TeXLayout.jl - possible successor to MathTeXEngine.jl?

[dawbarton]

I wanted to let you know about a new package I've been working on (rather, that I've been getting Claude to work on): TeXLayout.jl, an OpenType-aware LaTeX math typesetter for Makie.

I started from MathTeXEngine but fairly quickly concluded that achieving what I was after would require rewriting essentially all of it. So a new package made more sense than a PR. The main differences:

• Layout constants come entirely from the font's OpenType MATH table (no hard-coded fallbacks), so metrics are always consistent with the font in use.
• Any OpenType Math font works out of the box; eight families (New Computer Modern, STIX Two, Fira Math, TeX Gyre variants, Luciole) are bundled as lazily-downloaded Julia Artifacts — zero configuration for end users.
• Full TeX style cascade (D/T/S/SS with cramped variants), correct inter-atom spacing, italic correction, auto-sized delimiters, matrices, horizontal braces, \boldsymbol, and more.

It integrates with Makie via a package extension that intercepts generate_tex_elements(::LaTeXString), so existing Makie code works without changes (but type piracy abounds - I'd want to fix this if possible).

It's still early days with the code, but it's got pretty good TeX math coverage. I still need to wire up quite a few bits and pieces (including CI...).

I wanted to ask what you think makes most sense going forward. A few options as I see them:

1. Deprecate MathTeXEngine in favour of TeXLayout
2. Fold TeXLayout into MathTeXEngine — replace the layout engine under the existing package name.
3. Maintain both independently
4. Something else entirely?

No pressure at all, happy to discuss. (I haven't registered this package yet as I wanted to get your thoughts before doing that.)

[dawbarton]

See below for a stress test of the package. (Note I need to fix what's going on with some of the fonts since they use different names internally.)

[Kolaru]

That looks stellar!

If there is little to no overlap in the implementation (it looks like even the parser is also reimplemented), I would suggest keeping the two packages separated. Then, it will be up to the Makie dev (or others) to choose between their favorite backend. I would recommend opening a similar issue there to see their opinion, as they would be the main users (with maybe Luxor.jl).

I doubt that anyone will want to keep both backend available. Once that's clear, I'm happy to deprecate MathTeXEngine.jl, since TeXLayout aims at being a strict improvement (and I started this project to get access to underscore z, my love for TeX is in fact limited, I'm happy to let you take over with your implementation :D).

That being said, I don't know how the two fully compare currently, either in term produced image or performance (@ffreyer has recently been looking into the latter, I don't know if it has become a concern on the Makie side).

[dawbarton]

My day job means that TeX will remain useful to me for quite a while… Though I am looking forward to the day that Typst takes over, even if TeX notation is much more familiar to me.

Good question about the performance. I haven’t checked but it is something in the back of my mind to think about.

Once I’ve finished some of the wiring (CI, docs, etc), I’ll make an issue over at Makie / Luxor.

Thanks for all your work on MathTeXEngine so far!

[dawbarton]

As a side note, I got Claude to compare the master of MathTeXEngine against the master of TeXLayout for performance (in case @ffreyer is interested). The results are below. (I added some caching of font tables.)

Benchmark comparison: the TeXLayout Makie path is faster than plain MathTeXEngine on both test cases, with a much larger win on the more complex formula.

Case	Engine	Median time	Memory	Allocs
Small x^2+y^2=z^2	MathTeXEngine	155,864 ns	74,160 B	1,568
Small x^2+y^2=z^2	TeXLayout extension	48,040 ns	21,528 B	192
Large \int_0^\infty e^{-x^2}\,dx = \frac{\sqrt{\pi}}{2} + \sum_{n=1}^{\infty} \frac{(-1)^n}{n^2}	MathTeXEngine	1,131,836 ns	251,256 B	5,747
Large \int_0^\infty e^{-x^2}\,dx = \frac{\sqrt{\pi}}{2} + \sum_{n=1}^{\infty} \frac{(-1)^n}{n^2}	TeXLayout extension	147,548 ns	51,608 B	439

Relative result

• Small formula: TeXLayout is ~3.2x faster, with ~3.4x less memory and ~8.2x fewer allocations
• Large formula: TeXLayout is ~7.7x faster, with ~4.9x less memory and ~13.1x fewer allocations

This was measured in two separate Julia sessions: one with plain MathTeXEngine only, and one with TeXLayout + MathTeXEngine + GeometryBasics so the Makie extension path was active. The comparison is Makie-facing and end-to-end through MathTeXEngine.generate_tex_elements(::LaTeXString), so it is the relevant steady-state benchmark for repeated rendering calls.

[ffreyer]

That being said, I don't know how the two fully compare currently, either in term produced image or performance (@ffreyer has recently been looking into the latter, I don't know if it has become a concern on the Makie side).

That was more coincidental. I was benchmarking text to see if using SpecApi infrastructure for LaTeX lines would impact performance and noticed that the MathTexEngine was eating up way more time with joinpath/ispath calls. Without fixing that I can't really compare what I actually wanted to compare.

Benchmark comparison: the TeXLayout Makie path is faster than plain MathTeXEngine on both test cases, with a much larger win on the more complex formula.

Is this MathTexEngine 0.6.7? With 0.6.8 I get roughly 75µs and 330µs for the examples.

[dawbarton]

Is this MathTexEngine 0.6.7? With 0.6.8 I get roughly 75µs and 330µs for the examples.

Yes, good catch. I forgot that I was using my fork rather than the original repo, so it would have been 0.6.7 (or earlier). Updated numbers using current main branch (from my laptop rather than the machine I was working on above):

Expression	Engine	Time	Allocs
Small x^2+y^2=z^2	MathTeXEngine	76 μs	1569 allocations: 72.58 KiB
Small x^2+y^2=z^2	TeXLayout	21 μs	194 allocations: 22.23 KiB
Large \int_0^\infty e^{-x^2}\,dx = \frac{\sqrt{\pi}}{2} + \sum_{n=1}^{\infty} \frac{(-1)^n}{n^2}	MathTeXEngine	607 μs	5748 allocations: 245.60 KiB
Large \int_0^\infty e^{-x^2}\,dx = \frac{\sqrt{\pi}}{2} + \sum_{n=1}^{\infty} \frac{(-1)^n}{n^2}	TeXLayout	71 μs	438 allocations: 51.41 KiB

Not sure why I get a similar time for MathTeXEngine as you for the small example but not the large one 🤷