An Original Investigation into Compiler Optimisation Strategies for Heterogeneous Computing on Nigerian Resource-Constrained Embedded Platforms

The growing use of heterogeneous SoC platforms combining CPU and neural processing units in Nigerian embedded computing applications for agriculture, healthcare, and industry creates new compiler optimisation challenges, and no investigation addresses the specific optimisation requirements for workloads deployed on affordable heterogeneous platforms within Nigerian hardware budget constraints. This study conducted an original investigation into compiler optimisation strategies for heterogeneous embedded computing platforms accessible within typical Nigerian engineering research budgets, making novel contributions to both the theoretical understanding of cross-device optimisation and a practical compilation toolchain. Three platforms were selected: Raspberry Pi 4 (BCM2711, Cortex-A72 CPU), Orange Pi 5 (RK3588, Cortex-A76 with Mali GPU), and Kendryte K210 (RISC-V with KPU neural processor). An original Static Heterogeneous Workload Partitioning Algorithm (SHWPA) was developed, using a directed acyclic graph representation of application computation graphs with annotated device execution time and memory transfer overhead estimates, solving the optimal kernel-to-device assignment as an integer linear programme. SHWPA was applied to five representative workloads from Nigerian applications: IoT sensor fusion, speech keyword spotting, image classification for crop disease, ECG anomaly detection, and industrial vibration analysis. SHWPA achieved 74.3 percent reduction in total execution time on the Orange Pi 5 heterogeneous platform compared to CPU-only execution, outperforming vendor-provided TVM tuning by 21.4 percent on the same workloads. A novel cross-platform portability metric was defined and validated, enabling quantitative comparison of optimisation portability across the three target platforms. Expert review confirmed the study's original methodological and theoretical contributions.

Keywords: compiler optimisation, heterogeneous computing, embedded platforms, SHWPA, Nigeria