| name | verify-pmc-metrics |
| description | Mock-based rigorous PMC metric validation โ tests sentinel filtering, value ranges, and architecture-specific behavior across Perfetto and RocPD |
Mock-Based PMC Validation
Deterministic validation of AMD SMI Performance Monitoring Counter (PMC) metrics using the mock system (AMD_SMI_MOCK_CONFIG). Tests sentinel filtering, expected value ranges, positive presence of configured metrics, and negative absence of unsupported metrics across all output formats.
- Execute all profiling and validation automatically โ do NOT ask for user confirmation.
- Run ALL 6 test matrix entries (2 architectures x 3 formats).
- Sentinel detection is THE critical check โ a single leaked sentinel is a FAIL.
- Use rocprof-sys MCP tools for Perfetto traces, sqlite3 for RocPD databases.
- Mock configs are in the worktree at `~/worktree/rocm-systems/amd-smi-mocks/`.
Sentinel Value Reference
GPU firmware returns these values for unsupported/unavailable metrics. They must NEVER appear in output.
| Type | Hex | Decimal | Float (Perfetto) | Where Used |
|---|
uint16_t | 0xFFFF | 65535 | 65535.0 | Temperature, activity %, power, voltage, clocks, PCIe width/speed, VCN/JPEG arrays |
uint32_t | 0xFFFFFFFF | 4294967295 | 4294967295.0 | Various 32-bit fields |
uint64_t | 0xFFFFFFFFFFFFFFFF | 18446744073709551615 | 1.8446744073709552e+19 | PCIe bandwidth, XGMI data, memory |
Perfetto stores counter values as doubles, so UINT64_MAX appears as 1.8446744073709552e+19. Both exact and float representations must be checked.
Known Sentinel Filtering Pipeline
| Metric | device.hpp | perfetto_policy.hpp | perfetto_processor.cpp | rocpd_processor.cpp |
|---|
| VCN/JPEG (uint16) | Support check only | Filtered | Filtered | NOT filtered |
| XGMI (uint64) | Zeroed via populate_if_supported() | N/A | Double-checked | Filtered |
| PCIe (uint16/64) | Zeroed via populate_if_supported() | N/A | Relies on zero check | NOT filtered |
| NIC RDMA | N/A | N/A | No filtering | No filtering |
Known leakage risks:
- RocPD VCN/JPEG array values (no explicit sentinel filter)
- RocPD PCIe values (relies on device.hpp zeroing only)
- NIC metrics in all outputs (no sentinel filtering anywhere)
Mock Configurations
Two architectures test complementary metric subsets. Located in ~/worktree/rocm-systems/amd-smi-mocks/.
MI300X (mock_mi300x.yaml)
Data-center GPU with per-XCP metrics, HBM, and NIC.
| Category | What's Configured | Expected Range | Notes |
|---|
| Temperature | hotspot only | [55, 75] | No temperature_edge (sentineled) |
| HBM Temperature | 4 active / 4 total | [40, 60] | Array: active entries have values, rest sentineled |
| Power | current_socket_power | [300, 500] | Watts |
| GFX Busy | gfx_activity | [80, 100] | % |
| UMC Busy | umc_activity | [0, 10] | % |
| MM Busy | NOT configured | N/A | Should be absent (sentineled) |
| VCN Busy | per-XCP, 4 active engines | [0, 10] | Per-XCP format, NOT device-level VCN Activity |
| JPEG Busy | per-XCP, 32 active engines | [0, 10] | Per-XCP format, NOT device-level JPEG Activity |
| PCIe Width | Fixed 16 | 16 | Lanes |
| PCIe Speed | Fixed 32000 | 32000 | MT/s |
| XGMI | NOT configured | N/A | Should be absent (sentineled) |
| SDMA | 2 processes | varies | Cumulative byte counters |
| Memory | vram_used | varies | Derived from vram_used / vram_total |
| Voltages | NOT configured | N/A | Should be absent (sentineled) |
| Fan | NOT configured | N/A | Should be absent (sentineled) |
| NIC RDMA | 6 stats configured | [0, 1000000] | RX/TX bytes, packets, CNP |
Navi 48 (mock_navi48.yaml)
Consumer GPU with device-level metrics, no XCP, no NIC.
