Chris@69: /* Copyright (c) 2007-2008 CSIRO
Chris@69:    Copyright (c) 2007-2009 Xiph.Org Foundation
Chris@69:    Written by Jean-Marc Valin */
Chris@69: /*
Chris@69:    Redistribution and use in source and binary forms, with or without
Chris@69:    modification, are permitted provided that the following conditions
Chris@69:    are met:
Chris@69: 
Chris@69:    - Redistributions of source code must retain the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer.
Chris@69: 
Chris@69:    - Redistributions in binary form must reproduce the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer in the
Chris@69:    documentation and/or other materials provided with the distribution.
Chris@69: 
Chris@69:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
Chris@69:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
Chris@69:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
Chris@69:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
Chris@69:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
Chris@69:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
Chris@69:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
Chris@69:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
Chris@69:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
Chris@69:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
Chris@69:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Chris@69: */
Chris@69: 
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #include "mathops.h"
Chris@69: #include "cwrs.h"
Chris@69: #include "vq.h"
Chris@69: #include "arch.h"
Chris@69: #include "os_support.h"
Chris@69: #include "bands.h"
Chris@69: #include "rate.h"
Chris@69: #include "pitch.h"
Chris@69: 
Chris@69: #ifndef OVERRIDE_vq_exp_rotation1
Chris@69: static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s)
Chris@69: {
Chris@69:    int i;
Chris@69:    opus_val16 ms;
Chris@69:    celt_norm *Xptr;
Chris@69:    Xptr = X;
Chris@69:    ms = NEG16(s);
Chris@69:    for (i=0;i<len-stride;i++)
Chris@69:    {
Chris@69:       celt_norm x1, x2;
Chris@69:       x1 = Xptr[0];
Chris@69:       x2 = Xptr[stride];
Chris@69:       Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
Chris@69:       *Xptr++      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
Chris@69:    }
Chris@69:    Xptr = &X[len-2*stride-1];
Chris@69:    for (i=len-2*stride-1;i>=0;i--)
Chris@69:    {
Chris@69:       celt_norm x1, x2;
Chris@69:       x1 = Xptr[0];
Chris@69:       x2 = Xptr[stride];
Chris@69:       Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
Chris@69:       *Xptr--      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
Chris@69:    }
Chris@69: }
Chris@69: #endif /* OVERRIDE_vq_exp_rotation1 */
Chris@69: 
Chris@69: void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
Chris@69: {
Chris@69:    static const int SPREAD_FACTOR[3]={15,10,5};
Chris@69:    int i;
Chris@69:    opus_val16 c, s;
Chris@69:    opus_val16 gain, theta;
Chris@69:    int stride2=0;
Chris@69:    int factor;
Chris@69: 
Chris@69:    if (2*K>=len || spread==SPREAD_NONE)
Chris@69:       return;
Chris@69:    factor = SPREAD_FACTOR[spread-1];
Chris@69: 
Chris@69:    gain = celt_div((opus_val32)MULT16_16(Q15_ONE,len),(opus_val32)(len+factor*K));
Chris@69:    theta = HALF16(MULT16_16_Q15(gain,gain));
Chris@69: 
Chris@69:    c = celt_cos_norm(EXTEND32(theta));
Chris@69:    s = celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /*  sin(theta) */
Chris@69: 
Chris@69:    if (len>=8*stride)
Chris@69:    {
Chris@69:       stride2 = 1;
Chris@69:       /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
Chris@69:          It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
Chris@69:       while ((stride2*stride2+stride2)*stride + (stride>>2) < len)
Chris@69:          stride2++;
Chris@69:    }
Chris@69:    /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
Chris@69:       extract_collapse_mask().*/
Chris@69:    len = celt_udiv(len, stride);
Chris@69:    for (i=0;i<stride;i++)
Chris@69:    {
Chris@69:       if (dir < 0)
Chris@69:       {
Chris@69:          if (stride2)
Chris@69:             exp_rotation1(X+i*len, len, stride2, s, c);
