feat(rocpd): add AI-powered GPU trace analysis module (rocpd analyze)#4030

Open

ammarwa wants to merge 161 commits intodevelopfrom

aelwazir/rocpd-ai-analysis

Commits on Mar 13, 2026

Adding ROCpd Analysis
ammarwa
committed
Fixing ROCPD Analysis Tests
ammarwa
committed
Adding JSON Scheme
ammarwa
committed
Fix format→output_format mapping and GPU-less build failures

ammarwa
and
claude
committed
Add webview output format and fix file extension handling

ammarwa
and
claude
committed
docs(ai_analysis): document webview format and auto file extensions

ammarwa
and
claude
committed
feat(webview): add rich hover tooltips to all graphs and counters

ammarwa
and
claude
committed
docs(ai_analysis): document hover tooltips added to webview report

ammarwa
and
claude
committed
feat(webview): redesign webview UI/UX with AMD dashboard styling

ammarwa
and
claude
committed
fix(webview): use Unicode char instead of HTML entity in CSS content

ammarwa
and
claude
committed
docs(ai_analysis): add v0.1.4 changelog entry for CSS content fix

ammarwa
and
claude
committed
fix(webview): pin tooltip text color so it is readable in light theme

ammarwa
and
claude
committed
docs(ai_analysis): add v0.1.5 changelog entry for tooltip color fix

ammarwa
and
claude
committed
fix(recommendations): suppress redundant re-collection commands

ammarwa
and
claude
committed
fix(recommendations): also drop rocprof-sys --trace when sys-trace present

ammarwa
and
claude
committed
docs(llm-fence): add context-aware profiling recommendation rules

ammarwa
and
claude
committed
docs(llm-fence): replace explicit flag lists with reasoning directive

ammarwa
and
claude
committed
chore: add .gitignore entries for Claude session data and Python cache

ammarwa
and
claude
committed
fix(ai_analysis): address all audit issues + add --llm-model CLI flag

ammarwa
and
claude
committed
fix(ai_analysis): pass path as list to RocpdImportData

ammarwa
and
claude
committed
test(ai_analysis): register unit tests with CTest mechanism

ammarwa
and
claude
committed
fix(llm): handle max_completion_tokens for newer OpenAI models; clarify output format

ammarwa
and
claude
committed
fix(analyze): suppress recommendations for already-collected PMC counters

ammarwa
and
claude
committed
docs: update docs for PMC deduplication, OpenAI compat, CTest, format hint

ammarwa
and
claude
committed
docs(llm): enrich reference guide with AMD profiling workflow and counter knowledge

ammarwa
and
claude
committed
feat: add SessionStore + SessionData for interactive session persistence

ammarwa
and
claude
committed
fix: improve SessionStore robustness + add missing tests

ammarwa
and
claude
committed
feat(interactive): add InteractiveSession skeleton with main menu and resume prompt

ammarwa
and
claude
committed
fix(interactive): handle EOFError in run(), feedback on bad choices, remove duplicate import

ammarwa
and
claude
committed
feat(interactive): implement _path_profiling with helper methods (Task 3)

ammarwa
and
claude
committed
feat(llm): add local LLM (Ollama) support to LLMAnalyzer (Task 4)

ammarwa
and
claude
committed
fix(interactive,llm): apply code quality fixes for stdout pollution, import ordering, and error handling

ammarwa
and
claude
committed
feat(interactive): implement _path_optimize with two-stage LLM pipeline (Task 5)

ammarwa
and
claude
committed
fix(interactive): apply 6 code quality fixes (atomic write, shell safety, UX, dedup)

ammarwa
and
claude
committed
feat(cli): wire InteractiveSession to CLI — Task 7

ammarwa
and
claude
committed
feat: interactive session smoke test verified

ammarwa
and
claude
committed
fix(interactive): wire code_change_analysis source tag and deduplicate mock import

ammarwa
and
claude
committed
feat(llm): add AnalysisContext dataclass and _select_tags function

ammarwa
and
claude
committed
fix(llm): move relative import to top, add return type to _select_tags

ammarwa
and
claude
committed
feat(llm): add _filter_guide and wire context into _build_system_prompt

ammarwa
and
claude
committed
fix(llm): move _TAG_RE to module level, guard verbose div-by-zero

