使用執行期配置與轉譯技術達成可適性異質多核心計算

Abstract

本篇論文是在OpenCL的架構上，為異質多核心計算系統增添可適性。實作上我們先在執行時期取得系統環境資訊，再決定kernel該配置於何種裝置。若將kernel配置於CPU上，則進一步採用Whole Function Vectorization所提供的轉譯技術，使kernel能轉譯成SIMD指令，來提升kernel在CPU上執行的效能。此外，我們也利用過去歷史資料來協助預測裝置的附載量，讓該系統更智慧地配置。我們選擇AMD APP SDK所提供的範例程式來驗證增加可適性能改善OpenCL應用程式的效能。實驗結果指出，增加可適性於OpenCL系統可平均省下65%的時間。

This thesis is to quip adaptability for heterogeneous computing in OpenCL framework. The implementation is that first we obtain the statuses of devices at runtime and then determine on which device kernels execute. If the kernel is dispatched into CPU, in order to enhance the performance on CPU, it will be translated to SIMD instructions by the translation techniques provided by Whole Function Vectorization. Moreover, for the sake of more intelligent dispatch, we utilized the profile history to predict the loads of devices. With the improvement, experiments indicate that the average improvement of execution time is 65% on the samples provided by AMD APP SDK.