Chris@69:          exp_rotation1(X+i*len, len, 1, c, s);
Chris@69:       } else {
Chris@69:          exp_rotation1(X+i*len, len, 1, c, -s);
Chris@69:          if (stride2)
Chris@69:             exp_rotation1(X+i*len, len, stride2, s, -c);
Chris@69:       }
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: /** Takes the pitch vector and the decoded residual vector, computes the gain
Chris@69:     that will give ||p+g*y||=1 and mixes the residual with the pitch. */
Chris@69: static void normalise_residual(int * OPUS_RESTRICT iy, celt_norm * OPUS_RESTRICT X,
Chris@69:       int N, opus_val32 Ryy, opus_val16 gain)
Chris@69: {
Chris@69:    int i;
Chris@69: #ifdef FIXED_POINT
Chris@69:    int k;
Chris@69: #endif
Chris@69:    opus_val32 t;
Chris@69:    opus_val16 g;
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69:    k = celt_ilog2(Ryy)>>1;
Chris@69: #endif
Chris@69:    t = VSHR32(Ryy, 2*(k-7));
Chris@69:    g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
Chris@69: 
Chris@69:    i=0;
Chris@69:    do
Chris@69:       X[i] = EXTRACT16(PSHR32(MULT16_16(g, iy[i]), k+1));
Chris@69:    while (++i < N);
Chris@69: }
Chris@69: 
Chris@69: static unsigned extract_collapse_mask(int *iy, int N, int B)
Chris@69: {
Chris@69:    unsigned collapse_mask;
Chris@69:    int N0;
Chris@69:    int i;
Chris@69:    if (B<=1)
Chris@69:       return 1;
Chris@69:    /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
Chris@69:       exp_rotation().*/
Chris@69:    N0 = celt_udiv(N, B);
Chris@69:    collapse_mask = 0;
Chris@69:    i=0; do {
Chris@69:       int j;
Chris@69:       unsigned tmp=0;
Chris@69:       j=0; do {
Chris@69:          tmp |= iy[i*N0+j];
Chris@69:       } while (++j<N0);
Chris@69:       collapse_mask |= (tmp!=0)<<i;
Chris@69:    } while (++i<B);
Chris@69:    return collapse_mask;
Chris@69: }
Chris@69: 
Chris@69: opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
Chris@69: {
Chris@69:    VARDECL(celt_norm, y);
Chris@69:    VARDECL(int, signx);
Chris@69:    int i, j;
Chris@69:    int pulsesLeft;
Chris@69:    opus_val32 sum;
Chris@69:    opus_val32 xy;
Chris@69:    opus_val16 yy;
Chris@69:    SAVE_STACK;
Chris@69: 
Chris@69:    (void)arch;
Chris@69:    ALLOC(y, N, celt_norm);
Chris@69:    ALLOC(signx, N, int);
Chris@69: 
Chris@69:    /* Get rid of the sign */
Chris@69:    sum = 0;
Chris@69:    j=0; do {
Chris@69:       signx[j] = X[j]<0;
Chris@69:       /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */
Chris@69:       X[j] = ABS16(X[j]);
Chris@69:       iy[j] = 0;
Chris@69:       y[j] = 0;
Chris@69:    } while (++j<N);
Chris@69: 
Chris@69:    xy = yy = 0;
Chris@69: 
Chris@69:    pulsesLeft = K;
Chris@69: 
Chris@69:    /* Do a pre-search by projecting on the pyramid */
Chris@69:    if (K > (N>>1))
Chris@69:    {
Chris@69:       opus_val16 rcp;
Chris@69:       j=0; do {
Chris@69:          sum += X[j];
Chris@69:       }  while (++j<N);
Chris@69: 
Chris@69:       /* If X is too small, just replace it with a pulse at 0 */
Chris@69: #ifdef FIXED_POINT
Chris@69:       if (sum <= K)
Chris@69: #else
Chris@69:       /* Prevents infinities and NaNs from causing too many pulses
Chris@69:          to be allocated. 64 is an approximation of infinity here. */
Chris@69:       if (!(sum > EPSILON && sum < 64))
Chris@69: #endif
Chris@69:       {
Chris@69:          X[0] = QCONST16(1.f,14);
Chris@69:          j=1; do
Chris@69:             X[j]=0;
Chris@69:          while (++j<N);
Chris@69:          sum = QCONST16(1.f,14);
Chris@69:       }
Chris@69: #ifdef FIXED_POINT
Chris@69:       rcp = EXTRACT16(MULT16_32_Q16(K, celt_rcp(sum)));
Chris@69: #else
Chris@69:       /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
Chris@69:       rcp = EXTRACT16(MULT16_32_Q16(K+0.8f, celt_rcp(sum)));
Chris@69: #endif
Chris@69:       j=0; do {
Chris@69: #ifdef FIXED_POINT
Chris@69:          /* It's really important to round *towards zero* here */
Chris@69:          iy[j] = MULT16_16_Q15(X[j],rcp);
Chris@69: #else
Chris@69:          iy[j] = (int)floor(rcp*X[j]);
Chris@69: #endif
Chris@69:          y[j] = (celt_norm)iy[j];
