| | 12 | </p> |
| | 13 | |
| | 14 | <p> |
| | 15 | 175. |
| | 16 | <b>FFT better memory efficiency.</b><br> |
| | 17 | Benoit Zuber<br> |
| | 18 | April 13, 2009<br> |
| | 19 | An FFT of a 32-bit float map takes 8 times the memory of the original map |
| | 20 | because it produces a complex double precision results (4x) then the absolute |
| | 21 | value (2x), then cast to single-precision (1x), then recentering (1x). A variety |
| | 22 | of optimizations could reduce this although using a different FFT package like |
| | 23 | fftw would be needed to use single precision since our current numpy fft only |
| | 24 | supports double precision. |