raphlinus

Yup, I think slang is the future. Anyone on this thread willing to fund a Rust implementation?

We tried something like this with piet-gpu-hal. One problem is that spirv-cross is lossy, though gaps in target language support are getting better. For example, a device scoped barrier is just droppe…

So I would say skill at GPU assembly is in-demand for the elite tier of GPU performance work. Not necessarily writing much of it (though see [1] for an example, this is the kernel of multisplit as use…

The question of which assembly is best to learn is of course incredibly subjective, but I think the author gives short shrift to ARM32. It is historically important (especially for the Acorn computers…

I read LLVM (or one of its many GPU-flavored variants) reasonably often, mostly to figure out where in the chain a shader miscompilation is happening. But I've never personally had to write it, a…

Thanks so much, Peter, for writing this up. I think it adds a lot to the record about what exactly happened with the Cell. And, as with Larrabee, I have to wonder, what would an alternative universe l…

Thanks for posting this, I'll take a look. It wasn't on my radar, but the idea of doing a DSL specifically for SIMD is something I've been thinking about and also starting to explore my…

There are a lot of odd exclusions on that list. Just spot-checking, I see blog.plover.com, the blog of Mark Jason Dominus, who by the way is looking for a job[1]. Also, dtrace.org is excluded, which h…

Yup. A little more detail on the overheating part in particular is here: https://github.com/AsahiLinux/speakersafetyd …

Getting reasonable speaker support in Asahi Linux was a big deal. Part of the problem is that limiting the power usage to prevent overheating requires sophisticated DSP. Without that, you get very lim…

Regarding the SIMD optimizations, the authors may want to look into faer. I haven't had a great experience with its underlying library pulp, as I'm trying to things that go beyond its linear…

Absolutely. And the fact that we need to evolve both is one of the reasons progress has been difficult.

I consider Xeon Phi to be the shipping version of Larrabee. I've updated the post to mention it.

> It is an instance of Larrabee in the same sense as AMD Zen 4 is an instance of Larrabee. This is an odd claim. Clearly Xeon Phi is the shipping version of Larrabee, while Zen 4 is a completely di…

The problems I'm having are very different than those for raytracing. Sure, it's dynamic, but at a fine granularity, so the problems you run into are divergence, and often also wanting funct…

Possibly compilation and linking. That's very slow for big programs like Chromium. There's really interesting work on GPU compilers (co-dfns and Voetter's work). Optimization problems l…

The argument I have in mind is subtle and nuanced, and I didn't write clearly in that comment (the bit about the smart solo programmer was mostly sarcasm but with a grain of truth). But to try to…

Oh hey, I also have "in defense of undefined behavior" in the queue of blog posts I'd like to write some time, with that exact title. What a coincidence. That said, it's unlikely t…

Well, I have some qualifications in typography, a reasonable familiarity with ML techniques, and am fairly good at math (though I can't claim to have won the Putnam), and I can imagine how to app…

Agree with sibling comments. There's something very slippery and tricky going on with "perceptual area," it's not simple geometry. This is actually an area where I think machine le…

Yup, nothing wrong with clear exposition about simpler algorithms, there's definitely a place for that. I just thought HN readers should have some more context on whether we were looking at a pro…

The second one is Thomas Smith's independent reimplementation of Onesweep. For the official version, see https://github.com/NVIDIA/cccl . The Onesweep implementation is in c…

This is not a fast way to sort on GPU. The fastest known sorting algorithm on CUDA is Onesweep, which uses a lot of sophisticated techniques to take advantage of GPU-style parallelism and work around …

You might enjoy this talk by Erik Lindholm (now retired), who talks about Riva 128 and many of the other early Nvidia cards: https://ubc.ca.panopto.com/Panopto/Pages/Viewer.a…

I just want to say I'm rooting for you, and hope you enjoy the book and learn a lot from it. I had a bad experience with complex analysis as a teen (took a grad class that was a bit over my head)…

I see your point, but one issue is that you can't go from HLSL or GLSL to WGSL for compute work, as the way atomics are typed is incompatible[1]. That's a primary reason we (currently) use W…

In this case, GP has it very much right. Brotli was developed by people, not just faceless Googlers. Jyrki, who led the project, is as passionate an engineer as you're likely to meet. In this c…