Chris@69: /* Copyright (c) 2007-2008 CSIRO
Chris@69:    Copyright (c) 2007-2009 Xiph.Org Foundation
Chris@69:    Copyright (c) 2008-2009 Gregory Maxwell
Chris@69:    Written by Jean-Marc Valin and Gregory Maxwell */
Chris@69: /*
Chris@69:    Redistribution and use in source and binary forms, with or without
Chris@69:    modification, are permitted provided that the following conditions
Chris@69:    are met:
Chris@69: 
Chris@69:    - Redistributions of source code must retain the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer.
Chris@69: 
Chris@69:    - Redistributions in binary form must reproduce the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer in the
Chris@69:    documentation and/or other materials provided with the distribution.
Chris@69: 
Chris@69:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
Chris@69:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
Chris@69:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
Chris@69:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
Chris@69:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
Chris@69:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
Chris@69:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
Chris@69:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
Chris@69:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
Chris@69:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
Chris@69:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Chris@69: */
Chris@69: 
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #include <math.h>
Chris@69: #include "bands.h"
Chris@69: #include "modes.h"
Chris@69: #include "vq.h"
Chris@69: #include "cwrs.h"
Chris@69: #include "stack_alloc.h"
Chris@69: #include "os_support.h"
Chris@69: #include "mathops.h"
Chris@69: #include "rate.h"
Chris@69: #include "quant_bands.h"
Chris@69: #include "pitch.h"
Chris@69: 
Chris@69: int hysteresis_decision(opus_val16 val, const opus_val16 *thresholds, const opus_val16 *hysteresis, int N, int prev)
Chris@69: {
Chris@69:    int i;
Chris@69:    for (i=0;i<N;i++)
Chris@69:    {
Chris@69:       if (val < thresholds[i])
Chris@69:          break;
Chris@69:    }
Chris@69:    if (i>prev && val < thresholds[prev]+hysteresis[prev])
Chris@69:       i=prev;
Chris@69:    if (i<prev && val > thresholds[prev-1]-hysteresis[prev-1])
Chris@69:       i=prev;
Chris@69:    return i;
Chris@69: }
Chris@69: 
Chris@69: opus_uint32 celt_lcg_rand(opus_uint32 seed)
Chris@69: {
Chris@69:    return 1664525 * seed + 1013904223;
Chris@69: }
Chris@69: 
Chris@69: /* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness
Chris@69:    with this approximation is important because it has an impact on the bit allocation */
Chris@69: opus_int16 bitexact_cos(opus_int16 x)
Chris@69: {
Chris@69:    opus_int32 tmp;
Chris@69:    opus_int16 x2;
Chris@69:    tmp = (4096+((opus_int32)(x)*(x)))>>13;
Chris@69:    celt_sig_assert(tmp<=32767);
Chris@69:    x2 = tmp;
Chris@69:    x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2)))));
Chris@69:    celt_sig_assert(x2<=32766);
Chris@69:    return 1+x2;
Chris@69: }
Chris@69: 
Chris@69: int bitexact_log2tan(int isin,int icos)
Chris@69: {
Chris@69:    int lc;
Chris@69:    int ls;
Chris@69:    lc=EC_ILOG(icos);
Chris@69:    ls=EC_ILOG(isin);
Chris@69:    icos<<=15-lc;
Chris@69:    isin<<=15-ls;
Chris@69:    return (ls-lc)*(1<<11)
Chris@69:          +FRAC_MUL16(isin, FRAC_MUL16(isin, -2597) + 7932)
Chris@69:          -FRAC_MUL16(icos, FRAC_MUL16(icos, -2597) + 7932);
Chris@69: }
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69: /* Compute the amplitude (sqrt energy) in each of the bands */
Chris@69: void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch)
Chris@69: {
Chris@69:    int i, c, N;
Chris@69:    const opus_int16 *eBands = m->eBands;
Chris@69:    (void)arch;
Chris@69:    N = m->shortMdctSize<<LM;
Chris@69:    c=0; do {
Chris@69:       for (i=0;i<end;i++)
Chris@69:       {
Chris@69:          int j;
Chris@69:          opus_val32 maxval=0;
Chris@69:          opus_val32 sum = 0;
Chris@69: 
Chris@69:          maxval = celt_maxabs32(&X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM);
Chris@69:          if (maxval > 0)
Chris@69:          {
Chris@69:             int shift = celt_ilog2(maxval) - 14 + (((m->logN[i]>>BITRES)+LM+1)>>1);
Chris@69:             j=eBands[i]<<LM;
Chris@69:             if (shift>0)
Chris@69:             {
Chris@69:                do {
Chris@69:                   sum = MAC16_16(sum, EXTRACT16(SHR32(X[j+c*N],shift)),
Chris@69:                         EXTRACT16(SHR32(X[j+c*N],shift)));
Chris@69:                } while (++j<eBands[i+1]<<LM);
Chris@69:             } else {
Chris@69:                do {
Chris@69:                   sum = MAC16_16(sum, EXTRACT16(SHL32(X[j+c*N],-shift)),
Chris@69:                         EXTRACT16(SHL32(X[j+c*N],-shift)));
Chris@69:                } while (++j<eBands[i+1]<<LM);
Chris@69:             }
Chris@69:             /* We're adding one here to ensure the normalized band isn't larger than unity norm */
Chris@69:             bandE[i+c*m->nbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
Chris@69:          } else {
Chris@69:             bandE[i+c*m->nbEBands] = EPSILON;
Chris@69:          }
Chris@69:          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
Chris@69:       }
Chris@69:    } while (++c<C);
Chris@69:    /*printf ("\n");*/
Chris@69: }
Chris@69: 
Chris@69: /* Normalise each band such that the energy is one. */
Chris@69: void normalise_bands(const CELTMode *m, const celt_sig * OPUS_RESTRICT freq, celt_norm * OPUS_RESTRICT X, const celt_ener *bandE, int end, int C, int M)
Chris@69: {
Chris@69:    int i, c, N;
Chris@69:    const opus_int16 *eBands = m->eBands;
Chris@69:    N = M*m->shortMdctSize;
Chris@69:    c=0; do {
Chris@69:       i=0; do {
Chris@69:          opus_val16 g;
Chris@69:          int j,shift;
Chris@69:          opus_val16 E;
Chris@69:          shift = celt_zlog2(bandE[i+c*m->nbEBands])-13;
Chris@69:          E = VSHR32(bandE[i+c*m->nbEBands], shift);
Chris@69:          g = EXTRACT16(celt_rcp(SHL32(E,3)));
Chris@69:          j=M*eBands[i]; do {
Chris@69:             X[j+c*N] = MULT16_16_Q15(VSHR32(freq[j+c*N],shift-1),g);
Chris@69:          } while (++j<M*eBands[i+1]);
Chris@69:       } while (++i<end);
Chris@69:    } while (++c<C);
Chris@69: }
Chris@69: 
Chris@69: #else /* FIXED_POINT */
Chris@69: /* Compute the amplitude (sqrt energy) in each of the bands */
Chris@69: void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch)
Chris@69: {
Chris@69:    int i, c, N;
Chris@69:    const opus_int16 *eBands = m->eBands;
Chris@69:    N = m->shortMdctSize<<LM;
Chris@69:    c=0; do {
Chris@69:       for (i=0;i<end;i++)
Chris@69:       {
Chris@69:          opus_val32 sum;
Chris@69:          sum = 1e-27f + celt_inner_prod(&X[c*N+(eBands[i]<<LM)], &X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM, arch);
Chris@69:          bandE[i+c*m->nbEBands] = celt_sqrt(sum);
Chris@69:          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
Chris@69:       }
Chris@69:    } while (++c<C);
Chris@69:    /*printf ("\n");*/
Chris@69: }
Chris@69: 
Chris@69: /* Normalise each band such that the energy is one. */
Chris@69: void normalise_bands(const CELTMode *m, const celt_sig * OPUS_RESTRICT freq, celt_norm * OPUS_RESTRICT X, const celt_ener *bandE, int end, int C, int M)
Chris@69: {
Chris@69:    int i, c, N;
Chris@69:    const opus_int16 *eBands = m->eBands;
Chris@69:    N = M*m->shortMdctSize;
Chris@69:    c=0; do {
Chris@69:       for (i=0;i<end;i++)
Chris@69:       {
Chris@69:          int j;
Chris@69:          opus_val16 g = 1.f/(1e-27f+bandE[i+c*m->nbEBands]);
Chris@69:          for (j=M*eBands[i];j<M*eBands[i+1];j++)
Chris@69:             X[j+c*N] = freq[j+c*N]*g;
Chris@69:       }
Chris@69:    } while (++c<C);
Chris@69: }
Chris@69: 
Chris@69: #endif /* FIXED_POINT */
Chris@69: 
Chris@69: /* De-normalise the energy to produce the synthesis from the unit-energy bands */
Chris@69: void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X,
Chris@69:       celt_sig * OPUS_RESTRICT freq, const opus_val16 *bandLogE, int start,
Chris@69:       int end, int M, int downsample, int silence)
Chris@69: {
Chris@69:    int i, N;
Chris@69:    int bound;
Chris@69:    celt_sig * OPUS_RESTRICT f;
Chris@69:    const celt_norm * OPUS_RESTRICT x;
Chris@69:    const opus_int16 *eBands = m->eBands;
Chris@69:    N = M*m->shortMdctSize;
Chris@69:    bound = M*eBands[end];
Chris@69:    if (downsample!=1)
Chris@69:       bound = IMIN(bound, N/downsample);
Chris@69:    if (silence)
Chris@69:    {
Chris@69:       bound = 0;
Chris@69:       start = end = 0;
Chris@69:    }
Chris@69:    f = freq;
Chris@69:    x = X+M*eBands[start];
Chris@69:    for (i=0;i<M*eBands[start];i++)
Chris@69:       *f++ = 0;
Chris@69:    for (i=start;i<end;i++)
Chris@69:    {
Chris@69:       int j, band_end;
Chris@69:       opus_val16 g;
Chris@69:       opus_val16 lg;
Chris@69: #ifdef FIXED_POINT
Chris@69:       int shift;
Chris@69: #endif
Chris@69:       j=M*eBands[i];
Chris@69:       band_end = M*eBands[i+1];
Chris@69:       lg = SATURATE16(ADD32(bandLogE[i], SHL32((opus_val32)eMeans[i],6)));
Chris@69: #ifndef FIXED_POINT
Chris@69:       g = celt_exp2(MIN32(32.f, lg));
Chris@69: #else
Chris@69:       /* Handle the integer part of the log energy */
Chris@69:       shift = 16-(lg>>DB_SHIFT);
Chris@69:       if (shift>31)
Chris@69:       {
Chris@69:          shift=0;
Chris@69:          g=0;
Chris@69:       } else {
Chris@69:          /* Handle the fractional part. */
Chris@69:          g = celt_exp2_frac(lg&((1<<DB_SHIFT)-1));
Chris@69:       }
Chris@69:       /* Handle extreme gains with negative shift. */
Chris@69:       if (shift<0)
Chris@69:       {
Chris@69:          /* For shift <= -2 and g > 16384 we'd be likely to overflow, so we're
Chris@69:             capping the gain here, which is equivalent to a cap of 18 on lg.
