fix(amdsmi): report APU utilization as whole percent by adam360x · Pull Request #13540 · ROCm/rocm-systems

adam360x · 2026-10-15T14:03:00Z

Motivation

On integrated GPUs with a gpu_metrics v2.x table, GPU and multimedia activity read 100 times too high. A GPU at 45% busy read as 4545 in the C API; amd-smi metric --usage showed N/A, or up to 100% when the GPU was nearly idle.

The APU firmware reports activity in hundredths of a percent (the SMU headers mark GfxActivity/UvdActivity [centi]; amdgpu divides by 100 only for gpu_busy_percent), and amd-smi passed the value on as whole percent.

#11075 made these tables readable on v2.1-v2.3 APUs; this follows up on its post-merge review. v3.0 tables already report whole percent. Discrete GPUs are not affected.

Technical Details

Behavior change: v2.x GFX and MM activity are whole percent in amdsmi_gpu_metrics_t and in its APU sub-struct (was hundredths), so amdsmi_get_gpu_activity() follows. The not-applicable value is kept as is.
readDevInfoBinary(): passes on only the errnos that name a device state (suspended or resetting, runtime-inactive, access denied, interrupted, firmware without the metrics command); others report a failed read (EINVAL was reported as a caller error).
The APU metrics pointer is copied to thread-local storage while the device lock is held (a concurrent call could replace it while the caller read it).
amdsmi.h: return statuses of both metrics calls, the activity unit, and the partition call never returning APU metrics.
amd_smi.cc: three log statements rewrapped (formatting only).
Tests: GpuUnit.ApuActivityReachesThePublicFieldsAsWholePercent, NotApplicableApuActivityIsStillReportedAsNotApplicable, DriverRefusedMetricsReadIsNotReportedAsUnsupported; the oversized-header test now catches a missing read-size clamp; the Python APU field test spells out the v2.4 layout.
CHANGELOG: one Resolved Issues entry; the fix(amdsmi): return GPU metrics on APUs with metrics v2.1-v2.3 #11075 entry no longer names status constants.
Not changed here: per-metric query statuses, which test(amdsmi): fix APU cleanup #13412 changes.

Issue Tracking

JIRA ID: ROCM-29548

Test Plan

amdsmitst --gtest_filter='*Unit*' and the Python unit suite; each new test with its fix reverted.
Hawk Point APU (gfx1103, gpu_metrics v2.1) under a GPU load at a measured 45.9% duty. One captured table was bind-mounted over the sysfs node, so develop, the previous push and this branch read the same bytes: C API, Python and amd-smi metric --usage.

Test Result

Unit: 298 passed (18 skipped); Python: 485 passed. Each new test fails with its fix reverted.
Table holding 4545 hundredths (45.45%):

	C API (top / APU)	`GFX_ACTIVITY`	APU GFX row
develop	4545 / 4545	N/A	missing
previous push	45 / 4545	45 %	missing
this branch	45 / 45	45 %	45 %

Nearly idle table (18 hundredths): develop showed 18 % in both rows; this branch shows 0 %.

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

Every gpu_metrics v2.x table carries GFX and MM activity in hundredths of a percent. The SMU headers mark GfxActivity and UvdActivity "[centi]" for each ASIC that publishes one -- smu_v13_0_4, smu_v13_0_5 and yellow_carp for v2.1, renoir and vangogh for v2.2 to v2.4 -- and amdgpu stores the value unscaled, dividing by 100 only on the gpu_busy_percent path. amd-smi copied it straight into the public whole-percent fields, so amdsmi_get_gpu_activity() and amd-smi metric --usage reported a GPU at 45% busy as 4567. Convert on the way to the public fields, leaving the not-applicable sentinel alone so a value the device never answered for does not become 655. The APU sub-struct keeps the units the firmware published, as its power fields already do. A v3.0 table reports whole percent to begin with and is untouched. Two further fixes to the same path: - A per-metric query turned an unreadable metrics revision into unexpected data, contradicting the unsupported status the whole-table query reports for the same device. Keep the reason setup gave. - readDevInfoBinary() passed every errno from a refused read through to the caller, where EINVAL became INVALID_ARGS and blamed the caller for a fault on the driver's side. Pass on only the errnos that name a condition the caller can act on -- EBUSY, EPERM, EACCES, EINTR, and EOPNOTSUPP, which is what the SMU answers when the firmware does not know the metrics command at all -- and report the rest as the read failure they are. The APU metrics pointer handed back with the table pointed into the device's own metrics object, which a concurrent call on the same device could replace while the caller was still reading it. Copy it into thread-local storage while the device lock is still held, which is the lifetime the headers already promise for that pointer.

