michael@0: /* Copyright (c) 2007-2008 CSIRO
michael@0:    Copyright (c) 2007-2009 Xiph.Org Foundation
michael@0:    Written by Jean-Marc Valin */
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 "mathops.h"
michael@0: #include "cwrs.h"
michael@0: #include "vq.h"
michael@0: #include "arch.h"
michael@0: #include "os_support.h"
michael@0: #include "bands.h"
michael@0: #include "rate.h"
michael@0: 
michael@0: static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s)
michael@0: {
michael@0:    int i;
michael@0:    celt_norm *Xptr;
michael@0:    Xptr = X;
michael@0:    for (i=0;i<len-stride;i++)
michael@0:    {
michael@0:       celt_norm x1, x2;
michael@0:       x1 = Xptr[0];
michael@0:       x2 = Xptr[stride];
michael@0:       Xptr[stride] = EXTRACT16(SHR32(MULT16_16(c,x2) + MULT16_16(s,x1), 15));
michael@0:       *Xptr++      = EXTRACT16(SHR32(MULT16_16(c,x1) - MULT16_16(s,x2), 15));
michael@0:    }
michael@0:    Xptr = &X[len-2*stride-1];
michael@0:    for (i=len-2*stride-1;i>=0;i--)
michael@0:    {
michael@0:       celt_norm x1, x2;
michael@0:       x1 = Xptr[0];
michael@0:       x2 = Xptr[stride];
michael@0:       Xptr[stride] = EXTRACT16(SHR32(MULT16_16(c,x2) + MULT16_16(s,x1), 15));
michael@0:       *Xptr--      = EXTRACT16(SHR32(MULT16_16(c,x1) - MULT16_16(s,x2), 15));
michael@0:    }
michael@0: }
michael@0: 
michael@0: static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
michael@0: {
michael@0:    static const int SPREAD_FACTOR[3]={15,10,5};
michael@0:    int i;
michael@0:    opus_val16 c, s;
michael@0:    opus_val16 gain, theta;
michael@0:    int stride2=0;
michael@0:    int factor;
michael@0: 
michael@0:    if (2*K>=len || spread==SPREAD_NONE)
michael@0:       return;
michael@0:    factor = SPREAD_FACTOR[spread-1];
michael@0: 
michael@0:    gain = celt_div((opus_val32)MULT16_16(Q15_ONE,len),(opus_val32)(len+factor*K));
michael@0:    theta = HALF16(MULT16_16_Q15(gain,gain));
michael@0: 
michael@0:    c = celt_cos_norm(EXTEND32(theta));
michael@0:    s = celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /*  sin(theta) */
michael@0: 
michael@0:    if (len>=8*stride)
michael@0:    {
michael@0:       stride2 = 1;
michael@0:       /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
michael@0:          It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
michael@0:       while ((stride2*stride2+stride2)*stride + (stride>>2) < len)
michael@0:          stride2++;
michael@0:    }
michael@0:    /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
michael@0:       extract_collapse_mask().*/
michael@0:    len /= stride;
michael@0:    for (i=0;i<stride;i++)
michael@0:    {
michael@0:       if (dir < 0)
michael@0:       {
michael@0:          if (stride2)
michael@0:             exp_rotation1(X+i*len, len, stride2, s, c);
michael@0:          exp_rotation1(X+i*len, len, 1, c, s);
michael@0:       } else {
michael@0:          exp_rotation1(X+i*len, len, 1, c, -s);
michael@0:          if (stride2)
michael@0:             exp_rotation1(X+i*len, len, stride2, s, -c);
michael@0:       }
michael@0:    }
michael@0: }
michael@0: 
michael@0: /** Takes the pitch vector and the decoded residual vector, computes the gain
michael@0:     that will give ||p+g*y||=1 and mixes the residual with the pitch. */
michael@0: static void normalise_residual(int * OPUS_RESTRICT iy, celt_norm * OPUS_RESTRICT X,
michael@0:       int N, opus_val32 Ryy, opus_val16 gain)
michael@0: {
michael@0:    int i;
michael@0: #ifdef FIXED_POINT
michael@0:    int k;
michael@0: #endif
