Razporejevalnik opravil za vgrajeni sistem NXP FRDM-MCXN947

This thesis designs and evaluates a custom task scheduler for the NXP FRDM-MCXN947 embedded platform that integrates a general-purpose ARM Cortex-M33 CPU with an eIQ Neutron Neural Processing Unit (NPU). The objective is to enable concurrent execution of CPU-bound and NPU-bound workloads. The work addresses the need for resource-aware scheduling and precise timing on heterogenous embedded systems that run neural-network inference alongside conventional tasks. The scheduler is implemented on FreeRTOS and instrumented with a Data Watchpoint and Trace (DWT) cycle-counter timer to obtain NPU-safe, microsecond-resolution measurements. A custom task model augments classical real-time parameters (release time, period, deadline, execution time, priority) with additional fields. Type field enables the scheduler to partition tasks onto CPU or NPU resources. The dispatch policy is priority-based with Earliest-Deadline-First (EDF).Execution times are taken into account for task execution analysis. The evaluation combines synthetic CPU tasks and a NPU MobileNet inference task (TensorFlow Lite Micro model converted to a Neutron graph), exercised under periodic and mixed-priority scenarios. The synthetic test scenarios confirm concurrent use of heterogenous processing units, subsequently shortening overall execution time, and correct priority handling and soft/hard deadline modes.