Keep an iterative loop on the device

Synthetic English narration.

Different markets stop at different rounds
These four illustrative markets need two, four, three and five correction rounds. Their first stopping points differ. The solver must preserve those individual stopping decisions.

Launch and check each round
In a host-driven loop, the CPU repeatedly launches a device update and checks whether work remains. Many short rounds can add launch and synchronization overhead.

Retain each first stopping point
As a market converges, freeze its result. The remaining markets continue. Requiring every market to run five effective updates would change this stopping behavior.

Move the loop onto the device
A resident kernel keeps the correction loop and active masks on the device. One launch can cover several rounds while retaining each market's first stopping point.

Save coordination, retain the contract
The numerical target, precision and stopping rule remain fixed. The saving is in coordination. Register use and occupancy may change, so compare validated complete runs.