ifndef DISABLE_FLOAT_API
    856    if (analysis->valid)
    857    {
    858       trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8),
    859             (opus_val16)(QCONST16(2.f, 8)*(analysis->tonality_slope+.05f))));
    860    }
    861 #else
    862    (void)analysis;
    863 #endif
    864 
    865 #ifdef FIXED_POINT
    866    trim_index = PSHR32(trim, 8);
    867 #else
    868    trim_index = (int)floor(.5f+trim);
    869 #endif
    870    trim_index = IMAX(0, IMIN(10, trim_index));
    871    /*printf("%d\n", trim_index);*/
    872 #ifdef FUZZING
    873    trim_index = rand()%11;
    874 #endif
    875    return trim_index;
    876 }
    877 
    878 static int stereo_analysis(const CELTMode *m, const celt_norm *X,
    879       int LM, int N0)
    880 {
    881    int i;
    882    int thetas;
    883    opus_val32 sumLR = EPSILON, sumMS = EPSILON;
    884 
    885    /* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal */
    886    for (i=0;i<13;i++)
    887    {
    888       int j;
    889       for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++)
    890       {
    891          opus_val32 L, R, M, S;
    892          /* We cast to 32-bit first because of the -32768 case */
    893          L = EXTEND32(X[j]);
    894          R = EXTEND32(X[N0+j]);
    895          M = ADD32(L, R);
    896          S = SUB32(L, R);
    897          sumLR = ADD32(sumLR, ADD32(ABS32(L), ABS32(R)));
    898          sumMS = ADD32(sumMS, ADD32(ABS32(M), ABS32(S)));
    899       }
    900    }
    901    sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS);
    902    thetas = 13;
    903    /* We don't need thetas for lower bands with LM<=1 */
    904    if (LM<=1)
    905       thetas -= 8;
    906    return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS)
    907          > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR);
    908 }
    909 
    910 #define MSWAP(a,b) do {opus_val16 tmp = a;a=b;b=tmp;} while(0)
    911 static opus_val16 median_of_5(const opus_val16 *x)
    912 {
    913    opus_val16 t0, t1, t2, t3, t4;
    914    t2 = x[2];
    915    if (x[0] > x[1])
    916    {
    917       t0 = x[1];
    918       t1 = x[0];
    919    } else {
    920       t0 = x[0];
    921       t1 = x[1];
    922    }
    923    if (x[3] > x[4])
    924    {
    925       t3 = x[4];
    926       t4 = x[3];
    927    } else {
    928       t3 = x[3];
    929       t4 = x[4];
    930    }
    931    if (t0 > t3)
    932    {
    933       MSWAP(t0, t3);
    934       MSWAP(t1, t4);
    935    }
    936    if (t2 > t1)
    937    {
    938       if (t1 < t3)
    939          return MIN16(t2, t3);
    940       else
    941          return MIN16(t4, t1);
    942    } else {
    943       if (t2 < t3)
    944          return MIN16(t1, t3);
    945       else
    946          return MIN16(t2, t4);
    947    }
    948 }
    949 
    950 static opus_val16 median_of_3(const opus_val16 *x)
    951 {
    952    opus_val16 t0, t1, t2;
    953    if (x[0] > x[1])
    954    {
    955       t0 = x[1];
    956       t1 = x[0];
    957    } else {
    958       t0 = x[0];
    959       t1 = x[1];
    960    }
    961    t2 = x[2];
    962    if (t1 < t2)
    963       return t1;
    964    else if (t0 < t2)
    965       return t2;
    966    else
    967       return t0;
    968 }
    969 
    970 static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 *bandLogE2,
    971       int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN,
    972       int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM,
    973       int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc,
    974       AnalysisInfo *analysis, int *importance, int *spread_weight)
    975 {
    976    int i, c;
    977    opus_int32 tot_boost=0;
    978    opus_val16 maxDepth;
    979    VARDECL(opus_val16, follower);
    980    VARDECL(opus_val16, noise_floor);
    981    SAVE_STACK;
    982    ALLOC(follower, C*nbEBands, opus_val16);
    983    ALLOC(noise_floor, C*nbEBands, opus_val16);
    984    OPUS_CLEAR(offsets, nbEBands);
    985    /* Dynamic allocation code */
    986    maxDepth=-QCONST16(31.9f, DB_SHIFT);
    987    for (i=0;i<end;i++)
    988    {
    989       /* Noise floor must take into account eMeans, the depth, the width of the bands
    990          and the preemphasis filter (approx. square of bark band ID) */
    991       noise_floor[i] = MULT16_16(QCONST16(0.0625f, DB_SHIFT),logN[i])
    992             +QCONST16(.5f,DB_SHIFT)+SHL16(9-lsb_depth,DB_SHIFT)-SHL16(eMeans[i],6)
    993             +MULT16_16(QCONST16(.0062,DB_SHIFT),(i+5)*(i+5));
    994    }
    995    c=0;do
    996    {
    997       for (i=0;i<end;i++)
    998          maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]);
    999    } while (++c<C);
   1000    {
   1001       /* Compute a really simple masking model to avoid taking into account completely masked
   1002          bands when computing the spreading decision. */
   1003       VARDECL(opus_val16, mask);
   1004       VARDECL(opus_val16, sig);
   1005       ALLOC(mask, nbEBands, opus_val16);
   1006       ALLOC(sig, nbEBands, opus_val16);
   1007       for (i=0;i<end;i++)
   1008          mask[i] = bandLogE[i]-noise_floor[i];
   1009       if (C==2)
   1010       {
   1011          for (i=0;i<end;i++)
   1012             mask[i] = MAX16(mask[i], bandLogE[nbEBands+i]-noise_floor[i]);
   1013       }
   1014       OPUS_COPY(sig, mask, end);
   1015       for (i=1;i<end;i++)
   1016          mask[i] = MAX16(mask[i], mask[i-1] - QCONST16(2.f, DB_SHIFT));
   1017       for (i=end-2;i>=0;i--)
   1018          mask[i] = MAX16(mask[i], mask[i+1] - QCONST16(3.f, DB_SHIFT));
   1019       for (i=0;i<end;i++)
   1020       {
   1021          /* Compute SMR: Mask is never more than 72 dB below the peak and never below the noise floor.*/
   1022          opus_val16 smr = sig[i]-MAX16(MAX16(0, maxDepth-QCONST16(12.f, DB_SHIFT)), mask[i]);
   1023          /* Clamp SMR to make sure we're not shifting by something negative or too large. */
   1024 #ifdef FIXED_POINT
   1025          /* FIXME: Use PSHR16() instead */
   1026          int shift = -PSHR32(MAX16(-QCONST16(5.f, DB_SHIFT), MIN16(0, smr)), DB_SHIFT);
   1027 #else
   1028          int shift = IMIN(5, IMAX(0, -(int)floor(.5f + smr)));
   1029 #endif
   1030          spread_weight[i] = 32 >> shift;
   1031       }
   1032       /*for (i=0;i<end;i++)
   1033          printf("%d ", spread_weight[i]);
   1034       printf("\n");*/
   1035    }
   1036    /* Make sure that dynamic allocation can't make us bust the budget */
   1037    if (effectiveBytes > 50 && LM>=1 && !lfe)
   1038    {
   1039       int last=0;
   1040       c=0;do
   1041       {
   1042          opus_val16 offset;
   1043          opus_val16 tmp;
   1044          opus_val16 *f;
   1045          f = &follower[c*nbEBands];
   1046          f[0] = bandLogE2[c*nbEBands];
   1047          for (i=1;i<end;i++)
   1048          {
   1049             /* The last band to be at least 3 dB higher than the previous one
   1050                is the last we'll consider. Otherwise, we run into problems on
   1051                bandlimited signals. */
   1052             if (bandLogE2[c*nbEBands+i] > bandLogE2[c*nbEBands+i-1]+QCONST16(.5f,DB_SHIFT))
   1053                last=i;
   1054             f[i] = MIN16(f[i-1]+QCONST16(1.5f,DB_SHIFT), bandLogE2[c*nbEBands+i]);
   1055          }
   1056          for (i=last-1;i>=0;i--)
   1057             f[i] = MIN16(f[i], MIN16(f[i+1]+QCONST16(2.f,DB_SHIFT), bandLogE2[c*nbEBands+i]));
   1058 
   1059          /* Combine with a median filter to avoid dynalloc triggering unnecessarily.