Chris@69:             This shouldn't trigger unless the bitstream is already corrupted. */
Chris@69:          if (shift <= -2)
Chris@69:          {
Chris@69:             g = 16384;
Chris@69:             shift = -2;
Chris@69:          }
Chris@69:          do {
Chris@69:             *f++ = SHL32(MULT16_16(*x++, g), -shift);
Chris@69:          } while (++j<band_end);
Chris@69:       } else
Chris@69: #endif
Chris@69:          /* Be careful of the fixed-point "else" just above when changing this code */
Chris@69:          do {
Chris@69:             *f++ = SHR32(MULT16_16(*x++, g), shift);
Chris@69:          } while (++j<band_end);
Chris@69:    }
Chris@69:    celt_assert(start <= end);
Chris@69:    OPUS_CLEAR(&freq[bound], N-bound);
Chris@69: }
Chris@69: 
Chris@69: /* This prevents energy collapse for transients with multiple short MDCTs */
Chris@69: void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_masks, int LM, int C, int size,
Chris@69:       int start, int end, const opus_val16 *logE, const opus_val16 *prev1logE,
Chris@69:       const opus_val16 *prev2logE, const int *pulses, opus_uint32 seed, int arch)
Chris@69: {
Chris@69:    int c, i, j, k;
Chris@69:    for (i=start;i<end;i++)
Chris@69:    {
Chris@69:       int N0;
Chris@69:       opus_val16 thresh, sqrt_1;
Chris@69:       int depth;
Chris@69: #ifdef FIXED_POINT
Chris@69:       int shift;
Chris@69:       opus_val32 thresh32;
Chris@69: #endif
Chris@69: 
Chris@69:       N0 = m->eBands[i+1]-m->eBands[i];
Chris@69:       /* depth in 1/8 bits */
Chris@69:       celt_sig_assert(pulses[i]>=0);
Chris@69:       depth = celt_udiv(1+pulses[i], (m->eBands[i+1]-m->eBands[i]))>>LM;
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69:       thresh32 = SHR32(celt_exp2(-SHL16(depth, 10-BITRES)),1);
Chris@69:       thresh = MULT16_32_Q15(QCONST16(0.5f, 15), MIN32(32767,thresh32));
Chris@69:       {
Chris@69:          opus_val32 t;
Chris@69:          t = N0<<LM;
Chris@69:          shift = celt_ilog2(t)>>1;
Chris@69:          t = SHL32(t, (7-shift)<<1);
Chris@69:          sqrt_1 = celt_rsqrt_norm(t);
Chris@69:       }
Chris@69: #else
Chris@69:       thresh = .5f*celt_exp2(-.125f*depth);
Chris@69:       sqrt_1 = celt_rsqrt(N0<<LM);
Chris@69: #endif
Chris@69: 
Chris@69:       c=0; do
Chris@69:       {
Chris@69:          celt_norm *X;
Chris@69:          opus_val16 prev1;
Chris@69:          opus_val16 prev2;
Chris@69:          opus_val32 Ediff;
Chris@69:          opus_val16 r;
Chris@69:          int renormalize=0;
Chris@69:          prev1 = prev1logE[c*m->nbEBands+i];
Chris@69:          prev2 = prev2logE[c*m->nbEBands+i];
Chris@69:          if (C==1)
Chris@69:          {
Chris@69:             prev1 = MAX16(prev1,prev1logE[m->nbEBands+i]);
Chris@69:             prev2 = MAX16(prev2,prev2logE[m->nbEBands+i]);
Chris@69:          }
Chris@69:          Ediff = EXTEND32(logE[c*m->nbEBands+i])-EXTEND32(MIN16(prev1,prev2));
Chris@69:          Ediff = MAX32(0, Ediff);
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69:          if (Ediff < 16384)
Chris@69:          {
Chris@69:             opus_val32 r32 = SHR32(celt_exp2(-EXTRACT16(Ediff)),1);
Chris@69:             r = 2*MIN16(16383,r32);
Chris@69:          } else {
Chris@69:             r = 0;
Chris@69:          }
Chris@69:          if (LM==3)
Chris@69:             r = MULT16_16_Q14(23170, MIN32(23169, r));
Chris@69:          r = SHR16(MIN16(thresh, r),1);
Chris@69:          r = SHR32(MULT16_16_Q15(sqrt_1, r),shift);
Chris@69: #else
Chris@69:          /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
Chris@69:             short blocks don't have the same energy as long */
Chris@69:          r = 2.f*celt_exp2(-Ediff);
Chris@69:          if (LM==3)
Chris@69:             r *= 1.41421356f;
Chris@69:          r = MIN16(thresh, r);
Chris@69:          r = r*sqrt_1;
Chris@69: #endif
Chris@69:          X = X_+c*size+(m->eBands[i]<<LM);
Chris@69:          for (k=0;k<1<<LM;k++)
Chris@69:          {
Chris@69:             /* Detect collapse */
Chris@69:             if (!(collapse_masks[i*C+c]&1<<k))
Chris@69:             {
Chris@69:                /* Fill with noise */
Chris@69:                for (j=0;j<N0;j++)
Chris@69:                {
Chris@69:                   seed = celt_lcg_rand(seed);
Chris@69:                   X[(j<<LM)+k] = (seed&0x8000 ? r : -r);
Chris@69:                }
Chris@69:                renormalize = 1;
Chris@69:             }
Chris@69:          }
Chris@69:          /* We just added some energy, so we need to renormalise */
Chris@69:          if (renormalize)
Chris@69:             renormalise_vector(X, N0<<LM, Q15ONE, arch);
Chris@69:       } while (++c<C);
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: /* Compute the weights to use for optimizing normalized distortion across
Chris@69:    channels. We use the amplitude to weight square distortion, which means
Chris@69:    that we use the square root of the value we would have been using if we
Chris@69:    wanted to minimize the MSE in the non-normalized domain. This roughly
Chris@69:    corresponds to some quick-and-dirty perceptual experiments I ran to
Chris@69:    measure inter-aural masking (there doesn't seem to be any published data
Chris@69:    on the topic). */
Chris@69: static void compute_channel_weights(celt_ener Ex, celt_ener Ey, opus_val16 w[2])
Chris@69: {
Chris@69:    celt_ener minE;
Chris@69: #if FIXED_POINT
Chris@69:    int shift;
Chris@69: #endif
Chris@69:    minE = MIN32(Ex, Ey);
Chris@69:    /* Adjustment to make the weights a bit more conservative. */
Chris@69:    Ex = ADD32(Ex, minE/3);
Chris@69:    Ey = ADD32(Ey, minE/3);
Chris@69: #if FIXED_POINT
Chris@69:    shift = celt_ilog2(EPSILON+MAX32(Ex, Ey))-14;
Chris@69: #endif
Chris@69:    w[0] = VSHR32(Ex, shift);
Chris@69:    w[1] = VSHR32(Ey, shift);
Chris@69: }
Chris@69: 
Chris@69: static void intensity_stereo(const CELTMode *m, celt_norm * OPUS_RESTRICT X, const celt_norm * OPUS_RESTRICT Y, const celt_ener *bandE, int bandID, int N)
Chris@69: {
Chris@69:    int i = bandID;
Chris@69:    int j;
Chris@69:    opus_val16 a1, a2;
Chris@69:    opus_val16 left, right;
Chris@69:    opus_val16 norm;
Chris@69: #ifdef FIXED_POINT
Chris@69:    int shift = celt_zlog2(MAX32(bandE[i], bandE[i+m->nbEBands]))-13;
Chris@69: #endif
Chris@69:    left = VSHR32(bandE[i],shift);
Chris@69:    right = VSHR32(bandE[i+m->nbEBands],shift);
Chris@69:    norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right));
Chris@69:    a1 = DIV32_16(SHL32(EXTEND32(left),14),norm);
Chris@69:    a2 = DIV32_16(SHL32(EXTEND32(right),14),norm);
Chris@69:    for (j=0;j<N;j++)
Chris@69:    {
Chris@69:       celt_norm r, l;
Chris@69:       l = X[j];
Chris@69:       r = Y[j];
Chris@69:       X[j] = EXTRACT16(SHR32(MAC16_16(MULT16_16(a1, l), a2, r), 14));
Chris@69:       /* Side is not encoded, no need to calculate */
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: static void stereo_split(celt_norm * OPUS_RESTRICT X, celt_norm * OPUS_RESTRICT Y, int N)
Chris@69: {
Chris@69:    int j;
Chris@69:    for (j=0;j<N;j++)
Chris@69:    {
Chris@69:       opus_val32 r, l;
Chris@69:       l = MULT16_16(QCONST16(.70710678f, 15), X[j]);
Chris@69:       r = MULT16_16(QCONST16(.70710678f, 15), Y[j]);
Chris@69:       X[j] = EXTRACT16(SHR32(ADD32(l, r), 15));
Chris@69:       Y[j] = EXTRACT16(SHR32(SUB32(r, l), 15));
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: static void stereo_merge(celt_norm * OPUS_RESTRICT X, celt_norm * OPUS_RESTRICT Y, opus_val16 mid, int N, int arch)
Chris@69: {
Chris@69:    int j;
Chris@69:    opus_val32 xp=0, side=0;
Chris@69:    opus_val32 El, Er;
Chris@69:    opus_val16 mid2;
Chris@69: #ifdef FIXED_POINT
Chris@69:    int kl, kr;
Chris@69: #endif
Chris@69:    opus_val32 t, lgain, rgain;
Chris@69: 
Chris@69:    /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */
