Introduce RuntimeSeriesReward

[lqwk]

Introduce RuntimeSeriesReward

Introduce a new implementation of comparing program runtimes that computes the reward as the difference of the medians between the current set of runtimes and the previous set of runtimes only if the runtime series are significantly different (determined by the Kruskal–Wallis test).

Source: https://htor.inf.ethz.ch/publications/img/hoefler-scientific-benchmarking.pdf

Testing

I ran a series of tests comparing the new implementation with the existing implementation using the LLVM autotuner on the cbench-v1 benchmark. The rewards are shown below:

Benchmark	Runtime	Runtime Series
cbench-v1/bitcount	0.847564	0.780826
cbench-v1/blowfish	0.997087	0.997496
cbench-v1/bzip2	2.505352	2.516984
cbench-v1/crc32	1.003222	0.995800
cbench-v1/dijkstra	1.001125	1.014525
cbench-v1/gsm	0.785371	0.812009
cbench-v1/jpeg-c	0.976613	0.989570
cbench-v1/jpeg-d	1.004371	1.010437
cbench-v1/patricia	0.992379	1.010707
cbench-v1/qsort	0.958153	0.986507
cbench-v1/sha	1.000886	0.990302
cbench-v1/stringsearch	1.004281	1.006058
cbench-v1/stringsearch2	1.009237	1.050334
cbench-v1/susan	0.959058	0.975618
cbench-v1/tiff2bw	1.016761	1.018933
cbench-v1/tiff2rgba	1.105831	1.104215
cbench-v1/tiffdither	1.042336	1.022278
cbench-v1/tiffmedian	1.005050	1.010557

The new implementation is on par with the existing implementation, and even beats the existing implementation on 12/17 of the benchmarks.

I am proposing to merge this upstream and we can maybe work on other optimizations such as early stopping.

[codecov-commenter]

Codecov Report

Merging #762 (a53276a) into development (1c40e5b) will decrease coverage by 33.64%.
The diff coverage is 34.00%.

@@               Coverage Diff                @@
##           development     #762       +/-   ##
================================================
- Coverage        89.28%   55.63%   -33.65%     
================================================
  Files              130      131        +1     
  Lines             7912     7961       +49     
================================================
- Hits              7064     4429     -2635     
- Misses             848     3532     +2684     

Impacted Files	Coverage Δ
compiler_gym/spaces/runtime_series_reward.py	28.57% <28.57%> (ø)
compiler_gym/wrappers/llvm.py	44.82% <33.33%> (-55.18%)	⬇️
compiler_gym/spaces/__init__.py	100.00% <100.00%> (ø)
compiler_gym/wrappers/__init__.py	100.00% <100.00%> (ø)
compiler_gym/util/permutation.py	0.00% <0.00%> (-100.00%)	⬇️
compiler_gym/leaderboard/__init__.py	0.00% <0.00%> (-100.00%)	⬇️
compiler_gym/service/runtime/__init__.py	0.00% <0.00%> (-100.00%)	⬇️
compiler_gym/service/runtime/benchmark_cache.py	0.00% <0.00%> (-100.00%)	⬇️
...mpiler_gym/service/runtime/compiler_gym_service.py	0.00% <0.00%> (-100.00%)	⬇️
...ice/runtime/create_and_run_compiler_gym_service.py	0.00% <0.00%> (-100.00%)	⬇️
... and 83 more

[lqwk]

Note: needs #761 to land first

[ChrisCummins]

Hi @lqwk, I'm very sorry for my delay in reviewing this. I've not forgotten about it, I just have a backlog of issues to fix on the CI so that I can run the tests against these changes.

Cheers,
Chris

[ChrisCummins]

Hi @lqwk, okay, I finally pushed through the backlog of issues and have a newly stable v0.2.5 release. Sorry again that it me so long to getting around to this.

Are you still working on this? If so, could you please rebase this on top of the development branch so that we can use the CI to verify that all tests pass.

Cheers,
Chris