[sauverma93]

Motivation

This follows up #12212 for gfx1250 and gfx12.

gfx1250 (MI455) reports TA activity as TX_VCA_VCA_BUSY and TX_VCA_TOTAL_WAVEFRONTS. The JSON tool requested TA_TA_BUSY and TA_TOTAL_WAVEFRONTS, so those TA counters never appeared in the gfx1250 samples.

gfx1200 and gfx1201 do use the TA counters. TA_TA_BUSY was already defined. TA_TOTAL_WAVEFRONTS was not, so the counter-value tests compared TA_TA_BUSY with a missing wavefront counter.

The SQC I-cache check is the other failure, and it does not behave the same on every architecture. On MI300, SQC_ICACHE_REQ stays populated and hits and misses stay within that request count, so a miss-rate ratio is well defined. On gfx12 and gfx1250 these counters are nondeterministic per SQ and per bank. Early dispatches of a kernel still see I-cache traffic. Later dispatches of the same kernel can report no requests, hits, or misses, because the instruction cache is already warm. Dividing by SQC_ICACHE_REQ then fails.

Technical Details

The JSON tool requests TX_VCA_VCA_BUSY and TX_VCA_TOTAL_WAVEFRONTS alongside TA_TA_BUSY and TA_TOTAL_WAVEFRONTS. It collects only the counters the agent supports. validate.py reads the agent recorded on each dispatch:

• gfx1250: TX_VCA_VCA_BUSY and TX_VCA_TOTAL_WAVEFRONTS
• other agents, including gfx12: TA_TA_BUSY and TA_TOTAL_WAVEFRONTS

The VCA counters are already defined for gfx1250. TA_TOTAL_WAVEFRONTS is now defined for gfx12, gfx1200, and gfx1201 on the same TA event used by gfx908 and gfx90a.

The SQC check no longer divides by SQC_ICACHE_REQ. It requires SQC_ICACHE_HITS and SQC_ICACHE_MISSES to be no larger than SQC_ICACHE_REQ. A dispatch with no I-cache traffic passes. A hit or miss reported with no requests still fails. A missing counter fails with its name.

The tests this addresses:

• test-spm-dispatch-collection-validate.test_spm_counter_values
• test-spm-dispatch-collection-validate.test_spm_counter_consistency
• test-spm-dispatch-buffer-collection-validate.test_spm_counter_values

Issue Tracking

ISSUE ID: AIPROFSDK-1151

Follow-up to #12212.

Test Plan

gfx1250

• Re-run the SPM integration tests on gfx1250. The consistency checks pass.

• Confirm the gfx1250 samples include TX_VCA_VCA_BUSY and TX_VCA_TOTAL_WAVEFRONTS, and that the busy counter is greater than the wavefront counter.

• Re-run the two counter-value tests after the SQC comparison change:

  • test-spm-dispatch-collection-validate.test_spm_counter_values
  • test-spm-dispatch-buffer-collection-validate.test_spm_counter_values

  The VCA comparison already passed. The remaining failure was a warm I-cache: requests, hits, and misses were all absent, and the old check divided by the request count. That case is now accepted.

gfx12

• TA_TOTAL_WAVEFRONTS is collected, and TA_TA_BUSY is greater than it.
• These pass:
  • test-spm-dispatch-collection-validate.test_spm_counter_values
  • test-spm-dispatch-collection-validate.test_spm_counter_consistency
  • test-spm-dispatch-buffer-collection-validate.test_spm_counter_values
  • test-spm-dispatch-buffer-collection-validate.test_spm_counter_consistency
• SQC_ICACHE_* matches the expected gfx12 pattern. The first launches of a kernel still report requests, with hits and misses no larger than that request count. Repeat launches of the same kernel report no I-cache traffic. Both pass.
• Not changed here: rocprofv3-test-spm-json-validate.test_validate_spm still requires SQ_WAVES to match the PMC total exactly. SPM can come in slightly under that total, so this shader-clock check can still fail. The reference-clock run of the same check passed.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	⚠️ Warning	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like test_<name>.py / test_<name>.cpp (or <name>_test.*). Current: code file(s) changed: projects/rocprofiler-sdk/tests/spm/validate.py, projects/rocprofiler-sdk/tests/tools/json-tool.cpp; no test file found
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[github-actions]

Code Coverage Report

Code Coverage Report

Tests Only

Samples Only

Tests + Samples

[bing-ma]

Looks good to me; the gfx12/gfx1250 counter selection and the warm-I-cache SQC validation change both make sense.

[peymanr]

PR Pundit review

(Testing PR Pundit Code Review agent. Not all comments may be relevant. This is a one-time automated pass: please address what is useful, no need to request a re-review. Please reply to each point, whether it applies or not: we harvest these replies, and both kinds are enormously helpful for tuning the agent. Reach out to @peymanr for suggestions.)

#		Finding	Where
1	🟠	gfx1250 VCA busy/wavefronts assert fails with bare AssertionError: this PR reuses message-less TA heuristic	projects/rocprofiler-sdk/tests/spm/validate.py:128
2	🟠	SPM test passes with zero I-cache traffic on all dispatches: this PR replaced division check with value <= requests	projects/rocprofiler-sdk/tests/spm/validate.py:131-141

1. 🟠 Should fix: gfx1250 VCA busy/wavefronts assert fails with bare AssertionError: this PR reuses message-less TA heuristic

projects/rocprofiler-sdk/tests/spm/validate.py:128

assert float(busy) > float(wavefronts) in _validate_dispatch_counters has no failure message, unlike the other new asserts, so a CI failure does not show the dispatch id, counter names or values. The PR now also applies this strict relation to TX_VCA_VCA_BUSY (VCA busy cycles) vs TX_VCA_TOTAL_WAVEFRONTS (vec32 packets processed) on gfx1250, and nothing shows that relation holds for those counters. The message-less strict assert and its failure when both values are zero were already on the base for the TA counters. Consider adding a message with the names and values.

Who is affected: Developers reading SPM test failures, mainly on gfx1250 where the VCA counters are now compared.

Checked against the code by an independent pass.

2. 🟠 Should fix: SPM test passes with zero I-cache traffic on all dispatches: this PR replaced division check with value <= requests

projects/rocprofiler-sdk/tests/spm/validate.py:131-141

The PR now accepts SQC_ICACHE_REQ == 0; the base crashed with ZeroDivisionError on that input. Nothing requires any dispatch to have seen I-cache traffic, so a fault that zeroes only the SQC counters (while TA/TX_VCA busy stays non-zero) would pass. Bounding SQC_ICACHE_HITS and SQC_ICACHE_MISSES separately is unchanged from the base, which used the equivalent 100*x/REQ <= 100 form. Consider asserting that at least one dispatch reports SQC_ICACHE_REQ > 0.

Who is affected: Only the SPM validate.py test; a regression limited to the SQC I-cache counters would not fail it.

Checked against the code by an independent pass.

_{Shared from a pilot of PR Pundit, an agentic code reviewer that builds and tests the change and reads the project's past reviews. Reviewed at 67178eb4; findings checked and posted by @peymanr; points already raised on this PR are left out.}

[SrirakshaNag]

Why do we remove this altogether, why not check if Req != 0 and compute the ratio in that case