michael@0: /* Copyright (c) 2007-2008 CSIRO
michael@0:    Copyright (c) 2007-2009 Xiph.Org Foundation
michael@0:    Copyright (c) 2008-2009 Gregory Maxwell
michael@0:    Written by Jean-Marc Valin and Gregory Maxwell */
michael@0: /*
michael@0:    Redistribution and use in source and binary forms, with or without
michael@0:    modification, are permitted provided that the following conditions
michael@0:    are met:
michael@0: 
michael@0:    - Redistributions of source code must retain the above copyright
michael@0:    notice, this list of conditions and the following disclaimer.
michael@0: 
michael@0:    - Redistributions in binary form must reproduce the above copyright
michael@0:    notice, this list of conditions and the following disclaimer in the
michael@0:    documentation and/or other materials provided with the distribution.
michael@0: 
michael@0:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
michael@0:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
michael@0:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
michael@0:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
michael@0:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
michael@0:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
michael@0:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
michael@0:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
michael@0:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
michael@0:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
michael@0:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
michael@0: */
michael@0: 
michael@0: #ifdef HAVE_CONFIG_H
michael@0: #include "config.h"
michael@0: #endif
michael@0: 
michael@0: #include <math.h>
michael@0: #include "bands.h"
michael@0: #include "modes.h"
michael@0: #include "vq.h"
michael@0: #include "cwrs.h"
michael@0: #include "stack_alloc.h"
michael@0: #include "os_support.h"
michael@0: #include "mathops.h"
michael@0: #include "rate.h"
michael@0: #include "quant_bands.h"
michael@0: #include "pitch.h"
michael@0: 
michael@0: int hysteresis_decision(opus_val16 val, const opus_val16 *thresholds, const opus_val16 *hysteresis, int N, int prev)
michael@0: {
michael@0:    int i;
michael@0:    for (i=0;i<N;i++)
michael@0:    {
michael@0:       if (val < thresholds[i])
michael@0:          break;
michael@0:    }
michael@0:    if (i>prev && val < thresholds[prev]+hysteresis[prev])
michael@0:       i=prev;
michael@0:    if (i<prev && val > thresholds[prev-1]-hysteresis[prev-1])
michael@0:       i=prev;
michael@0:    return i;
michael@0: }
michael@0: 
michael@0: opus_uint32 celt_lcg_rand(opus_uint32 seed)
michael@0: {
michael@0:    return 1664525 * seed + 1013904223;
michael@0: }
michael@0: 
michael@0: /* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness
michael@0:    with this approximation is important because it has an impact on the bit allocation */
michael@0: static opus_int16 bitexact_cos(opus_int16 x)
michael@0: {
michael@0:    opus_int32 tmp;
michael@0:    opus_int16 x2;
michael@0:    tmp = (4096+((opus_int32)(x)*(x)))>>13;
michael@0:    celt_assert(tmp<=32767);
michael@0:    x2 = tmp;
michael@0:    x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2)))));
michael@0:    celt_assert(x2<=32766);
michael@0:    return 1+x2;
michael@0: }
michael@0: 
michael@0: static int bitexact_log2tan(int isin,int icos)
michael@0: {
michael@0:    int lc;
michael@0:    int ls;
michael@0:    lc=EC_ILOG(icos);
michael@0:    ls=EC_ILOG(isin);
michael@0:    icos<<=15-lc;
michael@0:    isin<<=15-ls;
michael@0:    return (ls-lc)*(1<<11)
michael@0:          +FRAC_MUL16(isin, FRAC_MUL16(isin, -2597) + 7932)
michael@0:          -FRAC_MUL16(icos, FRAC_MUL16(icos, -2597) + 7932);
michael@0: }
michael@0: 
michael@0: #ifdef FIXED_POINT
michael@0: /* Compute the amplitude (sqrt energy) in each of the bands */
michael@0: void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int M)
michael@0: {
michael@0:    int i, c, N;
michael@0:    const opus_int16 *eBands = m->eBands;
michael@0:    N = M*m->shortMdctSize;
michael@0:    c=0; do {
michael@0:       for (i=0;i<end;i++)
michael@0:       {
michael@0:          int j;
michael@0:          opus_val32 maxval=0;
michael@0:          opus_val32 sum = 0;
michael@0: 
michael@0:          j=M*eBands[i]; do {
michael@0:             maxval = MAX32(maxval, X[j+c*N]);
michael@0:             maxval = MAX32(maxval, -X[j+c*N]);
michael@0:          } while (++j<M*eBands[i+1]);
michael@0: 
michael@0:          if (maxval > 0)
michael@0:          {
michael@0:             int shift = celt_ilog2(maxval)-10;
michael@0:             j=M*eBands[i]; do {
michael@0:                sum = MAC16_16(sum, EXTRACT16(VSHR32(X[j+c*N],shift)),
michael@0:                                    EXTRACT16(VSHR32(X[j+c*N],shift)));
michael@0:             } while (++j<M*eBands[i+1]);
michael@0:             /* We're adding one here to ensure the normalized band isn't larger than unity norm */
michael@0:             bandE[i+c*m->nbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
michael@0:          } else {
michael@0:             bandE[i+c*m->nbEBands] = EPSILON;
michael@0:          }
michael@0:          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
michael@0:       }
michael@0:    } while (++c<C);
michael@0:    /*printf ("\n");*/
michael@0: }
michael@0: 
michael@0: /* Normalise each band such that the energy is one. */
michael@0: 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)
michael@0: {
michael@0:    int i, c, N;
michael@0:    const opus_int16 *eBands = m->eBands;
michael@0:    N = M*m->shortMdctSize;
michael@0:    c=0; do {
michael@0:       i=0; do {
michael@0:          opus_val16 g;
michael@0:          int j,shift;
michael@0:          opus_val16 E;
michael@0:          shift = celt_zlog2(bandE[i+c*m->nbEBands])-13;
michael@0:          E = VSHR32(bandE[i+c*m->nbEBands], shift);
michael@0:          g = EXTRACT16(celt_rcp(SHL32(E,3)));
michael@0:          j=M*eBands[i]; do {
michael@0:             X[j+c*N] = MULT16_16_Q15(VSHR32(freq[j+c*N],shift-1),g);
michael@0:          } while (++j<M*eBands[i+1]);
michael@0:       } while (++i<end);
michael@0:    } while (++c<C);
michael@0: }
michael@0: 
michael@0: #else /* FIXED_POINT */
michael@0: /* Compute the amplitude (sqrt energy) in each of the bands */
michael@0: void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int M)
michael@0: {
michael@0:    int i, c, N;
michael@0:    const opus_int16 *eBands = m->eBands;
michael@0:    N = M*m->shortMdctSize;
michael@0:    c=0; do {
michael@0:       for (i=0;i<end;i++)
michael@0:       {
michael@0:          int j;
michael@0:          opus_val32 sum = 1e-27f;
michael@0:          for (j=M*eBands[i];j<M*eBands[i+1];j++)
michael@0:             sum += X[j+c*N]*X[j+c*N];
michael@0:          bandE[i+c*m->nbEBands] = celt_sqrt(sum);
michael@0:          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
michael@0:       }
michael@0:    } while (++c<C);
michael@0:    /*printf ("\n");*/
michael@0: }
michael@0: 
michael@0: /* Normalise each band such that the energy is one. */
michael@0: 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)
michael@0: {
michael@0:    int i, c, N;
michael@0:    const opus_int16 *eBands = m->eBands;
michael@0:    N = M*m->shortMdctSize;
michael@0:    c=0; do {
michael@0:       for (i=0;i<end;i++)
michael@0:       {
michael@0:          int j;
michael@0:          opus_val16 g = 1.f/(1e-27f+bandE[i+c*m->nbEBands]);
michael@0:          for (j=M*eBands[i];j<M*eBands[i+1];j++)
michael@0:             X[j+c*N] = freq[j+c*N]*g;
michael@0:       }
michael@0:    } while (++c<C);
michael@0: }
michael@0: 
