Module device_benchmark 

STELLAR §A AH-track H1(a) — cross-circuit capability benchmark (decisions D26/D30).

“Rank by measured throughput, not a static device-type hierarchy” (Timothy). This runs an identical representative GEMV on every available compute circuit — each wgpu adapter (discrete GPU, integrated GPU) plus a native-Rust CPU path — and produces a capability matrix sorted fastest-first. The residency/device planner (H2) consumes this to decide where work goes; e.g. a weak iGPU can still beat PCIe-streaming for overflow, and a many-core CPU can beat an old iGPU — only the numbers decide.

Scope (honest): GPUs/iGPU via wgpu; CPU via a rayon GEMV. NPU is not benchmarked — NPU access is a platform API (DirectML / NNAPI / CoreML), not wgpu, and is reported as “not probed”. This is part (a) of H1 (probe + matrix); the human-key signing of the passport (part (b)) is blocked on the identity remediation (identity-governance-remediation.md) and lives elsewhere.

Native only.

Structs

CapabilityMatrix

The measured capability matrix — circuits sorted fastest-first. This IS the priority order.

CircuitBench

One benchmarked circuit.

DeviceBenchmarkRequest

Stable process boundary for one physical adapter/backend benchmark.

DeviceBenchmarkResponse

Enums

CircuitKind

A compute circuit’s class in the capability matrix.

Functions

benchmark_devices

Benchmark every available compute circuit and return the ranked capability matrix.

run_worker_from_env

Worker entry used by the dedicated binary and the unit-test subprocess route.