cannam@154: /* Copyright (c) 2007-2008 CSIRO
cannam@154:    Copyright (c) 2007-2009 Xiph.Org Foundation
cannam@154:    Copyright (c) 2008-2009 Gregory Maxwell
cannam@154:    Written by Jean-Marc Valin and Gregory Maxwell */
cannam@154: /*
cannam@154:    Redistribution and use in source and binary forms, with or without
cannam@154:    modification, are permitted provided that the following conditions
cannam@154:    are met:
cannam@154: 
cannam@154:    - Redistributions of source code must retain the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer.
cannam@154: 
cannam@154:    - Redistributions in binary form must reproduce the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer in the
cannam@154:    documentation and/or other materials provided with the distribution.
cannam@154: 
cannam@154:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
cannam@154:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
cannam@154:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
cannam@154:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
cannam@154:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
cannam@154:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
cannam@154:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
cannam@154:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
cannam@154:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
cannam@154:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
cannam@154:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
cannam@154: */
cannam@154: 
cannam@154: #ifdef HAVE_CONFIG_H
cannam@154: #include "config.h"
cannam@154: #endif
cannam@154: 
cannam@154: #include <math.h>
cannam@154: #include "bands.h"
cannam@154: #include "modes.h"
cannam@154: #include "vq.h"
cannam@154: #include "cwrs.h"
cannam@154: #include "stack_alloc.h"
cannam@154: #include "os_support.h"
cannam@154: #include "mathops.h"
cannam@154: #include "rate.h"
cannam@154: #include "quant_bands.h"
cannam@154: #include "pitch.h"
cannam@154: 
cannam@154: int hysteresis_decision(opus_val16 val, const opus_val16 *thresholds, const opus_val16 *hysteresis, int N, int prev)
cannam@154: {
cannam@154:    int i;
cannam@154:    for (i=0;i<N;i++)
cannam@154:    {
cannam@154:       if (val < thresholds[i])
cannam@154:          break;
cannam@154:    }
cannam@154:    if (i>prev && val < thresholds[prev]+hysteresis[prev])
cannam@154:       i=prev;
cannam@154:    if (i<prev && val > thresholds[prev-1]-hysteresis[prev-1])
cannam@154:       i=prev;
cannam@154:    return i;
cannam@154: }
cannam@154: 
cannam@154: opus_uint32 celt_lcg_rand(opus_uint32 seed)
cannam@154: {
cannam@154:    return 1664525 * seed + 1013904223;
cannam@154: }
cannam@154: 
cannam@154: /* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness
cannam@154:    with this approximation is important because it has an impact on the bit allocation */
cannam@154: opus_int16 bitexact_cos(opus_int16 x)
cannam@154: {
cannam@154:    opus_int32 tmp;
cannam@154:    opus_int16 x2;
cannam@154:    tmp = (4096+((opus_int32)(x)*(x)))>>13;
cannam@154:    celt_sig_assert(tmp<=32767);
cannam@154:    x2 = tmp;
cannam@154:    x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2)))));
cannam@154:    celt_sig_assert(x2<=32766);
cannam@154:    return 1+x2;
cannam@154: }
cannam@154: 
cannam@154: int bitexact_log2tan(int isin,int icos)
cannam@154: {
cannam@154:    int lc;
cannam@154:    int ls;
cannam@154:    lc=EC_ILOG(icos);
cannam@154:    ls=EC_ILOG(isin);
cannam@154:    icos<<=15-lc;
cannam@154:    isin<<=15-ls;
cannam@154:    return (ls-lc)*(1<<11)
cannam@154:          +FRAC_MUL16(isin, FRAC_MUL16(isin, -2597) + 7932)
cannam@154:          -FRAC_MUL16(icos, FRAC_MUL16(icos, -2597) + 7932);
cannam@154: }
cannam@154: 
cannam@154: #ifdef FIXED_POINT
cannam@154: /* Compute the amplitude (sqrt energy) in each of the bands */
cannam@154: void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch)
cannam@154: {
cannam@154:    int i, c, N;
cannam@154:    const opus_int16 *eBands = m->eBands;
cannam@154:    (void)arch;
cannam@154:    N = m->shortMdctSize<<LM;
cannam@154:    c=0; do {
cannam@154:       for (i=0;i<end;i++)
cannam@154:       {
cannam@154:          int j;
cannam@154:          opus_val32 maxval=0;
cannam@154:          opus_val32 sum = 0;
cannam@154: 
cannam@154:          maxval = celt_maxabs32(&X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM);
cannam@154:          if (maxval > 0)
cannam@154:          {
cannam@154:             int shift = celt_ilog2(maxval) - 14 + (((m->logN[i]>>BITRES)+LM+1)>>1);
cannam@154:             j=eBands[i]<<LM;
cannam@154:             if (shift>0)
cannam@154:             {
cannam@154:                do {
cannam@154:                   sum = MAC16_16(sum, EXTRACT16(SHR32(X[j+c*N],shift)),
cannam@154:                         EXTRACT16(SHR32(X[j+c*N],shift)));
cannam@154:                } while (++j<eBands[i+1]<<LM);
cannam@154:             } else {
cannam@154:                do {
cannam@154:                   sum = MAC16_16(sum, EXTRACT16(SHL32(X[j+c*N],-shift)),
cannam@154:                         EXTRACT16(SHL32(X[j+c*N],-shift)));
cannam@154:                } while (++j<eBands[i+1]<<LM);
cannam@154:             }
cannam@154:             /* We're adding one here to ensure the normalized band isn't larger than unity norm */
cannam@154:             bandE[i+c*m->nbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
cannam@154:          } else {
cannam@154:             bandE[i+c*m->nbEBands] = EPSILON;
cannam@154:          }
cannam@154:          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
cannam@154:       }
cannam@154:    } while (++c<C);
cannam@154:    /*printf ("\n");*/
cannam@154: }
cannam@154: 
cannam@154: /* Normalise each band such that the energy is one. */
cannam@154: 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)
cannam@154: {
cannam@154:    int i, c, N;
cannam@154:    const opus_int16 *eBands = m->eBands;
cannam@154:    N = M*m->shortMdctSize;
cannam@154:    c=0; do {
cannam@154:       i=0; do {
cannam@154:          opus_val16 g;
cannam@154:          int j,shift;
cannam@154:          opus_val16 E;
cannam@154:          shift = celt_zlog2(bandE[i+c*m->nbEBands])-13;
cannam@154:          E = VSHR32(bandE[i+c*m->nbEBands], shift);
cannam@154:          g = EXTRACT16(celt_rcp(SHL32(E,3)));
cannam@154:          j=M*eBands[i]; do {
cannam@154:             X[j+c*N] = MULT16_16_Q15(VSHR32(freq[j+c*N],shift-1),g);
cannam@154:          } while (++j<M*eBands[i+1]);
cannam@154:       } while (++i<end);
cannam@154:    } while (++c<C);
cannam@154: }
cannam@154: 
cannam@154: #else /* FIXED_POINT */
cannam@154: /* Compute the amplitude (sqrt energy) in each of the bands */
cannam@154: void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch)
cannam@154: {
cannam@154:    int i, c, N;
cannam@154:    const opus_int16 *eBands = m->eBands;
cannam@154:    N = m->shortMdctSize<<LM;
cannam@154:    c=0; do {
cannam@154:       for (i=0;i<end;i++)
cannam@154:       {
cannam@154:          opus_val32 sum;
cannam@154:          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);
cannam@154:          bandE[i+c*m->nbEBands] = celt_sqrt(sum);
cannam@154:          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
cannam@154:       }
cannam@154:    } while (++c<C);
cannam@154:    /*printf ("\n");*/
cannam@154: }
cannam@154: 
cannam@154: /* Normalise each band such that the energy is one. */
cannam@154: 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)
cannam@154: {
cannam@154:    int i, c, N;
cannam@154:    const opus_int16 *eBands = m->eBands;
cannam@154:    N = M*m->shortMdctSize;
cannam@154:    c=0; do {
cannam@154:       for (i=0;i<end;i++)
cannam@154:       {
cannam@154:          int j;
cannam@154:          opus_val16 g = 1.f/(1e-27f+bandE[i+c*m->nbEBands]);
cannam@154:          for (j=M*eBands[i];j<M*eBands[i+1];j++)
cannam@154:             X[j+c*N] = freq[j+c*N]*g;
cannam@154:       }
cannam@154:    } while (++c<C);
cannam@154: }
cannam@154: 
cannam@154: #endif /* FIXED_POINT */
cannam@154: 
cannam@154: /* De-normalise the energy to produce the synthesis from the unit-energy bands */
cannam@154: void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X,
cannam@154:       celt_sig * OPUS_RESTRICT freq, const opus_val16 *bandLogE, int start,
cannam@154:       int end, int M, int downsample, int silence)
cannam@154: {
cannam@154:    int i, N;
cannam@154:    int bound;
cannam@154:    celt_sig * OPUS_RESTRICT f;
cannam@154:    const celt_norm * OPUS_RESTRICT x;
cannam@154:    const opus_int16 *eBands = m->eBands;
cannam@154:    N = M*m->shortMdctSize;
cannam@154:    bound = M*eBands[end];
cannam@154:    if (downsample!=1)
cannam@154:       bound = IMIN(bound, N/downsample);
cannam@154:    if (silence)
cannam@154:    {
cannam@154:       bound = 0;
cannam@154:       start = end = 0;
cannam@154:    }
cannam@154:    f = freq;
cannam@154:    x = X+M*eBands[start];
cannam@154:    for (i=0;i<M*eBands[start];i++)
cannam@154:       *f++ = 0;
cannam@154:    for (i=start;i<end;i++)
cannam@154:    {
cannam@154:       int j, band_end;
cannam@154:       opus_val16 g;
cannam@154:       opus_val16 lg;
cannam@154: #ifdef FIXED_POINT
cannam@154:       int shift;
cannam@154: #endif
cannam@154:       j=M*eBands[i];
cannam@154:       band_end = M*eBands[i+1];
cannam@154:       lg = SATURATE16(ADD32(bandLogE[i], SHL32((opus_val32)eMeans[i],6)));
cannam@154: #ifndef FIXED_POINT
cannam@154:       g = celt_exp2(MIN32(32.f, lg));
cannam@154: #else
cannam@154:       /* Handle the integer part of the log energy */
cannam@154:       shift = 16-(lg>>DB_SHIFT);
cannam@154:       if (shift>31)
cannam@154:       {
cannam@154:          shift=0;
cannam@154:          g=0;
cannam@154:       } else {
cannam@154:          /* Handle the fractional part. */
cannam@154:          g = celt_exp2_frac(lg&((1<<DB_SHIFT)-1));
cannam@154:       }
cannam@154:       /* Handle extreme gains with negative shift. */
cannam@154:       if (shift<0)
cannam@154:       {
cannam@154:          /* For shift <= -2 and g > 16384 we'd be likely to overflow, so we're
cannam@154:             capping the gain here, which is equivalent to a cap of 18 on lg.
