Changelog

All notable changes to rocprof studio are recorded here. The format follows Keep a Changelog, and the project aims to follow Semantic Versioning.

Releases are recorded here without git tags. 0.3.0 is the current state of main.

[Unreleased]

[0.3.0] - 2026-10-14

Changed

• MCP queries select typed rows before JSON serialization, preserving row shapes, serialized-JSON search and nextOffset. Additive effectiveLimit exposes the 1000-row clamp; zero remains count-only. Explicit guards bound rows to 4 MiB and selected arrays/unpaged responses to 16 MiB.
• Private worker transport v2 carries counted dependency projections. Kernel aggregation streams name/start/end records through the GUI's shared incremental reducer, preserving exact arithmetic, interval unions and stable ordering; paging applies after aggregation.
• Timeline analysis projects the requested range, retains bounded windows and streams HTML with glyph, overlay and output budgets. Inline output permits 16 MiB / 100,000 glyphs; new-file output permits 512 MiB / two million glyphs. maxOutputBytes can lower the byte ceiling. Responses add clock, coverage and effective-budget context.

Fixed

• Timeline points at lower and split boundaries appear exactly once. Output publication validates the complete document and commits atomically without replacing an existing path; late cancellation preserves a completed publication.

Limits and compatibility

• Import memory is unchanged. Exact kernel aggregation still scans all records, O(N), with bounded group/name/interval state. Other analyses still materialize relevant projections under a 256 MiB encoded-input cap; diagnostics retains full-capture compatibility. This release does not make all analysis out-of-core or establish an RSS ceiling. Import overrides, the 1024 MiB Node heap and the 120-second deadline are unchanged.

[0.2.1] - 2026-10-02

A review-hardening release that applies the findings of an adversarial code review of the 0.2.0 remediation arc. There are no new features; the changes below make the live-profiling, import and tuning paths safer and more honest.

Fixed

• The desktop GUI no longer freezes for up to 180 seconds while a live profile is stopped or triggered. Status polls never block on the live-profile lock, and the detach wait runs without holding it.
• Live-profile detach escalates instead of killing outright: it writes a newline, then sends SIGINT after about 30 seconds, then SIGKILL only past 180 seconds. The previous SIGKILL-only path could truncate rocprofv3's synchronized output files and leave a surviving target in an unknown attach state.
• A failed attach-window launch no longer poisons the session with a "File exists" error; its directory and the collection it would have committed are rolled back. The rocprofv3 attach-argument cache under /tmp (rocprofv3_attach_<pid>.pkl) is removed when the target stops and again at shutdown, so a reused PID does not inherit a stale attach spec. Ending an already-exited timed window reports that it detached instead of a misleading error, and ending a window rejects stray collection options.
• The IRPF importer honors the producer's truncation contract. A truncated dispatch-events chunk is no longer silently relabelled as queue mode; truncated or dropped records and a failed session produce loud warnings; and a dynamic (all-zero) workgroup size renders as "Not recorded" instead of a fabricated grid.
• A Loom compile report within the per-file size cap can no longer exhaust memory before the record-cap check: the importer rejects an over-cap collection before allocating for it. The Loom importer is also cancellable now, and validates wait-reason inventories like the strict version-zero loader.
• ROCPROF_MAX_FILE_MIB now applies to .rgp imports. It was hard-coded to 256 MiB, so the variable was silently ineffective for RADV captures.
• kernelIterationRange rejects degenerate values such as -, , and 0-3,, that passed the old character-set check, and a blank or degenerate kernel name no longer crashes the tuning worker.
• The tuning and shader workers no longer hang on an uninterruptible descendant after a kill: a timed-out process tree is reaped with a bounded post-kill wait.
• Untracked bytecode and local scratch that had been swept into earlier commits are removed from version control.

