Update to opus-HEAD-66611f1.

celt/pitch.c

 /* Copyright (c) 2007-2008 CSIRO
    Copyright (c) 2007-2009 Xiph.Org Foundation
    Written by Jean-Marc Valin */
 /**
    @file pitch.c
    @brief Pitch analysis
  */
 
 /*
    Redistribution and use in source and binary forms, with or without
@@ skipping 196 matching lines @@
    }
    /* Add a zero */
    lpc2[0] = lpc[0] + QCONST16(.8f,SIG_SHIFT);
    lpc2[1] = lpc[1] + MULT16_16_Q15(c1,lpc[0]);
    lpc2[2] = lpc[2] + MULT16_16_Q15(c1,lpc[1]);
    lpc2[3] = lpc[3] + MULT16_16_Q15(c1,lpc[2]);
    lpc2[4] = MULT16_16_Q15(c1,lpc[3]);
    celt_fir5(x_lp, lpc2, x_lp, len>>1, mem);
 }
 
-#if 0 /* This is a simple version of the pitch correlation that should work
-         well on DSPs like Blackfin and TI C5x/C6x */
-
+/* Pure C implementation. */
 #ifdef FIXED_POINT
 opus_val32
 #else
 void
 #endif
-celt_pitch_xcorr(opus_val16 *x, opus_val16 *y, opus_val32 *xcorr, int len, int max_pitch)
+#if defined(OVERRIDE_PITCH_XCORR)
+celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y,
+      opus_val32 *xcorr, int len, int max_pitch)
+#else
+celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y,
+      opus_val32 *xcorr, int len, int max_pitch, int arch)
+#endif
 {
+
+#if 0 /* This is a simple version of the pitch correlation that should work
+         well on DSPs like Blackfin and TI C5x/C6x */
    int i, j;
 #ifdef FIXED_POINT
    opus_val32 maxcorr=1;
 #endif
+#if !defined(OVERRIDE_PITCH_XCORR)
+   (void)arch;
+#endif
    for (i=0;i<max_pitch;i++)
    {
       opus_val32 sum = 0;
       for (j=0;j<len;j++)
-         sum = MAC16_16(sum, x[j],y[i+j]);
+         sum = MAC16_16(sum, _x[j], _y[i+j]);
       xcorr[i] = sum;
 #ifdef FIXED_POINT
       maxcorr = MAX32(maxcorr, sum);
 #endif
    }
 #ifdef FIXED_POINT
    return maxcorr;
 #endif
-}
 
 #else /* Unrolled version of the pitch correlation -- runs faster on x86 and ARM */
-
-#ifdef FIXED_POINT
-opus_val32
-#else
-void
-#endif
-celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch)
-{
-   int i,j;
+   int i;
    /*The EDSP version requires that max_pitch is at least 1, and that _x is
       32-bit aligned.
      Since it's hard to put asserts in assembly, put them here.*/
-   celt_assert(max_pitch>0);
-   celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
 #ifdef FIXED_POINT
    opus_val32 maxcorr=1;
 #endif
+   celt_assert(max_pitch>0);
+   celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
    for (i=0;i<max_pitch-3;i+=4)
    {
       opus_val32 sum[4]={0,0,0,0};
-      xcorr_kernel(_x, _y+i, sum, len);
+#if defined(OVERRIDE_PITCH_XCORR)
+      xcorr_kernel_c(_x, _y+i, sum, len);
+#else
+      xcorr_kernel(_x, _y+i, sum, len, arch);
+#endif
       xcorr[i]=sum[0];
       xcorr[i+1]=sum[1];
       xcorr[i+2]=sum[2];
       xcorr[i+3]=sum[3];
 #ifdef FIXED_POINT
       sum[0] = MAX32(sum[0], sum[1]);
       sum[2] = MAX32(sum[2], sum[3]);
       sum[0] = MAX32(sum[0], sum[2]);
       maxcorr = MAX32(maxcorr, sum[0]);
 #endif
    }
    /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */
    for (;i<max_pitch;i++)
    {
-      opus_val32 sum = 0;
-      for (j=0;j<len;j++)
-         sum = MAC16_16(sum, _x[j],_y[i+j]);
+      opus_val32 sum;
+#if defined(OVERRIDE_PITCH_XCORR)
+      sum = celt_inner_prod_c(_x, _y+i, len);
+#else
+      sum = celt_inner_prod(_x, _y+i, len, arch);
+#endif
       xcorr[i] = sum;
 #ifdef FIXED_POINT
       maxcorr = MAX32(maxcorr, sum);
 #endif
    }
 #ifdef FIXED_POINT
    return maxcorr;
 #endif
+#endif
 }
 