cannam@154:             This shouldn't trigger unless the bitstream is already corrupted. */
cannam@154:          if (shift <= -2)
cannam@154:          {
cannam@154:             g = 16384;
cannam@154:             shift = -2;
cannam@154:          }
cannam@154:          do {
cannam@154:             *f++ = SHL32(MULT16_16(*x++, g), -shift);
cannam@154:          } while (++j<band_end);
cannam@154:       } else
cannam@154: #endif
cannam@154:          /* Be careful of the fixed-point "else" just above when changing this code */
cannam@154:          do {
cannam@154:             *f++ = SHR32(MULT16_16(*x++, g), shift);
cannam@154:          } while (++j<band_end);
cannam@154:    }
cannam@154:    celt_assert(start <= end);
cannam@154:    OPUS_CLEAR(&freq[bound], N-bound);
cannam@154: }
cannam@154: 
cannam@154: /* This prevents energy collapse for transients with multiple short MDCTs */
cannam@154: void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_masks, int LM, int C, int size,
cannam@154:       int start, int end, const opus_val16 *logE, const opus_val16 *prev1logE,
cannam@154:       const opus_val16 *prev2logE, const int *pulses, opus_uint32 seed, int arch)
cannam@154: {
cannam@154:    int c, i, j, k;
cannam@154:    for (i=start;i<end;i++)
cannam@154:    {
cannam@154:       int N0;
cannam@154:       opus_val16 thresh, sqrt_1;
cannam@154:       int depth;
cannam@154: #ifdef FIXED_POINT
cannam@154:       int shift;
cannam@154:       opus_val32 thresh32;
cannam@154: #endif
cannam@154: 
cannam@154:       N0 = m->eBands[i+1]-m->eBands[i];
cannam@154:       /* depth in 1/8 bits */
cannam@154:       celt_sig_assert(pulses[i]>=0);
cannam@154:       depth = celt_udiv(1+pulses[i], (m->eBands[i+1]-m->eBands[i]))>>LM;
cannam@154: 
cannam@154: #ifdef FIXED_POINT
cannam@154:       thresh32 = SHR32(celt_exp2(-SHL16(depth, 10-BITRES)),1);
cannam@154:       thresh = MULT16_32_Q15(QCONST16(0.5f, 15), MIN32(32767,thresh32));
cannam@154:       {
cannam@154:          opus_val32 t;
cannam@154:          t = N0<<LM;
cannam@154:          shift = celt_ilog2(t)>>1;
cannam@154:          t = SHL32(t, (7-shift)<<1);
cannam@154:          sqrt_1 = celt_rsqrt_norm(t);
cannam@154:       }
cannam@154: #else
cannam@154:       thresh = .5f*celt_exp2(-.125f*depth);
cannam@154:       sqrt_1 = celt_rsqrt(N0<<LM);
cannam@154: #endif
cannam@154: 
cannam@154:       c=0; do
cannam@154:       {
cannam@154:          celt_norm *X;
cannam@154:          opus_val16 prev1;
cannam@154:          opus_val16 prev2;
cannam@154:          opus_val32 Ediff;
cannam@154:          opus_val16 r;
cannam@154:          int renormalize=0;
cannam@154:          prev1 = prev1logE[c*m->nbEBands+i];
cannam@154:          prev2 = prev2logE[c*m->nbEBands+i];
cannam@154:          if (C==1)
cannam@154:          {
cannam@154:             prev1 = MAX16(prev1,prev1logE[m->nbEBands+i]);
cannam@154:             prev2 = MAX16(prev2,prev2logE[m->nbEBands+i]);
cannam@154:          }
cannam@154:          Ediff = EXTEND32(logE[c*m->nbEBands+i])-EXTEND32(MIN16(prev1,prev2));
cannam@154:          Ediff = MAX32(0, Ediff);
cannam@154: 
cannam@154: #ifdef FIXED_POINT
cannam@154:          if (Ediff < 16384)
cannam@154:          {
cannam@154:             opus_val32 r32 = SHR32(celt_exp2(-EXTRACT16(Ediff)),1);
cannam@154:             r = 2*MIN16(16383,r32);
cannam@154:          } else {
cannam@154:             r = 0;
cannam@154:          }
cannam@154:          if (LM==3)
cannam@154:             r = MULT16_16_Q14(23170, MIN32(23169, r));
cannam@154:          r = SHR16(MIN16(thresh, r),1);
cannam@154:          r = SHR32(MULT16_16_Q15(sqrt_1, r),shift);
cannam@154: #else
cannam@154:          /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
cannam@154:             short blocks don't have the same energy as long */
cannam@154:          r = 2.f*celt_exp2(-Ediff);
cannam@154:          if (LM==3)
cannam@154:             r *= 1.41421356f;
cannam@154:          r = MIN16(thresh, r);
cannam@154:          r = r*sqrt_1;
cannam@154: #endif
cannam@154:          X = X_+c*size+(m->eBands[i]<<LM);
cannam@154:          for (k=0;k<1<<LM;k++)
cannam@154:          {
cannam@154:             /* Detect collapse */
cannam@154:             if (!(collapse_masks[i*C+c]&1<<k))
cannam@154:             {
cannam@154:                /* Fill with noise */
cannam@154:                for (j=0;j<N0;j++)
cannam@154:                {
cannam@154:                   seed = celt_lcg_rand(seed);
cannam@154:                   X[(j<<LM)+k] = (seed&0x8000 ? r : -r);
cannam@154:                }
cannam@154:                renormalize = 1;
cannam@154:             }
cannam@154:          }
cannam@154:          /* We just added some energy, so we need to renormalise */
cannam@154:          if (renormalize)
cannam@154:             renormalise_vector(X, N0<<LM, Q15ONE, arch);
cannam@154:       } while (++c<C);
cannam@154:    }
cannam@154: }
cannam@154: 
cannam@154: /* Compute the weights to use for optimizing normalized distortion across
cannam@154:    channels. We use the amplitude to weight square distortion, which means
cannam@154:    that we use the square root of the value we would have been using if we
cannam@154:    wanted to minimize the MSE in the non-normalized domain. This roughly
cannam@154:    corresponds to some quick-and-dirty perceptual experiments I ran to
cannam@154:    measure inter-aural masking (there doesn't seem to be any published data
cannam@154:    on the topic). */
cannam@154: static void compute_channel_weights(celt_ener Ex, celt_ener Ey, opus_val16 w[2])
cannam@154: {
cannam@154:    celt_ener minE;
cannam@154: #if FIXED_POINT
cannam@154:    int shift;
cannam@154: #endif
cannam@154:    minE = MIN32(Ex, Ey);
cannam@154:    /* Adjustment to make the weights a bit more conservative. */
cannam@154:    Ex = ADD32(Ex, minE/3);
cannam@154:    Ey = ADD32(Ey, minE/3);
cannam@154: #if FIXED_POINT
cannam@154:    shift = celt_ilog2(EPSILON+MAX32(Ex, Ey))-14;
cannam@154: #endif
cannam@154:    w[0] = VSHR32(Ex, shift);
cannam@154:    w[1] = VSHR32(Ey, shift);
cannam@154: }
cannam@154: 
cannam@154: 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)
cannam@154: {
cannam@154:    int i = bandID;
cannam@154:    int j;
cannam@154:    opus_val16 a1, a2;
cannam@154:    opus_val16 left, right;
cannam@154:    opus_val16 norm;
cannam@154: #ifdef FIXED_POINT
cannam@154:    int shift = celt_zlog2(MAX32(bandE[i], bandE[i+m->nbEBands]))-13;
cannam@154: #endif
cannam@154:    left = VSHR32(bandE[i],shift);
cannam@154:    right = VSHR32(bandE[i+m->nbEBands],shift);
cannam@154:    norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right));
cannam@154:    a1 = DIV32_16(SHL32(EXTEND32(left),14),norm);
cannam@154:    a2 = DIV32_16(SHL32(EXTEND32(right),14),norm);
cannam@154:    for (j=0;j<N;j++)
cannam@154:    {
cannam@154:       celt_norm r, l;
cannam@154:       l = X[j];
cannam@154:       r = Y[j];
cannam@154:       X[j] = EXTRACT16(SHR32(MAC16_16(MULT16_16(a1, l), a2, r), 14));
cannam@154:       /* Side is not encoded, no need to calculate */
cannam@154:    }
cannam@154: }
cannam@154: 
cannam@154: static void stereo_split(celt_norm * OPUS_RESTRICT X, celt_norm * OPUS_RESTRICT Y, int N)
cannam@154: {
cannam@154:    int j;
cannam@154:    for (j=0;j<N;j++)
cannam@154:    {
cannam@154:       opus_val32 r, l;
cannam@154:       l = MULT16_16(QCONST16(.70710678f, 15), X[j]);
cannam@154:       r = MULT16_16(QCONST16(.70710678f, 15), Y[j]);
cannam@154:       X[j] = EXTRACT16(SHR32(ADD32(l, r), 15));
cannam@154:       Y[j] = EXTRACT16(SHR32(SUB32(r, l), 15));
cannam@154:    }
cannam@154: }
cannam@154: 
cannam@154: static void stereo_merge(celt_norm * OPUS_RESTRICT X, celt_norm * OPUS_RESTRICT Y, opus_val16 mid, int N, int arch)
cannam@154: {
cannam@154:    int j;
cannam@154:    opus_val32 xp=0, side=0;
cannam@154:    opus_val32 El, Er;
cannam@154:    opus_val16 mid2;
cannam@154: #ifdef FIXED_POINT
cannam@154:    int kl, kr;
cannam@154: #endif
cannam@154:    opus_val32 t, lgain, rgain;
cannam@154: 
cannam@154:    /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */
cannam@154:    dual_inner_prod(Y, X, Y, N, &xp, &side, arch);
cannam@154:    /* Compensating for the mid normalization */
