Oops. Thanks to Jim Crichton for pointing out that the complexity could end up
[speexdsp.git] / libspeex / cb_search.c
index 80412ff..5c68826 100644 (file)
-/* Copyright (C) 2002 Jean-Marc Valin 
+/* Copyright (C) 2002-2006 Jean-Marc Valin 
    File: cb_search.c
 
-   This library is free software; you can redistribute it and/or
-   modify it under the terms of the GNU Lesser General Public
-   License as published by the Free Software Foundation; either
-   version 2.1 of the License, or (at your option) any later version.
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
    
-   This library is distributed in the hope that it will be useful,
-   but WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   Lesser General Public License for more details.
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
    
-   You should have received a copy of the GNU Lesser General Public
-   License along with this library; if not, write to the Free Software
-   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+   
+   - Neither the name of the Xiph.org Foundation nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+   
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
+   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
 
-
-#include <stdlib.h>
 #include "cb_search.h"
 #include "filters.h"
-#include <math.h>
-#ifdef DEBUG
-#include <stdio.h>
-#endif
 #include "stack_alloc.h"
 #include "vq.h"
+#include "misc.h"
+
+#ifdef _USE_SSE
+#include "cb_search_sse.h"
+#elif defined(ARM4_ASM) || defined(ARM5E_ASM)
+#include "cb_search_arm4.h"
+#elif defined(BFIN_ASM)
+#include "cb_search_bfin.h"
+#endif
+
+#ifndef OVERRIDE_COMPUTE_WEIGHTED_CODEBOOK
+static void compute_weighted_codebook(const signed char *shape_cb, const spx_word16_t *r, spx_word16_t *resp, spx_word16_t *resp2, spx_word32_t *E, int shape_cb_size, int subvect_size, char *stack)
+{
+   int i, j, k;
+   VARDECL(spx_word16_t *shape);
+   ALLOC(shape, subvect_size, spx_word16_t);
+   for (i=0;i<shape_cb_size;i++)
+   {
+      spx_word16_t *res;
+      
+      res = resp+i*subvect_size;
+      for (k=0;k<subvect_size;k++)
+         shape[k] = (spx_word16_t)shape_cb[i*subvect_size+k];
+      E[i]=0;
+
+      /* Compute codeword response using convolution with impulse response */
+      for(j=0;j<subvect_size;j++)
+      {
+         spx_word32_t resj=0;
+         spx_word16_t res16;
+         for (k=0;k<=j;k++)
+            resj = MAC16_16(resj,shape[k],r[j-k]);
+#ifdef FIXED_POINT
+         res16 = EXTRACT16(SHR32(resj, 13));
+#else
+         res16 = 0.03125f*resj;
+#endif
+         /* Compute codeword energy */
+         E[i]=MAC16_16(E[i],res16,res16);
+         res[j] = res16;
+         /*printf ("%d\n", (int)res[j]);*/
+      }
+   }
+
+}
+#endif
+
+#ifndef OVERRIDE_TARGET_UPDATE
+static inline void target_update(spx_word16_t *t, spx_word16_t g, spx_word16_t *r, int len)
+{
+   int n;
+   for (n=0;n<len;n++)
+      t[n] = SUB16(t[n],PSHR32(MULT16_16(g,r[n]),13));
+}
+#endif
+
+
+
+static void split_cb_search_shape_sign_N1(
+spx_word16_t target[],                 /* target vector */
+spx_coef_t ak[],                       /* LPCs for this subframe */
+spx_coef_t awk1[],                     /* Weighted LPCs for this subframe */
+spx_coef_t awk2[],                     /* Weighted LPCs for this subframe */
+const void *par,                      /* Codebook/search parameters*/
+int   p,                        /* number of LPC coeffs */