   1060             The "offset" value controls how conservative we are -- a higher offset
   1061             reduces the impact of the median filter and makes dynalloc use more bits. */
   1062          offset = QCONST16(1.f, DB_SHIFT);
   1063          for (i=2;i<end-2;i++)
   1064             f[i] = MAX16(f[i], median_of_5(&bandLogE2[c*nbEBands+i-2])-offset);
   1065          tmp = median_of_3(&bandLogE2[c*nbEBands])-offset;
   1066          f[0] = MAX16(f[0], tmp);
   1067          f[1] = MAX16(f[1], tmp);
   1068          tmp = median_of_3(&bandLogE2[c*nbEBands+end-3])-offset;
   1069          f[end-2] = MAX16(f[end-2], tmp);
   1070          f[end-1] = MAX16(f[end-1], tmp);
   1071 
   1072          for (i=0;i<end;i++)
   1073             f[i] = MAX16(f[i], noise_floor[i]);
   1074       } while (++c<C);
   1075       if (C==2)
   1076       {
   1077          for (i=start;i<end;i++)
   1078          {
   1079             /* Consider 24 dB "cross-talk" */
   1080             follower[nbEBands+i] = MAX16(follower[nbEBands+i], follower[         i]-QCONST16(4.f,DB_SHIFT));
   1081             follower[         i] = MAX16(follower[         i], follower[nbEBands+i]-QCONST16(4.f,DB_SHIFT));
   1082             follower[i] = HALF16(MAX16(0, bandLogE[i]-follower[i]) + MAX16(0, bandLogE[nbEBands+i]-follower[nbEBands+i]));
   1083          }
   1084       } else {
   1085          for (i=start;i<end;i++)
   1086          {
   1087             follower[i] = MAX16(0, bandLogE[i]-follower[i]);
   1088          }
   1089       }
   1090       for (i=start;i<end;i++)
   1091          follower[i] = MAX16(follower[i], surround_dynalloc[i]);
   1092       for (i=start;i<end;i++)
   1093       {
   1094 #ifdef FIXED_POINT
   1095          importance[i] = PSHR32(13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))), 16);
   1096 #else
   1097          importance[i] = (int)floor(.5f+13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))));
   1098 #endif
   1099       }
   1100       /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */
   1101       if ((!vbr || constrained_vbr)&&!isTransient)
   1102       {
   1103          for (i=start;i<end;i++)
   1104             follower[i] = HALF16(follower[i]);
   1105       }
   1106       for (i=start;i<end;i++)
   1107       {
   1108          if (i<8)
   1109             follower[i] *= 2;
   1110          if (i>=12)
   1111             follower[i] = HALF16(follower[i]);
   1112       }
   1113 #ifdef DISABLE_FLOAT_API
   1114       (void)analysis;
   1115 #else
   1116       if (analysis->valid)
   1117       {
   1118          for (i=start;i<IMIN(LEAK_BANDS, end);i++)
   1119             follower[i] = follower[i] +  QCONST16(1.f/64.f, DB_SHIFT)*analysis->leak_boost[i];
   1120       }
   1121 #endif
   1122       for (i=start;i<end;i++)
   1123       {
   1124          int width;
   1125          int boost;
   1126          int boost_bits;
   1127 
   1128          follower[i] = MIN16(follower[i], QCONST16(4, DB_SHIFT));
   1129 
   1130          width = C*(eBands[i+1]-eBands[i])<<LM;
   1131          if (width<6)
   1132          {
   1133             boost = (int)SHR32(EXTEND32(follower[i]),DB_SHIFT);
   1134             boost_bits = boost*width<<BITRES;
   1135          } else if (width > 48) {
   1136             boost = (int)SHR32(EXTEND32(follower[i])*8,DB_SHIFT);
   1137             boost_bits = (boost*width<<BITRES)/8;
   1138          } else {
   1139             boost = (int)SHR32(EXTEND32(follower[i])*width/6,DB_SHIFT);
   1140             boost_bits = boost*6<<BITRES;
   1141          }
   1142          /* For CBR and non-transient CVBR frames, limit dynalloc to 2/3 of the bits */
   1143          if ((!vbr || (constrained_vbr&&!isTransient))
   1144                && (tot_boost+boost_bits)>>BITRES>>3 > 2*effectiveBytes/3)
   1145          {
   1146             opus_int32 cap = ((2*effectiveBytes/3)<<BITRES<<3);
   1147             offsets[i] = cap-tot_boost;
   1148             tot_boost = cap;
   1149             break;
   1150          } else {
   1151             offsets[i] = boost;
   1152             tot_boost += boost_bits;
   1153          }
   1154       }
   1155    } else {
   1156       for (i=start;i<end;i++)
   1157          importance[i] = 13;
   1158    }
   1159    *tot_boost_ = tot_boost;
   1160    RESTORE_STACK;
   1161    return maxDepth;
   1162 }
   1163 
   1164 
   1165 static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N,
   1166       int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes, AnalysisInfo *analysis)
   1167 {
   1168    int c;
   1169    VARDECL(celt_sig, _pre);
   1170    celt_sig *pre[2];
   1171    const CELTMode *mode;
   1172    int pitch_index;
   1173    opus_val16 gain1;
   1174    opus_val16 pf_threshold;
   1175    int pf_on;
   1176    int qg;
   1177    int overlap;
   1178    SAVE_STACK;
   1179 
   1180    mode = st->mode;
   1181    overlap = mode->overlap;
   1182    ALLOC(_pre, CC*(N+COMBFILTER_MAXPERIOD), celt_sig);
   1183 
   1184    pre[0] = _pre;
   1185    pre[1] = _pre + (N+COMBFILTER_MAXPERIOD);
   1186 
   1187 
   1188    c=0; do {
   1189       OPUS_COPY(pre[c], prefilter_mem+c*COMBFILTER_MAXPERIOD, COMBFILTER_MAXPERIOD);
   1190       OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+overlap)+overlap, N);
   1191    } while (++c<CC);
   1192 
   1193    if (enabled)
   1194    {
   1195       VARDECL(opus_val16, pitch_buf);
   1196       ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16);
   1197 
   1198       pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC, st->arch);
   1199       /* Don't search for the fir last 1.5 octave of the range because
   1200          there's too many false-positives due to short-term correlation */
   1201       pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N,
   1202             COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index,
   1203             st->arch);
   1204       pitch_index = COMBFILTER_MAXPERIOD-pitch_index;
   1205 
   1206       gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPERIOD,
   1207             N, &pitch_index, st->prefilter_period, st->prefilter_gain, st->arch);
   1208       if (pitch_index > COMBFILTER_MAXPERIOD-2)
   1209          pitch_index = COMBFILTER_MAXPERIOD-2;
   1210       gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1);
   1211       /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.tonality);*/
   1212       if (st->loss_rate>2)
   1213          gain1 = HALF32(gain1);
   1214       if (st->loss_rate>4)
   1215          gain1 = HALF32(gain1);
   1216       if (st->loss_rate>8)
   1217          gain1 = 0;
   1218    } else {
   1219       gain1 = 0;
   1220       pitch_index = COMBFILTER_MINPERIOD;
   1221    }
   1222 #ifndef DISABLE_FLOAT_API
   1223    if (analysis->valid)
   1224       gain1 = (opus_val16)(gain1 * analysis->max_pitch_ratio);
   1225 #else
   1226    (void)analysis;
   1227 #endif
   1228    /* Gain threshold for enabling the prefilter/postfilter */
   1229    pf_threshold = QCONST16(.2f,15);
   1230 
   1231    /* Adjusting the threshold based on rate and continuity */
   1232    if (abs(pitch_index-st->prefilter_period)*10>pitch_index)
   1233       pf_threshold += QCONST16(.2f,15);
   1234    if (nbAvailableBytes<25)
   1235       pf_threshold += QCONST16(.1f,15);
   1236    if (nbAvailableBytes<35)
   1237       pf_threshold += QCONST16(.1f,15);
   1238    if (st->prefilter_gain > QCONST16(.4f,15))
   1239       pf_threshold -= QCONST16(.1f,15);
   1240    if (st->prefilter_gain > QCONST16(.55f,15))
   1241       pf_threshold -= QCONST16(.1f,15);
   1242 
   1243    /* Hard threshold at 0.2 */
   1244    pf_threshold = MAX16(pf_threshold, QCONST16(.2f,15));
   1245    if (gain1<pf_threshold)
   1246    {
   1247       gain1 = 0;
   1248       pf_on = 0;
   1249       qg = 0;
   1250    } else {
   1251       /*This block is not gated by a total bits check only because
   1252         of the nbAvailableBytes check above.*/
   1253       if (ABS16(gain1-st->prefilter_gain)<QCONST16(.1f,15))
   1254          gain1=st->prefilter_gain;
   1255 
   1256 #ifdef FIXED_POINT
   1257       qg = ((gain1+1536)>>10)/3-1;
   1258 #else
   1259       qg = (int)floor(.5f+gain1*32/3)-1;
   1260 #endif
   1261       qg = IMAX(0, IMIN(7, qg));
   1262       gain1 = QCONST16(0.09375f,15)*(qg+1);
   1263       pf_on = 1;
   1264    }
   1265    /*printf("%d %f\n", pitch_index, gain1);*/
   1266 
   1267    c=0; do {
   1268       int offset = mode->shortMdctSize-overlap;
   1269       st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD);
   1270       OPUS_COPY(in+c*(N+overlap), st->in_mem+c*(overlap), overlap);
   1271       if (offset)
   1272          comb_filter(in+c*(N+overlap)+overlap, pre[c]+COMBFILTER_MAXPERIOD,
   1273                st->prefilter_period, st->prefilter_period, offset, -st->prefilter_gain, -st->prefilter_gain,
   1274                st->prefilter_tapset, st->prefilter_tapset, NULL, 0, st->arch);
   1275 