Chris@69:    dual_inner_prod(Y, X, Y, N, &xp, &side, arch);
Chris@69:    /* Compensating for the mid normalization */
Chris@69:    xp = MULT16_32_Q15(mid, xp);
Chris@69:    /* mid and side are in Q15, not Q14 like X and Y */
Chris@69:    mid2 = SHR16(mid, 1);
Chris@69:    El = MULT16_16(mid2, mid2) + side - 2*xp;
Chris@69:    Er = MULT16_16(mid2, mid2) + side + 2*xp;
Chris@69:    if (Er < QCONST32(6e-4f, 28) || El < QCONST32(6e-4f, 28))
Chris@69:    {
Chris@69:       OPUS_COPY(Y, X, N);
Chris@69:       return;
Chris@69:    }
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69:    kl = celt_ilog2(El)>>1;
Chris@69:    kr = celt_ilog2(Er)>>1;
Chris@69: #endif
Chris@69:    t = VSHR32(El, (kl-7)<<1);
Chris@69:    lgain = celt_rsqrt_norm(t);
Chris@69:    t = VSHR32(Er, (kr-7)<<1);
Chris@69:    rgain = celt_rsqrt_norm(t);
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69:    if (kl < 7)
Chris@69:       kl = 7;
Chris@69:    if (kr < 7)
Chris@69:       kr = 7;
Chris@69: #endif
Chris@69: 
Chris@69:    for (j=0;j<N;j++)
Chris@69:    {
Chris@69:       celt_norm r, l;
Chris@69:       /* Apply mid scaling (side is already scaled) */
Chris@69:       l = MULT16_16_P15(mid, X[j]);
Chris@69:       r = Y[j];
Chris@69:       X[j] = EXTRACT16(PSHR32(MULT16_16(lgain, SUB16(l,r)), kl+1));
Chris@69:       Y[j] = EXTRACT16(PSHR32(MULT16_16(rgain, ADD16(l,r)), kr+1));
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: /* Decide whether we should spread the pulses in the current frame */
Chris@69: int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
Chris@69:       int last_decision, int *hf_average, int *tapset_decision, int update_hf,
Chris@69:       int end, int C, int M, const int *spread_weight)
Chris@69: {
Chris@69:    int i, c, N0;
Chris@69:    int sum = 0, nbBands=0;
Chris@69:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
Chris@69:    int decision;
Chris@69:    int hf_sum=0;
Chris@69: 
Chris@69:    celt_assert(end>0);
Chris@69: 
Chris@69:    N0 = M*m->shortMdctSize;
Chris@69: 
Chris@69:    if (M*(eBands[end]-eBands[end-1]) <= 8)
Chris@69:       return SPREAD_NONE;
Chris@69:    c=0; do {
Chris@69:       for (i=0;i<end;i++)
Chris@69:       {
Chris@69:          int j, N, tmp=0;
Chris@69:          int tcount[3] = {0,0,0};
Chris@69:          const celt_norm * OPUS_RESTRICT x = X+M*eBands[i]+c*N0;
Chris@69:          N = M*(eBands[i+1]-eBands[i]);
Chris@69:          if (N<=8)
Chris@69:             continue;
Chris@69:          /* Compute rough CDF of |x[j]| */
Chris@69:          for (j=0;j<N;j++)
Chris@69:          {
Chris@69:             opus_val32 x2N; /* Q13 */
Chris@69: 
Chris@69:             x2N = MULT16_16(MULT16_16_Q15(x[j], x[j]), N);
Chris@69:             if (x2N < QCONST16(0.25f,13))
Chris@69:                tcount[0]++;
Chris@69:             if (x2N < QCONST16(0.0625f,13))
Chris@69:                tcount[1]++;
Chris@69:             if (x2N < QCONST16(0.015625f,13))
Chris@69:                tcount[2]++;
Chris@69:          }
Chris@69: 
Chris@69:          /* Only include four last bands (8 kHz and up) */
Chris@69:          if (i>m->nbEBands-4)
Chris@69:             hf_sum += celt_udiv(32*(tcount[1]+tcount[0]), N);
Chris@69:          tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N);
Chris@69:          sum += tmp*spread_weight[i];
Chris@69:          nbBands+=spread_weight[i];
Chris@69:       }
Chris@69:    } while (++c<C);
Chris@69: 
Chris@69:    if (update_hf)
Chris@69:    {
Chris@69:       if (hf_sum)
Chris@69:          hf_sum = celt_udiv(hf_sum, C*(4-m->nbEBands+end));
Chris@69:       *hf_average = (*hf_average+hf_sum)>>1;
Chris@69:       hf_sum = *hf_average;
Chris@69:       if (*tapset_decision==2)
Chris@69:          hf_sum += 4;
Chris@69:       else if (*tapset_decision==0)
Chris@69:          hf_sum -= 4;
Chris@69:       if (hf_sum > 22)
Chris@69:          *tapset_decision=2;
Chris@69:       else if (hf_sum > 18)
Chris@69:          *tapset_decision=1;
Chris@69:       else
Chris@69:          *tapset_decision=0;
Chris@69:    }
Chris@69:    /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/
Chris@69:    celt_assert(nbBands>0); /* end has to be non-zero */
Chris@69:    celt_assert(sum>=0);
Chris@69:    sum = celt_udiv((opus_int32)sum<<8, nbBands);
Chris@69:    /* Recursive averaging */
Chris@69:    sum = (sum+*average)>>1;
Chris@69:    *average = sum;
Chris@69:    /* Hysteresis */
Chris@69:    sum = (3*sum + (((3-last_decision)<<7) + 64) + 2)>>2;
Chris@69:    if (sum < 80)
Chris@69:    {
Chris@69:       decision = SPREAD_AGGRESSIVE;
Chris@69:    } else if (sum < 256)
Chris@69:    {
Chris@69:       decision = SPREAD_NORMAL;
Chris@69:    } else if (sum < 384)
Chris@69:    {
Chris@69:       decision = SPREAD_LIGHT;
Chris@69:    } else {
Chris@69:       decision = SPREAD_NONE;
Chris@69:    }
Chris@69: #ifdef FUZZING
Chris@69:    decision = rand()&0x3;
Chris@69:    *tapset_decision=rand()%3;
Chris@69: #endif
Chris@69:    return decision;
Chris@69: }
Chris@69: 
Chris@69: /* Indexing table for converting from natural Hadamard to ordery Hadamard
Chris@69:    This is essentially a bit-reversed Gray, on top of which we've added
Chris@69:    an inversion of the order because we want the DC at the end rather than
Chris@69:    the beginning. The lines are for N=2, 4, 8, 16 */
Chris@69: static const int ordery_table[] = {
Chris@69:        1,  0,
Chris@69:        3,  0,  2,  1,
Chris@69:        7,  0,  4,  3,  6,  1,  5,  2,
Chris@69:       15,  0,  8,  7, 12,  3, 11,  4, 14,  1,  9,  6, 13,  2, 10,  5,
Chris@69: };
Chris@69: 
Chris@69: static void deinterleave_hadamard(celt_norm *X, int N0, int stride, int hadamard)
Chris@69: {
Chris@69:    int i,j;
Chris@69:    VARDECL(celt_norm, tmp);
Chris@69:    int N;
Chris@69:    SAVE_STACK;
Chris@69:    N = N0*stride;
Chris@69:    ALLOC(tmp, N, celt_norm);
Chris@69:    celt_assert(stride>0);
Chris@69:    if (hadamard)
Chris@69:    {
Chris@69:       const int *ordery = ordery_table+stride-2;
Chris@69:       for (i=0;i<stride;i++)
Chris@69:       {
Chris@69:          for (j=0;j<N0;j++)
Chris@69:             tmp[ordery[i]*N0+j] = X[j*stride+i];
Chris@69:       }
Chris@69:    } else {
Chris@69:       for (i=0;i<stride;i++)
Chris@69:          for (j=0;j<N0;j++)
Chris@69:             tmp[i*N0+j] = X[j*stride+i];
Chris@69:    }
Chris@69:    OPUS_COPY(X, tmp, N);
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: 
Chris@69: static void interleave_hadamard(celt_norm *X, int N0, int stride, int hadamard)
Chris@69: {
Chris@69:    int i,j;
Chris@69:    VARDECL(celt_norm, tmp);
Chris@69:    int N;
Chris@69:    SAVE_STACK;
Chris@69:    N = N0*stride;
Chris@69:    ALLOC(tmp, N, celt_norm);
Chris@69:    if (hadamard)
Chris@69:    {
Chris@69:       const int *ordery = ordery_table+stride-2;
Chris@69:       for (i=0;i<stride;i++)
Chris@69:          for (j=0;j<N0;j++)
Chris@69:             tmp[j*stride+i] = X[ordery[i]*N0+j];
Chris@69:    } else {
Chris@69:       for (i=0;i<stride;i++)
Chris@69:          for (j=0;j<N0;j++)
Chris@69:             tmp[j*stride+i] = X[i*N0+j];
Chris@69:    }
Chris@69:    OPUS_COPY(X, tmp, N);
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: 
Chris@69: void haar1(celt_norm *X, int N0, int stride)
Chris@69: {
Chris@69:    int i, j;
Chris@69:    N0 >>= 1;
Chris@69:    for (i=0;i<stride;i++)
Chris@69:       for (j=0;j<N0;j++)
Chris@69:       {
Chris@69:          opus_val32 tmp1, tmp2;
Chris@69:          tmp1 = MULT16_16(QCONST16(.70710678f,15), X[stride*2*j+i]);
Chris@69:          tmp2 = MULT16_16(QCONST16(.70710678f,15), X[stride*(2*j+1)+i]);
Chris@69:          X[stride*2*j+i] = EXTRACT16(PSHR32(ADD32(tmp1, tmp2), 15));
Chris@69:          X[stride*(2*j+1)+i] = EXTRACT16(PSHR32(SUB32(tmp1, tmp2), 15));
Chris@69:       }
Chris@69: }
Chris@69: 
Chris@69: static int compute_qn(int N, int b, int offset, int pulse_cap, int stereo)
Chris@69: {
Chris@69:    static const opus_int16 exp2_table8[8] =