- The oversized-header case declared 4 KiB more than the layout we model but wrote only a table-sized file, so dropping the read clamp ended in a short read rather than an overrun: the case never reached past the buffer and ASAN saw nothing. The blob builder now always writes at least what the header declares, so the dangerous configuration is no longer reachable by leaving an argument out, and removing the clamp reports a heap-buffer-overflow. - Nothing covered the branch that tells a read the driver refused from a node that is simply empty -- both return zero bytes, and only the stream's error flag separates them. A directory in place of the node reproduces the refusal. The test checks that premise before drawing any conclusion from the status, so a platform where a directory stops behaving that way says so instead of passing for the wrong reason. - Nothing covered the activity unit conversion. One case walks three readings chosen to separate truncation, rounding away from zero, and rounding half to even, and gives the two fields different values so a crossed-over conversion cannot hide. Another checks the not-applicable sentinel is not scaled into a plausible-looking 655; it carries a live reading beside the sentinel, because the destination is pre-filled with that same sentinel and the check would otherwise also pass if the read had populated nothing at all. - The APU field-table test built its expectation with the same expression as the code it tested, so it held for whatever that table happened to contain. Spell the v2.4 layout out instead. - The empty-clock-file skip is gated on the driver's APU bit, which MI300A also carries. That is intentional, since those parts power-gate domains too, so say so rather than leaving a comment that reads as small-APU-only.

- The return documentation for both metrics getters listed only success, an unsupported revision and an unexpected size, while the call can also report a busy device or a permission problem. List what a caller can see. - Both getters carried the same APU-metrics paragraph, but a partition metrics table resolves only the XCP revisions and never carries APU metrics, so the pointer that paragraph describes is always null for that call. Say so, and keep the part that still applies: the call invalidates a pointer an earlier whole-table query handed out on the same thread. - Record the unit the APU sub-struct reports activity in, beside the existing note that its power fields stay in milliwatts. It follows the firmware, so it is hundredths of a percent on a v2.x table and whole percent on v3.0, while the top-level fields are whole percent for every revision. Say that the not-applicable sentinel is never scaled, since callers test for it by value. - A comment described the shorter APU revisions as byte-prefix supersets of the v2.4 struct, which is backwards and contradicted the paragraph stacked under it. They are prefixes of it. Fold the two into one and note that v3.0 is a separate layout rather than a prefix. - Restore three logging lines that were re-wrapped by hand while unrelated to the change that touched them. The pinned clang-format leaves both forms alone, so the churn was never required. - The changelog entry named status constants and described the change as APU-only. Say what a user sees instead, and cover the discrete-GPU PLX change it left out.

therock-pr-bot · 2026-10-15T14:04:09Z

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

Copilot

🟢 Approval recommended

The targeted fixes and strengthened regression tests support approval, with no unresolved blocking findings.

0 open findings

What changed in this PR

Corrects APU utilization reporting in amd-smi, following #11075, while improving metrics error handling and pointer lifetime safety.

Changes:

Converts v2.x activity to whole percent while preserving unavailable-value sentinels.
Preserves meaningful read errors and copies APU metrics into thread-local storage.
Strengthens regression tests and clarifies API documentation.