   1276       comb_filter(in+c*(N+overlap)+overlap+offset, pre[c]+COMBFILTER_MAXPERIOD+offset,
   1277             st->prefilter_period, pitch_index, N-offset, -st->prefilter_gain, -gain1,
   1278             st->prefilter_tapset, prefilter_tapset, mode->window, overlap, st->arch);
   1279       OPUS_COPY(st->in_mem+c*(overlap), in+c*(N+overlap)+N, overlap);
   1280 
   1281       if (N>COMBFILTER_MAXPERIOD)
   1282       {
   1283          OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD, pre[c]+N, COMBFILTER_MAXPERIOD);
   1284       } else {
   1285          OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, prefilter_mem+c*COMBFILTER_MAXPERIOD+N, COMBFILTER_MAXPERIOD-N);
   1286          OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD+COMBFILTER_MAXPERIOD-N, pre[c]+COMBFILTER_MAXPERIOD, N);
   1287       }
   1288    } while (++c<CC);
   1289 
   1290    RESTORE_STACK;
   1291    *gain = gain1;
   1292    *pitch = pitch_index;
   1293    *qgain = qg;
   1294    return pf_on;
   1295 }
   1296 
   1297 static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 base_target,
   1298       int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity,
   1299       int constrained_vbr, opus_val16 stereo_saving, int tot_boost,
   1300       opus_val16 tf_estimate, int pitch_change, opus_val16 maxDepth,
   1301       int lfe, int has_surround_mask, opus_val16 surround_masking,
   1302       opus_val16 temporal_vbr)
   1303 {
   1304    /* The target rate in 8th bits per frame */
   1305    opus_int32 target;
   1306    int coded_bins;
   1307    int coded_bands;
   1308    opus_val16 tf_calibration;
   1309    int nbEBands;
   1310    const opus_int16 *eBands;
   1311 
   1312    nbEBands = mode->nbEBands;
   1313    eBands = mode->eBands;
   1314 
   1315    coded_bands = lastCodedBands ? lastCodedBands : nbEBands;
   1316    coded_bins = eBands[coded_bands]<<LM;
   1317    if (C==2)
   1318       coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM;
   1319 
   1320    target = base_target;
   1321 
   1322    /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality, tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/
   1323 #ifndef DISABLE_FLOAT_API
   1324    if (analysis->valid && analysis->activity<.4)
   1325       target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity));
   1326 #endif
   1327    /* Stereo savings */
   1328    if (C==2)
   1329    {
   1330       int coded_stereo_bands;
   1331       int coded_stereo_dof;
   1332       opus_val16 max_frac;
   1333       coded_stereo_bands = IMIN(intensity, coded_bands);
   1334       coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands;
   1335       /* Maximum fraction of the bits we can save if the signal is mono. */
   1336       max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded_bins);
   1337       stereo_saving = MIN16(stereo_saving, QCONST16(1.f, 8));
   1338       /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/
   1339       target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target),
   1340                       SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8));
   1341    }
   1342    /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */
   1343    target += tot_boost-(19<<LM);
   1344    /* Apply transient boost, compensating for average boost. */
   1345    tf_calibration = QCONST16(0.044f,14);
   1346    target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1);
   1347 
   1348 #ifndef DISABLE_FLOAT_API
   1349    /* Apply tonality boost */
   1350    if (analysis->valid && !lfe)
   1351    {
   1352       opus_int32 tonal_target;
   1353       float tonal;
   1354 
   1355       /* Tonality boost (compensating for the average). */
   1356       tonal = MAX16(0.f,analysis->tonality-.15f)-0.12f;
   1357       tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal);
   1358       if (pitch_change)
   1359          tonal_target +=  (opus_int32)((coded_bins<<BITRES)*.8f);
   1360       /*printf("%f %f ", analysis->tonality, tonal);*/
   1361       target = tonal_target;
   1362    }
   1363 #else
   1364    (void)analysis;
   1365    (void)pitch_change;
   1366 #endif
   1367 
   1368    if (has_surround_mask&&!lfe)
   1369    {
   1370       opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(surround_masking,coded_bins<<BITRES), DB_SHIFT);
   1371       /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, st->intensity, surround_target, target, st->bitrate);*/
   1372       target = IMAX(target/4, surround_target);
   1373    }
   1374 
   1375    {
   1376       opus_int32 floor_depth;
   1377       int bins;
   1378       bins = eBands[nbEBands-2]<<LM;
   1379       /*floor_depth = SHR32(MULT16_16((C*bins<<BITRES),celt_log2(SHL32(MAX16(1,sample_max),13))), DB_SHIFT);*/
   1380       floor_depth = (opus_int32)SHR32(MULT16_16((C*bins<<BITRES),maxDepth), DB_SHIFT);
   1381       floor_depth = IMAX(floor_depth, target>>2);
   1382       target = IMIN(target, floor_depth);
   1383       /*printf("%f %d\n", maxDepth, floor_depth);*/
   1384    }
   1385 
   1386    /* Make VBR less aggressive for constrained VBR because we can't keep a higher bitrate
   1387       for long. Needs tuning. */
   1388    if ((!has_surround_mask||lfe) && constrained_vbr)
   1389    {
   1390       target = base_target + (opus_int32)MULT16_32_Q15(QCONST16(0.67f, 15), target-base_target);
   1391    }
   1392 
   1393    if (!has_surround_mask && tf_estimate < QCONST16(.2f, 14))
   1394    {
   1395       opus_val16 amount;
   1396       opus_val16 tvbr_factor;
   1397       amount = MULT16_16_Q15(QCONST16(.0000031f, 30), IMAX(0, IMIN(32000, 96000-bitrate)));
   1398       tvbr_factor = SHR32(MULT16_16(temporal_vbr, amount), DB_SHIFT);
   1399       target += (opus_int32)MULT16_32_Q15(tvbr_factor, target);
   1400    }
   1401 
   1402    /* Don't allow more than doubling the rate */
   1403    target = IMIN(2*base_target, target);
   1404 
   1405    return target;
   1406 }
   1407 
   1408 int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc)
   1409 {
   1410    int i, c, N;
   1411    opus_int32 bits;
   1412    ec_enc _enc;
   1413    VARDECL(celt_sig, in);
   1414    VARDECL(celt_sig, freq);
   1415    VARDECL(celt_norm, X);
   1416    VARDECL(celt_ener, bandE);
   1417    VARDECL(opus_val16, bandLogE);
   1418    VARDECL(opus_val16, bandLogE2);
   1419    VARDECL(int, fine_quant);
   1420    VARDECL(opus_val16, error);
   1421    VARDECL(int, pulses);
   1422    VARDECL(int, cap);
   1423    VARDECL(int, offsets);
   1424    VARDECL(int, importance);
   1425    VARDECL(int, spread_weight);
   1426    VARDECL(int, fine_priority);
   1427    VARDECL(int, tf_res);
   1428    VARDECL(unsigned char, collapse_masks);
   1429    celt_sig *prefilter_mem;
   1430    opus_val16 *oldBandE, *oldLogE, *oldLogE2, *energyError;
   1431    int shortBlocks=0;
   1432    int isTransient=0;
   1433    const int CC = st->channels;
   1434    const int C = st->stream_channels;
   1435    int LM, M;
   1436    int tf_select;
   1437    int nbFilledBytes, nbAvailableBytes;
   1438    int start;
   1439    int end;
   1440    int effEnd;
   1441    int codedBands;
   1442    int alloc_trim;
   1443    int pitch_index=COMBFILTER_MINPERIOD;
   1444    opus_val16 gain1 = 0;
   1445    int dual_stereo=0;
   1446    int effectiveBytes;
   1447    int dynalloc_logp;
   1448    opus_int32 vbr_rate;
   1449    opus_int32 total_bits;
   1450    opus_int32 total_boost;
   1451    opus_int32 balance;
   1452    opus_int32 tell;
   1453    opus_int32 tell0_frac;
   1454    int prefilter_tapset=0;
   1455    int pf_on;
   1456    int anti_collapse_rsv;
   1457    int anti_collapse_on=0;
   1458    int silence=0;
   1459    int tf_chan = 0;
   1460    opus_val16 tf_estimate;
   1461    int pitch_change=0;
   1462    opus_int32 tot_boost;
   1463    opus_val32 sample_max;
   1464    opus_val16 maxDepth;
   1465    const OpusCustomMode *mode;
   1466    int nbEBands;
   1467    int overlap;
   1468    const opus_int16 *eBands;
   1469    int secondMdct;
   1470    int signalBandwidth;
   1471    int transient_got_disabled=0;
   1472    opus_val16 surround_masking=0;
   1473    opus_val16 temporal_vbr=0;
   1474    opus_val16 surround_trim = 0;
   1475    opus_int32 equiv_rate;
   1476    int hybrid;
   1477    int weak_transient = 0;
   1478    int enable_tf_analysis;
   1479    VARDECL(opus_val16, surround_dynalloc);
   1480    ALLOC_STACK;
   1481 
   1482    mode = st->mode;
   1483    nbEBands = mode->nbEBands;
   1484    overlap = mode->overlap;
   1485    eBands = mode->eBands;
   1486    start = st->start;
   1487    end = st->end;
   1488    hybrid = start != 0;
   1489    tf_estimate = 0;
   1490    if (nbCompressedBytes<2 || pcm==NULL)
   1491    {
   1492       RESTORE_STACK;
   1493       return OPUS_BAD_ARG;
   1494    }
   1495 
   1496    frame_size *= st->upsample;
   1497    for (LM=0;LM<=mode->maxLM;LM++)
   1498       if (mode->shortMdctSize<<LM==frame_size)
   1499          break;
   1500    if (LM>mode->maxLM)
   1501    {
   1502       RESTORE_STACK;
