在Halide中運用機器學習方法來進行異質計算

Machine Learning Based Approach for Achieving Heterogeneous Computing in Halide

10.6342/NTU201701638

清水悠揮

Halide ； 影像處理 ； 異質計算 ； 機器學習 ； Halide ； Image Processing ； Heterogeneous Computing ； Machine Learning

國立臺灣大學資訊工程學系學位論文

2017年

碩士

廖世偉

英文

通用GPU（GPGPU）已成為運行一般平行數據之應用程序的通用方式，而異構多核平台由於其可用性和性能的提高，而顯著增加。此外，圖像處理管線正在成為廣泛應用中必不可少的計算組件，並且顯然地，將這些應用程序運行在異構環境中，可以獲得更好的性能。然而，因為每個管道都優先考慮不同的設備，因此很難找出不同設備之間的最佳任務劃分。在本文中，我提出了一種基於機器學習的方法來自動優化最佳分區，並為Halide —— 一種圖像處理管線的有效系統的DSL —— 開發框架。我們的實驗結果顯示出比以前的純粹動態方法更好的性能，並且大多數管線在運行單個設備方面皆獲得了性能提升。

General Purpose GPUs (GPGPU) have become common-place for running general purpose data-parallel applications, and Heterogeneous multi-core platforms are increasing being significant due to its availability and performance improvement. Additionally, the image processing pipelines are becoming essential computing components in a wide range of applications, and it is apparent that running these application on heterogeneous environment can gain better performance. However, it is very difficult to find out the best task partitioning among different devices since each pipeline prefers different devices following its characteristics. In this paper, I propose a machine learning based approach for fining out the best partitioning automatically, and develop a framework for Halide, which is a DSL proved to be an effective system for image processing pipelines. The result of our experiment shows the better performance than the previous work which is purely dynamic approach, and most of the pipelines gain performance improvement over running single devices.

1. [1]. Grewe, Dominik, and Michael F. P. OâBoyle. "A Static Task Partitioning Approach for Heterogeneous Systems Using OpenCL." Lecture Notes in Computer Science Compiler Construction (2011): 286-305.
   連結：
2. [3]. Ragan-Kelley, Jonathan, Andrew Adams, Sylvain Paris, Marc Levoy, Saman Amarasinghe, and Fredo Durand. "Decoupling Algorithms from Schedules for Easy Optimization of Image Processing Pipelines." ACM Transactions on Graphics 31.4 (2012)
   連結：
3. [4]. Wen, Yuan, Zheng Wang, and Michael F. P. O'boyle. "Smart Multi-task Scheduling for OpenCL Programs on CPU/GPU Heterogeneous Platforms." 2014 21st International Conference on High Performance Computing (HiPC) (2014)
   連結：
4. [5]. Mullapudi, Ravi Teja, Andrew Adams, Dillon Sharlet, Jonathan Ragan-Kelley, and Kayvon Fatahalian. "Automatically Scheduling Halide Image Processing Pipelines." ACM Transactions on Graphics 35.4 (2016): 1-11.
   連結：
5. [6]. Sun, Enqiang, Dana Schaa, Richard Bagley, Norman Rubin, and David Kaeli. "Enabling Task-level Scheduling on Heterogeneous Platforms." Proceedings of the 5th Annual Workshop on General Purpose Processing with Graphics Processing Units - GPGPU-5 (2012)
   連結：
6. [9]. Tomasi, C., and R. Manduchi. "Bilateral Filtering for Gray and Color Images." Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271)
   連結：
7. [10]. Stone, John E., David Gohara, and Guochun Shi. "OpenCL: A Parallel Programming Standard for Heterogeneous Computing Systems." Computing in Science & Engineering 12.3 (2010): 66-73.
   連結：
8. [2]. Rinker, R., J. Hammes, W.a. Najjar, W. Bohm, and B. Draper. "Compiling Image Processing Applications to Reconfigurable Hardware." Proceedings IEEE International Conference on Application-Specific Systems, Architectures, and Processors (n.d.)
9. [7]. Pandit, Prasanna, and R. Govindarajan. "Fluidic Kernels." Proceedings of Annual IEEE/ACM International Symposium on Code Generation and Optimization - CGO '14 (2014)
10. [8]. Kaleem, Rashid, Rajkishore Barik, Tatiana Shpeisman, Brian T. Lewis, Chunling Hu, and Keshav Pingali. "Adaptive Heterogeneous Scheduling for Integrated GPUs." Proceedings of the 23rd International Conference on Parallel Architectures and Compilation - PACT '14 (2014)
11. [11]. Paris, Sylvain, Samuel W. Hasinoff, and Jan Kautz. "Local Laplacian Filters." Communications of the ACM 58.3 (2015): 81-91