Table 1.
Framerate comparison between different processors.
| Duration of steps in sequential order (ms) | Occurrence number | CPU (i7-6700H) | GPU (GTX 960M) | GPU (GTX Titan Black) |
|---|---|---|---|---|
| FFT of interferogram | ×2 | 117.45 | 1.15 | 0.49 |
| Crop & IFFT | ×4 | 21.48 | 2.33 | 1.06 |
| Unwrapping | ×4 | 109.93 | 5.54 | 3.77 |
| FFT of shearing wavefront | ×4 | 19.21 | 1.73 | 0.87 |
| Least square fitting | ×1 | 547.97 | 7.59 | 5.12 |
| IFFT of wavefront spectrum | ×1 | 10.78 | 0.29 | 0.20 |
| Processing time without LE (ms) | 1417.92 | 49.61 | 29.15 | |
| Processing time with LE (ms) | — | 40.31 | 20.90 | |
| Framerate (fps) | 0.71 | 24.81 | 47.85 | |
The process steps are listed in the sequential order corresponding to the flowchart in Fig. 3, and their occurrence numbers indicate the times each process step occurs in the phase retrieval algorithm. Note that the duration of the second FFT on the 501 × 501 shearing wavefront was a bit larger than that of the first FFT on the 2048 × 2048 interferogram on GPU platforms, due to the fact that the zero padding of 501 × 501 pixels to the size of 2N for FFT also consumes time in memory transferring.