   1503       return OPUS_BAD_ARG;
   1504    }
   1505    M=1<<LM;
   1506    N = M*mode->shortMdctSize;
   1507 
   1508    prefilter_mem = st->in_mem+CC*(overlap);
   1509    oldBandE = (opus_val16*)(st->in_mem+CC*(overlap+COMBFILTER_MAXPERIOD));
   1510    oldLogE = oldBandE + CC*nbEBands;
   1511    oldLogE2 = oldLogE + CC*nbEBands;
   1512    energyError = oldLogE2 + CC*nbEBands;
   1513 
   1514    if (enc==NULL)
   1515    {
   1516       tell0_frac=tell=1;
   1517       nbFilledBytes=0;
   1518    } else {
   1519       tell0_frac=ec_tell_frac(enc);
   1520       tell=ec_tell(enc);
   1521       nbFilledBytes=(tell+4)>>3;
   1522    }
   1523 
   1524 #ifdef CUSTOM_MODES
   1525    if (st->signalling && enc==NULL)
   1526    {
   1527       int tmp = (mode->effEBands-end)>>1;
   1528       end = st->end = IMAX(1, mode->effEBands-tmp);
   1529       compressed[0] = tmp<<5;
   1530       compressed[0] |= LM<<3;
   1531       compressed[0] |= (C==2)<<2;
   1532       /* Convert "standard mode" to Opus header */
   1533       if (mode->Fs==48000 && mode->shortMdctSize==120)
   1534       {
   1535          int c0 = toOpus(compressed[0]);
   1536          if (c0<0)
   1537          {
   1538             RESTORE_STACK;
   1539             return OPUS_BAD_ARG;
   1540          }
   1541          compressed[0] = c0;
   1542       }
   1543       compressed++;
   1544       nbCompressedBytes--;
   1545    }
   1546 #else
   1547    celt_assert(st->signalling==0);
   1548 #endif
   1549 
   1550    /* Can't produce more than 1275 output bytes */
   1551    nbCompressedBytes = IMIN(nbCompressedBytes,1275);
   1552    nbAvailableBytes = nbCompressedBytes - nbFilledBytes;
   1553 
   1554    if (st->vbr && st->bitrate!=OPUS_BITRATE_MAX)
   1555    {
   1556       opus_int32 den=mode->Fs>>BITRES;
   1557       vbr_rate=(st->bitrate*frame_size+(den>>1))/den;
   1558 #ifdef CUSTOM_MODES
   1559       if (st->signalling)
   1560          vbr_rate -= 8<<BITRES;
   1561 #endif
   1562       effectiveBytes = vbr_rate>>(3+BITRES);
   1563    } else {
   1564       opus_int32 tmp;
   1565       vbr_rate = 0;
   1566       tmp = st->bitrate*frame_size;
   1567       if (tell>1)
   1568          tmp += tell;
   1569       if (st->bitrate!=OPUS_BITRATE_MAX)
   1570          nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes,
   1571                (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling));
   1572       effectiveBytes = nbCompressedBytes - nbFilledBytes;
   1573    }
   1574    equiv_rate = ((opus_int32)nbCompressedBytes*8*50 >> (3-LM)) - (40*C+20)*((400>>LM) - 50);
   1575    if (st->bitrate != OPUS_BITRATE_MAX)
   1576       equiv_rate = IMIN(equiv_rate, st->bitrate - (40*C+20)*((400>>LM) - 50));
   1577 
   1578    if (enc==NULL)
   1579    {
   1580       ec_enc_init(&_enc, compressed, nbCompressedBytes);
   1581       enc = &_enc;
   1582    }
   1583 
   1584    if (vbr_rate>0)
   1585    {
   1586       /* Computes the max bit-rate allowed in VBR mode to avoid violating the
   1587           target rate and buffering.
   1588          We must do this up front so that bust-prevention logic triggers
   1589           correctly if we don't have enough bits. */
   1590       if (st->constrained_vbr)
   1591       {
   1592          opus_int32 vbr_bound;
   1593          opus_int32 max_allowed;
   1594          /* We could use any multiple of vbr_rate as bound (depending on the
   1595              delay).
   1596             This is clamped to ensure we use at least two bytes if the encoder
   1597              was entirely empty, but to allow 0 in hybrid mode. */
   1598          vbr_bound = vbr_rate;
   1599          max_allowed = IMIN(IMAX(tell==1?2:0,
   1600                (vbr_rate+vbr_bound-st->vbr_reservoir)>>(BITRES+3)),
   1601                nbAvailableBytes);
   1602          if(max_allowed < nbAvailableBytes)
   1603          {
   1604             nbCompressedBytes = nbFilledBytes+max_allowed;
   1605             nbAvailableBytes = max_allowed;
   1606             ec_enc_shrink(enc, nbCompressedBytes);
   1607          }
   1608       }
   1609    }
   1610    total_bits = nbCompressedBytes*8;
   1611 
   1612    effEnd = end;
   1613    if (effEnd > mode->effEBands)
   1614       effEnd = mode->effEBands;
   1615 
   1616    ALLOC(in, CC*(N+overlap), celt_sig);
   1617 
   1618    sample_max=MAX32(st->overlap_max, celt_maxabs16(pcm, C*(N-overlap)/st->upsample));
   1619    st->overlap_max=celt_maxabs16(pcm+C*(N-overlap)/st->upsample, C*overlap/st->upsample);
   1620    sample_max=MAX32(sample_max, st->overlap_max);
   1621 #ifdef FIXED_POINT
   1622    silence = (sample_max==0);
   1623 #else
   1624    silence = (sample_max <= (opus_val16)1/(1<<st->lsb_depth));
   1625 #endif
   1626 #ifdef FUZZING
   1627    if ((rand()&0x3F)==0)
   1628       silence = 1;
   1629 #endif
   1630    if (tell==1)
   1631       ec_enc_bit_logp(enc, silence, 15);
   1632    else
   1633       silence=0;
   1634    if (silence)
   1635    {
   1636       /*In VBR mode there is no need to send more than the minimum. */
   1637       if (vbr_rate>0)
   1638       {
   1639          effectiveBytes=nbCompressedBytes=IMIN(nbCompressedBytes, nbFilledBytes+2);
   1640          total_bits=nbCompressedBytes*8;
   1641          nbAvailableBytes=2;
   1642          ec_enc_shrink(enc, nbCompressedBytes);
   1643       }
   1644       /* Pretend we've filled all the remaining bits with zeros
   1645             (that's what the initialiser did anyway) */
   1646       tell = nbCompressedBytes*8;
   1647       enc->nbits_total+=tell-ec_tell(enc);
   1648    }
   1649    c=0; do {
   1650       int need_clip=0;
   1651 #ifndef FIXED_POINT
   1652       need_clip = st->clip && sample_max>65536.f;
   1653 #endif
   1654       celt_preemphasis(pcm+c, in+c*(N+overlap)+overlap, N, CC, st->upsample,
   1655                   mode->preemph, st->preemph_memE+c, need_clip);
   1656    } while (++c<CC);
   1657 
   1658 
   1659 
   1660    /* Find pitch period and gain */
   1661    {
   1662       int enabled;
   1663       int qg;
   1664       enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && !hybrid && !silence && !st->disable_pf
   1665             && st->complexity >= 5;
   1666 
   1667       prefilter_tapset = st->tapset_decision;
   1668       pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes, &st->analysis);
   1669       if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) && (!st->analysis.valid || st->analysis.tonality > .3)
   1670             && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st->prefilter_period))
   1671          pitch_change = 1;
   1672       if (pf_on==0)
   1673       {
   1674          if(!hybrid && tell+16<=total_bits)
   1675             ec_enc_bit_logp(enc, 0, 1);
   1676       } else {
   1677          /*This block is not gated by a total bits check only because
   1678            of the nbAvailableBytes check above.*/
   1679          int octave;
   1680          ec_enc_bit_logp(enc, 1, 1);
   1681          pitch_index += 1;
   1682          octave = EC_ILOG(pitch_index)-5;
   1683          ec_enc_uint(enc, octave, 6);
   1684          ec_enc_bits(enc, pitch_index-(16<<octave), 4+octave);
   1685          pitch_index -= 1;
   1686          ec_enc_bits(enc, qg, 3);
   1687          ec_enc_icdf(enc, prefilter_tapset, tapset_icdf, 2);
   1688       }
   1689    }
   1690 
   1691    isTransient = 0;
   1692    shortBlocks = 0;
   1693    if (st->complexity >= 1 && !st->lfe)
   1694    {
   1695       /* Reduces the likelihood of energy instability on fricatives at low bitrate
   1696          in hybrid mode. It seems like we still want to have real transients on vowels
   1697          though (small SILK quantization offset value). */
   1698       int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.signalType != 2;
   1699       isTransient = transient_analysis(in, N+overlap, CC,
   1700             &tf_estimate, &tf_chan, allow_weak_transients, &weak_transient);
   1701    }
   1702    if (LM>0 && ec_tell(enc)+3<=total_bits)
   1703    {
   1704       if (isTransient)
   1705          shortBlocks = M;
   1706    } else {
   1707       isTransient = 0;
   1708       transient_got_disabled=1;
   1709    }
   1710 
   1711    ALLOC(freq, CC*N, celt_sig); /**< Interleaved signal MDCTs */
   1712    ALLOC(bandE,nbEBands*CC, celt_ener);
   1713    ALLOC(bandLogE,nbEBands*CC, opus_val16);
   1714 
   1715    secondMdct = shortBlocks && st->complexity>=8;
   1716    ALLOC(bandLogE2, C*nbEBands, opus_val16);
   1717    if (secondMdct)
   1718    {
   1719       compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample, st->arch);
   1720       compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
   1721       amp2Log2(mode, effEnd, end, bandE, bandLogE2, C);
   1722       for (i=0;i<C*nbEBands;i++)
   1723          bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT));
   1724    }
   1725 
   1726    compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch);
