rocm_smi: Initial event count and event table initialization event count upper bound mismatch & handling unsupported events

[djwoun]

Pull Request Description

Fix Event Table Initialization Bug

Tested using utility functions on:

• Dopamine (5.7.3 & 6.3.2)
• Odyssey (6.3.0)
• Histamine0 (5.7.3)

Example Test Commands:

• papi_command_line max_xgmi_internode_bw:device=3:target=0
• papi_component_avail

Issue and Fix

In papi/src/components/rocm_smi/rocs.c, the functions handle_derived_events_count and handle_derived_events & handle_xgmi_events_count and handle_xgmi_events have differing upper bounds for the number of events.

Problem:

for (i = 0; i < ROCS_PCI_BW_VARIANT__LANE_IDX - ROCS_PCI_BW_VARIANT__CURRENT + 1; ++i)

The loop iterates from i = 0 to ROCS_PCI_BW_VARIANT__LANE_IDX - ROCS_PCI_BW_VARIANT__CURRENT + 1, which causes it to run one extra iteration beyond the intended range.

Fix:

Removing the +1 ensures that the loop iterates exactly ROCS_PCI_BW_VARIANT__LANE_IDX - ROCS_PCI_BW_VARIANT__CURRENT times, correcting the off-by-one error.

Similarly, in handle_xgmi_events_count and handle_xgmi_events, there is a mismatch in the upper bounds for the number of events:

• RSMI_EVNT_XGMI_LAST - RSMI_EVNT_XGMI_FIRST and RSMI_EVNT_XGMI_DATA_OUT_LAST - RSMI_EVNT_XGMI_DATA_OUT_FIRST are used in handle_xgmi_events_count.
• However, the for loop in handle_xgmi_events goes beyond the expected step, leading to an inconsistency.

Additional Fix: Handle Unsupported PCIe Bandwidth Status

• Added check for RSMI_STATUS_NOT_SUPPORTED in PCIe bandwidth retrieval.
• Updated comment for future reference.

Author Checklist

• Description
  Why this PR exists. Reference all relevant information, including background, issues, test failures, etc
• Commits
  Commits are self contained and only do one thing
  Commits have a header of the form: module: short description
  Commits have a body (whenever relevant) containing a detailed description of the addressed problem and its solution
• Tests
  The PR needs to pass all the tests

[Treece-Burgess]

Testing results shown below:

Note: PAPI utilities used were papi_component_avail, papi_native_avail, and papi_command_line.

System	Device(s)	ROCm Version	Passing PAPI Utilities
Histamine0	2 * Radeon VII	5.7.3	✅
Dopamine	2 * MI210's	5.7.3	✅
Frontier	2 * MI250X's	5.7.1	✅
Odyssey	4 * MI300A's	6.3.0	✅

[Treece-Burgess]

I think this looks good, what do you think? @adanalis

[adanalis]

Yes, I agree.