michael@0:    opus_val32 t;
michael@0:    opus_val16 g;
michael@0: 
michael@0: #ifdef FIXED_POINT
michael@0:    k = celt_ilog2(Ryy)>>1;
michael@0: #endif
michael@0:    t = VSHR32(Ryy, 2*(k-7));
michael@0:    g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
michael@0: 
michael@0:    i=0;
michael@0:    do
michael@0:       X[i] = EXTRACT16(PSHR32(MULT16_16(g, iy[i]), k+1));
michael@0:    while (++i < N);
michael@0: }
michael@0: 
michael@0: static unsigned extract_collapse_mask(int *iy, int N, int B)
michael@0: {
michael@0:    unsigned collapse_mask;
michael@0:    int N0;
michael@0:    int i;
michael@0:    if (B<=1)
michael@0:       return 1;
michael@0:    /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
michael@0:       exp_rotation().*/
michael@0:    N0 = N/B;
michael@0:    collapse_mask = 0;
michael@0:    i=0; do {
michael@0:       int j;
michael@0:       j=0; do {
michael@0:          collapse_mask |= (iy[i*N0+j]!=0)<<i;
michael@0:       } while (++j<N0);
michael@0:    } while (++i<B);
michael@0:    return collapse_mask;
michael@0: }
michael@0: 
michael@0: unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc
michael@0: #ifdef RESYNTH
michael@0:    , opus_val16 gain
michael@0: #endif
michael@0:    )
michael@0: {
michael@0:    VARDECL(celt_norm, y);
michael@0:    VARDECL(int, iy);
michael@0:    VARDECL(opus_val16, signx);
michael@0:    int i, j;
michael@0:    opus_val16 s;
michael@0:    int pulsesLeft;
michael@0:    opus_val32 sum;
michael@0:    opus_val32 xy;
michael@0:    opus_val16 yy;
michael@0:    unsigned collapse_mask;
michael@0:    SAVE_STACK;
michael@0: 
michael@0:    celt_assert2(K>0, "alg_quant() needs at least one pulse");
michael@0:    celt_assert2(N>1, "alg_quant() needs at least two dimensions");
michael@0: 
michael@0:    ALLOC(y, N, celt_norm);
michael@0:    ALLOC(iy, N, int);
michael@0:    ALLOC(signx, N, opus_val16);
michael@0: 
michael@0:    exp_rotation(X, N, 1, B, K, spread);
michael@0: 
michael@0:    /* Get rid of the sign */
michael@0:    sum = 0;
michael@0:    j=0; do {
michael@0:       if (X[j]>0)
michael@0:          signx[j]=1;
michael@0:       else {
michael@0:          signx[j]=-1;
michael@0:          X[j]=-X[j];
michael@0:       }
michael@0:       iy[j] = 0;
michael@0:       y[j] = 0;
michael@0:    } while (++j<N);
michael@0: 
michael@0:    xy = yy = 0;
michael@0: 
michael@0:    pulsesLeft = K;
michael@0: 
michael@0:    /* Do a pre-search by projecting on the pyramid */
michael@0:    if (K > (N>>1))
michael@0:    {
michael@0:       opus_val16 rcp;
michael@0:       j=0; do {
michael@0:          sum += X[j];
michael@0:       }  while (++j<N);
michael@0: 
michael@0:       /* If X is too small, just replace it with a pulse at 0 */
michael@0: #ifdef FIXED_POINT
michael@0:       if (sum <= K)
michael@0: #else
michael@0:       /* Prevents infinities and NaNs from causing too many pulses
michael@0:          to be allocated. 64 is an approximation of infinity here. */
michael@0:       if (!(sum > EPSILON && sum < 64))
michael@0: #endif
michael@0:       {
michael@0:          X[0] = QCONST16(1.f,14);
michael@0:          j=1; do
michael@0:             X[j]=0;
michael@0:          while (++j<N);
michael@0:          sum = QCONST16(1.f,14);
michael@0:       }
michael@0:       rcp = EXTRACT16(MULT16_32_Q16(K-1, celt_rcp(sum)));
michael@0:       j=0; do {
michael@0: #ifdef FIXED_POINT
michael@0:          /* It's really important to round *towards zero* here */
michael@0:          iy[j] = MULT16_16_Q15(X[j],rcp);
michael@0: #else