   1727    /* This should catch any NaN in the CELT input. Since we're not supposed to see any (they're filtered
   1728       at the Opus layer), just abort. */
   1729    celt_assert(!celt_isnan(freq[0]) && (C==1 || !celt_isnan(freq[N])));
   1730    if (CC==2&&C==1)
   1731       tf_chan = 0;
   1732    compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
   1733 
   1734    if (st->lfe)
   1735    {
   1736       for (i=2;i<end;i++)
   1737       {
   1738          bandE[i] = IMIN(bandE[i], MULT16_32_Q15(QCONST16(1e-4f,15),bandE[0]));
   1739          bandE[i] = MAX32(bandE[i], EPSILON);
   1740       }
   1741    }
   1742    amp2Log2(mode, effEnd, end, bandE, bandLogE, C);
   1743 
   1744    ALLOC(surround_dynalloc, C*nbEBands, opus_val16);
   1745    OPUS_CLEAR(surround_dynalloc, end);
   1746    /* This computes how much masking takes place between surround channels */
   1747    if (!hybrid&&st->energy_mask&&!st->lfe)
   1748    {
   1749       int mask_end;
   1750       int midband;
   1751       int count_dynalloc;
   1752       opus_val32 mask_avg=0;
   1753       opus_val32 diff=0;
   1754       int count=0;
   1755       mask_end = IMAX(2,st->lastCodedBands);
   1756       for (c=0;c<C;c++)
   1757       {
   1758          for(i=0;i<mask_end;i++)
   1759          {
   1760             opus_val16 mask;
   1761             mask = MAX16(MIN16(st->energy_mask[nbEBands*c+i],
   1762                    QCONST16(.25f, DB_SHIFT)), -QCONST16(2.0f, DB_SHIFT));
   1763             if (mask > 0)
   1764                mask = HALF16(mask);
   1765             mask_avg += MULT16_16(mask, eBands[i+1]-eBands[i]);
   1766             count += eBands[i+1]-eBands[i];
   1767             diff += MULT16_16(mask, 1+2*i-mask_end);
   1768          }
   1769       }
   1770       celt_assert(count>0);
   1771       mask_avg = DIV32_16(mask_avg,count);
   1772       mask_avg += QCONST16(.2f, DB_SHIFT);
   1773       diff = diff*6/(C*(mask_end-1)*(mask_end+1)*mask_end);
   1774       /* Again, being conservative */
   1775       diff = HALF32(diff);
   1776       diff = MAX32(MIN32(diff, QCONST32(.031f, DB_SHIFT)), -QCONST32(.031f, DB_SHIFT));
   1777       /* Find the band that's in the middle of the coded spectrum */
   1778       for (midband=0;eBands[midband+1] < eBands[mask_end]/2;midband++);
   1779       count_dynalloc=0;
   1780       for(i=0;i<mask_end;i++)
   1781       {
   1782          opus_val32 lin;
   1783          opus_val16 unmask;
   1784          lin = mask_avg + diff*(i-midband);
   1785          if (C==2)
   1786             unmask = MAX16(st->energy_mask[i], st->energy_mask[nbEBands+i]);
   1787          else
   1788             unmask = st->energy_mask[i];
   1789          unmask = MIN16(unmask, QCONST16(.0f, DB_SHIFT));
   1790          unmask -= lin;
   1791          if (unmask > QCONST16(.25f, DB_SHIFT))
   1792          {
   1793             surround_dynalloc[i] = unmask - QCONST16(.25f, DB_SHIFT);
   1794             count_dynalloc++;
   1795          }
   1796       }
   1797       if (count_dynalloc>=3)
   1798       {
   1799          /* If we need dynalloc in many bands, it's probably because our
   1800             initial masking rate was too low. */
   1801          mask_avg += QCONST16(.25f, DB_SHIFT);
   1802          if (mask_avg>0)
   1803          {
   1804             /* Something went really wrong in the original calculations,
   1805                disabling masking. */
   1806             mask_avg = 0;
   1807             diff = 0;
   1808             OPUS_CLEAR(surround_dynalloc, mask_end);
   1809          } else {
   1810             for(i=0;i<mask_end;i++)
   1811                surround_dynalloc[i] = MAX16(0, surround_dynalloc[i]-QCONST16(.25f, DB_SHIFT));
   1812          }
   1813       }
   1814       mask_avg += QCONST16(.2f, DB_SHIFT);
   1815       /* Convert to 1/64th units used for the trim */
   1816       surround_trim = 64*diff;
   1817       /*printf("%d %d ", mask_avg, surround_trim);*/
   1818       surround_masking = mask_avg;
   1819    }
   1820    /* Temporal VBR (but not for LFE) */
   1821    if (!st->lfe)
   1822    {
   1823       opus_val16 follow=-QCONST16(10.0f,DB_SHIFT);
   1824       opus_val32 frame_avg=0;
   1825       opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0;
   1826       for(i=start;i<end;i++)
   1827       {
   1828          follow = MAX16(follow-QCONST16(1.f, DB_SHIFT), bandLogE[i]-offset);
   1829          if (C==2)
   1830             follow = MAX16(follow, bandLogE[i+nbEBands]-offset);
   1831          frame_avg += follow;
   1832       }
   1833       frame_avg /= (end-start);
   1834       temporal_vbr = SUB16(frame_avg,st->spec_avg);
   1835       temporal_vbr = MIN16(QCONST16(3.f, DB_SHIFT), MAX16(-QCONST16(1.5f, DB_SHIFT), temporal_vbr));
   1836       st->spec_avg += MULT16_16_Q15(QCONST16(.02f, 15), temporal_vbr);
   1837    }
   1838    /*for (i=0;i<21;i++)
   1839       printf("%f ", bandLogE[i]);
   1840    printf("\n");*/
   1841 
   1842    if (!secondMdct)
   1843    {
   1844       OPUS_COPY(bandLogE2, bandLogE, C*nbEBands);
   1845    }
   1846 
   1847    /* Last chance to catch any transient we might have missed in the
   1848       time-domain analysis */
   1849    if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe && !hybrid)
   1850    {
   1851       if (patch_transient_decision(bandLogE, oldBandE, nbEBands, start, end, C))
   1852       {
   1853          isTransient = 1;
   1854          shortBlocks = M;
   1855          compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch);
   1856          compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
   1857          amp2Log2(mode, effEnd, end, bandE, bandLogE, C);
   1858          /* Compensate for the scaling of short vs long mdcts */
   1859          for (i=0;i<C*nbEBands;i++)
   1860             bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT));
   1861          tf_estimate = QCONST16(.2f,14);
   1862       }
   1863    }
   1864 
   1865    if (LM>0 && ec_tell(enc)+3<=total_bits)
   1866       ec_enc_bit_logp(enc, isTransient, 3);
   1867 
   1868    ALLOC(X, C*N, celt_norm);         /**< Interleaved normalised MDCTs */
   1869 
   1870    /* Band normalisation */
   1871    normalise_bands(mode, freq, X, bandE, effEnd, C, M);
   1872 
   1873    enable_tf_analysis = effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe;
   1874 
   1875    ALLOC(offsets, nbEBands, int);
   1876    ALLOC(importance, nbEBands, int);
   1877    ALLOC(spread_weight, nbEBands, int);
   1878 
   1879    maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets,
   1880          st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr,
   1881          eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis, importance, spread_weight);
   1882 
   1883    ALLOC(tf_res, nbEBands, int);
   1884    /* Disable variable tf resolution for hybrid and at very low bitrate */
   1885    if (enable_tf_analysis)
   1886    {
   1887       int lambda;
   1888       lambda = IMAX(80, 20480/effectiveBytes + 2);
   1889       tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan, importance);
   1890       for (i=effEnd;i<end;i++)
   1891          tf_res[i] = tf_res[effEnd-1];
   1892    } else if (hybrid && weak_transient)
   1893    {
   1894       /* For weak transients, we rely on the fact that improving time resolution using
   1895          TF on a long window is imperfect and will not result in an energy collapse at
   1896          low bitrate. */
   1897       for (i=0;i<end;i++)
   1898          tf_res[i] = 1;
   1899       tf_select=0;
   1900    } else if (hybrid && effectiveBytes<15 && st->silk_info.signalType != 2)
   1901    {
   1902       /* For low bitrate hybrid, we force temporal resolution to 5 ms rather than 2.5 ms. */
   1903       for (i=0;i<end;i++)
   1904          tf_res[i] = 0;
   1905       tf_select=isTransient;
   1906    } else {
   1907       for (i=0;i<end;i++)
   1908          tf_res[i] = isTransient;
   1909       tf_select=0;
   1910    }
   1911 
   1912    ALLOC(error, C*nbEBands, opus_val16);
   1913    c=0;
   1914    do {
   1915       for (i=start;i<end;i++)
   1916       {
   1917          /* When the energy is stable, slightly bias energy quantization towards
   1918             the previous error to make the gain more stable (a constant offset is
   1919             better than fluctuations). */
   1920          if (ABS32(SUB32(bandLogE[i+c*nbEBands], oldBandE[i+c*nbEBands])) < QCONST16(2.f, DB_SHIFT))
   1921          {
   1922             bandLogE[i+c*nbEBands] -= MULT16_16_Q15(energyError[i+c*nbEBands], QCONST16(0.25f, 15));
   1923          }
   1924       }
   1925    } while (++c < C);
   1926    quant_coarse_energy(mode, start, end, effEnd, bandLogE,
   1927          oldBandE, total_bits, error, enc,
   1928          C, LM, nbAvailableBytes, st->force_intra,
   1929          &st->delayedIntra, st->complexity >= 4, st->loss_rate, st->lfe);
   1930 
   1931    tf_encode(start, end, isTransient, tf_res, LM, tf_select, enc);
   1932 
   1933    if (ec_tell(enc)+4<=total_bits)
