Repository navigation
Update formatting.yml for Rocprofiler-sdk - #7
Merged
Merged
Conversation
Changed runners to ubuntu-latest instead of AMD-ROCm-Internal-dev1
amd-hsivasun
changed the title
systems-assistant Bot
pushed a commit
that referenced
this pull request
Aug 5, 2025
Set execute permission for HSA allocated memory We need execute permission for HSA memory (req for IB buffers). Enforcement is upcoming which will break counter collection (see ticket). Change-Id: Id939bbdd8686915407bc60d0dfcd5110b0e91e90 Co-authored-by: Benjamin Welton <bewelton@amd.com> Co-authored-by: Bhardwaj, Gopesh <Gopesh.Bhardwaj@amd.com>
systems-assistant Bot
pushed a commit
that referenced
this pull request
Aug 5, 2025
* General logging updates
* ROCm 6.1.x fatal defect: unconditional abort if tool found but rocprofiler-sdk missing
- Adding rocprofiler-sdk support breaks backwards compatibility in ROCm 6.1.x because `rocprofiler_configure` symbol triggers looking for rocprofiler-sdk library (which was not released until ROCm 6.2)
- Bump version to 0.5.0
* revert to using fatal error
* Include header updates
* Fix CodeQL suggestions
* CMake and CI updates
- minimum cmake version is 3.22.0
- added requirements.txt
- improved rocprofiler_register_{formatting,linting}.cmake
* Disable deprecated declarations warnings
* Use "overwrite" instead of "override"
- override is a keyword in C++
* Disable REQUIRED for formatting and linting when BUILD_DEVELOPER=ON
---------
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
systems-assistant Bot
pushed a commit
that referenced
this pull request
Aug 5, 2025
* Obey CMAKE_BUILD_TYPE Obey CMAKE_BUILD_TYPE when it is passed on the command line. Previous behavior remains: Debug by default ROCRTST_BLD_TYPE can set build type * External CI: change trigger from amd-master to amd-mainline * Added gfx94x and gfx11xx IDs to enumerator (#5) * Use idomatic approach to extending CMAKE_MODULE_PATH. (#7) There are many reasons why there may already be a CMAKE_MODULE_PATH defined. The idiomatic way to extend it in a project is via list(APPEND). ROCm/rocminfo#97 submitted by @stellaraccident Co-authored-by: Stella Laurenzo <Stella.Laurenzo@amd.com> * Added release trigger for further releases * Fix for typo in ISA_TO_ID table (#6) Change-Id: Id0720988fb01079a3ca6f3287dc743d93db4fd0f * Remove WSL checks for displaying UUID and BDFID Signed-off-by: lyndonli <Lyndon.Li@amd.com> * Added KWS check for amd-mainline (#9) --------- Signed-off-by: lyndonli <Lyndon.Li@amd.com> Co-authored-by: benrichard-amd <ben.richard@amd.com> Co-authored-by: Su, Daniel <Daniel.Su@amd.com> Co-authored-by: Stella Laurenzo <Stella.Laurenzo@amd.com> Co-authored-by: Mallya, Ameya Keshava <AmeyaKeshava.Mallya@amd.com> Co-authored-by: Yat Sin, David <David.YatSin@amd.com> Co-authored-by: lyndonli <Lyndon.Li@amd.com>
jayhawk-commits
pushed a commit
that referenced
this pull request
Aug 5, 2025
* General logging updates
* ROCm 6.1.x fatal defect: unconditional abort if tool found but rocprofiler-sdk missing
- Adding rocprofiler-sdk support breaks backwards compatibility in ROCm 6.1.x because `rocprofiler_configure` symbol triggers looking for rocprofiler-sdk library (which was not released until ROCm 6.2)
- Bump version to 0.5.0
* revert to using fatal error
* Include header updates
* Fix CodeQL suggestions
* CMake and CI updates
- minimum cmake version is 3.22.0
- added requirements.txt
- improved rocprofiler_register_{formatting,linting}.cmake
* Disable deprecated declarations warnings
* Use "overwrite" instead of "override"
- override is a keyword in C++
* Disable REQUIRED for formatting and linting when BUILD_DEVELOPER=ON
---------
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
[ROCm/rocprofiler-register commit: cb1b430]
jayhawk-commits
pushed a commit
that referenced
this pull request
Aug 5, 2025
Set execute permission for HSA allocated memory We need execute permission for HSA memory (req for IB buffers). Enforcement is upcoming which will break counter collection (see ticket). Change-Id: Id939bbdd8686915407bc60d0dfcd5110b0e91e90 Co-authored-by: Benjamin Welton <bewelton@amd.com> Co-authored-by: Bhardwaj, Gopesh <Gopesh.Bhardwaj@amd.com> [ROCm/rocprofiler commit: 6cc5501]
jayhawk-commits
pushed a commit
that referenced
this pull request
Aug 11, 2025
There are many reasons why there may already be a CMAKE_MODULE_PATH defined. The idiomatic way to extend it in a project is via list(APPEND). ROCm/rocminfo#97 submitted by @stellaraccident Co-authored-by: Stella Laurenzo <Stella.Laurenzo@amd.com> [ROCm/rocminfo commit: 8503ec3]
jayhawk-commits
pushed a commit
that referenced
this pull request
Aug 11, 2025
* Obey CMAKE_BUILD_TYPE Obey CMAKE_BUILD_TYPE when it is passed on the command line. Previous behavior remains: Debug by default ROCRTST_BLD_TYPE can set build type * External CI: change trigger from amd-master to amd-mainline * Added gfx94x and gfx11xx IDs to enumerator (#5) * Use idomatic approach to extending CMAKE_MODULE_PATH. (#7) There are many reasons why there may already be a CMAKE_MODULE_PATH defined. The idiomatic way to extend it in a project is via list(APPEND). ROCm/rocminfo#97 submitted by @stellaraccident Co-authored-by: Stella Laurenzo <Stella.Laurenzo@amd.com> * Added release trigger for further releases * Fix for typo in ISA_TO_ID table (#6) Change-Id: Id0720988fb01079a3ca6f3287dc743d93db4fd0f * Remove WSL checks for displaying UUID and BDFID Signed-off-by: lyndonli <Lyndon.Li@amd.com> * Added KWS check for amd-mainline (#9) --------- Signed-off-by: lyndonli <Lyndon.Li@amd.com> Co-authored-by: benrichard-amd <ben.richard@amd.com> Co-authored-by: Su, Daniel <Daniel.Su@amd.com> Co-authored-by: Stella Laurenzo <Stella.Laurenzo@amd.com> Co-authored-by: Mallya, Ameya Keshava <AmeyaKeshava.Mallya@amd.com> Co-authored-by: Yat Sin, David <David.YatSin@amd.com> Co-authored-by: lyndonli <Lyndon.Li@amd.com> [ROCm/rocminfo commit: aec5d6f]
jayhawk-commits
pushed a commit
that referenced
this pull request
Aug 18, 2025
[ROCm/hipother commit: 3eb783e]
stellaraccident
added a commit
that referenced
this pull request
Feb 26, 2026
## Summary - Wire `ElfVerifier` into `kpack_offload_binary()` so corrupt output is caught at build time - Runs after all transform phases, before writing output file - Catches PHDR alignment violations, NOBITS collisions, and other structural issues - Verified: all 53 existing ELF tests pass, real binary transforms still succeed ## Motivation The kpack ELF transform has produced corrupt binaries in CI multiple times. The verifier already exists and is used in unit tests, but was never wired into the production code path. This means corruption is only discovered at runtime (`hipErrorInvalidImage`) long after the binary was generated. Confirmed the verifier catches both errors found in the latest CI artifacts: ``` Verification FAILED Errors (2): - PT_LOAD segments 1 and 2 share file offset 0x0 (NOBITS collision) - PT_LOAD segment 12 has misaligned offset/vaddr ``` ## Test plan - [x] All 53 existing ELF tests pass - [x] Pre-commit checks pass (black, etc.) - [x] Verified local transform of real librocrand.so still passes verification - [x] Verified corrupt CI artifact is correctly rejected by verifier 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
1 task
jeffqjiangNew
added a commit
that referenced
this pull request
Mar 13, 2026
