{"id":400,"date":"2023-03-03T18:28:23","date_gmt":"2023-03-03T10:28:23","guid":{"rendered":"https:\/\/blog.simbot.net\/?p=400"},"modified":"2023-03-03T18:32:58","modified_gmt":"2023-03-03T10:32:58","slug":"knn_search_acceleration_method","status":"publish","type":"post","link":"https:\/\/blog.simbot.net\/2023\/03\/03\/knn_search_acceleration_method\/","title":{"rendered":"knn \u641c\u7d22\u7684\u7ec8\u6781\u52a0\u901f\u65b9\u6cd5\uff0c\u4ece\u6b64\u7b97\u529b\u4e0d\u518d\u865a\u6807\uff01"},"content":{"rendered":"\n

knn \u641c\u7d22\u7684\u7ec8\u6781\u52a0\u901f\u65b9\u6cd5\uff0c\u4ece\u6b64\u7b97\u529b\u4e0d\u518d\u865a\u6807\uff01<\/h1>\n

\u4f5c\u8005\uff1ahws000\uff08hws.000#163.com\uff09
\u58f0\u660e\uff1a\u7248\u6743\u6240\u6709\uff0c\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u4f5c\u8005\u3002
\u51fa\u5904\uff1ahttps:\/\/blog.simbot.net\/2023\/03\/03\/knn_search_acceleration_method\/<\/a><\/p>\n

\u6458\u8981<\/h2>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u672c\u6587\u521b\u9020\u6027\u7684\u63d0\u51fa\u4e00\u79cd\u65b0\u7684knn\uff08k nearest neighbor\uff09\u641c\u7d22\u52a0\u901f\u7b97\u6cd5\uff0c\u53ef\u4ee5\u8ba9\u786c\u4ef6\u7684\u5229\u7528\u7387\u8fbe\u523090%\u4ee5\u4e0a\u3002\u672c\u7b97\u6cd5\u5c06knn\u641c\u7d22\u4ece\u8ba1\u7b97\u5bc6\u96c6\u3001\u5185\u5b58IO\u5bc6\u96c6\u7b97\u6cd5\uff0c\u6539\u8fdb\u4e3a\u7eaf\u8ba1\u7b97\u5bc6\u96c6\u7b97\u6cd5\uff0c\u5145\u5206\u6316\u6398\u786c\u4ef6\u80fd\u529b\uff0c\u53ea\u8981\u5806\u66f4\u591a\u7684\u786c\u4ef6\uff0c\u5c31\u80fd\u5b9e\u73b0\u66f4\u5feb\u7684\u8ba1\u7b97\u3002\u65b0\u7b97\u6cd5\u4e0d\u4f46\u53ef\u4ee5\u8fd0\u884c\u5728GPU\u4e0a\uff0c\u8fd8\u53ef\u4ee5\u9ad8\u6548\u8fd0\u884c\u4e8e\u5404\u79cd\u4e13\u4e1a\u77e9\u9635\u52a0\u901f\u5361\u4e0a\u3002<\/p>\n