michael@0: #endif /* FIXED_POINT */
michael@0: 
michael@0: /* De-normalise the energy to produce the synthesis from the unit-energy bands */
michael@0: void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X,
michael@0:       celt_sig * OPUS_RESTRICT freq, const opus_val16 *bandLogE, int start, int end, int C, int M)
michael@0: {
michael@0:    int i, c, N;
michael@0:    const opus_int16 *eBands = m->eBands;
michael@0:    N = M*m->shortMdctSize;
michael@0:    celt_assert2(C<=2, "denormalise_bands() not implemented for >2 channels");
michael@0:    c=0; do {
michael@0:       celt_sig * OPUS_RESTRICT f;
michael@0:       const celt_norm * OPUS_RESTRICT x;
michael@0:       f = freq+c*N;
michael@0:       x = X+c*N+M*eBands[start];
michael@0:       for (i=0;i<M*eBands[start];i++)
michael@0:          *f++ = 0;
michael@0:       for (i=start;i<end;i++)
michael@0:       {
michael@0:          int j, band_end;
michael@0:          opus_val16 g;
michael@0:          opus_val16 lg;
michael@0: #ifdef FIXED_POINT
michael@0:          int shift;
michael@0: #endif
michael@0:          j=M*eBands[i];
michael@0:          band_end = M*eBands[i+1];
michael@0:          lg = ADD16(bandLogE[i+c*m->nbEBands], SHL16((opus_val16)eMeans[i],6));
michael@0: #ifndef FIXED_POINT
michael@0:          g = celt_exp2(lg);
michael@0: #else
michael@0:          /* Handle the integer part of the log energy */
michael@0:          shift = 16-(lg>>DB_SHIFT);
michael@0:          if (shift>31)
michael@0:          {
michael@0:             shift=0;
michael@0:             g=0;
michael@0:          } else {
michael@0:             /* Handle the fractional part. */
michael@0:             g = celt_exp2_frac(lg&((1<<DB_SHIFT)-1));
michael@0:          }
michael@0:          /* Handle extreme gains with negative shift. */
michael@0:          if (shift<0)
michael@0:          {
michael@0:             /* For shift < -2 we'd be likely to overflow, so we're capping
michael@0:                the gain here. This shouldn't happen unless the bitstream is
michael@0:                already corrupted. */
michael@0:             if (shift < -2)
michael@0:             {
michael@0:                g = 32767;
michael@0:                shift = -2;
michael@0:             }
michael@0:             do {
michael@0:                *f++ = SHL32(MULT16_16(*x++, g), -shift);
michael@0:             } while (++j<band_end);
michael@0:          } else
michael@0: #endif
michael@0:          /* Be careful of the fixed-point "else" just above when changing this code */
michael@0:          do {
michael@0:             *f++ = SHR32(MULT16_16(*x++, g), shift);
michael@0:          } while (++j<band_end);
michael@0:       }
michael@0:       celt_assert(start <= end);
michael@0:       for (i=M*eBands[end];i<N;i++)
michael@0:          *f++ = 0;
michael@0:    } while (++c<C);
michael@0: }
michael@0: 
michael@0: /* This prevents energy collapse for transients with multiple short MDCTs */
michael@0: void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_masks, int LM, int C, int size,
michael@0:       int start, int end, opus_val16 *logE, opus_val16 *prev1logE,
michael@0:       opus_val16 *prev2logE, int *pulses, opus_uint32 seed)
michael@0: {
michael@0:    int c, i, j, k;
michael@0:    for (i=start;i<end;i++)
michael@0:    {
michael@0:       int N0;
michael@0:       opus_val16 thresh, sqrt_1;
michael@0:       int depth;
michael@0: #ifdef FIXED_POINT
michael@0:       int shift;
michael@0:       opus_val32 thresh32;
michael@0: #endif
michael@0: 
michael@0:       N0 = m->eBands[i+1]-m->eBands[i];
michael@0:       /* depth in 1/8 bits */
michael@0:       depth = (1+pulses[i])/((m->eBands[i+1]-m->eBands[i])<<LM);
michael@0: 
michael@0: #ifdef FIXED_POINT
michael@0:       thresh32 = SHR32(celt_exp2(-SHL16(depth, 10-BITRES)),1);
michael@0:       thresh = MULT16_32_Q15(QCONST16(0.5f, 15), MIN32(32767,thresh32));
michael@0:       {
michael@0:          opus_val32 t;
michael@0:          t = N0<<LM;
michael@0:          shift = celt_ilog2(t)>>1;
michael@0:          t = SHL32(t, (7-shift)<<1);
michael@0:          sqrt_1 = celt_rsqrt_norm(t);
michael@0:       }
michael@0: #else
michael@0:       thresh = .5f*celt_exp2(-.125f*depth);
michael@0:       sqrt_1 = celt_rsqrt(N0<<LM);
michael@0: #endif
michael@0: 
michael@0:       c=0; do
michael@0:       {
michael@0:          celt_norm *X;
michael@0:          opus_val16 prev1;
michael@0:          opus_val16 prev2;
michael@0:          opus_val32 Ediff;
michael@0:          opus_val16 r;
michael@0:          int renormalize=0;
michael@0:          prev1 = prev1logE[c*m->nbEBands+i];
michael@0:          prev2 = prev2logE[c*m->nbEBands+i];
michael@0:          if (C==1)
michael@0:          {
michael@0:             prev1 = MAX16(prev1,prev1logE[m->nbEBands+i]);
michael@0:             prev2 = MAX16(prev2,prev2logE[m->nbEBands+i]);
michael@0:          }
michael@0:          Ediff = EXTEND32(logE[c*m->nbEBands+i])-EXTEND32(MIN16(prev1,prev2));
michael@0:          Ediff = MAX32(0, Ediff);
michael@0: 
michael@0: #ifdef FIXED_POINT
michael@0:          if (Ediff < 16384)
michael@0:          {
michael@0:             opus_val32 r32 = SHR32(celt_exp2(-EXTRACT16(Ediff)),1);
michael@0:             r = 2*MIN16(16383,r32);
michael@0:          } else {
michael@0:             r = 0;
michael@0:          }
michael@0:          if (LM==3)
michael@0:             r = MULT16_16_Q14(23170, MIN32(23169, r));
michael@0:          r = SHR16(MIN16(thresh, r),1);
michael@0:          r = SHR32(MULT16_16_Q15(sqrt_1, r),shift);
michael@0: #else
michael@0:          /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
michael@0:             short blocks don't have the same energy as long */
michael@0:          r = 2.f*celt_exp2(-Ediff);
michael@0:          if (LM==3)
michael@0:             r *= 1.41421356f;
michael@0:          r = MIN16(thresh, r);
michael@0:          r = r*sqrt_1;
michael@0: #endif
michael@0:          X = X_+c*size+(m->eBands[i]<<LM);
michael@0:          for (k=0;k<1<<LM;k++)
michael@0:          {
michael@0:             /* Detect collapse */
michael@0:             if (!(collapse_masks[i*C+c]&1<<k))
michael@0:             {
michael@0:                /* Fill with noise */
michael@0:                for (j=0;j<N0;j++)
michael@0:                {
michael@0:                   seed = celt_lcg_rand(seed);
michael@0:                   X[(j<<LM)+k] = (seed&0x8000 ? r : -r);
michael@0:                }
michael@0:                renormalize = 1;
michael@0:             }
michael@0:          }
michael@0:          /* We just added some energy, so we need to renormalise */
michael@0:          if (renormalize)
michael@0:             renormalise_vector(X, N0<<LM, Q15ONE);
michael@0:       } while (++c<C);
michael@0:    }
michael@0: }
michael@0: 
michael@0: static void intensity_stereo(const CELTMode *m, celt_norm *X, celt_norm *Y, const celt_ener *bandE, int bandID, int N)
michael@0: {
michael@0:    int i = bandID;
michael@0:    int j;
michael@0:    opus_val16 a1, a2;
michael@0:    opus_val16 left, right;
michael@0:    opus_val16 norm;
michael@0: #ifdef FIXED_POINT
michael@0:    int shift = celt_zlog2(MAX32(bandE[i], bandE[i+m->nbEBands]))-13;
michael@0: #endif
michael@0:    left = VSHR32(bandE[i],shift);
michael@0:    right = VSHR32(bandE[i+m->nbEBands],shift);
michael@0:    norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right));
michael@0:    a1 = DIV32_16(SHL32(EXTEND32(left),14),norm);
michael@0:    a2 = DIV32_16(SHL32(EXTEND32(right),14),norm);
michael@0:    for (j=0;j<N;j++)
michael@0:    {
michael@0:       celt_norm r, l;
michael@0:       l = X[j];
michael@0:       r = Y[j];