+int   nsf,                      /* number of samples in subframe */
+spx_sig_t *exc,
+spx_word16_t *r,
+SpeexBits *bits,
+char *stack,
+int   update_target
+)
+{
+   int i,j,m,q;
+   VARDECL(spx_word16_t *resp);
+#ifdef _USE_SSE
+   VARDECL(__m128 *resp2);
+   VARDECL(__m128 *E);
+#else
+   spx_word16_t *resp2;
+   VARDECL(spx_word32_t *E);
+#endif
+   VARDECL(spx_word16_t *t);
+   VARDECL(spx_sig_t *e);
+   const signed char *shape_cb;
+   int shape_cb_size, subvect_size, nb_subvect;
+   const split_cb_params *params;
+   int best_index;
+   spx_word32_t best_dist;
+   int have_sign;
+   
+   params = (const split_cb_params *) par;
+   subvect_size = params->subvect_size;
+   nb_subvect = params->nb_subvect;
+   shape_cb_size = 1<<params->shape_bits;
+   shape_cb = params->shape_cb;
+   have_sign = params->have_sign;
+   ALLOC(resp, shape_cb_size*subvect_size, spx_word16_t);
+#ifdef _USE_SSE
+   ALLOC(resp2, (shape_cb_size*subvect_size)>>2, __m128);
+   ALLOC(E, shape_cb_size>>2, __m128);
+#else
+   resp2 = resp;
+   ALLOC(E, shape_cb_size, spx_word32_t);
+#endif
+   ALLOC(t, nsf, spx_word16_t);
+   ALLOC(e, nsf, spx_sig_t);
+   
+   /* FIXME: Do we still need to copy the target? */
+   for (i=0;i<nsf;i++)
+      t[i]=target[i];
+
+   compute_weighted_codebook(shape_cb, r, resp, resp2, E, shape_cb_size, subvect_size, stack);
+
+   for (i=0;i<nb_subvect;i++)
+   {
+      spx_word16_t *x=t+subvect_size*i;
+      /*Find new n-best based on previous n-best j*/
+      if (have_sign)
+         vq_nbest_sign(x, resp2, subvect_size, shape_cb_size, E, 1, &best_index, &best_dist, stack);
+      else
+         vq_nbest(x, resp2, subvect_size, shape_cb_size, E, 1, &best_index, &best_dist, stack);
+      
+      speex_bits_pack(bits,best_index,params->shape_bits+have_sign);
+      
+      {
+         int rind;
+         spx_word16_t *res;
+         spx_word16_t sign=1;
+         rind = best_index;
+         if (rind>=shape_cb_size)
+         {
+            sign=-1;
+            rind-=shape_cb_size;
+         }
+         res = resp+rind*subvect_size;
+         if (sign>0)
+            for (m=0;m<subvect_size;m++)
+               t[subvect_size*i+m] = SUB16(t[subvect_size*i+m], res[m]);
+         else
+            for (m=0;m<subvect_size;m++)
+               t[subvect_size*i+m] = ADD16(t[subvect_size*i+m], res[m]);
+
+#ifdef FIXED_POINT
+         if (sign)
+         {
+            for (j=0;j<subvect_size;j++)
+               e[subvect_size*i+j]=SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5);
+         } else {
+            for (j=0;j<subvect_size;j++)
+               e[subvect_size*i+j]=NEG32(SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5));
+         }
+#else
+         for (j=0;j<subvect_size;j++)
+            e[subvect_size*i+j]=sign*0.03125*shape_cb[rind*subvect_size+j];
+#endif
+      
+      }
+            
+      for (m=0;m<subvect_size;m++)
+      {
+         spx_word16_t g;
+         int rind;
+         spx_word16_t sign=1;
+         rind = best_index;
+         if (rind>=shape_cb_size)
+         {
+            sign=-1;
+            rind-=shape_cb_size;
+         }
+         
+         q=subvect_size-m;
+#ifdef FIXED_POINT
+         g=sign*shape_cb[rind*subvect_size+m];
+#else
+         g=sign*0.03125*shape_cb[rind*subvect_size+m];
+#endif
+         target_update(t+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
+      }
+   }
 
