Pete Cawley (@corsix) / X

Shitpost with AMD on X? Indifference. Shitpost with Tenstorrent on X? They’ll hire you to fix the problem (I’d like to see comprehensive low-level docs describing in detail everything you need to write a driver and a software stack for their HW … so that’s what I’ll be writing)

Pete Cawley

@corsix

Sep 2, 2022

Here it is: github.com/corsix/amx

Pete Cawley

@corsix

Aug 31, 2022

At long last, my reverse-engineered AMX instruction logic matches the hardware behaviour for my full barrage of random test cases. Research complete. Just some writing up to go. Given the size, this'll be a GitHub repo rather than a blog post.

Pete Cawley

@corsix

Mar 12, 2023

Turns out Apple made a few enhancements to AMX between M1 and M2. Most notably, adding bf16 support.

M2 Compatibility · Issue #5 · corsix/amx

Given: 1. crc32 has throughput 1 on port 1 2. pclmulqdq has throughput 1 on port 5 3. pclmulqdq+pxor can emulate crc32 It seems that fastest crc32 code should divide input in half and issue a crc32 _and_ a pclmulqdq every cycle. Code and numbers at corsix.org/content/fast-c…

Pete Cawley

@corsix

Apr 22, 2023

New bloggery: The many ways of converting FP32 to FP16 - corsix.org/content/conver…

Pete Cawley

@corsix

Nov 24, 2024

Replying to @jarredsumner

… and then TCP requires the checksum of the packet contents to appear before said contents. It’s turtles all the way down.

Pete Cawley

@corsix

May 9, 2024

I've wanted something like polyfill-glibc to exist for the last 10 years or so, and now it does: github.com/corsix/polyfil….

GitHub - corsix/polyfill-glibc: Patch Linux executables for compatibility with older glibc

Ready for #FOSDEM: got some hardware to talk about, and some local beer to drink.

Pete Cawley

@corsix

May 6, 2024

I couldn’t find a one-page summary of C23’s stdbit.h anywhere, so I wrote one:

corsix.org

C23 stdbit.h quick reference

Pete Cawley

@corsix

Jun 6, 2023

My cheat-sheet for AArch64 vector instructions:

corsix.org

A whirlwind tour of AArch64 vector instructions (NEON)

Pete Cawley

@corsix

Feb 19, 2024

Replying to @TartanLlama

Short answer: build dependency graph of ELF files (main executable + all referenced SOs, based on DT_NEEDED etc.), topological sort, iterate over files in reverse order (so main executable last). For each visited file, apply all its relocations. See github.com/andikleen/glib…