[Opus] Update to v1.2.1

third_party/opus/src/celt/vq.c

 /* Copyright (c) 2007-2008 CSIRO
    Copyright (c) 2007-2009 Xiph.Org Foundation
    Written by Jean-Marc Valin */
 /*
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions
    are met:
 
    - Redistributions of source code must retain the above copyright
    notice, this list of conditions and the following disclaimer.
@@ skipping 49 matching lines @@
    {
       celt_norm x1, x2;
       x1 = Xptr[0];
       x2 = Xptr[stride];
       Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
       *Xptr--      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
    }
 }
 #endif /* OVERRIDE_vq_exp_rotation1 */
 
-static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
+void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
 {
    static const int SPREAD_FACTOR[3]={15,10,5};
    int i;
    opus_val16 c, s;
    opus_val16 gain, theta;
    int stride2=0;
    int factor;
 
    if (2*K>=len || spread==SPREAD_NONE)
       return;
@@ skipping 70 matching lines @@
       int j;
       unsigned tmp=0;
       j=0; do {
          tmp |= iy[i*N0+j];
       } while (++j<N0);
       collapse_mask |= (tmp!=0)<<i;
    } while (++i<B);
    return collapse_mask;
 }
 
-unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc
-#ifdef RESYNTH
-   , opus_val16 gain
-#endif
-   )
+opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
 {
    VARDECL(celt_norm, y);
-   VARDECL(int, iy);
-   VARDECL(opus_val16, signx);
+   VARDECL(int, signx);
    int i, j;
-   opus_val16 s;
    int pulsesLeft;
    opus_val32 sum;
    opus_val32 xy;
    opus_val16 yy;
-   unsigned collapse_mask;
    SAVE_STACK;
 
-   celt_assert2(K>0, "alg_quant() needs at least one pulse");
-   celt_assert2(N>1, "alg_quant() needs at least two dimensions");
-
+   (void)arch;
    ALLOC(y, N, celt_norm);
-   ALLOC(iy, N, int);
-   ALLOC(signx, N, opus_val16);
-
-   exp_rotation(X, N, 1, B, K, spread);
+   ALLOC(signx, N, int);
 
    /* Get rid of the sign */
    sum = 0;
    j=0; do {
-      if (X[j]>0)
-         signx[j]=1;
-      else {
-         signx[j]=-1;
-         X[j]=-X[j];
-      }
+      signx[j] = X[j]<0;
+      /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */
+      X[j] = ABS16(X[j]);
       iy[j] = 0;
       y[j] = 0;
    } while (++j<N);
 
    xy = yy = 0;
 
    pulsesLeft = K;
 
    /* Do a pre-search by projecting on the pyramid */
    if (K > (N>>1))
@@ skipping 11 matching lines @@
          to be allocated. 64 is an approximation of infinity here. */
       if (!(sum > EPSILON && sum < 64))
 #endif
       {
          X[0] = QCONST16(1.f,14);
          j=1; do
             X[j]=0;
          while (++j<N);
          sum = QCONST16(1.f,14);
       }
-      rcp = EXTRACT16(MULT16_32_Q16(K-1, celt_rcp(sum)));
+#ifdef FIXED_POINT
+      rcp = EXTRACT16(MULT16_32_Q16(K, celt_rcp(sum)));
+#else
+      /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
+      rcp = EXTRACT16(MULT16_32_Q16(K+0.8f, celt_rcp(sum)));
+#endif
       j=0; do {
 #ifdef FIXED_POINT
          /* It's really important to round *towards zero* here */
          iy[j] = MULT16_16_Q15(X[j],rcp);
 #else
          iy[j] = (int)floor(rcp*X[j]);
 #endif
          y[j] = (celt_norm)iy[j];
          yy = MAC16_16(yy, y[j],y[j]);
          xy = MAC16_16(xy, X[j],y[j]);
          y[j] *= 2;
          pulsesLeft -= iy[j];
       }  while (++j<N);
    }
-   celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass");
+   celt_assert2(pulsesLeft>=0, "Allocated too many pulses in the quick pass");
 
    /* This should never happen, but just in case it does (e.g. on silence)
       we fill the first bin with pulses. */
 #ifdef FIXED_POINT_DEBUG
    celt_assert2(pulsesLeft<=N+3, "Not enough pulses in the quick pass");
 #endif
    if (pulsesLeft > N+3)
    {
       opus_val16 tmp = (opus_val16)pulsesLeft;
       yy = MAC16_16(yy, tmp, tmp);
       yy = MAC16_16(yy, tmp, y[0]);
       iy[0] += pulsesLeft;
       pulsesLeft=0;
    }
 
