Bryce Miller (Sep 16 2022 at 21:11): Bryce Miller (Sep 16 2022 at 21:11):I plotted the mandelbrot set using Roc Screen-Shot-2022-09-16-at-5.10.32-PM.png
Bryce Miller (Sep 16 2022 at 21:12):Would be very curious to do a performance comparison with other languages. I'm using the dumbest escape time algorithm possible.
Richard Feldman (Sep 16 2022 at 21:29):cooooooooool!!!
Brian Carroll (Sep 16 2022 at 21:29):Nice!
Bryce Miller (Sep 16 2022 at 21:29):I really want to try doing this properly perhaps using that Bevy platform @JanCVanB started working on.
jan kili (Sep 16 2022 at 22:10):Do you mean this Plotters platform? https://github.com/JanCVanB/roc-plotters
jan kili (Sep 16 2022 at 22:10):I'm currently updating it to work with latest Roc, so this is good timing
jan kili (Sep 16 2022 at 22:11):(huh, why did I name the Bevy one "bevies", plural? :thinking: foolish https://github.com/JanCVanB/roc-bevies)
jan kili (Sep 16 2022 at 22:12)::point_of_information: :surprise: Whoa, that ascii mandelbrot is PRETTY, great job!
Bryce Miller (Sep 16 2022 at 22:25):Thanks. I had to set the complex plane to stretch to fit the plot, and ensure that the plot was exactly three times as wide as it is tall.
Bryce Miller (Sep 16 2022 at 22:30):I saw the roc-plotters platform, but when I saw the line graphs I figured it might not be a good fit for plotting fractals. Maybe I was wrong about that tho! I figured the Bevy platform used for the breakout example would be a way to render actual pixels.
I don't know Bevy at all but if it provided access to the GPU that would really unlock some performance. These plots are float-heavy.
Bryce Miller (Sep 16 2022 at 22:33):When I get some time, I'd like to do ASCII versions in JS, Rust, and Haskell too. I'm extremely curious how Roc would compare considering how it did with quicksort.
Bryce Miller (Sep 16 2022 at 22:37):I doubt I am smart enough to do it, but it would be interesting to see how close we could get to Kalles Fraktaler with Roc
Bryce Miller (Sep 16 2022 at 22:38):But yeah if you think the plotters platform will work for this once you have it updated, I'll definitely do a version with that!
jan kili (Sep 16 2022 at 22:41):Take a look at the scatter example, it hints at how the Plotters lib can do WAY more than I integrated with
jan kili (Sep 16 2022 at 22:46):And the Plotters rust library in general - it might be simpler to extend that platform than bevy, but only if you have no bevy experience like I don't
jan kili (Sep 16 2022 at 22:46):Either way, help advancing any aspect of those visualization platforms is very welcome!!
jan kili (Sep 16 2022 at 22:47):Good point about the GPU...
jan kili (Sep 16 2022 at 22:47)::shrug:
jan kili (Sep 17 2022 at 01:31):@ me with any questions or ideas! I'm happy to review code, contribute, etc
Bryce Miller (Sep 17 2022 at 02:07):Sweet!
Bryce Miller (Sep 17 2022 at 19:03):Update: created a repo and added a couple additional implementations in Rust and JS. https://github.com/sandprickle/fROCtals
Bryce Miller (Sep 17 2022 at 19:09):Set up all three implementations to create identical plots, with max iterations set to 1M.
On my M1 Macbook Air, I got the following results:
JS (Node v16): ~10.3 s
Roc: ~ 4.62 s
Rust: ~ 4.57 s
So my question is: Is Roc just that fast, or did I fail to optimize the Rust version properly?
Folkert de Vries (Sep 17 2022 at 19:17):doesn't look like anything is obviously inefficient in the rust code
Folkert de Vries (Sep 17 2022 at 19:17):so if you ran a release build then yes
Folkert de Vries (Sep 17 2022 at 19:18):the code here is very simple, so roc and rust should produce pretty much the same thing (because both use llvm)
Bryce Miller (Sep 17 2022 at 19:22):Cool. Would you consider this to be an algorithm that would typically be pathological for a pure functional language, or not so much?
IOW, is this a good showcase of the optimizations you’ve done with the Roc compiler?
Folkert de Vries (Sep 17 2022 at 19:29):not in terms of performance I think? but the sort of "textbook implementation" is easier in roc I think
Folkert de Vries (Sep 17 2022 at 19:30):I'd be interested to see how haskell fares, in particular when you don't try too hard to optimize
Bryce Miller (Sep 17 2022 at 19:33):I was thinking I’d do one in Haskell next
Bryce Miller (Sep 17 2022 at 19:34):Do you think something along the lines of my Roc implementation would be sufficiently naive for the Haskell implementation?
