map kernel and copy events to logical stream ID instead of HSA queue ID (#1424)

[hannaxu]

Summary:

Previously, resourceId() propagated the HSA queue ID for kernel events, rather than the logical stream ID. The AMD default is GPU_MAX_HW_QUEUES=4, which follows why stream interleaving was noticeable only when number of workers (total) >= 5 which is common on models with remote (additional remote workers). Therefore, there was stream interleaving as we were mapping > 5 streams to 4 HSA queues. The separate backfillAsyncCopyStreams for H2D/D2D was also propagating HW queue IDs.

Switch to int64_t for stream ids to avoid truncation; HW queue IDs fit fine in int32. Return streamId with queueId fallback inresourceId(). Recover the real HIP stream from the correlated runtime row, resolving both kernels and copies to stream ID via backfillAsyncStreams. Remove separate backfillAsyncCopyStreams. Async rows carry only the HW queue, so we map correlation --> HIP stream from the runtime rows.

Process

1. Infer stream from queue when a queue maps unambiguously to one stream
2. Remap raw 64-bit HIP stream pointers to per-device indices since Perfetto misrenders large id values onto the same track
3. Kernels and copies share one per-device index space, so ops on the same HIP stream land on the same track. resourceId() returns the resolved streamId with a queueId <--> unambiguous stream map fallback

Add the HSA queue ID as a field in the trace, but not as the track provenance in next diff D107568026

Differential Revision: D107306915

[meta-codesync]

@hannaxu has exported this pull request. If you are a Meta employee, you can view the originating Diff in D107306915.

[sanrise]

LGTM, thanks for adding this -

[meta-codesync]

This pull request has been merged in 25856f7.