File	Description
`projects/amdsmi/tests/python/unit/gpu/test_cli_metrics_field_support.py`	Makes expected APU fields independent of production data.
`projects/amdsmi/tests/amd_smi_test/unit/gpu/dynamic_metrics_test.cc`	Tests scaling, sentinels, refused reads, and oversized tables.
`projects/amdsmi/tests/amd_smi_test/functional/gpu/clock/frequencies_read_write_test.cc`	Clarifies the existing APU skip rationale in comments.
`projects/amdsmi/src/amd_smi/amd_smi.cc`	Reverts unrelated logging-format changes.
`projects/amdsmi/rocm_smi/src/rocm_smi_gpu_metrics.cc`	Fixes activity units, status propagation, and APU pointer lifetime.
`projects/amdsmi/rocm_smi/src/rocm_smi_device.cc`	Limits read-error propagation to meaningful statuses.
`projects/amdsmi/include/amd_smi/amdsmi.h`	Documents units, sentinels, return statuses, and partition behavior.
`projects/amdsmi/CHANGELOG.md`	Records the metrics fixes.

🧠 Review effort: Balanced

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

The APU sub-struct kept v2.x GFX and MM activity in hundredths of a percent, while v3.0 reports whole percent and the Python interface and the CLI print both with a % sign. On a v2.1 APU more than 1% busy, the APU activity rows were dropped as N/A; below that they read up to 100 times too high. Convert them the same way as the top-level fields. Drop the per-metric query status change: another open pull request changes the same lines in a different way, so that change is left to it. Signed-off-by: Adam360x <Adam.pryor@amd.com>

- Leave v2.3/v2.4 GFX/MM activity as reported: only Van Gogh publishes them, its current firmware interface already reports whole percent, and dividing turned 48% GFX busy into 0% - Fold the 100x CHANGELOG entry into the integrated GPU metrics fix, since no release reported those APUs' activity before Old Van Gogh firmware can still pair v2.2 with whole percent or v2.3 with hundredths; the table carries nothing else to tell them apart. Signed-off-by: Maisam Arif <Maisam.Arif@amd.com>

- Divide v3.0 average_vcn_activity by 100 with the rounding helper and keep N/A: both v3.0 kernel producers report VCN in hundredths and divide it by 100 for vcn_busy_percent - Document the field as whole percent in amdsmi.h - Add a Resolved Issues entry, since v3.0 APU metrics shipped in 7.13.0 Signed-off-by: Maisam Arif <Maisam.Arif@amd.com>

## Motivation On integrated GPUs with a `gpu_metrics` v2.x table, GPU and multimedia activity read 100 times too high. A GPU at 45% busy read as 4545 in the C API; `amd-smi metric --usage` showed N/A, or up to 100% when the GPU was nearly idle. The APU firmware reports activity in hundredths of a percent (the SMU headers mark `GfxActivity`/`UvdActivity` `[centi]`; amdgpu divides by 100 only for `gpu_busy_percent`), and amd-smi passed the value on as whole percent. #11075 made these tables readable on v2.1-v2.3 APUs; this follows up on its post-merge review. v3.0 tables already report whole percent. Discrete GPUs are not affected. ## Technical Details - **Behavior change:** v2.x GFX and MM activity are whole percent in `amdsmi_gpu_metrics_t` and in its APU sub-struct (was hundredths), so `amdsmi_get_gpu_activity()` follows. The not-applicable value is kept as is. - `readDevInfoBinary()`: passes on only the errnos that name a device state (suspended or resetting, runtime-inactive, access denied, interrupted, firmware without the metrics command); others report a failed read (`EINVAL` was reported as a caller error). - The APU metrics pointer is copied to thread-local storage while the device lock is held (a concurrent call could replace it while the caller read it). - `amdsmi.h`: return statuses of both metrics calls, the activity unit, and the partition call never returning APU metrics. - `amd_smi.cc`: three log statements rewrapped (formatting only). - Tests: `GpuUnit.ApuActivityReachesThePublicFieldsAsWholePercent`, `NotApplicableApuActivityIsStillReportedAsNotApplicable`, `DriverRefusedMetricsReadIsNotReportedAsUnsupported`; the oversized-header test now catches a missing read-size clamp; the Python APU field test spells out the v2.4 layout. - CHANGELOG: one Resolved Issues entry; the #11075 entry no longer names status constants. - Not changed here: per-metric query statuses, which #13412 changes. ## Issue Tracking JIRA ID: ROCM-29548 ## Test Plan - `amdsmitst --gtest_filter='*Unit*'` and the Python unit suite; each new test with its fix reverted. - Hawk Point APU (gfx1103, `gpu_metrics` v2.1) under a GPU load at a measured 45.9% duty. One captured table was bind-mounted over the sysfs node, so develop, the previous push and this branch read the same bytes: C API, Python and `amd-smi metric --usage`. ## Test Result - Unit: 298 passed (18 skipped); Python: 485 passed. Each new test fails with its fix reverted. - Table holding 4545 hundredths (45.45%): | | C API (top / APU) | `GFX_ACTIVITY` | APU GFX row | |---|---|---|---| | develop | 4545 / 4545 | N/A | missing | | previous push | 45 / 4545 | 45 % | missing | | this branch | 45 / 45 | 45 % | 45 % | - Nearly idle table (18 hundredths): develop showed 18 % in both rows; this branch shows 0 %. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md. --------- Signed-off-by: Adam360x <Adam.pryor@amd.com> Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> Co-authored-by: Maisam Arif <Maisam.Arif@amd.com>

