void fwt1(unsigned int N, unsigned int d, const long dims[N], const long ostr[N], complex float* low, complex float* hgh, const long istr[N], const complex float* in, const long flen, const float filter[2][2][flen])
{
debug_printf(DP_DEBUG4, "fwt1: %d/%d\n", d, N);
debug_print_dims(DP_DEBUG4, N, dims);
assert(dims[d] >= 2);
long odims[N];
md_copy_dims(N, odims, dims);
odims[d] = bandsize(dims[d], flen);
debug_print_dims(DP_DEBUG4, N, odims);
long o = d + 1;
long u = N - o;
// 0 1 2 3 4 5 6|7
// --d-- * --u--|N
// ---o---
assert(d == md_calc_blockdim(d, dims + 0, istr + 0, CFL_SIZE));
assert(u == md_calc_blockdim(u, dims + o, istr + o, CFL_SIZE * md_calc_size(o, dims)));
assert(d == md_calc_blockdim(d, odims + 0, ostr + 0, CFL_SIZE));
assert(u == md_calc_blockdim(u, odims + o, ostr + o, CFL_SIZE * md_calc_size(o, odims)));
// merge dims
long wdims[3] = { md_calc_size(d, dims), dims[d], md_calc_size(u, dims + o) };
long wistr[3] = { CFL_SIZE, istr[d], CFL_SIZE * md_calc_size(o, dims) };
long wostr[3] = { CFL_SIZE, ostr[d], CFL_SIZE * md_calc_size(o, odims) };
#ifdef USE_CUDA
if (cuda_ondevice(in)) {
assert(cuda_ondevice(low));
assert(cuda_ondevice(hgh));
float* flow = md_gpu_move(1, MD_DIMS(flen), filter[0][0], FL_SIZE);
float* fhgh = md_gpu_move(1, MD_DIMS(flen), filter[0][1], FL_SIZE);
wl3_cuda_down3(wdims, wostr, low, wistr, in, flen, flow);
wl3_cuda_down3(wdims, wostr, hgh, wistr, in, flen, fhgh);
md_free(flow);
md_free(fhgh);
return;
}
#endif
// no clear needed
wavelet_down3(wdims, wostr, low, wistr, in, flen, filter[0][0]);
wavelet_down3(wdims, wostr, hgh, wistr, in, flen, filter[0][1]);
}
void iwt1(unsigned int N, unsigned int d, const long dims[N], const long ostr[N], complex float* out, const long istr[N], const complex float* low, const complex float* hgh, const long flen, const float filter[2][2][flen])
{
debug_printf(DP_DEBUG4, "ifwt1: %d/%d\n", d, N);
debug_print_dims(DP_DEBUG4, N, dims);
assert(dims[d] >= 2);
long idims[N];
md_copy_dims(N, idims, dims);
idims[d] = bandsize(dims[d], flen);
debug_print_dims(DP_DEBUG4, N, idims);
long o = d + 1;
long u = N - o;
// 0 1 2 3 4 5 6|7
// --d-- * --u--|N
// ---o---
assert(d == md_calc_blockdim(d, dims + 0, ostr + 0, CFL_SIZE));
assert(u == md_calc_blockdim(u, dims + o, ostr + o, CFL_SIZE * md_calc_size(o, dims)));
assert(d == md_calc_blockdim(d, idims + 0, istr + 0, CFL_SIZE));
assert(u == md_calc_blockdim(u, idims + o, istr + o, CFL_SIZE * md_calc_size(o, idims)));
long wdims[3] = { md_calc_size(d, dims), dims[d], md_calc_size(u, dims + o) };
long wistr[3] = { CFL_SIZE, istr[d], CFL_SIZE * md_calc_size(o, idims) };
long wostr[3] = { CFL_SIZE, ostr[d], CFL_SIZE * md_calc_size(o, dims) };
md_clear(3, wdims, out, CFL_SIZE); // we cannot clear because we merge outputs
#ifdef USE_CUDA
if (cuda_ondevice(out)) {
assert(cuda_ondevice(low));
assert(cuda_ondevice(hgh));
float* flow = md_gpu_move(1, MD_DIMS(flen), filter[1][0], FL_SIZE);
float* fhgh = md_gpu_move(1, MD_DIMS(flen), filter[1][1], FL_SIZE);
wl3_cuda_up3(wdims, wostr, out, wistr, low, flen, flow);
wl3_cuda_up3(wdims, wostr, out, wistr, hgh, flen, fhgh);
md_free(flow);
md_free(fhgh);
return;
}
#endif
wavelet_up3(wdims, wostr, out, wistr, low, flen, filter[1][0]);
wavelet_up3(wdims, wostr, out, wistr, hgh, flen, filter[1][1]);
}
/**
* compute minimal dimension of largest contiguous block(s)
*
*/
unsigned int min_blockdim(unsigned int D, unsigned int N, const long dims[N], long (*strs[D])[N], size_t size[D])
{
unsigned int mbd = N;
for (unsigned int i = 0; i < D; i++)
mbd = MIN(mbd, md_calc_blockdim(N, dims, *strs[i], size[i]));
return mbd;
}