-#define min(a,b) ((a) < (b) ? (a) : (b))
-#define max(a,b) ((a) > (b) ? (a) : (b))
+   /* Update excitation */
+   /* FIXME: We could update the excitation directly above */
+   for (j=0;j<nsf;j++)
+      exc[j]=ADD32(exc[j],e[j]);
+   
+   /* Update target: only update target if necessary */
+   if (update_target)
+   {
+      VARDECL(spx_sig_t *r2);
+      ALLOC(r2, nsf, spx_sig_t);
+      syn_percep_zero(e, ak, awk1, awk2, r2, nsf,p, stack);
+      for (j=0;j<nsf;j++)
+         target[j]=SUB16(target[j],EXTRACT16(PSHR32(r2[j],8)));
+   }
+}
 
 
 
-void split_cb_search_nogain(
-float target[],                        /* target vector */
-float ak[],                    /* LPCs for this subframe */
-float awk1[],                  /* Weighted LPCs for this subframe */
-float awk2[],                  /* Weighted LPCs for this subframe */
-void *par,                      /* Codebook/search parameters*/
+void split_cb_search_shape_sign(
+spx_word16_t target[],                 /* target vector */
+spx_coef_t ak[],                       /* LPCs for this subframe */
+spx_coef_t awk1[],                     /* Weighted LPCs for this subframe */
+spx_coef_t awk2[],                     /* Weighted LPCs for this subframe */
+const void *par,                      /* Codebook/search parameters*/
 int   p,                        /* number of LPC coeffs */
 int   nsf,                      /* number of samples in subframe */
-float *exc,
+spx_sig_t *exc,
+spx_word16_t *r,
 SpeexBits *bits,
-float *stack,
-int   complexity
+char *stack,
+int   complexity,
+int   update_target
 )
 {
    int i,j,k,m,n,q;
-   float *resp;
-   float *t, *r, *e, *E;
-   /*FIXME: Should make this dynamic*/
-   float *tmp, *ot[20], *nt[20];
-   float *ndist, *odist;
-   int *itmp, *nind[20], *oind[20];
-
-   int *ind;
-   float *shape_cb;
+   VARDECL(spx_word16_t *resp);
+#ifdef _USE_SSE
+   VARDECL(__m128 *resp2);
+   VARDECL(__m128 *E);
+#else
+   spx_word16_t *resp2;
+   VARDECL(spx_word32_t *E);
+#endif
+   VARDECL(spx_word16_t *t);
+   VARDECL(spx_sig_t *e);
+   VARDECL(spx_sig_t *r2);
+   VARDECL(spx_word16_t *tmp);
+   VARDECL(spx_word32_t *ndist);
+   VARDECL(spx_word32_t *odist);
+   VARDECL(int *itmp);
+   VARDECL(spx_word16_t **ot2);
+   VARDECL(spx_word16_t **nt2);
+   spx_word16_t **ot, **nt;
+   VARDECL(int **nind);
+   VARDECL(int **oind);
+   VARDECL(int *ind);
+   const signed char *shape_cb;
    int shape_cb_size, subvect_size, nb_subvect;
-   split_cb_params *params;
+   const split_cb_params *params;
    int N=2;
-   int *best_index;
-   float *best_dist;
-
+   VARDECL(int *best_index);
+   VARDECL(spx_word32_t *best_dist);
+   VARDECL(int *best_nind);
+   VARDECL(int *best_ntarget);
+   int have_sign;
    N=complexity;
-   if (N<1)
-      N=1;
    if (N>10)
       N=10;
+   /* Complexity isn't as important for the codebooks as it is for the pitch */
+   N=(2*N)/3;
+   if (N<1)
+      N=1;
+   if (N==1)
+   {
+      split_cb_search_shape_sign_N1(target,ak,awk1,awk2,par,p,nsf,exc,r,bits,stack,update_target);
+      return;
+   }
+   ALLOC(ot2, N, spx_word16_t*);
+   ALLOC(nt2, N, spx_word16_t*);
+   ALLOC(oind, N, int*);
+   ALLOC(nind, N, int*);
 
-   params = (split_cb_params *) par;
+   params = (const split_cb_params *) par;
    subvect_size = params->subvect_size;
    nb_subvect = params->nb_subvect;
    shape_cb_size = 1<<params->shape_bits;
    shape_cb = params->shape_cb;
-   resp = PUSH(stack, shape_cb_size*subvect_size);
-   t = PUSH(stack, nsf);
-   r = PUSH(stack, nsf);
-   e = PUSH(stack, nsf);
-   E = PUSH(stack, shape_cb_size);
-   ind = (int*)PUSH(stack, nb_subvect);
-
-   tmp = PUSH(stack, 2*N*nsf);
+   have_sign = params->have_sign;
+   ALLOC(resp, shape_cb_size*subvect_size, spx_word16_t);
+#ifdef _USE_SSE
+   ALLOC(resp2, (shape_cb_size*subvect_size)>>2, __m128);
+   ALLOC(E, shape_cb_size>>2, __m128);
+#else
+   resp2 = resp;
+   ALLOC(E, shape_cb_size, spx_word32_t);
+#endif
+   ALLOC(t, nsf, spx_word16_t);
+   ALLOC(e, nsf, spx_sig_t);
+   ALLOC(r2, nsf, spx_sig_t);
+   ALLOC(ind, nb_subvect, int);
+
+   ALLOC(tmp, 2*N*nsf, spx_word16_t);
    for (i=0;i<N;i++)
    {
-      ot[i]=tmp;
-      tmp += nsf;
-      nt[i]=tmp;
-      tmp += nsf;
+      ot2[i]=tmp+2*i*nsf;
+      nt2[i]=tmp+(2*i+1)*nsf;
    }
-
-   best_index = (int*)PUSH(stack, N);
-   best_dist = PUSH(stack, N);
-   ndist = PUSH(stack, N);
-   odist = PUSH(stack, N);
+   ot=ot2;
+   nt=nt2;
+   ALLOC(best_index, N, int);
+   ALLOC(best_dist, N, spx_word32_t);
+   ALLOC(best_nind, N, int);
+   ALLOC(best_ntarget, N, int);
+   ALLOC(ndist, N, spx_word32_t);
+   ALLOC(odist, N, spx_word32_t);
    
