Optimize Atari emulator by ~30%

[qstanczyk]

Eliminate variables which are constant.
Inline memory access methods.
Benchmarking shows avg. ~30% speedup on multiple Atari games.

[mgbellemare]

LGTM, thanks for this big performance improvement! I'm going to wait for @mcmachado to give his thumbs up before merging (==> he will merge).

[maxwellreuter]

@qstanczyk I'm looking to emulate my Atari games as fast as possible. Do you have a feeling for how much room there is for optimization, if any? If so, what sort of speedup % do you think is possible?

[qstanczyk]

Hi Max, I didn't have time myself to do that, but I had an idea of providing games' ROM converter to C++. Currently ALE implements full Atari instruction set and performs interpretation of ROM. Instead ROM could be converted to C++/intermediate representation and JIT-compiled. That could give order of magnitude speedup, but is a longer effort.

[maxwellreuter]

@qstanczyk thanks for your reply. Do you know how much of an effort that would be and what sort of skills/tools would be needed to do that?

[qstanczyk]

It requires C++ knowledge and some profiling/optimization experience. Would be best I guess if instead of generating C++ it would generate Intermediate Representation and JIT with LLVM (https://llvm.org/docs/LangRef.html). This approach is basically a re-write of ALE.

[maxwellreuter]

@qstanczyk ok, thanks -- do you know what sort of time commitment that would be? would it be less if I only focused on one game such as breakout? If so how much of a time commitment would that be?

[qstanczyk]

With some LLVM prior experience... I would say 3 weeks of actual work. Doing that for one specific game might be slightly easier, as ROMs have different versions with different operation sets (so doing that as a proof of concept for some specific game is a good start anyway).

[maxwellreuter]

@qstanczyk how long do you think it would take for someone to get up to speed without any experience with LLVM, working with ROMs, or profiling/optimization, so that they can be effective?

[JesseFarebro]

@maxwellreuter if you haven't heard of the CUDA port of the ALE you could also check that out. Depending on what you're trying to do that could provide some wall-time improvements.

[qstanczyk]

Hard to tell, depends on the background, approach (do you want to understand details or just get things done)... but Jesse has a point - GPU environment might be a fast option for you.

[maxwellreuter]

@JesseFarebro @qstanczyk thanks!