Chris@69:       {16384, 17866, 19483, 21247, 23170, 25267, 27554, 30048};
Chris@69:    int qn, qb;
Chris@69:    int N2 = 2*N-1;
Chris@69:    if (stereo && N==2)
Chris@69:       N2--;
Chris@69:    /* The upper limit ensures that in a stereo split with itheta==16384, we'll
Chris@69:        always have enough bits left over to code at least one pulse in the
Chris@69:        side; otherwise it would collapse, since it doesn't get folded. */
Chris@69:    qb = celt_sudiv(b+N2*offset, N2);
Chris@69:    qb = IMIN(b-pulse_cap-(4<<BITRES), qb);
Chris@69: 
Chris@69:    qb = IMIN(8<<BITRES, qb);
Chris@69: 
Chris@69:    if (qb<(1<<BITRES>>1)) {
Chris@69:       qn = 1;
Chris@69:    } else {
Chris@69:       qn = exp2_table8[qb&0x7]>>(14-(qb>>BITRES));
Chris@69:       qn = (qn+1)>>1<<1;
Chris@69:    }
Chris@69:    celt_assert(qn <= 256);
Chris@69:    return qn;
Chris@69: }
Chris@69: 
Chris@69: struct band_ctx {
Chris@69:    int encode;
Chris@69:    int resynth;
Chris@69:    const CELTMode *m;
Chris@69:    int i;
Chris@69:    int intensity;
Chris@69:    int spread;
Chris@69:    int tf_change;
Chris@69:    ec_ctx *ec;
Chris@69:    opus_int32 remaining_bits;
Chris@69:    const celt_ener *bandE;
Chris@69:    opus_uint32 seed;
Chris@69:    int arch;
Chris@69:    int theta_round;
Chris@69:    int disable_inv;
Chris@69:    int avoid_split_noise;
Chris@69: };
Chris@69: 
Chris@69: struct split_ctx {
Chris@69:    int inv;
Chris@69:    int imid;
Chris@69:    int iside;
Chris@69:    int delta;
Chris@69:    int itheta;
Chris@69:    int qalloc;
Chris@69: };
Chris@69: 
Chris@69: static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx,
Chris@69:       celt_norm *X, celt_norm *Y, int N, int *b, int B, int B0,
Chris@69:       int LM,
Chris@69:       int stereo, int *fill)
Chris@69: {
Chris@69:    int qn;
Chris@69:    int itheta=0;
Chris@69:    int delta;
Chris@69:    int imid, iside;
Chris@69:    int qalloc;
Chris@69:    int pulse_cap;
Chris@69:    int offset;
Chris@69:    opus_int32 tell;
Chris@69:    int inv=0;
Chris@69:    int encode;
Chris@69:    const CELTMode *m;
Chris@69:    int i;
Chris@69:    int intensity;
Chris@69:    ec_ctx *ec;
Chris@69:    const celt_ener *bandE;
Chris@69: 
Chris@69:    encode = ctx->encode;
Chris@69:    m = ctx->m;
Chris@69:    i = ctx->i;
Chris@69:    intensity = ctx->intensity;
Chris@69:    ec = ctx->ec;
Chris@69:    bandE = ctx->bandE;
Chris@69: 
Chris@69:    /* Decide on the resolution to give to the split parameter theta */
Chris@69:    pulse_cap = m->logN[i]+LM*(1<<BITRES);
Chris@69:    offset = (pulse_cap>>1) - (stereo&&N==2 ? QTHETA_OFFSET_TWOPHASE : QTHETA_OFFSET);
Chris@69:    qn = compute_qn(N, *b, offset, pulse_cap, stereo);
Chris@69:    if (stereo && i>=intensity)
Chris@69:       qn = 1;
Chris@69:    if (encode)
Chris@69:    {
Chris@69:       /* theta is the atan() of the ratio between the (normalized)
Chris@69:          side and mid. With just that parameter, we can re-scale both
Chris@69:          mid and side because we know that 1) they have unit norm and
Chris@69:          2) they are orthogonal. */
Chris@69:       itheta = stereo_itheta(X, Y, stereo, N, ctx->arch);
Chris@69:    }
Chris@69:    tell = ec_tell_frac(ec);
Chris@69:    if (qn!=1)
Chris@69:    {
Chris@69:       if (encode)
Chris@69:       {
Chris@69:          if (!stereo || ctx->theta_round == 0)
Chris@69:          {
Chris@69:             itheta = (itheta*(opus_int32)qn+8192)>>14;
Chris@69:             if (!stereo && ctx->avoid_split_noise && itheta > 0 && itheta < qn)
Chris@69:             {
Chris@69:                /* Check if the selected value of theta will cause the bit allocation
Chris@69:                   to inject noise on one side. If so, make sure the energy of that side
Chris@69:                   is zero. */
Chris@69:                int unquantized = celt_udiv((opus_int32)itheta*16384, qn);
Chris@69:                imid = bitexact_cos((opus_int16)unquantized);
Chris@69:                iside = bitexact_cos((opus_int16)(16384-unquantized));
Chris@69:                delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid));
Chris@69:                if (delta > *b)
Chris@69:                   itheta = qn;
Chris@69:                else if (delta < -*b)
Chris@69:                   itheta = 0;
Chris@69:             }
Chris@69:          } else {
Chris@69:             int down;
Chris@69:             /* Bias quantization towards itheta=0 and itheta=16384. */
Chris@69:             int bias = itheta > 8192 ? 32767/qn : -32767/qn;
Chris@69:             down = IMIN(qn-1, IMAX(0, (itheta*(opus_int32)qn + bias)>>14));
Chris@69:             if (ctx->theta_round < 0)
Chris@69:                itheta = down;
Chris@69:             else
Chris@69:                itheta = down+1;
Chris@69:          }
Chris@69:       }
Chris@69:       /* Entropy coding of the angle. We use a uniform pdf for the
Chris@69:          time split, a step for stereo, and a triangular one for the rest. */
Chris@69:       if (stereo && N>2)
Chris@69:       {
Chris@69:          int p0 = 3;
Chris@69:          int x = itheta;
Chris@69:          int x0 = qn/2;
Chris@69:          int ft = p0*(x0+1) + x0;
Chris@69:          /* Use a probability of p0 up to itheta=8192 and then use 1 after */
Chris@69:          if (encode)
Chris@69:          {
Chris@69:             ec_encode(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft);
Chris@69:          } else {
Chris@69:             int fs;
Chris@69:             fs=ec_decode(ec,ft);
Chris@69:             if (fs<(x0+1)*p0)
Chris@69:                x=fs/p0;
Chris@69:             else
Chris@69:                x=x0+1+(fs-(x0+1)*p0);
Chris@69:             ec_dec_update(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft);
Chris@69:             itheta = x;
Chris@69:          }
Chris@69:       } else if (B0>1 || stereo) {
Chris@69:          /* Uniform pdf */
Chris@69:          if (encode)
Chris@69:             ec_enc_uint(ec, itheta, qn+1);
Chris@69:          else
Chris@69:             itheta = ec_dec_uint(ec, qn+1);
Chris@69:       } else {
Chris@69:          int fs=1, ft;
Chris@69:          ft = ((qn>>1)+1)*((qn>>1)+1);
Chris@69:          if (encode)
Chris@69:          {
Chris@69:             int fl;
Chris@69: 
Chris@69:             fs = itheta <= (qn>>1) ? itheta + 1 : qn + 1 - itheta;
Chris@69:             fl = itheta <= (qn>>1) ? itheta*(itheta + 1)>>1 :
Chris@69:              ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1);
Chris@69: 
Chris@69:             ec_encode(ec, fl, fl+fs, ft);
Chris@69:          } else {
Chris@69:             /* Triangular pdf */
Chris@69:             int fl=0;
Chris@69:             int fm;
Chris@69:             fm = ec_decode(ec, ft);
Chris@69: 
Chris@69:             if (fm < ((qn>>1)*((qn>>1) + 1)>>1))
Chris@69:             {
Chris@69:                itheta = (isqrt32(8*(opus_uint32)fm + 1) - 1)>>1;
Chris@69:                fs = itheta + 1;
Chris@69:                fl = itheta*(itheta + 1)>>1;
Chris@69:             }
Chris@69:             else
Chris@69:             {
Chris@69:                itheta = (2*(qn + 1)
Chris@69:                 - isqrt32(8*(opus_uint32)(ft - fm - 1) + 1))>>1;
Chris@69:                fs = qn + 1 - itheta;
Chris@69:                fl = ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1);
Chris@69:             }
Chris@69: 
Chris@69:             ec_dec_update(ec, fl, fl+fs, ft);
Chris@69:          }
Chris@69:       }
Chris@69:       celt_assert(itheta>=0);
Chris@69:       itheta = celt_udiv((opus_int32)itheta*16384, qn);
Chris@69:       if (encode && stereo)
Chris@69:       {
Chris@69:          if (itheta==0)
Chris@69:             intensity_stereo(m, X, Y, bandE, i, N);
Chris@69:          else
Chris@69:             stereo_split(X, Y, N);
Chris@69:       }
Chris@69:       /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate.
