Fix a minor, but bitstream-affecting bug

[opus.git] / libcelt / vq.c

diff --git a/libcelt/vq.c b/libcelt/vq.c
index 62a8784..4876b59 100644 (file)
--- a/libcelt/vq.c
+++ b/libcelt/vq.c
@@ -1,5 +1,6 @@
-/* (C) 2007-2008 Jean-Marc Valin, CSIRO
-*/
+/* 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
 /*
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions


@@ -44,103 +45,168 @@
 #define M_PI 3.141592653
 #endif
 
 #define M_PI 3.141592653
 #endif
 

-static void exp_rotation(celt_norm_t *X, int len, int dir, int stride, int K)
+static celt_uint32 lcg_rand(celt_uint32 seed)

 {
 {

-   int i, k, iter;
-   celt_word16_t c, s;
-   celt_word16_t gain, theta;
-   celt_norm_t *Xptr;
-   gain = celt_div((celt_word32_t)MULT16_16(Q15_ONE,len),(celt_word32_t)(3+len+6*K));
-   /* FIXME: Make that HALF16 instead of HALF32 */
-   theta = SUB16(Q15ONE, HALF32(MULT16_16_Q15(gain,gain)));
-   /*if (len==30)
+   return 1664525 * seed + 1013904223;
+}
+
+static void exp_rotation1(celt_norm *X, int len, int stride, celt_word16 c, celt_word16 s)
+{
+   int i;
+   celt_norm *Xptr;
+   Xptr = X;
+   for (i=0;i<len-stride;i++)
+   {
+      celt_norm x1, x2;
+      x1 = Xptr[0];
+      x2 = Xptr[stride];
+      Xptr[stride] = EXTRACT16(SHR32(MULT16_16(c,x2) + MULT16_16(s,x1), 15));
+      *Xptr++      = EXTRACT16(SHR32(MULT16_16(c,x1) - MULT16_16(s,x2), 15));
+   }
+   Xptr = &X[len-2*stride-1];
+   for (i=len-2*stride-1;i>=0;i--)

    {
    {

-   for (i=0;i<len;i++)
-   X[i] = 0;
-   X[14] = 1;
-}*/ 
+      celt_norm x1, x2;
+      x1 = Xptr[0];
+      x2 = Xptr[stride];
+      Xptr[stride] = EXTRACT16(SHR32(MULT16_16(c,x2) + MULT16_16(s,x1), 15));
+      *Xptr--      = EXTRACT16(SHR32(MULT16_16(c,x1) - MULT16_16(s,x2), 15));
+   }
+}
+
+static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
+{
+   int i;
+   celt_word16 c, s;
+   celt_word16 gain, theta;
+   int stride2=0;
+   int factor;
+   /*int i;
+   if (len>=30)
+   {
+      for (i=0;i<len;i++)
+         X[i] = 0;
+      X[14] = 1;
+      K=5;
+   }*/
+   if (2*K>=len || spread==0)
+      return;
+   if (spread==1)
+      factor=10;
+   else if (spread==2)
+      factor=5;
+   else
+      factor=15;
+
+   gain = celt_div((celt_word32)MULT16_16(Q15_ONE,len),(celt_word32)(len+factor*K));
+   /* FIXME: Make that HALF16 instead of HALF32 */
+   theta = HALF32(MULT16_16_Q15(gain,gain));
+

    c = celt_cos_norm(EXTEND32(theta));
    c = celt_cos_norm(EXTEND32(theta));

-   s = dir*celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /*  sin(theta) */
-   if (len > 8*stride)
-      stride *= len/(8*stride);
-   iter = 1;
-   for (k=0;k<iter;k++)
+   s = celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /*  sin(theta) */
+
+   if (len>=8*stride)

    {
    {

-      /* We could use MULT16_16_P15 instead of MULT16_16_Q15 for more accuracy, 
-      but at this point, I really don't think it's necessary */
-      Xptr = X;
-      for (i=0;i<len-stride;i++)
-      {
-         celt_norm_t x1, x2;
-         x1 = Xptr[0];
-         x2 = Xptr[stride];
-         Xptr[stride] = MULT16_16_Q15(c,x2) + MULT16_16_Q15(s,x1);
-         *Xptr++      = MULT16_16_Q15(c,x1) - MULT16_16_Q15(s,x2);
-      }
-      Xptr = &X[len-2*stride-1];
-      for (i=len-2*stride-1;i>=0;i--)
+      stride2 = 1;
+      /* This is just a simple way of computing sqrt(len/stride) with rounding.
+         It's basically incrementing long as (stride2+0.5)^2 < len/stride.
+         I _think_ it is bit-exact */
+      while ((stride2*stride2+stride2)*stride + (stride>>2) < len)
+         stride2++;
+   }
+   len /= stride;
+   for (i=0;i<stride;i++)
+   {
+      if (dir < 0)

