- ranking metric acceleration on the gpu

[sriramch]

this is the last part of #5326 that has been split. the performance numbers are here

please note the following:

• metrics applicable to only ranking datasets are fully accelerated on gpu - map, ndcg, pre
• metrics (auc[pr]) applicable to ranking and non-ranking datasets work thusly:
  • these metrics when computed on non-ranking datasets aren't accelerated - i.e. they still run on cpu, but optimized on cpu (hence, should be better than what we had before - see here)
    • this can be worked on as a follow-up pr
  • these metrics when computed on ranking datasets are semi accelerated - meaning, processing multiple groups are still in parallel, but there is a linear iteration of predictions within each group to bucketize them
    • note: it is still much better than the version that runs on the cpu though (~ 6x-8x eval time improvement), even for datasets that have large number of elements/group and smaller group cardinality (which are atypical for ranking datasets). i tried a couple of 100 groups with ~250k items/group, and it still performed well

@RAMitchell @trivialfis - please review.

[mli]

Codecov Report

Merging #5398 into master will not change coverage by %.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #5398   +/-   ##
=======================================
  Coverage   84.07%   84.07%           
=======================================
  Files          11       11           
  Lines        2411     2411           
=======================================
  Hits         2027     2027           
  Misses        384      384           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a38e7bd...810e537. Read the comment docs.

[sriramch]

auc metric performance numbers

test environment

• 1 socket
• 6 cores/socket
• 2 threads/core
• 80 gb system memory
• v100 gpu

test

• uses all cpu threads
• builds 100 trees
• metric eval times are reported below

results

• no additional gpu memory was used
• all times are in seconds
• mortgage dataset used 60m training instances (to be able to fit into the available gpu memory)

dataset	eval time master	eval time this pr
higgs	70.79	1.42
mortgage	266.98	30.19

[sriramch]

@trivialfis i would appreciate your review when you get a chance.

[trivialfis]

LGTM! Sorry for the long wait. Previously I mentioned with @RAMitchell that maybe simple functions are more suitable for implementing the GPU metrics as I think the registry is just too tricky and unnecessary. But I won't block the PR for this as we can refactor them later when needed (like implementing other metrics).