## Motivation This PR adds entry and exit logs for core functions. ## Technical Details - Info level (3) logs are added at core function entry and exit points. - Monotonic clock stamp, process id, thread id, and file name are added to the logs. - The logs now have similar format to HIP logs for better viewing when both are present. - Direct error or warning logs to std::cerr. Other logs still go to std::cout. ## JIRA ID <!-- If applicable, mention the JIRA ID resolved by this PR (Example: Resolves SWDEV-12345). --> <!-- Do not post any JIRA links here. --> ## Test Plan Set env variable ROCDEC_LOG_LEVEL to control log output. All CTests and conformance tests must pass. Example summary of the test results: - CTests: Start 1: video_decodeRaw-HEVC 1/15 Test #1: video_decodeRaw-HEVC ............. Passed 2.05 sec Start 2: video_decodeRaw-AVC 2/15 Test #2: video_decodeRaw-AVC .............. Passed 1.89 sec Start 3: video_decodeRaw-AV1 3/15 Test #3: video_decodeRaw-AV1 .............. Passed 2.31 sec Start 4: video_decodeRaw-VP9 4/15 Test #4: video_decodeRaw-VP9 .............. Passed 2.04 sec Start 5: rocdec_Decode-HEVC 5/15 Test #5: rocdec_Decode-HEVC ............... Passed 0.92 sec Start 6: rocDecode_Negative_API_Tests 6/15 Test #6: rocDecode_Negative_API_Tests ..... Passed 0.78 sec Start 7: video_decode-HEVC 7/15 Test #7: video_decode-HEVC ................ Passed 2.74 sec Start 8: video_decode-AVC 8/15 Test #8: video_decode-AVC ................. Passed 2.64 sec Start 9: video_decode-AV1 9/15 Test #9: video_decode-AV1 ................. Passed 2.74 sec Start 10: video_decode-VP9 10/15 Test #10: video_decode-VP9 ................. Passed 2.78 sec Start 11: video_decodePerf-HEVC 11/15 Test #11: video_decodePerf-HEVC ............ Passed 2.92 sec Start 12: video_decodeBatch 12/15 Test #12: video_decodeBatch ................ Passed 4.34 sec Start 13: video_decodeRGB-HEVC 13/15 Test #13: video_decodeRGB-HEVC ............. Passed 33.51 sec Start 14: video_decodeMem-HEVC 14/15 Test #14: video_decodeMem-HEVC ............. Passed 2.09 sec Start 15: video_decodeRGB-Resize 15/15 Test #15: video_decodeRGB-Resize ........... Passed 33.94 sec 100% tests passed, 0 tests failed out of 15 - AVC conformance test: Conformance test completed on the 127 streams: - The number of passing streams is 127 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 - HEVC conformance test: Conformance test completed on the 135 streams: - The number of passing streams is 135 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 - AV1 conformance test: Conformance test completed on the 162 streams: - The number of passing streams is 162 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 - VP9 conformance test: Conformance test completed on the 108 streams: - The number of passing streams is 108 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 ## Test Result Function entry/exit logs should show up when ROCDEC_LOG_LEVEL >= 3. All conformance tests should pass. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Nikhil-Nunna
pushed a commit
that referenced
this pull request
Mar 13, 2026
## Motivation This PR adds entry and exit logs for core functions. ## Technical Details - Info level (3) logs are added at core function entry and exit points. - Monotonic clock stamp, process id, thread id, and file name are added to the logs. - The logs now have similar format to HIP logs for better viewing when both are present. - Direct error or warning logs to std::cerr. Other logs still go to std::cout. ## JIRA ID <!-- If applicable, mention the JIRA ID resolved by this PR (Example: Resolves SWDEV-12345). --> <!-- Do not post any JIRA links here. --> ## Test Plan Set env variable ROCDEC_LOG_LEVEL to control log output. All CTests and conformance tests must pass. Example summary of the test results: - CTests: Start 1: video_decodeRaw-HEVC 1/15 Test #1: video_decodeRaw-HEVC ............. Passed 2.05 sec Start 2: video_decodeRaw-AVC 2/15 Test #2: video_decodeRaw-AVC .............. Passed 1.89 sec Start 3: video_decodeRaw-AV1 3/15 Test #3: video_decodeRaw-AV1 .............. Passed 2.31 sec Start 4: video_decodeRaw-VP9 4/15 Test #4: video_decodeRaw-VP9 .............. Passed 2.04 sec Start 5: rocdec_Decode-HEVC 5/15 Test #5: rocdec_Decode-HEVC ............... Passed 0.92 sec Start 6: rocDecode_Negative_API_Tests 6/15 Test #6: rocDecode_Negative_API_Tests ..... Passed 0.78 sec Start 7: video_decode-HEVC 7/15 Test #7: video_decode-HEVC ................ Passed 2.74 sec Start 8: video_decode-AVC 8/15 Test #8: video_decode-AVC ................. Passed 2.64 sec Start 9: video_decode-AV1 9/15 Test #9: video_decode-AV1 ................. Passed 2.74 sec Start 10: video_decode-VP9 10/15 Test #10: video_decode-VP9 ................. Passed 2.78 sec Start 11: video_decodePerf-HEVC 11/15 Test #11: video_decodePerf-HEVC ............ Passed 2.92 sec Start 12: video_decodeBatch 12/15 Test #12: video_decodeBatch ................ Passed 4.34 sec Start 13: video_decodeRGB-HEVC 13/15 Test #13: video_decodeRGB-HEVC ............. Passed 33.51 sec Start 14: video_decodeMem-HEVC 14/15 Test #14: video_decodeMem-HEVC ............. Passed 2.09 sec Start 15: video_decodeRGB-Resize 15/15 Test #15: video_decodeRGB-Resize ........... Passed 33.94 sec 100% tests passed, 0 tests failed out of 15 - AVC conformance test: Conformance test completed on the 127 streams: - The number of passing streams is 127 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 - HEVC conformance test: Conformance test completed on the 135 streams: - The number of passing streams is 135 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 - AV1 conformance test: Conformance test completed on the 162 streams: - The number of passing streams is 162 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 - VP9 conformance test: Conformance test completed on the 108 streams: - The number of passing streams is 108 - The number of failing streams is 0 - The number of streams that did not finish decoding is 0 ## Test Result Function entry/exit logs should show up when ROCDEC_LOG_LEVEL >= 3. All conformance tests should pass. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
1 task
ammarwa
added a commit
that referenced
this pull request
Apr 21, 2026
Ships the informational diff variant (Confluence row #7): * ``perfxpert diff <baseline.db> <new.db> [--format …] [-o …] [-d …]`` — always returns rc=0. Calls the shared ``trace_diff.diff_runs`` engine and renders the result in one of four formats: - ``text`` — two-block overview + per-kernel table (stdout) - ``markdown`` — GitHub-flavored table + narrative (stdout) - ``json`` — the raw schema 0.3.1 dict - ``webview`` — standalone HTML with overview KPIs, stacked delta-bars (red = regression, green = improvement), sortable kernel table, narrative card * ``perfxpert analyze -i new.db --baseline baseline.db`` — end-of- session recap: after the agentic pipeline renders, splice a "Changed vs baseline" section into the main report. Works for all 4 formats (JSON adds a ``trace_diff`` top-level key; schema bumps to 0.3.1). One engine, four call sites: CLI, MCP tool, analyze-splice, and the upcoming ``perfxpert ci`` wrapper all route through ``trace_diff.diff_runs`` + ``cli.diff_cmd.render_diff``. Tests: 9 new (``test_diff_cmd`` × 4, ``test_analyze_baseline_splice`` × 5). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ammarwa
added a commit
that referenced
this pull request
Apr 21, 2026
…SSION_THRESHOLD
Adds a standalone gate rule ``trace_diff_regression_rule`` to
``perfxpert.runtime.gate_cascade`` so the correctness cascade, the
``perfxpert ci`` subcommand, and any future agentic caller share the
same regression-detection engine.
Behavior:
* Thin wrapper over ``tools.trace_diff.diff_runs``.
* Verdict ``fail`` when ``wall_delta_pct > threshold``; otherwise ``pass``.
* Threshold precedence: explicit ``threshold_pct=`` > env var
``PERFXPERT_CI_REGRESSION_THRESHOLD`` > :data:`DEFAULT_CI_REGRESSION_THRESHOLD_PCT`
(5.0%). Invalid env values fall back to the default (no exceptions
escape to the caller).
* Returns a machine-readable dict including the full ``diff_runs``
result so gate callers can show the same tables the CLI does.
The existing 5-gate ``evaluate()`` cascade is intentionally untouched —
this is a complementary standalone rule for the "compare two runs
without a candidate patch" flow that Confluence rows #7 + #8 need.