Chris@69:                Let's do that at higher complexity */
Chris@69:    } else if (stereo) {
Chris@69:       if (encode)
Chris@69:       {
Chris@69:          inv = itheta > 8192 && !ctx->disable_inv;
Chris@69:          if (inv)
Chris@69:          {
Chris@69:             int j;
Chris@69:             for (j=0;j<N;j++)
Chris@69:                Y[j] = -Y[j];
Chris@69:          }
Chris@69:          intensity_stereo(m, X, Y, bandE, i, N);
Chris@69:       }
Chris@69:       if (*b>2<<BITRES && ctx->remaining_bits > 2<<BITRES)
Chris@69:       {
Chris@69:          if (encode)
Chris@69:             ec_enc_bit_logp(ec, inv, 2);
Chris@69:          else
Chris@69:             inv = ec_dec_bit_logp(ec, 2);
Chris@69:       } else
Chris@69:          inv = 0;
Chris@69:       /* inv flag override to avoid problems with downmixing. */
Chris@69:       if (ctx->disable_inv)
Chris@69:          inv = 0;
Chris@69:       itheta = 0;
Chris@69:    }
Chris@69:    qalloc = ec_tell_frac(ec) - tell;
Chris@69:    *b -= qalloc;
Chris@69: 
Chris@69:    if (itheta == 0)
Chris@69:    {
Chris@69:       imid = 32767;
Chris@69:       iside = 0;
Chris@69:       *fill &= (1<<B)-1;
Chris@69:       delta = -16384;
Chris@69:    } else if (itheta == 16384)
Chris@69:    {
Chris@69:       imid = 0;
Chris@69:       iside = 32767;
Chris@69:       *fill &= ((1<<B)-1)<<B;
Chris@69:       delta = 16384;
Chris@69:    } else {
Chris@69:       imid = bitexact_cos((opus_int16)itheta);
Chris@69:       iside = bitexact_cos((opus_int16)(16384-itheta));
Chris@69:       /* This is the mid vs side allocation that minimizes squared error
Chris@69:          in that band. */
Chris@69:       delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid));
Chris@69:    }
Chris@69: 
Chris@69:    sctx->inv = inv;
Chris@69:    sctx->imid = imid;
Chris@69:    sctx->iside = iside;
Chris@69:    sctx->delta = delta;
Chris@69:    sctx->itheta = itheta;
Chris@69:    sctx->qalloc = qalloc;
Chris@69: }
Chris@69: static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, int b,
Chris@69:       celt_norm *lowband_out)
Chris@69: {
Chris@69:    int c;
Chris@69:    int stereo;
Chris@69:    celt_norm *x = X;
Chris@69:    int encode;
Chris@69:    ec_ctx *ec;
Chris@69: 
Chris@69:    encode = ctx->encode;
Chris@69:    ec = ctx->ec;
Chris@69: 
Chris@69:    stereo = Y != NULL;
Chris@69:    c=0; do {
Chris@69:       int sign=0;
Chris@69:       if (ctx->remaining_bits>=1<<BITRES)
Chris@69:       {
Chris@69:          if (encode)
Chris@69:          {
Chris@69:             sign = x[0]<0;
Chris@69:             ec_enc_bits(ec, sign, 1);
Chris@69:          } else {
Chris@69:             sign = ec_dec_bits(ec, 1);
Chris@69:          }
Chris@69:          ctx->remaining_bits -= 1<<BITRES;
Chris@69:          b-=1<<BITRES;
Chris@69:       }
Chris@69:       if (ctx->resynth)
Chris@69:          x[0] = sign ? -NORM_SCALING : NORM_SCALING;
Chris@69:       x = Y;
Chris@69:    } while (++c<1+stereo);
Chris@69:    if (lowband_out)
Chris@69:       lowband_out[0] = SHR16(X[0],4);
Chris@69:    return 1;
Chris@69: }
Chris@69: 
Chris@69: /* This function is responsible for encoding and decoding a mono partition.
Chris@69:    It can split the band in two and transmit the energy difference with
Chris@69:    the two half-bands. It can be called recursively so bands can end up being
Chris@69:    split in 8 parts. */
Chris@69: static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X,
Chris@69:       int N, int b, int B, celt_norm *lowband,
Chris@69:       int LM,
Chris@69:       opus_val16 gain, int fill)
Chris@69: {
Chris@69:    const unsigned char *cache;
Chris@69:    int q;
Chris@69:    int curr_bits;
Chris@69:    int imid=0, iside=0;
Chris@69:    int B0=B;
Chris@69:    opus_val16 mid=0, side=0;
Chris@69:    unsigned cm=0;
Chris@69:    celt_norm *Y=NULL;
Chris@69:    int encode;
Chris@69:    const CELTMode *m;
Chris@69:    int i;
Chris@69:    int spread;
Chris@69:    ec_ctx *ec;
Chris@69: 
Chris@69:    encode = ctx->encode;
Chris@69:    m = ctx->m;
Chris@69:    i = ctx->i;
Chris@69:    spread = ctx->spread;
Chris@69:    ec = ctx->ec;
Chris@69: 
Chris@69:    /* If we need 1.5 more bit than we can produce, split the band in two. */
Chris@69:    cache = m->cache.bits + m->cache.index[(LM+1)*m->nbEBands+i];
Chris@69:    if (LM != -1 && b > cache[cache[0]]+12 && N>2)
Chris@69:    {
Chris@69:       int mbits, sbits, delta;
Chris@69:       int itheta;
Chris@69:       int qalloc;
Chris@69:       struct split_ctx sctx;
Chris@69:       celt_norm *next_lowband2=NULL;
Chris@69:       opus_int32 rebalance;
Chris@69: 
Chris@69:       N >>= 1;
Chris@69:       Y = X+N;
Chris@69:       LM -= 1;
Chris@69:       if (B==1)
Chris@69:          fill = (fill&1)|(fill<<1);
Chris@69:       B = (B+1)>>1;
Chris@69: 
Chris@69:       compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, LM, 0, &fill);
Chris@69:       imid = sctx.imid;
Chris@69:       iside = sctx.iside;
Chris@69:       delta = sctx.delta;
Chris@69:       itheta = sctx.itheta;
Chris@69:       qalloc = sctx.qalloc;
Chris@69: #ifdef FIXED_POINT
Chris@69:       mid = imid;
Chris@69:       side = iside;
Chris@69: #else
Chris@69:       mid = (1.f/32768)*imid;
Chris@69:       side = (1.f/32768)*iside;
Chris@69: #endif
Chris@69: 
Chris@69:       /* Give more bits to low-energy MDCTs than they would otherwise deserve */
Chris@69:       if (B0>1 && (itheta&0x3fff))
Chris@69:       {
Chris@69:          if (itheta > 8192)
Chris@69:             /* Rough approximation for pre-echo masking */
Chris@69:             delta -= delta>>(4-LM);
Chris@69:          else
Chris@69:             /* Corresponds to a forward-masking slope of 1.5 dB per 10 ms */
Chris@69:             delta = IMIN(0, delta + (N<<BITRES>>(5-LM)));
Chris@69:       }
Chris@69:       mbits = IMAX(0, IMIN(b, (b-delta)/2));
Chris@69:       sbits = b-mbits;
Chris@69:       ctx->remaining_bits -= qalloc;
Chris@69: 
Chris@69:       if (lowband)
Chris@69:          next_lowband2 = lowband+N; /* >32-bit split case */
Chris@69: 
Chris@69:       rebalance = ctx->remaining_bits;
Chris@69:       if (mbits >= sbits)
Chris@69:       {
Chris@69:          cm = quant_partition(ctx, X, N, mbits, B, lowband, LM,
Chris@69:                MULT16_16_P15(gain,mid), fill);
Chris@69:          rebalance = mbits - (rebalance-ctx->remaining_bits);
Chris@69:          if (rebalance > 3<<BITRES && itheta!=0)
Chris@69:             sbits += rebalance - (3<<BITRES);
Chris@69:          cm |= quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM,
Chris@69:                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
Chris@69:       } else {
Chris@69:          cm = quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM,
Chris@69:                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
Chris@69:          rebalance = sbits - (rebalance-ctx->remaining_bits);
Chris@69:          if (rebalance > 3<<BITRES && itheta!=16384)
Chris@69:             mbits += rebalance - (3<<BITRES);
Chris@69:          cm |= quant_partition(ctx, X, N, mbits, B, lowband, LM,
Chris@69:                MULT16_16_P15(gain,mid), fill);
Chris@69:       }
Chris@69:    } else {
Chris@69:       /* This is the basic no-split case */
Chris@69:       q = bits2pulses(m, i, LM, b);
Chris@69:       curr_bits = pulses2bits(m, i, LM, q);
Chris@69:       ctx->remaining_bits -= curr_bits;
Chris@69: 
Chris@69:       /* Ensures we can never bust the budget */
Chris@69:       while (ctx->remaining_bits < 0 && q > 0)
Chris@69:       {
Chris@69:          ctx->remaining_bits += curr_bits;
Chris@69:          q--;
Chris@69:          curr_bits = pulses2bits(m, i, LM, q);
Chris@69:          ctx->remaining_bits -= curr_bits;
Chris@69:       }
Chris@69: 
Chris@69:       if (q!=0)
Chris@69:       {
Chris@69:          int K = get_pulses(q);
Chris@69: 
Chris@69:          /* Finally do the actual quantization */
Chris@69:          if (encode)
Chris@69:          {
Chris@69:             cm = alg_quant(X, N, K, spread, B, ec, gain, ctx->resynth, ctx->arch);
Chris@69:          } else {
Chris@69:             cm = alg_unquant(X, N, K, spread, B, ec, gain);
Chris@69:          }
Chris@69:       } else {
Chris@69:          /* If there's no pulse, fill the band anyway */
Chris@69:          int j;
Chris@69:          if (ctx->resynth)
Chris@69:          {
Chris@69:             unsigned cm_mask;
Chris@69:             /* B can be as large as 16, so this shift might overflow an int on a
