int signal_energy_nodc(int *input,unsigned int length) {
int i;
int temp,temp2;
register __m64 mm0,mm1,mm2,mm3;
__m64 *in = (__m64 *)input;
#ifdef MAIN
short *printb;
#endif
mm0 = _m_pxor(mm0,mm0);
mm3 = _m_pxor(mm3,mm3);
for (i=0;i<length>>1;i++) {
mm1 = in[i];
mm2 = mm1;
mm1 = _m_pmaddwd(mm1,mm1);// SIMD complex multiplication
mm1 = _m_psradi(mm1,shift);
mm0 = _m_paddd(mm0,mm1);
// temp2 = mm0;
// printf("%d %d\n",((int *)&in[i])[0],((int *)&in[i])[1]);
// printb = (short *)&mm2;
// printf("mm2 %d : %d %d %d %d\n",i,printb[0],printb[1],printb[2],printb[3]);
}
/*
#ifdef MAIN
printb = (short *)&mm3;
printf("%d %d %d %d\n",printb[0],printb[1],printb[2],printb[3]);
#endif
*/
mm1 = mm0;
mm0 = _m_psrlqi(mm0,32);
mm0 = _m_paddd(mm0,mm1);
temp = _m_to_int(mm0);
temp/=length;
temp<<=shift; // this is the average of x^2
#ifdef MAIN
printf("E x^2 = %d\n",temp);
#endif
_mm_empty();
_m_empty();
return((temp>0)?temp:1);
}
void main()
{
a = _m_paddd( b, c );
printf( "m1="AS_DWORDS"\n"
"m2="AS_DWORDS"\n"
"mm="AS_DWORDS"\n",
b._32[1], b._32[0],
c._32[1], c._32[0],
a._32[1], a._32[0] );
}
int signal_energy(int *input,unsigned int length) {
int i;
int temp,temp2;
register __m64 mm0,mm1,mm2,mm3;
__m64 *in = (__m64 *)input;
#ifdef MAIN
short *printb;
#endif
mm0 = _m_pxor(mm0,mm0);
mm3 = _m_pxor(mm3,mm3);
for (i=0;i<length>>1;i++) {
mm1 = in[i];
mm2 = mm1;
mm1 = _m_pmaddwd(mm1,mm1);
mm1 = _m_psradi(mm1,shift);// shift any 32 bits blocs of the word by the value shift
mm0 = _m_paddd(mm0,mm1);// add the two 64 bits words 4 bytes by 4 bytes
// temp2 = mm0;
// printf("%d %d\n",((int *)&temp2)[0],((int *)&temp2)[1]);
// printb = (short *)&mm2;
// printf("mm2 %d : %d %d %d %d\n",i,printb[0],printb[1],printb[2],printb[3]);
mm2 = _m_psrawi(mm2,shift_DC);
mm3 = _m_paddw(mm3,mm2);// add the two 64 bits words 2 bytes by 2 bytes
// printb = (short *)&mm3;
// printf("mm3 %d : %d %d %d %d\n",i,printb[0],printb[1],printb[2],printb[3]);
}
/*
#ifdef MAIN
printb = (short *)&mm3;
printf("%d %d %d %d\n",printb[0],printb[1],printb[2],printb[3]);
#endif
*/
mm1 = mm0;
mm0 = _m_psrlqi(mm0,32);
mm0 = _m_paddd(mm0,mm1);
temp = _m_to_int(mm0);
temp/=length;
temp<<=shift; // this is the average of x^2
// now remove the DC component
mm2 = _m_psrlqi(mm3,32);
mm2 = _m_paddw(mm2,mm3);
mm2 = _m_pmaddwd(mm2,mm2);
temp2 = _m_to_int(mm2);
temp2/=(length*length);
temp2<<=(2*shift_DC);
#ifdef MAIN
printf("E x^2 = %d\n",temp);
#endif
temp -= temp2;
#ifdef MAIN
printf("(E x)^2=%d\n",temp2);
#endif
_mm_empty();
_m_empty();
return((temp>0)?temp:1);
}