systems-assistant · 2026-10-22T19:13:17Z

TheRock Submodule Bump Activity

Newest first

2026-10-22 — Included in TheRock via ROCm/TheRock#9425.

Callers allocate most amdsmi structs and load the library by name, so a changed layout corrupts memory instead of failing to load. Public types changed size 46 times between ROCm 6.0 and 10.2, 19 of them within one SONAME; #11741 and #11821 are the latest. The ABI Compliance Check only labels such PRs. abi_gate.py parses amdsmi.h with libclang and records every public type's size, alignment and field offsets, enumerator values, function signatures, integer macros and the units, sentinels and directions in field and parameter doc comments. It snapshots a commit, its base and the latest release/therock-* branch from git, compares them, and prints each break as type, field, old and new value. gcc and g++ also compile static asserts of the base's layout against the new header. Additions and fields carved from reserved arrays pass. A break passes only when projects/amdsmi/abi/abi_exceptions.json gains an entry for that exact path, old and new value, and CHANGELOG.md names it, so it is reviewed by the owners of projects/amdsmi/. "abi_gate.py acknowledge" writes the entries. The file also lists append_only structs and notes on meaning the docs cannot express; today it holds the three % units #13540 added after release/therock-10.2. amdsmi-abi-gate.yml runs it on every PR to develop and release/therock-*, with no paths filter so it can become a required check, plus merge groups and pushes to develop. It runs the checker from the base commit, so a PR cannot edit it to pass; the workflow file itself still comes from the PR and needs a "require workflows" ruleset before the check is made required. Replayed on the 73 develop commits that changed amdsmi.h in the last year, it blocks all 19 layout and API breaks and misses none; it also blocks 4 doc-only unit corrections, which pass once acknowledged.

Callers allocate most amdsmi structs and load the library by name, so a changed layout corrupts memory instead of failing to load. Public types changed size 46 times between ROCm 6.0 and 10.2, 19 of them within one SONAME; #11741 and #11821 are the latest. The ABI Compliance Check only labels such PRs. abi_gate.py parses amdsmi.h with libclang and records every public type's size, alignment and field offsets, enumerator values, function signatures, integer macros and the units, sentinels and directions in field and parameter doc comments. It snapshots a commit, its base and the latest release/therock-* branch from git, compares them, and prints each break as type, field, old and new value. gcc and g++ also compile static asserts of the base's layout against the new header. Additions and fields carved from reserved arrays or bitfields pass. The header must also give C++ the same layouts, field types and functions, with C linkage. A break passes only when projects/amdsmi/abi/abi_exceptions.json gains an entry for that exact path, old and new value, and CHANGELOG.md names it, so it is reviewed by the owners of projects/amdsmi/. "abi_gate.py acknowledge" writes the entries. The file also lists append_only structs and notes on meaning the docs cannot express; today it holds the three % units #13540 added after release/therock-10.2. amdsmi-abi-gate.yml runs it on every PR to develop and release/therock-*, with no paths filter so it can become a required check, plus merge groups and pushes to develop. It runs the checker from the base commit, or from develop when the base has none, so a PR cannot edit it to pass; the workflow file itself still comes from the PR and needs a "require workflows" ruleset before the check is made required. Replayed on the 73 develop commits that changed amdsmi.h in the last year, it blocks all 19 layout and API breaks and misses none; it also blocks 4 doc-only unit corrections, which pass once acknowledged.