Chris@69:                16-bit platform; use a long to get defined behavior.*/
Chris@69:             cm_mask = (unsigned)(1UL<<B)-1;
Chris@69:             fill &= cm_mask;
Chris@69:             if (!fill)
Chris@69:             {
Chris@69:                OPUS_CLEAR(X, N);
Chris@69:             } else {
Chris@69:                if (lowband == NULL)
Chris@69:                {
Chris@69:                   /* Noise */
Chris@69:                   for (j=0;j<N;j++)
Chris@69:                   {
Chris@69:                      ctx->seed = celt_lcg_rand(ctx->seed);
Chris@69:                      X[j] = (celt_norm)((opus_int32)ctx->seed>>20);
Chris@69:                   }
Chris@69:                   cm = cm_mask;
Chris@69:                } else {
Chris@69:                   /* Folded spectrum */
Chris@69:                   for (j=0;j<N;j++)
Chris@69:                   {
Chris@69:                      opus_val16 tmp;
Chris@69:                      ctx->seed = celt_lcg_rand(ctx->seed);
Chris@69:                      /* About 48 dB below the "normal" folding level */
Chris@69:                      tmp = QCONST16(1.0f/256, 10);
Chris@69:                      tmp = (ctx->seed)&0x8000 ? tmp : -tmp;
Chris@69:                      X[j] = lowband[j]+tmp;
Chris@69:                   }
Chris@69:                   cm = fill;
Chris@69:                }
Chris@69:                renormalise_vector(X, N, gain, ctx->arch);
Chris@69:             }
Chris@69:          }
Chris@69:       }
Chris@69:    }
Chris@69: 
Chris@69:    return cm;
Chris@69: }
Chris@69: 
Chris@69: 
Chris@69: /* This function is responsible for encoding and decoding a band for the mono case. */
Chris@69: static unsigned quant_band(struct band_ctx *ctx, celt_norm *X,
Chris@69:       int N, int b, int B, celt_norm *lowband,
Chris@69:       int LM, celt_norm *lowband_out,
Chris@69:       opus_val16 gain, celt_norm *lowband_scratch, int fill)
Chris@69: {
Chris@69:    int N0=N;
Chris@69:    int N_B=N;
Chris@69:    int N_B0;
Chris@69:    int B0=B;
Chris@69:    int time_divide=0;
Chris@69:    int recombine=0;
Chris@69:    int longBlocks;
Chris@69:    unsigned cm=0;
Chris@69:    int k;
Chris@69:    int encode;
Chris@69:    int tf_change;
Chris@69: 
Chris@69:    encode = ctx->encode;
Chris@69:    tf_change = ctx->tf_change;
Chris@69: 
Chris@69:    longBlocks = B0==1;
Chris@69: 
Chris@69:    N_B = celt_udiv(N_B, B);
Chris@69: 
Chris@69:    /* Special case for one sample */
Chris@69:    if (N==1)
Chris@69:    {
Chris@69:       return quant_band_n1(ctx, X, NULL, b, lowband_out);
Chris@69:    }
Chris@69: 
Chris@69:    if (tf_change>0)
Chris@69:       recombine = tf_change;
Chris@69:    /* Band recombining to increase frequency resolution */
Chris@69: 
Chris@69:    if (lowband_scratch && lowband && (recombine || ((N_B&1) == 0 && tf_change<0) || B0>1))
Chris@69:    {
Chris@69:       OPUS_COPY(lowband_scratch, lowband, N);
Chris@69:       lowband = lowband_scratch;
Chris@69:    }
Chris@69: 
Chris@69:    for (k=0;k<recombine;k++)
Chris@69:    {
Chris@69:       static const unsigned char bit_interleave_table[16]={
Chris@69:             0,1,1,1,2,3,3,3,2,3,3,3,2,3,3,3
Chris@69:       };
Chris@69:       if (encode)
Chris@69:          haar1(X, N>>k, 1<<k);
Chris@69:       if (lowband)
Chris@69:          haar1(lowband, N>>k, 1<<k);
Chris@69:       fill = bit_interleave_table[fill&0xF]|bit_interleave_table[fill>>4]<<2;
Chris@69:    }
Chris@69:    B>>=recombine;
Chris@69:    N_B<<=recombine;
Chris@69: 
Chris@69:    /* Increasing the time resolution */
Chris@69:    while ((N_B&1) == 0 && tf_change<0)
Chris@69:    {
Chris@69:       if (encode)
Chris@69:          haar1(X, N_B, B);
Chris@69:       if (lowband)
Chris@69:          haar1(lowband, N_B, B);
Chris@69:       fill |= fill<<B;
Chris@69:       B <<= 1;
Chris@69:       N_B >>= 1;
Chris@69:       time_divide++;
Chris@69:       tf_change++;
Chris@69:    }
Chris@69:    B0=B;
Chris@69:    N_B0 = N_B;
Chris@69: 
Chris@69:    /* Reorganize the samples in time order instead of frequency order */
Chris@69:    if (B0>1)
Chris@69:    {
Chris@69:       if (encode)
Chris@69:          deinterleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
Chris@69:       if (lowband)
Chris@69:          deinterleave_hadamard(lowband, N_B>>recombine, B0<<recombine, longBlocks);
Chris@69:    }
Chris@69: 
Chris@69:    cm = quant_partition(ctx, X, N, b, B, lowband, LM, gain, fill);
Chris@69: 
Chris@69:    /* This code is used by the decoder and by the resynthesis-enabled encoder */
Chris@69:    if (ctx->resynth)
Chris@69:    {
Chris@69:       /* Undo the sample reorganization going from time order to frequency order */
Chris@69:       if (B0>1)
Chris@69:          interleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
Chris@69: 
Chris@69:       /* Undo time-freq changes that we did earlier */
Chris@69:       N_B = N_B0;
Chris@69:       B = B0;
Chris@69:       for (k=0;k<time_divide;k++)
Chris@69:       {
Chris@69:          B >>= 1;
Chris@69:          N_B <<= 1;
Chris@69:          cm |= cm>>B;
Chris@69:          haar1(X, N_B, B);
Chris@69:       }
Chris@69: 
Chris@69:       for (k=0;k<recombine;k++)
Chris@69:       {
Chris@69:          static const unsigned char bit_deinterleave_table[16]={
Chris@69:                0x00,0x03,0x0C,0x0F,0x30,0x33,0x3C,0x3F,
Chris@69:                0xC0,0xC3,0xCC,0xCF,0xF0,0xF3,0xFC,0xFF
Chris@69:          };
Chris@69:          cm = bit_deinterleave_table[cm];
Chris@69:          haar1(X, N0>>k, 1<<k);
Chris@69:       }
Chris@69:       B<<=recombine;
Chris@69: 
Chris@69:       /* Scale output for later folding */
Chris@69:       if (lowband_out)
Chris@69:       {
Chris@69:          int j;
Chris@69:          opus_val16 n;
Chris@69:          n = celt_sqrt(SHL32(EXTEND32(N0),22));
Chris@69:          for (j=0;j<N0;j++)
Chris@69:             lowband_out[j] = MULT16_16_Q15(n,X[j]);
Chris@69:       }
Chris@69:       cm &= (1<<B)-1;
Chris@69:    }
Chris@69:    return cm;
Chris@69: }
Chris@69: 
Chris@69: 
Chris@69: /* This function is responsible for encoding and decoding a band for the stereo case. */
Chris@69: static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm *Y,
Chris@69:       int N, int b, int B, celt_norm *lowband,
Chris@69:       int LM, celt_norm *lowband_out,
Chris@69:       celt_norm *lowband_scratch, int fill)
Chris@69: {
Chris@69:    int imid=0, iside=0;
Chris@69:    int inv = 0;
Chris@69:    opus_val16 mid=0, side=0;
Chris@69:    unsigned cm=0;
Chris@69:    int mbits, sbits, delta;
Chris@69:    int itheta;
Chris@69:    int qalloc;
Chris@69:    struct split_ctx sctx;
Chris@69:    int orig_fill;
Chris@69:    int encode;
Chris@69:    ec_ctx *ec;
Chris@69: 
Chris@69:    encode = ctx->encode;
Chris@69:    ec = ctx->ec;
Chris@69: 
Chris@69:    /* Special case for one sample */
Chris@69:    if (N==1)
Chris@69:    {
Chris@69:       return quant_band_n1(ctx, X, Y, b, lowband_out);
Chris@69:    }
Chris@69: 
Chris@69:    orig_fill = fill;
Chris@69: 
Chris@69:    compute_theta(ctx, &sctx, X, Y, N, &b, B, B, LM, 1, &fill);
Chris@69:    inv = sctx.inv;
Chris@69:    imid = sctx.imid;
Chris@69:    iside = sctx.iside;
Chris@69:    delta = sctx.delta;
Chris@69:    itheta = sctx.itheta;
Chris@69:    qalloc = sctx.qalloc;
Chris@69: #ifdef FIXED_POINT
Chris@69:    mid = imid;
Chris@69:    side = iside;
Chris@69: #else
Chris@69:    mid = (1.f/32768)*imid;
Chris@69:    side = (1.f/32768)*iside;
Chris@69: #endif
Chris@69: 
Chris@69:    /* This is a special case for N=2 that only works for stereo and takes
Chris@69:       advantage of the fact that mid and side are orthogonal to encode
Chris@69:       the side with just one bit. */
Chris@69:    if (N==2)
Chris@69:    {
Chris@69:       int c;
Chris@69:       int sign=0;
Chris@69:       celt_norm *x2, *y2;
Chris@69:       mbits = b;
Chris@69:       sbits = 0;
Chris@69:       /* Only need one bit for the side. */
Chris@69:       if (itheta != 0 && itheta != 16384)
Chris@69:          sbits = 1<<BITRES;
Chris@69:       mbits -= sbits;
Chris@69:       c = itheta > 8192;
Chris@69:       ctx->remaining_bits -= qalloc+sbits;
Chris@69: 
Chris@69:       x2 = c ? Y : X;
Chris@69:       y2 = c ? X : Y;
Chris@69:       if (sbits)
Chris@69:       {
Chris@69:          if (encode)
Chris@69:          {
Chris@69:             /* Here we only need to encode a sign for the side. */
Chris@69:             sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