michael@0:          iy[j] = (int)floor(rcp*X[j]);
michael@0: #endif
michael@0:          y[j] = (celt_norm)iy[j];
michael@0:          yy = MAC16_16(yy, y[j],y[j]);
michael@0:          xy = MAC16_16(xy, X[j],y[j]);
michael@0:          y[j] *= 2;
michael@0:          pulsesLeft -= iy[j];
michael@0:       }  while (++j<N);
michael@0:    }
michael@0:    celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass");
michael@0: 
michael@0:    /* This should never happen, but just in case it does (e.g. on silence)
michael@0:       we fill the first bin with pulses. */
michael@0: #ifdef FIXED_POINT_DEBUG
michael@0:    celt_assert2(pulsesLeft<=N+3, "Not enough pulses in the quick pass");
michael@0: #endif
michael@0:    if (pulsesLeft > N+3)
michael@0:    {
michael@0:       opus_val16 tmp = (opus_val16)pulsesLeft;
michael@0:       yy = MAC16_16(yy, tmp, tmp);
michael@0:       yy = MAC16_16(yy, tmp, y[0]);
michael@0:       iy[0] += pulsesLeft;
michael@0:       pulsesLeft=0;
michael@0:    }
michael@0: 
michael@0:    s = 1;
michael@0:    for (i=0;i<pulsesLeft;i++)
michael@0:    {
michael@0:       int best_id;
michael@0:       opus_val32 best_num = -VERY_LARGE16;
michael@0:       opus_val16 best_den = 0;
michael@0: #ifdef FIXED_POINT
michael@0:       int rshift;
michael@0: #endif
michael@0: #ifdef FIXED_POINT
michael@0:       rshift = 1+celt_ilog2(K-pulsesLeft+i+1);
michael@0: #endif
michael@0:       best_id = 0;
michael@0:       /* The squared magnitude term gets added anyway, so we might as well
michael@0:          add it outside the loop */
michael@0:       yy = ADD32(yy, 1);
michael@0:       j=0;
michael@0:       do {
michael@0:          opus_val16 Rxy, Ryy;
michael@0:          /* Temporary sums of the new pulse(s) */
michael@0:          Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift));
michael@0:          /* We're multiplying y[j] by two so we don't have to do it here */
michael@0:          Ryy = ADD16(yy, y[j]);
michael@0: 
michael@0:          /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
michael@0:             Rxy is positive because the sign is pre-computed) */
michael@0:          Rxy = MULT16_16_Q15(Rxy,Rxy);
michael@0:          /* The idea is to check for num/den >= best_num/best_den, but that way
michael@0:             we can do it without any division */
michael@0:          /* OPT: Make sure to use conditional moves here */
michael@0:          if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))
michael@0:          {
michael@0:             best_den = Ryy;
michael@0:             best_num = Rxy;
michael@0:             best_id = j;
michael@0:          }
michael@0:       } while (++j<N);
michael@0: 
michael@0:       /* Updating the sums of the new pulse(s) */
michael@0:       xy = ADD32(xy, EXTEND32(X[best_id]));
michael@0:       /* We're multiplying y[j] by two so we don't have to do it here */
michael@0:       yy = ADD16(yy, y[best_id]);
michael@0: 
michael@0:       /* Only now that we've made the final choice, update y/iy */
michael@0:       /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
michael@0:       y[best_id] += 2*s;
michael@0:       iy[best_id]++;
michael@0:    }
michael@0: 
michael@0:    /* Put the original sign back */
michael@0:    j=0;
michael@0:    do {
michael@0:       X[j] = MULT16_16(signx[j],X[j]);
michael@0:       if (signx[j] < 0)
michael@0:          iy[j] = -iy[j];
michael@0:    } while (++j<N);
michael@0:    encode_pulses(iy, N, K, enc);
michael@0: 
michael@0: #ifdef RESYNTH
michael@0:    normalise_residual(iy, X, N, yy, gain);