cannam@154:    xp = MULT16_32_Q15(mid, xp);
cannam@154:    /* mid and side are in Q15, not Q14 like X and Y */
cannam@154:    mid2 = SHR16(mid, 1);
cannam@154:    El = MULT16_16(mid2, mid2) + side - 2*xp;
cannam@154:    Er = MULT16_16(mid2, mid2) + side + 2*xp;
cannam@154:    if (Er < QCONST32(6e-4f, 28) || El < QCONST32(6e-4f, 28))
cannam@154:    {
cannam@154:       OPUS_COPY(Y, X, N);
cannam@154:       return;
cannam@154:    }
cannam@154: 
cannam@154: #ifdef FIXED_POINT
cannam@154:    kl = celt_ilog2(El)>>1;
cannam@154:    kr = celt_ilog2(Er)>>1;
cannam@154: #endif
cannam@154:    t = VSHR32(El, (kl-7)<<1);
cannam@154:    lgain = celt_rsqrt_norm(t);
cannam@154:    t = VSHR32(Er, (kr-7)<<1);
cannam@154:    rgain = celt_rsqrt_norm(t);
cannam@154: 
cannam@154: #ifdef FIXED_POINT
cannam@154:    if (kl < 7)
cannam@154:       kl = 7;
cannam@154:    if (kr < 7)
cannam@154:       kr = 7;
cannam@154: #endif
cannam@154: 
cannam@154:    for (j=0;j<N;j++)
cannam@154:    {
cannam@154:       celt_norm r, l;
cannam@154:       /* Apply mid scaling (side is already scaled) */
cannam@154:       l = MULT16_16_P15(mid, X[j]);
cannam@154:       r = Y[j];
cannam@154:       X[j] = EXTRACT16(PSHR32(MULT16_16(lgain, SUB16(l,r)), kl+1));
cannam@154:       Y[j] = EXTRACT16(PSHR32(MULT16_16(rgain, ADD16(l,r)), kr+1));
cannam@154:    }
cannam@154: }
cannam@154: 
cannam@154: /* Decide whether we should spread the pulses in the current frame */
cannam@154: int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
cannam@154:       int last_decision, int *hf_average, int *tapset_decision, int update_hf,
cannam@154:       int end, int C, int M, const int *spread_weight)
cannam@154: {
cannam@154:    int i, c, N0;
cannam@154:    int sum = 0, nbBands=0;
cannam@154:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
cannam@154:    int decision;
cannam@154:    int hf_sum=0;
cannam@154: 
cannam@154:    celt_assert(end>0);
cannam@154: 
cannam@154:    N0 = M*m->shortMdctSize;
cannam@154: 
cannam@154:    if (M*(eBands[end]-eBands[end-1]) <= 8)
cannam@154:       return SPREAD_NONE;
cannam@154:    c=0; do {
cannam@154:       for (i=0;i<end;i++)
cannam@154:       {
cannam@154:          int j, N, tmp=0;
cannam@154:          int tcount[3] = {0,0,0};
cannam@154:          const celt_norm * OPUS_RESTRICT x = X+M*eBands[i]+c*N0;
cannam@154:          N = M*(eBands[i+1]-eBands[i]);
cannam@154:          if (N<=8)
cannam@154:             continue;
cannam@154:          /* Compute rough CDF of |x[j]| */
cannam@154:          for (j=0;j<N;j++)
cannam@154:          {
cannam@154:             opus_val32 x2N; /* Q13 */
cannam@154: 
cannam@154:             x2N = MULT16_16(MULT16_16_Q15(x[j], x[j]), N);
cannam@154:             if (x2N < QCONST16(0.25f,13))
cannam@154:                tcount[0]++;
cannam@154:             if (x2N < QCONST16(0.0625f,13))
cannam@154:                tcount[1]++;
cannam@154:             if (x2N < QCONST16(0.015625f,13))
cannam@154:                tcount[2]++;
cannam@154:          }
cannam@154: 
cannam@154:          /* Only include four last bands (8 kHz and up) */
cannam@154:          if (i>m->nbEBands-4)
cannam@154:             hf_sum += celt_udiv(32*(tcount[1]+tcount[0]), N);
cannam@154:          tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N);
cannam@154:          sum += tmp*spread_weight[i];
cannam@154:          nbBands+=spread_weight[i];
cannam@154:       }
cannam@154:    } while (++c<C);
cannam@154: 
cannam@154:    if (update_hf)
cannam@154:    {
cannam@154:       if (hf_sum)
cannam@154:          hf_sum = celt_udiv(hf_sum, C*(4-m->nbEBands+end));
cannam@154:       *hf_average = (*hf_average+hf_sum)>>1;
cannam@154:       hf_sum = *hf_average;
cannam@154:       if (*tapset_decision==2)
cannam@154:          hf_sum += 4;
cannam@154:       else if (*tapset_decision==0)
cannam@154:          hf_sum -= 4;
cannam@154:       if (hf_sum > 22)
cannam@154:          *tapset_decision=2;
cannam@154:       else if (hf_sum > 18)
cannam@154:          *tapset_decision=1;
cannam@154:       else
cannam@154:          *tapset_decision=0;
cannam@154:    }
cannam@154:    /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/
cannam@154:    celt_assert(nbBands>0); /* end has to be non-zero */
cannam@154:    celt_assert(sum>=0);
cannam@154:    sum = celt_udiv((opus_int32)sum<<8, nbBands);
cannam@154:    /* Recursive averaging */
cannam@154:    sum = (sum+*average)>>1;
cannam@154:    *average = sum;
cannam@154:    /* Hysteresis */
cannam@154:    sum = (3*sum + (((3-last_decision)<<7) + 64) + 2)>>2;
cannam@154:    if (sum < 80)
cannam@154:    {
cannam@154:       decision = SPREAD_AGGRESSIVE;
cannam@154:    } else if (sum < 256)
cannam@154:    {
cannam@154:       decision = SPREAD_NORMAL;
cannam@154:    } else if (sum < 384)
cannam@154:    {
cannam@154:       decision = SPREAD_LIGHT;
cannam@154:    } else {
cannam@154:       decision = SPREAD_NONE;
cannam@154:    }
cannam@154: #ifdef FUZZING
cannam@154:    decision = rand()&0x3;
cannam@154:    *tapset_decision=rand()%3;
cannam@154: #endif
cannam@154:    return decision;
cannam@154: }
cannam@154: 
cannam@154: /* Indexing table for converting from natural Hadamard to ordery Hadamard
cannam@154:    This is essentially a bit-reversed Gray, on top of which we've added
cannam@154:    an inversion of the order because we want the DC at the end rather than
cannam@154:    the beginning. The lines are for N=2, 4, 8, 16 */
cannam@154: static const int ordery_table[] = {
cannam@154:        1,  0,
cannam@154:        3,  0,  2,  1,
cannam@154:        7,  0,  4,  3,  6,  1,  5,  2,
cannam@154:       15,  0,  8,  7, 12,  3, 11,  4, 14,  1,  9,  6, 13,  2, 10,  5,
cannam@154: };
cannam@154: 
cannam@154: static void deinterleave_hadamard(celt_norm *X, int N0, int stride, int hadamard)
cannam@154: {
cannam@154:    int i,j;
cannam@154:    VARDECL(celt_norm, tmp);
cannam@154:    int N;
cannam@154:    SAVE_STACK;
cannam@154:    N = N0*stride;
cannam@154:    ALLOC(tmp, N, celt_norm);
cannam@154:    celt_assert(stride>0);
cannam@154:    if (hadamard)
cannam@154:    {
cannam@154:       const int *ordery = ordery_table+stride-2;
cannam@154:       for (i=0;i<stride;i++)
cannam@154:       {
cannam@154:          for (j=0;j<N0;j++)
cannam@154:             tmp[ordery[i]*N0+j] = X[j*stride+i];
cannam@154:       }
cannam@154:    } else {
cannam@154:       for (i=0;i<stride;i++)
cannam@154:          for (j=0;j<N0;j++)
cannam@154:             tmp[i*N0+j] = X[j*stride+i];
cannam@154:    }
cannam@154:    OPUS_COPY(X, tmp, N);
cannam@154:    RESTORE_STACK;
cannam@154: }
cannam@154: 
cannam@154: static void interleave_hadamard(celt_norm *X, int N0, int stride, int hadamard)
cannam@154: {
cannam@154:    int i,j;
cannam@154:    VARDECL(celt_norm, tmp);
cannam@154:    int N;
cannam@154:    SAVE_STACK;
cannam@154:    N = N0*stride;
cannam@154:    ALLOC(tmp, N, celt_norm);
cannam@154:    if (hadamard)
cannam@154:    {
cannam@154:       const int *ordery = ordery_table+stride-2;
cannam@154:       for (i=0;i<stride;i++)
cannam@154:          for (j=0;j<N0;j++)
cannam@154:             tmp[j*stride+i] = X[ordery[i]*N0+j];
cannam@154:    } else {
cannam@154:       for (i=0;i<stride;i++)
cannam@154:          for (j=0;j<N0;j++)
cannam@154:             tmp[j*stride+i] = X[i*N0+j];
cannam@154:    }
cannam@154:    OPUS_COPY(X, tmp, N);
cannam@154:    RESTORE_STACK;
cannam@154: }
cannam@154: 
cannam@154: void haar1(celt_norm *X, int N0, int stride)
cannam@154: {
cannam@154:    int i, j;
cannam@154:    N0 >>= 1;
cannam@154:    for (i=0;i<stride;i++)
cannam@154:       for (j=0;j<N0;j++)
cannam@154:       {
cannam@154:          opus_val32 tmp1, tmp2;
cannam@154:          tmp1 = MULT16_16(QCONST16(.70710678f,15), X[stride*2*j+i]);
cannam@154:          tmp2 = MULT16_16(QCONST16(.70710678f,15), X[stride*(2*j+1)+i]);
cannam@154:          X[stride*2*j+i] = EXTRACT16(PSHR32(ADD32(tmp1, tmp2), 15));
cannam@154:          X[stride*(2*j+1)+i] = EXTRACT16(PSHR32(SUB32(tmp1, tmp2), 15));
cannam@154:       }
cannam@154: }
cannam@154: 
cannam@154: static int compute_qn(int N, int b, int offset, int pulse_cap, int stereo)
cannam@154: {
cannam@154:    static const opus_int16 exp2_table8[8] =