Tests: 8 new — threshold-resolution precedence (4) + verdict
correctness on baseline/tail-hurt/improved fixture pairs (4).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ammarwa
added a commit
that referenced
this pull request
Apr 21, 2026
…mped to 42 tools Documents Confluence rows #5 (severity refinement), #7 (trace_diff tool + subcommand + analyze splice), and #8 (perfxpert ci wrapper): - docs/guides/getting-started.md: new §7 "Compare runs + CI gating (``perfxpert diff`` / ``perfxpert ci``)" with GitHub Actions + GitLab CI snippets; remaining numbered sections shifted by +1. - docs/guides/python-api.md: document ``api.trace_diff_diff_runs`` mirror + expected output shape. - docs/integration/mcp-server.md: tool count 41 → 42; add ``trace_diff.diff_runs`` to the classifier/knowledge snapshot (34 → 35). - docs/contributing/schemas.md: describe the 0.3.1 ``trace_diff`` top-level field (shape, emission rules, own ``schema_version``); note severity + severity_label additions to ``hotspots[i].source_locations``. - docs/contributing/output_formats.md: contract checklist gains the conditional ``trace_diff`` top-level key; adds a severity- coloring section (bucket table + webview/text/markdown propagation rules) and a diff-reports contract requiring every format to surface regression-vs-improvement direction. Validation: - scripts/lint.sh --strict rc=0 - scripts/link-checker.py --strict rc=0 - scripts/test-samples.py --strict rc=0 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ammarwa
added a commit
that referenced
this pull request
Apr 23, 2026
… polish + contributing docs (#5174) ## Summary Proven-optimizations catalog, nightly benchmark harness, contributing guides + RFC process + architecture refresh, `perfxpert doctor` polish, Phase 7.1 `ai_analysis/` module deletion, and the convergence-review fixes that establish real legacy-vs-agentic parity. This is the last PR in the original retirement series; subsequent PRs (#5179, #5180) add the bundled opencode fork and the docs-audit toolchain. Stacks on #5173. Next in stack: #5179. ## What lands ### Phase 7.1 — `perfxpert/ai_analysis/` module removed Commit `cb03e82c36` deletes the entire `ai_analysis/` module — **21 files, ~11.7 KLOC**. All execution now routes through `perfxpert/agents/`; the dispatcher and the `PERFXPERT_LEGACY` env toggle are gone. This completes the retirement started in #5173: there is no longer a "legacy path" to opt back into. ### Proven-optimizations catalog (10 canonical case studies) - `knowledge/proven_optimizations.yaml`: vgpr_reduction, lds_tiling, mfma_enablement, memory_coalescing, fast_math, hip_stream_overlap, kernel_fusion, device_sync_removal, warp_primitives, cache_blocking - Dual-schema fixture integration: synthetic `duration_ns` + real rocprofv3 UUID-suffixed `(end_ns - start_ns)` tables - `tests/fixtures/*.db`: 5 fixtures re-profiled with `--pmc` on real MI300X hardware ### Nightly benchmark harness + orchestration ### Contributing guides + RFC process + architecture refresh ### `perfxpert doctor` polish (8 checks incl. opencode bundled + LLM providers) ## Review findings addressed (29 findings, 5 fix batches) ### Batch 1 — Gate cascade fidelity (3 commits) - `gate-2-sol-signature`: `_run_sol_gate` took wrong keyword; unit tests mocked the function so the TypeError went undetected. Two-tier check now: hard cap on speedup ratio (>50× rejected arch-independent) + optional absolute FLOPS check via `sol.sanity_check`. 13 new un-mocked integration tests (`TestRunSolGateUnmocked`). - `gate-4-hot-kernel-key-unit`: `regression.compare_runs` returns `per_kernel_deltas` (list of dicts with `delta_pct` as a **fraction**, not percent), not `hot_kernels`. Old code iterated a non-existent key so Gate 4 silently treated all regressions as OK. Fraction↔percent conversion corrected. - `gate-5-skip-verdict`: verdict strings no longer misleading on skip (`candidate_binary=None` path). ### Batch 2 — Schemas, validators, framework invariants (6 findings) - #4: Pydantic cross-field validators on `GateVerdictModel` (status/failing_gate coupling), `CorrectnessOutput` (verdict/action), confidence clamps `Field(..., ge=0.0, le=1.0)`, typed `Literal` for provider - #12: `Agent` dataclass `frozen=True`; `__post_init__` validates layer in (0,1,2) - #7: `_parse_headers` raises `ValueError` on malformed JSON instead of silent `{}` - #10, #18: tasks module logs resolved store path; stale READ_ONLY docstrings rewritten - #17: README clarifies opencode is system dependency; doctor warns when PERFXPERT_TASK_ROOT unset ### Batch 3 — External tool verifier + fence alignment (6 findings) - N28: anthropic/openai SDK imports routed through `require_tool()` with install-hint on missing - N29: all 8 agent fence slices aligned to actual agent tool allowlists - #5, #6, #9, #17: smoke-test opencode binary (not just presence), hard-fail on bad `PERFXPERT_OPENCODE_PATH` override, skip subdirs in config copy - `tooldep`: user-home path search + `PERFXPERT_ROCPROF_TRACE_DECODER_PATH` env support ### Batch 4 — Classifier + data-provenance fidelity (2 findings) - **#1**: classifier no longer swallows DB errors silently (was returning empty metrics → `data_insufficient`; now logs + propagates error markers) - **#26 (priority user-flagged)**: confidence now weighted by `evidence_strength × match_ratio` instead of raw rule-count normalization — partial-PMC classifications no longer over-confident. New `test_partial_pmc_confidence` covers trace-only vs full-PMC paths. ### Batch 5 — Tests coverage gaps (9 findings) - #20: gate cascade short-circuit coverage (Gates 2/3/4) - #21: `_TraceAnalysisAdapter` shim delegation tests - #22: trace-only `memcpy_pct` path in `classify_from_metrics` - #23: `anthropic.APITimeoutError` normalization path - #24: Root agent lambda stub guard (`test_root.py::test_lambda_stubs_never_invoked`) - #25: MCP tool coverage across 5 representative tools - #27: flaky `time.sleep` → `proc.wait(timeout=5)` - #14, #15, #16: stale comments + docstring rewrites (api.py, __main__.py, recommendation.py) ## Parity story **Problem found during review**: parity test wiring was toggling a dead flag (`PERFXPERT_USE_AGENTS`) after the retirement PR flipped the default — both legs were running the agentic path. "Parity" passed vacuously. **Fixes applied (Decision A+C):** - **A**: Parity runner now toggles the operative legacy opt-out so legacy and agentic paths are genuinely distinct; Tier-1 trace-only fixtures `xfail`ed until re-profiled with `--pmc`. - **C**: Re-profiled all 5 fixtures (`baseline`, `memory_bound`, `compute_bound`, `latency_bound`, `regressed`) on real MI300X with `rocprofv3 --pmc`: - `compute_bound`: VALU-bound kernel (not MFMA, to legitimately match Compute-Bound classification) - `memory_bound` / `baseline`: 128 MB H2D+D2H loop - `latency_bound` / `regressed`: 6,000 tiny kernels - `_build_analysis_result` returns `api_overhead` (not `latency`) for `overhead_pct > 25%` (legacy heuristic fix) - E2E + dashboard tests updated for evidence-weighted classification **Result**: `parity_snapshots/_aggregate.json` → **100% agreement on PMC fixtures** (10/10 bottleneck + rec_technique signals). `rec_type` signal skipped where the agentic path can't populate without an LLM (documented asymmetric None). ## Convergence review round (2026-04-18) - **Brand scrub**: zero-tolerance removal of the prior codename across code, tests, patches, env-var names, and documentation. - **CLI flag plumbing**: all 10 CLI flags (`--prompt`, `--llm`, `--llm-model`, `--att-dir`, `--pc-sampling`, `--source-dir`, `--format`, `-d/-o`) now wired end-to-end via an `analysis_options` side-channel on `Context`. - **PMC shard UNION** across multiple `.db` files uses column-intersection and emits a `RuntimeWarning` listing dropped columns. - **FallbackProvider** has a typed exception taxonomy (`LLMRateLimitError`, `LLMAuthenticationError`, `LLMTransientError`, `LLMFatalError`). Auth/fatal errors surface immediately; only transient + rate-limit cascade. The provider is wired into `build_session`; `PERFXPERT_LLM_FALLBACK_CHAIN` now has runtime effect. ## Test plan - [x] 1036 pass / 3 skip / 0 fail from `experimental/python/perfxpert/` - [x] Core suites — 100% pass: `tests/test_agents + test_providers + test_runtime + test_tools` - [x] Parity ship-gate — 100%: `tests/test_parity/` (12 passed, 5 rec_type skips documented) - [x] Gate cascade — 100%: 28 tests including 13 new un-mocked SOL integration tests (Batch 1 recovery) - [x] E2E brain: `perfxpert-code run -m openai/gpt-4o-mini "..."` → MCP → Python → `81.7 TFLOPS` + AMD footer - [x] `perfxpert doctor`: all 8 checks green (34 MCP tools at this PR boundary; 43 after #5179+#5180; opencode 1.4.11 bundled; 3/5 LLM providers configured) - [x] No legacy opt-out path remains (Phase 7.1 removed the dispatcher and the `PERFXPERT_LEGACY` toggle) - [ ] Live MI300X nightly run (requires hardware; nightly workflow will refresh baseline) - [ ] Full interactive TUI (requires TTY; E2E can only prove banner-emitted) --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