-#endif
 void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTRICT y,
                   int len, int max_pitch, int *pitch, int arch)
 {
    int i, j;
    int lag;
    int best_pitch[2]={0,0};
    VARDECL(opus_val16, x_lp4);
    VARDECL(opus_val16, y_lp4);
    VARDECL(opus_val32, xcorr);
 #ifdef FIXED_POINT
@@ skipping 48 matching lines @@
                    , 0, maxcorr
 #endif
                    );
 
    /* Finer search with 2x decimation */
 #ifdef FIXED_POINT
    maxcorr=1;
 #endif
    for (i=0;i<max_pitch>>1;i++)
    {
-      opus_val32 sum=0;
+      opus_val32 sum;
       xcorr[i] = 0;
       if (abs(i-2*best_pitch[0])>2 && abs(i-2*best_pitch[1])>2)
          continue;
+#ifdef FIXED_POINT
+      sum = 0;
       for (j=0;j<len>>1;j++)
          sum += SHR32(MULT16_16(x_lp[j],y[i+j]), shift);
+#else
+      sum = celt_inner_prod_c(x_lp, y+i, len>>1);
+#endif
       xcorr[i] = MAX32(-1, sum);
 #ifdef FIXED_POINT
       maxcorr = MAX32(maxcorr, sum);
 #endif
    }
    find_best_pitch(xcorr, y, len>>1, max_pitch>>1, best_pitch
 #ifdef FIXED_POINT
                    , shift+1, maxcorr
 #endif
                    );
@@ skipping 14 matching lines @@
    } else {
       offset = 0;
    }
    *pitch = 2*best_pitch[0]-offset;
 
    RESTORE_STACK;
 }
 
 static const int second_check[16] = {0, 0, 3, 2, 3, 2, 5, 2, 3, 2, 3, 2, 5, 2, 3, 2};
 opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
-      int N, int *T0_, int prev_period, opus_val16 prev_gain)
+      int N, int *T0_, int prev_period, opus_val16 prev_gain, int arch)
 {
    int k, i, T, T0;
    opus_val16 g, g0;
    opus_val16 pg;
    opus_val32 xy,xx,yy,xy2;
    opus_val32 xcorr[3];
    opus_val32 best_xy, best_yy;
    int offset;
    int minperiod0;
    VARDECL(opus_val32, yy_lookup);
@@ skipping 34 matching lines @@
 #else
       g = g0 = xy/celt_sqrt(1+xx*yy);
 #endif
    /* Look for any pitch at T/k */
    for (k=2;k<=15;k++)
    {
       int T1, T1b;
       opus_val16 g1;
       opus_val16 cont=0;
       opus_val16 thresh;
-      T1 = (2*T0+k)/(2*k);
+      T1 = celt_udiv(2*T0+k, 2*k);
       if (T1 < minperiod)
          break;
       /* Look for another strong correlation at T1b */
       if (k==2)
       {
          if (T1+T0>maxperiod)
             T1b = T0;
          else
             T1b = T0+T1;
       } else
       {
-         T1b = (2*second_check[k]*T0+k)/(2*k);
+         T1b = celt_udiv(2*second_check[k]*T0+k, 2*k);
       }
       dual_inner_prod(x, &x[-T1], &x[-T1b], N, &xy, &xy2);
       xy += xy2;
       yy = yy_lookup[T1] + yy_lookup[T1b];
 #ifdef FIXED_POINT
       {
          opus_val32 x2y2;
          int sh, t;
          x2y2 = 1+MULT32_32_Q31(xx,yy);
          sh = celt_ilog2(x2y2)>>1;
@@ skipping 24 matching lines @@
          g = g1;
       }
    }
    best_xy = MAX32(0, best_xy);
    if (best_yy <= best_xy)
       pg = Q15ONE;
    else
       pg = SHR32(frac_div32(best_xy,best_yy+1),16);
 
    for (k=0;k<3;k++)
-   {
-      int T1 = T+k-1;
-      xy = 0;
-      for (i=0;i<N;i++)
-         xy = MAC16_16(xy, x[i], x[i-T1]);
-      xcorr[k] = xy;
-   }
+      xcorr[k] = celt_inner_prod(x, x-(T+k-1), N, arch);
    if ((xcorr[2]-xcorr[0]) > MULT16_32_Q15(QCONST16(.7f,15),xcorr[1]-xcorr[0]))
       offset = 1;
    else if ((xcorr[0]-xcorr[2]) > MULT16_32_Q15(QCONST16(.7f,15),xcorr[1]-xcorr[2]))
       offset = -1;
    else
       offset = 0;
    if (pg > g)
       pg = g;
    *T0_ = 2*T+offset;
 
    if (*T0_<minperiod0)
       *T0_=minperiod0;
    RESTORE_STACK;
    return pg;
 }