基于用户的协同过滤算法的GPU加速.PDF

鍩轰簬鐢ㄦ埛鐨勫崗鍚岃繃婊ょ畻娉 鐨凣PU 鍔犻€ 鐜嬩腑浜氬垬鑾规潹瀹堕獜 涓浗绉戝闄㈠ぇ瀛﹁绠楁満涓庢帶鍒跺闄 澶х翰 鈥 寮曡█ 鈥 鐩稿叧鐮旂┒涓庤儗鏅粙缁 鈥 绠楁硶鍙婄摱棰堝垎鏋 鈥 CUDA浼樺寲绠楁硶鐨勮璁′笌瀹炵幇 鈥 瀹為獙缁撴灉 寮曡█ 鈥 鍗忓悓杩囨护绠楁硶鏈変粈涔堢敤锛屼负浠€涔堣鍔犻€燂紵 鈻 鎺ㄨ崘绯荤粺鐨勪富娴佹柟娉 鈻 闈复涓嶆柇澧為暱鐨勬暟鎹噺 鈻 鎬ц兘鏄噸瑕佺殑鐢ㄦ埛浣撻獙涔嬩竴 鈥 涓轰粈涔堥噰鐢℅PU杩涜鍔犻€燂紵 鈻 鍔犻€熸柟寮忓鏍峰寲(鍒嗗竷寮忚绠楃郴缁熷hadoop,澶欳PU 绛夌瓑) 鈻 GPU鏂规硶浠锋牸鏇翠綆寤 鈻 鎬ц兘鎻愬崌鏇存樉钁 鈻 寮€鍙戠幆澧冭秼浜庢垚鐔燂紙绫籆璇█寮€鍙戯級 鐩稿叧鐮旂┒涓庤儗鏅粙缁 鈥 鍗忓悓杩囨护绠楁硶绠€浠 鈻 鎺ㄨ崘绯荤粺涓绘祦绠楁硶 鈻 鍩烘湰鎬濇兂锛氱敤鎴烽綈蹇冨崗鍔涳紝閫氳繃涓嶆柇鍦板拰缃戠珯浜掑姩锛屼娇鑷繁 鐨勬帹鑽愬垪琛ㄨ兘澶熶笉鏂繃婊ゆ帀鑷繁涓嶆劅鍏磋叮鐨勭墿鍝侊紝瓒婃潵瓒婃弧 瓒宠嚜宸辩殑闇€姹傘€ 鈻 鍒嗕负涓ょ锛 飩 锛?锛夊熀浜庣敤鎴风殑鍗忓悓杩囨护锛圲ser-based Collaborative Filtering 锛 飩 鎺ㄨ崘鏂规硶锛 鈻 a)璁＄畻鐢ㄦ埛涔嬮棿鐨勭浉浼煎害 鈻 b)鏍规嵁閭诲眳锛岄娴嬪晢鍝佽瘎鍒 飩 锛?锛夊熀浜庨」鐩殑鍗忓悓杩囨护锛圛tem-based Collaborative Filtering 锛 飩 鎺ㄨ崘鏂规硶锛 鈻 a)璁＄畻鐗╁搧涔嬮棿鐨勭浉浼煎害 鈻 b)鏍规嵁鐩镐技鍟嗗搧锛岄娴嬪晢鍝佽瘎鍒 鐩稿叧鐮旂┒涓庤儗鏅粙缁 鈥 鐢ㄦ埛璇勫垎鐭╅樀 飩 鍩烘湰鍋囪锛氱浉浼肩殑鐢ㄦ埛瀵瑰悓涓€涓墿鍝佸叿鏈夌浉浼肩殑璇勫垎 鏀堕泦鐢ㄦ埛瀵 鐗╁搧鐨勮瘎鍒 璁＄畻鐢ㄦ埛 鎵惧嚭鐢ㄦ埛 棰勬祴鐢ㄦ埛璇勫垎 浜х敓鎺ㄨ崘缁撴灉 鏈€杩戣窛绂 杩戦偦 鐩稿叧鐮旂┒涓庤儗鏅粙缁 鈥 GPU濡備綍鍔犻€熻繍绠 鈻 CPU 锛氬ぇ閲忔櫠浣撶鐢ㄤ簬cache鍜屾帶鍒剁數璺紝闈㈠悜閫氱敤璁＄畻 鈻 GPU 锛氬ぇ閲忔櫠浣撶鐢ㄤ簬杩愮畻鍗曞厓銆 鈻 GPU鍜孋PU寮傛瀯宸ヤ綔锛岄€傚悎璁＄畻瀵嗛泦鍨嬬畻娉曠殑鍔犻€ 鐩墠鍗曚釜GPU鍙珮杈句笂鍗冧釜鏍?GPU 姣擟PU鏇撮€傚悎骞惰璁＄畻锛 鐩稿叧鐮旂┒涓庤儗鏅粙缁 鈥 CUDA 锛圕ompute Unified Device Architecture 锛孋UDA 锛 鈻 Nvidia鍏徃鎺ㄥ嚭锛屽熀浜嶨PU 鐨勮绠楁灦鏋 鈻 涓や釜灞傛骞惰:grid锛宐lock 鈻 Tile娉曞垝鍒?TileWidth=16 鈻 warp璋冨害鏈哄埗锛屽苟琛岀殑鏈€灏忓崟浣嶄负half-warp 锛?6涓嚎绋嬶級 飩 Memory coalescing 飩 Bank conflicts 绠楁硶鍙婄摱棰堝垎鏋 鈥 杈撳叆鏁版嵁 鈻 A 锛歮*n鍘熷璇勫垎鐭╅樀鏁扮粍, 姣忎竴椤逛负鐢ㄦ埛瀵瑰晢鍝佽瘎鍒 鈻 USER_K 锛氬父閲忥紝鍙朥SER_K涓渶杩戠敤鎴 鈻 ITEM_K 锛氬父閲忥紝鎺ㄨ崘ITEM_K涓晢鍝 鈥 杈撳嚭鏁版嵁 鈻 瀵规寚瀹氱敤鎴风殑ITEM_K涓帹鑽愬晢鍝 绠楁硶鍙婄摱棰堝垎鏋 鈥 涓昏姝ラ涓庣摱棰 鈻 Step 1.璁＄畻鐢ㄦ埛鐩镐技搴︾煩闃礢(m*m) 鈻 Step 2.鎵惧埌鐢ㄦ埛u 鐨刄SER_K涓繎閭 鈻 Step 3.璁＄畻鐢ㄦ埛u 鐨勬墍鏈夊晢鍝佹帹鑽愬害 鈻 Ste