ammarwa
and
claude
committed
feat(llm): propagate AnalysisContext through analyze_with_llm methods

ammarwa
and
claude
committed
fix(llm): remove unnecessary forward-reference quotes from AnalysisContext annotations
ammarwa
committed
feat(llm): export AnalysisContext from rocpd.ai_analysis public API

ammarwa
and
claude
committed
feat(llm): build and pass AnalysisContext from analyze.py LLM call sites

ammarwa
and
claude
committed
feat(llm): tag all reference guide sections with rocpd-context markers

ammarwa
and
claude
committed
docs(llm): add LLM_GUIDE_SECTIONS.md and AnalysisContext API docs

ammarwa
and
claude
committed
test(cmake): register guide filter unit tests with ctest

ammarwa
and
claude
committed
fix(security): shell quoting, path traversal, XSS, exception safety, overflow clamping

ammarwa
and
claude
committed
fix(security): fix C-2 NameError risk in Tier0 LLM finally block

ammarwa
and
claude
committed
fix(security): clamp overhead_time_ms in text formatters, document args dict quoting

ammarwa
and
claude
committed
fix(analysis): correct bottleneck classification, pass AnalysisContext in api, fix total_bytes

ammarwa
and
claude
committed
fix(analysis): add reason field to hardware counters exception path

ammarwa
and
claude
committed
fix(llm): defer key validation, honor self.model, fix exception arg type, add timeouts

ammarwa
and
claude
committed
fix(quality): dedup conversion helper, version constant, lazy import, re/os module fixes, LIMIT clause, to_json RuntimeError

ammarwa
and
claude
committed
test+docs(fixes): add tests and docs for all code review bug fixes

ammarwa
and
claude
committed
fix(review): address final code review findings

ammarwa
and
claude
committed
feat(llm): add --llm-thinking flag for extended thinking support

ammarwa
and
claude
committed
docs(llm-reference): fix critical errors, add PC sampling section, strengthen guardrails

ammarwa
and
claude
committed
fix(interactive): _select_hot_files fails when tier0 is SourceAnalysisResult

ammarwa
and
claude
committed
fix(interactive): bracket keys hidden by Rich markup, Enter not accepted as new session

ammarwa
and
claude
committed
fix(interactive): duplicate hot-files, no-LLM silent exit, show relative paths

ammarwa
and
claude
committed
fix(interactive): run commands directly, fix LLM format, fix dead code

ammarwa
and
claude
committed
feat: add WorkflowState dataclass for 7-phase workflow

ammarwa
and
claude
committed
feat: implement WorkflowSession 7-phase interactive workflow; add rocpd analyze --interactive

ammarwa
and
claude
committed
fix: WorkflowSession Phase 4 forwards llm_provider/llm_api_key to analyze_database()
ammarwa
committed
feat(tracelens): add tracelens_port.py with interval math and kernel categorization

ammarwa
and
claude
committed
test(tracelens): add _subtract_intervals unit tests; fix docstring

ammarwa
and
claude
committed
test(tracelens): add 13 DB-dependent unit tests + integration test for tracelens_port.py

ammarwa
and
claude
committed
feat(tracelens): wire compute_interval_timeline, analyze_kernels_by_category, analyze_short_kernels into analyze_performance() + 2 new recommendation rules

ammarwa
and
claude
committed
feat(tracelens): add kernel category + short kernel sections to all 4 output formats

ammarwa
and
claude
committed
feat(tracelens): add kernel_categories, short_kernels, interval_timeline to AnalysisResult and LLM format

ammarwa
and
claude
committed
feat(tracelens): wire TraceLens data into AnalysisContext, LLM prompt (_format_data_for_llm), and _select_tags()

ammarwa
and
claude
committed
docs(tracelens): add tracelens_metrics section to llm-reference-guide + bump schema to v0.3.0

ammarwa
and
claude
committed
fix(tracelens): pass TraceLens keys through _sanitize_data(); fix test; add Launch Efficiency to _CATEGORY_IDS; update docstring

ammarwa
and
claude
committed
feat: add SessionContext dataclass and SessionData.context field

ammarwa
and
claude
committed
test: fix misleading comment in backward-compat test
ammarwa
committed
refactor: _run_tier1_analysis returns (recs, breakdown) tuple