cannam@154:       {16384, 17866, 19483, 21247, 23170, 25267, 27554, 30048};
cannam@154:    int qn, qb;
cannam@154:    int N2 = 2*N-1;
cannam@154:    if (stereo && N==2)
cannam@154:       N2--;
cannam@154:    /* The upper limit ensures that in a stereo split with itheta==16384, we'll
cannam@154:        always have enough bits left over to code at least one pulse in the
cannam@154:        side; otherwise it would collapse, since it doesn't get folded. */
cannam@154:    qb = celt_sudiv(b+N2*offset, N2);
cannam@154:    qb = IMIN(b-pulse_cap-(4<<BITRES), qb);
cannam@154: 
cannam@154:    qb = IMIN(8<<BITRES, qb);
cannam@154: 
cannam@154:    if (qb<(1<<BITRES>>1)) {
cannam@154:       qn = 1;
cannam@154:    } else {
cannam@154:       qn = exp2_table8[qb&0x7]>>(14-(qb>>BITRES));
cannam@154:       qn = (qn+1)>>1<<1;
cannam@154:    }
cannam@154:    celt_assert(qn <= 256);
cannam@154:    return qn;
cannam@154: }
cannam@154: 
cannam@154: struct band_ctx {
cannam@154:    int encode;
cannam@154:    int resynth;
cannam@154:    const CELTMode *m;
cannam@154:    int i;
cannam@154:    int intensity;
cannam@154:    int spread;
cannam@154:    int tf_change;
cannam@154:    ec_ctx *ec;
cannam@154:    opus_int32 remaining_bits;
cannam@154:    const celt_ener *bandE;
cannam@154:    opus_uint32 seed;
cannam@154:    int arch;
cannam@154:    int theta_round;
cannam@154:    int disable_inv;
cannam@154:    int avoid_split_noise;
cannam@154: };
cannam@154: 
cannam@154: struct split_ctx {
cannam@154:    int inv;
cannam@154:    int imid;
cannam@154:    int iside;
cannam@154:    int delta;
cannam@154:    int itheta;
cannam@154:    int qalloc;
cannam@154: };
cannam@154: 
cannam@154: static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx,
cannam@154:       celt_norm *X, celt_norm *Y, int N, int *b, int B, int B0,
cannam@154:       int LM,
cannam@154:       int stereo, int *fill)
cannam@154: {
cannam@154:    int qn;
cannam@154:    int itheta=0;
cannam@154:    int delta;
cannam@154:    int imid, iside;
cannam@154:    int qalloc;
cannam@154:    int pulse_cap;
cannam@154:    int offset;
cannam@154:    opus_int32 tell;
cannam@154:    int inv=0;
cannam@154:    int encode;
cannam@154:    const CELTMode *m;
cannam@154:    int i;
cannam@154:    int intensity;
cannam@154:    ec_ctx *ec;
cannam@154:    const celt_ener *bandE;
cannam@154: 
cannam@154:    encode = ctx->encode;
cannam@154:    m = ctx->m;
cannam@154:    i = ctx->i;
cannam@154:    intensity = ctx->intensity;
cannam@154:    ec = ctx->ec;
cannam@154:    bandE = ctx->bandE;
cannam@154: 
cannam@154:    /* Decide on the resolution to give to the split parameter theta */
cannam@154:    pulse_cap = m->logN[i]+LM*(1<<BITRES);
cannam@154:    offset = (pulse_cap>>1) - (stereo&&N==2 ? QTHETA_OFFSET_TWOPHASE : QTHETA_OFFSET);
cannam@154:    qn = compute_qn(N, *b, offset, pulse_cap, stereo);
cannam@154:    if (stereo && i>=intensity)
cannam@154:       qn = 1;
cannam@154:    if (encode)
cannam@154:    {
cannam@154:       /* theta is the atan() of the ratio between the (normalized)
cannam@154:          side and mid. With just that parameter, we can re-scale both
cannam@154:          mid and side because we know that 1) they have unit norm and
cannam@154:          2) they are orthogonal. */
cannam@154:       itheta = stereo_itheta(X, Y, stereo, N, ctx->arch);
cannam@154:    }
cannam@154:    tell = ec_tell_frac(ec);
cannam@154:    if (qn!=1)
cannam@154:    {
cannam@154:       if (encode)
cannam@154:       {
cannam@154:          if (!stereo || ctx->theta_round == 0)
cannam@154:          {
cannam@154:             itheta = (itheta*(opus_int32)qn+8192)>>14;
cannam@154:             if (!stereo && ctx->avoid_split_noise && itheta > 0 && itheta < qn)
cannam@154:             {
cannam@154:                /* Check if the selected value of theta will cause the bit allocation
cannam@154:                   to inject noise on one side. If so, make sure the energy of that side
cannam@154:                   is zero. */
cannam@154:                int unquantized = celt_udiv((opus_int32)itheta*16384, qn);
cannam@154:                imid = bitexact_cos((opus_int16)unquantized);
cannam@154:                iside = bitexact_cos((opus_int16)(16384-unquantized));
cannam@154:                delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid));
cannam@154:                if (delta > *b)
cannam@154:                   itheta = qn;
cannam@154:                else if (delta < -*b)
cannam@154:                   itheta = 0;
cannam@154:             }
cannam@154:          } else {
cannam@154:             int down;
cannam@154:             /* Bias quantization towards itheta=0 and itheta=16384. */
cannam@154:             int bias = itheta > 8192 ? 32767/qn : -32767/qn;
cannam@154:             down = IMIN(qn-1, IMAX(0, (itheta*(opus_int32)qn + bias)>>14));
cannam@154:             if (ctx->theta_round < 0)
cannam@154:                itheta = down;
cannam@154:             else
cannam@154:                itheta = down+1;
cannam@154:          }
cannam@154:       }
cannam@154:       /* Entropy coding of the angle. We use a uniform pdf for the
cannam@154:          time split, a step for stereo, and a triangular one for the rest. */
cannam@154:       if (stereo && N>2)
cannam@154:       {
cannam@154:          int p0 = 3;
cannam@154:          int x = itheta;
cannam@154:          int x0 = qn/2;
cannam@154:          int ft = p0*(x0+1) + x0;
cannam@154:          /* Use a probability of p0 up to itheta=8192 and then use 1 after */
cannam@154:          if (encode)
cannam@154:          {
cannam@154:             ec_encode(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft);
cannam@154:          } else {
cannam@154:             int fs;
cannam@154:             fs=ec_decode(ec,ft);
cannam@154:             if (fs<(x0+1)*p0)
cannam@154:                x=fs/p0;
cannam@154:             else
cannam@154:                x=x0+1+(fs-(x0+1)*p0);
cannam@154:             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);
cannam@154:             itheta = x;
cannam@154:          }
cannam@154:       } else if (B0>1 || stereo) {
cannam@154:          /* Uniform pdf */
cannam@154:          if (encode)
cannam@154:             ec_enc_uint(ec, itheta, qn+1);
cannam@154:          else
cannam@154:             itheta = ec_dec_uint(ec, qn+1);
cannam@154:       } else {
cannam@154:          int fs=1, ft;
cannam@154:          ft = ((qn>>1)+1)*((qn>>1)+1);
cannam@154:          if (encode)
cannam@154:          {
cannam@154:             int fl;
cannam@154: 
cannam@154:             fs = itheta <= (qn>>1) ? itheta + 1 : qn + 1 - itheta;
cannam@154:             fl = itheta <= (qn>>1) ? itheta*(itheta + 1)>>1 :
cannam@154:              ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1);
cannam@154: 
cannam@154:             ec_encode(ec, fl, fl+fs, ft);
cannam@154:          } else {
cannam@154:             /* Triangular pdf */
cannam@154:             int fl=0;
cannam@154:             int fm;
cannam@154:             fm = ec_decode(ec, ft);
cannam@154: 
cannam@154:             if (fm < ((qn>>1)*((qn>>1) + 1)>>1))
cannam@154:             {
cannam@154:                itheta = (isqrt32(8*(opus_uint32)fm + 1) - 1)>>1;
cannam@154:                fs = itheta + 1;
cannam@154:                fl = itheta*(itheta + 1)>>1;
cannam@154:             }
cannam@154:             else
cannam@154:             {
cannam@154:                itheta = (2*(qn + 1)
cannam@154:                 - isqrt32(8*(opus_uint32)(ft - fm - 1) + 1))>>1;
cannam@154:                fs = qn + 1 - itheta;
cannam@154:                fl = ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1);
cannam@154:             }
cannam@154: 
cannam@154:             ec_dec_update(ec, fl, fl+fs, ft);
cannam@154:          }
cannam@154:       }
cannam@154:       celt_assert(itheta>=0);
cannam@154:       itheta = celt_udiv((opus_int32)itheta*16384, qn);
cannam@154:       if (encode && stereo)
cannam@154:       {
cannam@154:          if (itheta==0)
cannam@154:             intensity_stereo(m, X, Y, bandE, i, N);
cannam@154:          else
cannam@154:             stereo_split(X, Y, N);
cannam@154:       }
cannam@154:       /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate.