Chris@69:          yy = MAC16_16(yy, y[j],y[j]);
Chris@69:          xy = MAC16_16(xy, X[j],y[j]);
Chris@69:          y[j] *= 2;
Chris@69:          pulsesLeft -= iy[j];
Chris@69:       }  while (++j<N);
Chris@69:    }
Chris@69:    celt_sig_assert(pulsesLeft>=0);
Chris@69: 
Chris@69:    /* This should never happen, but just in case it does (e.g. on silence)
Chris@69:       we fill the first bin with pulses. */
Chris@69: #ifdef FIXED_POINT_DEBUG
Chris@69:    celt_sig_assert(pulsesLeft<=N+3);
Chris@69: #endif
Chris@69:    if (pulsesLeft > N+3)
Chris@69:    {
Chris@69:       opus_val16 tmp = (opus_val16)pulsesLeft;
Chris@69:       yy = MAC16_16(yy, tmp, tmp);
Chris@69:       yy = MAC16_16(yy, tmp, y[0]);
Chris@69:       iy[0] += pulsesLeft;
Chris@69:       pulsesLeft=0;
Chris@69:    }
Chris@69: 
Chris@69:    for (i=0;i<pulsesLeft;i++)
Chris@69:    {
Chris@69:       opus_val16 Rxy, Ryy;
Chris@69:       int best_id;
Chris@69:       opus_val32 best_num;
Chris@69:       opus_val16 best_den;
Chris@69: #ifdef FIXED_POINT
Chris@69:       int rshift;
Chris@69: #endif
Chris@69: #ifdef FIXED_POINT
Chris@69:       rshift = 1+celt_ilog2(K-pulsesLeft+i+1);
Chris@69: #endif
Chris@69:       best_id = 0;
Chris@69:       /* The squared magnitude term gets added anyway, so we might as well
Chris@69:          add it outside the loop */
Chris@69:       yy = ADD16(yy, 1);
Chris@69: 
Chris@69:       /* Calculations for position 0 are out of the loop, in part to reduce
Chris@69:          mispredicted branches (since the if condition is usually false)
Chris@69:          in the loop. */
Chris@69:       /* Temporary sums of the new pulse(s) */
Chris@69:       Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[0])),rshift));
Chris@69:       /* We're multiplying y[j] by two so we don't have to do it here */
Chris@69:       Ryy = ADD16(yy, y[0]);
Chris@69: 
Chris@69:       /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
Chris@69:          Rxy is positive because the sign is pre-computed) */
Chris@69:       Rxy = MULT16_16_Q15(Rxy,Rxy);
Chris@69:       best_den = Ryy;
Chris@69:       best_num = Rxy;
Chris@69:       j=1;
Chris@69:       do {
Chris@69:          /* Temporary sums of the new pulse(s) */
Chris@69:          Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift));
Chris@69:          /* We're multiplying y[j] by two so we don't have to do it here */
Chris@69:          Ryy = ADD16(yy, y[j]);
Chris@69: 
Chris@69:          /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
Chris@69:             Rxy is positive because the sign is pre-computed) */
Chris@69:          Rxy = MULT16_16_Q15(Rxy,Rxy);
Chris@69:          /* The idea is to check for num/den >= best_num/best_den, but that way
Chris@69:             we can do it without any division */
Chris@69:          /* OPT: It's not clear whether a cmov is faster than a branch here
Chris@69:             since the condition is more often false than true and using
Chris@69:             a cmov introduces data dependencies across iterations. The optimal
Chris@69:             choice may be architecture-dependent. */
Chris@69:          if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)))
Chris@69:          {
Chris@69:             best_den = Ryy;
Chris@69:             best_num = Rxy;
Chris@69:             best_id = j;
Chris@69:          }
Chris@69:       } while (++j<N);
Chris@69: 
Chris@69:       /* Updating the sums of the new pulse(s) */
Chris@69:       xy = ADD32(xy, EXTEND32(X[best_id]));
Chris@69:       /* We're multiplying y[j] by two so we don't have to do it here */
Chris@69:       yy = ADD16(yy, y[best_id]);
Chris@69: 
Chris@69:       /* Only now that we've made the final choice, update y/iy */
Chris@69:       /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
Chris@69:       y[best_id] += 2;
Chris@69:       iy[best_id]++;
Chris@69:    }
Chris@69: 
Chris@69:    /* Put the original sign back */
Chris@69:    j=0;
Chris@69:    do {
Chris@69:       /*iy[j] = signx[j] ? -iy[j] : iy[j];*/
Chris@69:       /* OPT: The is more likely to be compiled without a branch than the code above
Chris@69:          but has the same performance otherwise. */