michael@0:       X[j] = MULT16_16_Q14(a1,l) + MULT16_16_Q14(a2,r);
michael@0:       /* Side is not encoded, no need to calculate */
michael@0:    }
michael@0: }
michael@0: 
michael@0: static void stereo_split(celt_norm *X, celt_norm *Y, int N)
michael@0: {
michael@0:    int j;
michael@0:    for (j=0;j<N;j++)
michael@0:    {
michael@0:       celt_norm r, l;
michael@0:       l = MULT16_16_Q15(QCONST16(.70710678f,15), X[j]);
michael@0:       r = MULT16_16_Q15(QCONST16(.70710678f,15), Y[j]);
michael@0:       X[j] = l+r;
michael@0:       Y[j] = r-l;
michael@0:    }
michael@0: }
michael@0: 
michael@0: static void stereo_merge(celt_norm *X, celt_norm *Y, opus_val16 mid, int N)
michael@0: {
michael@0:    int j;
michael@0:    opus_val32 xp=0, side=0;
michael@0:    opus_val32 El, Er;
michael@0:    opus_val16 mid2;
michael@0: #ifdef FIXED_POINT
michael@0:    int kl, kr;
michael@0: #endif
michael@0:    opus_val32 t, lgain, rgain;
michael@0: 
michael@0:    /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */
michael@0:    dual_inner_prod(Y, X, Y, N, &xp, &side);
michael@0:    /* Compensating for the mid normalization */
michael@0:    xp = MULT16_32_Q15(mid, xp);
michael@0:    /* mid and side are in Q15, not Q14 like X and Y */
michael@0:    mid2 = SHR32(mid, 1);
michael@0:    El = MULT16_16(mid2, mid2) + side - 2*xp;
michael@0:    Er = MULT16_16(mid2, mid2) + side + 2*xp;
michael@0:    if (Er < QCONST32(6e-4f, 28) || El < QCONST32(6e-4f, 28))
michael@0:    {
michael@0:       for (j=0;j<N;j++)
michael@0:          Y[j] = X[j];
michael@0:       return;
michael@0:    }
michael@0: 
michael@0: #ifdef FIXED_POINT
michael@0:    kl = celt_ilog2(El)>>1;
michael@0:    kr = celt_ilog2(Er)>>1;
michael@0: #endif
michael@0:    t = VSHR32(El, (kl-7)<<1);
michael@0:    lgain = celt_rsqrt_norm(t);
michael@0:    t = VSHR32(Er, (kr-7)<<1);
michael@0:    rgain = celt_rsqrt_norm(t);
michael@0: 
michael@0: #ifdef FIXED_POINT
michael@0:    if (kl < 7)
michael@0:       kl = 7;
michael@0:    if (kr < 7)
michael@0:       kr = 7;
michael@0: #endif
michael@0: 
michael@0:    for (j=0;j<N;j++)
michael@0:    {
michael@0:       celt_norm r, l;
michael@0:       /* Apply mid scaling (side is already scaled) */
michael@0:       l = MULT16_16_Q15(mid, X[j]);
michael@0:       r = Y[j];
michael@0:       X[j] = EXTRACT16(PSHR32(MULT16_16(lgain, SUB16(l,r)), kl+1));
michael@0:       Y[j] = EXTRACT16(PSHR32(MULT16_16(rgain, ADD16(l,r)), kr+1));
michael@0:    }
michael@0: }
michael@0: 
michael@0: /* Decide whether we should spread the pulses in the current frame */
michael@0: int spreading_decision(const CELTMode *m, celt_norm *X, int *average,
michael@0:       int last_decision, int *hf_average, int *tapset_decision, int update_hf,
michael@0:       int end, int C, int M)
michael@0: {
michael@0:    int i, c, N0;
michael@0:    int sum = 0, nbBands=0;
michael@0:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
michael@0:    int decision;
michael@0:    int hf_sum=0;
michael@0: 
michael@0:    celt_assert(end>0);
michael@0: 
michael@0:    N0 = M*m->shortMdctSize;
michael@0: 
michael@0:    if (M*(eBands[end]-eBands[end-1]) <= 8)
michael@0:       return SPREAD_NONE;
michael@0:    c=0; do {
michael@0:       for (i=0;i<end;i++)
michael@0:       {
michael@0:          int j, N, tmp=0;
michael@0:          int tcount[3] = {0,0,0};
michael@0:          celt_norm * OPUS_RESTRICT x = X+M*eBands[i]+c*N0;
michael@0:          N = M*(eBands[i+1]-eBands[i]);
michael@0:          if (N<=8)
michael@0:             continue;
michael@0:          /* Compute rough CDF of |x[j]| */
michael@0:          for (j=0;j<N;j++)
michael@0:          {
michael@0:             opus_val32 x2N; /* Q13 */
michael@0: 
michael@0:             x2N = MULT16_16(MULT16_16_Q15(x[j], x[j]), N);
michael@0:             if (x2N < QCONST16(0.25f,13))
michael@0:                tcount[0]++;
michael@0:             if (x2N < QCONST16(0.0625f,13))
michael@0:                tcount[1]++;
michael@0:             if (x2N < QCONST16(0.015625f,13))
michael@0:                tcount[2]++;
michael@0:          }
michael@0: 
michael@0:          /* Only include four last bands (8 kHz and up) */
michael@0:          if (i>m->nbEBands-4)
michael@0:             hf_sum += 32*(tcount[1]+tcount[0])/N;
michael@0:          tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N);
michael@0:          sum += tmp*256;
michael@0:          nbBands++;
michael@0:       }
michael@0:    } while (++c<C);
michael@0: 
michael@0:    if (update_hf)
michael@0:    {
michael@0:       if (hf_sum)
michael@0:          hf_sum /= C*(4-m->nbEBands+end);
michael@0:       *hf_average = (*hf_average+hf_sum)>>1;
michael@0:       hf_sum = *hf_average;
michael@0:       if (*tapset_decision==2)
michael@0:          hf_sum += 4;
michael@0:       else if (*tapset_decision==0)
michael@0:          hf_sum -= 4;
michael@0:       if (hf_sum > 22)
michael@0:          *tapset_decision=2;
michael@0:       else if (hf_sum > 18)
michael@0:          *tapset_decision=1;
michael@0:       else
michael@0:          *tapset_decision=0;
michael@0:    }
michael@0:    /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/
michael@0:    celt_assert(nbBands>0); /* end has to be non-zero */
michael@0:    sum /= nbBands;
michael@0:    /* Recursive averaging */
michael@0:    sum = (sum+*average)>>1;
michael@0:    *average = sum;
michael@0:    /* Hysteresis */
michael@0:    sum = (3*sum + (((3-last_decision)<<7) + 64) + 2)>>2;
michael@0:    if (sum < 80)
michael@0:    {
michael@0:       decision = SPREAD_AGGRESSIVE;
michael@0:    } else if (sum < 256)
michael@0:    {
michael@0:       decision = SPREAD_NORMAL;
michael@0:    } else if (sum < 384)
michael@0:    {
michael@0:       decision = SPREAD_LIGHT;
michael@0:    } else {
michael@0:       decision = SPREAD_NONE;
michael@0:    }
michael@0: #ifdef FUZZING
michael@0:    decision = rand()&0x3;
michael@0:    *tapset_decision=rand()%3;
michael@0: #endif
michael@0:    return decision;
michael@0: }
michael@0: 
michael@0: /* Indexing table for converting from natural Hadamard to ordery Hadamard
michael@0:    This is essentially a bit-reversed Gray, on top of which we've added
michael@0:    an inversion of the order because we want the DC at the end rather than
michael@0:    the beginning. The lines are for N=2, 4, 8, 16 */
michael@0: static const int ordery_table[] = {
michael@0:        1,  0,
michael@0:        3,  0,  2,  1,
michael@0:        7,  0,  4,  3,  6,  1,  5,  2,
michael@0:       15,  0,  8,  7, 12,  3, 11,  4, 14,  1,  9,  6, 13,  2, 10,  5,
michael@0: };
michael@0: 
michael@0: static void deinterleave_hadamard(celt_norm *X, int N0, int stride, int hadamard)
michael@0: {
michael@0:    int i,j;
michael@0:    VARDECL(celt_norm, tmp);
michael@0:    int N;
michael@0:    SAVE_STACK;
michael@0:    N = N0*stride;
michael@0:    ALLOC(tmp, N, celt_norm);
michael@0:    celt_assert(stride>0);
michael@0:    if (hadamard)
michael@0:    {
michael@0:       const int *ordery = ordery_table+stride-2;
michael@0:       for (i=0;i<stride;i++)
michael@0:       {
michael@0:          for (j=0;j<N0;j++)
michael@0:             tmp[ordery[i]*N0+j] = X[j*stride+i];
michael@0:       }
michael@0:    } else {
michael@0:       for (i=0;i<stride;i++)
michael@0:          for (j=0;j<N0;j++)
michael@0:             tmp[i*N0+j] = X[j*stride+i];
michael@0:    }
michael@0:    for (j=0;j<N;j++)
michael@0:       X[j] = tmp[j];
michael@0:    RESTORE_STACK;
michael@0: }
michael@0: 
michael@0: static void interleave_hadamard(celt_norm *X, int N0, int stride, int hadamard)
michael@0: {
michael@0:    int i,j;
michael@0:    VARDECL(celt_norm, tmp);
michael@0:    int N;
