[llvm_instcount] Leaderboard Submission: DQN trained on test set

[phesse001]

This adds the results of training a dqn on the benchmarks from the test set (cbench-v1) for 4000 episodes using the llvm-ic-v0 environment with InstCountNorm as the observation passed to the dqn.

The trained model was then saved and loaded for evaluation of the policy. Since this is not a general policy, I timed the training process with the time command and got a total of 2085.99s. I averaged this over the 23 benchmarks, getting ~90.700s per benchmark, and added the average walltime of .318s from eval_llvm_instcount_policy() to get a total of 91.018s walltime(mean).

My code for the dqn can be found here

What is in this PR?

• A write-up of the approach based on the leaderboard submission template and the tabular_q writeup.
• A csv file of the results produced by eval_llvm_instcount_policy()
• A leaderboard entry for results achieved using the dqn

Walltime (mean)	Codesize Reduction (geomean)
91.018s	1.029×

[ChrisCummins]

Hi @phesse001, nice submission!

Could you add a little bit of an explanation to the "What range of values were considered for the above parameters?" section of the write up to describe how you arrived at those values? E.g. was it by trial and error, what was the experimental setup for each "trial". Just a sentence or two would be enough :)

Other than that, looks great. A few minor nitpick comments left inline.

I have reproduced your results locally, so this is me signing off on your results! Once you've had a chance to take a look at those small things I'd say this is looking good to merge.

$ python -m compiler_gym.bin.validate --env=llvm-ic-v0 CompilerGym/results.csv
✅  cbench-v1/rijndael                        1.1027
✅  cbench-v1/adpcm                           1.0084
✅  cbench-v1/adpcm                           1.0084
✅  cbench-v1/adpcm                           1.0084
✅  cbench-v1/ispell                          0.9887
✅  cbench-v1/rijndael                        1.1027
✅  cbench-v1/rijndael                        1.1027
...
✅  cbench-v1/tiffmedian                      1.0324
----------------------------------------------------
Number of validated results: 230 of 230
Mean walltime per benchmark: 0.367s (std: 0.781s)
Geometric mean IrInstructionCountOz: 1.029 (std: 0.112)

Cheers,
Chris

[ChrisCummins]

LGTM, thanks @phesse001!