   1934    {
   1935       if (st->lfe)
   1936       {
   1937          st->tapset_decision = 0;
   1938          st->spread_decision = SPREAD_NORMAL;
   1939       } else if (hybrid)
   1940       {
   1941          if (st->complexity == 0)
   1942             st->spread_decision = SPREAD_NONE;
   1943          else if (isTransient)
   1944             st->spread_decision = SPREAD_NORMAL;
   1945          else
   1946             st->spread_decision = SPREAD_AGGRESSIVE;
   1947       } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C)
   1948       {
   1949          if (st->complexity == 0)
   1950             st->spread_decision = SPREAD_NONE;
   1951          else
   1952             st->spread_decision = SPREAD_NORMAL;
   1953       } else {
   1954          /* Disable new spreading+tapset estimator until we can show it works
   1955             better than the old one. So far it seems like spreading_decision()
   1956             works best. */
   1957 #if 0
   1958          if (st->analysis.valid)
   1959          {
   1960             static const opus_val16 spread_thresholds[3] = {-QCONST16(.6f, 15), -QCONST16(.2f, 15), -QCONST16(.07f, 15)};
   1961             static const opus_val16 spread_histeresis[3] = {QCONST16(.15f, 15), QCONST16(.07f, 15), QCONST16(.02f, 15)};
   1962             static const opus_val16 tapset_thresholds[2] = {QCONST16(.0f, 15), QCONST16(.15f, 15)};
   1963             static const opus_val16 tapset_histeresis[2] = {QCONST16(.1f, 15), QCONST16(.05f, 15)};
   1964             st->spread_decision = hysteresis_decision(-st->analysis.tonality, spread_thresholds, spread_histeresis, 3, st->spread_decision);
   1965             st->tapset_decision = hysteresis_decision(st->analysis.tonality_slope, tapset_thresholds, tapset_histeresis, 2, st->tapset_decision);
   1966          } else
   1967 #endif
   1968          {
   1969             st->spread_decision = spreading_decision(mode, X,
   1970                   &st->tonal_average, st->spread_decision, &st->hf_average,
   1971                   &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M, spread_weight);
   1972          }
   1973          /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/
   1974          /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision, st->analysis.tonality_slope, st->tapset_decision);*/
   1975       }
   1976       ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5);
   1977    }
   1978 
   1979    /* For LFE, everything interesting is in the first band */
   1980    if (st->lfe)
   1981       offsets[0] = IMIN(8, effectiveBytes/3);
   1982    ALLOC(cap, nbEBands, int);
   1983    init_caps(mode,cap,LM,C);
   1984 
   1985    dynalloc_logp = 6;
   1986    total_bits<<=BITRES;
   1987    total_boost = 0;
   1988    tell = ec_tell_frac(enc);
   1989    for (i=start;i<end;i++)
   1990    {
   1991       int width, quanta;
   1992       int dynalloc_loop_logp;
   1993       int boost;
   1994       int j;
   1995       width = C*(eBands[i+1]-eBands[i])<<LM;
   1996       /* quanta is 6 bits, but no more than 1 bit/sample
   1997          and no less than 1/8 bit/sample */
   1998       quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width));
   1999       dynalloc_loop_logp = dynalloc_logp;
   2000       boost = 0;
   2001       for (j = 0; tell+(dynalloc_loop_logp<<BITRES) < total_bits-total_boost
   2002             && boost < cap[i]; j++)
   2003       {
   2004          int flag;
   2005          flag = j<offsets[i];
   2006          ec_enc_bit_logp(enc, flag, dynalloc_loop_logp);
   2007          tell = ec_tell_frac(enc);
   2008          if (!flag)
   2009             break;
   2010          boost += quanta;
   2011          total_boost += quanta;
   2012          dynalloc_loop_logp = 1;
   2013       }
   2014       /* Making dynalloc more likely */
   2015       if (j)
   2016          dynalloc_logp = IMAX(2, dynalloc_logp-1);
   2017       offsets[i] = boost;
   2018    }
   2019 
   2020    if (C==2)
   2021    {
   2022       static const opus_val16 intensity_thresholds[21]=
   2023       /* 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19  20  off*/
   2024         {  1, 2, 3, 4, 5, 6, 7, 8,16,24,36,44,50,56,62,67,72,79,88,106,134};
   2025       static const opus_val16 intensity_histeresis[21]=
   2026         {  1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 4, 5, 6,  8, 8};
   2027 
   2028       /* Always use MS for 2.5 ms frames until we can do a better analysis */
   2029       if (LM!=0)
   2030          dual_stereo = stereo_analysis(mode, X, LM, N);
   2031 
   2032       st->intensity = hysteresis_decision((opus_val16)(equiv_rate/1000),
   2033             intensity_thresholds, intensity_histeresis, 21, st->intensity);
   2034       st->intensity = IMIN(end,IMAX(start, st->intensity));
   2035    }
   2036 
   2037    alloc_trim = 5;
   2038    if (tell+(6<<BITRES) <= total_bits - total_boost)
   2039    {
   2040       if (start > 0 || st->lfe)
   2041       {
   2042          st->stereo_saving = 0;
   2043          alloc_trim = 5;
   2044       } else {
   2045          alloc_trim = alloc_trim_analysis(mode, X, bandLogE,
   2046             end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate,
   2047             st->intensity, surround_trim, equiv_rate, st->arch);
   2048       }
   2049       ec_enc_icdf(enc, alloc_trim, trim_icdf, 7);
   2050       tell = ec_tell_frac(enc);
   2051    }
   2052 
   2053    /* Variable bitrate */
   2054    if (vbr_rate>0)
   2055    {
   2056      opus_val16 alpha;
   2057      opus_int32 delta;
   2058      /* The target rate in 8th bits per frame */
   2059      opus_int32 target, base_target;
   2060      opus_int32 min_allowed;
   2061      int lm_diff = mode->maxLM - LM;
   2062 
   2063      /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms.
   2064         The CELT allocator will just not be able to use more than that anyway. */
   2065      nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM));
   2066      if (!hybrid)
   2067      {
   2068         base_target = vbr_rate - ((40*C+20)<<BITRES);
   2069      } else {
   2070         base_target = IMAX(0, vbr_rate - ((9*C+4)<<BITRES));
   2071      }
   2072 
   2073      if (st->constrained_vbr)
   2074         base_target += (st->vbr_offset>>lm_diff);
   2075 
   2076      if (!hybrid)
   2077      {
   2078         target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate,
   2079            st->lastCodedBands, C, st->intensity, st->constrained_vbr,
   2080            st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth,
   2081            st->lfe, st->energy_mask!=NULL, surround_masking,
   2082            temporal_vbr);
   2083      } else {
   2084         target = base_target;
   2085         /* Tonal frames (offset<100) need more bits than noisy (offset>100) ones. */
   2086         if (st->silk_info.offset < 100) target += 12 << BITRES >> (3-LM);
   2087         if (st->silk_info.offset > 100) target -= 18 << BITRES >> (3-LM);
   2088         /* Boosting bitrate on transients and vowels with significant temporal
   2089            spikes. */
   2090         target += (opus_int32)MULT16_16_Q14(tf_estimate-QCONST16(.25f,14), (50<<BITRES));
   2091         /* If we have a strong transient, let's make sure it has enough bits to code
   2092            the first two bands, so that it can use folding rather than noise. */
   2093         if (tf_estimate > QCONST16(.7f,14))
   2094            target = IMAX(target, 50<<BITRES);
   2095      }
   2096      /* The current offset is removed from the target and the space used
   2097         so far is added*/
   2098      target=target+tell;
   2099      /* In VBR mode the frame size must not be reduced so much that it would
   2100          result in the encoder running out of bits.
   2101         The margin of 2 bytes ensures that none of the bust-prevention logic
   2102          in the decoder will have triggered so far. */
   2103      min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2;
   2104      /* Take into account the 37 bits we need to have left in the packet to
   2105         signal a redundant frame in hybrid mode. Creating a shorter packet would
   2106         create an entropy coder desync. */
   2107      if (hybrid)
   2108         min_allowed = IMAX(min_allowed, (tell0_frac+(37<<BITRES)+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3));
   2109 
   2110      nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3);
   2111      nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes);
   2112      nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes);
   2113 
   2114      /* By how much did we "miss" the target on that frame */
   2115      delta = target - vbr_rate;
   2116 
   2117      target=nbAvailableBytes<<(BITRES+3);
   2118 
   2119      /*If the frame is silent we don't adjust our drift, otherwise
   2120        the encoder will shoot to very high rates after hitting a
   2121        span of silence, but we do allow the bitres to refill.