michael@0:    SAVE_STACK;
michael@0:    N = N0*stride;
michael@0:    ALLOC(tmp, N, celt_norm);
michael@0:    if (hadamard)
michael@0:    {
michael@0:       const int *ordery = ordery_table+stride-2;
michael@0:       for (i=0;i<stride;i++)
michael@0:          for (j=0;j<N0;j++)
michael@0:             tmp[j*stride+i] = X[ordery[i]*N0+j];
michael@0:    } else {
michael@0:       for (i=0;i<stride;i++)
michael@0:          for (j=0;j<N0;j++)
michael@0:             tmp[j*stride+i] = X[i*N0+j];
michael@0:    }
michael@0:    for (j=0;j<N;j++)
michael@0:       X[j] = tmp[j];
michael@0:    RESTORE_STACK;
michael@0: }
michael@0: 
michael@0: void haar1(celt_norm *X, int N0, int stride)
michael@0: {
michael@0:    int i, j;
michael@0:    N0 >>= 1;
michael@0:    for (i=0;i<stride;i++)
michael@0:       for (j=0;j<N0;j++)
michael@0:       {
michael@0:          celt_norm tmp1, tmp2;
michael@0:          tmp1 = MULT16_16_Q15(QCONST16(.70710678f,15), X[stride*2*j+i]);
michael@0:          tmp2 = MULT16_16_Q15(QCONST16(.70710678f,15), X[stride*(2*j+1)+i]);
michael@0:          X[stride*2*j+i] = tmp1 + tmp2;
michael@0:          X[stride*(2*j+1)+i] = tmp1 - tmp2;
michael@0:       }
michael@0: }
michael@0: 
michael@0: static int compute_qn(int N, int b, int offset, int pulse_cap, int stereo)
michael@0: {
michael@0:    static const opus_int16 exp2_table8[8] =
michael@0:       {16384, 17866, 19483, 21247, 23170, 25267, 27554, 30048};
michael@0:    int qn, qb;
michael@0:    int N2 = 2*N-1;
michael@0:    if (stereo && N==2)
michael@0:       N2--;
michael@0:    /* The upper limit ensures that in a stereo split with itheta==16384, we'll
michael@0:        always have enough bits left over to code at least one pulse in the
michael@0:        side; otherwise it would collapse, since it doesn't get folded. */
michael@0:    qb = IMIN(b-pulse_cap-(4<<BITRES), (b+N2*offset)/N2);
michael@0: 
michael@0:    qb = IMIN(8<<BITRES, qb);
michael@0: 
michael@0:    if (qb<(1<<BITRES>>1)) {
michael@0:       qn = 1;
michael@0:    } else {
michael@0:       qn = exp2_table8[qb&0x7]>>(14-(qb>>BITRES));
michael@0:       qn = (qn+1)>>1<<1;
michael@0:    }
michael@0:    celt_assert(qn <= 256);
michael@0:    return qn;
michael@0: }
michael@0: 
michael@0: struct band_ctx {
michael@0:    int encode;
michael@0:    const CELTMode *m;
michael@0:    int i;
michael@0:    int intensity;
michael@0:    int spread;
michael@0:    int tf_change;
michael@0:    ec_ctx *ec;
michael@0:    opus_int32 remaining_bits;
michael@0:    const celt_ener *bandE;
michael@0:    opus_uint32 seed;
michael@0: };
michael@0: 
michael@0: struct split_ctx {
michael@0:    int inv;
michael@0:    int imid;
michael@0:    int iside;
michael@0:    int delta;
michael@0:    int itheta;
michael@0:    int qalloc;
michael@0: };
michael@0: 
michael@0: static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx,
michael@0:       celt_norm *X, celt_norm *Y, int N, int *b, int B, int B0,
michael@0:       int LM,
michael@0:       int stereo, int *fill)
michael@0: {
michael@0:    int qn;
michael@0:    int itheta=0;
michael@0:    int delta;
michael@0:    int imid, iside;
michael@0:    int qalloc;
michael@0:    int pulse_cap;
michael@0:    int offset;
michael@0:    opus_int32 tell;
michael@0:    int inv=0;
michael@0:    int encode;
michael@0:    const CELTMode *m;
michael@0:    int i;
michael@0:    int intensity;
michael@0:    ec_ctx *ec;
michael@0:    const celt_ener *bandE;
michael@0: 
michael@0:    encode = ctx->encode;
michael@0:    m = ctx->m;
michael@0:    i = ctx->i;
michael@0:    intensity = ctx->intensity;
michael@0:    ec = ctx->ec;
michael@0:    bandE = ctx->bandE;
michael@0: 
michael@0:    /* Decide on the resolution to give to the split parameter theta */
michael@0:    pulse_cap = m->logN[i]+LM*(1<<BITRES);
michael@0:    offset = (pulse_cap>>1) - (stereo&&N==2 ? QTHETA_OFFSET_TWOPHASE : QTHETA_OFFSET);
michael@0:    qn = compute_qn(N, *b, offset, pulse_cap, stereo);
michael@0:    if (stereo && i>=intensity)
michael@0:       qn = 1;
michael@0:    if (encode)
michael@0:    {
michael@0:       /* theta is the atan() of the ratio between the (normalized)
michael@0:          side and mid. With just that parameter, we can re-scale both
michael@0:          mid and side because we know that 1) they have unit norm and
michael@0:          2) they are orthogonal. */
michael@0:       itheta = stereo_itheta(X, Y, stereo, N);
michael@0:    }
michael@0:    tell = ec_tell_frac(ec);
michael@0:    if (qn!=1)
michael@0:    {
michael@0:       if (encode)
michael@0:          itheta = (itheta*qn+8192)>>14;
michael@0: 
michael@0:       /* Entropy coding of the angle. We use a uniform pdf for the
michael@0:          time split, a step for stereo, and a triangular one for the rest. */
michael@0:       if (stereo && N>2)
michael@0:       {
michael@0:          int p0 = 3;
michael@0:          int x = itheta;
michael@0:          int x0 = qn/2;
michael@0:          int ft = p0*(x0+1) + x0;
michael@0:          /* Use a probability of p0 up to itheta=8192 and then use 1 after */
michael@0:          if (encode)
michael@0:          {
michael@0:             ec_encode(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft);
michael@0:          } else {
michael@0:             int fs;
michael@0:             fs=ec_decode(ec,ft);
michael@0:             if (fs<(x0+1)*p0)
michael@0:                x=fs/p0;
michael@0:             else
michael@0:                x=x0+1+(fs-(x0+1)*p0);
michael@0:             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);
michael@0:             itheta = x;
michael@0:          }
michael@0:       } else if (B0>1 || stereo) {
michael@0:          /* Uniform pdf */
michael@0:          if (encode)
michael@0:             ec_enc_uint(ec, itheta, qn+1);
michael@0:          else
michael@0:             itheta = ec_dec_uint(ec, qn+1);
michael@0:       } else {
michael@0:          int fs=1, ft;
michael@0:          ft = ((qn>>1)+1)*((qn>>1)+1);
michael@0:          if (encode)
michael@0:          {
michael@0:             int fl;
michael@0: 
michael@0:             fs = itheta <= (qn>>1) ? itheta + 1 : qn + 1 - itheta;
michael@0:             fl = itheta <= (qn>>1) ? itheta*(itheta + 1)>>1 :
michael@0:              ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1);
michael@0: 
michael@0:             ec_encode(ec, fl, fl+fs, ft);
michael@0:          } else {
michael@0:             /* Triangular pdf */
michael@0:             int fl=0;
michael@0:             int fm;
michael@0:             fm = ec_decode(ec, ft);
michael@0: 
michael@0:             if (fm < ((qn>>1)*((qn>>1) + 1)>>1))
michael@0:             {
michael@0:                itheta = (isqrt32(8*(opus_uint32)fm + 1) - 1)>>1;
michael@0:                fs = itheta + 1;
michael@0:                fl = itheta*(itheta + 1)>>1;
michael@0:             }
michael@0:             else
michael@0:             {
michael@0:                itheta = (2*(qn + 1)
michael@0:                 - isqrt32(8*(opus_uint32)(ft - fm - 1) + 1))>>1;
michael@0:                fs = qn + 1 - itheta;
michael@0:                fl = ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1);
michael@0:             }
michael@0: 
michael@0:             ec_dec_update(ec, fl, fl+fs, ft);
michael@0:          }
michael@0:       }
michael@0:       itheta = (opus_int32)itheta*16384/qn;
michael@0:       if (encode && stereo)
michael@0:       {
michael@0:          if (itheta==0)
michael@0:             intensity_stereo(m, X, Y, bandE, i, N);
michael@0:          else
michael@0:             stereo_split(X, Y, N);
michael@0:       }
michael@0:       /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate.
