Machine Learning Based Auto-Tuning for Enhanced OpenCL Performance Portability

Abstract

Heterogeneous computing, which combines devices with different architectures, is rising in popularity, and promises increased performance combined with reduced energy consumption. OpenCL has been proposed as a standard for programing such systems, and offers functional portability. It does, however, suffer from poor performance portability, code tuned for one device must be re-tuned to achieve good performance on another device. In this paper, we use machine learning-based auto-tuning to address this problem. Benchmarks are run on a random subset of the entire tuning parameter configuration space, and the results are used to build an artificial neural network based model. The model can then be used to find interesting parts of the parameter space for further search. We evaluate our method with different benchmarks, on several devices, including an Intel i7 3770 CPU, an Nvidia K40 GPU and an AMD Radeon HD 7970 GPU. Our model achieves a mean relative error as low as 6.1%, and is able to find configurations as little as 1.3% worse than the global minimum.

ImageCL: Language and source‐to‐source compiler for performance portability, load balancing, and scalability prediction on heterogeneous systems

“…Despite good results, the time required for auto‐tuning, even using our machine learning‐based auto‐tuner, () can be a significant drawback. For the cases presented, the auto‐tuner executed around 1700 valid candidate implementations during its search for each device/benchmark combination.…”

“…The manual tuning was therefore carried out by systematically trying out different possible Halide schedules for each device/benchmark combination. The ImageCL implementations were auto‐tuned with the machine learning‐based auto‐tuner from our previous work,() which is described in Section 4.…”

“…This is particularly true for highly different devices, such as a CPU and a GPU, but also remains the case for two GPUs from the same vendor. ()…”

“…Performance portability is improved, since porting to a new device simply requires the auto‐tuning to be re‐done. While a high number of candidate implementations can lead to prohibitive search times, techniques based on machine learning, as proposed in our previous work,() can be used to speed up the search. This kind of auto‐tuning should also include the generation of candidate implementations, as writing these manually is tedious, time‐consuming, and error prone.…”

End-to-End Deep Learning of Optimization Heuristics

“…Researchers have focused on machine learning as a means to constructing high quality heuristics that often outperform their handcrafted equivalents [1][2][3][4]. A predictive model is trained, using supervised machine learning, on empirical performance data and important quantifiable properties, or features, of representative programs.…”