-   itmp = (int*)PUSH(stack, 2*N*nb_subvect);
+   ALLOC(itmp, 2*N*nb_subvect, int);
    for (i=0;i<N;i++)
    {
-      nind[i]=itmp;
-      itmp+=nb_subvect;
-      oind[i]=itmp;
-      itmp+=nb_subvect;
-      for (j=0;j<nb_subvect;j++)
-         nind[i][j]=oind[i][j]=-1;
+      nind[i]=itmp+2*i*nb_subvect;
+      oind[i]=itmp+(2*i+1)*nb_subvect;
    }
-
-   for (j=0;j<N;j++)
-      for (i=0;i<nsf;i++)
-         ot[j][i]=target[i];
-
+   
    for (i=0;i<nsf;i++)
       t[i]=target[i];
 
-   e[0]=1;
-   for (i=1;i<nsf;i++)
-      e[i]=0;
-   residue_zero(e, awk1, r, nsf, p);
-   syn_filt_zero(r, ak, r, nsf, p);
-   syn_filt_zero(r, awk2, r, nsf,p);
-   
-   /* Pre-compute codewords response and energy */
-   for (i=0;i<shape_cb_size;i++)
-   {
-      float *res = resp+i*subvect_size;
+   for (j=0;j<N;j++)
+      speex_move(&ot[j][0], t, nsf*sizeof(spx_word16_t));
 
-      /* Compute codeword response */
-      int k;
-      for(j=0;j<subvect_size;j++)
-         res[j]=0;
-      for(j=0;j<subvect_size;j++)
-      {
-         for (k=j;k<subvect_size;k++)
-            res[k]+=shape_cb[i*subvect_size+j]*r[k-j];
-      }
-      E[i]=0;
-      for(j=0;j<subvect_size;j++)
-         E[i]+=res[j]*res[j];
-   }
+   /* Pre-compute codewords response and energy */
+   compute_weighted_codebook(shape_cb, r, resp, resp2, E, shape_cb_size, subvect_size, stack);
 
    for (j=0;j<N;j++)
       odist[j]=0;
+   
    /*For all subvectors*/
    for (i=0;i<nb_subvect;i++)
    {
       /*"erase" nbest list*/
       for (j=0;j<N;j++)
-         ndist[j]=-1;
+         ndist[j]=VERY_LARGE32;
 
       /*For all n-bests of previous subvector*/
       for (j=0;j<N;j++)
       {
-         float *x=ot[j]+subvect_size*i;
+         spx_word16_t *x=ot[j]+subvect_size*i;
+         spx_word32_t tener = 0;
+         for (m=0;m<subvect_size;m++)
+            tener = MAC16_16(tener, x[m],x[m]);
+#ifdef FIXED_POINT
+         tener = SHR32(tener,1);
+#else
+         tener *= .5;
+#endif
          /*Find new n-best based on previous n-best j*/
-         vq_nbest(x, resp, subvect_size, shape_cb_size, E, N, best_index, best_dist);
+         if (have_sign)
+            vq_nbest_sign(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
+         else
+            vq_nbest(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
 
          /*For all new n-bests*/
          for (k=0;k<N;k++)
          {
-            float err=0;
-            /*previous target*/
-            for (m=0;m<nsf;m++)
-               t[m]=ot[j][m];
-            /*update target*/
-            for (m=0;m<subvect_size;m++)
-            {
-               float g=shape_cb[best_index[k]*subvect_size+m];
-               for (n=subvect_size*i+m,q=0;n<nsf;n++,q++)
-                  t[n] -= g*r[q];
-            }
+            /* Compute total distance (including previous sub-vectors */
+            spx_word32_t err = ADD32(ADD32(odist[j],best_dist[k]),tener);
             
