Alexey Akimov

I have checked db content. All kernel launches and memory copies exists. Could you please provide small reproducer?

Hi, hipMemcpyKind enum is defined here: https://github.com/ROCm-Developer-Tools/HIP/blob/develop/include/hip/driver_types.h#L344

You can try to specify line buffering for stdout explicitly `setvbuf(stdout, NULL, _IOLBF, 4096);`

Could you please check small fix? https://github.com/ROCm-Developer-Tools/rocprofiler/pull/77 Value of feature_count in https://github.com/ROCm-Developer-Tools/rocprofiler/files/7933103/vectoradd_hip.cpp.txt. can't be greater than actual number of features in vector.

You are adding 2 features to profile, but setting feature count to 1. I have fixed it locally, but forgot to report. Could you please check with feature_count set to...

My bad. It's not related to feature_count. You have a typo in metric name. "SQ_INSTS_VALU", not "SQ_INST_VALU".

Can't reproduce with relatively large kernel. Rocprofiler submits additional packets to hsa_queue forcing sched_yield(). There is no additional switches to kernel space.

Could you please provide small synthetic test, build instructions and profiler options used to run?