cannam@154: /* Copyright (c) 2007-2008 CSIRO
cannam@154:    Copyright (c) 2007-2009 Xiph.Org Foundation
cannam@154:    Written by Jean-Marc Valin */
cannam@154: /*
cannam@154:    Redistribution and use in source and binary forms, with or without
cannam@154:    modification, are permitted provided that the following conditions
cannam@154:    are met:
cannam@154: 
cannam@154:    - Redistributions of source code must retain the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer.
cannam@154: 
cannam@154:    - Redistributions in binary form must reproduce the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer in the
cannam@154:    documentation and/or other materials provided with the distribution.
cannam@154: 
cannam@154:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
cannam@154:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
cannam@154:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
cannam@154:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
cannam@154:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
cannam@154:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
cannam@154:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
cannam@154:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
cannam@154:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
cannam@154:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
cannam@154:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
cannam@154: */
cannam@154: 
cannam@154: #ifdef HAVE_CONFIG_H
cannam@154: #include "config.h"
cannam@154: #endif
cannam@154: 
cannam@154: #include "mathops.h"
cannam@154: #include "cwrs.h"
cannam@154: #include "vq.h"
cannam@154: #include "arch.h"
cannam@154: #include "os_support.h"
cannam@154: #include "bands.h"
cannam@154: #include "rate.h"
cannam@154: #include "pitch.h"
cannam@154: 
cannam@154: #ifndef OVERRIDE_vq_exp_rotation1
cannam@154: static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s)
cannam@154: {
cannam@154:    int i;
cannam@154:    opus_val16 ms;
cannam@154:    celt_norm *Xptr;
cannam@154:    Xptr = X;
cannam@154:    ms = NEG16(s);
cannam@154:    for (i=0;i<len-stride;i++)
cannam@154:    {
cannam@154:       celt_norm x1, x2;
cannam@154:       x1 = Xptr[0];
cannam@154:       x2 = Xptr[stride];
cannam@154:       Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
cannam@154:       *Xptr++      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
cannam@154:    }
cannam@154:    Xptr = &X[len-2*stride-1];
cannam@154:    for (i=len-2*stride-1;i>=0;i--)
cannam@154:    {
cannam@154:       celt_norm x1, x2;
cannam@154:       x1 = Xptr[0];
cannam@154:       x2 = Xptr[stride];
cannam@154:       Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
cannam@154:       *Xptr--      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
cannam@154:    }
cannam@154: }
cannam@154: #endif /* OVERRIDE_vq_exp_rotation1 */
cannam@154: 
cannam@154: void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
cannam@154: {
cannam@154:    static const int SPREAD_FACTOR[3]={15,10,5};
cannam@154:    int i;
cannam@154:    opus_val16 c, s;
cannam@154:    opus_val16 gain, theta;
cannam@154:    int stride2=0;
cannam@154:    int factor;
cannam@154: 
cannam@154:    if (2*K>=len || spread==SPREAD_NONE)
cannam@154:       return;
cannam@154:    factor = SPREAD_FACTOR[spread-1];
cannam@154: 
cannam@154:    gain = celt_div((opus_val32)MULT16_16(Q15_ONE,len),(opus_val32)(len+factor*K));
cannam@154:    theta = HALF16(MULT16_16_Q15(gain,gain));