-            /*compute error (distance)*/
-            err=odist[j];
-            for (m=i*subvect_size;m<(i+1)*subvect_size;m++)
-               err += t[m]*t[m];
             /*update n-best list*/
-            if (err<ndist[N-1] || ndist[N-1]<-.5)
+            if (err<ndist[N-1])
             {
                for (m=0;m<N;m++)
                {
-                  if (err < ndist[m] || ndist[m]<-.5)
+                  if (err < ndist[m])
                   {
                      for (n=N-1;n>m;n--)
                      {
-                        for (q=0;q<nsf;q++)
-                           nt[n][q]=nt[n-1][q];
-                        for (q=0;q<nb_subvect;q++)
-                           nind[n][q]=nind[n-1][q];
-                        ndist[n]=ndist[n-1];
+                        ndist[n] = ndist[n-1];
+                        best_nind[n] = best_nind[n-1];
+                        best_ntarget[n] = best_ntarget[n-1];
                      }
-                     for (q=0;q<nsf;q++)
-                        nt[m][q]=t[q];
-                     for (q=0;q<nb_subvect;q++)
-                        nind[m][q]=oind[j][q];
-                     nind[m][i]=best_index[k];
-                     ndist[m]=err;
+                     ndist[m] = err;
+                     best_nind[n] = best_index[k];
+                     best_ntarget[n] = j;
                      break;
                   }
                }
             }
          }
          if (i==0)
-           break;
+            break;
       }
-
-      /*opdate old-new data*/
       for (j=0;j<N;j++)
-         for (m=0;m<nsf;m++)
-            ot[j][m]=nt[j][m];
+      {
+         /*previous target (we don't care what happened before*/
+         for (m=(i+1)*subvect_size;m<nsf;m++)
+            nt[j][m]=ot[best_ntarget[j]][m];
+         
+         /* New code: update the rest of the target only if it's worth it */
+         for (m=0;m<subvect_size;m++)
+         {
+            spx_word16_t g;
+            int rind;
+            spx_word16_t sign=1;
+            rind = best_nind[j];
+            if (rind>=shape_cb_size)
+            {
+               sign=-1;
+               rind-=shape_cb_size;
+            }
+
+            q=subvect_size-m;
+#ifdef FIXED_POINT
+            g=sign*shape_cb[rind*subvect_size+m];
+#else
+            g=sign*0.03125*shape_cb[rind*subvect_size+m];
+#endif
+            target_update(nt[j]+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
+         }
+
+         for (q=0;q<nb_subvect;q++)
+            nind[j][q]=oind[best_ntarget[j]][q];
+         nind[j][i]=best_nind[j];
+      }
+
+      /*update old-new data*/
+      /* just swap pointers instead of a long copy */
+      {
+         spx_word16_t **tmp2;
+         tmp2=ot;
+         ot=nt;
+         nt=tmp2;
+      }
       for (j=0;j<N;j++)
          for (m=0;m<nb_subvect;m++)
             oind[j][m]=nind[j][m];
@@ -219,343 +460,147 @@ int   complexity
    for (i=0;i<nb_subvect;i++)
    {
       ind[i]=nind[0][i];
-      speex_bits_pack(bits,ind[i],params->shape_bits);
+      speex_bits_pack(bits,ind[i],params->shape_bits+have_sign);
    }
    
    /* Put everything back together */
    for (i=0;i<nb_subvect;i++)
+   {
+      int rind;
+      spx_word16_t sign=1;
+      rind = ind[i];
+      if (rind>=shape_cb_size)
+      {
+         sign=-1;
+         rind-=shape_cb_size;
+      }
+#ifdef FIXED_POINT
+      if (sign==1)
+      {
+         for (j=0;j<subvect_size;j++)
+            e[subvect_size*i+j]=SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5);
+      } else {
+         for (j=0;j<subvect_size;j++)
+            e[subvect_size*i+j]=NEG32(SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5));
+      }
+#else
       for (j=0;j<subvect_size;j++)
-         e[subvect_size*i+j]=shape_cb[ind[i]*subvect_size+j];
-
+         e[subvect_size*i+j]=sign*0.03125*shape_cb[rind*subvect_size+j];
+#endif
+   }   
    /* Update excitation */
    for (j=0;j<nsf;j++)
-      exc[j]+=e[j];
+      exc[j]=ADD32(exc[j],e[j]);
    
