本論文針對GPU (Graphics Processing Unit)上之哼唱選歌 (Query by Singing/Humming, QBSH)系統做出效能上的改良。此系統可將使用者數秒鐘之哼唱片段，與資料庫近兩萬首之歌曲進行比對，並找出最相似之前十名歌曲。
線性伸縮為此系統其中一種比對方法。我們設計了不同於前人之實作方式，在存放資料庫歌曲之音高向量方面，以GPU上之shared memory來取代local memory，使得系統更加貼近GPU之硬體特性，並在資料庫歌曲中加入額外的音高點以減少bank conflict之問題，此改進可將搜尋時間減少32%。
另一方面，為減少多餘的比對，我們也從資料庫中刪除了1265首之重複歌曲以及佔總長約22%之歌曲最長重複片段。
最後，我們也將此系統改良成可支援多張GPU，將比對工作平均分配給各GPU來運算，以最直觀的方式改善效能。當配備N張GPU時，系統可幾乎獲得N倍的加速。

This research proposes a modification to speed up a QBSH (query by singing/humming) system running on a GPU. The system can compare the user's singing or humming with about 20000 songs in the database and retrieve the top-10 most likely candidates.
Linear scaling is one of the comparison methods in our QBSH system. We design a new implementation that uses shared memory instead of local memory for storing pitch vectors of songs in the database so that it is more closely conform to the design of GPU. Furthermore, we add extra pitch points to songs in the database to eliminate bank conflict. After these improvements, we can achieve a 32% reduction in retrieve time.
Besides, in order to reduce redundant comparisons, we remove 1265 repeating songs from the database and the longest repeating pattern of each song, comprising around 22% of the total length of all songs in the database.
Finally, we enhance the system to support multi-GPU so that the work load can be distributed among all GPUs to increase overall efficiency. A speedup factor close to N is achieved when the task is distributed among N GPUs.

摘要 I
Abstract II
謝誌 III
目錄 IV
表目次 VI
圖目次 VII
第一章 緒論 9
1.1 研究主題 9
1.2 相關研究 9
1.3 研究方向及主要成果 10
1.4 章節概要 11
第二章 旋律辨識系統MIRACLE 12
2.1 沿革 12
2.2 CUDA (Compute Unified Device Architecture)簡介 16
2.3 線性伸縮 (Linear Scaling) 23
2.4 動態時間扭曲 (Dynamic Time Warping, DTW) 25
2.5 MIRACLE於CUDA上實作 27
2.5.1 載入歌曲資料庫 27
2.5.2 資料庫歌曲資料前處理 27
2.5.3 使用者輸入音高向量前處理 28
2.5.4 GPU上的比對工作 28
2.5.5 回傳並合併辨識結果 28
第三章 研究方法及實作 29
3.1 線性伸縮於GPU上之執行 29
3.2 線性伸縮之不同實作 30
3.2.1 線性伸縮 (一) 31
3.2.2 線性伸縮 (二) 33
3.2.3 線性伸縮 (三) 35
3.3 刪除資料庫重複歌曲 37
3.3.1 第一階段過濾 37
3.3.2 第二階段過濾 38
3.4 刪除歌曲之重複片段 39
3.5 多GPU之MIRACLE系統 42
第四章 實驗結果與分析討論 44
4.1 實驗使用的測試語料及資料庫 44
4.2 實驗環境設定 44
4.3 不同線性伸縮實作方式之效能分析 45
4.4 GPU數量對效能影響之分析 51
第五章 結論與未來工作 52
5.1 結論 52
5.2 未來工作 53
參考文獻 54

