fixed-point: simplifying the arithmetic in alg_quant()
authorJean-Marc Valin <Jean-Marc.Valin@csiro.au>
Thu, 13 Mar 2008 00:18:15 +0000 (11:18 +1100)
committerJean-Marc Valin <Jean-Marc.Valin@csiro.au>
Thu, 13 Mar 2008 00:18:15 +0000 (11:18 +1100)
libcelt/fixed_generic.h
libcelt/mathops.h
libcelt/vq.c

index f0490a1..44b1f79 100644 (file)
@@ -41,6 +41,8 @@
 
 #define MULT32_32_Q31(a,b) ADD32(ADD32(SHL(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),15)), SHR(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),15))
 
+#define MULT32_32_Q32(a,b) ADD32(ADD32(MULT16_16(SHR((a),16),SHR((b),16)), SHR(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),16)), SHR(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),16))
+
 
 #define QCONST16(x,bits) ((celt_word16_t)(.5+(x)*(((celt_word32_t)1)<<(bits))))
 #define QCONST32(x,bits) ((celt_word32_t)(.5+(x)*(((celt_word32_t)1)<<(bits))))
index 94cd2ba..726d718 100644 (file)
@@ -164,7 +164,7 @@ static inline celt_word32_t celt_exp2(celt_word16_t x)
    return VSHR32(EXTEND32(frac), -integer-2);
 }
 
-static inline celt_word32_t celt_rcp(celt_word16_t x)
+static inline celt_word32_t celt_rcp(celt_word32_t x)
 {
    int i, neg=0;
    celt_word16_t n, frac;
@@ -174,15 +174,13 @@ static inline celt_word32_t celt_rcp(celt_word16_t x)
       neg = 1;
       x = NEG16(x);
    }
-   if (x==0)
-      return 0;
    i = celt_ilog2(x);
    n = VSHR32(x,i-16)-SHL32(EXTEND32(3),15);
    frac = ADD16(C[0], MULT16_16_Q15(n, ADD16(C[1], MULT16_16_Q15(n, ADD16(C[2], 
                 MULT16_16_Q15(n, ADD16(C[3], MULT16_16_Q15(n, (C[4])))))))));
    if (neg)
       frac = -frac;
-   return SHL32(EXTEND32(frac),16-i);
+   return VSHR32(EXTEND32(frac),i-16);
 }
 
 #endif /* FIXED_POINT */
index f265a88..7c89d02 100644 (file)
@@ -218,24 +218,24 @@ void alg_quant(celt_norm_t *X, celt_mask_t *W, int N, int K, const celt_norm_t *
                if (iy[m][j]*sign < 0)
                   continue;
 
-               spj = MULT16_16_P14(s, P[j]);
-               aspj = MULT16_16_P15(alpha, spj);
+               spj = MULT16_16_Q14(s, P[j]);
+               aspj = MULT16_16_Q15(alpha, spj);
                /* Updating the sums of the new pulse(s) */
-               Rxy = xy[m] + MULT16_16(s,X[j])     - MULT16_16(MULT16_16_P15(alpha,spj),Rxp);
+               Rxy = xy[m] + MULT16_16(s,X[j])     - MULT16_16(MULT16_16_Q15(alpha,spj),Rxp);
                Ryy = yy[m] + 2*MULT16_16(s,y[m][j]) + MULT16_16(s,s)   +MULT16_16(aspj,MULT16_16_Q14(aspj,Rpp)) - 2*MULT16_32_Q14(aspj,yp[m]) - 2*MULT16_16(s,MULT16_16_Q14(aspj,P[j]));
                Ryp = yp[m] + MULT16_16(spj, SUB16(QCONST16(1.f,14),MULT16_16_Q15(alpha,Rpp)));
                
                /* Compute the gain such that ||p + g*y|| = 1 */
-               g = MULT32_32_Q31(
-                     SHL32(celt_sqrt(MULT16_16(ROUND(Ryp,14),ROUND(Ryp,14)) + Ryy -
-                                     MULT16_16(ROUND(Ryy,14),Rpp))
-                           - ROUND(Ryp,14), 14),
-                     celt_rcp(ROUND(Ryy,14)));
+               g = MULT16_32_Q15(
+                        celt_sqrt(MULT16_16(ROUND(Ryp,14),ROUND(Ryp,14)) + Ryy -
+                                  MULT16_16(ROUND(Ryy,14),Rpp))
+                        - ROUND(Ryp,14),
+                   celt_rcp(SHR32(Ryy,12)));
                /* Knowing that gain, what's the error: (x-g*y)^2 
                   (result is negated and we discard x^2 because it's constant) */
                /*score = 2.f*g*Rxy - 1.f*g*g*Ryy*NORM_SCALING_1;*/
-               score = 2*MULT16_32_Q14(ROUND(Rxy,14),g) -
-                     MULT16_32_Q14(EXTRACT16(MULT16_32_Q14(ROUND(Ryy,14),g)),g);
+               score = 2*MULT16_32_Q14(ROUND(Rxy,14),g)
+                       - MULT16_32_Q14(EXTRACT16(MULT16_32_Q14(ROUND(Ryy,14),g)),g);
 
                if (score>nbest[Lupdate-1]->score)
                {