Facebook Is Aiming To Make Compilers Faster Using Machine Learning With CompilerGym

[phoronix]

Phoronix: Facebook Is Aiming To Make Compilers Faster Using Machine Learning With CompilerGym

Facebook this week announced the open-sourcing of CompilerGym as their effort to improve compiler performance by leveraging machine learning to tackle optimization work...

https://www.phoronix.com/scan.php?pa...k-Compiler-Gym

[George99]

typo: 104461 -> 10^4461 or 1E4461

[ms178]

Pretty amazing. I hope it doesn't take a decade to see the fruits of that work incorporated into LLVM/GCC.

[skeevy420]

Unlike Intel, they hired a proof reader. I was actually able to read their press release without getting hung-up on my usual dumbassities.

[ZuseZ4]

ms178 One of this years LLVM gsoc projects already looked into something related:
https://summerofcode.withgoogle.com/...26575691792384

[ddriver]

Great, maybe they can use AI to figure how to not be such a sh1tty company.

Also, having used react, I somehow doubt performance and efficiency are true priorities.

[timofonic]

I hope efforts get directly mainlined in compilers in some way and we can see a lot better optimization along them.

[david-nk]

GCC also needs to get configurable phase ordering, so it can benefit from such work. The problem with GCC right now is that the optimization phases break each other or are simply not applied because they happen in the wrong order. Often it would be necessary to run an optimization phase multiple times throughout the optimization process and with reinforcement learning, you could figure out the best pipeline setup for your application. And find a better default setting as well.

[lamka02sk]

Great, maybe they can use AI to figure how to not be such a sh1tty company.

Also, having used react, I somehow doubt performance and efficiency are true priorities.

React is a whole different story and the performance is quite similar to other frameworks like Vue or Angular so I wouldn't criticise it too much. The problem of these frameworks is the overhead/glue code and you can't really do anything about that. Just a short time ago, everyone thought that virtual DOM is the only way to go forward, but now, Svelte is the proof, that you don't need virtual DOM to make dynamic web apps fast. I think the compiler based approach is a much better solution.

[bregma]

This seems very confusing. Their goal is to have compiler produce more "optimized" (for what?) programs by calling out to some Python programs that will iterate over combinatoric solutions using inscrutable and unknown heuristics and that will speed up the compiler. It strikes me that (1) adding a whole lot of out-of-process processing is unlikely to make the compilation faster, (2) how it is possible to prove the correctness of the output of a ML-generated algorithm laden with intellectual debt, and (3) how faster compilers equate with faster programs is beyond my understanding.

I guess the proof will be in the pudding. Or at least, the results will be in the pudding. I wouldn't trust my life or my liability to ML-generated software. Azimov's laws are fiction.