vlaindic
added a commit
that referenced
this pull request
May 18, 2026
Translation parser (#3, #6): Replace the residual ROCP_FATAL on wave_id mismatch in the GFX1250 V2/V4 and stochastic copySample paths with ROCP_ERROR + returning a value-initialized record so it is filtered by is_invalid_sample() instead of aborting the process. Drop the redundant std::memset() that followed value-initialization of the invalid-sample sentinel in the GFX9/GFX12/GFX1250 V2 and GFX1250 V4 paths, and document the timestamp-zero invariant. query_agent_configurations_v2 (#5): Replace the dead "sort preferred-method configs first" pass with the same PREFER semantics already implemented by configure_v2: if any preferred-method config exists, return only those; otherwise fall back to returning all configs. REQUIRE remains a hard filter. Callback signature: Change rocprofiler_available_pc_sampling_configurations_v2_cb_t to deliver configurations as a const T** array of pointers, matching the convention used by rocprofiler_query_available_agents_cb_t. Update the only consumer (samples/pc_sampling_v2/pcs.cpp) accordingly. Documentation (#7, #8): Remove the V5 reference (only V0-V4 exist today); expand the api_flags doc with PREFER/REQUIRE behavior for both query and configure entry points and add per-enumerator @var blocks; add an Example 4 to configure_v2 showing conflicting flags being rejected; mark configuration_v2_t as output-only; add per-flag filtering rules and three usage examples to query_v2's doc; expand _get_parse_func_for_record_kind's docstring to specify when nullptr is returned. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
vlaindic
added a commit
that referenced
this pull request
May 27, 2026
Translation parser (#3, #6): Replace the residual ROCP_FATAL on wave_id mismatch in the GFX1250 V2/V4 and stochastic copySample paths with ROCP_ERROR + returning a value-initialized record so it is filtered by is_invalid_sample() instead of aborting the process. Drop the redundant std::memset() that followed value-initialization of the invalid-sample sentinel in the GFX9/GFX12/GFX1250 V2 and GFX1250 V4 paths, and document the timestamp-zero invariant. query_agent_configurations_v2 (#5): Replace the dead "sort preferred-method configs first" pass with the same PREFER semantics already implemented by configure_v2: if any preferred-method config exists, return only those; otherwise fall back to returning all configs. REQUIRE remains a hard filter. Callback signature: Change rocprofiler_available_pc_sampling_configurations_v2_cb_t to deliver configurations as a const T** array of pointers, matching the convention used by rocprofiler_query_available_agents_cb_t. Update the only consumer (samples/pc_sampling_v2/pcs.cpp) accordingly. Documentation (#7, #8): Remove the V5 reference (only V0-V4 exist today); expand the api_flags doc with PREFER/REQUIRE behavior for both query and configure entry points and add per-enumerator @var blocks; add an Example 4 to configure_v2 showing conflicting flags being rejected; mark configuration_v2_t as output-only; add per-flag filtering rules and three usage examples to query_v2's doc; expand _get_parse_func_for_record_kind's docstring to specify when nullptr is returned. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
AryanSalmanpour
added a commit
that referenced
this pull request
Jun 10, 2026
…error diagnostics in parser (#7010) ## Motivation This PR enhances rocJPEG logging by adding structured, offset-annotated debug logging for each major JPEG marker (SOF0, DQT, DHT, SOS, EOI), improves clarity of error messages for unsupported formats, and explicitly rejects progressive JPEGs (SOF2). sample output: ``` ROCJPEG_LOG_LEVEL=4 ./jpegdecode -i ~/install/share/rocjpeg/images/mug_420.jpg Reading images from disk, please wait! Using GPU device 0: AMD Radeon RX 7900 XT[gfx1100] on PCI bus 03:00.0 [3, Info] rocjpeg_api.cpp:123: 2274381572050 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegCreate( 0, 0, 0x7fff0596b718 ): entry ... [3, Info] rocjpeg_decoder.cpp:81: 2274381572066 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitializeDecoder( ): entry ... [3, Info] rocjpeg_decoder.cpp:50: 2274381572069 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitHIP( 0 ): entry ... [3, Info] rocjpeg_decoder.cpp:50: 2274381592272 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitHIP( 0 ): exit (20203 us) ... [3, Info] rocjpeg_vaapi_decoder.cpp:416: 2274381592356 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitializeDecoder(): gpu_uuids_to_render_nodes_map_: {e0b64ad979bec1fd: 128} [3, Info] rocjpeg_vaapi_decoder.cpp:417: 2274381592362 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitializeDecoder(): Selected GPU UUID: e0b64ad979bec1fd [3, Info] rocjpeg_vaapi_decoder.cpp:475: 2274381592366 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitVAAPI(): Opening DRM node: /dev/dri/renderD128 [3, Info] rocjpeg_vaapi_decoder.cpp:500: 2274381593624 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitVAAPI(): VA-API version 1.22 [3, Info] rocjpeg_vaapi_decoder.cpp:502: 2274381593634 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitVAAPI(): VA-API vendor: Mesa Gallium driver 26.0.0-devel for AMD Radeon RX 7900 XT (radeonsi, navi31, , DRM 3.64, 6.17.0-23-generic) [3, Info] rocjpeg_vaapi_decoder.cpp:504: 2274381593638 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitVAAPI(): Trying to open /home/aryan/install/lib/rocm_sysdeps/lib/radeonsi_drv_video.so [3, Info] rocjpeg_decoder.cpp:81: 2274381593728 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] InitializeDecoder( ): exit (21662 us) ... [3, Info] rocjpeg_api.cpp:123: 2274381593732 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegCreate( 0, 0, 0x7fff0596b718 ): exit (21682 us) ... [3, Info] rocjpeg_api.cpp:40: 2274381593735 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegStreamCreate( 0x7fff0596b710 ): entry ... [3, Info] rocjpeg_api.cpp:40: 2274381593738 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegStreamCreate( 0x7fff0596b710 ): exit (3 us) ... Input file name: /home/aryan/install/share/rocjpeg/images/mug_420.jpg [3, Info] rocjpeg_api.cpp:75: 2274381594731 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegStreamParse( 0x70b1f41c4010, 2340067, 0x55f2ebfec840 ): entry ... [3, Info] rocjpeg_parser.cpp:48: 2274381594734 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseJpegStream( 0x70b1f41c4010, 2340067 ): entry ... [4, Debug] rocjpeg_parser.cpp:408: 2274381594759 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseDQT(): *** Marker: DQT (xFFDB) *** Define a Quantization Table. OFFSET: 0x00001296 Table length = 67 ---- Precision=8 bits Destination ID=0 (Luminance) DQT, Row #0: 1 1 1 1 1 1 1 1 DQT, Row #1: 1 1 1 1 1 1 1 1 DQT, Row #2: 1 1 1 1 1 1 1 1 DQT, Row #3: 1 1 1 1 1 1 1 1 DQT, Row #4: 1 1 1 1 1 1 1 1 DQT, Row #5: 1 1 1 1 1 1 1 1 DQT, Row #6: 1 1 1 1 1 1 1 1 DQT, Row #7: 1 1 1 1 1 1 1 1 Approx quality factor = 100.00 (scaling=2.99 variance=6.13) [4, Debug] rocjpeg_parser.cpp:408: 2274381594772 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseDQT(): *** Marker: DQT (xFFDB) *** Define a Quantization Table. OFFSET: 0x000012DB Table length = 67 ---- Precision=8 bits Destination ID=1 (Chrominance) DQT, Row #0: 1 1 1 1 1 1 1 1 DQT, Row #1: 1 1 1 1 1 1 1 1 DQT, Row #2: 1 1 1 1 1 1 1 1 DQT, Row #3: 1 1 1 1 1 1 1 1 DQT, Row #4: 1 1 1 1 1 1 1 1 DQT, Row #5: 1 1 1 1 1 1 1 1 DQT, Row #6: 1 1 1 1 1 1 1 1 DQT, Row #7: 1 1 1 1 1 1 1 1 Approx quality factor = 100.00 (scaling=1.54 variance=1.58) [4, Debug] rocjpeg_parser.cpp:297: 2274381594790 