Comment by ribit
4 years ago
I don't know about you, but I was really laughing out loud reading that GitHub conversation.
GPUs: able to render millions of triangles with complex geometrical transformations and non-trivial per-pixel programs in real time
MS engineers: drawing colored text is SLOW, what do you expect
P.S. And yes, I know, text rendering is a non-trivial problem. But it is a largely solved problem. We have text editors that can render huge files with real-time syntax highlighting, browsers that can quickly layout much more complex text, and, obviously Linux and Mac terminal emulators that somehow have no issue whatsoever rendering large amount of colored text.
To be fair to the MS engineers, from their background experience with things like DirectText, they would have an ingrained rule-of-thumb that text is slow.
That's because it is slow in the most general case: If you have to support arbitrary effects, transformations, ligatures, subpixel hinting, and smooth animations simultaneously, there's no quick and simple approach.
The Windows Terminal is a special case that doesn't have all of those features: No animation and no arbitrary transforms dramatically simplifies things. Having a constant font size helps a lot with caching. The regularity of the fixed-width font grid placement eliminates kerning and any code path that deals with subpixel level hinting or alignment. Etc...
It really is a simple problem: it's a grid of rectangles with a little sprite on each one.
Supercomputer or not, it's a terminal.
In the real-CRT-terminal days of the 1970's & 1980's of course the interface to the local or remote mainframe or PC was IO-bound but not much else could slow it down.
UI elements like the keyboard/screen combo have been expected to perform at the speed of light for decades using only simple hardware to begin with.
The UX of a modern terminal app would best be not much different than a real CRT unit unless the traditional keyboard/display UI could actually be improved in some way.
Even adding a "mouse" didn't slow down the Atari 400 (which was only an 8-bit personal computer) when I programmed it to use the gaming trackball to point & click plus drag & drop. That was using regular Atari Basic, no assembly code. And I'm no software engineer.
A decade later once the mouse had been recognized and brought into the mainstream it didn't seem to slow down DOS at all, compared to a rodent-free environment.
Using modern electronics surely there should not be any perceptable lag compared to non-intelligent CRT's over dial-up.
Unless maybe the engineers are not as advanced as they used to be decades ago.
Or maybe the management/approach is faulty, all it takes is one non-leader in a leadership position to negate the abilities of all talented operators working under that sub-hierarchy.
Exactly. Fast terminal rendering on bitmap displays has been a solved problem for at least 35+ years. Lower resolutions, sure, but also magnitudes slower hardware.
It's more subtle than that. What the Microsoft engineers are saying is that the console's current approach to drawing text is inherently slow in this particular case, due to the way the text drawing library it's based on uses the GPU. The proposed solution requires the terminal to have its own text drawing code specific to the task of rendering a terminal, including handling all the nasty subtlties and edge-cases of Unicode, which must be maintained forever. This is not trivial at all; every piece of code ever written to handle this seems to end up having endless subtle bugs involving weird edge-cases (remember all those stories about character strings that crash iPhones and other devices - and the open source equivalents are no better). It's relatively easy to write one that seems to work for the cases that happen to be tested by the developer, but that's only a tiny part of the work.
I’m fairly sure noone in question wrote a new font renderer, but just rendered all available fonts upfront with a system library, and uploaded it to the GPU and let it use it as a bitmap.
Text rendering is still done mostly on the CPU side in the great majority of applications, since vector graphics are hard to do efficiently on GPUs.
Simply shaping text using state of the art libraries (like harfbuzz) can take an INCREDIBLE amount of time in some cases. If you're used to rendering text in western character sets you may think it can always be fast, but there are cases where it's actually quite slow! You get a sense for this if you try to write something like a web browser or a word processor and have to support people other than github posters.
Of course in this case it seems like it was possible to make it very fast, but people who think proper text rendering is naturally going to be somewhat slow aren't always wrong.
Saying that text rendering is "largely solved" is also incorrect. There are still changes and improvements being made to the state of the art and there are still unhappy users who don't get good text rendering and layout in their favorite applications when using a language other than English.
> naturally going to be somewhat slow
Naturally slow for a text renderer might mean it renders in 4ms instead of 0.1.
Yes, and for a terminal like discussed in the article, 4ms is considered unacceptably slow
1 reply →
You are right in the general case. But terminals are a specific niche, not requiring the full extent of text rendering edge cases as a browser, wysiwyg editor, etc “experience”. It renders “strictly”* monospaced fonts, which makes it trivial to cache and parallelize.
* as it was brought up, one might use a non-monospace font, but that case can just use the slow path and let the “normal” people use a fast terminal