cannam@154: 
cannam@154:    c = celt_cos_norm(EXTEND32(theta));
cannam@154:    s = celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /*  sin(theta) */
cannam@154: 
cannam@154:    if (len>=8*stride)
cannam@154:    {
cannam@154:       stride2 = 1;
cannam@154:       /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
cannam@154:          It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
cannam@154:       while ((stride2*stride2+stride2)*stride + (stride>>2) < len)
cannam@154:          stride2++;
cannam@154:    }
cannam@154:    /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
cannam@154:       extract_collapse_mask().*/
cannam@154:    len = celt_udiv(len, stride);
cannam@154:    for (i=0;i<stride;i++)
cannam@154:    {
cannam@154:       if (dir < 0)
cannam@154:       {
cannam@154:          if (stride2)
cannam@154:             exp_rotation1(X+i*len, len, stride2, s, c);
cannam@154:          exp_rotation1(X+i*len, len, 1, c, s);
cannam@154:       } else {
cannam@154:          exp_rotation1(X+i*len, len, 1, c, -s);
cannam@154:          if (stride2)
cannam@154:             exp_rotation1(X+i*len, len, stride2, s, -c);
cannam@154:       }
cannam@154:    }
cannam@154: }
cannam@154: 
cannam@154: /** Takes the pitch vector and the decoded residual vector, computes the gain
cannam@154:     that will give ||p+g*y||=1 and mixes the residual with the pitch. */
cannam@154: static void normalise_residual(int * OPUS_RESTRICT iy, celt_norm * OPUS_RESTRICT X,
cannam@154:       int N, opus_val32 Ryy, opus_val16 gain)
cannam@154: {
cannam@154:    int i;
cannam@154: #ifdef FIXED_POINT
cannam@154:    int k;
cannam@154: #endif
cannam@154:    opus_val32 t;
cannam@154:    opus_val16 g;
cannam@154: 
cannam@154: #ifdef FIXED_POINT
cannam@154:    k = celt_ilog2(Ryy)>>1;
cannam@154: #endif
cannam@154:    t = VSHR32(Ryy, 2*(k-7));
cannam@154:    g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
cannam@154: 
cannam@154:    i=0;
cannam@154:    do
cannam@154:       X[i] = EXTRACT16(PSHR32(MULT16_16(g, iy[i]), k+1));
cannam@154:    while (++i < N);
cannam@154: }
cannam@154: 
cannam@154: static unsigned extract_collapse_mask(int *iy, int N, int B)
cannam@154: {
cannam@154:    unsigned collapse_mask;
cannam@154:    int N0;
cannam@154:    int i;
cannam@154:    if (B<=1)
cannam@154:       return 1;
cannam@154:    /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
cannam@154:       exp_rotation().*/
cannam@154:    N0 = celt_udiv(N, B);
cannam@154:    collapse_mask = 0;
cannam@154:    i=0; do {
cannam@154:       int j;
cannam@154:       unsigned tmp=0;
cannam@154:       j=0; do {
cannam@154:          tmp |= iy[i*N0+j];
cannam@154:       } while (++j<N0);
cannam@154:       collapse_mask |= (tmp!=0)<<i;
cannam@154:    } while (++i<B);
cannam@154:    return collapse_mask;
cannam@154: }
cannam@154: 
cannam@154: opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
cannam@154: {
cannam@154:    VARDECL(celt_norm, y);
cannam@154:    VARDECL(int, signx);
cannam@154:    int i, j;
cannam@154:    int pulsesLeft;
cannam@154:    opus_val32 sum;
cannam@154:    opus_val32 xy;
cannam@154:    opus_val16 yy;
cannam@154:    SAVE_STACK;
cannam@154: 
cannam@154:    (void)arch;
cannam@154:    ALLOC(y, N, celt_norm);
cannam@154:    ALLOC(signx, N, int);
cannam@154: 
cannam@154:    /* Get rid of the sign */
cannam@154:    sum = 0;
cannam@154:    j=0; do {
cannam@154:       signx[j] = X[j]<0;