ammarwa
and
claude
committed
feat: add context update methods and wire at analysis/command call sites

ammarwa
and
claude
committed
feat: add _format_context_block and inject into LLM call sites

ammarwa
and
claude
committed
feat: add _extract_ai_commands and _offer_run_ai_commands

ammarwa
and
claude
committed
feat: persist SessionContext on save; restore on session resume

ammarwa
and
claude
committed
fix: update test_interactive mocks for tuple return; add integration test

ammarwa
and
claude
committed
docs: update README and CLAUDE.md for interactive session context feature
ammarwa
committed
fix(interactive): correct re-profiling UX — commands, menu, and default flags

ammarwa
and
claude
committed
fix(interactive): strip --kernel-names flag and retry LLM rewrite on failure

ammarwa
and
claude
committed
fix(interactive): strip hallucinated rocprofv3 flags from AI-recommended commands

ammarwa
and
claude
committed
fix: replace invalid --hip-api-trace with --hip-trace in recommendations

ammarwa
and
claude
committed
fix(llm): increase timeout to 300s for file rewrite calls

ammarwa
and
claude
committed
docs: add LLMConversation persistent streaming session spec

ammarwa
and
claude
committed
docs: add LLMConversation implementation plan

ammarwa
and
claude
committed
feat: add LLMConversation streaming persistent session class

ammarwa
and
claude
committed
feat: expose load_reference_guide() and LLMConversation in ai_analysis public API

ammarwa
and
claude
committed
feat: replace SessionContext with LLMConversation in InteractiveSession

ammarwa
and
claude
committed
feat: add --llm-compact-every CLI flag for LLMConversation compaction interval
ammarwa
committed
feat: add integration tests and source-dedup for LLMConversation

ammarwa
and
claude
committed
feat: add private/enterprise OpenAI-compatible LLM server support

ammarwa
and
claude
committed
feat: add SSL verification control for private LLM provider

ammarwa
and
claude
committed
fix: accept single-quoted Python dict literals in ROCPD_LLM_PRIVATE_HEADERS

ammarwa
and
claude
committed
docs(fence): add vLLM ROCm pitfalls — pin_memory is not an LLM() param

ammarwa
and
claude
committed
fix(interactive): route 'private' provider to _call_private() for file rewrites

ammarwa
and
claude
committed
docs: update README and API docs for private LLM provider and LLMConversation

ammarwa
and
claude
committed
fix(llm): address 4 code-review issues in private provider and session persistence

ammarwa
and
claude
committed
feat(interactive): cycle-break + go-deeper + AI-edit revert

ammarwa
and
claude
committed
fix(llm+interactive): ban __builtin_amdgcn_sin/cos; teach session from revert

ammarwa
and
claude
committed
fix(interactive): code-review fixes + docs update for cycle-break/revert

ammarwa
and
claude
committed
feat(interactive): Phase 1b quick workload analysis for starter command

ammarwa
and
claude
committed
fix(interactive): add session persistence for WorkflowSession + Ctrl+C save

ammarwa
and
claude
committed
fix(interactive): real LLM failure analysis on revert instead of fake injection

ammarwa
and
claude
committed
feat(interactive): what-next menu after revert + retry loop in Phase 6

ammarwa
and
claude
committed
fix(interactive): code-review fixes — env-strip, allow_retry, ask-before-LLM, "r" alias

ammarwa
and
claude
committed
build(python): add tracelens_port.py, MAKE_DIRECTORY fix, logo asset + PNG glob

ammarwa
and
claude
committed
feat(interactive): multi-process profiling with --process-sync, %nid% output + DB merge

ammarwa
and
claude
committed
chore: remove CLAUDE.md and superpowers plan/spec docs from branch tracking

ammarwa
and
claude
committed
fix(build): use file(COPY) instead of configure_file for AI analysis assets

ammarwa
and
claude
committed
style(cmake): fix formatting in utilities.cmake

ammarwa
and
claude
committed
fix(lint): resolve all flake8 F401/F541/F841/E225/E231 issues across Python modules

ammarwa
and
claude
committed
fix(lint): black formatting + reviewer code fixes

ammarwa
and
claude
committed
fix(test): use urlparse to validate \$schema domain exactly

ammarwa
and
claude
committed
fix(lint): apply cmake-format, black, and flake8 fixes to remaining files

