Today I get to do something that’s exceedingly fun for me: “well, actually” my old self. In this case, I get to post a correction to my recent Dyalog APL Sudoku solver.
A few days ago, Adám Brudzewsky from the Dyalog APL team reached out with detailed notes and a cleaned-up rewrite. He had some very kind words about my solution, but pointed out some things to improve, which is just incredible.
He also graciously granted me permission to quote and attribute, so here we go.
In the original post I leaned into golfing to make the code feel “APL-like”. That aesthetic is definitely common online, but Adám reminded me that it doesn’t have to be golfed to look APL-like, and that production APL tends to be much less dense. That’s a fair point. I got carried away a bit when shrinking my solution, and truth be told, it was quite fun! But I feel like it’s important to point out that that’s optional.
The more useful lesson is: APL rewards good representations. If you pick shapes that align with array operations, the code often becomes short without trying. I tried anyway, but I should disclose that it isn’t a requirement for idiomatic APL.
Adám also came to me with some concrete improvements that I want to dig into to make sure I understand them correctly.
Adám suggested a tighter S that removes redundant
parentheses (in line with his
style guide), replaces my “find minimum” with a sort-based pick,
simplifies how @ is applied, and uses a more tacit step for
picking the first solution:
S←{
b←P ⍵ ⋄ 0=≢b:⍬ ⋄ ~0∊b:b
i←⍸0=b ⋄ cs←↑b∘C¨i ⋄ l←+/cs ⋄ k←⊃⍋l ⋄ idx←i[k]
sols←⊂∘∇¨⊣@idx∘b¨cs[k;]/D
⊃(×∘≢∘⊃¨⍛/sols),⊂⍬
}I also got a nicer candidate function C: it avoids
explicit box-index arithmetic by extracting the 3x3 box via drop/take
(↓/↑), which reads very “APL” once you know
the idiom (which, uh, I didn’t really):
C←{
m←9 9⍴⍺ ⋄ r c←9 9⊤⍵
~D∊0~⍨∪m[r;],m[;c],,3 3↑(3×⌊r c÷3)↓m
}If you compare this to my original C, the big change is
that the box is obtained by:
3×⌊r c÷3)↓m)3 3↑…)Much cooler.
My original solver boxed sols just to unbox it again.
This was a defensive step against shooting myself in the foot with my
own limited understanding. Dyalog may sometimes mix same-shaped results
into a plain array because it can—and I don’t understand the rules it
follows completely—which breaks a pipeline that filters non-empty
solutions and then takes the first remaining one.
The rewrite above makes the selection step cleaner by doing away with the need for explicit unboxing. It feels less brittle and is more readable to boot!
Thanks again to Adám for the review. He mentioned he’s reachable in the APL Orchard (which I understand he started!). He was exceedingly kind and helpful, so I encourage anyone who’s even remotely interested in APL to hop in and say hi!