Half-precision implementation

[woensug-choi]

Since the final data of the sonar calculation is typical int values, if carefully adopted, the half-precision calculation could boost the refresh rate of the sonar image. This is a note for later retry.

Issues on Half-precision CUDA calculation implementation

• Usage of a half-precision library is very limited to certain types of graphics cards.
  • It's sometimes not supported at all or partially supported.
  • If partially supported, the multiplication (__hmul()) and division(__hdiv()) function cannot be used. Only addition and subtraction are allowed. Also, some mathematical functions are not usable.
• Calculation time boost in the sonar_calculation CUDA kernel function does not affect the final refresh rate.
  • The bottleneck of the sonar calculation is the ray summation (reduction). It's what CPUs are good at compared to GPUs. Effective reduction of parallelization is very complex. Some considerations are already included in the code.

Benchmark results

• at workstation (i9-9900K 3.6GHz, Nvidia GeForce RTX 2080Ti)
• 10 Hz achievable with Ray and Range reduced

512 Beams	Refresh Rate [Hz]	Total Time [s]	Core	Sum	Corr	FFT
Full Calculation 60 m Range 114 Rays	0.5 Hz	1.7 s	0.3	1.26	0.05	0.03
Ray Reduced 60 m Range 11 Rays	3.0 Hz	0.27 s	0.02	0.16	0.05	0.03
Ray/Range Reduced 10 m Range 11 Rays	10 Hz	0.06 s	0.00	0.04	0.01	0.00