       {
       {

-         celt_norm_t x1, x2;
-         x1 = Xptr[0];
-         x2 = Xptr[stride];
-         Xptr[stride] = MULT16_16_Q15(c,x2) + MULT16_16_Q15(s,x1);
-         *Xptr--      = MULT16_16_Q15(c,x1) - MULT16_16_Q15(s,x2);
+         if (stride2)
+            exp_rotation1(X+i*len, len, stride2, s, c);
+         exp_rotation1(X+i*len, len, 1, c, s);
+      } else {
+         exp_rotation1(X+i*len, len, 1, c, -s);
+         if (stride2)
+            exp_rotation1(X+i*len, len, stride2, s, -c);

       }
    }
       }
    }

-   /*if (len==30)
+   /*if (len>=30)

    {
    {

-   for (i=0;i<len;i++)
-   printf ("%f ", X[i]);
-   printf ("\n");
-   exit(0);
-}*/
+      for (i=0;i<len;i++)
+         printf ("%f ", X[i]);
+      printf ("\n");
+      exit(0);
+   }*/

 }
 
 }
 

-

 /** Takes the pitch vector and the decoded residual vector, computes the gain
     that will give ||p+g*y||=1 and mixes the residual with the pitch. */
 /** Takes the pitch vector and the decoded residual vector, computes the gain
     that will give ||p+g*y||=1 and mixes the residual with the pitch. */

-static void mix_pitch_and_residual(int * restrict iy, celt_norm_t * restrict X, int N, int K, celt_word32_t Ryy)
+static void normalise_residual(int * restrict iy, celt_norm * restrict X, int N, int K, celt_word32 Ryy)

 {
    int i;
 {
    int i;

-   celt_word32_t g;
+#ifdef FIXED_POINT
+   int k;
+#endif
+   celt_word32 t;
+   celt_word16 g;

 
 

-   g = celt_rsqrt(Ryy);
+#ifdef FIXED_POINT
+   k = celt_ilog2(Ryy)>>1;
+#endif
+   t = VSHR32(Ryy, (k-7)<<1);
+   g = celt_rsqrt_norm(t);

 
    i=0;
    do
 
    i=0;
    do

-      X[i] = MULT16_32_P15(g, SHL32(EXTEND32(iy[i]),14));
+      X[i] = EXTRACT16(PSHR32(MULT16_16(g, iy[i]), k+1));

    while (++i < N);
 }
 
    while (++i < N);
 }
 

-void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
+void alg_quant(celt_norm *X, int N, int K, int spread, int B, celt_norm *lowband, int resynth, ec_enc *enc, celt_int32 *seed)

 {
 {

-   VARDECL(celt_norm_t, y);
+   VARDECL(celt_norm, y);

    VARDECL(int, iy);
    VARDECL(int, iy);

-   VARDECL(celt_word16_t, signx);
+   VARDECL(celt_word16, signx);

    int j, is;
    int j, is;

-   celt_word16_t s;
+   celt_word16 s;

    int pulsesLeft;
    int pulsesLeft;

-   celt_word32_t sum;
-   celt_word32_t xy, yy;
+   celt_word32 sum;
+   celt_word32 xy, yy;

    int N_1; /* Inverse of N, in Q14 format (even for float) */
 #ifdef FIXED_POINT
    int yshift;
 #endif
    SAVE_STACK;
 
    int N_1; /* Inverse of N, in Q14 format (even for float) */
 #ifdef FIXED_POINT
    int yshift;
 #endif
    SAVE_STACK;
 

+   /* When there's no pulse, fill with noise or folded spectrum */
+   if (K==0)
+   {
+      if (lowband != NULL && resynth)
+      {
+         for (j=0;j<N;j++)
+            X[j] = lowband[j];
+      } else {
+         /* This is important for encoding the side in stereo mode */
+         for (j=0;j<N;j++)
+         {
+            *seed = lcg_rand(*seed);
+            X[j] = (int)(*seed)>>20;
+         }
+      }
+      renormalise_vector(X, Q15ONE, N, 1);
+      return;
+   }

    K = get_pulses(K);
 #ifdef FIXED_POINT
    yshift = 13-celt_ilog2(K);
 #endif
 
    K = get_pulses(K);
 #ifdef FIXED_POINT
    yshift = 13-celt_ilog2(K);
 #endif
 

-   ALLOC(y, N, celt_norm_t);
+   ALLOC(y, N, celt_norm);

    ALLOC(iy, N, int);
    ALLOC(iy, N, int);

-   ALLOC(signx, N, celt_word16_t);
+   ALLOC(signx, N, celt_word16);

    N_1 = 512/N;
    
    N_1 = 512/N;
    