Chris@69:             ec_enc_bits(ec, sign, 1);
Chris@69:          } else {
Chris@69:             sign = ec_dec_bits(ec, 1);
Chris@69:          }
Chris@69:       }
Chris@69:       sign = 1-2*sign;
Chris@69:       /* We use orig_fill here because we want to fold the side, but if
Chris@69:          itheta==16384, we'll have cleared the low bits of fill. */
Chris@69:       cm = quant_band(ctx, x2, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
Chris@69:             lowband_scratch, orig_fill);
Chris@69:       /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
Chris@69:          and there's no need to worry about mixing with the other channel. */
Chris@69:       y2[0] = -sign*x2[1];
Chris@69:       y2[1] = sign*x2[0];
Chris@69:       if (ctx->resynth)
Chris@69:       {
Chris@69:          celt_norm tmp;
Chris@69:          X[0] = MULT16_16_Q15(mid, X[0]);
Chris@69:          X[1] = MULT16_16_Q15(mid, X[1]);
Chris@69:          Y[0] = MULT16_16_Q15(side, Y[0]);
Chris@69:          Y[1] = MULT16_16_Q15(side, Y[1]);
Chris@69:          tmp = X[0];
Chris@69:          X[0] = SUB16(tmp,Y[0]);
Chris@69:          Y[0] = ADD16(tmp,Y[0]);
Chris@69:          tmp = X[1];
Chris@69:          X[1] = SUB16(tmp,Y[1]);
Chris@69:          Y[1] = ADD16(tmp,Y[1]);
Chris@69:       }
Chris@69:    } else {
Chris@69:       /* "Normal" split code */
Chris@69:       opus_int32 rebalance;
Chris@69: 
Chris@69:       mbits = IMAX(0, IMIN(b, (b-delta)/2));
Chris@69:       sbits = b-mbits;
Chris@69:       ctx->remaining_bits -= qalloc;
Chris@69: 
Chris@69:       rebalance = ctx->remaining_bits;
Chris@69:       if (mbits >= sbits)
Chris@69:       {
Chris@69:          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
Chris@69:             mid for folding later. */
Chris@69:          cm = quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
Chris@69:                lowband_scratch, fill);
Chris@69:          rebalance = mbits - (rebalance-ctx->remaining_bits);
Chris@69:          if (rebalance > 3<<BITRES && itheta!=0)
Chris@69:             sbits += rebalance - (3<<BITRES);
Chris@69: 
Chris@69:          /* For a stereo split, the high bits of fill are always zero, so no
Chris@69:             folding will be done to the side. */
Chris@69:          cm |= quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B);
Chris@69:       } else {
Chris@69:          /* For a stereo split, the high bits of fill are always zero, so no
Chris@69:             folding will be done to the side. */
Chris@69:          cm = quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B);
Chris@69:          rebalance = sbits - (rebalance-ctx->remaining_bits);
Chris@69:          if (rebalance > 3<<BITRES && itheta!=16384)
Chris@69:             mbits += rebalance - (3<<BITRES);
Chris@69:          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
Chris@69:             mid for folding later. */
Chris@69:          cm |= quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
Chris@69:                lowband_scratch, fill);
Chris@69:       }
Chris@69:    }
Chris@69: 
Chris@69: 
Chris@69:    /* This code is used by the decoder and by the resynthesis-enabled encoder */
Chris@69:    if (ctx->resynth)
Chris@69:    {
Chris@69:       if (N!=2)
Chris@69:          stereo_merge(X, Y, mid, N, ctx->arch);
Chris@69:       if (inv)
Chris@69:       {
Chris@69:          int j;
Chris@69:          for (j=0;j<N;j++)
Chris@69:             Y[j] = -Y[j];
Chris@69:       }
Chris@69:    }
Chris@69:    return cm;
Chris@69: }
Chris@69: 
Chris@69: static void special_hybrid_folding(const CELTMode *m, celt_norm *norm, celt_norm *norm2, int start, int M, int dual_stereo)
Chris@69: {
Chris@69:    int n1, n2;
Chris@69:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
Chris@69:    n1 = M*(eBands[start+1]-eBands[start]);
Chris@69:    n2 = M*(eBands[start+2]-eBands[start+1]);
Chris@69:    /* Duplicate enough of the first band folding data to be able to fold the second band.
Chris@69:       Copies no data for CELT-only mode. */
Chris@69:    OPUS_COPY(&norm[n1], &norm[2*n1 - n2], n2-n1);
Chris@69:    if (dual_stereo)
Chris@69:       OPUS_COPY(&norm2[n1], &norm2[2*n1 - n2], n2-n1);
Chris@69: }
Chris@69: 
Chris@69: void quant_all_bands(int encode, const CELTMode *m, int start, int end,
Chris@69:       celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks,
Chris@69:       const celt_ener *bandE, int *pulses, int shortBlocks, int spread,
Chris@69:       int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits,
Chris@69:       opus_int32 balance, ec_ctx *ec, int LM, int codedBands,
Chris@69:       opus_uint32 *seed, int complexity, int arch, int disable_inv)
Chris@69: {
Chris@69:    int i;
Chris@69:    opus_int32 remaining_bits;
Chris@69:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
Chris@69:    celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2;
Chris@69:    VARDECL(celt_norm, _norm);
Chris@69:    VARDECL(celt_norm, _lowband_scratch);
Chris@69:    VARDECL(celt_norm, X_save);
Chris@69:    VARDECL(celt_norm, Y_save);
Chris@69:    VARDECL(celt_norm, X_save2);
Chris@69:    VARDECL(celt_norm, Y_save2);
Chris@69:    VARDECL(celt_norm, norm_save2);
Chris@69:    int resynth_alloc;
Chris@69:    celt_norm *lowband_scratch;
Chris@69:    int B;
Chris@69:    int M;
Chris@69:    int lowband_offset;
Chris@69:    int update_lowband = 1;
Chris@69:    int C = Y_ != NULL ? 2 : 1;
Chris@69:    int norm_offset;
Chris@69:    int theta_rdo = encode && Y_!=NULL && !dual_stereo && complexity>=8;
Chris@69: #ifdef RESYNTH
Chris@69:    int resynth = 1;
Chris@69: #else
Chris@69:    int resynth = !encode || theta_rdo;
Chris@69: #endif
Chris@69:    struct band_ctx ctx;
Chris@69:    SAVE_STACK;
Chris@69: 
Chris@69:    M = 1<<LM;
Chris@69:    B = shortBlocks ? M : 1;
Chris@69:    norm_offset = M*eBands[start];
Chris@69:    /* No need to allocate norm for the last band because we don't need an
Chris@69:       output in that band. */
Chris@69:    ALLOC(_norm, C*(M*eBands[m->nbEBands-1]-norm_offset), celt_norm);
Chris@69:    norm = _norm;
Chris@69:    norm2 = norm + M*eBands[m->nbEBands-1]-norm_offset;
Chris@69: 
Chris@69:    /* For decoding, we can use the last band as scratch space because we don't need that
Chris@69:       scratch space for the last band and we don't care about the data there until we're
Chris@69:       decoding the last band. */
Chris@69:    if (encode && resynth)
Chris@69:       resynth_alloc = M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]);
Chris@69:    else
Chris@69:       resynth_alloc = ALLOC_NONE;
Chris@69:    ALLOC(_lowband_scratch, resynth_alloc, celt_norm);
Chris@69:    if (encode && resynth)
Chris@69:       lowband_scratch = _lowband_scratch;
Chris@69:    else
Chris@69:       lowband_scratch = X_+M*eBands[m->nbEBands-1];
Chris@69:    ALLOC(X_save, resynth_alloc, celt_norm);
Chris@69:    ALLOC(Y_save, resynth_alloc, celt_norm);
Chris@69:    ALLOC(X_save2, resynth_alloc, celt_norm);
Chris@69:    ALLOC(Y_save2, resynth_alloc, celt_norm);
Chris@69:    ALLOC(norm_save2, resynth_alloc, celt_norm);
Chris@69: 
Chris@69:    lowband_offset = 0;
Chris@69:    ctx.bandE = bandE;
Chris@69:    ctx.ec = ec;
Chris@69:    ctx.encode = encode;
Chris@69:    ctx.intensity = intensity;
Chris@69:    ctx.m = m;
Chris@69:    ctx.seed = *seed;
Chris@69:    ctx.spread = spread;
Chris@69:    ctx.arch = arch;
Chris@69:    ctx.disable_inv = disable_inv;
Chris@69:    ctx.resynth = resynth;
Chris@69:    ctx.theta_round = 0;
Chris@69:    /* Avoid injecting noise in the first band on transients. */
Chris@69:    ctx.avoid_split_noise = B > 1;
Chris@69:    for (i=start;i<end;i++)
Chris@69:    {
Chris@69:       opus_int32 tell;
Chris@69:       int b;
Chris@69:       int N;
Chris@69:       opus_int32 curr_balance;
Chris@69:       int effective_lowband=-1;
Chris@69:       celt_norm * OPUS_RESTRICT X, * OPUS_RESTRICT Y;
Chris@69:       int tf_change=0;
Chris@69:       unsigned x_cm;
Chris@69:       unsigned y_cm;
Chris@69:       int last;
Chris@69: 
Chris@69:       ctx.i = i;
Chris@69:       last = (i==end-1);
Chris@69: 
Chris@69:       X = X_+M*eBands[i];
Chris@69:       if (Y_!=NULL)
Chris@69:          Y = Y_+M*eBands[i];
Chris@69:       else
Chris@69:          Y = NULL;
Chris@69:       N = M*eBands[i+1]-M*eBands[i];
Chris@69:       celt_assert(N > 0);
Chris@69:       tell = ec_tell_frac(ec);
Chris@69: 
Chris@69:       /* Compute how many bits we want to allocate to this band */