   2122        This means that we'll undershoot our target in CVBR/VBR modes
   2123        on files with lots of silence. */
   2124      if(silence)
   2125      {
   2126        nbAvailableBytes = 2;
   2127        target = 2*8<<BITRES;
   2128        delta = 0;
   2129      }
   2130 
   2131      if (st->vbr_count < 970)
   2132      {
   2133         st->vbr_count++;
   2134         alpha = celt_rcp(SHL32(EXTEND32(st->vbr_count+20),16));
   2135      } else
   2136         alpha = QCONST16(.001f,15);
   2137      /* How many bits have we used in excess of what we're allowed */
   2138      if (st->constrained_vbr)
   2139         st->vbr_reservoir += target - vbr_rate;
   2140      /*printf ("%d\n", st->vbr_reservoir);*/
   2141 
   2142      /* Compute the offset we need to apply in order to reach the target */
   2143      if (st->constrained_vbr)
   2144      {
   2145         st->vbr_drift += (opus_int32)MULT16_32_Q15(alpha,(delta*(1<<lm_diff))-st->vbr_offset-st->vbr_drift);
   2146         st->vbr_offset = -st->vbr_drift;
   2147      }
   2148      /*printf ("%d\n", st->vbr_drift);*/
   2149 
   2150      if (st->constrained_vbr && st->vbr_reservoir < 0)
   2151      {
   2152         /* We're under the min value -- increase rate */
   2153         int adjust = (-st->vbr_reservoir)/(8<<BITRES);
   2154         /* Unless we're just coding silence */
   2155         nbAvailableBytes += silence?0:adjust;
   2156         st->vbr_reservoir = 0;
   2157         /*printf ("+%d\n", adjust);*/
   2158      }
   2159      nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes);
   2160      /*printf("%d\n", nbCompressedBytes*50*8);*/
   2161      /* This moves the raw bits to take into account the new compressed size */
   2162      ec_enc_shrink(enc, nbCompressedBytes);
   2163    }
   2164 
   2165    /* Bit allocation */
   2166    ALLOC(fine_quant, nbEBands, int);
   2167    ALLOC(pulses, nbEBands, int);
   2168    ALLOC(fine_priority, nbEBands, int);
   2169 
   2170    /* bits =           packet size                    - where we are - safety*/
   2171    bits = (((opus_int32)nbCompressedBytes*8)<<BITRES) - ec_tell_frac(enc) - 1;
   2172    anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES) : 0;
   2173    bits -= anti_collapse_rsv;
   2174    signalBandwidth = end-1;
   2175 #ifndef DISABLE_FLOAT_API
   2176    if (st->analysis.valid)
   2177    {
   2178       int min_bandwidth;
   2179       if (equiv_rate < (opus_int32)32000*C)
   2180          min_bandwidth = 13;
   2181       else if (equiv_rate < (opus_int32)48000*C)
   2182          min_bandwidth = 16;
   2183       else if (equiv_rate < (opus_int32)60000*C)
   2184          min_bandwidth = 18;
   2185       else  if (equiv_rate < (opus_int32)80000*C)
   2186          min_bandwidth = 19;
   2187       else
   2188          min_bandwidth = 20;
   2189       signalBandwidth = IMAX(st->analysis.bandwidth, min_bandwidth);
   2190    }
   2191 #endif
   2192    if (st->lfe)
   2193       signalBandwidth = 1;
   2194    codedBands = clt_compute_allocation(mode, start, end, offsets, cap,
   2195          alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses,
   2196          fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBandwidth);
   2197    if (st->lastCodedBands)
   2198       st->lastCodedBands = IMIN(st->lastCodedBands+1,IMAX(st->lastCodedBands-1,codedBands));
   2199    else
   2200       st->lastCodedBands = codedBands;
   2201 
   2202    quant_fine_energy(mode, start, end, oldBandE, error, fine_quant, enc, C);
   2203 
   2204    /* Residual quantisation */
   2205    ALLOC(collapse_masks, C*nbEBands, unsigned char);
   2206    quant_all_bands(1, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks,
   2207          bandE, pulses, shortBlocks, st->spread_decision,
   2208          dual_stereo, st->intensity, tf_res, nbCompressedBytes*(8<<BITRES)-anti_collapse_rsv,
   2209          balance, enc, LM, codedBands, &st->rng, st->complexity, st->arch, st->disable_inv);
   2210 
   2211    if (anti_collapse_rsv > 0)
   2212    {
   2213       anti_collapse_on = st->consec_transient<2;
   2214 #ifdef FUZZING
   2215       anti_collapse_on = rand()&0x1;
   2216 #endif
   2217       ec_enc_bits(enc, anti_collapse_on, 1);
   2218    }
   2219    quant_energy_finalise(mode, start, end, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C);
   2220    OPUS_CLEAR(energyError, nbEBands*CC);
   2221    c=0;
   2222    do {
   2223       for (i=start;i<end;i++)
   2224       {
   2225          energyError[i+c*nbEBands] = MAX16(-QCONST16(0.5f, 15), MIN16(QCONST16(0.5f, 15), error[i+c*nbEBands]));
   2226       }
   2227    } while (++c < C);
   2228 
   2229    if (silence)
   2230    {
   2231       for (i=0;i<C*nbEBands;i++)
   2232          oldBandE[i] = -QCONST16(28.f,DB_SHIFT);
   2233    }
   2234 
   2235 #ifdef RESYNTH
   2236    /* Re-synthesis of the coded audio if required */
   2237    {
   2238       celt_sig *out_mem[2];
   2239 
   2240       if (anti_collapse_on)
   2241       {
   2242          anti_collapse(mode, X, collapse_masks, LM, C, N,
   2243                start, end, oldBandE, oldLogE, oldLogE2, pulses, st->rng);
   2244       }
   2245 
   2246       c=0; do {
   2247          OPUS_MOVE(st->syn_mem[c], st->syn_mem[c]+N, 2*MAX_PERIOD-N+overlap/2);
   2248       } while (++c<CC);
   2249 
   2250       c=0; do {
   2251          out_mem[c] = st->syn_mem[c]+2*MAX_PERIOD-N;
   2252       } while (++c<CC);
   2253 
   2254       celt_synthesis(mode, X, out_mem, oldBandE, start, effEnd,
   2255                      C, CC, isTransient, LM, st->upsample, silence, st->arch);
   2256 
   2257       c=0; do {
   2258          st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD);
   2259          st->prefilter_period_old=IMAX(st->prefilter_period_old, COMBFILTER_MINPERIOD);
   2260          comb_filter(out_mem[c], out_mem[c], st->prefilter_period_old, st->prefilter_period, mode->shortMdctSize,
   2261                st->prefilter_gain_old, st->prefilter_gain, st->prefilter_tapset_old, st->prefilter_tapset,
   2262                mode->window, overlap);
   2263          if (LM!=0)
   2264             comb_filter(out_mem[c]+mode->shortMdctSize, out_mem[c]+mode->shortMdctSize, st->prefilter_period, pitch_index, N-mode->shortMdctSize,
   2265                   st->prefilter_gain, gain1, st->prefilter_tapset, prefilter_tapset,
   2266                   mode->window, overlap);
   2267       } while (++c<CC);
   2268 
   2269       /* We reuse freq[] as scratch space for the de-emphasis */
   2270       deemphasis(out_mem, (opus_val16*)pcm, N, CC, st->upsample, mode->preemph, st->preemph_memD);
   2271       st->prefilter_period_old = st->prefilter_period;
   2272       st->prefilter_gain_old = st->prefilter_gain;
   2273       st->prefilter_tapset_old = st->prefilter_tapset;
   2274    }
   2275 #endif
   2276 
   2277    st->prefilter_period = pitch_index;
   2278    st->prefilter_gain = gain1;
   2279    st->prefilter_tapset = prefilter_tapset;
   2280 #ifdef RESYNTH
   2281    if (LM!=0)
   2282    {
   2283       st->prefilter_period_old = st->prefilter_period;
   2284       st->prefilter_gain_old = st->prefilter_gain;
   2285       st->prefilter_tapset_old = st->prefilter_tapset;
   2286    }
   2287 #endif
   2288 
   2289    if (CC==2&&C==1) {
   2290       OPUS_COPY(&oldBandE[nbEBands], oldBandE, nbEBands);
   2291    }
   2292 
   2293    if (!isTransient)
   2294    {
   2295       OPUS_COPY(oldLogE2, oldLogE, CC*nbEBands);
   2296       OPUS_COPY(oldLogE, oldBandE, CC*nbEBands);
   2297    } else {
   2298       for (i=0;i<CC*nbEBands;i++)
   2299          oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]);
   2300    }
   2301    /* In case start or end were to change */
   2302    c=0; do
   2303    {
   2304       for (i=0;i<start;i++)
   2305       {
   2306          oldBandE[c*nbEBands+i]=0;
   2307          oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
   2308       }
   2309       for (i=end;i<nbEBands;i++)
   2310       {
   2311          oldBandE[c*nbEBands+i]=0;
   2312          oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
   2313       }
   2314    } while (++c<CC);
   2315 
   2316    if (isTransient || transient_got_disabled)
   2317       st->consec_transient++;
   2318    else
   2319       st->consec_transient=0;
   2320    st->rng = enc->rng;
   2321 
   2322    /* If there's any room left (can only happen for very high rates),
   2323       it's already filled with zeros */
   2324    ec_enc_done(enc);
   2325 
   2326 #ifdef CUSTOM_MODES
   2327    if (st->signalling)
   2328       nbCompressedBytes++;
   2329 #endif
   2330 
   2331    RESTORE_STACK;
   2332    if (ec_get_error(enc))
   2333       return OPUS_INTERNAL_ERROR;
   2334    else
   2335       return nbCompressedBytes;
   2336 }
   2337 
   2338 
   2339 #ifdef CUSTOM_MODES
   2340 
   2341 #ifdef FIXED_POINT
   2342 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
   2343 {
   2344    return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL);
   2345 }
   2346 
   2347 #ifndef DISABLE_FLOAT_API
   2348 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
   2349 {
   2350    int j, ret, C, N;
   2351    VARDECL(opus_int16, in);
   2352    ALLOC_STACK;
   2353 
   2354    if (pcm==NULL)
   2355       return OPUS_BAD_ARG;
   2356 
   2357    C = st->channels;
   2358    N = frame_size;
   2359    ALLOC(in, C*N, opus_int16);
   2360 
   2361    for (j=0;j<C*N;j++)
   2362      in[j] = FLOAT2INT16(pcm[j]);
   2363 
   2364    ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL);
   2365 #ifdef RESYNTH
   2366    for (j=0;j<C*N;j++)
   2367       ((float*)pcm)[j]=in[j]*(1.f/32768.f);
   2368 #endif
   2369    RESTORE_STACK;
   2370    return ret;
   2371 }
   2372 #endif /* DISABLE_FLOAT_API */
   2373 #else
   2374 
   2375 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
   2376 {
   2377    int j, ret, C, N;
   2378    VARDECL(celt_sig, in);
   2379    ALLOC_STACK;
   2380 
   2381    if (pcm==NULL)
   2382       return OPUS_BAD_ARG;
   2383 
   2384    C=st->channels;
   2385    N=frame_size;
   2386    ALLOC(in, C*N, celt_sig);
   2387    for (j=0;j<C*N;j++) {
   2388      in[j] = SCALEOUT(pcm[j]);
   2389    }
   2390 
   2391    ret = celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL);
   2392 #ifdef RESYNTH
   2393    for (j=0;j<C*N;j++)
   2394       ((opus_int16*)pcm)[j] = FLOAT2INT16(in[j]);
   2395 #endif
   2396    RESTORE_STACK;
   2397    return ret;
   2398 }
   2399 
   2400 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
   2401 {
   2402    return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL);
   2403 }
   2404 
   2405 #endif
   2406 
   2407 #endif /* CUSTOM_MODES */
   2408 
   2409 int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...)
