P32 priority control microkernel
A high-priority control microkernel can update the cancellation epoch while a long P32 compute kernel is resident.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
A high-priority control microkernel can update the cancellation epoch while a long P32 compute kernel is resident.
Why the test is meaningful
CUDA stream priority affects queued scheduling but does not imply mid-kernel preemption when the running grid occupies device resources.
control_response=t_epoch_seen−t_control_launchgate: epoch changes during computelatency target≤15 msHow it was tested
Launch exact P32 compute, enqueue a minimal control microkernel on a high-priority stream, and timestamp eight attempts while preserving full-domain correctness and paired throughput checks.
What happened
The control kernel launched but was starved until compute completed. Zero of eight signals were observed during work; responses were 156.66–157.30 ms. Retention was 0.999055 with CI95 0.998749–0.999361.
Exactness and statistical controls
B32, 2^32 work, zero discrepancies and zero spills passed. The control-path technical gate failed 0/8.
What the result means
Stream priority is not preemption for this resident workload. The mechanism is rejected without an operational campaign.
Limitations
- The conclusion applies to this occupancy and scheduling configuration.
- Different cooperative or preemptible kernels are separate hypotheses.
- Correctness does not rescue failed control latency.
Evidence trail
Use sealed compute/control kernels, confirm the control launch timestamp, and distinguish launch from epoch observation in every trial.
Canonical variants
CANONICAL-EXP-170PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.