A New Method to Generate Software Test Libraries for In-Field GPU Testing Resorting to High-Level Languages

Self-Test Libraries (STLs) are widely used by companies for in-field fault detection in CPU devices. Their usage is now extending to GPUs, due to their increasing adoption in safety-critical applications. Using STLs provided by GPU manufacturers, system companies can effectively test these devices during their operative life, as required by functional safety standards. In the automotive domain, GPUs are often used to process a high amount of sensitive information in real-time (e.g., object recognition and path tracking). Thus, GPU devices in this field must guarantee functional safety features (e.g., ISO26262) by using one or more functional safety mechanisms. This paper presents a methodology to develop STLs resorting to High-Level Languages (HLLs) (e.g., CUDA), reducing the complexity of encoding at the assembly level. Moreover, we describe the main advantages and discuss the challenges and constraints when developing STLs with HLLs for GPUs. In particular, we describe those cases that demand the usage of a Low-Level Language (LLL). Additionally, we highlight a method to develop STLs resorting to HLLs, at least for some modules. The FlexGripPlus GPU model was employed to evaluate and validate the proposed strategies experimentally. The results show that STLs based on HLLs can be effectively developed for regular modules in the GPU.