-   /* Update target */
-   residue_zero(e, awk1, r, nsf, p);
-   syn_filt_zero(r, ak, r, nsf, p);
-   syn_filt_zero(r, awk2, r, nsf,p);
-   for (j=0;j<nsf;j++)
-      target[j]-=r[j];
-
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
+   /* Update target: only update target if necessary */
+   if (update_target)
+   {
+      syn_percep_zero(e, ak, awk1, awk2, r2, nsf,p, stack);
+      for (j=0;j<nsf;j++)
+         target[j]=SUB16(target[j],EXTRACT16(PSHR32(r2[j],8)));
+   }
 }
 
 
-
-
-
-void split_cb_search_shape_sign(
-float target[],                        /* target vector */
-float ak[],                    /* LPCs for this subframe */
-float awk1[],                  /* Weighted LPCs for this subframe */
-float awk2[],                  /* Weighted LPCs for this subframe */
-void *par,                      /* Codebook/search parameters*/
-int   p,                        /* number of LPC coeffs */
+void split_cb_shape_sign_unquant(
+spx_sig_t *exc,
+const void *par,                      /* non-overlapping codebook */
 int   nsf,                      /* number of samples in subframe */
-float *exc,
 SpeexBits *bits,
-float *stack,
-int   complexity
+char *stack,
+spx_int32_t *seed
 )
 {
    int i,j;
-   float *resp;
-   float *t, *r, *e, *E;
-   int *ind, *signs;
-   float *shape_cb;
+   VARDECL(int *ind);
+   VARDECL(int *signs);
+   const signed char *shape_cb;
    int shape_cb_size, subvect_size, nb_subvect;
-   split_cb_params *params;
+   const split_cb_params *params;
+   int have_sign;
 
-   params = (split_cb_params *) par;
+   params = (const split_cb_params *) par;
    subvect_size = params->subvect_size;
    nb_subvect = params->nb_subvect;
    shape_cb_size = 1<<params->shape_bits;
    shape_cb = params->shape_cb;
-   resp = PUSH(stack, shape_cb_size*subvect_size);
-   t = PUSH(stack, nsf);
-   r = PUSH(stack, nsf);
-   e = PUSH(stack, nsf);
-   E = PUSH(stack, shape_cb_size);
-   ind = (int*)PUSH(stack, nb_subvect);
-   signs = (int*)PUSH(stack, nb_subvect);
+   have_sign = params->have_sign;
 
-   for (i=0;i<nsf;i++)
-      t[i]=target[i];
-
-   e[0]=1;
-   for (i=1;i<nsf;i++)
-      e[i]=0;
-   residue_zero(e, awk1, r, nsf, p);
-   syn_filt_zero(r, ak, r, nsf, p);
-   syn_filt_zero(r, awk2, r, nsf,p);
-   
-   /* Pre-compute codewords response and energy */
-   for (i=0;i<shape_cb_size;i++)
-   {
-      float *res = resp+i*subvect_size;
-
-      /* Compute codeword response */
-      int k;
-      for(j=0;j<subvect_size;j++)
-         res[j]=0;
-      for(j=0;j<subvect_size;j++)
-      {
-         for (k=j;k<subvect_size;k++)
-            res[k]+=shape_cb[i*subvect_size+j]*r[k-j];
-      }
-      E[i]=0;
-      for(j=0;j<subvect_size;j++)
-         E[i]+=res[j]*res[j];
-   }
+   ALLOC(ind, nb_subvect, int);
+   ALLOC(signs, nb_subvect, int);
 