ammarwa
and
claude
committed
fix(lint): reformat with black --target-version py310 to match CI

ammarwa
and
claude
committed
fix(lint): reformat with Python 3.10 black and remove hardcoded paths

ammarwa
and
claude
committed
fix(lint): apply exact black formatting from CI (Python 3.10 clean env)

ammarwa
and
claude
committed
fix(review): address new Copilot PR comments

ammarwa
and
claude
committed
fix(rocpd): address PR review comments on LLMConversation and schema

ammarwa
and
claude
committed
fix(rocpd): address code review findings from proactive review

ammarwa
and
claude
committed
fix(rocpd): address remaining PR inline review comments

ammarwa
and
claude
committed
style(rocpd): cmake-format utilities.cmake

ammarwa
and
claude
committed
fix(schema): add Tier 0 compatibility and schema validation tests

ammarwa
and
claude
committed
fix(python36): replace re.Pattern and importlib.resources with py3.6-compat equivalents

ammarwa
and
claude
committed
fix(tests): remove gfx000 placeholder; require GPU for hardware tests

ammarwa
and
claude
committed
fix(llm): validate ROCPD_LLM_PRIVATE_HEADERS is a JSON object

ammarwa
and
claude
committed
perf(llm): accumulate stream chunks in list; join once at end

ammarwa
and
claude
committed
feat(checkpoints): add CheckpointRecord dataclass and CheckpointError

ammarwa
and
claude
committed
fix(tests): move pytest import to top of test_workflow.py
ammarwa
committed
feat(checkpoints): add checkpoint fields to WorkflowState and _EditRecord

ammarwa
and
claude
committed
feat(checkpoints): add GitCheckpointManager class

ammarwa
and
claude
committed
test(checkpoints): add missing GCM tests and fix import placement

ammarwa
and
claude
committed
feat(checkpoints): add _init_checkpoints — git detection and dirty check

ammarwa
and
claude
committed
fix(checkpoints): handle get_head() failure gracefully in _init_checkpoints

ammarwa
and
claude
committed
feat(checkpoints): add _create_checkpoint and hook into Phase 6

ammarwa
and
claude
committed
fix(checkpoints): compute _rel_path once in Phase 6 checkpoint hook
ammarwa
committed
feat(checkpoints): record run index and performance delta per checkpoint

ammarwa
and
claude
committed
fix(checkpoints): split run_index (Phase3) and delta (Phase4) into separate methods

ammarwa
and
claude
committed
feat(checkpoints): add rollback, blacklist, and blacklist-block methods

ammarwa
and
claude
committed
fix(checkpoints): rollback baseline always clears state; blacklist saves session

ammarwa
and
claude
committed
feat(checkpoints): add [b] rollback option to Phase 5 menu

ammarwa
and
claude
committed
fix(checkpoints): cp_id validation, baseline blacklist, _input() usage

ammarwa
and
claude
committed
feat(checkpoints): inject blacklist block into Phase 6 LLM prompt

ammarwa
and
claude
committed
feat(checkpoints): worktree teardown on exit and stale pruning on start

ammarwa
and
claude
committed
fix(checkpoints): guard teardown against exceptions and empty worktree path

ammarwa
and
claude
committed
feat(checkpoints): complete session checkpoint system with git worktrees

ammarwa
and
claude
committed
fix(checkpoints): critical and important quality fixes from code review

ammarwa
and
claude
committed
docs(checkpoints): document session checkpoint system in AI_ANALYSIS_API.md
ammarwa
committed
fix(checkpoints): silently disable checkpoints on dirty tree instead of aborting
ammarwa
committed
fix(checkpoints): remove dirty-tree check — only AI-modified files are committed
ammarwa
committed
refactor(checkpoints): remove unused is_dirty method and update docs
ammarwa
committed
fix(checkpoints): remove dirty-tree check, fix flake8 F841, apply black formatting

ammarwa
and
claude
committed
fix(review): validate ROCPD_LLM_PRIVATE_HEADERS is a dict; fix Py3.6 compat in CMake schema test

ammarwa
and
claude
committed
fix(tests): fix two ctest failures — wrong tool filter and missing known tag

ammarwa
and
claude
committed
fix(cmake): replace configure_file COPYONLY with build-time staging for test scripts

ammarwa
and
claude
committed