cannam@154:       /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */
cannam@154:       X[j] = ABS16(X[j]);
cannam@154:       iy[j] = 0;
cannam@154:       y[j] = 0;
cannam@154:    } while (++j<N);
cannam@154: 
cannam@154:    xy = yy = 0;
cannam@154: 
cannam@154:    pulsesLeft = K;
cannam@154: 
cannam@154:    /* Do a pre-search by projecting on the pyramid */
cannam@154:    if (K > (N>>1))
cannam@154:    {
cannam@154:       opus_val16 rcp;
cannam@154:       j=0; do {
cannam@154:          sum += X[j];
cannam@154:       }  while (++j<N);
cannam@154: 
cannam@154:       /* If X is too small, just replace it with a pulse at 0 */
cannam@154: #ifdef FIXED_POINT
cannam@154:       if (sum <= K)
cannam@154: #else
cannam@154:       /* Prevents infinities and NaNs from causing too many pulses
cannam@154:          to be allocated. 64 is an approximation of infinity here. */
cannam@154:       if (!(sum > EPSILON && sum < 64))
cannam@154: #endif
cannam@154:       {
cannam@154:          X[0] = QCONST16(1.f,14);
cannam@154:          j=1; do
cannam@154:             X[j]=0;
cannam@154:          while (++j<N);
cannam@154:          sum = QCONST16(1.f,14);
cannam@154:       }
cannam@154: #ifdef FIXED_POINT
cannam@154:       rcp = EXTRACT16(MULT16_32_Q16(K, celt_rcp(sum)));
cannam@154: #else
cannam@154:       /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
cannam@154:       rcp = EXTRACT16(MULT16_32_Q16(K+0.8f, celt_rcp(sum)));
cannam@154: #endif
cannam@154:       j=0; do {
cannam@154: #ifdef FIXED_POINT
cannam@154:          /* It's really important to round *towards zero* here */
cannam@154:          iy[j] = MULT16_16_Q15(X[j],rcp);
cannam@154: #else
cannam@154:          iy[j] = (int)floor(rcp*X[j]);
cannam@154: #endif
cannam@154:          y[j] = (celt_norm)iy[j];
cannam@154:          yy = MAC16_16(yy, y[j],y[j]);
cannam@154:          xy = MAC16_16(xy, X[j],y[j]);
cannam@154:          y[j] *= 2;
cannam@154:          pulsesLeft -= iy[j];
cannam@154:       }  while (++j<N);
cannam@154:    }
cannam@154:    celt_sig_assert(pulsesLeft>=0);
cannam@154: 
cannam@154:    /* This should never happen, but just in case it does (e.g. on silence)
cannam@154:       we fill the first bin with pulses. */
cannam@154: #ifdef FIXED_POINT_DEBUG
cannam@154:    celt_sig_assert(pulsesLeft<=N+3);
cannam@154: #endif
cannam@154:    if (pulsesLeft > N+3)
cannam@154:    {
cannam@154:       opus_val16 tmp = (opus_val16)pulsesLeft;
cannam@154:       yy = MAC16_16(yy, tmp, tmp);
cannam@154:       yy = MAC16_16(yy, tmp, y[0]);
cannam@154:       iy[0] += pulsesLeft;
cannam@154:       pulsesLeft=0;
cannam@154:    }
cannam@154: 
cannam@154:    for (i=0;i<pulsesLeft;i++)
cannam@154:    {
cannam@154:       opus_val16 Rxy, Ryy;
cannam@154:       int best_id;
cannam@154:       opus_val32 best_num;
cannam@154:       opus_val16 best_den;
cannam@154: #ifdef FIXED_POINT
cannam@154:       int rshift;
cannam@154: #endif
cannam@154: #ifdef FIXED_POINT
cannam@154:       rshift = 1+celt_ilog2(K-pulsesLeft+i+1);
cannam@154: #endif
cannam@154:       best_id = 0;
cannam@154:       /* The squared magnitude term gets added anyway, so we might as well
cannam@154:          add it outside the loop */
cannam@154:       yy = ADD16(yy, 1);
cannam@154: 
cannam@154:       /* Calculations for position 0 are out of the loop, in part to reduce
cannam@154:          mispredicted branches (since the if condition is usually false)
cannam@154:          in the loop. */
cannam@154:       /* Temporary sums of the new pulse(s) */