michael@0:                Let's do that at higher complexity */
michael@0:    } else if (stereo) {
michael@0:       if (encode)
michael@0:       {
michael@0:          inv = itheta > 8192;
michael@0:          if (inv)
michael@0:          {
michael@0:             int j;
michael@0:             for (j=0;j<N;j++)
michael@0:                Y[j] = -Y[j];
michael@0:          }
michael@0:          intensity_stereo(m, X, Y, bandE, i, N);
michael@0:       }
michael@0:       if (*b>2<<BITRES && ctx->remaining_bits > 2<<BITRES)
michael@0:       {
michael@0:          if (encode)
michael@0:             ec_enc_bit_logp(ec, inv, 2);
michael@0:          else
michael@0:             inv = ec_dec_bit_logp(ec, 2);
michael@0:       } else
michael@0:          inv = 0;
michael@0:       itheta = 0;
michael@0:    }
michael@0:    qalloc = ec_tell_frac(ec) - tell;
michael@0:    *b -= qalloc;
michael@0: 
michael@0:    if (itheta == 0)
michael@0:    {
michael@0:       imid = 32767;
michael@0:       iside = 0;
michael@0:       *fill &= (1<<B)-1;
michael@0:       delta = -16384;
michael@0:    } else if (itheta == 16384)
michael@0:    {
michael@0:       imid = 0;
michael@0:       iside = 32767;
michael@0:       *fill &= ((1<<B)-1)<<B;
michael@0:       delta = 16384;
michael@0:    } else {
michael@0:       imid = bitexact_cos((opus_int16)itheta);
michael@0:       iside = bitexact_cos((opus_int16)(16384-itheta));
michael@0:       /* This is the mid vs side allocation that minimizes squared error
michael@0:          in that band. */
michael@0:       delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid));
michael@0:    }
michael@0: 
michael@0:    sctx->inv = inv;
michael@0:    sctx->imid = imid;
michael@0:    sctx->iside = iside;
michael@0:    sctx->delta = delta;
michael@0:    sctx->itheta = itheta;
michael@0:    sctx->qalloc = qalloc;
michael@0: }
michael@0: static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, int b,
michael@0:       celt_norm *lowband_out)
michael@0: {
michael@0: #ifdef RESYNTH
michael@0:    int resynth = 1;
michael@0: #else
michael@0:    int resynth = !ctx->encode;
michael@0: #endif
michael@0:    int c;
michael@0:    int stereo;
michael@0:    celt_norm *x = X;
michael@0:    int encode;
michael@0:    ec_ctx *ec;
michael@0: 
michael@0:    encode = ctx->encode;
michael@0:    ec = ctx->ec;
michael@0: 
michael@0:    stereo = Y != NULL;
michael@0:    c=0; do {
michael@0:       int sign=0;
michael@0:       if (ctx->remaining_bits>=1<<BITRES)
michael@0:       {
michael@0:          if (encode)
michael@0:          {
michael@0:             sign = x[0]<0;
michael@0:             ec_enc_bits(ec, sign, 1);
michael@0:          } else {
michael@0:             sign = ec_dec_bits(ec, 1);
michael@0:          }
michael@0:          ctx->remaining_bits -= 1<<BITRES;
michael@0:          b-=1<<BITRES;
michael@0:       }
michael@0:       if (resynth)
michael@0:          x[0] = sign ? -NORM_SCALING : NORM_SCALING;
michael@0:       x = Y;
michael@0:    } while (++c<1+stereo);
michael@0:    if (lowband_out)
michael@0:       lowband_out[0] = SHR16(X[0],4);
michael@0:    return 1;
michael@0: }
michael@0: 
michael@0: /* This function is responsible for encoding and decoding a mono partition.
michael@0:    It can split the band in two and transmit the energy difference with
michael@0:    the two half-bands. It can be called recursively so bands can end up being
michael@0:    split in 8 parts. */
michael@0: static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X,
michael@0:       int N, int b, int B, celt_norm *lowband,
michael@0:       int LM,
michael@0:       opus_val16 gain, int fill)
michael@0: {
michael@0:    const unsigned char *cache;
michael@0:    int q;
michael@0:    int curr_bits;
michael@0:    int imid=0, iside=0;
michael@0:    int B0=B;
michael@0:    opus_val16 mid=0, side=0;
michael@0:    unsigned cm=0;
michael@0: #ifdef RESYNTH
michael@0:    int resynth = 1;
michael@0: #else
michael@0:    int resynth = !ctx->encode;
michael@0: #endif
michael@0:    celt_norm *Y=NULL;
michael@0:    int encode;
michael@0:    const CELTMode *m;
michael@0:    int i;
michael@0:    int spread;
michael@0:    ec_ctx *ec;
michael@0: 
michael@0:    encode = ctx->encode;
michael@0:    m = ctx->m;
michael@0:    i = ctx->i;
michael@0:    spread = ctx->spread;
michael@0:    ec = ctx->ec;
michael@0: 
michael@0:    /* If we need 1.5 more bit than we can produce, split the band in two. */
michael@0:    cache = m->cache.bits + m->cache.index[(LM+1)*m->nbEBands+i];
michael@0:    if (LM != -1 && b > cache[cache[0]]+12 && N>2)
michael@0:    {
michael@0:       int mbits, sbits, delta;
michael@0:       int itheta;
michael@0:       int qalloc;
michael@0:       struct split_ctx sctx;
michael@0:       celt_norm *next_lowband2=NULL;
michael@0:       opus_int32 rebalance;
michael@0: 
michael@0:       N >>= 1;
michael@0:       Y = X+N;
michael@0:       LM -= 1;
michael@0:       if (B==1)
michael@0:          fill = (fill&1)|(fill<<1);
michael@0:       B = (B+1)>>1;
michael@0: 
michael@0:       compute_theta(ctx, &sctx, X, Y, N, &b, B, B0,
michael@0:             LM, 0, &fill);
michael@0:       imid = sctx.imid;
michael@0:       iside = sctx.iside;
michael@0:       delta = sctx.delta;
michael@0:       itheta = sctx.itheta;
michael@0:       qalloc = sctx.qalloc;
michael@0: #ifdef FIXED_POINT
michael@0:       mid = imid;
michael@0:       side = iside;
michael@0: #else
michael@0:       mid = (1.f/32768)*imid;
michael@0:       side = (1.f/32768)*iside;
michael@0: #endif
michael@0: 
michael@0:       /* Give more bits to low-energy MDCTs than they would otherwise deserve */
michael@0:       if (B0>1 && (itheta&0x3fff))
michael@0:       {
michael@0:          if (itheta > 8192)
michael@0:             /* Rough approximation for pre-echo masking */
michael@0:             delta -= delta>>(4-LM);
michael@0:          else
michael@0:             /* Corresponds to a forward-masking slope of 1.5 dB per 10 ms */
michael@0:             delta = IMIN(0, delta + (N<<BITRES>>(5-LM)));
michael@0:       }
michael@0:       mbits = IMAX(0, IMIN(b, (b-delta)/2));
michael@0:       sbits = b-mbits;
michael@0:       ctx->remaining_bits -= qalloc;
michael@0: 
michael@0:       if (lowband)
michael@0:          next_lowband2 = lowband+N; /* >32-bit split case */
michael@0: 
michael@0:       rebalance = ctx->remaining_bits;
michael@0:       if (mbits >= sbits)
michael@0:       {
michael@0:          cm = quant_partition(ctx, X, N, mbits, B,
michael@0:                lowband, LM,
michael@0:                MULT16_16_P15(gain,mid), fill);
michael@0:          rebalance = mbits - (rebalance-ctx->remaining_bits);
michael@0:          if (rebalance > 3<<BITRES && itheta!=0)
michael@0:             sbits += rebalance - (3<<BITRES);
michael@0:          cm |= quant_partition(ctx, Y, N, sbits, B,
michael@0:                next_lowband2, LM,
michael@0:                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
michael@0:       } else {
michael@0:          cm = quant_partition(ctx, Y, N, sbits, B,
michael@0:                next_lowband2, LM,
michael@0:                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
michael@0:          rebalance = sbits - (rebalance-ctx->remaining_bits);
