Guo Yejun

I wrote some code to verify it, but looks that some values are not as expected. ``` #include #include void print_event_profiling(const sycl::event& event, const char* prefix = "NULL") { auto...

anyone can help? thanks

thanks @bader. I tried the same code twice on another machine for level zero backend and opencl backend, see below. Looks that the opencl backend has another explanation of the...

what is the starting time (the origin of the time line) for opencl backend, is it the time when the machine is power on? thanks.

thanks @bader , OpenCL spec says: A 64-bit value that describes the current device time counter in nanoseconds when the command identified by event ..., and so time 0 means...

> As I mentioned in my first comment - it's known issue of level zero backend. it is a known issue for submit time of level zero backend. My concern...

yes, and so I added `event.wait()` for explicit sync in the code, and so `sleep` consumes the time at both host side and device side, and so the `device time...

i verified that `event.get_profiling_info` returns value in nanoseconds on my system for level zero backend. I would say that there's a bug of level zero backend when the case is...

> * show as much of the full path as possible on 1 line("/.../dirA/subdir/foo.py"). This will help providing more context. We can start from root or from an arbitrary dir...

to make the logic clear. The peek stats should be cleared after cache is empty.