Performance Behavior of GEMV in Single vs. Multi-Threading

[TheGorge7]

Hi @martin-frbg

I noticed that the gemv function in OpenBLAS performs equally or even better in a single thread compared to multiple threads. Are there specific factors like memory access patterns, workload distribution, or threading overhead affecting this behavior?

[martin-frbg]

I think it is platform-dependent, the current x86_64 code in particular performing worse than, say, arm64. But if I knew why exactly and how to fix it, we wouldn't be having this discussion

[Harishmcw]

Hi @martin-frbg,

I have also observed this behavior on a Windows on ARM64 (WoA) device, where GEMV performs better in single-threaded mode compared to multi-threaded execution.