Changed

• "Stop and collect" is cancellable from the desktop, through a shared cancellation flag on the live-profile control.
• The headless CLI uses the same cancellable, input-fingerprinted import router as the GUI and MCP, and accepts .rocprof-att manifests. The earlier CLI/desktop import divergence is gone.
• The desktop Live profiling panel validates the window duration against the 1 to 600000 ms bound the MCP schema enforces, exposes the ATT sub-options (targetCu, simdSelect, bufferSize, shaderEngineMask) and the kernel iteration range, and reports errors in the panel's own vocabulary rather than raw MCP tool names.
• Wave-budget and row-cap rejection messages interpolate their effective constants, so the message a user sees cannot drift from the enforced bound.
• The rocprofv3 option vocabulary lives in one key table, and a contract test fails on a one-sided edit instead of letting a rocprofv3-only option be silently accepted on a RADV run.

Security

• The tuning and shader workers no longer inherit LD_PRELOAD, HSA_TOOLS_LIB or HSA_TOOLS_REPORT_LOAD_FAILURE from the launching environment, and reject them if passed back in env, so a profiler library the app itself was launched under cannot perturb or instrument the kernel being measured.

Upgrade notes

• An installed binary, including one run as a systemd service, must be rebuilt and reinstalled to pick up 0.2.1.
• The headless CLI now computes the same input fingerprint as the desktop and MCP. A capture opened through the CLI participates in the same input-change detection: a file that changes during import is rejected, and the capture-archive overwrite guard refuses to overwrite an archive whose source file has since changed.

[0.2.0] - 2026-10-02

A remediation release driven by user feedback, grouped by theme: live-decode profiling, input caps, HRX/loom support, the source flamegraph, and kerncap parity.

Added

• Live-decode profiling: a desktop Live profiling sidebar panel for launching an application and collecting gated windows without a human at the MCP keyboard. It drives the same core as the MCP tools through four live_profile_start/trigger/status/stop Tauri commands.
• Kernel filters on profile_run and profile_trigger: kernelIncludeRegex and kernelExcludeRegex, mapping to rocprofv3's --kernel-include-regex and --kernel-exclude-regex, so a window can trace one expensive dispatch instead of the first small kernels that share its name.
• kernelIterationRange on profile_run and profile_trigger, mapping to rocprofv3's --kernel-iteration-range (a value like 64 or 0-3). It is the only way to reach a dispatch other than the first occurrence of a kernel name, which a default per-dispatch ATT trace samples once.
• Triggered profiling windows: profile_trigger with triggered: true and no windowMs stays attached until you end it with {"end": true} (which writes a newline to rocprofv3's stdin so it detaches) or profile_stop, so you can capture exactly one request without guessing a duration. profile_status reports windowTriggered for such a window.
• Environment-configurable import and analysis caps: ROCPROF_MAX_RECORDS (default 2,000,000), ROCPROF_MAX_FILE_MIB (256), ROCPROF_MAX_HIDDEN_TOKENS (4,000,000), ROCPROF_MAX_ATT_TRACES (256) and ROCPROF_MAX_ATT_CODE_OBJECTS (256), resolved in the new crates/profile-core/src/caps.rs and applied across desktop, headless CLI and MCP.
• HRX/loom support: a new IRPF v2.0 importer (crates/profile-core/src/irpf.rs) for HRX native profile bundles produced with HRX_PROFILE_MODE=queue or dispatch. It reconstructs resolved dispatch and function names, grids, code-object identity and clock correlations. The .irpf extension is wired into the import dispatcher and the desktop file picker, and bundles are also recognized by content when the filename differs.
• Loom compile-report import: a new importer (crates/profile-core/src/loom_report.rs) for Loom's version-zero compile reports from loom-compile (*.compile-report.json) and from an iree-benchmark-loom benchmark envelope. It mirrors the strict version-zero loader (loom/py/loom/reporting/compile_report.py), rejecting any other schema version, an unknown mode, a non-object or failed status, entry count or index mismatches, duplicate entry identity and the residency-constraint invariant violations, and it refuses the pre-schema public sample. Accepted reports are surfaced as static compile facts (identity, static and dynamic instruction mix, target resources, allocation and spill facts, wait plan, workload geometry, residency and constraints, move causes, detail rows) on a compileReports capture surface kept separate from dispatches and measured data. Every field is optional and absent stays absent; the importer does not infer cycles, dynamic counts, hardware register counts, occupancy as throughput or bank-conflict freedom. Provenance is read out of band from tool-provenance.json and commands.json sidecars. The report is recognized by the .compile-report.json suffix or a loom.compile_report content sniff. The imported facts are queryable over MCP with capture_query and dataset compileReports, which returns one labelled row per report (unit static, version-zero/same-compiler-horizon) with report- and per-entry identity, detail flags, target resources (SGPR/VGPR, subgroup geometry, static occupancy percent, limiting resource, residency), instruction mix, allocation/spill and emitted-instruction counters, wait plan and workload geometry, never presented as measurements. A schema-mismatched or pre-schema report is rejected with regeneration guidance under both filenames — naming the expected kind/schema_version 0 and the regenerate-with-this-compiler instruction — while unrelated JSON keeps its existing message. The importer is verified end-to-end against a real iree-benchmark-loom envelope produced on gfx1151 by the provider-enabled tool (pinned hrx-system 32b86d42), vendored with provenance at fixtures/loom/. See HRX profiling.
• HRX kernel-name families: ggml_* names are classified as hrx-ggml, and iree_hal_amdgpu_device_* names as internal runtime kernels (hidden from the Tune button like RADV's internal shaders).
• A shared read-side kernel-name classifier, src/kernelName.ts (classifyKernelName, kernelBase, isInternalKernel, registerKernelFamily), as the single plug-in point for backend families.
• A "Group by" control in the Source flamegraph viewport, grouping by source location (the default) or by code object (a pipeline/kernel label for RADV captures, the ELF code-object id for HIP/ATT captures). The grouping is saved per capture, each grouping keeps its own viewport, and the SVG export caption records which produced it.
• HIP kernel-tuning replay conditions sustained (Kerncap's --no-recopy path) and repeats, with a mode echo-verify: a session refuses a result whose reported mode does not match the request, so an un-rebuilt runtime cannot silently ignore a condition.
• New documentation: docs/hrx-profiling.md, docs/kernel-name-layer.md and docs/kernel-tuning-parity.md, plus a documentation site changelog page.