Chris@69:       iy[j] = (iy[j]^-signx[j]) + signx[j];
Chris@69:    } while (++j<N);
Chris@69:    RESTORE_STACK;
Chris@69:    return yy;
Chris@69: }
Chris@69: 
Chris@69: unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
Chris@69:       opus_val16 gain, int resynth, int arch)
Chris@69: {
Chris@69:    VARDECL(int, iy);
Chris@69:    opus_val16 yy;
Chris@69:    unsigned collapse_mask;
Chris@69:    SAVE_STACK;
Chris@69: 
Chris@69:    celt_assert2(K>0, "alg_quant() needs at least one pulse");
Chris@69:    celt_assert2(N>1, "alg_quant() needs at least two dimensions");
Chris@69: 
Chris@69:    /* Covers vectorization by up to 4. */
Chris@69:    ALLOC(iy, N+3, int);
Chris@69: 
Chris@69:    exp_rotation(X, N, 1, B, K, spread);
Chris@69: 
Chris@69:    yy = op_pvq_search(X, iy, K, N, arch);
Chris@69: 
Chris@69:    encode_pulses(iy, N, K, enc);
Chris@69: 
Chris@69:    if (resynth)
Chris@69:    {
Chris@69:       normalise_residual(iy, X, N, yy, gain);
Chris@69:       exp_rotation(X, N, -1, B, K, spread);
Chris@69:    }
Chris@69: 
Chris@69:    collapse_mask = extract_collapse_mask(iy, N, B);
Chris@69:    RESTORE_STACK;
Chris@69:    return collapse_mask;
Chris@69: }
Chris@69: 
Chris@69: /** Decode pulse vector and combine the result with the pitch vector to produce
Chris@69:     the final normalised signal in the current band. */
Chris@69: unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
Chris@69:       ec_dec *dec, opus_val16 gain)
Chris@69: {
Chris@69:    opus_val32 Ryy;
Chris@69:    unsigned collapse_mask;
Chris@69:    VARDECL(int, iy);
Chris@69:    SAVE_STACK;
Chris@69: 
Chris@69:    celt_assert2(K>0, "alg_unquant() needs at least one pulse");
Chris@69:    celt_assert2(N>1, "alg_unquant() needs at least two dimensions");
Chris@69:    ALLOC(iy, N, int);
Chris@69:    Ryy = decode_pulses(iy, N, K, dec);
Chris@69:    normalise_residual(iy, X, N, Ryy, gain);
Chris@69:    exp_rotation(X, N, -1, B, K, spread);
Chris@69:    collapse_mask = extract_collapse_mask(iy, N, B);
Chris@69:    RESTORE_STACK;
Chris@69:    return collapse_mask;
Chris@69: }
Chris@69: 
Chris@69: #ifndef OVERRIDE_renormalise_vector
Chris@69: void renormalise_vector(celt_norm *X, int N, opus_val16 gain, int arch)
Chris@69: {
Chris@69:    int i;
Chris@69: #ifdef FIXED_POINT
Chris@69:    int k;
Chris@69: #endif
Chris@69:    opus_val32 E;
Chris@69:    opus_val16 g;
Chris@69:    opus_val32 t;
Chris@69:    celt_norm *xptr;
Chris@69:    E = EPSILON + celt_inner_prod(X, X, N, arch);
Chris@69: #ifdef FIXED_POINT
Chris@69:    k = celt_ilog2(E)>>1;
Chris@69: #endif
Chris@69:    t = VSHR32(E, 2*(k-7));
Chris@69:    g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
Chris@69: 
Chris@69:    xptr = X;
Chris@69:    for (i=0;i<N;i++)
Chris@69:    {
Chris@69:       *xptr = EXTRACT16(PSHR32(MULT16_16(g, *xptr), k+1));
Chris@69:       xptr++;
Chris@69:    }
Chris@69:    /*return celt_sqrt(E);*/
Chris@69: }
Chris@69: #endif /* OVERRIDE_renormalise_vector */
Chris@69: 
Chris@69: int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int arch)
Chris@69: {
Chris@69:    int i;
Chris@69:    int itheta;
Chris@69:    opus_val16 mid, side;
Chris@69:    opus_val32 Emid, Eside;
Chris@69: 
Chris@69:    Emid = Eside = EPSILON;
Chris@69:    if (stereo)
Chris@69:    {
Chris@69:       for (i=0;i<N;i++)
Chris@69:       {
Chris@69:          celt_norm m, s;
Chris@69:          m = ADD16(SHR16(X[i],1),SHR16(Y[i],1));
Chris@69:          s = SUB16(SHR16(X[i],1),SHR16(Y[i],1));
Chris@69:          Emid = MAC16_16(Emid, m, m);
Chris@69:          Eside = MAC16_16(Eside, s, s);
Chris@69:       }
Chris@69:    } else {
Chris@69:       Emid += celt_inner_prod(X, X, N, arch);
Chris@69:       Eside += celt_inner_prod(Y, Y, N, arch);
Chris@69:    }
Chris@69:    mid = celt_sqrt(Emid);
Chris@69:    side = celt_sqrt(Eside);
Chris@69: #ifdef FIXED_POINT
Chris@69:    /* 0.63662 = 2/pi */
Chris@69:    itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid));
Chris@69: #else
Chris@69:    itheta = (int)floor(.5f+16384*0.63662f*fast_atan2f(side,mid));
Chris@69: #endif
Chris@69: 
Chris@69:    return itheta;
Chris@69: }