Use rocprofiler-sdk for RCCL-API tracing (#126)

- Add support for RCCL API tracing through rocprofiler-sdk.
- Refactored the comm_data code to use the SDK RCCL_API callbacks.
- Add a runtime version check for SDK to gate callback enablement, rather than just the compile-time check.
- Fixed: SAMPLING_TIMEOUT was not being handled correctly in add_test.

[ROCm/rocprofiler-systems commit: af77d93f75]
This commit is contained in:
David Galiffi
2025-06-06 11:36:17 -04:00
zatwierdzone przez GitHub
rodzic f718bd907c
commit c7c3c3f97e
31 zmienionych plików z 350 dodań i 1252 usunięć
@@ -298,6 +298,8 @@ comm_data::audit(const gotcha_data& _data, audit::incoming, const void*, int sen
#endif
#if defined(ROCPROFSYS_USE_RCCL)
// Kept for reference, but now gathered throught the SDK callbacks.
// ncclReduce
void
comm_data::audit(const gotcha_data& _data, audit::incoming, const void*, const void*,
@@ -403,6 +405,7 @@ comm_data::audit(const gotcha_data& _data, audit::incoming, const void*, const v
}
// ncclAllGather
// ncclAllToAll
void
comm_data::audit(const gotcha_data& _data, audit::incoming, const void*, const void*,
size_t count, ncclDataType_t datatype, ncclComm_t, hipStream_t)