[SYCL][CUDA] Optimal way of invoking device functions from SYCL kernel

[TejaX-Alaghari]

I'm trying to call a function from inside a SYCL kernel, defined in a different namespace and class. This code is compiled with clang++ built for CUDA backend.

A drop in performance has been observed after invoking that function from the SYCL kernel, when compared to native CUDA implementation.
What could be the reasons causing the drop in performance and what optimizations could be employed in this scenario to make this operation performant?

[bader]

inlining?

[TejaX-Alaghari]

Thanks for the quick response. "inlining" the function didn't reduce the time and only affected the runtime slightly.
Are there any other tips that I can try?

Also, the function that I'm trying to call performs operations on sycl::vec objects.
Are there any optimizations that are related to vector operations?

[bader]

Could you share the code to reproduce your problem, please?
I don't think speculating about the root cause using provided bites of information can be efficient.

[zjin-lcf]

I read #6583 just now.

[npmiller]

Could you try inlining again? #7046 may have helped with that

[npmiller]

Adding the stale label, there hasn't been updates in a while and the issue might be partially solve, feel free to comment if the issue persist, otherwise the issue will be automatically closed.