void mp_shiftr1(digit_t* x, const unsigned int nwords)
{ // Multiprecision right shift by one.
unsigned int i;
for (i = 0; i < nwords-1; i++) {
SHIFTR(x[i+1], x[i], 1, x[i], RADIX);
}
x[nwords-1] >>= 1;
}
void fdct_int32( short *const In )
{
short *pIn;
int i;
pIn = In;
for(i=8; i>0; --i)
{
int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill;
// even
LOAD_BUTF(mm1,mm6, 1, 6, mm0, pIn);
LOAD_BUTF(mm2,mm5, 2, 5, mm0, pIn);
LOAD_BUTF(mm3,mm4, 3, 4, mm0, pIn);
LOAD_BUTF(mm0,mm7, 0, 7, Spill, pIn);
BUTF(mm1, mm2, Spill);
BUTF(mm0, mm3, Spill);
ROTATE(mm3, mm2, ROT6_C, ROT6_SmC, -ROT6_SpC, Spill, FIX-FPASS, HALF(FIX-FPASS));
pIn[2] = mm3;
pIn[6] = mm2;
BUTF(mm0, mm1, Spill);
pIn[0] = SHIFTL(mm0, FPASS);
pIn[4] = SHIFTL(mm1, FPASS);
// odd
mm3 = mm5 + mm7;
mm2 = mm4 + mm6;
ROTATE(mm2, mm3, ROT17_C, -ROT17_SpC, -ROT17_SmC, mm0, FIX-FPASS, HALF(FIX-FPASS));
ROTATE(mm4, mm7, -ROT37_C, ROT37_SpC, ROT37_SmC, mm0, FIX-FPASS, HALF(FIX-FPASS));
mm7 += mm3;
mm4 += mm2;
pIn[1] = mm7;
pIn[7] = mm4;
ROTATE(mm5, mm6, -ROT13_C, ROT13_SmC, ROT13_SpC, mm0, FIX-FPASS, HALF(FIX-FPASS));
mm5 += mm3;
mm6 += mm2;
pIn[3] = mm6;
pIn[5] = mm5;
pIn += 8;
}
pIn = In;
for(i=8; i>0; --i)
{
int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, Spill;
// even
LOAD_BUTF(mm1,mm6, 1*8, 6*8, mm0, pIn);
LOAD_BUTF(mm2,mm5, 2*8, 5*8, mm0, pIn);
BUTF(mm1, mm2, mm0);
LOAD_BUTF(mm3,mm4, 3*8, 4*8, mm0, pIn);
LOAD_BUTF(mm0,mm7, 0*8, 7*8, Spill, pIn);
BUTF(mm0, mm3, Spill);
ROTATE(mm3, mm2, ROT6_C, ROT6_SmC, -ROT6_SpC, Spill, 0, HALF(FIX+FPASS+3));
pIn[2*8] = (int16_t)SHIFTR(mm3,FIX+FPASS+3);
pIn[6*8] = (int16_t)SHIFTR(mm2,FIX+FPASS+3);
mm0 += HALF(FPASS+3) - 1;
BUTF(mm0, mm1, Spill);
pIn[0*8] = (int16_t)SHIFTR(mm0, FPASS+3);
pIn[4*8] = (int16_t)SHIFTR(mm1, FPASS+3);
// odd
mm3 = mm5 + mm7;
mm2 = mm4 + mm6;
ROTATE(mm2, mm3, ROT17_C, -ROT17_SpC, -ROT17_SmC, mm0, 0, HALF(FIX+FPASS+3));
ROTATE2(mm4, mm7, -ROT37_C, ROT37_SpC, ROT37_SmC, mm0);
mm7 += mm3;
mm4 += mm2;
pIn[7*8] = (int16_t)SHIFTR(mm4,FIX+FPASS+3);
pIn[1*8] = (int16_t)SHIFTR(mm7,FIX+FPASS+3);
ROTATE2(mm5, mm6, -ROT13_C, ROT13_SmC, ROT13_SpC, mm0);
mm5 += mm3;
mm6 += mm2;
pIn[5*8] = (int16_t)SHIFTR(mm5,FIX+FPASS+3);
pIn[3*8] = (int16_t)SHIFTR(mm6,FIX+FPASS+3);
pIn++;
}
}