Changed

• Import and analysis caps are now configurable rather than fixed, but they remain explicit rejection boundaries: defaults are unchanged, nothing is silently sampled, and the rejection message interpolates the effective value. The wave-viewport budgets (128 waves / 500,000 tokens), the 4,096-wave statistics budget and the six-open-capture limit were deliberately left fixed, because those paths are not yet paged.
• The same-collection-per-process rule for rocprofv3 attach windows now also covers kernel filters, iteration range and window mode (timed versus triggered), in addition to streams, counters and ATT options. rocprofv3 pickles each process's attach options and reattaches only with identical ones, so a mismatch is rejected up front.
• profile_trigger now retries a busy-queue attach abort (the ROCm set_profiler_active_on_queue handshake) with backoff while the target is alive, reporting attachAttempts. A persistent failure is surfaced as an attachWarning, with a per-window warning at profile_stop carrying rocprofv3's stderr tail. This is a known ROCm limitation, not a fix for it.
• docs/profiling-caveats.md records that a default per-dispatch ATT trace samples the first occurrence of each kernel name, and that kernelIterationRange is how to target a later dispatch.
• docs/kernel-tuning-parity.md records what was deliberately not done for HIP tuning (--memory and --flush-mb) and why Kerncap's fixed-address restore on a unified-memory APU makes them meaningless there.

Fixed

• rocprofv3 counter-collection CSVs keep their grid and workgroup. The importer now reads whichever columns the header provides: the kernel trace's _X/_Y/_Z triple, or the single scalar Grid_Size/Workgroup_Size column a counter CSV carries. Counter CSVs previously rendered an empty × × because the importer always formatted the triple.

Upgrade notes

• The import and analysis caps are environment-configurable now. No action is required, because the defaults equal the previously shipped values; set a ROCPROF_MAX_* variable before launch only if a large capture is being rejected.
• An installed binary, including one run as a systemd service, must be rebuilt and reinstalled to pick up 0.2.0 and the new behavior.