Since the input is copied to the real and imaginary components, you don’t actually need to copy it at all.
Think of it like having a pointer (in the C sense) to the real and imaginary image buffers. Set both pointers to the input signal, then run your FFT algo on that.
The key is that it minimizes memory bandwidth, not computation. It’s highly cache coherent.
Thanks, so the real and imag components always map to the same value during the processing, i.e. After every butterfly the real and iamg will be the same value? Cause again if the real and imag components ever differ it means more storage needed.
Think of it like having a pointer (in the C sense) to the real and imaginary image buffers. Set both pointers to the input signal, then run your FFT algo on that.
The key is that it minimizes memory bandwidth, not computation. It’s highly cache coherent.