cannam@154:                Let's do that at higher complexity */
cannam@154:    } else if (stereo) {
cannam@154:       if (encode)
cannam@154:       {
cannam@154:          inv = itheta > 8192 && !ctx->disable_inv;
cannam@154:          if (inv)
cannam@154:          {
cannam@154:             int j;
cannam@154:             for (j=0;j<N;j++)
cannam@154:                Y[j] = -Y[j];
cannam@154:          }
cannam@154:          intensity_stereo(m, X, Y, bandE, i, N);
cannam@154:       }
cannam@154:       if (*b>2<<BITRES && ctx->remaining_bits > 2<<BITRES)
cannam@154:       {
cannam@154:          if (encode)
cannam@154:             ec_enc_bit_logp(ec, inv, 2);
cannam@154:          else
cannam@154:             inv = ec_dec_bit_logp(ec, 2);
cannam@154:       } else
cannam@154:          inv = 0;
cannam@154:       /* inv flag override to avoid problems with downmixing. */
cannam@154:       if (ctx->disable_inv)
cannam@154:          inv = 0;
cannam@154:       itheta = 0;
cannam@154:    }
cannam@154:    qalloc = ec_tell_frac(ec) - tell;
cannam@154:    *b -= qalloc;
cannam@154: 
cannam@154:    if (itheta == 0)
cannam@154:    {
cannam@154:       imid = 32767;
cannam@154:       iside = 0;
cannam@154:       *fill &= (1<<B)-1;
cannam@154:       delta = -16384;
cannam@154:    } else if (itheta == 16384)
cannam@154:    {
cannam@154:       imid = 0;
cannam@154:       iside = 32767;
cannam@154:       *fill &= ((1<<B)-1)<<B;
cannam@154:       delta = 16384;
cannam@154:    } else {
cannam@154:       imid = bitexact_cos((opus_int16)itheta);
cannam@154:       iside = bitexact_cos((opus_int16)(16384-itheta));
cannam@154:       /* This is the mid vs side allocation that minimizes squared error
cannam@154:          in that band. */
cannam@154:       delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid));
cannam@154:    }
cannam@154: 
cannam@154:    sctx->inv = inv;
cannam@154:    sctx->imid = imid;
cannam@154:    sctx->iside = iside;
cannam@154:    sctx->delta = delta;
cannam@154:    sctx->itheta = itheta;
cannam@154:    sctx->qalloc = qalloc;
cannam@154: }
cannam@154: static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, int b,
cannam@154:       celt_norm *lowband_out)
cannam@154: {
cannam@154:    int c;
cannam@154:    int stereo;
cannam@154:    celt_norm *x = X;
cannam@154:    int encode;
cannam@154:    ec_ctx *ec;
cannam@154: 
cannam@154:    encode = ctx->encode;
cannam@154:    ec = ctx->ec;
cannam@154: 
cannam@154:    stereo = Y != NULL;
cannam@154:    c=0; do {
cannam@154:       int sign=0;
cannam@154:       if (ctx->remaining_bits>=1<<BITRES)
cannam@154:       {
cannam@154:          if (encode)
cannam@154:          {
cannam@154:             sign = x[0]<0;
cannam@154:             ec_enc_bits(ec, sign, 1);
cannam@154:          } else {
cannam@154:             sign = ec_dec_bits(ec, 1);
cannam@154:          }
cannam@154:          ctx->remaining_bits -= 1<<BITRES;
cannam@154:          b-=1<<BITRES;
cannam@154:       }
cannam@154:       if (ctx->resynth)
cannam@154:          x[0] = sign ? -NORM_SCALING : NORM_SCALING;
cannam@154:       x = Y;
cannam@154:    } while (++c<1+stereo);
cannam@154:    if (lowband_out)
cannam@154:       lowband_out[0] = SHR16(X[0],4);
cannam@154:    return 1;
cannam@154: }
cannam@154: 
cannam@154: /* This function is responsible for encoding and decoding a mono partition.
cannam@154:    It can split the band in two and transmit the energy difference with
cannam@154:    the two half-bands. It can be called recursively so bands can end up being
cannam@154:    split in 8 parts. */
cannam@154: static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X,
cannam@154:       int N, int b, int B, celt_norm *lowband,
cannam@154:       int LM,
cannam@154:       opus_val16 gain, int fill)
cannam@154: {
cannam@154:    const unsigned char *cache;
cannam@154:    int q;
cannam@154:    int curr_bits;
cannam@154:    int imid=0, iside=0;
cannam@154:    int B0=B;
cannam@154:    opus_val16 mid=0, side=0;
cannam@154:    unsigned cm=0;
cannam@154:    celt_norm *Y=NULL;
cannam@154:    int encode;
cannam@154:    const CELTMode *m;
cannam@154:    int i;
cannam@154:    int spread;
cannam@154:    ec_ctx *ec;
cannam@154: 
cannam@154:    encode = ctx->encode;
cannam@154:    m = ctx->m;
cannam@154:    i = ctx->i;
cannam@154:    spread = ctx->spread;
cannam@154:    ec = ctx->ec;
cannam@154: 
cannam@154:    /* If we need 1.5 more bit than we can produce, split the band in two. */
cannam@154:    cache = m->cache.bits + m->cache.index[(LM+1)*m->nbEBands+i];
cannam@154:    if (LM != -1 && b > cache[cache[0]]+12 && N>2)
cannam@154:    {
cannam@154:       int mbits, sbits, delta;
cannam@154:       int itheta;
cannam@154:       int qalloc;
cannam@154:       struct split_ctx sctx;
cannam@154:       celt_norm *next_lowband2=NULL;
cannam@154:       opus_int32 rebalance;
cannam@154: 
cannam@154:       N >>= 1;
cannam@154:       Y = X+N;
cannam@154:       LM -= 1;
cannam@154:       if (B==1)
cannam@154:          fill = (fill&1)|(fill<<1);
cannam@154:       B = (B+1)>>1;
cannam@154: 
cannam@154:       compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, LM, 0, &fill);
cannam@154:       imid = sctx.imid;
cannam@154:       iside = sctx.iside;
cannam@154:       delta = sctx.delta;
cannam@154:       itheta = sctx.itheta;
cannam@154:       qalloc = sctx.qalloc;
cannam@154: #ifdef FIXED_POINT
cannam@154:       mid = imid;
cannam@154:       side = iside;
cannam@154: #else
cannam@154:       mid = (1.f/32768)*imid;
cannam@154:       side = (1.f/32768)*iside;
cannam@154: #endif
cannam@154: 
cannam@154:       /* Give more bits to low-energy MDCTs than they would otherwise deserve */
cannam@154:       if (B0>1 && (itheta&0x3fff))
cannam@154:       {
cannam@154:          if (itheta > 8192)
cannam@154:             /* Rough approximation for pre-echo masking */
cannam@154:             delta -= delta>>(4-LM);
cannam@154:          else
cannam@154:             /* Corresponds to a forward-masking slope of 1.5 dB per 10 ms */
cannam@154:             delta = IMIN(0, delta + (N<<BITRES>>(5-LM)));
cannam@154:       }
cannam@154:       mbits = IMAX(0, IMIN(b, (b-delta)/2));
cannam@154:       sbits = b-mbits;
cannam@154:       ctx->remaining_bits -= qalloc;
cannam@154: 
cannam@154:       if (lowband)
cannam@154:          next_lowband2 = lowband+N; /* >32-bit split case */
cannam@154: 
cannam@154:       rebalance = ctx->remaining_bits;
cannam@154:       if (mbits >= sbits)
cannam@154:       {
cannam@154:          cm = quant_partition(ctx, X, N, mbits, B, lowband, LM,
cannam@154:                MULT16_16_P15(gain,mid), fill);
cannam@154:          rebalance = mbits - (rebalance-ctx->remaining_bits);
cannam@154:          if (rebalance > 3<<BITRES && itheta!=0)
cannam@154:             sbits += rebalance - (3<<BITRES);
cannam@154:          cm |= quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM,
cannam@154:                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
cannam@154:       } else {
cannam@154:          cm = quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM,
cannam@154:                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
cannam@154:          rebalance = sbits - (rebalance-ctx->remaining_bits);
cannam@154:          if (rebalance > 3<<BITRES && itheta!=16384)
cannam@154:             mbits += rebalance - (3<<BITRES);
cannam@154:          cm |= quant_partition(ctx, X, N, mbits, B, lowband, LM,
cannam@154:                MULT16_16_P15(gain,mid), fill);
cannam@154:       }
cannam@154:    } else {
cannam@154:       /* This is the basic no-split case */
cannam@154:       q = bits2pulses(m, i, LM, b);
cannam@154:       curr_bits = pulses2bits(m, i, LM, q);
cannam@154:       ctx->remaining_bits -= curr_bits;
cannam@154: 
cannam@154:       /* Ensures we can never bust the budget */
cannam@154:       while (ctx->remaining_bits < 0 && q > 0)
cannam@154:       {
cannam@154:          ctx->remaining_bits += curr_bits;
cannam@154:          q--;
cannam@154:          curr_bits = pulses2bits(m, i, LM, q);
cannam@154:          ctx->remaining_bits -= curr_bits;
cannam@154:       }
cannam@154: 
cannam@154:       if (q!=0)
cannam@154:       {
cannam@154:          int K = get_pulses(q);
cannam@154: 
cannam@154:          /* Finally do the actual quantization */
cannam@154:          if (encode)
cannam@154:          {
cannam@154:             cm = alg_quant(X, N, K, spread, B, ec, gain, ctx->resynth, ctx->arch);
cannam@154:          } else {
cannam@154:             cm = alg_unquant(X, N, K, spread, B, ec, gain);
cannam@154:          }
cannam@154:       } else {
cannam@154:          /* If there's no pulse, fill the band anyway */
cannam@154:          int j;
cannam@154:          if (ctx->resynth)
cannam@154:          {
cannam@154:             unsigned cm_mask;
