[kragen]

aha, i finally figured out that yes, indeed, my elisp code is being compiled to native code and saved as shared library files in /usr/lib/emacs/28.2/native-lisp/28.2-e4556eb6. the functions have names like F76632d63616c6c2d6261636b656e64_vc_call_backend_0@@Base (the hex string decodes to 'vc-call-backend'), and all the calls are indirected through registers, but `objdump -d` can decode it into recognizable amd64 assembly language, with function prologues and epilogues, nop-padding to 8-byte boundaries, tests followed by conditional jumps, that kind of thing. so emacs doesn't really have an excuse for being so slow

i thought that maybe the earlier version of left in c would be more efficient, so i git cloned https://git.sr.ht/~rabbits/left, checked out 4f127602e4e9c27171ef8f6c11f2bc7698c6157c, and built the last c version of the editor. a simple startup and shutdown cost 407 million instructions, and writing the same c program in it again in more or less the same way took 2½ minutes and 1.8 billion instructions. 1.4 billion instructions in 150 seconds are about 9 million instructions per second, which is surprisingly about the same as the version running in uxn. but just leaving it open for 2¼ minutes also used only 424 million instructions, so maybe it makes more sense to compare the 1.4 billion against the 8.8 billion of emacs, the 0.039 billion of ae, the 3.4 billion of uxn left, and the 0.68 billion of vim, since it seems to grow primarily with activity rather than just time open

[entaloneralie]

Hey kragen,

Someone just told me to have a look at the thread and I'm very happy I did! It's been really good for me to read back you two's exchange. I'm not here to defend uxn or anything like that, I was only wondering, could you do the same test with uxn11(instead of uxnemu)? I don't personally use uxnemu, I find it's too demanding for my laptop, I would love to have the data for uxn11 in comparison to uxnemu from your system if you can run X11 programs.

Let me know :) Cheers!

[kragen]

oh hi! delighted to hear from you! i hope it's clear i'm not here to attack uxn either

let's see about uxn11... initial signs are good, 175711 instructions to start up and shutdown rather than hundreds of millions. but after using it to go through the whole left editing session, it's 175645 instructions; that suggests the real work is being done in a child process

yeah, strace confirms there's a child process being spawned off. i'll have to see if i can disable that to get more accurate measurements with valgrind, later today

i'm interested to hear your thoughts about how left and other uxn apps might be made usable under the constraints of the zorzpad display hardware: two 400×240 screens, each 35×58mm, with only black and white (no greyscale). left itself seems like it might be relatively easy to run usably on one of them?

[entaloneralie]

Of course, I've really enjoyed your exploration of uxn, and all your comments are accurate.

Uxn11 has a special device to spawn linux processes(like playing a mp3 on disk with aplay), it can be disabled in devices/console, I wonder why it would be acting up, left doesn't do any request to the console's special ports I think, I'll have to double-check.

Left is pretty heavy, it does A LOT, it's definitely not a simple text editor, not only does it do a lot of work locating symbols and navigation strings, the front-end uses a proportional font and does a lot of positioning for it.

I don't think left would be a good candidate for zorzpad, I can imagine a simpler IDE that uses fixed width font, and doesn't try to make any sense of the tal code, syntax highlight, and doesn't support utf-8 glyphs - But Left is not it.

I've ported the classic macintosh notepad application, which I now use daily for taking notes, it has proportional font support, but is monochrome and has a small window, doesn't do anything fancy. Expanding this into a proper text editor with scrolling might be more realistic than trying to fit left to a monochrome screen.

https://git.sr.ht/~rabbits/uxn-utils/tree/main/item/gui/note...

But really, I think it'd be better to write a whole new text editor specifically for zorzpad, left has a very specific goal in mind, and I can imagine one like it designed specifically for the zorzpad. Writing text editors, or block editors, is an art that shouldn't be forgotten, anyone who's got an opportunity to write one, should do it.

I've read as much of your works as I could find, you've been a big influence on the work I do nowadays, it's an honor to read a message from you. Thank you.

[kragen]

aw, thanks, i'm really flattered—i admire your work a lot, and i'm glad to hear i've contributed somewhat to it. you're very welcome

as for proportional fonts, a few years back i wrote a microbenchmark for proportional font layout with word wrap to see how cheap i could get it. it's http://canonical.org/~kragen/sw/dev3/propfont.c (maybe make sure you redirect stdout to a file if you run it), and it uses the n×6 font i designed for http://canonical.org/~kragen/bible-columns, derived from janne kujala's 4×6 font. the notes in the comments say that on my old laptop it ran at 70 megabytes per second, which is probably about 60 instructions per byte. if you redrew 32 lines of text with 32 characters each (about what you could fit on the zorzpad's two screens) it would be about 60000 instructions, about a microsecond at the apollo3's theoretical 60 dmips; that might be a reasonable thing to do after each keystroke. and there are some notes in there about how you could optimize it more

running the same benchmark on my current laptop i get numbers from 64 to 68 megabytes per second, but for comparability with the editors, i should say that cachegrind measures 1,482,006,535 instructions for running 100000 iterations of rendering 126 bytes, which works out to about 118 instructions per byte and 120 000 instructions for the screen-redrawing example

(of course that computation takes about 2 microseconds, while actually updating the screen will take 50000 microseconds according to the datasheet, 16700 microseconds according to some other people's experimental results, so the computation may not be the tall pole in the tent here)

propfont.c might get a lot faster if its pixels were bits instead of bytes, especially on something like the thumb-2 apollo3, which has bitfield update instructions. then again, it would get slower with a larger pixel font size

[entaloneralie]

I'm currently away from reliable network and I can't manage to load the bible-columns image, I will try once we return to Canada.

I've added proportional fonts to Left because, after writing loads of THINK Pascal, I felt like I needed it, and I still enjoy this a lot more than fixed-width, but it annoys a lot of people, even even I like to write space-padded files from time to time so Left has to support both.

If I was golfing an editor(one that I would actually have to use each day), I think I would make a block editor, fixed-width, and by its block-paginated design would do away with gap buffers altogether by needed to move very little data around.

An early version of Left tried something different where it would only display one @symbol at a time. In Uxn, "scope" or objects are defined with @label, and methods by &method, making it so it's well usable if you have a list of objects, and methods a-la Smalltalk's System Navigator. During editing, you never insert into the text file, but in a gap buffer made of only the object or method being edited. I always found that design pretty neat, the main downsize is when you want to add non-source files, then you have to contend with editing to a single large buffer.

The font encoding for Left is https://wiki.xxiivv.com/site/ufx_format.html, which allows me to calculate the sprite address glyph width sort-of quickly, the issue is that varvara has no blitter, everything is drawn as tiles so it creates a bit of overdraw for each glyph. If you have a blitter, then I think there would be pretty smart encodings for drawing proportional fonts that would make them more efficient.

[pushfoo]

Can I consider the UXF format stable?

UXF was one of the intended eventual targets of a very bad font converter I need to eventually rewrite.

[kragen]

usually the issue with that image is not its byte size (it's 4 megs) but its ram usage (about 100 megapixels)

my best thought on how to handle proportional fonts is a slight variant on nick gravgaard's 'elastic tabstops'; i think we should use tabs to separate columns and \v and \f as delimiters to nest tables, so that we can do pre-css-style html nested table layout in plain ascii text files

[kragen]

with respect to moving very little data around, the bulk of the data that needs to be moved is pixels to display; a character is 1–3 bytes in the editor buffer, but in a 16-pixel-tall font, it averages about 128 pixels. that's 512 bytes in 32-bit rgba or 16 bytes in 1 bit per pixel. on the zorzpad, without interpretation overhead, i think i can roughly guesstimate about 100 instructions, and at a nominal 25 picojoules per instruction, that's about 2.5 nanojoules

i still haven't measured, but if i recall correctly, the ls027b7dh01 memory lcd datasheet says that maintaining the display statically costs about 50 microwatts, while inverting all the pixels once a second raises that to 175 microwatts, which is to say 125 microjoules to invert all 96000 pixels, which works out to 1.3 nanojoules per pixel. so updating those 128 or so pixels will cost about 170 nanojoules, which is a lot more than 2.5 nanojoules

i'm unclear on how much of this 125 microjoules is due to simply shifting in new lines of 400 pixels, regardless of their contents, and how much is due to the actual inversion of color of the pixels; after all, they invert polarity several times a second just to maintain the display without damaging it. if the expensive part is shifting in the pixels, then to display a single new 16-pixel-high letterform, we're updating 16 × 400 = 6400 pixels rather than 128, costing 8300 nanojoules. typing at 60 words per minute (5 letters plus a space) would then dissipate another 50 microwatts if you fully updated the display after every letter. when you're typing into previously empty space, you could maybe update only a third of the pixel rows after each letter to cut down on this, completing the delayed updates after a few hundred milliseconds if you stop typing

it might turn out that the most important thing to do to minimize power usage is to minimize the amount of the screen that gets regularly updated, and last night i was thinking about how in bsd talk and on mit its, text would wrap around from the bottom of the window back to the top rather than scrolling. emacs does a less aggressive version of this where it does scroll, but it scrolls by half a screenful at once when the cursor goes off the screen, so that updates to the whole screen only happen every ten or twenty lines of movement or typing, rather than after every line. this was important on terminals like the adm3a that didn't have escape sequences to insert and delete lines

narrowing the view to a single symbol might avoid the need to spend energy drawing lines of text from other symbols, perhaps repeatedly as line insertion or deletion moves them around on the screen

the adm3a also didn't have escape sequences to insert or delete characters, and vi (strongly influenced by the adm3a) does a peculiar thing with its 'c' change command; if you say, for example, ct. to change all the text before the following period, it doesn't erase that text immediately, but puts a $ at the end of it so that you know what you're changing. this avoids the need to shift the rest of the line right after each character. (vim does not do this.) if the big energy suck is actually changing pixels, as i think it is on e-ink, rather than shifting them through a shift register, that might be a useful strategy to adopt: open up space to type into, with some kind of indicator that it doesn't contain spaces or any other characters. like a visible manifestation of a buffer gap, but measured in pixels rather than bytes

i don't think there's a reasonable way to hack a hardware blitter into the zorzpad, but doing it in software is very reasonable