Folkert de Vries (Sep 17 2022 at 19:34):yes I think that would be interesting
Bryce Miller (Sep 17 2022 at 19:34):Cool. That’s next on my list then.
Folkert de Vries (Sep 17 2022 at 19:35):in particular: we see how ghc does versus llvm on this problem, and how good/bad the roc stdlib is versus the haskell one
Bryce Miller (Sep 17 2022 at 19:41):I made it to chapter 4 or so of Graham Hutton’s Haskell book, and that was probably a year ago or more. We’ll see how far I get with it.
Bryce Miller (Sep 17 2022 at 19:41):So standard prelude is what you’re more curious about? That’s what Haskell calls their stdlib right?
Folkert de Vries (Sep 17 2022 at 19:41):yeah
Bryce Miller (Sep 17 2022 at 19:42):Cool, I don’t know how to use alternate preludes anyway 😅
Folkert de Vries (Sep 17 2022 at 19:42):one advantage that the more modern preludes may have is that they use Text instead of String (i.e. a linked list of characters, a bad idea)
Folkert de Vries (Sep 17 2022 at 19:43):for the rest of the code, I don't think a more modern prelude would change anything
Folkert de Vries (Sep 17 2022 at 19:43):but haskell does have mutable vectors in the prelude somewhere, which might be an improvement, but those are so obscure that I'd say it doesn't really count as idiomatic haskell
Bryce Miller (Sep 17 2022 at 19:44):I was about to ask whether I should use the default list or one of the vectors.
Bryce Miller (Sep 17 2022 at 19:44):I’m curious to try both.
Bryce Miller (Sep 17 2022 at 19:46):Using a list puts idiomatic Haskell against idiomatic Roc, while using a vector would highlight more of the overhead/optimization differences when using the same data structures.
Bryce Miller (Sep 17 2022 at 19:47):Or similar, I guess. Not sure if haskell vectors would be on the stack like the arrays I used in the Rust version.
Folkert de Vries (Sep 17 2022 at 19:53):I don't think they would be
Richard Feldman (Sep 17 2022 at 21:23):@Bryce Miller that's a very cool result! Would you mind if I share a screenshot of your post with the numbers?
Bryce Miller (Sep 17 2022 at 21:25):@Richard Feldman Don't mind at all! I also have some results from my desktop (Ryzen 3800X) which shows a larger difference between Roc and Rust.
Bryce Miller (Sep 17 2022 at 21:28):3800X results (Fedora)
JS: ~ 9 s (not super consistent)
Roc: ~ 3.16 s
Rust: ~ 2.86 s
Bryce Miller (Sep 17 2022 at 21:29):Odd that JS didn't get much faster :thinking:
jan kili (Sep 17 2022 at 21:47):How was/is your initial/current developer experience? Docs, syntax, compiler, etc
Richard Feldman (Sep 17 2022 at 21:57):@Bryce Miller I bet the gap will be much bigger if you do like 1000 elements, because then the jit wouldn't kick in
Richard Feldman (Sep 17 2022 at 21:58):(the gap between JS and the others, I mean - Roc doesn't use a JIT)
Bryce Miller (Sep 17 2022 at 22:34):@Richard Feldman You mean using a smaller array, or fewer iterations, or both? Maybe I can add a cli arg to all the implementations to set the iteration count.
Bryce Miller (Sep 17 2022 at 22:41):@JanCVanB the biggest thing I wanted was hover definitions so I could check type signatures. Format on save would be really nice, but at least Roc has a built-in formatter. I’m still fighting to get a haskell formatter working with nvim. Coffeescript syntax highlighting kinda breaks with backpassing, so I switched to the VScode extension that someone made. Less robust, but at least I don’t get a lot of red with backpassed functions.
Speaking of backpassing, that’s going to take some getting used to. Coming from Elm, a lot of the other characteristics of the language are pretty intuitive. Backpassing syntax still hurts my brain a little bit.
Folkert de Vries (Sep 17 2022 at 22:43):you don't need to use it, especially for the list stuff
Bryce Miller (Sep 17 2022 at 22:44):Yeah, I don't feel like this little script really needed it. I included it in one place just to try getting used to the syntax. Once my brain is able to decode it quickly, I think it will be a very welcome feature. Indentation hell is real.
Bryce Miller (Sep 17 2022 at 22:46):On that note, really grateful that local values don't have an additional level of indentation like they do in Elm's let...in.