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseSOF(): *** Marker: SOF0 (Baseline DCT) (xFFC0) *** OFFSET: 0x00001320 Frame header length = 17 Precision = 8 Number of Lines = 2160 Samples per Line = 3840 Image Size = 3840 x 2160 Chroma subsampling: 4:2:0 Raw Image Orientation = Landscape Number of Img components = 3 Component[1]: ID=0x01, Samp Fac=0x22 (Subsamp 1 x 1), Quant Tbl Sel=0x00 (Lum: Y) Component[2]: ID=0x02, Samp Fac=0x11 (Subsamp 2 x 2), Quant Tbl Sel=0x01 (Chrom: Cb) Component[3]: ID=0x03, Samp Fac=0x11 (Subsamp 2 x 2), Quant Tbl Sel=0x01 (Chrom: Cr) [4, Debug] rocjpeg_parser.cpp:515: 2274381594802 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseDHT(): *** Marker: DHT (Define Huffman Table) (xFFC4) *** OFFSET: 0x00001333 Huffman table length = 31 ---- Destination ID = 0 Class = 0 (DC / Lossless Table) Codes of length 01 bits (000 total): Codes of length 02 bits (002 total): 03 04 Codes of length 03 bits (002 total): 01 02 Codes of length 04 bits (003 total): 00 05 06 Codes of length 05 bits (001 total): 07 Codes of length 06 bits (001 total): 08 Codes of length 07 bits (001 total): 09 Codes of length 08 bits (001 total): 0A Codes of length 09 bits (001 total): 0B Codes of length 0A bits (000 total): Codes of length 11 bits (000 total): Codes of length 12 bits (000 total): Codes of length 13 bits (000 total): Codes of length 14 bits (000 total): Codes of length 15 bits (000 total): Codes of length 16 bits (000 total): Total number of codes: 012 [4, Debug] rocjpeg_parser.cpp:515: 2274381594816 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseDHT(): *** Marker: DHT (Define Huffman Table) (xFFC4) *** OFFSET: 0x00001354 Huffman table length = 91 ---- Destination ID = 0 Class = 1 (AC Table) Codes of length 01 bits (000 total): Codes of length 02 bits (002 total): 01 02 Codes of length 03 bits (001 total): 11 Codes of length 04 bits (002 total): 03 21 Codes of length 05 bits (004 total): 00 04 12 31 Codes of length 06 bits (003 total): 05 41 51 Codes of length 07 bits (005 total): 06 13 22 61 71 Codes of length 08 bits (005 total): 32 81 91 A1 D1 Codes of length 09 bits (005 total): 07 52 B1 C1 F0 Codes of length 0A bits (005 total): 08 14 23 42 E1 Codes of length 0B bits (005 total): 62 72 92 A2 D2 Codes of length 0C bits (007 total): 09 15 33 53 82 E2 F1 Codes of length 0D bits (002 total): 24 B2 Codes of length 0E bits (001 total): C2 Codes of length 0F bits (002 total): 16 43 Codes of length 10 bits (023 total): 93 63 D3 0A 17 25 73 A3 34 54 83 18 19 B3 26 27 44 F2 35 64 C3 E3 46 Total number of codes: 072 [4, Debug] rocjpeg_parser.cpp:515: 2274381594829 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseDHT(): *** Marker: DHT (Define Huffman Table) (xFFC4) *** OFFSET: 0x000013B1 Huffman table length = 29 ---- Destination ID = 1 Class = 0 (DC / Lossless Table) Codes of length 01 bits (001 total): 00 Codes of length 02 bits (000 total): Codes of length 03 bits (003 total): 01 02 03 Codes of length 04 bits (001 total): 04 Codes of length 05 bits (001 total): 05 Codes of length 06 bits (001 total): 06 Codes of length 07 bits (001 total): 07 Codes of length 08 bits (001 total): 08 Codes of length 09 bits (001 total): 09 Codes of length 0A bits (000 total): Codes of length 11 bits (000 total): Codes of length 12 bits (000 total): Codes of length 13 bits (000 total): Codes of length 14 bits (000 total): Codes of length 15 bits (000 total): Codes of length 16 bits (000 total): Total number of codes: 010 [4, Debug] rocjpeg_parser.cpp:515: 2274381594842 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseDHT(): *** Marker: DHT (Define Huffman Table) (xFFC4) *** OFFSET: 0x000013D0 Huffman table length = 83 ---- Destination ID = 1 Class = 1 (AC Table) Codes of length 01 bits (000 total): Codes of length 02 bits (001 total): 01 Codes of length 03 bits (001 total): 11 Codes of length 04 bits (004 total): 00 02 21 61 Codes of length 05 bits (007 total): 03 31 41 51 71 91 F0 Codes of length 06 bits (006 total): 12 22 52 81 A1 D1 Codes of length 07 bits (004 total): 04 62 B1 E1 Codes of length 08 bits (005 total): 13 32 82 A2 C1 Codes of length 09 bits (004 total): 05 42 92 F1 Codes of length 0A bits (000 total): Codes of length 11 bits (004 total): 06 14 23 72 Codes of length 0C bits (004 total): 07 33 53 B2 Codes of length 0D bits (005 total): 15 43 C2 D2 E2 Codes of length 0E bits (001 total): 16 Codes of length 0F bits (001 total): 63 Codes of length 10 bits (017 total): 93 24 34 54 08 73 83 17 18 44 A3 35 F2 25 74 B3 D3 Total number of codes: 064 [4, Debug] rocjpeg_parser.cpp:606: 2274381594856 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseSOS(): *** Marker: SOS (Start of Scan) (xFFDA) *** OFFSET: 0x00001425 Scan header length = 12 Number of img components = 3 Component[1]: selector=0x01, table=0(DC),0(AC) Component[2]: selector=0x02, table=1(DC),1(AC) Component[3]: selector=0x03, table=1(DC),1(AC) Spectral selection = 0 .. 63 Successive approximation = 0x00 [4, Debug] rocjpeg_parser.cpp:669: 2274381595378 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseEOI(): *** Marker: EOI (End of Image) (xFFD9) *** OFFSET: 0x0023B4E1 [3, Info] rocjpeg_parser.cpp:48: 2274381595383 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] ParseJpegStream( 0x70b1f41c4010, 2340067 ): exit (649 us) ... [3, Info] rocjpeg_api.cpp:75: 2274381595386 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegStreamParse( 0x70b1f41c4010, 2340067, 0x55f2ebfec840 ): exit (655 us) ... [3, Info] rocjpeg_api.cpp:188: 2274381595393 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegGetImageInfo( 0x55f2ebf6dde0, 0x55f2ebfec840, 0x7fff0596b82a, 0x7fff0596b728, 0x7fff0596b810, 0x7fff0596b800 ): entry ... [3, Info] rocjpeg_decoder.cpp:332: 2274381595397 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] GetImageInfo( 0x55f2ebfec840, 0x7fff0596b82a, 0x7fff0596b728, 0x7fff0596b810, 0x7fff0596b800 ): entry ... [3, Info] rocjpeg_decoder.cpp:332: 2274381595400 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] GetImageInfo( 0x55f2ebfec840, 0x7fff0596b82a, 0x7fff0596b728, 0x7fff0596b810, 0x7fff0596b800 ): exit (3 us) ... [3, Info] rocjpeg_api.cpp:188: 2274381595403 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegGetImageInfo( 0x55f2ebf6dde0, 0x55f2ebfec840, 0x7fff0596b82a, 0x7fff0596b728, 0x7fff0596b810, 0x7fff0596b800 ): exit (10 us) ... Input image resolution: 3840x2160 Chroma subsampling: YUV 4:2:0 Decoding started, please wait! ... [3, Info] rocjpeg_api.cpp:222: 2274381595488 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegDecode( 0x55f2ebf6dde0, 0x55f2ebfec840, 0x7fff0596b6c0, 0x7fff0596b6e0 ): entry ... [3, Info] rocjpeg_decoder.cpp:120: 2274381595492 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] Decode( 0x55f2ebfec840, 0x7fff0596b6c0, 0x7fff0596b6e0 ): entry ... [3, Info] rocjpeg_decoder.cpp:120: 2274381607931 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] Decode( 0x55f2ebfec840, 0x7fff0596b6c0, 0x7fff0596b6e0 ): exit (12439 us) ... [3, Info] rocjpeg_api.cpp:222: 2274381607937 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegDecode( 0x55f2ebf6dde0, 0x55f2ebfec840, 0x7fff0596b6c0, 0x7fff0596b6e0 ): exit (12449 us) ... Average processing time per image (ms): 12.4542 Average images per sec: 80.2941 [3, Info] rocjpeg_api.cpp:154: 2274381608091 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegDestroy( 0x55f2ebf6dde0 ): entry ... [3, Info] rocjpeg_api.cpp:154: 2274381610464 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegDestroy( 0x55f2ebf6dde0 ): exit (2373 us) ... [3, Info] rocjpeg_api.cpp:99: 2274381610471 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegStreamDestroy( 0x55f2ebfec840 ): entry ... [3, Info] rocjpeg_api.cpp:99: 2274381610474 us: [pid:1740825 tid:1740825 hashid:0xbc0a7] rocJpegStreamDestroy( 0x55f2ebfec840 ): exit (3 us) ... Decoding completed! ``` ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## JIRA ID <!-- If applicable, mention the JIRA ID resolved by this PR (Example: Resolves SWDEV-12345). --> <!-- Do not post any JIRA links here. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
abouteiller
added a commit
that referenced
this pull request
Jun 26, 2026
## Motivation
On certain test systems we do not have a shared filesystem, the
deadlock_analysis script harness would depend on a shared filesystem to
collect traces and artifacts. Remove that limitation.
