As far as I understand, that should be practically zero-cost on x86, and quite cheap also on ARM - and it's also only paid during acquisition and destruction. I haven't yet caught refcounting in profiling as being the cause of unacceptable performance.
It’s practically zero cost when uncontested. When contested, the instructions with that `lock` prefix are concurrently accessing a single cache line from multiple cores at the same time.
These cache coherency protocols are relatively slow. On some processors, reading a cache line recently modified by another core can cost 300-500 cycles, way more expensive than a cache miss and roundtrip to off-chip memory.
Your requirements are probably similar to this C# queues class: https://github.com/Const-me/Vrmac/blob/master/VrmacVideo/Aud... That library decodes and plays realtime video + audio, both low latency and consistent performance over prolonged use were rather important. BTW that code runs on Raspberry Pi4, CPU performance is a fraction of what you’d expect on modern desktops or servers.