cannam@154:             /* B can be as large as 16, so this shift might overflow an int on a
cannam@154:                16-bit platform; use a long to get defined behavior.*/
cannam@154:             cm_mask = (unsigned)(1UL<<B)-1;
cannam@154:             fill &= cm_mask;
cannam@154:             if (!fill)
cannam@154:             {
cannam@154:                OPUS_CLEAR(X, N);
cannam@154:             } else {
cannam@154:                if (lowband == NULL)
cannam@154:                {
cannam@154:                   /* Noise */
cannam@154:                   for (j=0;j<N;j++)
cannam@154:                   {
cannam@154:                      ctx->seed = celt_lcg_rand(ctx->seed);
cannam@154:                      X[j] = (celt_norm)((opus_int32)ctx->seed>>20);
cannam@154:                   }
cannam@154:                   cm = cm_mask;
cannam@154:                } else {
cannam@154:                   /* Folded spectrum */
cannam@154:                   for (j=0;j<N;j++)
cannam@154:                   {
cannam@154:                      opus_val16 tmp;
cannam@154:                      ctx->seed = celt_lcg_rand(ctx->seed);
cannam@154:                      /* About 48 dB below the "normal" folding level */
cannam@154:                      tmp = QCONST16(1.0f/256, 10);
cannam@154:                      tmp = (ctx->seed)&0x8000 ? tmp : -tmp;
cannam@154:                      X[j] = lowband[j]+tmp;
cannam@154:                   }
cannam@154:                   cm = fill;
cannam@154:                }
cannam@154:                renormalise_vector(X, N, gain, ctx->arch);
cannam@154:             }
cannam@154:          }
cannam@154:       }
cannam@154:    }
cannam@154: 
cannam@154:    return cm;
cannam@154: }
cannam@154: 
cannam@154: 
cannam@154: /* This function is responsible for encoding and decoding a band for the mono case. */
cannam@154: static unsigned quant_band(struct band_ctx *ctx, celt_norm *X,
cannam@154:       int N, int b, int B, celt_norm *lowband,
cannam@154:       int LM, celt_norm *lowband_out,
cannam@154:       opus_val16 gain, celt_norm *lowband_scratch, int fill)
cannam@154: {
cannam@154:    int N0=N;
cannam@154:    int N_B=N;
cannam@154:    int N_B0;
cannam@154:    int B0=B;
cannam@154:    int time_divide=0;
cannam@154:    int recombine=0;
cannam@154:    int longBlocks;
cannam@154:    unsigned cm=0;
cannam@154:    int k;
cannam@154:    int encode;
cannam@154:    int tf_change;
cannam@154: 
cannam@154:    encode = ctx->encode;
cannam@154:    tf_change = ctx->tf_change;
cannam@154: 
cannam@154:    longBlocks = B0==1;
cannam@154: 
cannam@154:    N_B = celt_udiv(N_B, B);
cannam@154: 
cannam@154:    /* Special case for one sample */
cannam@154:    if (N==1)
cannam@154:    {
cannam@154:       return quant_band_n1(ctx, X, NULL, b, lowband_out);
cannam@154:    }
cannam@154: 
cannam@154:    if (tf_change>0)
cannam@154:       recombine = tf_change;
cannam@154:    /* Band recombining to increase frequency resolution */
cannam@154: 
cannam@154:    if (lowband_scratch && lowband && (recombine || ((N_B&1) == 0 && tf_change<0) || B0>1))
cannam@154:    {
cannam@154:       OPUS_COPY(lowband_scratch, lowband, N);
cannam@154:       lowband = lowband_scratch;
cannam@154:    }
cannam@154: 
cannam@154:    for (k=0;k<recombine;k++)
cannam@154:    {
cannam@154:       static const unsigned char bit_interleave_table[16]={
cannam@154:             0,1,1,1,2,3,3,3,2,3,3,3,2,3,3,3
cannam@154:       };
cannam@154:       if (encode)
cannam@154:          haar1(X, N>>k, 1<<k);
cannam@154:       if (lowband)
cannam@154:          haar1(lowband, N>>k, 1<<k);
cannam@154:       fill = bit_interleave_table[fill&0xF]|bit_interleave_table[fill>>4]<<2;
cannam@154:    }
cannam@154:    B>>=recombine;
cannam@154:    N_B<<=recombine;
cannam@154: 
cannam@154:    /* Increasing the time resolution */
cannam@154:    while ((N_B&1) == 0 && tf_change<0)
cannam@154:    {
cannam@154:       if (encode)
cannam@154:          haar1(X, N_B, B);
cannam@154:       if (lowband)
cannam@154:          haar1(lowband, N_B, B);
cannam@154:       fill |= fill<<B;
cannam@154:       B <<= 1;
cannam@154:       N_B >>= 1;
cannam@154:       time_divide++;
cannam@154:       tf_change++;
cannam@154:    }
cannam@154:    B0=B;
cannam@154:    N_B0 = N_B;
cannam@154: 
cannam@154:    /* Reorganize the samples in time order instead of frequency order */
cannam@154:    if (B0>1)
cannam@154:    {
cannam@154:       if (encode)
cannam@154:          deinterleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
cannam@154:       if (lowband)
cannam@154:          deinterleave_hadamard(lowband, N_B>>recombine, B0<<recombine, longBlocks);
cannam@154:    }
cannam@154: 
cannam@154:    cm = quant_partition(ctx, X, N, b, B, lowband, LM, gain, fill);
cannam@154: 
cannam@154:    /* This code is used by the decoder and by the resynthesis-enabled encoder */
cannam@154:    if (ctx->resynth)
cannam@154:    {
cannam@154:       /* Undo the sample reorganization going from time order to frequency order */
cannam@154:       if (B0>1)
cannam@154:          interleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
cannam@154: 
cannam@154:       /* Undo time-freq changes that we did earlier */
cannam@154:       N_B = N_B0;
cannam@154:       B = B0;
cannam@154:       for (k=0;k<time_divide;k++)
cannam@154:       {
cannam@154:          B >>= 1;
cannam@154:          N_B <<= 1;
cannam@154:          cm |= cm>>B;
cannam@154:          haar1(X, N_B, B);
cannam@154:       }
cannam@154: 
cannam@154:       for (k=0;k<recombine;k++)
cannam@154:       {
cannam@154:          static const unsigned char bit_deinterleave_table[16]={
cannam@154:                0x00,0x03,0x0C,0x0F,0x30,0x33,0x3C,0x3F,
cannam@154:                0xC0,0xC3,0xCC,0xCF,0xF0,0xF3,0xFC,0xFF
cannam@154:          };
cannam@154:          cm = bit_deinterleave_table[cm];
cannam@154:          haar1(X, N0>>k, 1<<k);
cannam@154:       }
cannam@154:       B<<=recombine;
cannam@154: 
cannam@154:       /* Scale output for later folding */
cannam@154:       if (lowband_out)
cannam@154:       {
cannam@154:          int j;
cannam@154:          opus_val16 n;
cannam@154:          n = celt_sqrt(SHL32(EXTEND32(N0),22));
cannam@154:          for (j=0;j<N0;j++)
cannam@154:             lowband_out[j] = MULT16_16_Q15(n,X[j]);
cannam@154:       }
cannam@154:       cm &= (1<<B)-1;
cannam@154:    }
cannam@154:    return cm;
cannam@154: }
cannam@154: 
cannam@154: 
cannam@154: /* This function is responsible for encoding and decoding a band for the stereo case. */
cannam@154: static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm *Y,
cannam@154:       int N, int b, int B, celt_norm *lowband,
cannam@154:       int LM, celt_norm *lowband_out,
cannam@154:       celt_norm *lowband_scratch, int fill)
cannam@154: {
cannam@154:    int imid=0, iside=0;
cannam@154:    int inv = 0;
cannam@154:    opus_val16 mid=0, side=0;
cannam@154:    unsigned cm=0;
cannam@154:    int mbits, sbits, delta;
cannam@154:    int itheta;
cannam@154:    int qalloc;
cannam@154:    struct split_ctx sctx;
cannam@154:    int orig_fill;
cannam@154:    int encode;
cannam@154:    ec_ctx *ec;
cannam@154: 
cannam@154:    encode = ctx->encode;
cannam@154:    ec = ctx->ec;
cannam@154: 
cannam@154:    /* Special case for one sample */
cannam@154:    if (N==1)
cannam@154:    {
cannam@154:       return quant_band_n1(ctx, X, Y, b, lowband_out);
cannam@154:    }
cannam@154: 
cannam@154:    orig_fill = fill;
cannam@154: 
cannam@154:    compute_theta(ctx, &sctx, X, Y, N, &b, B, B, LM, 1, &fill);
cannam@154:    inv = sctx.inv;
cannam@154:    imid = sctx.imid;
cannam@154:    iside = sctx.iside;
cannam@154:    delta = sctx.delta;
cannam@154:    itheta = sctx.itheta;
cannam@154:    qalloc = sctx.qalloc;
cannam@154: #ifdef FIXED_POINT
cannam@154:    mid = imid;
cannam@154:    side = iside;
cannam@154: #else
cannam@154:    mid = (1.f/32768)*imid;
cannam@154:    side = (1.f/32768)*iside;
cannam@154: #endif
cannam@154: 
cannam@154:    /* This is a special case for N=2 that only works for stereo and takes
cannam@154:       advantage of the fact that mid and side are orthogonal to encode
cannam@154:       the side with just one bit. */
cannam@154:    if (N==2)
cannam@154:    {
cannam@154:       int c;
cannam@154:       int sign=0;
cannam@154:       celt_norm *x2, *y2;
cannam@154:       mbits = b;
cannam@154:       sbits = 0;
cannam@154:       /* Only need one bit for the side. */
cannam@154:       if (itheta != 0 && itheta != 16384)
cannam@154:          sbits = 1<<BITRES;
cannam@154:       mbits -= sbits;
cannam@154:       c = itheta > 8192;
cannam@154:       ctx->remaining_bits -= qalloc+sbits;
cannam@154: 
cannam@154:       x2 = c ? Y : X;
cannam@154:       y2 = c ? X : Y;
cannam@154:       if (sbits)
cannam@154:       {
cannam@154:          if (encode)
cannam@154:          {
cannam@154:             /* Here we only need to encode a sign for the side. */