## Technical Details
Use MPI stdout redirection to coalesce the trace artifacts to the
headnode in a headnode local output directory, perform analysis on the
headnode.
Prior filesystem based mode remains available and default.
## JIRA ID
## Test Plan
Intentionally added a deadlock to a functional test, compiled with
Debug.
Run mpiexec_deadlock_attach script and observe artifacts.
## Test Result
```
ROCSHMEM_DEBUG_LEVEL=trace:stats ~/ompi/bin/mpirun -n 2 tests/functional_tests/rocshmem_functional_tests -n 1 -a 26 & (sleep 2; /home/abouteil/rocm-systems/projects/rocshmem/scripts/debug/mpiexec_deadlock_analysis.sh rocshmem_functional_tests --stats --color --no-shared-fs )
### Creating Test: Blocking WG level Puts B=ipc PE=2 W=1 Z=64 ###
Deploying deadlock analysis...
Executable: rocshmem_functional_tests
Output directory: ./rocshmem_deadlock_analysis_20260530_013437
Filesystem mode: no shared fs (--output-filename)
Stats: yes
Color: --color
Found 4 process(es) matching 'rocshmem_functional_tests': 717472 717473 751554 751555
Output directory: /tmp/rocshmem_deadlock_20260530_013437
Per-process timeout: 120s
Analyzing PID 717472 (PE 0) ... OK
Analyzing PID 717473 (PE 1) ... OK (no GPU threads)
Analyzing PID 751554 (PE 0) ... OK
Analyzing PID 751555 (PE 1) ... OK (no GPU threads)
=== Attach and Analyze Summary ===
Processes analyzed: 4 / 4
PE 0 PID 751554 groups=1 \033[1;31m1\033[0m wf(s) in rocSHMEM; \033[1;36mWaiting for remote PE memory update (barrier/sync or user rocshmem_wait_until). Check if the remote PE is alive and making progress. (at: #0 0x0000792efa5af3c8 in rocshmem::Context::wait_until<long> (this=0x792eed407000, ivars=0x792efac04008, cmp=0, val=1) at src/context_tmpl_device.hpp:224)\033[0m
PE 1 PID 751555 groups=? (no GPU threads found)
=== BEGIN FILE: pe0_pid_717472.txt ===
[New LWP 717493]
[New LWP 717490]
[New LWP 717488]
[New LWP 717485]
[New LWP 717481]
[New LWP 717480]
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
0x000077a09e06695f in ?? () from /opt/rocm-7.2.1/lib/libhsa-runtime64.so.1
=== rocSHMEM Deadlock Analysis ===
Process(es): 717472
Total GPU wavefronts analyzed: 1
Unique backtrace groups: 1
--- Group 1 (1 wavefront(s)) ---
WFs: (0,0,0)/0
Backtrace:
#0 0x0000776005fafce0 in rocshmem::uncached_load<long volatile> (src=0x776006604008) at src/assembly.hpp:125
#1 rocshmem::Context::test<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:432
#2 rocshmem::Context::wait_until<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:224
#3 rocshmem::IPCContext::internal_direct_barrier (this=0x775ff8e07000, pe=0, PE_start=0, stride=1, n_pes=2, pSync=0x776006604000) at build/src/ipc/context_ipc_device_coll.cpp:44
#4 0x0000776005fb7c24 in rocshmem::IPCContext::internal_sync_wg (this=0x775ff8e07000, pe=0, PE_start=0, stride=1, PE_size=2, pSync=0x776006604000) at build/src/ipc/context_ipc_device_coll.cpp:115
#5 0x000077600603d018 in rocshmem::IPCContext::sync_all_wg (this=0x775ff8e07000) at build/src/ipc/context_ipc_device_coll.cpp:168
#6 0x000077600603d558 in rocshmem::IPCContext::barrier_all_wg (this=0x775ff8e07000) at build/src/ipc/context_ipc_device_coll.cpp:187
#7 0x000077600603dedc in rocshmem::Context::barrier_all_wg (this=0x775ff8e07000) at build/src/context_device.cpp:151
#8 0x000077600603e33c in rocshmem::rocshmem_barrier_all_wg () at build/src/rocshmem_gpu.cpp:885
#9 0x00007760060437cc in WorkGroupPrimitiveTest (loop=1, skip=10, start_time=0x775ff8e31000, end_time=0x775ff8e32000, source=0x776013c00480 "abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqr"..., dest=0x776013d00500 "1", size=1, type=WGPutTestType, ctx_type=4) at /home/abouteil/rocm-systems/projects/rocshmem/tests/functional_tests/workgroup_primitives.cpp:72
[HINT] Waiting for remote PE memory update (barrier/sync or user rocshmem_wait_until). Check if the remote PE is alive and making progress. (at: #2 rocshmem::Context::wait_until<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:224)
...
=== END FILE: pe1_pid_751555.txt ===
Reconstructing per-PE output files from mpiexec output...
Reconstructed output from 1 rank file(s).
Running cross-rank coalescing pass...
=== Cross-Rank Deadlock Analysis ===
Directory: ./rocshmem_deadlock_analysis_20260530_013437
Ranks with analysis files: 2 (0, 1)
Unique backtrace patterns: 2
--- Pattern 1 [UNIQUE: rank 0] ---
Ranks: 0
Total wavefronts: 1
Backtrace:
#0 0x0000776005fafce0 in rocshmem::uncached_load<long volatile> (src=0x776006604008) at src/assembly.hpp:125
#1 rocshmem::Context::test<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:432
#2 rocshmem::Context::wait_until<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:224
#3 rocshmem::IPCContext::internal_direct_barrier (this=0x775ff8e07000, pe=0, PE_start=0, stride=1, n_pes=2, pSync=0x776006604000) at build/src/ipc/context_ipc_device_coll.cpp:44
#4 0x0000776005fb7c24 in rocshmem::IPCContext::internal_sync_wg (this=0x775ff8e07000, pe=0, PE_start=0, stride=1, PE_size=2, pSync=0x776006604000) at build/src/ipc/context_ipc_device_coll.cpp:115
#5 0x000077600603d018 in rocshmem::IPCContext::sync_all_wg (this=0x775ff8e07000) at build/src/ipc/context_ipc_device_coll.cpp:168
#6 0x000077600603d558 in rocshmem::IPCContext::barrier_all_wg (this=0x775ff8e07000) at build/src/ipc/context_ipc_device_coll.cpp:187
#7 0x000077600603dedc in rocshmem::Context::barrier_all_wg (this=0x775ff8e07000) at build/src/context_device.cpp:151
#8 0x000077600603e33c in rocshmem::rocshmem_barrier_all_wg () at build/src/rocshmem_gpu.cpp:885
#9 0x00007760060437cc in WorkGroupPrimitiveTest (loop=1, skip=10, start_time=0x775ff8e31000, end_time=0x775ff8e32000, source=0x776013c00480 "abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqr"..., dest=0x776013d00500 "1", size=1, type=WGPutTestType, ctx_type=4) at /home/abouteil/rocm-systems/projects/rocshmem/tests/functional_tests/workgroup_primitives.cpp:72
[HINT] Waiting for remote PE memory update (barrier/sync or user rocshmem_wait_until). Check if the remote PE is alive and making progress. (at: #2 rocshmem::Context::wait_until<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:224)
=== Per-Rank rocSHMEM Wavefront Count ===
Rank 0: 2 wf(s) inside rocSHMEM
Rank 1: 0 wf(s) inside rocSHMEM
=== Summary ===
0 pattern(s) common to all 2 rank(s)
1 pattern(s) present in exactly one rank
Most common hint: Waiting for remote PE memory update (barrier/sync or user rocshmem_wait_until). Check if the remote PE is alive and making progress. (at: #2 rocshmem::Context::wait_until<long> (this=0x775ff8e07000, ivars=0x776006604008, cmp=0, val=1) at src/context_tmpl_device.hpp:224)
```
abouteiller
added a commit
that referenced
this pull request
Jun 29, 2026
