zeux - Network

Some file formats like OBJ and PPM are delightfully simple to write, which makes them great as a debugging tool since you can use an...

Some file formats like OBJ and PPM are delightfully simple to write, which makes them great as a debugging tool since you can use an existing viewer, but if you decide to parse them the delight evaporates instantly.

(I guess that's generally true of many file formats? Some formats are painful to read *and* write though!)

ah, nothing like finding a bug that was introduced in 2018 with an optimization commit

ah, nothing like finding a bug that was introduced in 2018 with an optimization commit

I've ran my benchmarks on a few more CPUs out of curiosity, and unlike previous posts I've normalized the throughput by CPU...

I've ran my benchmarks on a few more CPUs out of curiosity, and unlike previous posts I've normalized the throughput by CPU frequency. (this is not *quite* correct to do! for a variety of reasons. but it's interesting.)

On one hand, it's striking how little normalized performance gain there is in Apple Mn. On the other hand, it's striking just how good M1 was and is by this measure.

Single thread, vtx0/1 are SIMD heavy, idx is scalar. M1/v0.22 = version without Apple specific optimizations.

morning audio track: tales of yore by @arashttps://open.spotify.com/episode/4IVlBhlTMznd7G9YPgYhK3

morning audio track: tales of yore by @aras

https://open.spotify.com/episode/4IVlBhlTMznd7G9YPgYhK3

Does anyone want to help make a great .OBJ parser even better? :) .obj files apparently support line directive, and fast_obj doesn't...

Does anyone want to help make a great .OBJ parser even better? :) .obj files apparently support line directive, and fast_obj doesn't handle it atm.

https://github.com/thisistherk/fast_obj/issues/53

In light of AVX10 news I've ran the benchmarks again with recent code (some tweaks to idx) & AVX512 config.I've written this...

In light of AVX10 news I've ran the benchmarks again with recent code (some tweaks to idx) & AVX512 config.

I've written this (128bit) AVX512 code in 2019. I haven't gone ahead and enabled it by default because of muted gains and AVX512 deployment story on clients which continues to be Weird, but hey at least there's hope one day client Intel CPUs will get it :) the Intel chip here is a server Sapphire Rapids where AVX512 has been doing quite well for a while now.

MSVC issue was due to some seriously broken codegen; I was able to work around that thankfully, with perf closer to clang on index decoder...

MSVC issue was due to some seriously broken codegen; I was able to work around that thankfully, with perf closer to clang on index decoder now.

Clang has reasonable codegen and turns out gcc is an outlier here, using a surprising scalar to vector promotion that surprisingly saves time by increasing IPC. Not sure I want to replicate that for now…

The index codec on Zen4 runs into some perf issues with clang though - don’t think Zen4 should necessarily be slower here. That said, MSVC...

The index codec on Zen4 runs into some perf issues with clang though - don’t think Zen4 should necessarily be slower here. That said, MSVC is unfortunately even worse… we’ll see if these can be fixed by tweaking the code further.

Fun to see just how good Apple CPUs are in single core performance, especially if you tune the code.

Fun to see just how good Apple CPUs are in single core performance, especially if you tune the code.

At gdc wed-thu and maybe fri, dm if you’d like to grab coffee

At gdc wed-thu and maybe fri, dm if you’d like to grab coffee

hot take: vim with arrows is actually great and hjkl is unnecessary cognitive overhead

hot take: vim with arrows is actually great and hjkl is unnecessary cognitive overhead

achievement unlocked: used “radii” in the APII wish I had a usecase for “frusta”

achievement unlocked: used “radii” in the API

I wish I had a usecase for “frusta”

With every release meshoptimizer documentation keeps improving; GitHub has an up to date version, but you can also always go to this website...

With every release meshoptimizer documentation keeps improving; GitHub has an up to date version, but you can also always go to this website for the latest release:

https://meshoptimizer.org/

Boosts / GitHub stars appreciated!https://github.com/zeux/meshoptimizer/releases/tag/v0.23

Boosts / GitHub stars appreciated!

https://github.com/zeux/meshoptimizer/releases/tag/v0.23

meshoptimizer 0.23 is out! Featuring vertex codec v1 (increased compression ratio, faster decompression), even faster decompression on Apple...

meshoptimizer 0.23 is out! Featuring vertex codec v1 (increased compression ratio, faster decompression), even faster decompression on Apple Silicon, improvements to meshlet builder with fewer disconnected clusters and more options for clustered raytracing and hierarchical clusterization, new algorithms for hierarchical clusterization, simplification quality improvements and more.

Thanks to Valve for sponsoring most of the work on the core library in this release!

TypeScript compiler being rewritten in Go (from TypeScript): https://devblogs.microsoft.com/typescript/typescript-native-port/Curious choice...

TypeScript compiler being rewritten in Go (from TypeScript): https://devblogs.microsoft.com/typescript/typescript-native-port/

Curious choice of the target language, and interesting development regardless.

P.S. I checked and this affects luau as well. If you use clang-19 or clang-20 in production, I recommend using `-mllvm...

P.S. I checked and this affects luau as well. If you use clang-19 or clang-20 in production, I recommend using `-mllvm -tail-dup-pred-size=256 -mllvm -tail-dup-succ-size=256` as a workaround in the meantime. I've let the Luau folks know, hopefully they keep track of the pending LLVM fix and/or adjust the CMake settings accordingly if this never gets fixed...

Great article. I read the tail call article when it appeared and was also very confused by the results - no relation to Python but I tuned...

Great article. I read the tail call article when it appeared and was also very confused by the results - no relation to Python but I tuned luau interpreter to bits a few years ago and there just wasn’t a 10% opportunity there based on very careful inspection of assembly; I filed this away as “maybe CPython interpreter loop is weird” but the reality is more interesting.
https://mendeddrum.org/@fanf/114135155417228009

"... the legal officers in companies sometimes create the worst possible setup for companies because the legal officers don't...

"... the legal officers in companies sometimes create the worst possible setup for companies because the legal officers don't really try to balance the legal risk with the company potential.
They are mostly going to err on the safe side in order to be right later. If you are super conservative, you never, never do anything wrong from your point of view, However, you're paralyzing the company."

(from https://changelog.com/podcast/631)

Are there special considerations wrt FFI for languages like C#, Rust, Zig, Odin, Nim, Go for C functions that take array of pointers - is...

Are there special considerations wrt FFI for languages like C#, Rust, Zig, Odin, Nim, Go for C functions that take array of pointers - is that substantially worse vs an array of values? (eg imagine a function that takes an arbitrary number of float arrays, vs a function that takes a single float array that has all streams combined and interleaved)