Data Capture#

The capture subsystem collects PMU counter data from the target MCU over RTT, USB CDC, or SWO, and optionally power measurements via Joulescope.

SWO capture pipeline#

graph LR
    A[Firmware<br/>SWO printf] -->|ITM Channel 0| B[J-Link probe]
    B -->|USB| C[pylink]
    C --> D[transport/swo.py]
    D -->|raw lines| E[parser.py]
    E -->|PmuResult| F[Pipeline context]

SWO capture backend#

SWO capture uses the J-Link probe through the pylink Python bindings. The reader configures SWO in-process and then drains stimulus port 0:

The reader collects lines until it sees HPX_END or a timeout expires.

HPX protocol#

The firmware prints a structured text protocol over SWO. The protocol is line-oriented and human-readable for debugging:

HPX_START
HPX_META arena_size=131072
HPX_META model_size=45632
HPX_META tensor_count=24
HPX_META layer_count=13
HPX_PRESET cpu
HPX_COUNTERS CPU_CYCLES,INST_RETIRED,LD_RETIRED,ST_RETIRED,BR_RETIRED,BR_MIS_PRED,STALL_FRONTEND,STALL_BACKEND
HPX_ITER 0
0,CONV_2D,1234567,456789,123456,78901,23456,1234,5678,9012
1,DEPTHWISE_CONV_2D,987654,321098,76543,21098,7654,321,987,654
...
HPX_ITER 1
0,CONV_2D,1234568,456790,123457,78902,23457,1235,5679,9013
...
HPX_PRESET_DONE
HPX_PRESET cache
HPX_COUNTERS L1D_CACHE,L1D_CACHE_RD,L1D_CACHE_REFILL,...
HPX_ITER 0
...
HPX_PRESET_DONE
HPX_END

Protocol elements#

Parser#

File: capture/parser.py

The parser processes the raw SWO lines into structured data:

def parse_hpx_output(lines: list[str]) -> PmuResult:
    """Parse HPX protocol lines into a PmuResult."""

Parsing steps#

1. Extract metadata — HPX_META lines → dict[str, str]
2. Group by preset — lines between HPX_PRESET and HPX_PRESET_DONE
3. Parse CSV rows — each data line → layer index, op name, counter values
4. Average iterations — for each layer, average counters across iterations
5. Build PresetResult — one per counter preset

Result structure#

@dataclass
class PmuResult:
    metadata: dict[str, str]         # HPX_META key-value pairs
    presets: list[PresetResult]      # One per counter pass

@dataclass
class PresetResult:
    name: str                        # Preset name (e.g. "cpu")
    counter_names: list[str]         # Column headers
    layers: list[LayerResult]        # Per-layer averaged data

@dataclass
class LayerResult:
    index: int                       # Layer index (0-based)
    op_name: str                     # TFLite op name
    counters: dict[str, float]       # counter_name → averaged value

Multi-pass merging#

When profiling requires more counters than the hardware supports (8 on Cortex-M55), the pipeline runs multiple passes. After all passes complete, the report stage merges them:

# Merge presets into unified layer results
for preset in pmu_result.presets:
    for layer in preset.layers:
        merged[layer.index].counters.update(layer.counters)

The merge assumes: - Layer ordering is stable — same model produces same layer indices - CPU_CYCLES is repeated in every preset for cross-validation - Layer count is identical across presets

If layer counts don't match (firmware issue), the merge raises a CaptureError with diagnostics.

Timeouts and error handling#

Power capture#

When power.enabled is true, the power stage runs after PMU capture:

1. Reset target — power-cycle via Joulescope
2. Start capture — begin current/voltage sampling
3. Wait for duration — power.duration_s seconds
4. Stop and compute — average current, peak current, energy

The power result is independent of PMU data — they capture different aspects of the same inference workload. Correlating them is done at the report level, not during capture.