cannam@154:             sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
cannam@154:             ec_enc_bits(ec, sign, 1);
cannam@154:          } else {
cannam@154:             sign = ec_dec_bits(ec, 1);
cannam@154:          }
cannam@154:       }
cannam@154:       sign = 1-2*sign;
cannam@154:       /* We use orig_fill here because we want to fold the side, but if
cannam@154:          itheta==16384, we'll have cleared the low bits of fill. */
cannam@154:       cm = quant_band(ctx, x2, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
cannam@154:             lowband_scratch, orig_fill);
cannam@154:       /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
cannam@154:          and there's no need to worry about mixing with the other channel. */
cannam@154:       y2[0] = -sign*x2[1];
cannam@154:       y2[1] = sign*x2[0];
cannam@154:       if (ctx->resynth)
cannam@154:       {
cannam@154:          celt_norm tmp;
cannam@154:          X[0] = MULT16_16_Q15(mid, X[0]);
cannam@154:          X[1] = MULT16_16_Q15(mid, X[1]);
cannam@154:          Y[0] = MULT16_16_Q15(side, Y[0]);
cannam@154:          Y[1] = MULT16_16_Q15(side, Y[1]);
cannam@154:          tmp = X[0];
cannam@154:          X[0] = SUB16(tmp,Y[0]);
cannam@154:          Y[0] = ADD16(tmp,Y[0]);
cannam@154:          tmp = X[1];
cannam@154:          X[1] = SUB16(tmp,Y[1]);
cannam@154:          Y[1] = ADD16(tmp,Y[1]);
cannam@154:       }
cannam@154:    } else {
cannam@154:       /* "Normal" split code */
cannam@154:       opus_int32 rebalance;
cannam@154: 
cannam@154:       mbits = IMAX(0, IMIN(b, (b-delta)/2));
cannam@154:       sbits = b-mbits;
cannam@154:       ctx->remaining_bits -= qalloc;
cannam@154: 
cannam@154:       rebalance = ctx->remaining_bits;
cannam@154:       if (mbits >= sbits)
cannam@154:       {
cannam@154:          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
cannam@154:             mid for folding later. */
cannam@154:          cm = quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
cannam@154:                lowband_scratch, fill);
cannam@154:          rebalance = mbits - (rebalance-ctx->remaining_bits);
cannam@154:          if (rebalance > 3<<BITRES && itheta!=0)
cannam@154:             sbits += rebalance - (3<<BITRES);
cannam@154: 
cannam@154:          /* For a stereo split, the high bits of fill are always zero, so no
cannam@154:             folding will be done to the side. */
cannam@154:          cm |= quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B);
cannam@154:       } else {
cannam@154:          /* For a stereo split, the high bits of fill are always zero, so no
cannam@154:             folding will be done to the side. */
cannam@154:          cm = quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B);
cannam@154:          rebalance = sbits - (rebalance-ctx->remaining_bits);
cannam@154:          if (rebalance > 3<<BITRES && itheta!=16384)
cannam@154:             mbits += rebalance - (3<<BITRES);
cannam@154:          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
cannam@154:             mid for folding later. */
cannam@154:          cm |= quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
cannam@154:                lowband_scratch, fill);
cannam@154:       }
cannam@154:    }
cannam@154: 
cannam@154: 
cannam@154:    /* This code is used by the decoder and by the resynthesis-enabled encoder */
cannam@154:    if (ctx->resynth)
cannam@154:    {
cannam@154:       if (N!=2)
cannam@154:          stereo_merge(X, Y, mid, N, ctx->arch);
cannam@154:       if (inv)
cannam@154:       {
cannam@154:          int j;
cannam@154:          for (j=0;j<N;j++)
cannam@154:             Y[j] = -Y[j];
cannam@154:       }
cannam@154:    }
cannam@154:    return cm;
cannam@154: }
cannam@154: 
cannam@154: static void special_hybrid_folding(const CELTMode *m, celt_norm *norm, celt_norm *norm2, int start, int M, int dual_stereo)
cannam@154: {
cannam@154:    int n1, n2;
cannam@154:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
cannam@154:    n1 = M*(eBands[start+1]-eBands[start]);
cannam@154:    n2 = M*(eBands[start+2]-eBands[start+1]);
cannam@154:    /* Duplicate enough of the first band folding data to be able to fold the second band.
cannam@154:       Copies no data for CELT-only mode. */
cannam@154:    OPUS_COPY(&norm[n1], &norm[2*n1 - n2], n2-n1);
cannam@154:    if (dual_stereo)
cannam@154:       OPUS_COPY(&norm2[n1], &norm2[2*n1 - n2], n2-n1);
cannam@154: }
cannam@154: 
cannam@154: void quant_all_bands(int encode, const CELTMode *m, int start, int end,
cannam@154:       celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks,
cannam@154:       const celt_ener *bandE, int *pulses, int shortBlocks, int spread,
cannam@154:       int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits,
cannam@154:       opus_int32 balance, ec_ctx *ec, int LM, int codedBands,
cannam@154:       opus_uint32 *seed, int complexity, int arch, int disable_inv)
cannam@154: {
cannam@154:    int i;
cannam@154:    opus_int32 remaining_bits;
cannam@154:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
cannam@154:    celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2;
cannam@154:    VARDECL(celt_norm, _norm);
cannam@154:    VARDECL(celt_norm, _lowband_scratch);
cannam@154:    VARDECL(celt_norm, X_save);
cannam@154:    VARDECL(celt_norm, Y_save);
cannam@154:    VARDECL(celt_norm, X_save2);
cannam@154:    VARDECL(celt_norm, Y_save2);
cannam@154:    VARDECL(celt_norm, norm_save2);
cannam@154:    int resynth_alloc;
cannam@154:    celt_norm *lowband_scratch;
cannam@154:    int B;
cannam@154:    int M;
cannam@154:    int lowband_offset;
cannam@154:    int update_lowband = 1;
cannam@154:    int C = Y_ != NULL ? 2 : 1;
cannam@154:    int norm_offset;
cannam@154:    int theta_rdo = encode && Y_!=NULL && !dual_stereo && complexity>=8;
cannam@154: #ifdef RESYNTH
cannam@154:    int resynth = 1;
cannam@154: #else
cannam@154:    int resynth = !encode || theta_rdo;
cannam@154: #endif
cannam@154:    struct band_ctx ctx;
cannam@154:    SAVE_STACK;
cannam@154: 
cannam@154:    M = 1<<LM;
cannam@154:    B = shortBlocks ? M : 1;
cannam@154:    norm_offset = M*eBands[start];
cannam@154:    /* No need to allocate norm for the last band because we don't need an
cannam@154:       output in that band. */
cannam@154:    ALLOC(_norm, C*(M*eBands[m->nbEBands-1]-norm_offset), celt_norm);
cannam@154:    norm = _norm;
cannam@154:    norm2 = norm + M*eBands[m->nbEBands-1]-norm_offset;
cannam@154: 
cannam@154:    /* For decoding, we can use the last band as scratch space because we don't need that
cannam@154:       scratch space for the last band and we don't care about the data there until we're
cannam@154:       decoding the last band. */
cannam@154:    if (encode && resynth)
cannam@154:       resynth_alloc = M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]);
cannam@154:    else
cannam@154:       resynth_alloc = ALLOC_NONE;
cannam@154:    ALLOC(_lowband_scratch, resynth_alloc, celt_norm);
cannam@154:    if (encode && resynth)
cannam@154:       lowband_scratch = _lowband_scratch;
cannam@154:    else
cannam@154:       lowband_scratch = X_+M*eBands[m->nbEBands-1];
cannam@154:    ALLOC(X_save, resynth_alloc, celt_norm);
cannam@154:    ALLOC(Y_save, resynth_alloc, celt_norm);
cannam@154:    ALLOC(X_save2, resynth_alloc, celt_norm);
cannam@154:    ALLOC(Y_save2, resynth_alloc, celt_norm);
cannam@154:    ALLOC(norm_save2, resynth_alloc, celt_norm);
cannam@154: 
cannam@154:    lowband_offset = 0;
cannam@154:    ctx.bandE = bandE;
cannam@154:    ctx.ec = ec;
cannam@154:    ctx.encode = encode;
cannam@154:    ctx.intensity = intensity;
cannam@154:    ctx.m = m;
cannam@154:    ctx.seed = *seed;
cannam@154:    ctx.spread = spread;
cannam@154:    ctx.arch = arch;
cannam@154:    ctx.disable_inv = disable_inv;
cannam@154:    ctx.resynth = resynth;
cannam@154:    ctx.theta_round = 0;
cannam@154:    /* Avoid injecting noise in the first band on transients. */
cannam@154:    ctx.avoid_split_noise = B > 1;
cannam@154:    for (i=start;i<end;i++)
cannam@154:    {
cannam@154:       opus_int32 tell;
cannam@154:       int b;
cannam@154:       int N;
cannam@154:       opus_int32 curr_balance;
cannam@154:       int effective_lowband=-1;
cannam@154:       celt_norm * OPUS_RESTRICT X, * OPUS_RESTRICT Y;
cannam@154:       int tf_change=0;
cannam@154:       unsigned x_cm;
cannam@154:       unsigned y_cm;
cannam@154:       int last;
cannam@154: 
cannam@154:       ctx.i = i;
cannam@154:       last = (i==end-1);
cannam@154: 
cannam@154:       X = X_+M*eBands[i];
cannam@154:       if (Y_!=NULL)
cannam@154:          Y = Y_+M*eBands[i];
cannam@154:       else
cannam@154:          Y = NULL;
cannam@154:       N = M*eBands[i+1]-M*eBands[i];
cannam@154:       celt_assert(N > 0);
cannam@154:       tell = ec_tell_frac(ec);
cannam@154: 
