Return ResourceInfo with sycl_queue_id by xpu profiler #1210
Conversation
|
CI results are there: |
| device_id, | ||
| sycl_queue_id, | ||
| sycl_queue_id, | ||
| fmt::format("sycl queue {}", sycl_queue_id)); |
There was a problem hiding this comment.
The "sycl queue" will be shown in profiling chart?
Similar to CUDA, SYCL should not have fixed queue id. This sycl_queue_id should be queue object address or handler provided by PTI?
Let's call it Stream, instead, to avoid confusing user. Below is the reference from AIUPTI.
There was a problem hiding this comment.
It is sycl_queue_id provided by PTI.
I'm not sure if 'Stream' is meaningful on XPU architecture and would not cause confusion.
There was a problem hiding this comment.
That should be OK. For PyTorch profiler user, the queue ID is meaningless, as it's runtime value.
In our design, there's 1 on 1 mapping between PyTorch Stream and SYCL queue. And, in most case, we are not going to expose SYCL queue to user. The user only wants to know there're different PyTorch Streams.
2 participants
It allows Kineto to label xpu related plots in perfetto with 'sycl queue' label