Bryce Miller (Sep 17 2022 at 22:47):Oh, and docs were pretty good. The function signature for List.mapWithIndex I think could use a set of parens for the callback tho?
Richard Feldman (Sep 17 2022 at 22:47):@Bryce Miller I meant smaller number of iterations - like 1000 instead of 1M
Folkert de Vries (Sep 17 2022 at 22:48):wow a time-traveling message
Bryce Miller (Sep 17 2022 at 22:49):Might be useful to have a link to roc-lang.org/builtins tutorial though? I had to rummage through Zulip to find it.
Richard Feldman (Sep 17 2022 at 22:49):yeah that was weird haha, not sure why it repeated much later
Richard Feldman (Sep 17 2022 at 22:49):@Bryce Miller that's a good call! I can add a link to it in the readme
Bryce Miller (Sep 17 2022 at 22:49):you've heard of time-traveling debuggers. now get ready for... time traveling zulip chats!
Bryce Miller (Sep 17 2022 at 22:50):I'll do some runs with 1k iterations and see what happens!
Folkert de Vries (Sep 17 2022 at 22:53):btw hyperfine is an excellent tool for running benchmarks like this
Folkert de Vries (Sep 17 2022 at 22:54):don't want to be doing eyeball statistics
Bryce Miller (Sep 17 2022 at 23:05):I think I was looking at user time only, not total time for the numbers I provided above (using the time command I think built in to zsh?). I'll definitely re-run these benchmarks using hyperfine. Thanks for the suggestion!
Looking at total time using the time command, 1k iterations makes JS roughly an order of magnitude slower. Roc and Rust are hovering around 0.005s, where JS is around 0.045s.
Brendan Hansknecht (Sep 17 2022 at 23:53):I wonder how much of that is js engine startup time and such.
Brendan Hansknecht (Sep 17 2022 at 23:55):Like how long does Js take with 0 and 1 loop iterations?
Bryce Miller (Sep 18 2022 at 02:06):Added some benchmarks via hyperfine to the repo.
Bryce Miller (Sep 18 2022 at 02:08):I’ll try to do some with 1 iteration too. I mainly included JS as a reference “slow” language.
Brendan Hansknecht (Sep 18 2022 at 02:11):Awesome. Interested to see the results. Can make some pretty graphs as the number of iterations increase
Bryce Miller (Sep 18 2022 at 02:23):I have the Haskell implementation nearly done too. Either I messed up the algorithm or I’m making some rookie Haskell mistakes. Or both.
Bryce Miller (Sep 18 2022 at 17:45):Screen-Shot-2022-09-18-at-1.44.25-PM.png Apple M1 results with 1K iterations including Haskell, via hyperfine
Bryce Miller (Sep 18 2022 at 17:46):Screen-Shot-2022-09-18-at-1.45.59-PM.png AMD Ryzen 3800X results including Haskell, via hyperfine
Bryce Miller (Sep 18 2022 at 17:47):Using default Prelude for Haskell, no vectors. Completely unoptimized to the best of my knowledge.
Bryce Miller (Sep 18 2022 at 17:48):Next up is to accept cli param for iteration count so I can use hyperfine to generate data
Brendan Hansknecht (Sep 18 2022 at 17:57):From a random post online about Haskel:
There is no such thing as "release mode". Perhaps you want to enable more optimization? In that case, pass in --ghc-options -O2. Alternatively, add -O2 to ghc-options: in your cabal file.
Bryce Miller (Sep 18 2022 at 18:16):Oh interesting. I was wondering whether there was a flag I should be using for a production build. I know almost nothing about Haskell.
Bryce Miller (Sep 18 2022 at 18:20):Ok that's actually a substantial improvement.
Bryce Miller (Sep 18 2022 at 18:23):Screen-Shot-2022-09-18-at-2.22.50-PM.png Apple M1 with -O2 GHC option
Brendan Hansknecht (Sep 18 2022 at 18:26):Oh wow, about 2x faster
Bryce Miller (Sep 18 2022 at 18:27):Yeah pretty surprising.
Bryce Miller (Sep 18 2022 at 18:29):Would a substantially larger plot hurt the linked lists in Haskell more? Maybe I can do some supersampling if the larger memory footprint would reveal something interesting.
Brendan Hansknecht (Sep 18 2022 at 19:05):I depends on the algorithm. Haven't actually read through your code, but i wouldn't expect any huge slowdown from larger lists
Brendan Hansknecht (Sep 18 2022 at 19:05):But again, heavily depends on how the lists are being used
Last updated: Jul 06 2025 at 12:14 UTC