Shuffle

[RAMitchell]

Implementation of thrust::shuffle following std::shuffle API.

Todo:

• Code style consistency
• Documentation
• Document/justify selection of round function (currently uses linear feedback shift but with somewhat arbitrary parameters)
• Define thrust::shuffle_copy

I'm wondering if I should add another version of this function specifying an output iterator to avoid the copy?

Benchmark results for shuffle (on DGX1 V100):

Algorithm	Element Type	Elements per Trial	STL Average Walltime	STL Walltime Uncertainty	STL Average Throughput	Thrust Average Walltime	Thrust Walltime Uncertainty	Thrust Average Throughput
shuffle	char	67108864	1.86	0.04	3.60E+07	0.02	0.03	4.00E+09
shuffle	char	134217728	5.33	0.03	2.52E+07	0.04	0.04	4.00E+09
shuffle	double	8388608	0.2402	0.0003	3.49E+07	0.0041	0.0003	2.10E+09
shuffle	double	16777216	0.66	0.01	2.52E+07	0.0063	0.001	2.70E+09
shuffle	float	16777216	0.49	0.01	3.40E+07	0.01	0.03	2.00E+09
shuffle	float	33554432	1.36	0.02	2.47E+07	0.01	0.03	3.00E+09
shuffle	int	16777216	0.475	0.007	3.53E+07	0.01	0.03	1.00E+09
shuffle	int	33554432	1.3	0.01	2.58E+07	0.01	0.03	2.00E+09
shuffle	int16_t	33554432	0.95	0.01	3.53E+07	0.01	0.04	3.00E+09
shuffle	int16_t	67108864	2.61	0.02	2.57E+07	0.02	0.04	3.00E+09
shuffle	int32_t	16777216	0.477	0.007	3.52E+07	0.01	0.04	1.00E+09
shuffle	int32_t	33554432	1.3	0.02	2.58E+07	0.01	0.03	3.00E+09
shuffle	int64_t	8388608	0.2414	0.0007	3.48E+07	0.01	0.04	8.00E+08
shuffle	int64_t	16777216	0.6686	0.0006	2.51E+07	0.01	0.03	1.00E+09
shuffle	int8_t	67108864	1.89	0.01	3.55E+07	0.02	0.03	4.00E+09
shuffle	int8_t	134217728	5.23	0.05	2.57E+07	0.04	0.04	4.00E+09

[jaredhoberock]

I'm wondering if I should add another version of this function specifying an output iterator to avoid the copy?

Yes, I'd suggest shuffle and shuffle_copy.

[griwes]

This looks pretty good, I only have a few comments. I've noticed some typos in the doxygen, I'll try to do another pass through that soon.

Looking at the individual commits, I think we should squash this down to a single one before we merge.

[alliepiper]

This will cause issues with C++11 compilers, but once that's cleared up this patch LGTM.

[alliepiper]

Thank for the quick updates! I'll take it from here and starting validating this against our internal CI.

Just a heads up, I'll need to squash this into a single commit before merging. I'll push an update that does that soon.

[alliepiper]

Local build is clean, tests pass. CI started under shelve 28225884.