static complex double fftmod_phase2(long n, int j, bool inv, double phase)
{
phase += fftmod_phase(n, j);
double rem = phase - floor(phase);
double sgn = inv ? -1. : 1.;
#if 1
if (rem == 0.)
return 1.;
if (rem == 0.5)
return -1.;
if (rem == 0.25)
return 1.i * sgn;
if (rem == 0.75)
return -1.i * sgn;
#endif
return cexp(M_PI * 2.i * sgn * rem);
}
static void fftmod2_r(unsigned int N, const long dims[N], unsigned long flags, const long ostrs[N], complex float* dst, const long istrs[N], const complex float* src, bool inv, double phase)
{
if (0 == flags) {
md_zsmul2(N, dims, ostrs, dst, istrs, src, cexp(M_PI * 2.i * (inv ? -phase : phase)));
return;
}
/* this will also currently be slow on the GPU because we do not
* support strides there on the lowest level */
unsigned int i = N - 1;
while (!MD_IS_SET(flags, i))
i--;
#if 1
// If there is only one dimensions left and it is the innermost
// which is contiguous optimize using md_zfftmod2
if ((0u == MD_CLEAR(flags, i)) && (1 == md_calc_size(i, dims))
&& (CFL_SIZE == ostrs[i]) && (CFL_SIZE == istrs[i])) {
md_zfftmod2(N - i, dims + i, ostrs + i, dst, istrs + i, src, inv, phase);
return;
}
#endif
long tdims[N];
md_select_dims(N, ~MD_BIT(i), tdims, dims);
#pragma omp parallel for
for (int j = 0; j < dims[i]; j++)
fftmod2_r(N, tdims, MD_CLEAR(flags, i),
ostrs, (void*)dst + j * ostrs[i], istrs, (void*)src + j * istrs[i],
inv, phase + fftmod_phase(dims[i], j));
}
