Repository navigation
Update .gitmodules - #16
Merged
Merged
Conversation
Added path for rocprofiler-systems/external/papi
jayhawk-commits
pushed a commit
that referenced
this pull request
Aug 5, 2025
[ROCm/rocprofiler-register commit: 19fe192]
ammallya
pushed a commit
that referenced
this pull request
Aug 6, 2025
- Fix for proto files not being viewable in Perfetto UI - Ported from ROCm/omnitrace#411 - Update Workflows - Use V47 trace_processor_shell for certain OS releases. - RedHat 8, SUSE 15.5, and Ubuntu 20.04 are no longer compatible with the latest trace_processor_shell. - Incompatible version of GLIBC. - Remove notes about Perfetto workaround in documentation. --------- Signed-off-by: David Galiffi <David.Galiffi@amd.com> [ROCm/rocprofiler-systems commit: 21d9ab7]
ammallya
pushed a commit
that referenced
this pull request
Aug 6, 2025
- Fix for proto files not being viewable in Perfetto UI - Ported from ROCm/omnitrace#411 - Update Workflows - Use V47 trace_processor_shell for certain OS releases. - RedHat 8, SUSE 15.5, and Ubuntu 20.04 are no longer compatible with the latest trace_processor_shell. - Incompatible version of GLIBC. - Remove notes about Perfetto workaround in documentation. --------- Signed-off-by: David Galiffi <David.Galiffi@amd.com> [ROCm/rocprofiler-systems commit: 46ca5ca]
ammallya
pushed a commit
that referenced
this pull request
Aug 7, 2025
* Update kfd ioctl header - Adds new event for dropped events - Mirrors kernel update by Philip Yang * Add error code for page migration events - Adds support for new error code field for page migration end events - Page migration end event is now generated for migration failure - Error code is zero for successful migration * Add dropped event SMI event - New event type indicates if events were dropped - Events are dropped if the buffer is full [ROCm/rocprofiler-sdk commit: fc25138]
ammallya
pushed a commit
that referenced
this pull request
Jan 21, 2026
Create unit tests for simple buffer transfers using coarse-grained memory
ammallya
pushed a commit
that referenced
this pull request
Jan 21, 2026
Fix signal calculation bug for fine-tiled unit tests
ammallya
pushed a commit
that referenced
this pull request
Jan 21, 2026
Create unit tests for simple buffer transfers using coarse-grained memory [ROCm/rocshmem commit: 733cb09]
ammallya
pushed a commit
that referenced
this pull request
Jan 21, 2026
Fix signal calculation bug for fine-tiled unit tests [ROCm/rocshmem commit: 7fd6666]
dayatsin-amd
pushed a commit
that referenced
this pull request
Apr 7, 2026
ammallya
pushed a commit
that referenced
this pull request
Apr 9, 2026
ammallya
pushed a commit
that referenced
this pull request
Apr 9, 2026
[ROCm/hipFile commit: 99742ca]
7 of 9 tasks
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>
EvanPhoukong
added a commit
that referenced
this pull request
Aug 4, 2026
…, 11D) The exploratory lane is only useful if a model cannot escape it, so this attacks the boundary instead of assuming it holds. Two red-team attacks (14 -> 16). #15 tries to promote a proposal into vetted advice from every reachable direction: writing invented techniques into the vetted lane, relabelling status to "recommended", claiming the runtime-owned identity/provenance fields, and exceeding the confidence ceiling to borrow the authority of a measured result. #16 fabricates the evidence underneath a proposal — citing an uncalled tool, an unprofiled kernel, a near-miss kernel name, and no evidence at all. The parity suite needed the same treatment. It mocked the model echoing the air-gap answer back, so it passed whether or not the runtime consulted the model and proved nothing. The model now returns hostile output on every deterministic field, parity is asserted across classification, specialist techniques, and recommendation ranking, and the fixture records invocations so a test fails rather than silently passing if a live session ever falls back to air-gap. Also states the two-lane contract in the bundled AGENTS.md. The orchestrator decides how proposals reach the user, and every guardrail behind it is moot if it presents them as benchmarked advice. Corrects the fence description in agent-hierarchy.md, which still claimed prompts load verbatim with no runtime concatenation — 11A changed that to composition of always.md plus the role slice.
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).
1 task done
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.
Added path for rocprofiler-systems/external/papi