michael@0:    exp_rotation(X, N, -1, B, K, spread);
michael@0: #endif
michael@0: 
michael@0:    collapse_mask = extract_collapse_mask(iy, N, B);
michael@0:    RESTORE_STACK;
michael@0:    return collapse_mask;
michael@0: }
michael@0: 
michael@0: /** Decode pulse vector and combine the result with the pitch vector to produce
michael@0:     the final normalised signal in the current band. */
michael@0: unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
michael@0:       ec_dec *dec, opus_val16 gain)
michael@0: {
michael@0:    int i;
michael@0:    opus_val32 Ryy;
michael@0:    unsigned collapse_mask;
michael@0:    VARDECL(int, iy);
michael@0:    SAVE_STACK;
michael@0: 
michael@0:    celt_assert2(K>0, "alg_unquant() needs at least one pulse");
michael@0:    celt_assert2(N>1, "alg_unquant() needs at least two dimensions");
michael@0:    ALLOC(iy, N, int);
michael@0:    decode_pulses(iy, N, K, dec);
michael@0:    Ryy = 0;
michael@0:    i=0;
michael@0:    do {
michael@0:       Ryy = MAC16_16(Ryy, iy[i], iy[i]);
michael@0:    } while (++i < N);
michael@0:    normalise_residual(iy, X, N, Ryy, gain);
michael@0:    exp_rotation(X, N, -1, B, K, spread);
michael@0:    collapse_mask = extract_collapse_mask(iy, N, B);
michael@0:    RESTORE_STACK;
michael@0:    return collapse_mask;
michael@0: }
michael@0: 
michael@0: void renormalise_vector(celt_norm *X, int N, opus_val16 gain)
michael@0: {
michael@0:    int i;
michael@0: #ifdef FIXED_POINT
michael@0:    int k;
michael@0: #endif
michael@0:    opus_val32 E = EPSILON;
michael@0:    opus_val16 g;
michael@0:    opus_val32 t;
michael@0:    celt_norm *xptr = X;
michael@0:    for (i=0;i<N;i++)
michael@0:    {
michael@0:       E = MAC16_16(E, *xptr, *xptr);
michael@0:       xptr++;
michael@0:    }
michael@0: #ifdef FIXED_POINT
michael@0:    k = celt_ilog2(E)>>1;
michael@0: #endif
michael@0:    t = VSHR32(E, 2*(k-7));
michael@0:    g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
michael@0: 
michael@0:    xptr = X;
michael@0:    for (i=0;i<N;i++)
michael@0:    {
michael@0:       *xptr = EXTRACT16(PSHR32(MULT16_16(g, *xptr), k+1));
michael@0:       xptr++;
michael@0:    }
michael@0:    /*return celt_sqrt(E);*/
michael@0: }
michael@0: 
michael@0: int stereo_itheta(celt_norm *X, celt_norm *Y, int stereo, int N)
michael@0: {
michael@0:    int i;
michael@0:    int itheta;
michael@0:    opus_val16 mid, side;
michael@0:    opus_val32 Emid, Eside;
michael@0: 
michael@0:    Emid = Eside = EPSILON;
michael@0:    if (stereo)
michael@0:    {
michael@0:       for (i=0;i<N;i++)
michael@0:       {
michael@0:          celt_norm m, s;
michael@0:          m = ADD16(SHR16(X[i],1),SHR16(Y[i],1));
michael@0:          s = SUB16(SHR16(X[i],1),SHR16(Y[i],1));
michael@0:          Emid = MAC16_16(Emid, m, m);
michael@0:          Eside = MAC16_16(Eside, s, s);
michael@0:       }
michael@0:    } else {
michael@0:       for (i=0;i<N;i++)
michael@0:       {
michael@0:          celt_norm m, s;
michael@0:          m = X[i];
michael@0:          s = Y[i];
michael@0:          Emid = MAC16_16(Emid, m, m);
michael@0:          Eside = MAC16_16(Eside, s, s);
michael@0:       }
michael@0:    }
michael@0:    mid = celt_sqrt(Emid);
michael@0:    side = celt_sqrt(Eside);
michael@0: #ifdef FIXED_POINT
michael@0:    /* 0.63662 = 2/pi */
michael@0:    itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid));
michael@0: #else
michael@0:    itheta = (int)floor(.5f+16384*0.63662f*atan2(side,mid));
michael@0: #endif
michael@0: 
michael@0:    return itheta;
michael@0: }