cannam@154:       Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[0])),rshift));
cannam@154:       /* We're multiplying y[j] by two so we don't have to do it here */
cannam@154:       Ryy = ADD16(yy, y[0]);
cannam@154: 
cannam@154:       /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
cannam@154:          Rxy is positive because the sign is pre-computed) */
cannam@154:       Rxy = MULT16_16_Q15(Rxy,Rxy);
cannam@154:       best_den = Ryy;
cannam@154:       best_num = Rxy;
cannam@154:       j=1;
cannam@154:       do {
cannam@154:          /* Temporary sums of the new pulse(s) */
cannam@154:          Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift));
cannam@154:          /* We're multiplying y[j] by two so we don't have to do it here */
cannam@154:          Ryy = ADD16(yy, y[j]);
cannam@154: 
cannam@154:          /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
cannam@154:             Rxy is positive because the sign is pre-computed) */
cannam@154:          Rxy = MULT16_16_Q15(Rxy,Rxy);
cannam@154:          /* The idea is to check for num/den >= best_num/best_den, but that way
cannam@154:             we can do it without any division */
cannam@154:          /* OPT: It's not clear whether a cmov is faster than a branch here
cannam@154:             since the condition is more often false than true and using
cannam@154:             a cmov introduces data dependencies across iterations. The optimal
cannam@154:             choice may be architecture-dependent. */
cannam@154:          if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)))
cannam@154:          {
cannam@154:             best_den = Ryy;
cannam@154:             best_num = Rxy;
cannam@154:             best_id = j;
cannam@154:          }
cannam@154:       } while (++j<N);
cannam@154: 
cannam@154:       /* Updating the sums of the new pulse(s) */
cannam@154:       xy = ADD32(xy, EXTEND32(X[best_id]));
cannam@154:       /* We're multiplying y[j] by two so we don't have to do it here */
cannam@154:       yy = ADD16(yy, y[best_id]);
cannam@154: 
cannam@154:       /* Only now that we've made the final choice, update y/iy */
cannam@154:       /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
cannam@154:       y[best_id] += 2;
cannam@154:       iy[best_id]++;
cannam@154:    }
cannam@154: 
cannam@154:    /* Put the original sign back */
cannam@154:    j=0;
cannam@154:    do {
cannam@154:       /*iy[j] = signx[j] ? -iy[j] : iy[j];*/
cannam@154:       /* OPT: The is more likely to be compiled without a branch than the code above
cannam@154:          but has the same performance otherwise. */
cannam@154:       iy[j] = (iy[j]^-signx[j]) + signx[j];
cannam@154:    } while (++j<N);
cannam@154:    RESTORE_STACK;
cannam@154:    return yy;
cannam@154: }
cannam@154: 
cannam@154: unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
cannam@154:       opus_val16 gain, int resynth, int arch)
cannam@154: {
cannam@154:    VARDECL(int, iy);
cannam@154:    opus_val16 yy;
cannam@154:    unsigned collapse_mask;
cannam@154:    SAVE_STACK;
cannam@154: 
cannam@154:    celt_assert2(K>0, "alg_quant() needs at least one pulse");
cannam@154:    celt_assert2(N>1, "alg_quant() needs at least two dimensions");
cannam@154: 
cannam@154:    /* Covers vectorization by up to 4. */
cannam@154:    ALLOC(iy, N+3, int);
cannam@154: 
cannam@154:    exp_rotation(X, N, 1, B, K, spread);
cannam@154: 
cannam@154:    yy = op_pvq_search(X, iy, K, N, arch);
cannam@154: 