   2410 {
   2411    va_list ap;
   2412 
   2413    va_start(ap, request);
   2414    switch (request)
   2415    {
   2416       case OPUS_SET_COMPLEXITY_REQUEST:
   2417       {
   2418          int value = va_arg(ap, opus_int32);
   2419          if (value<0 || value>10)
   2420             goto bad_arg;
   2421          st->complexity = value;
   2422       }
   2423       break;
   2424       case CELT_SET_START_BAND_REQUEST:
   2425       {
   2426          opus_int32 value = va_arg(ap, opus_int32);
   2427          if (value<0 || value>=st->mode->nbEBands)
   2428             goto bad_arg;
   2429          st->start = value;
   2430       }
   2431       break;
   2432       case CELT_SET_END_BAND_REQUEST:
   2433       {
   2434          opus_int32 value = va_arg(ap, opus_int32);
   2435          if (value<1 || value>st->mode->nbEBands)
   2436             goto bad_arg;
   2437          st->end = value;
   2438       }
   2439       break;
   2440       case CELT_SET_PREDICTION_REQUEST:
   2441       {
   2442          int value = va_arg(ap, opus_int32);
   2443          if (value<0 || value>2)
   2444             goto bad_arg;
   2445          st->disable_pf = value<=1;
   2446          st->force_intra = value==0;
   2447       }
   2448       break;
   2449       case OPUS_SET_PACKET_LOSS_PERC_REQUEST:
   2450       {
   2451          int value = va_arg(ap, opus_int32);
   2452          if (value<0 || value>100)
   2453             goto bad_arg;
   2454          st->loss_rate = value;
   2455       }
   2456       break;
   2457       case OPUS_SET_VBR_CONSTRAINT_REQUEST:
   2458       {
   2459          opus_int32 value = va_arg(ap, opus_int32);
   2460          st->constrained_vbr = value;
   2461       }
   2462       break;
   2463       case OPUS_SET_VBR_REQUEST:
   2464       {
   2465          opus_int32 value = va_arg(ap, opus_int32);
   2466          st->vbr = value;
   2467       }
   2468       break;
   2469       case OPUS_SET_BITRATE_REQUEST:
   2470       {
   2471          opus_int32 value = va_arg(ap, opus_int32);
   2472          if (value<=500 && value!=OPUS_BITRATE_MAX)
   2473             goto bad_arg;
   2474          value = IMIN(value, 260000*st->channels);
   2475          st->bitrate = value;
   2476       }
   2477       break;
   2478       case CELT_SET_CHANNELS_REQUEST:
   2479       {
   2480          opus_int32 value = va_arg(ap, opus_int32);
   2481          if (value<1 || value>2)
   2482             goto bad_arg;
   2483          st->stream_channels = value;
   2484       }
   2485       break;
   2486       case OPUS_SET_LSB_DEPTH_REQUEST:
   2487       {
   2488           opus_int32 value = va_arg(ap, opus_int32);
   2489           if (value<8 || value>24)
   2490              goto bad_arg;
   2491           st->lsb_depth=value;
   2492       }
   2493       break;
   2494       case OPUS_GET_LSB_DEPTH_REQUEST:
   2495       {
   2496           opus_int32 *value = va_arg(ap, opus_int32*);
   2497           *value=st->lsb_depth;
   2498       }
   2499       break;
   2500       case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST:
   2501       {
   2502           opus_int32 value = va_arg(ap, opus_int32);
   2503           if(value<0 || value>1)
   2504           {
   2505              goto bad_arg;
   2506           }
   2507           st->disable_inv = value;
   2508       }
   2509       break;
   2510       case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST:
   2511       {
   2512           opus_int32 *value = va_arg(ap, opus_int32*);
   2513           if (!value)
   2514           {
   2515              goto bad_arg;
   2516           }
   2517           *value = st->disable_inv;
   2518       }
   2519       break;
   2520       case OPUS_RESET_STATE:
   2521       {
   2522          int i;
   2523          opus_val16 *oldBandE, *oldLogE, *oldLogE2;
   2524          oldBandE = (opus_val16*)(st->in_mem+st->channels*(st->mode->overlap+COMBFILTER_MAXPERIOD));
   2525          oldLogE = oldBandE + st->channels*st->mode->nbEBands;
   2526          oldLogE2 = oldLogE + st->channels*st->mode->nbEBands;
   2527          OPUS_CLEAR((char*)&st->ENCODER_RESET_START,
   2528                opus_custom_encoder_get_size(st->mode, st->channels)-
   2529                ((char*)&st->ENCODER_RESET_START - (char*)st));
   2530          for (i=0;i<st->channels*st->mode->nbEBands;i++)
   2531             oldLogE[i]=oldLogE2[i]=-QCONST16(28.f,DB_SHIFT);
   2532          st->vbr_offset = 0;
   2533          st->delayedIntra = 1;
   2534          st->spread_decision = SPREAD_NORMAL;
   2535          st->tonal_average = 256;
   2536          st->hf_average = 0;
   2537          st->tapset_decision = 0;
   2538       }
   2539       break;
   2540 #ifdef CUSTOM_MODES
   2541       case CELT_SET_INPUT_CLIPPING_REQUEST:
   2542       {
   2543          opus_int32 value = va_arg(ap, opus_int32);
   2544          st->clip = value;
   2545       }
   2546       break;
   2547 #endif
   2548       case CELT_SET_SIGNALLING_REQUEST:
   2549       {
   2550          opus_int32 value = va_arg(ap, opus_int32);
   2551          st->signalling = value;
   2552       }
   2553       break;
   2554       case CELT_SET_ANALYSIS_REQUEST:
   2555       {
   2556          AnalysisInfo *info = va_arg(ap, AnalysisInfo *);
   2557          if (info)
   2558             OPUS_COPY(&st->analysis, info, 1);
   2559       }
   2560       break;
   2561       case CELT_SET_SILK_INFO_REQUEST:
   2562       {
   2563          SILKInfo *info = va_arg(ap, SILKInfo *);
   2564          if (info)
   2565             OPUS_COPY(&st->silk_info, info, 1);
   2566       }
   2567       break;
   2568       case CELT_GET_MODE_REQUEST:
   2569       {
   2570          const CELTMode ** value = va_arg(ap, const CELTMode**);
   2571          if (value==0)
   2572             goto bad_arg;
   2573          *value=st->mode;
   2574       }
   2575       break;
   2576       case OPUS_GET_FINAL_RANGE_REQUEST:
   2577       {
   2578          opus_uint32 * value = va_arg(ap, opus_uint32 *);
   2579          if (value==0)
   2580             goto bad_arg;
   2581          *value=st->rng;
   2582       }
   2583       break;
   2584       case OPUS_SET_LFE_REQUEST:
   2585       {
   2586           opus_int32 value = va_arg(ap, opus_int32);
   2587           st->lfe = value;
   2588       }
   2589       break;
   2590       case OPUS_SET_ENERGY_MASK_REQUEST:
   2591       {
   2592           opus_val16 *value = va_arg(ap, opus_val16*);
   2593           st->energy_mask = value;
   2594       }
   2595       break;
   2596       default:
   2597          goto bad_request;
   2598    }
   2599    va_end(ap);
   2600    return OPUS_OK;
   2601 bad_arg:
   2602    va_end(ap);
   2603    return OPUS_BAD_ARG;
   2604 bad_request:
   2605    va_end(ap);
   2606    return OPUS_UNIMPLEMENTED;
   2607 }
   2608