michael@0:          if (rebalance > 3<<BITRES && itheta!=16384)
michael@0:             mbits += rebalance - (3<<BITRES);
michael@0:          cm |= quant_partition(ctx, X, N, mbits, B,
michael@0:                lowband, LM,
michael@0:                MULT16_16_P15(gain,mid), fill);
michael@0:       }
michael@0:    } else {
michael@0:       /* This is the basic no-split case */
michael@0:       q = bits2pulses(m, i, LM, b);
michael@0:       curr_bits = pulses2bits(m, i, LM, q);
michael@0:       ctx->remaining_bits -= curr_bits;
michael@0: 
michael@0:       /* Ensures we can never bust the budget */
michael@0:       while (ctx->remaining_bits < 0 && q > 0)
michael@0:       {
michael@0:          ctx->remaining_bits += curr_bits;
michael@0:          q--;
michael@0:          curr_bits = pulses2bits(m, i, LM, q);
michael@0:          ctx->remaining_bits -= curr_bits;
michael@0:       }
michael@0: 
michael@0:       if (q!=0)
michael@0:       {
michael@0:          int K = get_pulses(q);
michael@0: 
michael@0:          /* Finally do the actual quantization */
michael@0:          if (encode)
michael@0:          {
michael@0:             cm = alg_quant(X, N, K, spread, B, ec
michael@0: #ifdef RESYNTH
michael@0:                  , gain
michael@0: #endif
michael@0:                  );
michael@0:          } else {
michael@0:             cm = alg_unquant(X, N, K, spread, B, ec, gain);
michael@0:          }
michael@0:       } else {
michael@0:          /* If there's no pulse, fill the band anyway */
michael@0:          int j;
michael@0:          if (resynth)
michael@0:          {
michael@0:             unsigned cm_mask;
michael@0:             /* B can be as large as 16, so this shift might overflow an int on a
michael@0:                16-bit platform; use a long to get defined behavior.*/
michael@0:             cm_mask = (unsigned)(1UL<<B)-1;
michael@0:             fill &= cm_mask;
michael@0:             if (!fill)
michael@0:             {
michael@0:                for (j=0;j<N;j++)
michael@0:                   X[j] = 0;
michael@0:             } else {
michael@0:                if (lowband == NULL)
michael@0:                {
michael@0:                   /* Noise */
michael@0:                   for (j=0;j<N;j++)
michael@0:                   {
michael@0:                      ctx->seed = celt_lcg_rand(ctx->seed);
michael@0:                      X[j] = (celt_norm)((opus_int32)ctx->seed>>20);
michael@0:                   }
michael@0:                   cm = cm_mask;
michael@0:                } else {
michael@0:                   /* Folded spectrum */
michael@0:                   for (j=0;j<N;j++)
michael@0:                   {
michael@0:                      opus_val16 tmp;
michael@0:                      ctx->seed = celt_lcg_rand(ctx->seed);
michael@0:                      /* About 48 dB below the "normal" folding level */
michael@0:                      tmp = QCONST16(1.0f/256, 10);
michael@0:                      tmp = (ctx->seed)&0x8000 ? tmp : -tmp;
michael@0:                      X[j] = lowband[j]+tmp;
michael@0:                   }
michael@0:                   cm = fill;
michael@0:                }
michael@0:                renormalise_vector(X, N, gain);
michael@0:             }
michael@0:          }
michael@0:       }
michael@0:    }
michael@0: 
michael@0:    return cm;
michael@0: }
michael@0: 
michael@0: 
michael@0: /* This function is responsible for encoding and decoding a band for the mono case. */
michael@0: static unsigned quant_band(struct band_ctx *ctx, celt_norm *X,
michael@0:       int N, int b, int B, celt_norm *lowband,
michael@0:       int LM, celt_norm *lowband_out,
michael@0:       opus_val16 gain, celt_norm *lowband_scratch, int fill)
michael@0: {
michael@0:    int N0=N;
michael@0:    int N_B=N;
michael@0:    int N_B0;
michael@0:    int B0=B;
michael@0:    int time_divide=0;
michael@0:    int recombine=0;
michael@0:    int longBlocks;
michael@0:    unsigned cm=0;
michael@0: #ifdef RESYNTH
michael@0:    int resynth = 1;
michael@0: #else
michael@0:    int resynth = !ctx->encode;
michael@0: #endif
michael@0:    int k;
michael@0:    int encode;
michael@0:    int tf_change;
michael@0: 
michael@0:    encode = ctx->encode;
michael@0:    tf_change = ctx->tf_change;
michael@0: 
michael@0:    longBlocks = B0==1;
michael@0: 
michael@0:    N_B /= B;
michael@0: 
michael@0:    /* Special case for one sample */
michael@0:    if (N==1)
michael@0:    {
michael@0:       return quant_band_n1(ctx, X, NULL, b, lowband_out);
michael@0:    }
michael@0: 
michael@0:    if (tf_change>0)
michael@0:       recombine = tf_change;
michael@0:    /* Band recombining to increase frequency resolution */
michael@0: 
michael@0:    if (lowband_scratch && lowband && (recombine || ((N_B&1) == 0 && tf_change<0) || B0>1))
michael@0:    {
michael@0:       int j;
michael@0:       for (j=0;j<N;j++)
michael@0:          lowband_scratch[j] = lowband[j];
michael@0:       lowband = lowband_scratch;
michael@0:    }
michael@0: 
michael@0:    for (k=0;k<recombine;k++)
michael@0:    {
michael@0:       static const unsigned char bit_interleave_table[16]={
michael@0:             0,1,1,1,2,3,3,3,2,3,3,3,2,3,3,3
michael@0:       };
michael@0:       if (encode)
michael@0:          haar1(X, N>>k, 1<<k);
michael@0:       if (lowband)
michael@0:          haar1(lowband, N>>k, 1<<k);
michael@0:       fill = bit_interleave_table[fill&0xF]|bit_interleave_table[fill>>4]<<2;
michael@0:    }
michael@0:    B>>=recombine;
michael@0:    N_B<<=recombine;
michael@0: 
michael@0:    /* Increasing the time resolution */
michael@0:    while ((N_B&1) == 0 && tf_change<0)
michael@0:    {
michael@0:       if (encode)
michael@0:          haar1(X, N_B, B);
michael@0:       if (lowband)
michael@0:          haar1(lowband, N_B, B);
michael@0:       fill |= fill<<B;
michael@0:       B <<= 1;
michael@0:       N_B >>= 1;
michael@0:       time_divide++;
michael@0:       tf_change++;
michael@0:    }
michael@0:    B0=B;
michael@0:    N_B0 = N_B;
michael@0: 
michael@0:    /* Reorganize the samples in time order instead of frequency order */
michael@0:    if (B0>1)
michael@0:    {
michael@0:       if (encode)
michael@0:          deinterleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
michael@0:       if (lowband)
michael@0:          deinterleave_hadamard(lowband, N_B>>recombine, B0<<recombine, longBlocks);
michael@0:    }
michael@0: 
michael@0:    cm = quant_partition(ctx, X, N, b, B, lowband,
michael@0:          LM, gain, fill);
michael@0: 
michael@0:    /* This code is used by the decoder and by the resynthesis-enabled encoder */
michael@0:    if (resynth)
michael@0:    {
michael@0:       /* Undo the sample reorganization going from time order to frequency order */
michael@0:       if (B0>1)
michael@0:          interleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
michael@0: 
michael@0:       /* Undo time-freq changes that we did earlier */
michael@0:       N_B = N_B0;
michael@0:       B = B0;
michael@0:       for (k=0;k<time_divide;k++)
michael@0:       {
michael@0:          B >>= 1;
michael@0:          N_B <<= 1;
michael@0:          cm |= cm>>B;
michael@0:          haar1(X, N_B, B);
michael@0:       }
michael@0: 
michael@0:       for (k=0;k<recombine;k++)
