00:15 <bloom> https://rosenzweig.io/blog/asahi-gpu-part-3.html

00:15 <bloom> Have a field day y'all :)

00:16 <bloom> (Disclaimer: all the code for that is published but it's split across 3 repos and it'd probably be hard to reproduce as-is. This will be rectified over time..)

00:16 <bloom> https://gist.github.com/alyssarosenzweig/4dc380184a4959fe609bfa26bf96dd6c are the shaders, from which you'll learn, knowing how to write a shader compiler does *not* mean you know how to write shaders that don't suck :p

00:17 <bloom> The frontend is a fork of the usual demo app https://github.com/AsahiLinux/gpu/tree/wip

00:17 <bloom> but the code has a bunch of hacks in it .. again, all cleaned up in time

00:18 <bloom> JTL: ^ there you go :p

00:22 <JTL> cool, will check it out

09:29 <radex> bloom: just curious, how do we know that AGX has "some form of out of order execution"? I thought OoO isn't really a thing in GPUs, and instead, heavy multithreading is used to hide latency, but I'm probably uninformed

12:21 <bloom> radex: Heuristic argument.. we know GPUs need some form of perthread SIMD or VLIW or multi-issue or straight up superscalar for high perf

12:21 <bloom> and we know it's not the first two.. and we know AGX perf is unparalleled..

12:22 <bloom> Also, the linked LLVM presentation from Apple has a brief discussion on their scheduling.

13:11 <radex> bloom: huh, I thought "superscalar" just means "can issue multiple instructions in one cycle" and is orthogonal to OoO

13:18 <Yuzu> i was wondering about the same thing

13:19 <Yuzu> it seems like the structures needed to support OoO would be expensive when architectural state is as large as it tends to be in GPUs

13:27 <Yuzu> and if anything, the stuff in the linked presentation sounds in-order (the optimizations in slide 57, for instance, would be of fairly limited use in a true OoO machine, since independent ops would happily execute while waiting for blocked ops)

18:05 <bloom> yeah, dougall pointed out similar issues.. working theory is that it's dual-issue ALU, with memory loads as messagepassing (quick)