| Category | What's Configured | Expected Range | Notes |
|---|
| Temperature | edge, hotspot, mem | [35, 75] | Has temperature_edge |
| HBM Temperature | NOT configured | N/A | Should be absent (sentineled) |
| Power | current_socket_power | [20, 30] | Watts |
| GFX Busy | gfx_activity | [0, 10] | % |
| UMC Busy | umc_activity | [0, 5] | % |
| MM Busy | mm_activity | [0, 10] | % (present, unlike MI300X) |
| VCN Activity | device-level, 1 active / 4 total | [0, 10] | Device-level, NOT per-XCP |
| JPEG Activity | NOT configured | N/A | Should be absent (sentineled) |
| PCIe Width | Fixed 16 | 16 | Lanes |
| PCIe Speed | Fixed 2500 | 2500 | MT/s |
| XGMI | NOT configured | N/A | Should be absent |
| SDMA | 1 process | varies | Cumulative byte counters |
| Memory | vram_used | varies | Derived |
| Voltages | gfx, soc, mem configured | [700, 1200] | mV |
| Fan | rpm and max configured | [800, 1200] | RPM |
| NIC | No NIC devices | N/A | No NIC tracks should appear |
Output Formats
| # | Format | Environment Variables |
|---|
| 1 | Perfetto Standard | ROCPROFSYS_TRACE=1 |
| 2 | Perfetto Legacy | ROCPROFSYS_TRACE=1 ROCPROFSYS_TRACE_LEGACY=true |
| 3 | RocPD | ROCPROFSYS_USE_ROCPD=1 |
Perfetto Track Name Differences
| Format | GPU Pattern | NIC Pattern |
|---|
| Standard | GPU [N] <Metric> (S) | NIC [N] <Metric> (S) |
| Legacy | GPU <Metric> [N] (S) | NIC <device_name> <Metric> [N] (S) |
Legacy also uses abbreviated forms: Pkts vs Packets.
Test Matrix
| Run ID | Config | Format | Extra Env |
|---|
| mi300-perfetto | mock_mi300x.yaml | Perfetto Standard | ROCPROFSYS_SAMPLING_AINICS="all" |
| mi300-legacy | mock_mi300x.yaml | Perfetto Legacy | ROCPROFSYS_SAMPLING_AINICS="all" |
| mi300-rocpd | mock_mi300x.yaml | RocPD | ROCPROFSYS_SAMPLING_AINICS="all" |
| navi-perfetto | mock_navi48.yaml | Perfetto Standard | โ |
| navi-legacy | mock_navi48.yaml | Perfetto Legacy | โ |
| navi-rocpd | mock_navi48.yaml | RocPD | โ |
Common env for ALL runs:
AMD_SMI_MOCK_CONFIG=<path-to-yaml>ROCPROFSYS_AMD_SMI_METRICS="all"
Workload: ./transpose (mock generates all metric values regardless of workload type).
Execution
Phase 1: Setup
- Auto-discover build directory (
build/debug, build/release, or build/)
cd into build dir, source share/rocprofiler-systems/setup-env.sh- Verify mock support:
ldd bin/rocprof-sys-run | grep yaml (yaml-cpp must be linked)
- Locate mock configs:
~/worktree/rocm-systems/amd-smi-mocks/mock_mi300x.yaml and mock_navi48.yaml
- Verify
transpose workload exists
Phase 2: Execute Test Matrix
Run 6 profiling sessions. Use /tmp/pmc-validate-<run-id>/ output paths.
