Survey and Overview on Code Optimization and Hardware Adaptation of Heterogeneous Platforms (CPU vs GPU) For Real-Time and Accelerated Simulation to Achieve Level 5 of Autonomous Driving

​"The complexity and computing requirements for autonomous driving compelled many academic researchers, automotive makers and especially electronic device manufacturers to innovate with their respective products to face the various challenges of self-driving vehicles. These challenges consist of driving the vehicles over long distances: about twelve billion miles in a simulated environment and around a hundred million miles on the public roads. To this end, many resort to the parallel computer on homogeneous and heterogeneous platforms. A good example of these platforms is modern high-performance computers capitalizing, at the same time, on computing and graphics computing units’ capabilities and resources. These units form eterogeneous platforms that are dedicated to high-performance and parallel computing. The main gap of such frameworks is the data

transfer latency between the host and its associated accelerators. This is mainly due to the process of transferring data from the host’s Random-Access Memory (RAM) to the RAM of an attached device. The data transfer process goes through different stages, from retrieving data from Pageable and Pinned memory, then transferring them to devices, generally through the Peripheral Component Interconnect Express (PCIe) bus. This paper introduces a general overview of the related work and the different concepts that are put forward to reduce the data transferring latency. It introduces some new concepts that could overcome or at least circumvent the main limitation of heterogeneous platforms. Thus, the last part of the paper tables performances of different data transfer methods on heterogeneous platforms. The experiments were conducted on the NVIDIA Jetson TX2 developments board and on an NVS 510 GPU."