Fixes permissions on arm2gnu.pl

[opus.git] / celt / pitch.c

diff --git a/celt/pitch.c b/celt/pitch.c
index cffee19..2d63a5a 100644 (file)
--- a/celt/pitch.c
+++ b/celt/pitch.c
@@ -145,7 +145,7 @@ static void celt_fir5(const opus_val16 *x,
 
 
 void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x_lp,
-      int len, int C)
+      int len, int C, int arch)
 {
    int i;
    opus_val32 ac[5];
@@ -180,7 +180,7 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x
    }
 
    _celt_autocorr(x_lp, ac, NULL, 0,
-                  4, len>>1);
+                  4, len>>1, arch);
 
    /* Noise floor -40 dB */
 #ifdef FIXED_POINT
@@ -222,7 +222,7 @@ opus_val32
 #else
 void
 #endif
-pitch_xcorr(opus_val16 *x, opus_val16 *y, opus_val32 *xcorr, int len, int max_pitch)
+celt_pitch_xcorr(opus_val16 *x, opus_val16 *y, opus_val32 *xcorr, int len, int max_pitch)
 {
    int i, j;
 #ifdef FIXED_POINT
@@ -250,7 +250,7 @@ opus_val32
 #else
 void
 #endif
-pitch_xcorr(opus_val16 *_x, opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch)
+celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch)
 {
    int i,j;
 #ifdef FIXED_POINT
@@ -258,83 +258,17 @@ pitch_xcorr(opus_val16 *_x, opus_val16 *_y, opus_val32 *xcorr, int len, int max_
 #endif
    for (i=0;i<max_pitch-3;i+=4)
    {
-      /* Compute correlation*/
-      /*corr[nb_pitch-1-i]=inner_prod(x, _y+i, len);*/
-      opus_val32 sum1=0;
-      opus_val32 sum2=0;
-      opus_val32 sum3=0;
-      opus_val32 sum4=0;
-      const opus_val16 *y = _y+i;
-      const opus_val16 *x = _x;
-      opus_val16 y_0, y_1, y_2, y_3;
-      y_3=0; /* gcc doesn't realize that y_3 can't be used uninitialized */
-      y_0=*y++;
-      y_1=*y++;
-      y_2=*y++;
-      for (j=0;j<len-3;j+=4)
-      {
-         opus_val16 tmp;
-         tmp = *x++;
-         y_3=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_0);
-         sum2 = MAC16_16(sum2,tmp,y_1);
-         sum3 = MAC16_16(sum3,tmp,y_2);
-         sum4 = MAC16_16(sum4,tmp,y_3);
-         tmp=*x++;
-         y_0=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_1);
-         sum2 = MAC16_16(sum2,tmp,y_2);
-         sum3 = MAC16_16(sum3,tmp,y_3);
-         sum4 = MAC16_16(sum4,tmp,y_0);
-         tmp=*x++;
-         y_1=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_2);
-         sum2 = MAC16_16(sum2,tmp,y_3);
-         sum3 = MAC16_16(sum3,tmp,y_0);
-         sum4 = MAC16_16(sum4,tmp,y_1);
-         tmp=*x++;
-         y_2=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_3);
-         sum2 = MAC16_16(sum2,tmp,y_0);
-         sum3 = MAC16_16(sum3,tmp,y_1);
-         sum4 = MAC16_16(sum4,tmp,y_2);
-      }
-      if (j++<len)
-      {
-         opus_val16 tmp = *x++;
-         y_3=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_0);
-         sum2 = MAC16_16(sum2,tmp,y_1);
-         sum3 = MAC16_16(sum3,tmp,y_2);
-         sum4 = MAC16_16(sum4,tmp,y_3);
-      }
-      if (j++<len)
-      {
-         opus_val16 tmp=*x++;
-         y_0=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_1);
-         sum2 = MAC16_16(sum2,tmp,y_2);
-         sum3 = MAC16_16(sum3,tmp,y_3);
-         sum4 = MAC16_16(sum4,tmp,y_0);
-      }
-      if (j<len)
-      {
-         opus_val16 tmp=*x++;
-         y_1=*y++;
-         sum1 = MAC16_16(sum1,tmp,y_2);
-         sum2 = MAC16_16(sum2,tmp,y_3);
-         sum3 = MAC16_16(sum3,tmp,y_0);
-         sum4 = MAC16_16(sum4,tmp,y_1);
-      }
-      xcorr[i]=sum1;
-      xcorr[i+1]=sum2;
-      xcorr[i+2]=sum3;
-      xcorr[i+3]=sum4;
+      opus_val32 sum[4]={0,0,0,0};
+      xcorr_kernel(_x, _y+i, sum, len);
+      xcorr[i]=sum[0];
+      xcorr[i+1]=sum[1];
+      xcorr[i+2]=sum[2];
+      xcorr[i+3]=sum[3];
 #ifdef FIXED_POINT
-      sum1 = MAX32(sum1, sum2);
-      sum3 = MAX32(sum3, sum4);
-      sum1 = MAX32(sum1, sum3);
-      maxcorr = MAX32(maxcorr, sum1);
+      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. */
@@ -355,7 +289,7 @@ pitch_xcorr(opus_val16 *_x, opus_val16 *_y, opus_val32 *xcorr, int len, int max_
 
 #endif
 void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTRICT y,
-                  int len, int max_pitch, int *pitch)
+                  int len, int max_pitch, int *pitch, int arch)
 {
    int i, j;
    int lag;
@@ -408,7 +342,7 @@ void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTR
 #ifdef FIXED_POINT
    maxcorr =
 #endif
-   pitch_xcorr(x_lp4, y_lp4, xcorr, len>>2, max_pitch>>2);
+   celt_pitch_xcorr(x_lp4, y_lp4, xcorr, len>>2, max_pitch>>2, arch);
 
    find_best_pitch(xcorr, y_lp4, len>>2, max_pitch>>2, best_pitch
 #ifdef FIXED_POINT
@@ -467,7 +401,7 @@ opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
    int k, i, T, T0;
    opus_val16 g, g0;
    opus_val16 pg;
-   opus_val32 xy,xx,yy;
+   opus_val32 xy,xx,yy,xy2;
    opus_val32 xcorr[3];
    opus_val32 best_xy, best_yy;
    int offset;
@@ -487,12 +421,7 @@ opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
 
    T = T0 = *T0_;
    ALLOC(yy_lookup, maxperiod+1, opus_val32);
-   xy=xx=0;
-   for (i=0;i<N;i++)
-   {
-      xx = MAC16_16(xx, x[i], x[i]);
-      xy = MAC16_16(xy, x[i], x[i-T0]);
-   }
+   dual_inner_prod(x, x, x-T0, N, &xx, &xy);
    yy_lookup[0] = xx;
    yy=xx;
    for (i=1;i<=maxperiod;i++)
@@ -536,12 +465,8 @@ opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
       {
          T1b = (2*second_check[k]*T0+k)/(2*k);
       }
-      xy=0;
-      for (i=0;i<N;i++)
-      {
-         xy = MAC16_16(xy, x[i], x[i-T1]);
-         xy = MAC16_16(xy, x[i], x[i-T1b]);
-      }
+      dual_inner_prod(x, &x[-T1], &x[-T1b], N, &xy, &xy2);
+      xy += xy2;
       yy = yy_lookup[T1] + yy_lookup[T1b];
 #ifdef FIXED_POINT
       {