[amd-vkale]

Motivation

--kernel-iteration-range selects application-observed kernel launches. Kernel replay creates
multiple executions of one logical dispatch so profiling can collect a different group on each
pass. Before this fix, counter_dispatch_callback consulted the stateful iteration filter on every
pass. The filter therefore mistook replay passes for new application iterations.

The validation report reproduced this with independent HIP and Kokkos workloads. Both the
application and rocprofv3 exited successfully, with no warning, even though the resulting counter
records were incomplete and appeared valid.

For example, with iteration range 2-3 and two replay passes, the expected output is two selected
logical dispatches, each containing passes [0, 1]: four counter records in total. Instead, the
filter counted each pass as an iteration. Pass 1 from one logical dispatch and pass 0 from the next
were selected separately, leaving both dispatches incomplete. Numeric dispatch IDs may vary; the
missing pass in each dispatch is the failure.

Intuition: one dispatch, several physical executions

Profiling hardware cannot always measure every requested event simultaneously, so tools divide the
request into groups and execute the same dispatch once per group. That creates two different units:

• A logical dispatch is the launch the application requested. Kernel names, launch filters, and
  iteration ranges describe this unit.
• A replay pass is one physical execution used to collect one profiling group. Pass selection
  and per-pass configuration describe this unit.

The bug came from applying a logical-dispatch filter at the replay-pass boundary.

flowchart TB
  subgraph beforeFix [Before: filter evaluated per replay pass]
    launch1Before["Logical launch 1"]
    launch1Pass0["Pass 0 calls filter; iteration 1; rejected"]
    launch1Pass1["Pass 1 calls filter; iteration 2; selected"]
    incomplete1["Dispatch D1 contains pass 1 only"]
    launch2Before["Logical launch 2"]
    launch2Pass0["Pass 0 calls filter; iteration 3; selected"]
    launch2Pass1["Pass 1 calls filter; iteration 4; rejected"]
    incomplete2["Dispatch D2 contains pass 0 only"]
    launch1Before --> launch1Pass0 --> launch1Pass1 --> incomplete1
    launch2Before --> launch2Pass0 --> launch2Pass1 --> incomplete2
  end

  subgraph afterFix [After: filter evaluated per logical dispatch]
    launch1After["Logical launch 1; filter iteration 1; rejected"]
    launch2After["Logical launch 2; filter iteration 2; selected"]
    launch2Passes["Passes 0 and 1 reuse selected"]
    complete2["Dispatch D2 contains passes 0 and 1"]
    launch3After["Logical launch 3; filter iteration 3; selected"]
    launch3Passes["Passes 0 and 1 reuse selected"]
    complete3["Dispatch D3 contains passes 0 and 1"]
    launch1After --> launch2After --> launch2Passes --> complete2
    launch2After --> launch3After --> launch3Passes --> complete3
  end

Loading

Expected record shape after the fix:

iteration range 2-3, two profiling groups

logical launch 1  -> rejected
logical launch 2  -> dispatch D2: [pass 0, pass 1]
logical launch 3  -> dispatch D3: [pass 0, pass 1]

total: 2 selected dispatches x 2 passes = 4 counter records

Code map:

• The stateful iteration filter advances the per-kernel count:
  tool.cpp L391-L409.
• The replay pass and cached dispatch decision are thread-local:
  tool.cpp L1641-L1657.
• Pass 0 consults the filter and later passes reuse the result:
  tool.cpp L2023-L2036.
• CONFIG entry clears the previous dispatch decision; PASS entry/exit publishes and clears the
  authoritative pass index:
  tool.cpp L2561-L2593.

Technical Details

The iteration-range decision now belongs to the dispatch:

1. CONFIG entry resets tl_replay_dispatch_targeted, preventing the previous dispatch's decision
   from leaking into the next replay loop.
2. Pass 0 calls is_targeted_kernel(), advancing the per-kernel iteration count exactly once, and
   caches the result.
3. Later passes reuse the cached result and do not advance the iteration count.
4. Non-replayed dispatches continue to call the filter normally.

Replay PASS callbacks and the counter dispatch callback execute synchronously and in order on the
enqueuing thread, so the decision can be stored thread-locally without a process-wide map or lock.

Issue Tracking

JIRA ID: AIPROFSDK-1130

Test Plan

Adds an integration test combining kernel replay with --kernel-iteration-range. It selects a
subset of logical launches and verifies that every selected dispatch still contains every expected
replay pass. This combination was previously uncovered: iteration-range tests did not enable
replay, while replay tests did not specify an iteration range.

The regression test mirrors AIPROFSDK-1130 with four launches, range 2-3, and two counter groups.
It asserts exactly two selected logical dispatches, passes [0,1] on each, and a distinct counter
group on each pass. Numeric dispatch IDs are deliberately not pinned.

Regression coverage:
CMakeLists.txt L148-L202.

Test Result

CI restarted for the latest test-only update. As of September 8, 2026: 8 passed, 29 pending,
0 failed, and 8 skipped.

View the live CI status.

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/source/lib/rocprofiler-sdk-tool/tool.cpp, projects/rocprofiler-sdk/tests/rocprofv3/kernel-replay/conftest.py, projects/rocprofiler-sdk/tests/rocprofv3/kernel-replay/validate.py; 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?

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ PR Description

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.

[Copilot]

🟢 Approval recommended

The dispatch-level caching approach aligns with the stated replay/pass execution model in this tool and the added integration test directly guards against the reported regression.

Pull request overview

This PR fixes an interaction bug between kernel replay and --kernel-iteration-range in the rocprofiler-sdk tool: the iteration-range filter was being consulted once per replay pass (instead of once per user-visible dispatch), causing a single kernel launch to consume multiple “iterations” and leading to later replay passes being silently dropped. It also adds an integration regression test to ensure filtered dispatches still include all replay passes.

Changes:

• Cache the kernel-iteration-range targeting decision per dispatch during kernel replay so only replay pass 0 advances the per-kernel iteration counter.
• Reset the cached decision at the start of each replay loop to prevent cross-dispatch leakage.
• Add an integration test covering --replay-mode kernel combined with --kernel-iteration-range, validating pass completeness.

File summaries

File	Description
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/tool.cpp	Ensures kernel-iteration-range filtering is applied once per dispatch (not per replay pass) by caching the decision in a thread-local during replay.
projects/rocprofiler-sdk/tests/rocprofv3/kernel-replay/CMakeLists.txt	Adds an integration regression test combining kernel replay with --kernel-iteration-range and validating that surviving dispatches still have all replay passes.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Code Coverage Report

Code Coverage Report

Tests Only

Samples Only

Tests + Samples

[Copilot]

🟢 Approval recommended

The implementation matches replay callback ordering and the integration test covers the reported pass-loss regression.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[bgopesh]

Claude flags the position of the fix. worth looking

[amd-vkale]

Claude flags the position of the fix. worth looking

Thanks for that. Working on fixing and then testing again.

[github-actions]

This pull request has been inactive for 25 days and will be marked as stale.

If you would like to keep this PR open, please:

• Add new commits
• Add a comment explaining why it should remain open

This PR will be automatically closed in 5 days if no further activity occurs.

[amd-vkale]

@bgopesh can you please re-review?

[systems-assistant]

TheRock Submodule Bump Activity

Newest first

• 2026-10-09 — Included in TheRock via ROCm/TheRock#8885.