Making decode_pulses() compute the L2-norm on the fly

3 files changed, 31 insertions, 19 deletions

diff --git a/celt/cwrs.c b/celt/cwrs.c
index ad980cc7..2fa9f89c 100644
--- a/celt/cwrs.c
+++ b/celt/cwrs.c
@@ -460,10 +460,12 @@ void encode_pulses(const int *_y,int _n,int _k,ec_enc *_enc){
   ec_enc_uint(_enc,icwrs(_n,_y),CELT_PVQ_V(_n,_k));
 }
 
-static void cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
+static opus_val32 cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
   opus_uint32 p;
   int         s;
   int         k0;
+  opus_int16  val;
+  opus_val32  yy=0;
   celt_assert(_k>0);
   celt_assert(_n>1);
   while(_n>2){
@@ -487,7 +489,9 @@ static void cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
       }
       else for(p=row[_k];p>_i;p=row[_k])_k--;
       _i-=p;
-      *_y++=(k0-_k+s)^s;
+      val=(k0-_k+s)^s;
+      *_y++=val;
+      yy=MAC16_16(yy,val,val);
     }
     /*Lots of dimensions case:*/
     else{
@@ -507,7 +511,9 @@ static void cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
         do p=CELT_PVQ_U_ROW[--_k][_n];
         while(p>_i);
         _i-=p;
-        *_y++=(k0-_k+s)^s;
+        val=(k0-_k+s)^s;
+        *_y++=val;
+        yy=MAC16_16(yy,val,val);
       }
     }
     _n--;
@@ -519,14 +525,19 @@ static void cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
   k0=_k;
   _k=(_i+1)>>1;
   if(_k)_i-=2*_k-1;
-  *_y++=(k0-_k+s)^s;
+  val=(k0-_k+s)^s;
+  *_y++=val;
+  yy=MAC16_16(yy,val,val);
   /*_n==1*/
   s=-(int)_i;
-  *_y=(_k+s)^s;
+  val=(_k+s)^s;
+  *_y=val;
+  yy=MAC16_16(yy,val,val);
+  return yy;
 }
 
-void decode_pulses(int *_y,int _n,int _k,ec_dec *_dec){
-  cwrsi(_n,_k,ec_dec_uint(_dec,CELT_PVQ_V(_n,_k)),_y);
+opus_val32 decode_pulses(int *_y,int _n,int _k,ec_dec *_dec){
+  return cwrsi(_n,_k,ec_dec_uint(_dec,CELT_PVQ_V(_n,_k)),_y);
 }
 
 #else /* SMALL_FOOTPRINT */
@@ -591,8 +602,10 @@ static opus_uint32 ncwrs_urow(unsigned _n,unsigned _k,opus_uint32 *_u){
   _y: Returns the vector of pulses.
   _u: Must contain entries [0..._k+1] of row _n of U() on input.
       Its contents will be destructively modified.*/
-static void cwrsi(int _n,int _k,opus_uint32 _i,int *_y,opus_uint32 *_u){
+static opus_val32 cwrsi(int _n,int _k,opus_uint32 _i,int *_y,opus_uint32 *_u){
   int j;
+  opus_int16 val;
+  opus_val32 yy=0;
   celt_assert(_n>0);
   j=0;
   do{
@@ -607,10 +620,13 @@ static void cwrsi(int _n,int _k,opus_uint32 _i,int *_y,opus_uint32 *_u){
     while(p>_i)p=_u[--_k];
     _i-=p;
     yj-=_k;
-    _y[j]=(yj+s)^s;
+    val=(yj+s)^s;
+    _y[j]=val;
+    yy=MAC16_16(yy,val,val);
     uprev(_u,_k+2,0);
   }
   while(++j<_n);
+  return yy;
 }
 
 /*Returns the index of the given combination of K elements chosen from a set
@@ -685,13 +701,15 @@ void encode_pulses(const int *_y,int _n,int _k,ec_enc *_enc){
   RESTORE_STACK;
 }
 
-void decode_pulses(int *_y,int _n,int _k,ec_dec *_dec){
+opus_val32 decode_pulses(int *_y,int _n,int _k,ec_dec *_dec){
   VARDECL(opus_uint32,u);
+  int ret;
   SAVE_STACK;
   celt_assert(_k>0);
   ALLOC(u,_k+2U,opus_uint32);
-  cwrsi(_n,_k,ec_dec_uint(_dec,ncwrs_urow(_n,_k,u)),_y,u);
+  ret = cwrsi(_n,_k,ec_dec_uint(_dec,ncwrs_urow(_n,_k,u)),_y,u);
   RESTORE_STACK;
+  return ret;
 }
 
 #endif /* SMALL_FOOTPRINT */
diff --git a/celt/cwrs.h b/celt/cwrs.h
index 7dfbd076..7cd47174 100644
--- a/celt/cwrs.h
+++ b/celt/cwrs.h
@@ -43,6 +43,6 @@ void get_required_bits(opus_int16 *bits, int N, int K, int frac);
 
 void encode_pulses(const int *_y, int N, int K, ec_enc *enc);
 
-void decode_pulses(int *_y, int N, int K, ec_dec *dec);
+opus_val32 decode_pulses(int *_y, int N, int K, ec_dec *dec);
 
 #endif /* CWRS_H */
diff --git a/celt/vq.c b/celt/vq.c
index 914ec861..a064df00 100644
--- a/celt/vq.c
+++ b/celt/vq.c
@@ -325,7 +325,6 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc
 unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
       ec_dec *dec, opus_val16 gain)
 {
-   int i;
    opus_val32 Ryy;
    unsigned collapse_mask;
    VARDECL(int, iy);
@@ -334,12 +333,7 @@ unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
    celt_assert2(K>0, "alg_unquant() needs at least one pulse");
    celt_assert2(N>1, "alg_unquant() needs at least two dimensions");
    ALLOC(iy, N, int);
-   decode_pulses(iy, N, K, dec);
-   Ryy = 0;
-   i=0;
-   do {
-      Ryy = MAC16_16(Ryy, iy[i], iy[i]);
-   } while (++i < N);
+   Ryy = decode_pulses(iy, N, K, dec);
    normalise_residual(iy, X, N, Ryy, gain);
    exp_rotation(X, N, -1, B, K, spread);
    collapse_mask = extract_collapse_mask(iy, N, B);