static inline void TRANSFORM(FFTComplex * z, unsigned int n, FFTSample wre, FFTSample wim)
{
register FFTSample t1,t2,t5,t6,r_re,r_im;
r_re = z[n*2].re;
r_im = z[n*2].im;
XPROD31_R(r_re, r_im, wre, wim, t1,t2);
r_re = z[n*3].re;
r_im = z[n*3].im;
XNPROD31_R(r_re, r_im, wre, wim, t5,t6);
BUTTERFLIES(z[0],z[n],z[n*2],z[n*3]);
}
void ff_imdct_half(unsigned int nbits, fixed32 *output, const fixed32 *input)
{
int n8, n4, n2, n, j;
const fixed32 *in1, *in2;
n = 1 << nbits;
n2 = n >> 1;
n4 = n >> 2;
n8 = n >> 3;
FFTComplex *z = (FFTComplex *)output;
/* pre rotation */
in1 = input;
in2 = input + n2 - 1;
/* revtab comes from the fft; revtab table is sized for N=4096 size fft = 2^12.
The fft is size N/4 so s->nbits-2, so our shift needs to be (12-(nbits-2)) */
const int revtab_shift = (14- nbits);
/* bitreverse reorder the input and rotate; result here is in OUTPUT ... */
/* (note that when using the current split radix, the bitreverse ordering is
complex, meaning that this reordering cannot easily be done in-place) */
/* Using the following pdf, you can see that it is possible to rearrange
the 'classic' pre/post rotate with an alternative one that enables
us to use fewer distinct twiddle factors.
http://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568980508.pdf
For prerotation, the factors are just sin,cos(2PI*i/N)
For postrotation, the factors are sin,cos(2PI*(i+1/4)/N)
Therefore, prerotation can immediately reuse the same twiddles as fft
(for postrotation it's still a bit complex, we reuse the fft trig tables
where we can, or a special table for N=2048, or interpolate between
trig tables for N>2048)
*/
const int32_t *T = sincos_lookup0;
const int step = 2<<(12-nbits);
const uint16_t * p_revtab=revtab;
{
const uint16_t * const p_revtab_end = p_revtab + n8;
while(LIKELY(p_revtab < p_revtab_end))
{
j = (*p_revtab)>>revtab_shift;
XNPROD31(*in2, *in1, T[1], T[0], &z[j].re, &z[j].im );
T += step;
in1 += 2;
in2 -= 2;
p_revtab++;
j = (*p_revtab)>>revtab_shift;
XNPROD31(*in2, *in1, T[1], T[0], &z[j].re, &z[j].im );
T += step;
in1 += 2;
in2 -= 2;
p_revtab++;
}
}
{
const uint16_t * const p_revtab_end = p_revtab + n8;
while(LIKELY(p_revtab < p_revtab_end))
{
j = (*p_revtab)>>revtab_shift;
XNPROD31(*in2, *in1, T[0], T[1], &z[j].re, &z[j].im);
T -= step;
in1 += 2;
in2 -= 2;
p_revtab++;
j = (*p_revtab)>>revtab_shift;
XNPROD31(*in2, *in1, T[0], T[1], &z[j].re, &z[j].im);
T -= step;
in1 += 2;
in2 -= 2;
p_revtab++;
}
}
/* ... and so fft runs in OUTPUT buffer */
ff_fft_calc_c(nbits-2, z);
/* post rotation + reordering. now keeps the result within the OUTPUT buffer */
switch( nbits )
{
default:
{
fixed32 * z1 = (fixed32 *)(&z[0]);
fixed32 * z2 = (fixed32 *)(&z[n4-1]);
int magic_step = step>>2;
int newstep;
if(n<=1024)
{
T = sincos_lookup0 + magic_step;
newstep = step>>1;
}
else
{
T = sincos_lookup1;
newstep = 2;
}
while(z1<z2)
{
fixed32 r0,i0,r1,i1;
XNPROD31_R(z1[1], z1[0], T[0], T[1], r0, i1 ); T+=newstep;
XNPROD31_R(z2[1], z2[0], T[1], T[0], r1, i0 ); T+=newstep;
z1[0] = -r0;
z1[1] = -i0;
z2[0] = -r1;
z2[1] = -i1;
z1+=2;
z2-=2;
}
break;
}
