Decimate QEXT PLC pitch search by 4 instead of 2

Reduces the PLC complexity

Author: jmvalin

celt/celt_decoder.c

@@ -542,8 +542,6 @@ static void tf_decode(int start, int end, int isTransient, int *tf_res, int LM,
 static int celt_plc_pitch_search(CELTDecoder *st, celt_sig *decode_mem[2], int C, int arch)
 {
    int pitch_index;
-   int decode_buffer_size;
-   int plc_pitch_lag_max, plc_pitch_lag_min;
 #ifdef ENABLE_QEXT
    int qext_scale;
 #endif
@@ -552,18 +550,15 @@ static int celt_plc_pitch_search(CELTDecoder *st, celt_sig *decode_mem[2], int C
 #ifdef ENABLE_QEXT
    qext_scale = st->qext_scale;
 #endif
-   decode_buffer_size = QEXT_SCALE(DECODE_BUFFER_SIZE);
-   plc_pitch_lag_max = QEXT_SCALE(PLC_PITCH_LAG_MAX);
-   plc_pitch_lag_min = QEXT_SCALE(PLC_PITCH_LAG_MIN);
-   ALLOC( lp_pitch_buf, decode_buffer_size>>1, opus_val16 );
+   ALLOC( lp_pitch_buf, DECODE_BUFFER_SIZE>>1, opus_val16 );
    pitch_downsample(decode_mem, lp_pitch_buf,
-         decode_buffer_size, C, arch);
-   pitch_search(lp_pitch_buf+(plc_pitch_lag_max>>1), lp_pitch_buf,
-         decode_buffer_size-plc_pitch_lag_max,
-         plc_pitch_lag_max-plc_pitch_lag_min, &pitch_index, arch);
-   pitch_index = plc_pitch_lag_max-pitch_index;
+         DECODE_BUFFER_SIZE>>1, C, QEXT_SCALE(2), arch);
+   pitch_search(lp_pitch_buf+(PLC_PITCH_LAG_MAX>>1), lp_pitch_buf,
+         DECODE_BUFFER_SIZE-PLC_PITCH_LAG_MAX,
+         PLC_PITCH_LAG_MAX-PLC_PITCH_LAG_MIN, &pitch_index, arch);
+   pitch_index = PLC_PITCH_LAG_MAX-pitch_index;
    RESTORE_STACK;
-   return pitch_index;
+   return QEXT_SCALE(pitch_index);
 }
 
 static void prefilter_and_fold(CELTDecoder * OPUS_RESTRICT st, int N)

celt/celt_encoder.c

@@ -1450,7 +1450,7 @@ static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem,
       VARDECL(opus_val16, pitch_buf);
       ALLOC(pitch_buf, (max_period+N)>>1, opus_val16);
 
-      pitch_downsample(pre, pitch_buf, max_period+N, CC, st->arch);
+      pitch_downsample(pre, pitch_buf, (max_period+N)>>1, CC, 2, st->arch);
       /* Don't search for the fir last 1.5 octave of the range because
          there's too many false-positives due to short-term correlation */
       pitch_search(pitch_buf+(max_period>>1), pitch_buf, N,

celt/pitch.c

@@ -138,20 +138,25 @@ static void celt_fir5(opus_val16 *x,
 
 
 void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x_lp,
-      int len, int C, int arch)
+      int len, int C, int factor, int arch)
 {
    int i;
    opus_val32 ac[5];
    opus_val16 tmp=Q15ONE;
    opus_val16 lpc[4];
    opus_val16 lpc2[5];
    opus_val16 c1 = QCONST16(.8f,15);
+   int offset;
 #ifdef FIXED_POINT
    int shift;
-   opus_val32 maxabs = celt_maxabs32(x[0], len);
+   opus_val32 maxabs;
+#endif
+   offset = factor/2;
+#ifdef FIXED_POINT
+   maxabs = celt_maxabs32(x[0], len*factor);
    if (C==2)
    {
-      opus_val32 maxabs_1 = celt_maxabs32(x[1], len);
+      opus_val32 maxabs_1 = celt_maxabs32(x[1], len*factor);
       maxabs = MAX32(maxabs, maxabs_1);
    }
    if (maxabs<1)
@@ -161,28 +166,28 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x
       shift=0;
    if (C==2)
       shift++;
-   for (i=1;i<len>>1;i++)
-      x_lp[i] = SHR32(x[0][(2*i-1)], shift+2) + SHR32(x[0][(2*i+1)], shift+2) + SHR32(x[0][2*i], shift+1);
-   x_lp[0] = SHR32(x[0][1], shift+2) + SHR32(x[0][0], shift+1);
+   for (i=1;i<len;i++)
+      x_lp[i] = SHR32(x[0][(factor*i-offset)], shift+2) + SHR32(x[0][(factor*i+offset)], shift+2) + SHR32(x[0][factor*i], shift+1);
+   x_lp[0] = SHR32(x[0][offset], shift+2) + SHR32(x[0][0], shift+1);
    if (C==2)
    {
-      for (i=1;i<len>>1;i++)
-         x_lp[i] += SHR32(x[1][(2*i-1)], shift+2) + SHR32(x[1][(2*i+1)], shift+2) + SHR32(x[1][2*i], shift+1);
-      x_lp[0] += SHR32(x[1][1], shift+2) + SHR32(x[1][0], shift+1);
+      for (i=1;i<len;i++)
+         x_lp[i] += SHR32(x[1][(factor*i-offset)], shift+2) + SHR32(x[1][(factor*i+offset)], shift+2) + SHR32(x[1][factor*i], shift+1);
+      x_lp[0] += SHR32(x[1][offset], shift+2) + SHR32(x[1][0], shift+1);
    }
 #else
-   for (i=1;i<len>>1;i++)
-      x_lp[i] = .25f*x[0][(2*i-1)] + .25f*x[0][(2*i+1)] + .5f*x[0][2*i];
-   x_lp[0] = .25f*x[0][1] + .5f*x[0][0];
+   for (i=1;i<len;i++)
+      x_lp[i] = .25f*x[0][(factor*i-offset)] + .25f*x[0][(factor*i+offset)] + .5f*x[0][factor*i];
+   x_lp[0] = .25f*x[0][offset] + .5f*x[0][0];
    if (C==2)
    {
-      for (i=1;i<len>>1;i++)
-         x_lp[i] += .25f*x[1][(2*i-1)] + .25f*x[1][(2*i+1)] + .5f*x[1][2*i];
-      x_lp[0] += .25f*x[1][1] + .5f*x[1][0];
+      for (i=1;i<len;i++)
+         x_lp[i] += .25f*x[1][(factor*i-offset)] + .25f*x[1][(factor*i+offset)] + .5f*x[1][factor*i];
+      x_lp[0] += .25f*x[1][offset] + .5f*x[1][0];
    }
 #endif
    _celt_autocorr(x_lp, ac, NULL, 0,
-                  4, len>>1, arch);
+                  4, len, arch);
 
    /* Noise floor -40 dB */
 #ifdef FIXED_POINT
@@ -213,7 +218,7 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x
    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, len>>1);
+   celt_fir5(x_lp, lpc2, len);
 }
 
 /* Pure C implementation. */

celt/pitch.h

@@ -51,7 +51,7 @@
 #endif
 
 void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x_lp,
-      int len, int C, int arch);
+      int len, int C, int factor, int arch);
 
 void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTRICT y,
                   int len, int max_pitch, int *pitch, int arch);