臺灣博碩士論文加值系統

隨著需要計算的資料量越來越大，我們使用了通用圖型處理器來輔助中央處理器以提升整體的計算速度的例子越來越多，也直接導致了異質計算語言在近幾年變得越來越普及。
其中又由非盈利性技術組織Khronos所掌管開放工業標準的OpenCL即為一例，OpenCL是由幾家處理器大廠的技術團隊所制定出來的標準，專為異質架構平台編寫程式的框架，由設定硬體資訊的OpenCL API，和撰寫欲平常的kernel (OpenCL C)所組成。
雖然有了OpenCL能夠控制GPU，CPU跟GPU之間的還是有許多限制以及缺陷。像是CPU跟GPU之間的虛擬記憶體位址的不同，資料在PCIe匯流排上傳輸成了運算的瓶頸。所以在2012年由異質系統架構基金會提出了一個將CPU以及GPU整合在一起的新架構，改進了異質計算一些缺陷，企圖徹底釋放運算能力及能源使用效率。
在此篇論文中，實作了一些OpenCL 2.0新制定的特性，如Shared Virtual Memory, C11 atomic, Pipe…。並在OpenCL的runtime的實作中運用了HSA runtime 規格書中所定義的特性，如User mode queue, HSA signal, HSA AQL packet…etc來實作，且能夠在HSAemu上用AMD APP SDK裡面OpenCL benchmark, Intel ocl2.0 benchmark, Rodinia OpenCL benchmark來驗證實作的正確性。

In the recent years GPGPU and heterogeneous computing become more and more popular. One of the popular case is OpenCL. OpenCL 2.0 brings many features that can reduce overhead while writing OpenCL programs. For example: Shared Virtual Memory, Dynamic Parallelism, Pipes. These enhance the interaction between host and devices. And HSA foundation released a new architecture name Heterogeneous System Architecture (HSA). HSA integrates CPU and GPU into one chip try to reduce the overhead of communication between CPU and GPU or other acceralators, improve the energy-efficiency of heterogeneous system.
In this thesis we accomplish a full time system HSAemu 2.0 which supports most OpenCL 2.0 features. We will introduce the OpenCL 2.0 runtime and HSA runtime v1.0. And give an overview how we integrates OpenCL runtime and HSA runtime together to meet the requirements. Finally we use OpenCL benchmarks in AMD APP SDK 3.0 Beta to verify the OpenCL 2.0 features.

Chapter 1 Introduction 8
Chapter 2 Background 10
2.1 OpenCL 2.0 10
2.2 Heterogeneous System Architecture Runtime 10
Chapter 3 Related work 13
Chapter 4 Architecture of HSAemu 15
4.1 HSAemu overview 15
4.2 OpenCL 2.0 runtime 16
Chapter 5 Implementation 18
5.1 Shared Virtual Memory 18
5.2 C11 atomic 20
5.3 Pipes 21
5.4 Nested Parallelism 22
5.5 Relation between OpenCL runtime and HSA runtime 23
Chapter 6 Experiment 27
6.1 Shared Virtual Memory 27
6.2 Dynamic Parallelism 29
6.3 Pipes 30
Chapter 7 Conclusion and Future work 32
Reference 33

1. OpenCL https://www.khronos.org/opencl/.
2. HSA foundation http://www.hsafoundation.com/
3. Ding, J.-H., et al. HSAemu: a full system emulator for HSA platforms. in Proceedings of the 2014 International Conference on Hardware/Software Codesign and System Synthesis. 2014. ACM.
4. Jeng, B.-C., 異質計算模擬器框架. 清華大學資訊工程學系學位論文, 2014: p. 1-29.
5. Lin, K.-M., HSA 編譯框架. 清華大學資訊工程學系學位論文, 2014: p. 1-26.
6. HSA Platform System Architecture Specification 1.0 http://www.hsafoundation.com/standards/.
7. HSA Programmer Reference Manual Specification 1.0 http://www.hsafoundation.com/standards/.
8. HSA Runtime Specification 1.0 http://www.hsafoundation.com/standards/.
9. AMD APP SDK v3.0 Beta http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/.
10. NVIDIA OpenCL SDK https://developer.nvidia.com/opencl.
11. Intel OpenCL https://software.intel.com/en-us/intel-opencl.
12. POCL http://portablecl.org/.
13. HSA-Runtime-AMD https://github.com/HSAFoundation/HSA-Runtime-AMD.
14. Okra-Interface-to-HSA-Device https://github.com/HSAFoundation/Okra-Interface-to-HSA-Device.
15. HSA-Runtime-Reference-Source https://github.com/HSAFoundation/HSA-Runtime-Reference-Source.
16. POCL-HSA https://github.com/HSAFoundation/POCL-HSA.
17. Multi2Sim https://www.multi2sim.org/.
18. SVMOverview https://software.intel.com/en-us/articles/opencl-20-shared-virtual-memory-overview.
19. GNU atomic built-in https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html.
20. VPMU https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/7/html/Virtualization_Tuning_and_Optimization_Guide/sect-Virtualization_Tuning_Optimization_Guide-Monitoring_Tools-vPMU.html.