-   if (spread)
-      exp_rotation(X, N, 1, spread, K);
+   exp_rotation(X, N, 1, B, K, spread);

 
 

+   /* Get rid of the sign */

    sum = 0;
    j=0; do {
       if (X[j]>0)
    sum = 0;
    j=0; do {
       if (X[j]>0)


@@ -160,12 +226,12 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
    /* Do a pre-search by projecting on the pyramid */
    if (K > (N>>1))
    {
    /* Do a pre-search by projecting on the pyramid */
    if (K > (N>>1))
    {

-      celt_word16_t rcp;
-      sum=0;
+      celt_word16 rcp;

       j=0; do {
          sum += X[j];
       }  while (++j<N);
 
       j=0; do {
          sum += X[j];
       }  while (++j<N);
 

+      /* If X is too small, just replace it with a pulse at 0 */

 #ifdef FIXED_POINT
       if (sum <= K)
 #else
 #ifdef FIXED_POINT
       if (sum <= K)
 #else


@@ -185,7 +251,7 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
          /* It's really important to round *towards zero* here */
          iy[j] = MULT16_16_Q15(X[j],rcp);
 #else
          /* It's really important to round *towards zero* here */
          iy[j] = MULT16_16_Q15(X[j],rcp);
 #else

-         iy[j] = floor(rcp*X[j]);
+         iy[j] = (int)floor(rcp*X[j]);

 #endif
          y[j] = SHL16(iy[j],yshift);
          yy = MAC16_16(yy, y[j],y[j]);
 #endif
          y[j] = SHL16(iy[j],yshift);
          yy = MAC16_16(yy, y[j],y[j]);


@@ -200,9 +266,9 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
    {
       int pulsesAtOnce=1;
       int best_id;
    {
       int pulsesAtOnce=1;
       int best_id;

-      celt_word16_t magnitude;
-      celt_word32_t best_num = -VERY_LARGE16;
-      celt_word16_t best_den = 0;
+      celt_word16 magnitude;
+      celt_word32 best_num = -VERY_LARGE16;
+      celt_word16 best_den = 0;

 #ifdef FIXED_POINT
       int rshift;
 #endif
 #ifdef FIXED_POINT
       int rshift;
 #endif


@@ -223,7 +289,7 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
          /* This should ensure that anything we can process will have a better score */
       j=0;
       do {
          /* This should ensure that anything we can process will have a better score */
       j=0;
       do {

-         celt_word16_t Rxy, Ryy;
+         celt_word16 Rxy, Ryy;

          /* Select sign based on X[j] alone */
          s = magnitude;
          /* Temporary sums of the new pulse(s) */
          /* Select sign based on X[j] alone */
          s = magnitude;
          /* Temporary sums of the new pulse(s) */


@@ -231,11 +297,11 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
          /* We're multiplying y[j] by two so we don't have to do it here */
          Ryy = EXTRACT16(SHR32(MAC16_16(yy, s,y[j]),rshift));
             
          /* We're multiplying y[j] by two so we don't have to do it here */
          Ryy = EXTRACT16(SHR32(MAC16_16(yy, s,y[j]),rshift));
             

-            /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that 
-         Rxy is positive because the sign is pre-computed) */
+         /* 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);
          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 */
+         /* 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: Make sure to use conditional moves here */
          if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))
          {


@@ -260,6 +326,8 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
       iy[j] += is;
       pulsesLeft -= pulsesAtOnce;
    }
       iy[j] += is;
       pulsesLeft -= pulsesAtOnce;
    }

+
+   /* Put the original sign back */

    j=0;
    do {
       X[j] = MULT16_16(signx[j],X[j]);
    j=0;
    do {
       X[j] = MULT16_16(signx[j],X[j]);


@@ -268,23 +336,41 @@ void alg_quant(celt_norm_t *X, int N, int K, int spread, ec_enc *enc)
    } while (++j<N);
    encode_pulses(iy, N, K, enc);
    
    } while (++j<N);
    encode_pulses(iy, N, K, enc);
    

-   /* Recompute the gain in one pass to reduce the encoder-decoder mismatch
-   due to the recursive computation used in quantisation. */
-   mix_pitch_and_residual(iy, X, N, K, EXTRACT16(SHR32(yy,2*yshift)));
-   if (spread)
-      exp_rotation(X, N, -1, spread, K);
+   if (resynth)
+   {
+      normalise_residual(iy, X, N, K, EXTRACT16(SHR32(yy,2*yshift)));
+      exp_rotation(X, N, -1, B, K, spread);
+   }

    RESTORE_STACK;
 }
 
 
 /** Decode pulse vector and combine the result with the pitch vector to produce
     the final normalised signal in the current band. */
    RESTORE_STACK;
 }
 
 
 /** Decode pulse vector and combine the result with the pitch vector to produce
     the final normalised signal in the current band. */