## Motivation Callers allocate most amdsmi structs and load the library by name, so a changed layout corrupts memory in programs built against the old header instead of failing to load. Recent examples: - #11741 grew `amdsmi_fabric_info_t` from 320 to 552 bytes. - #11821 grew `amdsmi_driver_info_t` from 768 to 1792 bytes. - #10226 grew `amdsmi_enumeration_info_t` from 276 to 280 bytes. All three landed within SONAME 27. #14425 made the existing `ABI Compliance Check` report such changes, but it only labels a PR and cannot block it. This PR also covers the follow-ups from #14425's review. ## Technical Details **ABI Compliance Check (#14425 follow-ups)** - `abi_check.py` exits 1 for a break and 2 when it could not compare. That covers a checker error, a missing merge base, an old header it cannot fetch, or a report without the 2.3 summary table. `abi-compliance-check.yml` records exit 2 as `error`, which leaves the labels unchanged instead of adding a false `MAJOR ABI BREAKAGE`. - It defines `ENABLE_ESMI_LIB` before parsing, so the CPU API (69 functions, 12 types) is checked. - It fails on a changed non-string constant, such as `AMDSMI_MAX_MM_IP_COUNT` going from 8 to 16. - Smaller fixes: - a size row no longer inherits a type name from an earlier section; - it works on Python 3.6; - the patterns point at `ACC_COMMIT`; - new tests cover High severity, symbol problems and Low type problems that are not size changes. **AMDSMI ABI Gate (new, blocking once required)** - `tests/abi_check/abi_gate.py` parses `amdsmi.h` with libclang and records: - each type's size, alignment and field offsets; - enumerator values and function signatures; - integer macros; - the units, sentinel values and in/out directions in field and parameter doc comments. It compares the PR with its base and the newest `release/therock-*` branch, each read from git. Additions pass, and so do fields carved from `reserved*` arrays or bitfields when the struct keeps its size. - An intended break is recorded with `abi_gate.py acknowledge --why "#NNNN: ..."`, which adds one line per change (path, old and new value, reason) to `projects/amdsmi/abi/abi_exceptions.json`, together with a CHANGELOG line naming it. The owners of `projects/amdsmi/` review both. The file starts with the three `%` units #13540 documented on develop that 10.2 does not have. - `.github/workflows/amdsmi-abi-gate.yml` runs on every PR to develop and `release/therock-*`, on merge groups, and on pushes to develop. - It has a read-only token and no secrets. - It runs the checker from the base commit, or from develop when the base has none, such as an older release branch. - There is no `paths:` filter, so it can become a required check. PRs that do not touch amdsmi's headers pass in seconds. **TheRock CI** - `.github/skip-ci-config.toml` skips the TheRock build for `projects/amdsmi/abi/*` and `projects/amdsmi/tests/abi_check/*`, which no build reads. This PR does not make the gate required. That is a ruleset change after an advisory period, together with a "require workflows" rule, because on pull requests GitHub runs the PR's own copy of the workflow file. ## Issue Tracking JIRA ID: TBD Follows up #14425. ## Test Plan - `abi_check_test.py`, `abi_gate_test.py` and the amdsmi pre-commit hooks. - `abi_check.py` with abi-compliance-checker 2.3 on real headers: no change, an enum value change, a header that does not compile, and a missing base ref. - Gate replay on the 73 develop commits that changed `amdsmi.h` in the last 12 months, the `release/therock-*` transitions, and 25 bypass fixtures. - The gate workflow run locally with `act` for same-repo and fork PRs, merge groups and pushes to develop. ## Test Result - 29 and 86 unit tests pass. - `abi_check.py` exits 0, 1, 2 and 2 for the four headers. - The gate blocks all 19 layout and API breaks in the replay and misses none. Among them: #11741, #11821, #13926, #8575, #10226, the `AMDSMI_LINK_TYPE_PCIE`/`XGMI` value swap, and the same-size `amdsmi_version_t` change. - It also blocks 4 doc-only unit corrections, which pass once acknowledged. - A PR that replaces the checker with `exit 0` still fails. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md. --------- Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> Co-authored-by: Maisam Arif <Maisam.Arif@amd.com>