source share/rocprofiler-systems/setup-env.sh
AMD_SMI_MOCK_CONFIG=~/worktree/rocm-systems/amd-smi-mocks/mock_mi300x.yaml \
ROCPROFSYS_OUTPUT_PATH=/tmp/pmc-validate-mi300-perfetto \
ROCPROFSYS_TRACE=1 \
ROCPROFSYS_AMD_SMI_METRICS="all" \
ROCPROFSYS_SAMPLING_AINICS="all" \
bin/rocprof-sys-run -- ./transpose
AMD_SMI_MOCK_CONFIG=~/worktree/rocm-systems/amd-smi-mocks/mock_mi300x.yaml \
ROCPROFSYS_OUTPUT_PATH=/tmp/pmc-validate-mi300-legacy \
ROCPROFSYS_TRACE=1 \
ROCPROFSYS_TRACE_LEGACY=true \
ROCPROFSYS_AMD_SMI_METRICS="all" \
ROCPROFSYS_SAMPLING_AINICS="all" \
bin/rocprof-sys-run -- ./transpose
AMD_SMI_MOCK_CONFIG=~/worktree/rocm-systems/amd-smi-mocks/mock_mi300x.yaml \
ROCPROFSYS_OUTPUT_PATH=/tmp/pmc-validate-mi300-rocpd \
ROCPROFSYS_USE_ROCPD=1 \
ROCPROFSYS_AMD_SMI_METRICS="all" \
ROCPROFSYS_SAMPLING_AINICS="all" \
bin/rocprof-sys-run -- ./transpose
AMD_SMI_MOCK_CONFIG=~/worktree/rocm-systems/amd-smi-mocks/mock_navi48.yaml \
ROCPROFSYS_OUTPUT_PATH=/tmp/pmc-validate-navi-perfetto \
ROCPROFSYS_TRACE=1 \
ROCPROFSYS_AMD_SMI_METRICS="all" \
bin/rocprof-sys-run -- ./transpose
AMD_SMI_MOCK_CONFIG=~/worktree/rocm-systems/amd-smi-mocks/mock_navi48.yaml \
ROCPROFSYS_OUTPUT_PATH=/tmp/pmc-validate-navi-legacy \
ROCPROFSYS_TRACE=1 \
ROCPROFSYS_TRACE_LEGACY=true \
ROCPROFSYS_AMD_SMI_METRICS="all" \
bin/rocprof-sys-run -- ./transpose
AMD_SMI_MOCK_CONFIG=~/worktree/rocm-systems/amd-smi-mocks/mock_navi48.yaml \
ROCPROFSYS_OUTPUT_PATH=/tmp/pmc-validate-navi-rocpd \
ROCPROFSYS_USE_ROCPD=1 \
ROCPROFSYS_AMD_SMI_METRICS="all" \
bin/rocprof-sys-run -- ./transpose
Phase 3: Sentinel Detection (CRITICAL)
This is the most important validation phase. A single leaked sentinel value is a FAIL.
Perfetto Traces (Standard and Legacy)
Load each trace with MCP tools, then run:
SELECT t.name, c.value, c.ts
FROM counter c JOIN counter_track t ON c.track_id = t.id
WHERE t.name LIKE '%(S)%'
AND (c.value = 65535
OR c.value = 4294967295
OR c.value = 18446744073709551615
OR c.value = 1.8446744073709552e+19
OR c.value >= 1.844e+19)
LIMIT 20;
RocPD Databases
sqlite3 <db-file> "
SELECT p.name, e.value
FROM rocpd_pmc_event e JOIN rocpd_info_pmc p ON e.pmc_id = p.id
WHERE e.value IN (65535, 4294967295, 18446744073709551615)
OR e.value > 1e18
LIMIT 20;
"
Phase 4: Positive Validation (Configured Metrics Present)
Verify that all configured metrics appear in output with expected value ranges.
Perfetto SQL
SELECT t.name,
COUNT(*) as samples,
MIN(c.value) as min_val,
MAX(c.value) as max_val,
ROUND(AVG(c.value), 2) as avg_val
FROM counter c JOIN counter_track t ON c.track_id = t.id
WHERE t.name LIKE '%(S)%'
GROUP BY t.name
ORDER BY t.name;
RocPD SQL
sqlite3 <db-file> "
SELECT p.name, COUNT(*) as samples,
MIN(e.value) as min_val, MAX(e.value) as max_val,
ROUND(AVG(e.value), 2) as avg_val
FROM rocpd_pmc_event e JOIN rocpd_info_pmc p ON e.pmc_id = p.id
GROUP BY p.name ORDER BY p.name;
"
MI300X Expected Metrics
| Metric | Perfetto Standard Track | Expected Range | Notes |