… progress tracker (#6616) ## Motivation When the program deadlocks it can be hard to know how much work was performed thus far, for example did it deadlock after only a couple of communication calls or after a long time, is there only 1 call of 1 type? etc. ## Technical Details builds with PROFILE=ON can produce call statistics (slowing down execution btw), this can be triggered from the gdb attachement to get the stat dump. ## JIRA ID <!-- If applicable, mention the JIRA ID resolved by this PR (Example: Resolves SWDEV-12345). --> <!-- Do not post any JIRA links here. --> ## Test Plan modified a functional test to intentionally deadlock. build with Debug and PROFILE=ON. ``` ROCSHMEM_DEBUG_LEVEL=trace:stats ~/ompi/bin/mpirun -n 2 tests/functional_tests/rocshmem_functional_tests -n 1 -a 26 & (sleep 2; ../scripts/debug/attach_deadlock_analysis.sh rocshmem_functional_tests --stats ) ``` Output file ``` === rocSHMEM Deadlock Analysis === Process(es): 845512 Total GPU wavefronts analyzed: 1 Unique backtrace groups: 1 --- Group 1 (1 wavefront(s)) --- WFs: (0,0,0)/0 Backtrace: #0 0x00007ae55f5afcf0 in rocshmem::uncached_load<long volatile> (src=0x7ae55fc04008) at src/assembly.hpp:125 #1 rocshmem::Context::test<long> (this=0x7ae552407000, ivars=0x7ae55fc04008, cmp=0, val=1) at src/context_tmpl_device.hpp:432 #2 rocshmem::Context::wait_until<long> (this=0x7ae552407000, ivars=0x7ae55fc04008, cmp=0, val=1) at src/context_tmpl_device.hpp:224 #3 rocshmem::IPCContext::internal_direct_barrier (this=0x7ae552407000, pe=0, PE_start=0, stride=1, n_pes=2, pSync=0x7ae55fc04000) at build/src/ipc/context_ipc_device_coll.cpp:44 #4 0x00007ae55f5b7c24 in rocshmem::IPCContext::internal_sync_wg (this=0x7ae552407000, pe=0, PE_start=0, stride=1, PE_size=2, pSync=0x7ae55fc04000) at build/src/ipc/context_ipc_device_coll.cpp:115 #5 0x00007ae55f63d018 in rocshmem::IPCContext::sync_all_wg (this=0x7ae552407000) at build/src/ipc/context_ipc_device_coll.cpp:168 #6 0x00007ae55f63d558 in rocshmem::IPCContext::barrier_all_wg (this=0x7ae552407000) at build/src/ipc/context_ipc_device_coll.cpp:187 #7 0x00007ae55f63dedc in rocshmem::Context::barrier_all_wg (this=0x7ae552407000) at build/src/context_device.cpp:151 #8 0x00007ae55f63e33c in rocshmem::rocshmem_barrier_all_wg () at build/src/rocshmem_gpu.cpp:885 #9 0x00007ae55f6437cc in WorkGroupPrimitiveTest (loop=1, skip=10, start_time=0x7ae552431000, end_time=0x7ae552432000, source=0x7ae56d200480 "abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqr"..., dest=0x7ae56d300500 "1", size=1, type=WGPutTestType, ctx_type=4) at /home/abouteil/rocm-systems/projects/rocshmem/tests/functional_tests/workgroup_primitives.cpp:72 [HINT] Waiting for remote PE memory update (barrier/sync or user rocshmem_wait_until). Check if the remote PE is alive and making progress. (at: #2 rocshmem::Context::wait_until<long> (this=0x7ae552407000, ivars=0x7ae55fc04008, cmp=0, val=1) at src/context_tmpl_device.hpp:224) === rocSHMEM Stats === PE 0 (counters reflect activity since process start) Device: quiet 1 put_wg 704 Host: (all zero) === Summary === 1 wavefront(s) inside rocSHMEM 0 wavefront(s) in user code / other ``` Note the == rocshmem Stats == section contains the number of put and quiet calls. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Signed-off-by: Aurelien Bouteiller <aurelien.bouteiller@amd.com> Signed-off-by: Aurelien Bouteiller <abouteil@amd.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
dlamd1dai
referenced
this pull request
in dlamd1dai/rocm-systems
Jun 29, 2026
300x, 350x, 355x machines. Several issues combined to produce null/invalid pointer writes on MI350X (and could affect any MI300-class machine in containers or with partial builds): Host/device struct skew — ncclGinAnvilSdmaGPUContext gained layoutMagic, queueHandles, and sdmaDirty at the front. Without layoutMagic set on the host, the device read wrong fields as signals → near-null pointer. Fused OSS7 copy+signal (#7) — putSignal() on MI350X (__gfx950__) with an invalid remote signal VA. Silent HIP failures — gin_anvil_factory logged errors but still returned success, leaving null queueHandles/sdmaDirty. getOamId() throws — missing sysfs xgmi_physical_id in Docker broke Anvil connect on some machines. readSignal on line 1 of GinAlltoAllKernel — generic readSignal dereferences GetSignalPtr with no null guard. modified: projects/rccl/src/gin/gin_plugin_rocshmem_anvil.cc modified: projects/rccl/src/include/nccl_device/gin/anvil_sdma/gin_anvil_sdma.h modified: projects/rccl/src/include/nccl_device/gin/anvil_sdma/gin_anvil_sdma_device_host_common.h modified: projects/rocshmem/src/sdma/anvil.cpp modified: projects/rocshmem/src/sdma/gin_anvil_factory.cpp
dlamd1dai
referenced
this pull request
in dlamd1dai/rocm-systems
Jul 5, 2026
300x, 350x, 355x machines. Several issues combined to produce null/invalid pointer writes on MI350X (and could affect any MI300-class machine in containers or with partial builds): Host/device struct skew — ncclGinAnvilSdmaGPUContext gained layoutMagic, queueHandles, and sdmaDirty at the front. Without layoutMagic set on the host, the device read wrong fields as signals → near-null pointer. Fused OSS7 copy+signal (#7) — putSignal() on MI350X (__gfx950__) with an invalid remote signal VA. Silent HIP failures — gin_anvil_factory logged errors but still returned success, leaving null queueHandles/sdmaDirty. getOamId() throws — missing sysfs xgmi_physical_id in Docker broke Anvil connect on some machines. readSignal on line 1 of GinAlltoAllKernel — generic readSignal dereferences GetSignalPtr with no null guard. modified: projects/rccl/src/gin/gin_plugin_rocshmem_anvil.cc modified: projects/rccl/src/include/nccl_device/gin/anvil_sdma/gin_anvil_sdma.h modified: projects/rccl/src/include/nccl_device/gin/anvil_sdma/gin_anvil_sdma_device_host_common.h modified: projects/rocshmem/src/sdma/anvil.cpp modified: projects/rocshmem/src/sdma/gin_anvil_factory.cpp
1 task done
prasanna-amd
added a commit
that referenced
this pull request
Aug 20, 2026
… real headers - Fix double-deref in KernelCh stop: combine per-channel and kernel-completion ref drops into a single atomic operation under the coll mutex, preventing negative refcount and use-after-free (#1) - Add mutex to ProxyOpInfo, protect step timing accumulation from concurrent proxy thread writes (#2) - Lock ProxyOp before memcpy into coll->proxyOps to prevent reading partially-written step data (#3) - Move all pools (coll, proxyOp, proxyStep) from process-global statics into per-communicator acclCommContext, eliminating cross-comm pool starvation (#4) - Fix fmt_size to show fractional values (1.5M instead of 1M) (#6) - Fix bottleneck classifier denominator to use launch + max(kernel, proxy) instead of exec + launch, preventing percentages exceeding 100% due to kernel/proxy overlap (#7) - Add --output flag to 'single' subcommand and clean up hasattr guard (#8) - Switch output file from append to write mode to avoid stale data from PID reuse (#9) - Replace vendored nccl/profiler.h with accl_shim.h that provides minimal types and includes RCCL's real profiler_v5.h; add CMakeLists.txt for in-tree builds (#10)
pvallem
added a commit
that referenced
this pull request
Aug 25, 2026
Defaulting --jobs to 4 broke the Jenkins precheckin Test stage. That pipeline
invokes the runner as
test_runner.py -c configs/ci-precheckin.json --skip-mpi-check \
--rerun-failed --stop-on-rerun-failure --verbose
with no --jobs, so validate_arg_combinations rejected --rerun-failed against the
new default and exited 1. Build #8 failed in 833 ms / 1405 ms / 2167 ms across
the three arch legs -- an argument-validation exit, not a test failure -- where
build #7 ran the identical command and passed in 428-1880 s.
Reverting rather than relaxing the guard, because on the evidence the default
bought nothing. The only current consumers of this runner are that precheckin
job and .github/workflows/ainic_test_runner.yml; CoCo has migrated off it (see
its own scripts/convert_test_runner_config.py, which reads the "legacy
test_runner/configs/*.json format"). ainic.json is now entirely serial_only, and
62 of ci-precheckin's 66 tests are single-process gtests that enumerate all 8
GPUs, so each fills the co-tenancy budget and runs alone regardless -- only 4 of
its tests could ever share the node. Concurrency stays available behind an
explicit --jobs N for anyone running a full functional config directly, which is
where the measured 13.7% (1 node) / 15.9% (2 nodes) actually came from.
Keeps two things from the reverted commit that stand on their own: -j/--jobs and
--max-parallel-gpus are now listed in the README options table, which they were
not before.