-void alg_unquant(celt_norm_t *X, int N, int K, int spread, ec_dec *dec)
+void alg_unquant(celt_norm *X, int N, int K, int spread, int B, celt_norm *lowband, ec_dec *dec, celt_int32 *seed)

 {
    int i;
 {
    int i;

-   celt_word32_t Ryy;
+   celt_word32 Ryy;

    VARDECL(int, iy);
    SAVE_STACK;
    VARDECL(int, iy);
    SAVE_STACK;

+
+   if (K==0)
+   {
+      if (lowband != NULL)
+      {
+         for (i=0;i<N;i++)
+            X[i] = lowband[i];
+      } else {
+         /* This is important for encoding the side in stereo mode */
+         for (i=0;i<N;i++)
+         {
+            *seed = lcg_rand(*seed);
+            X[i] = (int)(*seed)>>20;
+         }
+      }
+      renormalise_vector(X, Q15ONE, N, 1);
+      return;
+   }

    K = get_pulses(K);
    ALLOC(iy, N, int);
    decode_pulses(iy, N, K, dec);
    K = get_pulses(K);
    ALLOC(iy, N, int);
    decode_pulses(iy, N, K, dec);


@@ -293,69 +379,38 @@ void alg_unquant(celt_norm_t *X, int N, int K, int spread, ec_dec *dec)
    do {
       Ryy = MAC16_16(Ryy, iy[i], iy[i]);
    } while (++i < N);
    do {
       Ryy = MAC16_16(Ryy, iy[i], iy[i]);
    } while (++i < N);

-   mix_pitch_and_residual(iy, X, N, K, Ryy);
-   if (spread)
-      exp_rotation(X, N, -1, spread, K);
+   normalise_residual(iy, X, N, K, Ryy);
+   exp_rotation(X, N, -1, B, K, spread);

    RESTORE_STACK;
 }
 
    RESTORE_STACK;
 }
 

-celt_word16_t renormalise_vector(celt_norm_t *X, celt_word16_t value, int N, int stride)
+celt_word16 renormalise_vector(celt_norm *X, celt_word16 value, int N, int stride)

 {
    int i;
 {
    int i;

-   celt_word32_t E = EPSILON;
-   celt_word16_t rE;
-   celt_word16_t g;
-   celt_norm_t *xptr = X;
+#ifdef FIXED_POINT
+   int k;
+#endif
+   celt_word32 E = EPSILON;
+   celt_word16 g;
+   celt_word32 t;
+   celt_norm *xptr = X;

    for (i=0;i<N;i++)
    {
       E = MAC16_16(E, *xptr, *xptr);
       xptr += stride;
    }
    for (i=0;i<N;i++)
    {
       E = MAC16_16(E, *xptr, *xptr);
       xptr += stride;
    }

-
-   rE = celt_sqrt(E);

 #ifdef FIXED_POINT
 #ifdef FIXED_POINT

-   if (rE <= 128)
-      g = Q15ONE;
-   else
+   k = celt_ilog2(E)>>1;

 #endif
 #endif

-      g = MULT16_16_Q15(value,celt_rcp(SHL32(rE,9)));
+   t = VSHR32(E, (k-7)<<1);
+   g = MULT16_16_Q15(value, celt_rsqrt_norm(t));
+

    xptr = X;
    for (i=0;i<N;i++)
    {
    xptr = X;
    for (i=0;i<N;i++)
    {

-      *xptr = PSHR32(MULT16_16(g, *xptr),8);
+      *xptr = EXTRACT16(PSHR32(MULT16_16(g, *xptr), k+1));

       xptr += stride;
    }
       xptr += stride;
    }

-   return rE;
-}
-
-static void fold(const CELTMode *m, int N, const celt_norm_t * restrict Y, celt_norm_t * restrict P, int N0, int B)
-{
-   int j;
-   int id = N0 % B;
-   /* Here, we assume that id will never be greater than N0, i.e. that 
-      no band is wider than N0. In the unlikely case it happens, we set
-      everything to zero */
-   /*{
-          int offset = (N0*C - (id+C*N))/2;
-          if (offset > C*N0/16)
-                  offset = C*N0/16;
-          offset -= offset % (C*B);
-          if (offset < 0)
-                  offset = 0;
-          //printf ("%d\n", offset);
-          id += offset;
-   }*/
-   if (id+N>N0)
-      for (j=0;j<N;j++)
-         P[j] = 0;
-   else
-      for (j=0;j<N;j++)
-         P[j] = Y[id++];
-}
-
-void intra_fold(const CELTMode *m, int N, const celt_norm_t * restrict Y, celt_norm_t * restrict P, int N0, int B)
-{
-   fold(m, N, Y, P, N0, B);
-   renormalise_vector(P, Q15ONE, N, 1);
+   return celt_sqrt(E);

 }
 
 }