1.\u5f15\u8a00<\/h2>\n

\u00a0 \u00a0 \u00a0 \u00a0 knn\u641c\u7d22\u662f\u7cbe\u786e\u641c\u7d22\uff0c\u901a\u5e38\u4f5c\u4e3a\u590d\u6742\u7b97\u6cd5\u7684\u57fa\u7840\uff0c\u5982knn\u56fe\u6784\u5efa[6]\u3001\u805a\u7c7b\u7b97\u6cd5\uff08\u5982k-means\uff09\u7b49\u3002\u8fd9\u4e9b\u7b97\u6cd5\u5e38\u7528\u4e8e\u4eba\u8138\u8bc6\u522b[2]\u3001\u56fe\u50cf\u641c\u7d22[3]\u3001\u6570\u636e\u6316\u6398\u7b49\u573a\u666f\u3002knn\u641c\u7d22\u4e2d\u6709\u4e24\u79cd\u5e38\u7528\u7684\u8ddd\u79bb\u8ba1\u7b97\u65b9\u5f0f\uff1a\u5185\u79ef\u3001\u6b27\u5f0f\u8ddd\u79bb\u3002\u5728\u6279\u91cf\u8ba1\u7b97\u65f6\uff0c\u67e5\u8be2\u6570\u636e\u8868\u793a\u4e3aM\uff0c\u6570\u636e\u5e93\u6570\u636e\u8868\u793a\u4e3aN\uff0c\u6570\u636e\u7ef4\u5ea6\u8868\u793a\u4e3aK\uff0c\u5185\u79ef\u8ddd\u79bb\u8ba1\u7b97\u65b9\u6cd5\u5373MN\uff0c\u6b64\u516c\u5f0f\u4e0e\u77e9\u9635\u4e58\u7b49\u4ef7\uff1b\u6b27\u5f0f\u8ddd\u79bb\u7684\u8ba1\u7b97\u65b9\u6cd5\u4e3aMnorm+Nnorm-2MN\uff0c\u53ef\u4ee5\u53d8\u6362\u4e3a2(Mnorm\/2+Nnorm\/2-MN)\uff0cMnorm\/2\u548cNnorm\/2\u53ef\u4ee5\u63d0\u524d\u8ba1\u7b97\u5e76\u4f5c\u4e3a\u5e38\u91cf\u6570\u636e\u8bfb\u53d6\uff0c\u6b64\u516c\u5f0f\u7b49\u540c\u4e8e\u77e9\u9635\u4e58\u548c\u4e09\u5143\u52a0\u3002\u56e0\u6b64\u4e24\u79cd\u8ddd\u79bb\u8ba1\u7b97\u65b9\u5f0f\u5747\u53ef\u901a\u8fc7\u77e9\u9635\u4e58\u6cd5\u52a0\u901f\u3002knn\u641c\u7d22\u6570\u636e\u96c6\u7684\u7ef4\u5ea6\u901a\u5e38\u4ecb\u4e8e96\uff5e512\u4e4b\u95f4\uff0c\u5f88\u5c11\u6709\u6570\u636e\u96c6\u80fd\u8d85\u8fc7512\u7684\u7ef4\u5ea6\u3002Jia Zhe\u7b49\u4eba\u5206\u6790\u4e86\u77e9\u9635\u4e58\u6cd5\u5728NVidia Turing T4 GPU\u4e0a\u7684\u6267\u884c\u6548\u7387[1]\uff0c\u5728K\u6bd4\u8f83\u4f4e\u65f6\uff0c\u968f\u7740K\u503c\u589e\u5927\u8ba1\u7b97\u6548\u7387\u9010\u6e10\u5347\u9ad8\u3002\u786c\u4ef6\u7684\u8ba1\u7b97\u80fd\u529b\u662f\u56fa\u5b9a\u7684\uff0c\u4e3a\u4ec0\u4e48\u4e0d\u540c\u7684K\u503c\u4f1a\u8868\u73b0\u51fa\u4e0d\u540c\u7684\u8ba1\u7b97\u6548\u7387\uff1f\u4e3b\u8981\u95ee\u9898\u51fa\u5728\u7ed3\u679c\u7684\u8f93\u51fa\u4e0a\uff0c\u76f8\u540c\u7684\u8ba1\u7b97\u91cf\uff08\u53732MNK\uff09\uff0c\u4e0d\u540c\u7684K\u4f1a\u4ea7\u751f\u4e0d\u540c\u7684\u7ed3\u679c\u4e2a\u6570\uff0c\u5185\u5b58\u5e26\u5bbd\u74f6\u9888\u5bfc\u81f4K\u503c\u6bd4\u8f83\u4f4e\u65f6\uff0c\u8ba1\u7b97\u7ed3\u679c\u4e0d\u80fd\u53ca\u65f6\u8f93\u51fa\u3002knn\u641c\u7d22\u4e0d\u4ec5\u8981\u8ba1\u7b97\u8ddd\u79bb\uff0c\u8fd8\u8981\u5bf9\u8ddd\u79bb\u8fdb\u884c\u6392\u5e8f\uff0c\u624d\u80fd\u6700\u7ec8\u5f97\u5230TopK\u4e2a\u7ed3\u679c\uff0c\u800c\u6392\u5e8f\u5fc5\u5b9a\u9700\u8981\u518d\u6b21\u8bfb\u53d6\u8ba1\u7b97\u7ed3\u679c\uff0c\u4f7f\u5185\u5b58\u5e26\u5bbd\u7684\u74f6\u9888\u88ab\u8fdb\u4e00\u6b65\u653e\u5927\u3002Jeff Johnson\u7b49\u4eba\u5c06GPU\u7528\u4e8e\u5411\u91cf\u641c\u7d22[6]\u65f6\uff0cGPU\u7b97\u529b\u8fdc\u4f4e\u4e8e\u73b0\u5728\uff0c\u5185\u5b58\u5e26\u5bbd\u7684\u95ee\u9898\u5e76\u4e0d\u660e\u663e\uff0c\u4f46TensorCore\u51fa\u73b0\u4ee5\u540e\uff0cGPU\u7b97\u529b\u76f4\u63a5\u63d0\u5347\u6570\u500d\uff0c\u800c\u5185\u5b58\u5e26\u5bbd\u7684\u63d0\u5347\u5374\u6709\u9650\u3002\u6211\u4eec\u63d0\u51fa\u4e00\u79cd\u8ba1\u7b97\u8ddd\u79bb\u5e76\u7acb\u5373\u8fc7\u6ee4\u7684\u590d\u5408\u51fd\u6570\uff0c\u5728K\u503c\u8f83\u4f4e\u65f6\u4e5f\u53ef\u4ee5\u6709\u5f88\u9ad8\u7684\u8ba1\u7b97\u6548\u7387\uff0c\u5e76\u4f7f\u7528\u5f88\u4f4e\u7684\u5185\u5b58\u5e26\u5bbd\uff0c\u800c\u4e14\u80fd\u5bf9\u8fc7\u6ee4\u7ed3\u679c\u7684\u6570\u91cf\u8fdb\u884c\u63a7\u5236\uff0c\u4ece\u800c\u63d0\u9ad8\u6392\u5e8f\u6027\u80fd\u3002<\/p>\n

