Re-introduce double buffer in UpdatePosition, to fix perf regression in gpu_hist

[hcho3]

Closes #6552 by partially reverting the commit f779980

[hcho3]

I suspect building TemporaryArray at every invocation of UpdatePosition gets expensive in the particular example provided.

[trivialfis]

Could you please share some benchmark results?

[hcho3]

@trivialfis The posted example in #6552 won't complete in a reasonable amount of time with the latest master (and 1.3.0). I can leave it up over night and see if it finishes by tomorrow.

On the other hand, here's the result with the proposed fix:

[0]     convergence-merror:0.10066
[1]     convergence-merror:0.10053
[2]     convergence-merror:0.09998
[3]     convergence-merror:0.09928
[4]     convergence-merror:0.09851
[5]     convergence-merror:0.09770
[6]     convergence-merror:0.09689
[7]     convergence-merror:0.09601
[8]     convergence-merror:0.09517
[9]     convergence-merror:0.09438
Time elapsed = 673.7375219200039 s

[trivialfis]

Thanks for sharing, what about more conventional dataset?

[trivialfis]

I can leave it up over night and see if it finishes by tomorrow.

Nah, no need. I'm just worrying about there might be regression on other types of data.

[hcho3]

what about more conventional dataset?

Any suggestion? Should I try gbm-bench?

[trivialfis]

Any suggestion? Should I try gbm-bench?

That would be great!

[trivialfis]

Looks good as long as there's no regression on other datasets.

[hcho3]

@trivialfis Here is the benchmark results on gbm-bench. I do not see any performance degradation:

Runtime (s)	Before	After
airline	40.44	40.57
bosch	2.77	2.77
covtype	4.37	4.38
epsilon	16.12	16.19
fraud	0.44	0.41
higgs	6.50	6.53
year	0.75	0.72