univector<T> correlate(const univector_ref<const T>& src1, const univector_ref<const T>& src2)
{
const size_t size = next_poweroftwo(src1.size() + src2.size() - 1);
univector<complex<T>> src1padded = src1;
univector<complex<T>> src2padded = reverse(src2);
src1padded.resize(size, 0);
src2padded.resize(size, 0);
dft_plan_ptr<T> dft = dft_cache::instance().get(ctype_t<T>(), size);
univector<u8> temp(dft->temp_size);
dft->execute(src1padded, src1padded, temp);
dft->execute(src2padded, src2padded, temp);
src1padded = src1padded * src2padded;
dft->execute(src1padded, src1padded, temp, true);
const T invsize = reciprocal<T>(size);
return truncate(real(src1padded), src1.size() + src2.size() - 1) * invsize;
}
size_t process(univector<T, Tag>& output, univector_ref<const T> input)
{
const itype required_input_size = input_size_for_output(output.size());
const itype input_size = input.size();
for (size_t i = 0; i < output.size(); i++)
{
const itype intermediate_index =
output_position_to_intermediate(static_cast<itype>(i) + output_position);
const itype intermediate_start = intermediate_index - taps + 1;
const std::lldiv_t input_pos =
floor_div(intermediate_start + interpolation_factor - 1, interpolation_factor);
const itype input_start = input_pos.quot; // first input sample
const itype tap_start = interpolation_factor - 1 - input_pos.rem;
const univector_ref<T> tap_ptr = filter.slice(static_cast<size_t>(tap_start * depth));
if (input_start >= input_position + input_size)
{
output[i] = T(0);
}
else if (input_start >= input_position)
{
output[i] = dotproduct(input.slice(input_start - input_position, depth), tap_ptr);
}
else
{
const itype prev_count = input_position - input_start;
output[i] = dotproduct(delay.slice(size_t(depth - prev_count)), tap_ptr) +
dotproduct(input.slice(0, size_t(depth - prev_count)),
tap_ptr.slice(size_t(prev_count), size_t(depth - prev_count)));
}
}
if (required_input_size >= depth)
{
delay.slice(0, delay.size()) = padded(input.slice(size_t(required_input_size - depth)));
}
else
{
delay.truncate(size_t(depth - required_input_size)) = delay.slice(size_t(required_input_size));
delay.slice(size_t(depth - required_input_size)) = padded(input);
}
input_position += required_input_size;
output_position += output.size();
return required_input_size;
}