cannam@154:       /* Compute how many bits we want to allocate to this band */
cannam@154:       if (i != start)
cannam@154:          balance -= tell;
cannam@154:       remaining_bits = total_bits-tell-1;
cannam@154:       ctx.remaining_bits = remaining_bits;
cannam@154:       if (i <= codedBands-1)
cannam@154:       {
cannam@154:          curr_balance = celt_sudiv(balance, IMIN(3, codedBands-i));
cannam@154:          b = IMAX(0, IMIN(16383, IMIN(remaining_bits+1,pulses[i]+curr_balance)));
cannam@154:       } else {
cannam@154:          b = 0;
cannam@154:       }
cannam@154: 
cannam@154: #ifndef DISABLE_UPDATE_DRAFT
cannam@154:       if (resynth && (M*eBands[i]-N >= M*eBands[start] || i==start+1) && (update_lowband || lowband_offset==0))
cannam@154:             lowband_offset = i;
cannam@154:       if (i == start+1)
cannam@154:          special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
cannam@154: #else
cannam@154:       if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0))
cannam@154:             lowband_offset = i;
cannam@154: #endif
cannam@154: 
cannam@154:       tf_change = tf_res[i];
cannam@154:       ctx.tf_change = tf_change;
cannam@154:       if (i>=m->effEBands)
cannam@154:       {
cannam@154:          X=norm;
cannam@154:          if (Y_!=NULL)
cannam@154:             Y = norm;
cannam@154:          lowband_scratch = NULL;
cannam@154:       }
cannam@154:       if (last && !theta_rdo)
cannam@154:          lowband_scratch = NULL;
cannam@154: 
cannam@154:       /* Get a conservative estimate of the collapse_mask's for the bands we're
cannam@154:          going to be folding from. */
cannam@154:       if (lowband_offset != 0 && (spread!=SPREAD_AGGRESSIVE || B>1 || tf_change<0))
cannam@154:       {
cannam@154:          int fold_start;
cannam@154:          int fold_end;
cannam@154:          int fold_i;
cannam@154:          /* This ensures we never repeat spectral content within one band */
cannam@154:          effective_lowband = IMAX(0, M*eBands[lowband_offset]-norm_offset-N);
cannam@154:          fold_start = lowband_offset;
cannam@154:          while(M*eBands[--fold_start] > effective_lowband+norm_offset);
cannam@154:          fold_end = lowband_offset-1;
cannam@154: #ifndef DISABLE_UPDATE_DRAFT
cannam@154:          while(++fold_end < i && M*eBands[fold_end] < effective_lowband+norm_offset+N);
cannam@154: #else
cannam@154:          while(M*eBands[++fold_end] < effective_lowband+norm_offset+N);
cannam@154: #endif
cannam@154:          x_cm = y_cm = 0;
cannam@154:          fold_i = fold_start; do {
cannam@154:            x_cm |= collapse_masks[fold_i*C+0];
cannam@154:            y_cm |= collapse_masks[fold_i*C+C-1];
cannam@154:          } while (++fold_i<fold_end);
cannam@154:       }
cannam@154:       /* Otherwise, we'll be using the LCG to fold, so all blocks will (almost
cannam@154:          always) be non-zero. */
cannam@154:       else
cannam@154:          x_cm = y_cm = (1<<B)-1;
cannam@154: 
cannam@154:       if (dual_stereo && i==intensity)
cannam@154:       {
cannam@154:          int j;
cannam@154: 
cannam@154:          /* Switch off dual stereo to do intensity. */
cannam@154:          dual_stereo = 0;
cannam@154:          if (resynth)
cannam@154:             for (j=0;j<M*eBands[i]-norm_offset;j++)
cannam@154:                norm[j] = HALF32(norm[j]+norm2[j]);
cannam@154:       }
cannam@154:       if (dual_stereo)
cannam@154:       {
cannam@154:          x_cm = quant_band(&ctx, X, N, b/2, B,
cannam@154:                effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
cannam@154:                last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm);
cannam@154:          y_cm = quant_band(&ctx, Y, N, b/2, B,
cannam@154:                effective_lowband != -1 ? norm2+effective_lowband : NULL, LM,
cannam@154:                last?NULL:norm2+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, y_cm);
cannam@154:       } else {
cannam@154:          if (Y!=NULL)
cannam@154:          {
cannam@154:             if (theta_rdo && i < intensity)
cannam@154:             {
cannam@154:                ec_ctx ec_save, ec_save2;
cannam@154:                struct band_ctx ctx_save, ctx_save2;
cannam@154:                opus_val32 dist0, dist1;
cannam@154:                unsigned cm, cm2;
cannam@154:                int nstart_bytes, nend_bytes, save_bytes;
cannam@154:                unsigned char *bytes_buf;
cannam@154:                unsigned char bytes_save[1275];
cannam@154:                opus_val16 w[2];
cannam@154:                compute_channel_weights(bandE[i], bandE[i+m->nbEBands], w);
cannam@154:                /* Make a copy. */
cannam@154:                cm = x_cm|y_cm;
cannam@154:                ec_save = *ec;
cannam@154:                ctx_save = ctx;
cannam@154:                OPUS_COPY(X_save, X, N);
cannam@154:                OPUS_COPY(Y_save, Y, N);
cannam@154:                /* Encode and round down. */
cannam@154:                ctx.theta_round = -1;
cannam@154:                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
cannam@154:                      effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
cannam@154:                      last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm);
cannam@154:                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));
cannam@154: 
cannam@154:                /* Save first result. */
cannam@154:                cm2 = x_cm;
cannam@154:                ec_save2 = *ec;
cannam@154:                ctx_save2 = ctx;
cannam@154:                OPUS_COPY(X_save2, X, N);
cannam@154:                OPUS_COPY(Y_save2, Y, N);
cannam@154:                if (!last)
cannam@154:                   OPUS_COPY(norm_save2, norm+M*eBands[i]-norm_offset, N);
cannam@154:                nstart_bytes = ec_save.offs;
cannam@154:                nend_bytes = ec_save.storage;
cannam@154:                bytes_buf = ec_save.buf+nstart_bytes;
cannam@154:                save_bytes = nend_bytes-nstart_bytes;
cannam@154:                OPUS_COPY(bytes_save, bytes_buf, save_bytes);
cannam@154: 
cannam@154:                /* Restore */
cannam@154:                *ec = ec_save;
cannam@154:                ctx = ctx_save;
cannam@154:                OPUS_COPY(X, X_save, N);
cannam@154:                OPUS_COPY(Y, Y_save, N);
cannam@154: #ifndef DISABLE_UPDATE_DRAFT
cannam@154:                if (i == start+1)
cannam@154:                   special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
cannam@154: #endif
cannam@154:                /* Encode and round up. */
cannam@154:                ctx.theta_round = 1;
cannam@154:                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
cannam@154:                      effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
cannam@154:                      last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm);
cannam@154:                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));
cannam@154:                if (dist0 >= dist1) {
cannam@154:                   x_cm = cm2;
cannam@154:                   *ec = ec_save2;
cannam@154:                   ctx = ctx_save2;
cannam@154:                   OPUS_COPY(X, X_save2, N);
cannam@154:                   OPUS_COPY(Y, Y_save2, N);
cannam@154:                   if (!last)
cannam@154:                      OPUS_COPY(norm+M*eBands[i]-norm_offset, norm_save2, N);
cannam@154:                   OPUS_COPY(bytes_buf, bytes_save, save_bytes);
cannam@154:                }
cannam@154:             } else {
cannam@154:                ctx.theta_round = 0;
cannam@154:                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
cannam@154:                      effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
cannam@154:                      last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm);
cannam@154:             }
cannam@154:          } else {
cannam@154:             x_cm = quant_band(&ctx, X, N, b, B,
cannam@154:                   effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
cannam@154:                   last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm);
cannam@154:          }
cannam@154:          y_cm = x_cm;
cannam@154:       }
cannam@154:       collapse_masks[i*C+0] = (unsigned char)x_cm;
cannam@154:       collapse_masks[i*C+C-1] = (unsigned char)y_cm;
cannam@154:       balance += pulses[i] + tell;
cannam@154: 
cannam@154:       /* Update the folding position only as long as we have 1 bit/sample depth. */
cannam@154:       update_lowband = b>(N<<BITRES);
cannam@154:       /* We only need to avoid noise on a split for the first band. After that, we
cannam@154:          have folding. */
cannam@154:       ctx.avoid_split_noise = 0;
cannam@154:    }
cannam@154:    *seed = ctx.seed;
cannam@154: 
cannam@154:    RESTORE_STACK;
cannam@154: }
cannam@154: