Max Godefroy

Graphics & engine programmer

[Max Godefroy @MaxGodefroy]

@lisyarus I also abstracts some of the more complex aspects like resource barriers which, while important to learn at some point, may be too much for a starter course.

[Max Godefroy @MaxGodefroy]

@lisyarus I think I lack experience using WebGPU to give you a definitive answer, but I think that going with a more modern API is definitively the way to go.
Most of these APIs share a lot of common features and WebGPU is a higher level abstraction for learning as a student.

[Max Godefroy @MaxGodefroy]

In that regard, I prefer NEON a lot. It has a few extensions but I don't need them as the ISA already has a ton of features.
Maybe ARM was able to design it a lot better thanks to insight from the x86 mess. Also they didn't need to make two manufacturers agree to common specs.

[Max Godefroy @MaxGodefroy]

While I'm at it, I want to rant a bit about how bad the SIMD support is for x86_64 processors IMO. By default, only SSE2 is assumed to be supported, but then a lot of features actually require additionnal ISA support.

[Max Godefroy @MaxGodefroy]

I've been doing some background work on my engine for the last few weeks. I added Github workflow support to make sure I push less breaking changes to the codebase. Testing both the build of the engine core, its modules, and the samples; then I run all available unit tests.

[Max Godefroy @MaxGodefroy]

But maybe it doesn't, I'm not sure. And I don't like to be unsure about this sort of thing for code that is on the hot path; where performance is critical.

[Max Godefroy @MaxGodefroy]

For instance you can arbitrarly shuffle vectors in SSE at a very low cost, while NEON only has a few specific shuffle-like functions, with hard-coded shuffle patterns. For esoteric shuffles, you often need to do a few additional steps in NEON.

[Max Godefroy @MaxGodefroy]

I've been doing some background work on my engine for the last few weeks. I added Github workflow support to make sure I push less breaking changes to the codebase. Testing both the build of the engine core, its modules, and the samples; then I run all available unit tests.

[Max Godefroy @MaxGodefroy]

That took a lot of work. Most SSE and NEON features have direct parallel, but not all. To avoid performance loss due to abstraction, you need specific paths for certain functions.

[Rune Skovbo Johansen →🦋🐘 @runevision]

My erosion filter is out! Video, blog post, and shader source.

It emulates erosion without simulation, so it's fast, GPU friendly, and trivial to generate in chunks.

Explainer video:
https://t.co/4PbOmFNfDf

Companion blog post:
https://t.co/JkLbHHdMBC

#ProcGen #vfx #GameDev

[Max Godefroy @MaxGodefroy]

Loaded glyphs look correct, so I should be able to safely assume all is fixed on that front

[Max Godefroy @MaxGodefroy]

As expected, something fishy was going on. Turn out only a part of the glyph MSDF bitmpas where packed into the file. With fixed version, a 64 font size MSDF bitmap compressed version of the file is 3.5MB, much more coherent. (uncompressed version is 8.9MB)

[Max Godefroy @MaxGodefroy]

Well, at least I got it working well. And file sizes are pretty good:
Original .ttf: 277KB
Uncompressed(MSDF (64px font size) + outlines): 832KB (+200%)
Compressed(M+O): 391KB (+41%)
Compressed(M): 120KB (-57%)
Compressed(O): 236KB (-15%)

[Max Godefroy @MaxGodefroy]

As the old programmer saying goes: "If it's too good to be true, it probably is". A bit late now so I'll investigate tomorrow

[Max Godefroy @MaxGodefroy]

Well, at least I got it working well. And file sizes are pretty good:
Original .ttf: 277KB
Uncompressed(MSDF (64px font size) + outlines): 832KB (+200%)
Compressed(M+O): 391KB (+41%)
Compressed(M): 120KB (-57%)
Compressed(O): 236KB (-15%)

[Max Godefroy @MaxGodefroy]

Working on a game engine in a deadline-free environment has a tendency to slowly but surely lead you to a death by a thousand over-engineering urges.

[Max Godefroy @MaxGodefroy]

I'm actually very plesantly surprised by the file size reduction from the MSDF bitmap version. And it makes me nervous that there is some underlying bug...