Paul Khuong
I guess technically it's now 2026, but I'm not changing the date on this one https://pvk.ca/Blog/2025/12/30/six-versions-accessing/
Six versions accessing: wait-free protected versions with bounded cardinality - Paul Khuong: some Lisp

Paul Khuong's personal blog. Some Lisp, some optimisation, mathematical or computer.

Jan 2 at 3:33pmWeb
Show threadPaul KhuongJan 2
@tobinbaker i feel like you'll be interested in ^
Show threadWilliam D. JonesJan 2

@pkhuong Looks neat, not sure I can really understand it well tho :P.

I'll just stick to Bakery Algorithm :D.

Show threadAlexander MonakovJan 3

@pkhuong I can't follow all of it, but I wonder if this bit:

> It can also be helpful to use cache line-wide stores (e.g., AVX-512 or FSRM stores) to avoid reads for ownership.

should be clarified? First, wouldn't a full line write need a "no data RFO" in any case? Also, do you know if any CPUs implement this optimization for avx512? As I recall, Skylakes didn't use rfo_nodata for aligned avx512 writes, and I haven't heard that changed since.

Show threadPaul KhuongJan 3

@amonakov Yeah, you're right, I'll clarify to avoiding a full RFO. The chip still needs to send a read request for ownership, but no data.

Re the AVX-512 path, I'm pretty sure it works on SPR+ (but can be hard to confirm with all the prefetchers).