This talk focuses on the challenge of software self-configurability and system adaptivity. I introduce a methodology based on statistical machine learning for systematic automated optimisation of algorithmic and implementation parameters to achieve end-to-end quality-of-result objectives as well as energy and performance goals. I will show the effectiveness of our approach on next generation real-time 3D scene understanding applications where configurability is especially important. I briefly introduce SLAMBench, a publicly-available benchmarking framework which represents a starting point for quantitative, comparable and validatable experimental research to investigate trade-offs in performance, accuracy and energy consumption of 3D scene understanding systems. I then examine how it can be mapped to state-of-the-art desktop and mobile GPU-accelerated systems using a design configurability strategy based on Random forest and Active learning. The goal of this work is to take the human out of the loop and to provide self-optimization of complex software and hardware pipelines.

Bio:
Dr Luigi Nardi is a post-doctoral research associate at Imperial College London in the Software Performance Optimization group. Luigi's primary role is to work in the co-design of high-performance low-power computer vision systems where performance, power and accuracy are part of the same optimization space. Luigi earned his Ph.D. in computer science from Pierre et Marie Curie University, France, creating a new domain-specific language in the context of automatic code generation for numerical simulations in inverse problems. He has almost 10 years of experience in parallel computing and more than 6 years of experience developing GPU enabled codes using CUDA and OpenCL from desktop to embedded. Prior to his current position, Luigi was a permanent researcher at the financial firm Murex S.A.S., France, leading the high-performance computing effort.
Web page: http://wp.doc.ic.ac.uk/lnardi/