michael@0:       {
michael@0:          static const unsigned char bit_deinterleave_table[16]={
michael@0:                0x00,0x03,0x0C,0x0F,0x30,0x33,0x3C,0x3F,
michael@0:                0xC0,0xC3,0xCC,0xCF,0xF0,0xF3,0xFC,0xFF
michael@0:          };
michael@0:          cm = bit_deinterleave_table[cm];
michael@0:          haar1(X, N0>>k, 1<<k);
michael@0:       }
michael@0:       B<<=recombine;
michael@0: 
michael@0:       /* Scale output for later folding */
michael@0:       if (lowband_out)
michael@0:       {
michael@0:          int j;
michael@0:          opus_val16 n;
michael@0:          n = celt_sqrt(SHL32(EXTEND32(N0),22));
michael@0:          for (j=0;j<N0;j++)
michael@0:             lowband_out[j] = MULT16_16_Q15(n,X[j]);
michael@0:       }
michael@0:       cm &= (1<<B)-1;
michael@0:    }
michael@0:    return cm;
michael@0: }
michael@0: 
michael@0: 
michael@0: /* This function is responsible for encoding and decoding a band for the stereo case. */
michael@0: static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm *Y,
michael@0:       int N, int b, int B, celt_norm *lowband,
michael@0:       int LM, celt_norm *lowband_out,
michael@0:       celt_norm *lowband_scratch, int fill)
michael@0: {
michael@0:    int imid=0, iside=0;
michael@0:    int inv = 0;
michael@0:    opus_val16 mid=0, side=0;
michael@0:    unsigned cm=0;
michael@0: #ifdef RESYNTH
michael@0:    int resynth = 1;
michael@0: #else
michael@0:    int resynth = !ctx->encode;
michael@0: #endif
michael@0:    int mbits, sbits, delta;
michael@0:    int itheta;
michael@0:    int qalloc;
michael@0:    struct split_ctx sctx;
michael@0:    int orig_fill;
michael@0:    int encode;
michael@0:    ec_ctx *ec;
michael@0: 
michael@0:    encode = ctx->encode;
michael@0:    ec = ctx->ec;
michael@0: 
michael@0:    /* Special case for one sample */
michael@0:    if (N==1)
michael@0:    {
michael@0:       return quant_band_n1(ctx, X, Y, b, lowband_out);
michael@0:    }
michael@0: 
michael@0:    orig_fill = fill;
michael@0: 
michael@0:    compute_theta(ctx, &sctx, X, Y, N, &b, B, B,
michael@0:          LM, 1, &fill);
michael@0:    inv = sctx.inv;
michael@0:    imid = sctx.imid;
michael@0:    iside = sctx.iside;
michael@0:    delta = sctx.delta;
michael@0:    itheta = sctx.itheta;
michael@0:    qalloc = sctx.qalloc;
michael@0: #ifdef FIXED_POINT
michael@0:    mid = imid;
michael@0:    side = iside;
michael@0: #else
michael@0:    mid = (1.f/32768)*imid;
michael@0:    side = (1.f/32768)*iside;
michael@0: #endif
michael@0: 
michael@0:    /* This is a special case for N=2 that only works for stereo and takes
michael@0:       advantage of the fact that mid and side are orthogonal to encode
michael@0:       the side with just one bit. */
michael@0:    if (N==2)
michael@0:    {
michael@0:       int c;
michael@0:       int sign=0;
michael@0:       celt_norm *x2, *y2;
michael@0:       mbits = b;
michael@0:       sbits = 0;
michael@0:       /* Only need one bit for the side. */
michael@0:       if (itheta != 0 && itheta != 16384)
michael@0:          sbits = 1<<BITRES;
michael@0:       mbits -= sbits;
michael@0:       c = itheta > 8192;
michael@0:       ctx->remaining_bits -= qalloc+sbits;
michael@0: 
michael@0:       x2 = c ? Y : X;
michael@0:       y2 = c ? X : Y;
michael@0:       if (sbits)
michael@0:       {
michael@0:          if (encode)
michael@0:          {
michael@0:             /* Here we only need to encode a sign for the side. */
michael@0:             sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
michael@0:             ec_enc_bits(ec, sign, 1);
michael@0:          } else {
michael@0:             sign = ec_dec_bits(ec, 1);
michael@0:          }
michael@0:       }
michael@0:       sign = 1-2*sign;
michael@0:       /* We use orig_fill here because we want to fold the side, but if
michael@0:          itheta==16384, we'll have cleared the low bits of fill. */
michael@0:       cm = quant_band(ctx, x2, N, mbits, B, lowband,
michael@0:             LM, lowband_out, Q15ONE, lowband_scratch, orig_fill);
michael@0:       /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
michael@0:          and there's no need to worry about mixing with the other channel. */
michael@0:       y2[0] = -sign*x2[1];
michael@0:       y2[1] = sign*x2[0];
michael@0:       if (resynth)
michael@0:       {
michael@0:          celt_norm tmp;
michael@0:          X[0] = MULT16_16_Q15(mid, X[0]);
michael@0:          X[1] = MULT16_16_Q15(mid, X[1]);
michael@0:          Y[0] = MULT16_16_Q15(side, Y[0]);
michael@0:          Y[1] = MULT16_16_Q15(side, Y[1]);
michael@0:          tmp = X[0];
michael@0:          X[0] = SUB16(tmp,Y[0]);
michael@0:          Y[0] = ADD16(tmp,Y[0]);
michael@0:          tmp = X[1];
michael@0:          X[1] = SUB16(tmp,Y[1]);
michael@0:          Y[1] = ADD16(tmp,Y[1]);
michael@0:       }
michael@0:    } else {
michael@0:       /* "Normal" split code */
michael@0:       opus_int32 rebalance;
michael@0: 
michael@0:       mbits = IMAX(0, IMIN(b, (b-delta)/2));
michael@0:       sbits = b-mbits;
michael@0:       ctx->remaining_bits -= qalloc;
michael@0: 
michael@0:       rebalance = ctx->remaining_bits;
michael@0:       if (mbits >= sbits)
michael@0:       {
michael@0:          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
michael@0:             mid for folding later. */
michael@0:          cm = quant_band(ctx, X, N, mbits, B,
michael@0:                lowband, LM, lowband_out,
michael@0:                Q15ONE, lowband_scratch, fill);
michael@0:          rebalance = mbits - (rebalance-ctx->remaining_bits);
michael@0:          if (rebalance > 3<<BITRES && itheta!=0)
michael@0:             sbits += rebalance - (3<<BITRES);
michael@0: 
michael@0:          /* For a stereo split, the high bits of fill are always zero, so no
michael@0:             folding will be done to the side. */
michael@0:          cm |= quant_band(ctx, Y, N, sbits, B,
michael@0:                NULL, LM, NULL,
michael@0:                side, NULL, fill>>B);
michael@0:       } else {
michael@0:          /* For a stereo split, the high bits of fill are always zero, so no
michael@0:             folding will be done to the side. */
michael@0:          cm = quant_band(ctx, Y, N, sbits, B,
michael@0:                NULL, LM, NULL,
michael@0:                side, NULL, fill>>B);
michael@0:          rebalance = sbits - (rebalance-ctx->remaining_bits);
michael@0:          if (rebalance > 3<<BITRES && itheta!=16384)
michael@0:             mbits += rebalance - (3<<BITRES);
michael@0:          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
michael@0:             mid for folding later. */
michael@0:          cm |= quant_band(ctx, X, N, mbits, B,
michael@0:                lowband, LM, lowband_out,
michael@0:                Q15ONE, lowband_scratch, fill);
michael@0:       }
michael@0:    }
michael@0: 
michael@0: 
michael@0:    /* This code is used by the decoder and by the resynthesis-enabled encoder */
michael@0:    if (resynth)
michael@0:    {
michael@0:       if (N!=2)
michael@0:          stereo_merge(X, Y, mid, N);
michael@0:       if (inv)
michael@0:       {
michael@0:          int j;
michael@0:          for (j=0;j<N;j++)
michael@0:             Y[j] = -Y[j];
michael@0:       }
