Re: Convolution based regtest comparison

On Sat, Jun 21, 2025 at 10:54 PM Luca Fascione <[email protected]> wrote:
> Here are some partial timings, taken on my mac laptop (2019, 2,3 GHz 8-Core Intel Core i9),
> all code single-threaded.
>
> Here's the execution log:
>
> Parsing scene '/Users/lukes/Documents/src/undvips/tex/schintro_outline.udps'
> Parse complete 1156.21 ms


This is plenty fast. IIRC, lilypond itself takes about 240s to run
through ~1200 regtests, so as long as time per test is around or below
200ms, we should be fine.

> Missing bits: if people think this idea has merit, we need something to stand in the role of my little UDPS file format. I see a couple different avenues

It doesn't matter too much, but if I may suggest a shade of bikeshed,
something based on json would be nice, so we can easily load data
across languages/systems.

>1. [..] Measure its performance (it seems to fail on GPL Ghostscript 10.05.1, I don't know why at the moment). Regrettably I have no notes on how long this step took when I was last working on this
>
> 2. Alternatively, and possibly much better, pivot and implement a similar idea using mupdf library, but using the PDF directly, sidestepping the PostScript entirely
> 3. Alternatively still, instrument some late layer of ... "Cairo" (say) to emit UDPS directly from the C++

If your PS hack provides final coordinates, I think that will be the
least amount of work. While mupdf is a nice, reading the dicts from
the PDF file is not enough: you still have to write a 2D graphics
stack so you can handle transformations (rotation, scaling,
translation), clipping, coloring etc.

In theory, 3. is also a nice approach, except Cairo doesn't support
this type of instrumentation (at least, I couldn't find any support
for it). You'd have to add a feature to Cairo in order to capture
commands after all the transformations have been applied, and just
before they go onto the raster.

> Lastly, my little test is already built to support multiple pages, and this could be advantageous in that it would save us to run
> all the warmup code over and over again. I guess I'm saying: if we can make it work for the whole set at once we could be looking at

which warmup are you concerned about? Except for font-loads, lilypond
outputs self-contained (e)ps in the ps backend. In cairo, the files
are fully contained. So you have to process the page and font setup
repeated for each file.


--
Han-Wen Nienhuys - [email protected] - http://www.xs4all.nl/~hanwen