JIRA ID: AICOMRCCL-2053
prasanna-amd
added a commit
that referenced
this pull request
Aug 26, 2026
…ntegration Add monotonic generation counter per pool slot to detect stale callbacks from reclaimed collectives, resolving the 0==0 race in pool reclaim and all callback-lifetime concerns (threads #4, #6, #8, #9). Add atomic ref-count to acclCommContext so acclPluginFinalize defers free() until all outstanding callbacks have drained (thread #2). Wire accl plugin into the parent RCCL CMake build (thread #20), add CXX_VISIBILITY_PRESET/VISIBILITY_INLINES_HIDDEN (thread #1), and bump ACCL_MAX_CHANNELS from 64 to 256 (thread #11). Fix acclShouldFinalize: nChannels==0 collectives (onerank path) finalize immediately on coll stop; guard channelId against nChannels (thread #10, #12). Free leaked ProxyOp when coll's proxyOpIndices is full (thread #13). Rename firstCpuStart to firstKernelCpuStartUs for clarity (thread #14). Fix warmup off-by-one (sn <= warmup → sn < warmup, thread #7), use try/finally for output file handle (thread #15), fix test fixtures to start at coll_sn=0 (thread #16), remove stale ACCL_PROFILER_WARMUP_ITERS docs (thread #17), sync bottleneck categories with classify() (thread #18), add NVIDIA Apache-2.0 header to accl_shim.h (thread #19). Add tests: ProxyStep lifecycle, channel bounds, kernel timing assertions (threads #3, #12, #21).
mch
added a commit
that referenced
this pull request
Aug 26, 2026
Add Test #6 -- when credit is available for every op in a PutSignalGroup, ncclRmaProxyPollNonPersistDesc issues all of them in one pass, moves the group descriptor to the in-progress queue, and advances the consumer index once. Adds the PushPendingPutGroup fixture helper the group-path tests (#7/#8) build on. Co-authored-by: Pi Agent
bwelton
added a commit
that referenced
this pull request
Aug 27, 2026
…blished pipe first and is re-probed under the backstop
1 task done
bwelton
added a commit
that referenced
this pull request
Aug 31, 2026
…blished pipe first and is re-probed under the backstop
bwelton
added a commit
that referenced
this pull request
Aug 31, 2026
…blished pipe first and is re-probed under the backstop
bwelton
added a commit
that referenced
this pull request
Sep 1, 2026
…blished pipe first and is re-probed under the backstop
bwelton
added a commit
that referenced
this pull request
Sep 1, 2026
…blished pipe first and is re-probed under the backstop
venkat1361
pushed a commit
that referenced
this pull request
Sep 1, 2026
…10809) ## Summary Addresses a 15-point static re-review of the KFD dispatch-log signal-less path (on top of `review-fixes`). Each finding was evaluated against the code and the prior design history; this PR **fixes the genuine correctness/safety/resource issues** (potential misattribution, UAF on terminal teardown, unbounded pairer-map growth, plus FATAL-handling, env-race, and robustness items) and **documents the ones intentionally left as accepted residuals or pre-existing**, with rationale. `+365 / −49` across 9 files. ## Fixed - **#4 (misattribution) — validate identity before claiming an unknown-START EOP.** The unknown-START path in `dispatch_hub.hpp::record_kernel_end` took the sole first-owner entry with no timestamp check, and terminal-time validation ran only *after* the entry was consumed/retired — so a stale EOP could claim a newer dispatch's entry. Now rejects an EOP whose `end_ticks <= window.t_open` (a kernel cannot end at/before its window opened; same GPU-tick domain the known-START path already uses) **before** taking the entry, so a rejected EOP consumes/retires nothing. Fail-first verified. - **#3 + #2 (UAF) — the terminal fences are now truly terminal (fail-closed).** `signal_less_fence_completions()` previously quiesced visible work but did not stop the reader on success, so a still-PENDING kernel could complete after the client's callback/buffer state (detach/finalize, F23) or the HSA runtime (final `hsa_shut_down`, F24) was freed. A new `signal_less_terminal_stop()` (factored from teardown) latches disable → stops+joins the reader/processor → drains+ledgers PENDING, and is called from the F23/F24 sites. Verified deadlock-free (never joins its own task group, holds no lock across the join, joined threads never block on the joiner) and 5/5 clean under a shutdown watchdog. The existing `get_fini_status()` re-entry guard (the earlier crash fix) is preserved. - **#11 (unbounded growth) — hard size cap on the pairer START map.** `pair_state::pending_starts` was bounded only by a 30-minute age timer. Added `ROCPROFILER_KFD_DISPATCH_LOG_MAX_PENDING_STARTS` (default 2M, shared reject-and-default parser) with oldest-first eviction (recurring key exempted so the `ambiguous` latch is preserved). - **#7 (coverage) + #10 (liveness) — FATAL handling.** FATAL now latches admission and waits for the processor to consume the already-published pipe **before** draining the hub (so a fatal session's final EOP and healthy-GPU EOPs behind it aren't dropped), and the FATAL status is re-probed while the empty-wake backstop is tripped (so a fatal stream that produces no records is still detected). - **#5 (robustness) — master switch is a strict fail-closed opt-in.** `ROCPROFILER_KFD_DISPATCH_LOG_SIGNAL_LESS` now accepts only explicit true values; anything else is disabled with a one-time warning (no throw on malformed input). - **#9 (arg-order) — deterministic GC-gate operand ordering.** `gc_gate_open`'s two operands are read into named locals in the intended order rather than relying on unspecified argument-evaluation order. - **#12 (race) — env knobs primed at serialized init.** The age/cap/quiesce/switch env reads are populated during reader start-up rather than lazily on a worker thread. - **#14(b) — fork child no longer hangs in `interposition_fini()`.** A normal (non-exec) child now short-circuits `interposition_fini()` under the existing `fork_stale()` guard, avoiding a wlock on a registry a vanished parent thread may hold. ## Intentionally not changed (accepted residuals / design decisions / pre-existing) - **#1 — teardown coverage loss** (a kernel in flight at exit may emit no record): the accepted D8/F1 coverage residual. The *UAF* variant of this concern is #3, which is fixed. - **#8 — equal-tick START/EOP**: the accepted D1 residual. The code emits a **no-timing / zero-duration drop**, not a fabricated 1 ns; no misattribution. - **#9-main — GC gate "race"**: the copied-but-still-in-ring window is the accepted **R10** residual (an EOP still in the ring is intentionally not gated), and cap-eviction's bypass of the in-flight gate is D9's documented coverage-loss residual. The operand-ordering nit is fixed (#9-arg above); the residual is by design (coverage loss, never misattribution). - **#13 — cross-batch inversion**: the explicitly-accepted D1 residual (coverage loss under queue churn, never misattribution). A cross-batch reorder buffer is a larger design change weighed against measured coverage impact, which the benchmarks do not show hitting. - **#6 — completion-callback self-deadlock**: a **pre-existing** reentrancy defect on the generic signal path (a client callback that calls its own finalize). The new terminal-stop path was specifically designed **not** to add another instance of it. - **#14(a) — session mid-construction skipped by a concurrent fork**: a child-lifetime resource leak (not a hang); the parent is unaffected and the child drops it at exit. The "fix" would introduce a formal data race on an `smap` pointer a vanished thread was mid-write on, to reclaim memory about to be freed — not worth it. Left as-is. - **#15 — eager arming on a registered (vs active) context**: an efficiency consideration, not a correctness issue; arming stays gated on a registered kernel-dispatch context. ## Known residuals in the fixes (documented) - #4 adds no upper time bound (an implausibly-future EOP is still taken and rejected by the finalizer as no-timing, never misattributed). - #10 restores FATAL *detection* under the backstop; a *persistently failing* status ioctl still isn't fail-closed (separate follow-up). - #11 eviction is O(live) per over-cap insert (same tradeoff as the D9 cap; only reached under pathological EOP loss). ## Testing (MI350, host ROCm 7.2.1, build `fd06e1407c`) - Build clean at the project warning level — **0 new warnings**. - `ctest -L kfd-dlog` — **90/90 pass** (incl. 2 new tests: pairer-map cap, strict opt-in). - ThreadSanitizer, all 8 kfd-dlog suites — **90/90, zero reports** (covers the terminal-stop join, FATAL drain-wait, GC-gate ordering). - Shutdown crash/hang reproducer (1040), **5 runs under a watchdog — 5/5 clean, no hang, no crash** (the decisive check for the terminal-stop rework). - 1041-d (20k kernels ×3), 1031, 1041-a, 1041-c — all pass, no attribution regression. - graph-perf — **−44.21%** overhead reduction, `register-refused=0`/block (no throughput regression). ## Tracking JIRA ID: AIPROFSDK-1012 AIPROFSDK-1040 AIPROFSDK-1041
3 of 11 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Changed runners to ubuntu-latest instead of AMD-ROCm-Internal-dev1