shish / rosettaboy

A gameboy emulator in several different languages
MIT License
466 stars 22 forks source link

RosettaBoy

C C++ Go Nim PHP Python Rust Zig

Trying to implement a gameboy emulator in a bunch of languages for my own amusement and education; also giving people an opportunity to compare the same code written in different languages, similar to Rosetta Code but with a non-trivial codebase :)

The main goals are:

Notably, 100% accuracy is not a goal - if Tetris works perfectly then I'm happy, if other games require more obscure hardware features, then I'll weigh up whether or not the feature is worth the complexity.

Also yes, "consistent across languages" and "idiomatic" can be at odds - there are subjective compromises to be made, but for the most part that doesn't seem to be a huge problem. Rust uses Result, Python uses Exception, Go uses error - but so far it's always been pretty obvious that eg NewCart() in go and Cart::new() in rust are doing fundamentally the same thing in the same way.

So far all the implementations follow a fairly standard layout, with each module teaching me how to do a new thing. In fact they're all so similar, I wrote one copy of the documentation for all the implementations:

Pull requests to translate into new languages, or fleshing out existing languages, are very welcome :)

Dev Guide

I want to keep the build processes as simple as possible:

Ideally the build script should also fetch (if needed) any dependencies, the only assumption I want to make is that the user has the standard language dev kits installed (eg we assume anyone who wants to work on the Rust version will have Cargo installed; anyone who wants to work on Python will have virtualenv + pip; etc)

If you have Docker available, you can run ./utils/shell.sh to create and run an environment with all necessary dev tools pre-installed. — ./build.sh && ./rosettaboy-release --headless --silent should be able to pass tests for all languages.

Benchmarks

Warning: These implementations aren't 100% in-sync, so take numbers with a large grain of salt. For example, as of this writing, the PHP version is using a stub SDL mock instead of calling the real C library, because I couldn't find an SDL library that worked.

If somebody knows how to measure CPU instructions instead of clock time, that seems fairer; especially if we can get the measurement included automatically via github actions. Pull requests welcome :)

Running on an M1 Macbook Pro, using (to my knowledge) the latest version of each compiler, with standard "release mode" flags (see each language's build.sh for exactly which flags are used):

$ ./all.py bench | ./utils/sort.py
  nim / lto    : Emulated 16438 frames in 10.00s (1643fps)
   rs / lto    : Emulated 16029 frames in 10.00s (1603fps)
  cpp / lto    : Emulated 15064 frames in 10.00s (1506fps)
    c / lto    : Emulated 15054 frames in 10.00s (1505fps)
   rs / release: Emulated 13904 frames in 10.00s (1390fps)
    c / release: Emulated 13368 frames in 10.00s (1337fps)
  cpp / release: Emulated 12967 frames in 10.00s (1297fps)
  nim / release: Emulated 11860 frames in 10.00s (1185fps)
  zig / release: Emulated  8844 frames in 10.00s (884fps)
  zig / safe   : Emulated  7206 frames in 10.00s (721fps)
    c / debug  : Emulated  5914 frames in 10.00s (591fps)
  cpp / debug  : Emulated  5663 frames in 10.00s (566fps)
   go / release: Emulated  5029 frames in 10.00s (503fps)
  pxd / release: Emulated  3610 frames in 10.00s (361fps)
  nim / debug  : Emulated  2587 frames in 10.00s (258fps)
   rs / debug  : Emulated  1808 frames in 10.01s (181fps)
   py / mypyc  : Emulated   789 frames in 10.01s (79fps)
  php / opcache: Emulated   655 frames in 10.00s (65fps)
  php / release: Emulated   257 frames in 10.02s (26fps)
   py / release: Emulated   165 frames in 10.02s (16fps)

Also if you spot some bit of code that is weirdly slow and making your favourite language look bad, pull requests to fix that might be welcome too, but "simplicity and consistency" are going to take priority (eg an "add an inline flag to this function" would be great but "replace python's CPU interpreter with a JIT compiler written as a C extension module" would probably be rejected[0])

[0] That said if somebody wanted to come up with a separate "python but all the slow parts are replaced with C modules like they would be in a real app" implementation, that could be interesting...