臺灣博碩士論文加值系統

棘波序列度量法(Spike Train Metrics)之目的，主要是計算兩筆棘波序列的距離(差異性)，依據棘波序列比對的差異性，研究神經活動的變化及其觸發的相似性。由於棘波序列度量法中利用動態規劃的方式，比對棘波序列之相似度，在處理過程中產生巨量且相似的運算步驟。因此，本論文將利用平行處理改善動態規劃之計算時間，並使用統一運算架構(Compute Unified Device Architecture; CUDA)的通用圖形處理器(General-Purpose Computing on Graphics Processing Units; GPGPU)執行計算。另外，我們對於如何善用圖形處理器的記憶體亦作深入的瞭解，並於程式中適當規劃記憶體的配置，以加速程式的執行時間。為了測試本研究的增速效果，本論文將使用實際資料於本研究方法與循序棘波序列度量法進行驗證。實驗結果證實本研究方法在時間的執行效率上優於循序棘波序列度量法，並可增加約300倍的執行效率。

A spike train metric is used to calculate the distance of two spike sequences and measures their similarity. It is helpful to investigate the relation between neural responses. The computational methodology of spike train metric is based on dynamic programming, which is very time consuming for long length of spike trains. Therefore, this research uses the technique of parallel processing on computer graphics to improve the performance. In this research, we explore the use of CUDA (Compute Unified Device Architecture), which is based on the GPGPU (General-Purpose Computing on Graphics Processing Units) technology. We deeply study the architecture of CUDA, and develop a parallel version of dynamic programming using efficient memory distribution to speed up the execution time. A real data set of spike trains was applied on the parallel implementation to verify its efficiency. From the results, our implementation achieved a speedup of approximately 300 over a sequential implementation.

第1章 緒論 1

第2章 研究背景知識 3
第2.1節 棘波序列度量法 3
第2.2節 CUDA 8
第2.2.1節 程式模型 8
第2.2.2節 記憶體模型 11

第3章 研究方法 16
第3.1節 平行演算法 16
第3.1.1節 動態規劃之平行化 16
第3.1.2節 時間精準度參數q之平行化 23
第3.2節 記憶體配置 24
第3.2.1節 紋理記憶體 25
第3.2.2節 共享記憶體 25

第4章 實驗結果 32
第4.1節 實驗設計 32
第4.1.1節 資料簡介 32
第4.1.2節 硬體介紹 33
第4.2節 實驗結果與探討 34
第4.2.1節 Shift與Modulus方法之比較 34
第4.2.2節 單一Block執行一層平行化 37
第4.2.3節 多個Block執行兩層平行化 41

第5章 結論與未來展望 46
第5.1節 結論 46
第5.2節 未來展望 47

參考文獻 48

[1]Aronov, D., Reich, D.S., Mechler, F., Victor, J.D. “Neural coding of spatial phase in V1 of the macaque monkey.” Journal of Neurophysiology. 89:3304-3327 (2003).

[2]Aronov, D. “Fast algorithm for the metric-space analysis of simultaneous responses of multiple single neurons.” Journal of Neuroscience Methods. 124:175-179 (2003).

[3]Chichilnisky, E.J., Rieke, F. “Detection sensitivity and temporal resolution of visual signals near absolute threshold in the salamander retina.” The Journal of Neuroscience, 25(2):318-330 (2005).

[4]Machens, C.K., Prinz, P., Stemmler, M.B., Ronacher, B., Herz, A.V.M. “Discrimination of behaviorally relevant signals by auditory receptor neurons.” Neurocomputing, 38:263-268 (2001).

[5]Manocha, D. “General-purpose computations using graphics processors.” IEEE Computer Society. 38(8):85-88 (2005).

[6]Liu, W., Schmidt, B., Voss, G., Schroeder, A., Muller-Wittig, W. “Bio-sequence database scanning on GPU.” IEEE Int. Workshop on High Performance Computational Biology (2006).

[7]Victor, J.D., Purpura, K.P. “Metric-space analysis of spike trains: theory, algorithms and application.” Network. 8:127-64 (1997).

[8]Victor, J.D. “Spike train metrics.” Current Opinion in Neurobiology. 15(5):585-592 (2005).

[9]Victor, J.D., Goldberg, D.H., Gardner, D. “Dynamic programming algorithms for comparing multi-neuronal spike trains via cost-based metrics and alignments.” Journal of Neuroscience Methods. 161(2):351-360 (2007).

[10]CUDA Programming Guide 1.1, http://developer.download.nvidia.com/compute/cuda/1_1/NVIDIA_CUDA_Programming_Guide_1.1.pdf (2007).

[11]CUDA Programming Guide 3.0, http://developer.download.nvidia.com/compute/cuda/3_0/toolkit/docs/NVIDIA_CUDA_ProgrammingGuide.pdf (2010).

[12]NVIDIA Developer: GPU Computing: CUDA Toolkit, http://developer.nvidia.com/object/cuda_3_0_downloads.html (2010).

[13]Spike Train Analysis Toolkit, http://neuroanalysis.org/neuroanalysis/downloads/spike.zip (2010).