adam360x added 3 commits October 15, 2026 09:00

adam360x requested review from a team and a balanced review from Copilot October 15, 2026 14:03

github-actions Bot added the project: amdsmi label Oct 15, 2026

Copilot started reviewing on behalf of adam360x October 15, 2026 14:04

View session

Copilot AI reviewed Oct 15, 2026

View reviewed changes

systems-assistant Bot added the organization: ROCm label Oct 15, 2026

dmitrii-galantsev approved these changes Oct 15, 2026

View reviewed changes

marifamd force-pushed the users/adam360x/ROCM-29548-followup branch from 2e1738c to 6db10cc Compare October 16, 2026 14:19

marifamd approved these changes Oct 16, 2026

View reviewed changes

koushikbillakanti-amd mentioned this pull request Oct 16, 2026

test(amdsmi): fix APU cleanup #13412

Draft

1 task done

marifamd mentioned this pull request Oct 16, 2026

fix(amdsmi): report absent APU cores as N/A, apply review follow-ups #13678

Open

1 task done

adam360x merged commit 58468a2 into develop Oct 21, 2026
126 of 127 checks passed

adam360x deleted the users/adam360x/ROCM-29548-followup branch October 21, 2026 01:09

This was referenced Oct 21, 2026

[rocjitsu] Correct legacy floating min/max and median selection #13898

Merged

[rocjitsu] Expose backed CDNA3 and CDNA4 hardware identity fields #13899

Merged

[rocjitsu] Implement scratch LDS transfers and scalar cache discard #13900

Merged

This was referenced Oct 27, 2026

chore(amdsmi): bump version to 27.3.0 and open the 10.3.0 changelog #14391

Closed

chore(amdsmi): bump version to 27.3.0 and correct the 10.x changelog #14390

Open

dmitrii-galantsev mentioned this pull request Oct 30, 2026

feat(amdsmi): block ABI breaks in amdsmi.h #14676

Merged

1 task

pmoutsias-amd mentioned this pull request Nov 2, 2026

Move AMD SMI changelog entries for code merged after the ROCm 10.2.0 branch to the amd_smi_lib for ROCm 10.3.0 section #14866

Closed

Navigation Menu

Repository navigation

fix(amdsmi): report APU utilization as whole percent - #13540

fix(amdsmi): report APU utilization as whole percent#13540
adam360x merged 6 commits into
developfrom
users/adam360x/ROCM-29548-followup

adam360x commented Oct 15, 2026 •

edited

Loading

Uh oh!

therock-pr-bot Bot commented Oct 15, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

systems-assistant Bot commented Oct 22, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

adam360x commented Oct 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Issue Tracking

Test Plan

Test Result

Submission Checklist

Uh oh!

therock-pr-bot Bot commented Oct 15, 2026

✅ All Policy Checks Passed

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟢 Approval recommended

Uh oh!

Uh oh!

systems-assistant Bot commented Oct 22, 2026

TheRock Submodule Bump Activity

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

adam360x commented Oct 15, 2026 •

edited

Loading