-   s = 1;
    for (i=0;i<pulsesLeft;i++)
    {
+      opus_val16 Rxy, Ryy;
       int best_id;
-      opus_val32 best_num = -VERY_LARGE16;
-      opus_val16 best_den = 0;
+      opus_val32 best_num;
+      opus_val16 best_den;
 #ifdef FIXED_POINT
       int rshift;
 #endif
 #ifdef FIXED_POINT
       rshift = 1+celt_ilog2(K-pulsesLeft+i+1);
 #endif
       best_id = 0;
       /* The squared magnitude term gets added anyway, so we might as well
          add it outside the loop */
       yy = ADD16(yy, 1);
-      j=0;
+
+      /* Calculations for position 0 are out of the loop, in part to reduce
+         mispredicted branches (since the if condition is usually false)
+         in the loop. */
+      /* Temporary sums of the new pulse(s) */
+      Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[0])),rshift));
+      /* We're multiplying y[j] by two so we don't have to do it here */
+      Ryy = ADD16(yy, y[0]);
+
+      /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
+         Rxy is positive because the sign is pre-computed) */
+      Rxy = MULT16_16_Q15(Rxy,Rxy);
+      best_den = Ryy;
+      best_num = Rxy;
+      j=1;
       do {
-         opus_val16 Rxy, Ryy;
          /* Temporary sums of the new pulse(s) */
          Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift));
          /* We're multiplying y[j] by two so we don't have to do it here */
          Ryy = ADD16(yy, y[j]);
 
          /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
             Rxy is positive because the sign is pre-computed) */
          Rxy = MULT16_16_Q15(Rxy,Rxy);
          /* The idea is to check for num/den >= best_num/best_den, but that way
             we can do it without any division */
-         /* OPT: Make sure to use conditional moves here */
-         if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))
+         /* OPT: It's not clear whether a cmov is faster than a branch here
+            since the condition is more often false than true and using
+            a cmov introduces data dependencies across iterations. The optimal
+            choice may be architecture-dependent. */
+         if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)))
          {
             best_den = Ryy;
             best_num = Rxy;
             best_id = j;
          }
       } while (++j<N);
 
       /* Updating the sums of the new pulse(s) */
       xy = ADD32(xy, EXTEND32(X[best_id]));
       /* We're multiplying y[j] by two so we don't have to do it here */
       yy = ADD16(yy, y[best_id]);
 
       /* Only now that we've made the final choice, update y/iy */
       /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
-      y[best_id] += 2*s;
+      y[best_id] += 2;
       iy[best_id]++;
    }
 
    /* Put the original sign back */
    j=0;
    do {
-      X[j] = MULT16_16(signx[j],X[j]);
-      if (signx[j] < 0)
-         iy[j] = -iy[j];
+      /*iy[j] = signx[j] ? -iy[j] : iy[j];*/
+      /* OPT: The is more likely to be compiled without a branch than the code above
+         but has the same performance otherwise. */
+      iy[j] = (iy[j]^-signx[j]) + signx[j];
    } while (++j<N);
+   RESTORE_STACK;
+   return yy;
+}
+
+unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
+      opus_val16 gain, int resynth, int arch)
+{
+   VARDECL(int, iy);
+   opus_val16 yy;
+   unsigned collapse_mask;
+   SAVE_STACK;
+
+   celt_assert2(K>0, "alg_quant() needs at least one pulse");
+   celt_assert2(N>1, "alg_quant() needs at least two dimensions");
+
+   /* Covers vectorization by up to 4. */
+   ALLOC(iy, N+3, int);
+
+   exp_rotation(X, N, 1, B, K, spread);
+
+   yy = op_pvq_search(X, iy, K, N, arch);
+
    encode_pulses(iy, N, K, enc);
 
-#ifdef RESYNTH
-   normalise_residual(iy, X, N, yy, gain);
-   exp_rotation(X, N, -1, B, K, spread);
-#endif
+   if (resynth)
+   {
+      normalise_residual(iy, X, N, yy, gain);
+      exp_rotation(X, N, -1, B, K, spread);
+   }
 
    collapse_mask = extract_collapse_mask(iy, N, B);
    RESTORE_STACK;
    return collapse_mask;
 }
 
 /** Decode pulse vector and combine the result with the pitch vector to produce
     the final normalised signal in the current band. */
 unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
       ec_dec *dec, opus_val16 gain)
@@ skipping 63 matching lines @@
    } else {
       Emid += celt_inner_prod(X, X, N, arch);
       Eside += celt_inner_prod(Y, Y, N, arch);
    }
    mid = celt_sqrt(Emid);
    side = celt_sqrt(Eside);
 #ifdef FIXED_POINT
    /* 0.63662 = 2/pi */
    itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid));
 #else
-   itheta = (int)floor(.5f+16384*0.63662f*atan2(side,mid));
+   itheta = (int)floor(.5f+16384*0.63662f*fast_atan2f(side,mid));
 #endif
 
    return itheta;
 }