HetroCV: Auto-tuning Framework and Runtime for Image Processing and Computer Vision Applications on Heterogeneous Platform

With the wide adoption of high-performance processors and accelerators, large-scale computer vision applications have gained great performance improvement. However, it often requires extensive experiments and expertise to achieve optimal performance from manually-tuned programs, and the programs often need to be re-tuned when transplanted to a different platform, or using a different system configuration. To overcome this problem, in this paper we proposed Hetro CV, a programmer-directed auto-tuning framework and runtime for computer vision applications on heterogeneous CPU-MIC platform. In Hetro CV auto-tuning framework, computation units in the application pipeline are categorized in to one of three patterns: Map, Stencil and MapReduce, and program statistics are extracted from units' meta-information. Machine learning is adopted to train models for each pattern using the tuned parameters and program statistics from trial run sets, so that when a new unit is presented, Hetro CV auto tuner can use the corresponding trained model to generate optimized tuning parameters. In Hetro CV runtime, performance models for processor and co-processor are built to predict the prospective execution time of each computation unit in the application pipeline. We adopted the maximum-throughput mapping strategy, thus each unit would be mapped dynamically to the processor/co-processor queue, which would generate the minimum overall execution time. Experiments on two medical image processing applications running on heterogeneous platform composed of Intel Xeon CPU and Intel Phi co-processor showed advanced performance over naive Open MP tuning and Genetic Algorithm (GA) based heuristic tuning.

I. Introduction

As the rapid development of digital imaging devices and the large increase of images and video generated from social media, there has been a surge of interest in image processing applications and softwares. At the same time, the image processing algorithms they depended on are evolving in both complexity and scale, which made them in great need of high performance computing implementations. Meanwhile, as the growing of high-performance field, there have been more and more types of high-performance computing devices developed and applied into HPC appications, from the early SMP devices, GPGPUs, to the more recent many-core architecture Intel Phi coprocessors [1]. To take advantage the computing power, extensive work have been devoted to accelerate the computational intensive image processing [2]–[5] and computer vision [6], [7] applications on highperformance platforms.