ImageCL: An Image Processing Language for Performance Portability on Heterogeneous Systems

摘要

Modern computer systems typically conbine multicore CPUs with acceleratorslike GPUs for inproved performance and energy efficiency. However, these sys-tems suffer from poor performance portability, code tuned for one device mustbe retuned to achieve high performance on another. Image processing is increas-ing in importance , with applications ranging from seismology and medicine toPhotoshop. Based on our experience with medical image processing, we proposeImageCL, a high-level domain-specific language and source-to-source compiler,targeting heterogeneous hardware. ImageCL resembles OpenCL, but abstracts awayper- formance optimization details, allowing the programmer to focus onalgorithm development, rather than performance tuning. The latter is left toour source-to- source compiler and auto-tuner. From high-level ImageCL kernels,our source- to-source compiler can generate multiple OpenCL implementationswith different optimizations applied. We rely on auto-tuning rather thanmachine models or ex- pert programmer knowledge to determine whichoptimizations to apply, making our tuning procedure highly robust. Furthermore,we can generate high perform- ing implementations for different devices from asingle source code, thereby im- proving performance portability. We evaluateour approach on three image processing benchmarks, on different GPU and CPUdevices, and are able to outperform other state of the art solutions in severalcases, achieving speedups of up to 4.57x.