Chris@69:       if (i != start)
Chris@69:          balance -= tell;
Chris@69:       remaining_bits = total_bits-tell-1;
Chris@69:       ctx.remaining_bits = remaining_bits;
Chris@69:       if (i <= codedBands-1)
Chris@69:       {
Chris@69:          curr_balance = celt_sudiv(balance, IMIN(3, codedBands-i));
Chris@69:          b = IMAX(0, IMIN(16383, IMIN(remaining_bits+1,pulses[i]+curr_balance)));
Chris@69:       } else {
Chris@69:          b = 0;
Chris@69:       }
Chris@69: 
Chris@69: #ifndef DISABLE_UPDATE_DRAFT
Chris@69:       if (resynth && (M*eBands[i]-N >= M*eBands[start] || i==start+1) && (update_lowband || lowband_offset==0))
Chris@69:             lowband_offset = i;
Chris@69:       if (i == start+1)
Chris@69:          special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
Chris@69: #else
Chris@69:       if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0))
Chris@69:             lowband_offset = i;
Chris@69: #endif
Chris@69: 
Chris@69:       tf_change = tf_res[i];
Chris@69:       ctx.tf_change = tf_change;
Chris@69:       if (i>=m->effEBands)
Chris@69:       {
Chris@69:          X=norm;
Chris@69:          if (Y_!=NULL)
Chris@69:             Y = norm;
Chris@69:          lowband_scratch = NULL;
Chris@69:       }
Chris@69:       if (last && !theta_rdo)
Chris@69:          lowband_scratch = NULL;
Chris@69: 
Chris@69:       /* Get a conservative estimate of the collapse_mask's for the bands we're
Chris@69:          going to be folding from. */
Chris@69:       if (lowband_offset != 0 && (spread!=SPREAD_AGGRESSIVE || B>1 || tf_change<0))
Chris@69:       {
Chris@69:          int fold_start;
Chris@69:          int fold_end;
Chris@69:          int fold_i;
Chris@69:          /* This ensures we never repeat spectral content within one band */
Chris@69:          effective_lowband = IMAX(0, M*eBands[lowband_offset]-norm_offset-N);
Chris@69:          fold_start = lowband_offset;
Chris@69:          while(M*eBands[--fold_start] > effective_lowband+norm_offset);
Chris@69:          fold_end = lowband_offset-1;
Chris@69: #ifndef DISABLE_UPDATE_DRAFT
Chris@69:          while(++fold_end < i && M*eBands[fold_end] < effective_lowband+norm_offset+N);
Chris@69: #else
Chris@69:          while(M*eBands[++fold_end] < effective_lowband+norm_offset+N);
Chris@69: #endif
Chris@69:          x_cm = y_cm = 0;
Chris@69:          fold_i = fold_start; do {
Chris@69:            x_cm |= collapse_masks[fold_i*C+0];
Chris@69:            y_cm |= collapse_masks[fold_i*C+C-1];
Chris@69:          } while (++fold_i<fold_end);
Chris@69:       }
Chris@69:       /* Otherwise, we'll be using the LCG to fold, so all blocks will (almost
Chris@69:          always) be non-zero. */
Chris@69:       else
Chris@69:          x_cm = y_cm = (1<<B)-1;
Chris@69: 
Chris@69:       if (dual_stereo && i==intensity)
Chris@69:       {
Chris@69:          int j;
Chris@69: 
Chris@69:          /* Switch off dual stereo to do intensity. */
Chris@69:          dual_stereo = 0;
Chris@69:          if (resynth)
Chris@69:             for (j=0;j<M*eBands[i]-norm_offset;j++)
Chris@69:                norm[j] = HALF32(norm[j]+norm2[j]);
Chris@69:       }
Chris@69:       if (dual_stereo)
Chris@69:       {
Chris@69:          x_cm = quant_band(&ctx, X, N, b/2, B,
Chris@69:                effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
Chris@69:                last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm);
Chris@69:          y_cm = quant_band(&ctx, Y, N, b/2, B,
Chris@69:                effective_lowband != -1 ? norm2+effective_lowband : NULL, LM,
Chris@69:                last?NULL:norm2+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, y_cm);
Chris@69:       } else {
Chris@69:          if (Y!=NULL)
Chris@69:          {
Chris@69:             if (theta_rdo && i < intensity)
Chris@69:             {
Chris@69:                ec_ctx ec_save, ec_save2;
Chris@69:                struct band_ctx ctx_save, ctx_save2;
Chris@69:                opus_val32 dist0, dist1;
Chris@69:                unsigned cm, cm2;
Chris@69:                int nstart_bytes, nend_bytes, save_bytes;
Chris@69:                unsigned char *bytes_buf;
Chris@69:                unsigned char bytes_save[1275];
Chris@69:                opus_val16 w[2];
Chris@69:                compute_channel_weights(bandE[i], bandE[i+m->nbEBands], w);
Chris@69:                /* Make a copy. */
Chris@69:                cm = x_cm|y_cm;
Chris@69:                ec_save = *ec;
Chris@69:                ctx_save = ctx;
Chris@69:                OPUS_COPY(X_save, X, N);
Chris@69:                OPUS_COPY(Y_save, Y, N);
Chris@69:                /* Encode and round down. */
Chris@69:                ctx.theta_round = -1;
Chris@69:                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
Chris@69:                      effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
Chris@69:                      last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm);
Chris@69:                dist0 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch));
Chris@69: 
Chris@69:                /* Save first result. */
Chris@69:                cm2 = x_cm;
Chris@69:                ec_save2 = *ec;
Chris@69:                ctx_save2 = ctx;
Chris@69:                OPUS_COPY(X_save2, X, N);
Chris@69:                OPUS_COPY(Y_save2, Y, N);
Chris@69:                if (!last)
Chris@69:                   OPUS_COPY(norm_save2, norm+M*eBands[i]-norm_offset, N);
Chris@69:                nstart_bytes = ec_save.offs;
Chris@69:                nend_bytes = ec_save.storage;
Chris@69:                bytes_buf = ec_save.buf+nstart_bytes;
Chris@69:                save_bytes = nend_bytes-nstart_bytes;
Chris@69:                OPUS_COPY(bytes_save, bytes_buf, save_bytes);
Chris@69: 
Chris@69:                /* Restore */
Chris@69:                *ec = ec_save;
Chris@69:                ctx = ctx_save;
Chris@69:                OPUS_COPY(X, X_save, N);
Chris@69:                OPUS_COPY(Y, Y_save, N);
Chris@69: #ifndef DISABLE_UPDATE_DRAFT
Chris@69:                if (i == start+1)
Chris@69:                   special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
Chris@69: #endif
Chris@69:                /* Encode and round up. */
Chris@69:                ctx.theta_round = 1;
Chris@69:                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
Chris@69:                      effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
Chris@69:                      last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm);
Chris@69:                dist1 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch));
Chris@69:                if (dist0 >= dist1) {
Chris@69:                   x_cm = cm2;
Chris@69:                   *ec = ec_save2;
Chris@69:                   ctx = ctx_save2;
Chris@69:                   OPUS_COPY(X, X_save2, N);
Chris@69:                   OPUS_COPY(Y, Y_save2, N);
Chris@69:                   if (!last)
Chris@69:                      OPUS_COPY(norm+M*eBands[i]-norm_offset, norm_save2, N);
Chris@69:                   OPUS_COPY(bytes_buf, bytes_save, save_bytes);
Chris@69:                }
Chris@69:             } else {
Chris@69:                ctx.theta_round = 0;
Chris@69:                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
Chris@69:                      effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
Chris@69:                      last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm);
Chris@69:             }
Chris@69:          } else {
Chris@69:             x_cm = quant_band(&ctx, X, N, b, B,
Chris@69:                   effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
Chris@69:                   last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm);
Chris@69:          }
Chris@69:          y_cm = x_cm;
Chris@69:       }
Chris@69:       collapse_masks[i*C+0] = (unsigned char)x_cm;
Chris@69:       collapse_masks[i*C+C-1] = (unsigned char)y_cm;
Chris@69:       balance += pulses[i] + tell;
Chris@69: 
Chris@69:       /* Update the folding position only as long as we have 1 bit/sample depth. */
Chris@69:       update_lowband = b>(N<<BITRES);
Chris@69:       /* We only need to avoid noise on a split for the first band. After that, we
Chris@69:          have folding. */
Chris@69:       ctx.avoid_split_noise = 0;
Chris@69:    }
Chris@69:    *seed = ctx.seed;
Chris@69: 
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: