Thanks for the report! For performance issues normally we want to try to figure out whether the regression is coming from the compiler or the runtime or somewhere else. So you can check a couple of things: --Check if the number of work groups/work items the runtime uses has changed for that kernel. You can do that with `iprof -k ./a.out...` and look at the number of work groups/work items for the kernel. --Try using the current SDK but with the runtime from the last SDK and vice versa (for example, module load oneapi; module switch -f intel_compute_runtime/release/agama-prerelease-316), and maybe we can pinpoint if it's from the compiler or the runtime. Other people on the list might have ideas, but this is where I'd start. If you have time, we can have a call and look at it together. Thanks, Colleen ________________________________ From: Ecp-aurora-sdk-discuss <[email protected]> on behalf of Moore, Stan via Ecp-aurora-sdk-discuss <[email protected]> Sent: Wednesday, March 30, 2022 9:56 AM To: [email protected] <[email protected]> Subject: [Ecp-aurora-sdk-discuss] Large slowdown with latest oneapi compiler Hi everyone, I'm seeing a ~6x overall slowdown for the EXAALT KPP benchmark (LAMMPS SNAP) when switching from oneapi/eng-compiler/2022.01.30.002 to oneapi/eng-compiler/2022.01.30.003. This is using the Kokkos SYCL backend, keeping both Kokkos and LAMMPS versions constant. I profiled the kernels and the most expensive kernel (ComputeYi) became ~10x slower due to the compiler update. Most of the other kernels have similar performance between the two compiler versions. I have set SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file" but it doesn't seem to have any effect. Any ideas? Thank you, Stan