Conferences and supporting programme
Performance Profiling and Optimization for Autonomous Driving Workload
Autonomous driving system consists of multiple modules to process the data over the perception, localization, planning and control stages. The performance is the key fact which could impact the hardware selection, functional safety, user experience and also the total cost of the solution. In this paper, we select the open source Apollo (http://apollo.auto) project as the reference project, use the EM motion planning module as the reference module, and run the benchmark and performance profiling on one of Intel Atom Architecture based low cost hardware platform. We introduced the general methodologies for running the profiling inside Apollo framework and the detailed analysis based on the profiling data such as hotspot, CPU pipeline execution efficiency, parallelization etc. Based on the performance profiling data, we are trying to extract the best performance on the selected hardware by using different optimization technologies including the compiler optimization, high performance libraries, vectorization and parallelization to optimize the planning module. With these optimizations applied, we could receive close to 2x performance speedup compared to the initial setup. The optimization work approves that even for the low cost, power efficiency hardware platform, as long as the workload is highly optimized, we can still get the expected performance, which breaks the door for the autonomous driving solution vendor to select the cost saving hardware platforms.
--- Date: 27.02.2019 Time: 10:00 AM - 10:30 AM Location: Conference Counter NCC Ost