Vehicle Motion Control on SIMD: Traditional and AI based models on the edge
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Systems, control and mechatronics (MPSYS), MSc
Embedded electronic system design (MPEES), MSc
Embedded electronic system design (MPEES), MSc
Publicerad
2022
Författare
Suresh, Madhu
Sudarshan, Saurubh
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Recent advancements in technology such as Artificial Intelligence (AI) and Non-
Linear Model Predictive Control (NMPC) have led to its use in the field of motion
control in vehicles. When it comes to the implementation of the models related to
these technologies, they are expected to be executed within hard timing deadlines as
they are performance critical. Further, due to their high computational cost, coupled
with the strict deadlines, they are usually deployed on accelerators like the Graphics
Processing Unit (GPU). However, resource-constrained embedded platforms cannot
afford to have such accelerators. Therefore considering these limitations, it’s crucial
to thoroughly investigate the implementation of these models entirely on CPU without
any dedicated accelerator, while meeting the strict requirements. This thesis
investigates the method by analyzing two different models, viz. AI and NMPC models,
in which the Single Instruction Multiple Data (SIMD) component of an Arm
processor is exploited. The SIMD units are commonly used for vector operations
in a modern CPU. By using these models, various Arm’s SIMD implementation
techniques such as Arm Neon intrinsics, Ne10 library and Auto-vectorization are
investigated. When compared to the traditional approach of sequential computing
implementation, the proposed method implemented with Neon Intrinsics was found
to be more efficient and gave an execution time reduction of 61.9% for an AI model,
while the NMPC model gave an increase in execution time of 8.3%.
Beskrivning
Ämne/nyckelord
AI , NMPC , Graphical Processing Unit , Internet of Things , SIMD , Neon Intrinsics , Neon enabled library , CUDA