透過您的圖書館登入
IP:52.15.183.117
  • 學位論文

基於中央處理器及圖形處理單元之協同運算

A Synergistic Framework base on Halide

指導教授 : 廖世偉

摘要


在這篇論文裡,我們打造了一個協同運算的框架做為 Halide 的 附加元件,藉以利用 Halide 之優勢試圖去創造一個較好的異質計算 環境。在此框架中我們提供兩種的分配工作量方法,這兩種工作法 方可以提供不同程度的加速優化。我們會在這篇論文中測量兩種分 配方法提供的優化程度以及在不同輸入大小的優化程度。在實驗中 使用我們的框架搭配最佳化過的 Halide 的 CPU 及 GPU 程式可以得 到 1.21x,1.55x,1.16x 的加速在 Nexus 7,Intel CPU 搭載 ATI GPU 和 Intel CPU 搭載 Nvidia GPU.

並列摘要


In this thesis,we create a synergistic framework as Halide extension to build a better heterogeneous computing environment with the advantage of Halide.Our framework support 2 methods - dynamic dispatch and static dis- patch to assign different workload to CPU and GPU with those dispatch method we will get different level speedup.And we will measure those dispatch meth- ods with different input sizes at mobile,ATI GPU and NVIDIA GPU.In our experiment, using our framework with optimized Halide CPU and GPU funci- ton can get 1.21x speedup at Nexus 7, 1.55x speedup at Inetel CPU with ATI GPU and 1.16x speedup at Intel CPU with Nvidia GPU.

參考文獻


[1] Sylvain Paris Marc Levoy Saman Amarasinghe Fredo Durand. Jonathan Ragan- Kelley, Andrew Adams. Decoupling algorithms from schedules for easy optimiza- tion of image processing pipelines.
[2] Jonathan Ragan-Kelley,Connelly Barnes,Andrew Adams,Sylvain Paris,Fredo Du- rand,Saman Amarasinghe. Halide: a language and compiler for optimizing paral- lelism, locality, and recomputation in image processing pipelines. In Proceedings of the 34th ACM SIGPLAN Conference on Programming Language Design and Im- plementation.
[3] R. Govindarajan Prasanna Pandit. Fluidic kernels: Cooperative execution of opencl programs on multiple heterogeneous devices. In Proceedings of Annual IEEE/ACM International Symposium on Code Generation and Optimization.
[4] Khronos Group. Opencl - the open standard for parallel programming of heteroge- neous systems. 2013.
[5] C. omasi and R. Manduchi. Bilateral filtering for gray and color images. In In Proceedings of the IEEE International Conference on Computer.

延伸閱讀