michael@0:    }
michael@0:    return cm;
michael@0: }
michael@0: 
michael@0: 
michael@0: void quant_all_bands(int encode, const CELTMode *m, int start, int end,
michael@0:       celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks, const celt_ener *bandE, int *pulses,
michael@0:       int shortBlocks, int spread, int dual_stereo, int intensity, int *tf_res,
michael@0:       opus_int32 total_bits, opus_int32 balance, ec_ctx *ec, int LM, int codedBands, opus_uint32 *seed)
michael@0: {
michael@0:    int i;
michael@0:    opus_int32 remaining_bits;
michael@0:    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
michael@0:    celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2;
michael@0:    VARDECL(celt_norm, _norm);
michael@0:    celt_norm *lowband_scratch;
michael@0:    int B;
michael@0:    int M;
michael@0:    int lowband_offset;
michael@0:    int update_lowband = 1;
michael@0:    int C = Y_ != NULL ? 2 : 1;
michael@0:    int norm_offset;
michael@0: #ifdef RESYNTH
michael@0:    int resynth = 1;
michael@0: #else
michael@0:    int resynth = !encode;
michael@0: #endif
michael@0:    struct band_ctx ctx;
michael@0:    SAVE_STACK;
michael@0: 
michael@0:    M = 1<<LM;
michael@0:    B = shortBlocks ? M : 1;
michael@0:    norm_offset = M*eBands[start];
michael@0:    /* No need to allocate norm for the last band because we don't need an
michael@0:       output in that band. */
michael@0:    ALLOC(_norm, C*(M*eBands[m->nbEBands-1]-norm_offset), celt_norm);
michael@0:    norm = _norm;
michael@0:    norm2 = norm + M*eBands[m->nbEBands-1]-norm_offset;
michael@0:    /* We can use the last band as scratch space because we don't need that
michael@0:       scratch space for the last band. */
michael@0:    lowband_scratch = X_+M*eBands[m->nbEBands-1];
michael@0: 
michael@0:    lowband_offset = 0;
michael@0:    ctx.bandE = bandE;
michael@0:    ctx.ec = ec;
michael@0:    ctx.encode = encode;
michael@0:    ctx.intensity = intensity;
michael@0:    ctx.m = m;
michael@0:    ctx.seed = *seed;
michael@0:    ctx.spread = spread;
michael@0:    for (i=start;i<end;i++)
michael@0:    {
michael@0:       opus_int32 tell;
michael@0:       int b;
michael@0:       int N;
michael@0:       opus_int32 curr_balance;
michael@0:       int effective_lowband=-1;
michael@0:       celt_norm * OPUS_RESTRICT X, * OPUS_RESTRICT Y;
michael@0:       int tf_change=0;
michael@0:       unsigned x_cm;
michael@0:       unsigned y_cm;
michael@0:       int last;
michael@0: 
michael@0:       ctx.i = i;
michael@0:       last = (i==end-1);
michael@0: 
michael@0:       X = X_+M*eBands[i];
michael@0:       if (Y_!=NULL)
michael@0:          Y = Y_+M*eBands[i];
michael@0:       else
michael@0:          Y = NULL;
michael@0:       N = M*eBands[i+1]-M*eBands[i];
michael@0:       tell = ec_tell_frac(ec);
michael@0: 
michael@0:       /* Compute how many bits we want to allocate to this band */
michael@0:       if (i != start)
michael@0:          balance -= tell;
michael@0:       remaining_bits = total_bits-tell-1;
michael@0:       ctx.remaining_bits = remaining_bits;
michael@0:       if (i <= codedBands-1)
michael@0:       {
michael@0:          curr_balance = balance / IMIN(3, codedBands-i);
michael@0:          b = IMAX(0, IMIN(16383, IMIN(remaining_bits+1,pulses[i]+curr_balance)));
michael@0:       } else {
michael@0:          b = 0;
michael@0:       }
michael@0: 
michael@0:       if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0))
michael@0:             lowband_offset = i;
michael@0: 
michael@0:       tf_change = tf_res[i];
michael@0:       ctx.tf_change = tf_change;
michael@0:       if (i>=m->effEBands)
michael@0:       {
michael@0:          X=norm;
michael@0:          if (Y_!=NULL)
michael@0:             Y = norm;
michael@0:          lowband_scratch = NULL;
michael@0:       }
michael@0:       if (i==end-1)
michael@0:          lowband_scratch = NULL;
michael@0: 
michael@0:       /* Get a conservative estimate of the collapse_mask's for the bands we're
michael@0:          going to be folding from. */
michael@0:       if (lowband_offset != 0 && (spread!=SPREAD_AGGRESSIVE || B>1 || tf_change<0))
michael@0:       {
michael@0:          int fold_start;
michael@0:          int fold_end;
michael@0:          int fold_i;
michael@0:          /* This ensures we never repeat spectral content within one band */
michael@0:          effective_lowband = IMAX(0, M*eBands[lowband_offset]-norm_offset-N);
michael@0:          fold_start = lowband_offset;
michael@0:          while(M*eBands[--fold_start] > effective_lowband+norm_offset);
michael@0:          fold_end = lowband_offset-1;
michael@0:          while(M*eBands[++fold_end] < effective_lowband+norm_offset+N);
michael@0:          x_cm = y_cm = 0;
michael@0:          fold_i = fold_start; do {
michael@0:            x_cm |= collapse_masks[fold_i*C+0];
michael@0:            y_cm |= collapse_masks[fold_i*C+C-1];
michael@0:          } while (++fold_i<fold_end);
michael@0:       }
michael@0:       /* Otherwise, we'll be using the LCG to fold, so all blocks will (almost
michael@0:          always) be non-zero. */
michael@0:       else
michael@0:          x_cm = y_cm = (1<<B)-1;
michael@0: 
michael@0:       if (dual_stereo && i==intensity)
michael@0:       {
michael@0:          int j;
michael@0: 
michael@0:          /* Switch off dual stereo to do intensity. */
michael@0:          dual_stereo = 0;
michael@0:          if (resynth)
michael@0:             for (j=0;j<M*eBands[i]-norm_offset;j++)
michael@0:                norm[j] = HALF32(norm[j]+norm2[j]);
michael@0:       }
michael@0:       if (dual_stereo)
michael@0:       {
michael@0:          x_cm = quant_band(&ctx, X, N, b/2, B,
michael@0:                effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
michael@0:                last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm);
michael@0:          y_cm = quant_band(&ctx, Y, N, b/2, B,
michael@0:                effective_lowband != -1 ? norm2+effective_lowband : NULL, LM,
michael@0:                last?NULL:norm2+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, y_cm);
michael@0:       } else {
michael@0:          if (Y!=NULL)
michael@0:          {
michael@0:             x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
michael@0:                   effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
michael@0:                         last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm);
michael@0:          } else {
michael@0:             x_cm = quant_band(&ctx, X, N, b, B,
michael@0:                   effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
michael@0:                         last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm);
michael@0:          }
michael@0:          y_cm = x_cm;
michael@0:       }
michael@0:       collapse_masks[i*C+0] = (unsigned char)x_cm;
michael@0:       collapse_masks[i*C+C-1] = (unsigned char)y_cm;
michael@0:       balance += pulses[i] + tell;
michael@0: 
michael@0:       /* Update the folding position only as long as we have 1 bit/sample depth. */
michael@0:       update_lowband = b>(N<<BITRES);
michael@0:    }
michael@0:    *seed = ctx.seed;
michael@0: 
michael@0:    RESTORE_STACK;
michael@0: }
michael@0: