Adds a 3 Hz high-pass filter and boost allocation on leakage
authorJean-Marc Valin <jmvalin@jmvalin.ca>
Mon, 19 Dec 2011 03:12:42 +0000 (22:12 -0500)
committerJean-Marc Valin <jmvalin@jmvalin.ca>
Fri, 13 Jul 2012 18:50:35 +0000 (14:50 -0400)
Fixes two leakage problems on the wood blocks sample
- Removes DC which causes leakage with no masking
- Detect leakage by comparing short-MDCT energy to long-MDCT energy
  and boost allocation for bands with leakage

celt/celt.c
celt/mdct.c
src/opus_encoder.c

index 0e953b2..7645b22 100644 (file)
@@ -1325,9 +1325,44 @@ int celt_encode_with_ec(CELTEncoder * restrict st, const opus_val16 * pcm, int f
    compute_band_energies(st->mode, freq, bandE, effEnd, C, M);
 
    amp2Log2(st->mode, effEnd, st->end, bandE, bandLogE, C);
-   /*for (i=0;i<17;i++)
+   /*for (i=0;i<21;i++)
       printf("%f ", bandLogE[i]);
    printf("\n");*/
+
+   ALLOC(bandLogE2, C*st->mode->nbEBands, opus_val16);
+   if (shortBlocks)
+   {
+      ALLOC(freq2, C*N, celt_sig);
+      compute_mdcts(st->mode, 0, in, freq2, CC, LM);
+      if (CC==2&&C==1)
+      {
+         for (i=0;i<N;i++)
+            freq2[i] = ADD32(HALF32(freq2[i]), HALF32(freq2[N+i]));
+      }
+      if (st->upsample != 1)
+      {
+         c=0; do
+         {
+            int bound = N/st->upsample;
+            for (i=0;i<bound;i++)
+               freq2[c*N+i] *= st->upsample;
+            for (;i<N;i++)
+               freq2[c*N+i] = 0;
+         } while (++c<C);
+      }
+      ALLOC(bandE2, C*st->mode->nbEBands, opus_val32);
+      compute_band_energies(st->mode, freq2, bandE2, effEnd, C, M);
+      amp2Log2(st->mode, effEnd, st->end, bandE2, bandLogE2, C);
+      for (i=0;i<C*st->mode->nbEBands;i++)
+         bandLogE2[i] += LM/2.;
+   } else {
+      for (i=0;i<C*st->mode->nbEBands;i++)
+         bandLogE2[i] = bandLogE[i];
+   }
+   /*for (i=0;i<C*st->mode->nbEBands;i++)
+      printf("%f ", MAX16(0,bandLogE[i]-bandLogE2[i]-LM/2.));
+   printf("\n");*/
+
    {
       opus_val16 follower[42]={0};
       c=0;do
@@ -1403,11 +1438,11 @@ int celt_encode_with_ec(CELTEncoder * restrict st, const opus_val16 * pcm, int f
       opus_val16 follower[42]={0};
       c=0;do
       {
-         follower[c*st->mode->nbEBands] = bandLogE[c*st->mode->nbEBands];
+         follower[c*st->mode->nbEBands] = bandLogE2[c*st->mode->nbEBands];
          for (i=1;i<st->mode->nbEBands;i++)
-            follower[c*st->mode->nbEBands+i] = MIN16(follower[c*st->mode->nbEBands+i-1]+2, bandLogE[c*st->mode->nbEBands+i]);
+            follower[c*st->mode->nbEBands+i] = MIN16(follower[c*st->mode->nbEBands+i-1]+2, bandLogE2[c*st->mode->nbEBands+i]);
          for (i=st->mode->nbEBands-2;i>=0;i--)
-            follower[c*st->mode->nbEBands+i] = MIN16(follower[c*st->mode->nbEBands+i], MIN16(follower[c*st->mode->nbEBands+i+1]+2, bandLogE[c*st->mode->nbEBands+i]));
+            follower[c*st->mode->nbEBands+i] = MIN16(follower[c*st->mode->nbEBands+i], MIN16(follower[c*st->mode->nbEBands+i+1]+2, bandLogE2[c*st->mode->nbEBands+i]));
       } while (++c<2);
       if (C==2)
       {
@@ -1423,13 +1458,13 @@ int celt_encode_with_ec(CELTEncoder * restrict st, const opus_val16 * pcm, int f
             follower[i] = MAX16(0, bandLogE[i]-follower[i]);
          }
       }
-      opus_val32 tot_boost=effectiveBytes*8/6;
+      opus_val32 tot_boost=(effectiveBytes*8-20-40*C)/5;
       for (i=st->start;i<st->end-1;i++)
       {
          int width;
          int boost;
 
-         follower[i] = MIN16(follower[i], QCONST16(2, DB_SHIFT));
+         follower[i] = MIN16(2*follower[i], i<10 ? QCONST16(4, DB_SHIFT) : QCONST16(2, DB_SHIFT));
          width = C*(st->mode->eBands[i+1]-st->mode->eBands[i])<<LM;
          if (width<6)
          {
@@ -1446,43 +1481,10 @@ int celt_encode_with_ec(CELTEncoder * restrict st, const opus_val16 * pcm, int f
          offsets[i] = boost;
       }
       /*for (i=st->start;i<st->end-1;i++)
-         printf("%f ", follower[i]);*/
-      //printf("%f\n", tot_boost);
-#if 0
-      if (LM <= 1)
-      {
-         t1 = 3;
-         t2 = 5;
-      } else {
-         t1 = 2;
-         t2 = 4;
-      }
-      for (i=st->start+1;i<st->end-1;i++)
-      {
-         opus_val32 d2;
-         d2 = 2*bandLogE[i]-bandLogE[i-1]-bandLogE[i+1];
-         if (C==2)
-            d2 = HALF32(d2 + 2*bandLogE[i+st->mode->nbEBands]-
-                  bandLogE[i-1+st->mode->nbEBands]-bandLogE[i+1+st->mode->nbEBands]);
-#ifdef FUZZING
-         if((rand()&0xF)==0)
-         {
-            offsets[i] += 1;
-            if((rand()&0x3)==0)
-               offsets[i] += 1+(rand()&0x3);
-         }
-#else
-         if (d2 > SHL16(t1,DB_SHIFT))
-            offsets[i] += 1;
-         if (d2 > SHL16(t2,DB_SHIFT))
-            offsets[i] += 1;
-#endif
-      }
-#endif
+         printf("%f ", follower[i]);
+      printf("%f\n", tot_boost);*/
    }
 #ifndef FIXED_POINT
-   //offsets[4]  += 12;
-   //offsets[10] += 12;
    if (0 && st->analysis.valid)
    {
       if (st->analysis.boost_amount[0]>.2)
index 3193b7c..acccb8f 100644 (file)
@@ -109,12 +109,14 @@ void clt_mdct_forward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar
    int N, N2, N4;
    kiss_twiddle_scalar sine;
    VARDECL(kiss_fft_scalar, f);
+   VARDECL(kiss_fft_scalar, f2);
    SAVE_STACK;
    N = l->n;
    N >>= shift;
    N2 = N>>1;
    N4 = N>>2;
    ALLOC(f, N2, kiss_fft_scalar);
+   ALLOC(f2, N2, kiss_fft_scalar);
    /* sin(x) ~= x here */
 #ifdef FIXED_POINT
    sine = TRIG_UPSCALE*(QCONST16(0.7853981f, 15)+N2)/N;
@@ -180,12 +182,12 @@ void clt_mdct_forward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar
    }
 
    /* N/4 complex FFT, down-scales by 4/N */
-   opus_fft(l->kfft[shift], (kiss_fft_cpx *)f, (kiss_fft_cpx *)in);
+   opus_fft(l->kfft[shift], (kiss_fft_cpx *)f, (kiss_fft_cpx *)f2);
 
    /* Post-rotate */
    {
       /* Temp pointers to make it really clear to the compiler what we're doing */
-      const kiss_fft_scalar * restrict fp = in;
+      const kiss_fft_scalar * restrict fp = f2;
       kiss_fft_scalar * restrict yp1 = out;
       kiss_fft_scalar * restrict yp2 = out+stride*(N2-1);
       const kiss_twiddle_scalar *t = &l->trig[0];
index 2a14c07..6dee15d 100644 (file)
@@ -859,8 +859,9 @@ opus_int32 opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_s
     {
        hp_cutoff(pcm, cutoff_Hz, &pcm_buf[total_buffer*st->channels], st->hp_mem, frame_size, st->channels, st->Fs);
     } else {
-       for (i=0;i<frame_size*st->channels;i++)
-          pcm_buf[total_buffer*st->channels + i] = pcm[i];
+       hp_cutoff(pcm, 3, &pcm_buf[total_buffer*st->channels], st->hp_mem, frame_size, st->channels, st->Fs);
+       /*for (i=0;i<frame_size*st->channels;i++)
+          pcm_buf[total_buffer*st->channels + i] = pcm[i];*/
     }
 
 #ifndef FIXED_POINT