Oh man this is so much better.

Thanks! Want to grab a beer at some point after the holidays if you're still in Paris btw?

Thank you very much 😁

Oh I didn't know it had Tint and Naga now, nice! Thanks, TIL!

I'm not sure if there is something, I use PIX on windows to get the HLSL.
It's ok if you don't have an easy way to get the code, I won't be able to test on Mac anyway.

😱😱😱

I know the explanation is short but it would take quite a few tweets to fully go through it 😅

V0 is worse than original for testing, V1 is the original code, V2 is a failed fix, V3 fixes the execution sync issue but lacks a read after write synchronization. It seems the writes make it to memory fast enough though!

The code is here if you want to take a look github.com/frguthmann/w...

I was looking at some code synchronizing multiple threadgroups within the same dispatch and I had a hunch something was off. There are actually multiple causes for the original race(s) so I wrote increasingly less racy versions of it, to validate my hunch and finally fix the original code.

Thanks! What did you test this on?

Oh wow, interesting! I only inspected the generated HLSL and AMD ISA but I would have expected this to work on Apple devices as well, thanks for testing!
Would it be easy for you to send me the generated MSL code?

Here's what I get:
V0: ❌, V1: ❌, V2: ❌, V3: ✅, V4: ✅

Yes, definitely! When rehearsing I always tried to phrase occupancy as *one* way to hide latency, but not sure how I did when presenting that day.

A fun night 😁

Sorry for the terrible photo quality, knees were weak and arms heavy :p.
Also, shout-out to the first row who noticed I was taking a picture and decided to have fun with it 😁.

I would love a write up explaining the design choices, pitfalls avoided and so on afterwards!

Nice, see you there!