[1] J. Ghias, D. C. Logan, and B. C. Smith, “Query by humming-musical information retrieval in an audio database,” in Proc. ACM Multimedia’95, San Francisco, pp. 216–221, 1995.
[2] R. J. McNab, L. A. Smith, I. H. Witten, C. L. Henderson, and S. J. Cunningham, “Toward the digital music library: Tune retrieval from acoustic input,” in Proc. ACM Digital Libraries, pp. 11–18, 1996.
[3] J.-S. R. Jang, H.-R. Lee, and M.-Y. Kao, “Content-based Music Retrieval Using Linear Scaling and Branch-and-Bound Tree search,” in Proc. of IEEE International Conference on Multimedia and Expo, August 2001.
[4] J.-S. R. Jang, J.-C. Chen, and M.-Y. Kao. “MIRACLE: A Music Information Retrieval System with Clustered Computing Engines,” in Proceedings of the 2nd International Conference on Music Information Retrieval, ISMIR 2001, 2001.
[5] G. Poli, A. L. M. Levada, J. F. Mari, J. H. Satio, “Voice Command Recognition with Dynamic Time Warping (DTW) using Graphics Processing Units (GPU) with Compute Unified Device Architecture (CUDA),” in Proceedings of the 19th International Symposium on Computer Architecture and High Performance Computing , SBAC-PAD 2007, Brazil, pp. 19–25, 2007.
[6] Jun Li, Shuangping Chen, Yanhui Li, “The Fast Evaluation of Hidden Markov Models on GPU,” in IEEE International Conference on Intelligent Computing and Intelligent Systems, Shanghai, vol. 4:426-430, Nov., 2009.
[7] D. Sart, A. Mueen, W. Najjar, E. Keogh, and V. Niennattrakul, “Accelerating Dynamic Time Warping Subsequence Search with GPUs and FPGAs,” in ICDM '10 Proceedings of the 2010 IEEE International Conference on Data Mining, pp. 1001-1006, 2010.
[8] P. Ferraro, P. Hanna, L. Imbert, and T. Izart, “Accelerating Query-by-Humming on GPU,” in Proceedings of the 10th International Conference on Music Information Retrieval, ISMIR 2009, pp. 279–284, 2009.
[9] Chung-Che Wang, Chieh-Hsing Chen, Chin-Yang Kuo, Li-Ting Chiu, and Jyh-Shing Roger Jang, “Accelerating Query by Singing/Humming on GPU: Optimization for Web Deployment,” The 36th International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Kyoto, Japan, March 2012.
[10] Kuo, “Accelerating Query By Singing/Humming on GPU”, National Tsing Hua Univ. 2013
[11] Kao, “A Two-Stage Query by Singing/Humming System on GPU“, National Tsing Hua Univ. 2013
[12] Jang, J.-S Roger, Ming-Yang Kao, “A Query-by-Singing System based on Dynamic Programming,” International Workshop on Intelligent Systems Resolutions (the 8th Bellman Continuum), PP. 85-89, Hsinchu, Taiwan, Dec 2000.
[13] Chieh-Hsing Chen,“Speedup Mechanism for Comparison of Query by Singing/Humming over GPUs,”National Tsing Hua Univ.2012.
[14] 張舒, 褚豔利, 趙開勇, 張鈺勃, “GPU高效能運算之CUDA,”碁峰資訊, 2011
[15] 林俊淵, 周嘉益, 林郁翔, 李昇達, 陳昱蓉, 黃宣穎, 李天齡, “CUDA輕鬆上手 新世代GPU應用技術,” 松崗資訊, 2011
[16] CVG @ ETHZ - GP-GPU: General Purpose Programming on the Graphics Processing Unit, http://www.cvg.ethz.ch/teaching/2011spring/gpgpu/, 2012
[17] NVIDIA, “NVIDIA CUDA C Programming Guide Version 4.2.”
[18] CUDA -维基百科，自由的百科全书, http://zh.wikipedia.org/wiki/CUDA, 2014
[19] Rob Farber, “CUDA APPLICATION DESIGN AND DEVELOPMENT, ”Elsevier Inc., 2011
[20] Borda Count – http://en.wikipedia.org/wiki/Borda_count
[21] Jia-Lien Hsu, Chih-Chin Liu, and Arbee L. P. Chen, “Efficient Repeating Pattern Finding in Music Databases,” International Conference on Information and Knowledge Management (ACM CIKM)，ACM ，pp. 281-288, November 03. 1998.