cannam@154:    encode_pulses(iy, N, K, enc);
cannam@154: 
cannam@154:    if (resynth)
cannam@154:    {
cannam@154:       normalise_residual(iy, X, N, yy, gain);
cannam@154:       exp_rotation(X, N, -1, B, K, spread);
cannam@154:    }
cannam@154: 
cannam@154:    collapse_mask = extract_collapse_mask(iy, N, B);
cannam@154:    RESTORE_STACK;
cannam@154:    return collapse_mask;
cannam@154: }
cannam@154: 
cannam@154: /** Decode pulse vector and combine the result with the pitch vector to produce
cannam@154:     the final normalised signal in the current band. */
cannam@154: unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
cannam@154:       ec_dec *dec, opus_val16 gain)
cannam@154: {
cannam@154:    opus_val32 Ryy;
cannam@154:    unsigned collapse_mask;
cannam@154:    VARDECL(int, iy);
cannam@154:    SAVE_STACK;
cannam@154: 
cannam@154:    celt_assert2(K>0, "alg_unquant() needs at least one pulse");
cannam@154:    celt_assert2(N>1, "alg_unquant() needs at least two dimensions");
cannam@154:    ALLOC(iy, N, int);
cannam@154:    Ryy = decode_pulses(iy, N, K, dec);
cannam@154:    normalise_residual(iy, X, N, Ryy, gain);
cannam@154:    exp_rotation(X, N, -1, B, K, spread);
cannam@154:    collapse_mask = extract_collapse_mask(iy, N, B);
cannam@154:    RESTORE_STACK;
cannam@154:    return collapse_mask;
cannam@154: }
cannam@154: 
cannam@154: #ifndef OVERRIDE_renormalise_vector
cannam@154: void renormalise_vector(celt_norm *X, int N, opus_val16 gain, int arch)
cannam@154: {
cannam@154:    int i;
cannam@154: #ifdef FIXED_POINT
cannam@154:    int k;
cannam@154: #endif
cannam@154:    opus_val32 E;
cannam@154:    opus_val16 g;
cannam@154:    opus_val32 t;
cannam@154:    celt_norm *xptr;
cannam@154:    E = EPSILON + celt_inner_prod(X, X, N, arch);
cannam@154: #ifdef FIXED_POINT
cannam@154:    k = celt_ilog2(E)>>1;
cannam@154: #endif
cannam@154:    t = VSHR32(E, 2*(k-7));
cannam@154:    g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
cannam@154: 
cannam@154:    xptr = X;
cannam@154:    for (i=0;i<N;i++)
cannam@154:    {
cannam@154:       *xptr = EXTRACT16(PSHR32(MULT16_16(g, *xptr), k+1));
cannam@154:       xptr++;
cannam@154:    }
cannam@154:    /*return celt_sqrt(E);*/
cannam@154: }
cannam@154: #endif /* OVERRIDE_renormalise_vector */
cannam@154: 
cannam@154: int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int arch)
cannam@154: {
cannam@154:    int i;
cannam@154:    int itheta;
cannam@154:    opus_val16 mid, side;
cannam@154:    opus_val32 Emid, Eside;
cannam@154: 
cannam@154:    Emid = Eside = EPSILON;
cannam@154:    if (stereo)
cannam@154:    {
cannam@154:       for (i=0;i<N;i++)
cannam@154:       {
cannam@154:          celt_norm m, s;
cannam@154:          m = ADD16(SHR16(X[i],1),SHR16(Y[i],1));
cannam@154:          s = SUB16(SHR16(X[i],1),SHR16(Y[i],1));
cannam@154:          Emid = MAC16_16(Emid, m, m);
cannam@154:          Eside = MAC16_16(Eside, s, s);
cannam@154:       }
cannam@154:    } else {
cannam@154:       Emid += celt_inner_prod(X, X, N, arch);
cannam@154:       Eside += celt_inner_prod(Y, Y, N, arch);
cannam@154:    }
cannam@154:    mid = celt_sqrt(Emid);
cannam@154:    side = celt_sqrt(Eside);
cannam@154: #ifdef FIXED_POINT
cannam@154:    /* 0.63662 = 2/pi */
cannam@154:    itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid));
cannam@154: #else
cannam@154:    itheta = (int)floor(.5f+16384*0.63662f*fast_atan2f(side,mid));
cannam@154: #endif
cannam@154: 
cannam@154:    return itheta;
cannam@154: }