Various CUDA Optimizations

[lilinitsy]

I've added in a ton of CUDA optimizations for LS/CE/AOV.

Largely, I've done some lightcurve batching on lomb-scargle, and made use of CUDA asynchronous streams (including async memory transfers) to get some nice performance improvements.

Smaller improvements come from using restrict on pointers that are appropriate for it, and using some GPU intrinsic functions for slow math calls (ie,, using __sincosf).

All testings were done on a V100 on the SDSC Expanse GPU cluster.

TIMINGS

METHOD	VERSION	MACHINE	DATA	TIME	% GAIN
LS	Baseline	EXPANSE	all time, all mags, 100k periods	139.75323605537415
LS	OPTIMIZED	EXPANSE	all time, all mags, 100k periods	122.99888670444	+11.98%
CE	Baseline	EXPANSE	all time, all mags, all	237.9152114391327
CE	OPTIMIZED	EXPANSE	all time, all mags, all	224.75576400756836	+5.53%
AOV	Baseline	EXPANSE	all time, all mags, 100k periods	244.3776876926422
AOV	OPTIMIZED	EXPANSE	all time, all mags, 100k periods	214.41236901283264	+12.26%

METHOD	VERSION	MACHINE	DATA	TIME	% GAIN
LS	Baseline	EXPANSE	1000 lightcurves, all periods	17.119052171707153
LS	OPTIMIZED	EXPANSE	1000 lightcurves, all periods	16.61055612564087	+2.97%
CE	Baseline	EXPANSE	1000 lightcurves, all periods	28.453676223754883
CE	OPTIMIZED	EXPANSE	1000 lightcurves, all periods	27.26391577720642	+4.18%
AOV	Baseline	EXPANSE	1000 lightcurves, all periods	30.421194076538086
AOV	OPTIMIZED	EXPANSE	1000 lightcurves, all periods	26.356033086776733	+13.36%

The performance gains decrease as the GPU memory bandwidth increases -- on a GTX 1080, the lomb-scargle gains were in the mid 20%'s.

[ejaszewski]

Happy to review the changes this weekend if you would like another pair of eyes on it!

[ejaszewski]

The changed .clang-format makes it very difficult to actually tell what has changed because the diff is picking up all of the whitespace changes. If possible, can you re-format this with the original clang-format so the diff is meaningful?

[lilinitsy]

@ejaszewski Sure thing. I'll try and do that this weekend.