+   /* Decode codewords and gains */
    for (i=0;i<nb_subvect;i++)
    {
-      int best_index[2]={0,0}, k, m;
-      float g, dist, best_dist[2]={-1,-1}, best_sign[2]={0,0};
-      float *a, *x;
-      float energy=0;
-      x=t+subvect_size*i;
-
-      for (k=0;k<subvect_size;k++)
-         energy+=x[k]*x[k];
-      /* Find best codeword for current sub-vector */
-      for (j=0;j<shape_cb_size;j++)
-      {
-         int sign;
-         dist=0;
-         a=resp+j*subvect_size;
-         dist=0;
-         for (k=0;k<subvect_size;k++)
-            dist -= 2*a[k]*x[k];
-         if (dist > 0)
-         {
-            sign=1;
-            dist =- dist;
-         } else
-            sign=0;
-         dist += energy+E[j];
-         if (dist<best_dist[0] || best_dist[0]<0)
-         {
-            best_dist[1]=best_dist[0];
-            best_index[1]=best_index[0];
-            best_sign[1]=best_sign[0];
-            best_dist[0]=dist;
-            best_index[0]=j;
-            best_sign[0]=sign;
-         } else if (dist<best_dist[1] || best_dist[1]<0)
-         {
-            best_dist[1]=dist;
-            best_index[1]=j;
-            best_sign[1]=sign;
-         }
-      }
-      if (i<nb_subvect-1)
-      {
-         int nbest;
-         float *tt, err[2];
-         float best_score[2];
-         tt=PUSH(stack,nsf);
-         for (nbest=0;nbest<2;nbest++)
-         {
-            float s=1;
-            if (best_sign[nbest])
-               s=-1;
-            for (j=0;j<nsf;j++)
-               tt[j]=t[j];
-            for (j=0;j<subvect_size;j++)
-            {
-               g=s*shape_cb[best_index[nbest]*subvect_size+j];
-               for (k=subvect_size*i+j,m=0;k<nsf;k++,m++)
-                  tt[k] -= g*r[m];
-            }
-            
-            {
-               int best_index2=0, best_sign2=0, sign2;
-               float  best_dist2=0;
-               x=t+subvect_size*(i+1);
-               for (j=0;j<shape_cb_size;j++)
-               {
-                  a=resp+j*subvect_size;
-                  dist = 0;
-                  for (k=0;k<subvect_size;k++)
-                     dist -= 2*a[k]*x[k];
-                  if (dist > 0)
-                  {
-                     sign2=1;
-                     dist =- dist;
-                  } else
-                     sign2=0;
-                  dist += energy+E[j];
-                  if (dist<best_dist2 || j==0)
-                  {
-                     best_dist2=dist;
-                     best_index2=j;
-                     best_sign2=sign2;
-                  }
-               }
-               s=1;
-               if (best_sign2)
-                  s=-1;
-               /*int i2=vq_index(&tt[subvect_size*(i+1)], resp, subvect_size, shape_cb_size);*/
-               
-               for (j=0;j<subvect_size;j++)
-               {
-                  g=s*shape_cb[best_index2*subvect_size+j];
-                  for (k=subvect_size*(i+1)+j,m=0;k<nsf;k++,m++)
-                     tt[k] -= g*r[m];
-               }
-            }
-
-            err[nbest]=0;
-            for (j=subvect_size*i;j<subvect_size*(i+2);j++)
-               err[nbest]-=tt[j]*tt[j];
-            
-            best_score[nbest]=err[nbest];
-         }
-
-         if (best_score[1]>best_score[0])
-         {
-            best_sign[0]=best_sign[1];
-            best_index[0]=best_index[1];
-            best_score[0]=best_score[1];
-         }
-         POP(stack);
-
-      }
-
-      ind[i]=best_index[0];
-      signs[i] = best_sign[0];
-
-      /*printf ("best index: %d/%d\n", best_index, shape_cb_size);*/
-      speex_bits_pack(bits,signs[i],1);
-      speex_bits_pack(bits,ind[i],params->shape_bits);
-
-      /* Update target for next subvector */
-      for (j=0;j<subvect_size;j++)
-      {
-         g=shape_cb[ind[i]*subvect_size+j];
-         if (signs[i])
-            g=-g;
-         for (k=subvect_size*i+j,m=0;k<nsf;k++,m++)
-            t[k] -= g*r[m];
-      }
+      if (have_sign)
+         signs[i] = speex_bits_unpack_unsigned(bits, 1);
+      else
+         signs[i] = 0;
+      ind[i] = speex_bits_unpack_unsigned(bits, params->shape_bits);
    }
-   
-   /* Put everything back together */
+   /* Compute decoded excitation */
    for (i=0;i<nb_subvect;i++)
    {
-      float s=1;
+      spx_word16_t s=1;
       if (signs[i])
          s=-1;
+#ifdef FIXED_POINT
+      if (s==1)
+      {
+         for (j=0;j<subvect_size;j++)
+            exc[subvect_size*i+j]=SHL32(EXTEND32(shape_cb[ind[i]*subvect_size+j]),SIG_SHIFT-5);
+      } else {
+         for (j=0;j<subvect_size;j++)
+            exc[subvect_size*i+j]=NEG32(SHL32(EXTEND32(shape_cb[ind[i]*subvect_size+j]),SIG_SHIFT-5));
+      }
+#else
       for (j=0;j<subvect_size;j++)
-         e[subvect_size*i+j]=s*shape_cb[ind[i]*subvect_size+j];
+         exc[subvect_size*i+j]+=s*0.03125*shape_cb[ind[i]*subvect_size+j];      
+#endif
    }
-   /* Update excitation */
-   for (j=0;j<nsf;j++)
-      exc[j]+=e[j];
-   
-   /* Update target */
-   residue_zero(e, awk1, r, nsf, p);
-   syn_filt_zero(r, ak, r, nsf, p);
-   syn_filt_zero(r, awk2, r, nsf,p);
-   for (j=0;j<nsf;j++)
-      target[j]-=r[j];
-
-   
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
-   POP(stack);
 }
 
