@jakob really enjoyed this post; thanks for sharing. I have a bunch of questions but first of all I have to ask: what on earth was it like to have your dad teach you Emacs at age nine? (I have two kids age ten and twelve, and I've been teaching them TIC80 and love2d but sticking with more conventional editors so far)

@technomancy Haha, I knew someone would mention that. I wasn't technically apt enough at that age to have learned the typical CUA-style keyboard shortcuts, so learning it didn't feel foreign or unnatural in the way that it might for an experienced programmer being introduced to Emacs. I was mostly writing HTML by hand at the time, and I remember being able to use the menu bar to insert tags. It was a nice experience.

I should find my old laptop and see if I can pull the config off of it.

@technomancy I think you're doing the right thing by sticking to conventional editors for right now. My situation probably made more sense for the time -- I'm not sure of any other accessible options that existed when I was nine years old.

@jakob haha, awesome.

anyway, a few notes about the Fennel section.

> Unfortunately, you can't destructure everywhere. You can't write this, for example

the reason that example fails is not because each doesn't support destructuring but because the arguments are in the wrong order; put the _ first and it'll work fine.

I suspect the reason luajit is running so slowly for you is that the JIT hasn't had time to warm up and perform optimization traces. does the performance improve if you run (say) thousands of passes instead of just one?

> there's no way to send something from a source code buffer to the REPL.

C-c C-k should reload the entire module once you have a repl buffer open. The repl buffer just uses inferior-lisp-mode, so all the normal bindings for that should work the same. if you run into issues with that please open an issue!


> I suspect the reason luajit is running so slowly for you is that the JIT hasn't had time to warm up and perform optimization traces. does the performance improve if you run (say) thousands of passes instead of just one?

I actually thought the luajit performance was quite good. I wanted to avoid rigor here because I'm a bit of a statistics noob, but I'll do a run where I try to get to i.i.d iterations. I should probably do something similar for Racket, which is also JITed.

@technomancy As for the other comments, I'll get working on putting those into the errata. Thanks a bunch!

@jakob I wrote a similar "comparative lisps" post in 2013 but it was quite a bit more superficial: https://technomancy.us/169
@jakob yeah, the problem is that measuring performance is such a nuanced problem; it's hard to do a comprehensive job.

that said, with tuning LuaJIT should be able to approach within 1.1x the speed of C++ according to this guy who built a smalltalk VM in it: https://old.reddit.com/r/lua/comments/htqn0t/luajit_once_again_nearly_as_fast_as_the/

@technomancy For sure. I've done a bit of reading on statistically-sound benchmarking, since we do some comparisons of sequence alignment algorithms in my lab, but I still defer to my statistician friends for help with the design and interpretation. I thought the most appropriate thing to do was to just wave hands and give the disclaimer that it is complete hand-waving.

Benchmark is still running at the moment. Ten iterations since I've started.

@technomancy One difference between that success story and mine is that I haven't done much to optimize my raytracer. In all implementations, it's a rather naïve translation of the underlying equations.

@jakob @technomancy Speaking of naive, I'm wondering now how much sense it would make to contribute an Emacs Lisp solution...

@wasamasa @technomancy I considered it out of morbid curiosity. I'm sure the Common Lisp version wouldn't be too hard to port over

@jakob @technomancy I thought so as well, but nope. There are glaring deficiencies in cl-lib/eieio that make this surprisingly hard:
- with-slots is incompatible with cl-defstruct and expects cl-defclass.
- cl-defclass can be made to create functions like cl-defstruct, but fails when using cl-defmethod because it for some reason creates predicates and clashes with what cl-defclass generated.
- with-slots can be made to work with cl-defstruct, but my macrology is more of a sham than shim and fails with even wilder errors than with the previous approach.

I'm pondering to instead rewrite this in a more idiomatic way without option types, but that would require actually understanding the code. Maybe another time.

@wasamasa @technomancy thanks for trying, at least. Perhaps I'll have a go at it this weekend when I have a little more time.

@technomancy I've incorporated your corrections to the article. My erratum is becoming more of a "change log" at this point, since I worry people would take my word for it and ignore the erratum. I'd feel bad misleading people about Fennel.

Pending results on benchmark, which I'll incorporate to the article when I get them. I've moved it to a different machine that doesn't have a faulty RAM stick.

@technomancy Running it on the machine with `time' had a real duration of 0m57.982. Using lua_benchmarking,

Mean: 86.477030 +/ 9.725198, min 59.790374, max 108.472983

So... I suspect there are some confounding factors, but there isn't a massive, noticeable difference coming from the JIT warmup.

I tried with --nogc, but it panicked rather quickly. Again, I suspect this is more of an issue with my code than LuaJIT or Fennel. I seem to be consing quite a bit.

[1]: github.com/softdevteam/lua_ben

@jakob some style suggestions: https://p.hagelb.org/0001-Some-style-changes.patch.html

mostly based on a cursory reading of the code. removing pack in particular seems to have a noticeable performance improvement impact; it went from 2m30s to 2m12s on my machine in a very unscientific test. since you mentioned a lot of consing is happening I'm going to test it out with the new generational GC in Lua 5.4 to see how that changes things.

@technomancy thanks for the patch! I'll merge this in.

Since you're actually looking at the code and bound to notice it at some point, I'm going note that `degrees->radians' is a really embarrassing mistake I made while going through my university notebook, seeing π/360 in a formula, and thinking "oh yeah it's converting to radians". It works fine because I'm not using it to convert any values outside of that formula, but it's probably very confusing if you were to come across the function.

Sign in to participate in the conversation
Mastodon @ SDF

"I appreciate SDF but it's a general-purpose server and the name doesn't make it obvious that it's about art." - Eugen Rochko