nvPtxCompiler still serializing?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello, I ran into heavy serialization while using the nvPtxCompiler api. I noticed that in the docs for cuda 11.2 for the nvPtxCompiler api it says

”All PTX Compiler API functions are thread safe and may be invoked by multiple threads concurrently. However, due to implementation limitations, compilation is serialized if multiple threads attempt to compile PTX code concurrently; compilation is done for a single thread at a time.”

This was changed and now in 12.9.1 (the version I’m running) it says:
”All PTX Compiler API functions are thread safe and may be invoked by multiple threads concurrently.”

Is it a known limitation that nvPtxCompiler is still serializing all the calls when made in the same process? Does nvJitLink with spilt-compile seem to do okay as a replacement?

Thanks,
Charlie Durham

2

I apologize, I did a clean reimplementation of just this part and it seems to be scaling really well. Scales well for physical cores but not so well with hyper-threading cores on my AMD system.