-
-void split_cb_nogain_unquant(
-float *exc,
-void *par,                      /* non-overlapping codebook */
+void noise_codebook_quant(
+spx_word16_t target[],                 /* target vector */
+spx_coef_t ak[],                       /* LPCs for this subframe */
+spx_coef_t awk1[],                     /* Weighted LPCs for this subframe */
+spx_coef_t awk2[],                     /* Weighted LPCs for this subframe */
+const void *par,                      /* Codebook/search parameters*/
+int   p,                        /* number of LPC coeffs */
 int   nsf,                      /* number of samples in subframe */
+spx_sig_t *exc,
+spx_word16_t *r,
 SpeexBits *bits,
-float *stack
+char *stack,
+int   complexity,
+int   update_target
 )
 {
-   int i,j;
-   int *ind;
-   float *shape_cb;
-   int shape_cb_size, subvect_size, nb_subvect;
-   split_cb_params *params;
-
-   params = (split_cb_params *) par;
-   subvect_size = params->subvect_size;
-   nb_subvect = params->nb_subvect;
-   shape_cb_size = 1<<params->shape_bits;
-   shape_cb = params->shape_cb;
-   
-   ind = (int*)PUSH(stack, nb_subvect);
-
-   /* Decode codewords and gains */
-   for (i=0;i<nb_subvect;i++)
-      ind[i] = speex_bits_unpack_unsigned(bits, params->shape_bits);
-
-   /* Compute decoded excitation */
-   for (i=0;i<nb_subvect;i++)
-      for (j=0;j<subvect_size;j++)
-         exc[subvect_size*i+j]+=shape_cb[ind[i]*subvect_size+j];
+   int i;
+   VARDECL(spx_sig_t *tmp);
+   ALLOC(tmp, nsf, spx_sig_t);
+   for (i=0;i<nsf;i++)
+      tmp[i]=PSHR32(EXTEND32(target[i]),SIG_SHIFT);
+   residue_percep_zero(tmp, ak, awk1, awk2, tmp, nsf, p, stack);
 
-   POP(stack);
+   for (i=0;i<nsf;i++)
+      exc[i]+=tmp[i];
+   for (i=0;i<nsf;i++)
+      target[i]=0;
 }
 
-void split_cb_shape_sign_unquant(
-float *exc,
-void *par,                      /* non-overlapping codebook */
+
+void noise_codebook_unquant(
+spx_sig_t *exc,
+const void *par,                      /* non-overlapping codebook */
 int   nsf,                      /* number of samples in subframe */
 SpeexBits *bits,
-float *stack
+char *stack,
+spx_int32_t *seed
 )
 {
-   int i,j;
-   int *ind, *signs;
-   float *shape_cb;
-   int shape_cb_size, subvect_size, nb_subvect;
-   split_cb_params *params;
-
-   params = (split_cb_params *) par;
-   subvect_size = params->subvect_size;
-   nb_subvect = params->nb_subvect;
-   shape_cb_size = 1<<params->shape_bits;
-   shape_cb = params->shape_cb;
-   
-   ind = (int*)PUSH(stack, nb_subvect);
-   signs = (int*)PUSH(stack, nb_subvect);
-
-   /* Decode codewords and gains */
-   for (i=0;i<nb_subvect;i++)
-   {
-      signs[i] = speex_bits_unpack_unsigned(bits, 1);
-      ind[i] = speex_bits_unpack_unsigned(bits, params->shape_bits);
-   }
-   /* Compute decoded excitation */
-   for (i=0;i<nb_subvect;i++)
-   {
-      float s=1;
-      if (signs[i])
-         s=-1;
-      for (j=0;j<subvect_size;j++)
-         exc[subvect_size*i+j]+=s*shape_cb[ind[i]*subvect_size+j];
-   }
-   POP(stack);
-   POP(stack);
+   int i;
+   /* FIXME: This is bad, but I don't think the function ever gets called anyway */
+   for (i=0;i<nsf;i++)
+      exc[i]=SHL32(EXTEND32(speex_rand(1, seed)),SIG_SHIFT);
 }