News for hackers. Stuff that matters.
Re:Towards fearless SIMD, 7 years later by zozbot234 on Sunday March 30, 2025 @12:41PM (#43525476)
> Atomics are very expensive operations.

Atomics are very expensive when contended - which is also the case where locks would introduce blocking and reduced performance. Uncontended atomics are relatively cheap.

Re:Towards fearless SIMD, 7 years later by pclmulqdq on Sunday March 30, 2025 @01:11PM (#43525726)
I suggest you run a microbenchmark. Uncontended atomics are some of the most expensive assembly instructions out there. They also acquire global system locks on certain stages of execution.
Re:Towards fearless SIMD, 7 years later by janwas on Sunday March 30, 2025 @03:35PM (#43526875)
Yep, including draining the store buffer.

We've gotten our ThreadPool barrier+wait to use only acq/rel, but not yet the work stealing. Does anyone have experience with that already?