\u672c\u6587\u4f5c\u51fa\u4e86\u4ee5\u4e0b\u8d21\u732e\uff1a
\u63d0\u51fa\u4e00\u79cd\u8ba1\u7b97\u8ddd\u79bb\u5e76\u7acb\u5373\u8fc7\u6ee4\u7684\u590d\u5408\u51fd\u6570\uff0c\u53ea\u8f93\u51fa\u7b26\u5408\u8fc7\u6ee4\u6761\u4ef6\u7684\u7ed3\u679c\uff0c\u51cf\u5c11\u5bf9\u5185\u5b58\u5e26\u5bbd\u7684\u4f9d\u8d56\uff1b
\u8bbe\u8ba1\u4e00\u5957\u9ad8\u6548\u7684\u66f4\u65b0\u8fc7\u6ee4\u9608\u503c\u7684\u65b9\u6cd5\uff0c\u4fdd\u8bc1\u8fc7\u6ee4\u7ed3\u679c\u51c6\u786e\u3001\u53ef\u63a7\uff0c\u540c\u65f6\u51cf\u5c11\u6392\u5e8f\u7684\u8ba1\u7b97\u91cf\uff1b
\u9488\u5bf9\u7ef4\u5ea6\u8f83\u4f4e\u7684\u6570\u636e\uff0c\u63d0\u51fa\u4e00\u79cd\u8ba9\u67e5\u8be2\u6570\u636e\u5e38\u9a7b\u5bc4\u5b58\u5668\u7684\u65b9\u6cd5\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u786c\u4ef6\u5229\u7528\u7387\u3002<\/p>\n

2.\u7814\u7a76\u65b9\u6cd5<\/h2>\n

\u7b97\u529b\u4e0e\u5185\u5b58\u5e26\u5bbd<\/h3>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u5728\u8fdb\u884c\u5185\u79ef\u8ddd\u79bb\u7684knn\u641c\u7d22\u65f6\uff0c\u5047\u8bbe\u6bcf\u6b21\u5904\u7406\u7684\u67e5\u8be2\u91cf\u4e3aM\u6761\uff0c\u6570\u636e\u5e93\u7684\u6570\u636e\u91cf\u4e3aN\u6761\uff0c\u6570\u636e\u7ef4\u5ea6\u4e3aK\uff0c\u8ddd\u79bb\u7684\u8ba1\u7b97\u91cf\u4e3a2MNK FLOPS\uff0c\u8f93\u5165\u3001\u8f93\u51fa\u7c7b\u578b\u5747\u4e3ahalf\uff08FP16 \u5360\u75282\u5b57\u8282\uff09\uff0c\u7ed3\u679c\u8f93\u51fa\u4e3a2MN\u5b57\u8282\uff0c\u8f93\u5165\u6570\u636e\u4e3a2(MK+NK)\u5b57\u8282\u3002\u5f53M\u8fdc\u5927\u4e8eK\uff0c\u5e76\u8fdc\u5c0f\u4e8eN\u65f6\uff0c\u8f93\u5165\u6570\u636e\u8fdc\u5c0f\u4e8e\u8f93\u51fa\u6570\u636e\uff0c\u5185\u5b58\u8bfb\u5199\u91cf\u8fd1\u4f3c\u4e3a2MN\u5b57\u8282\u3002\u5373\u5f53\u8ba1\u7b97\u91cf\u4e3a2MNK FLOPS \u65f6\uff0c\u5185\u5b58\u8bfb\u5199\u91cf\u4e3a2MN\u5b57\u8282\u3002RTX 2080ti\u7684\u5185\u5b58\u5e26\u5bbd\u4e3a0.6TB\/s\uff0c\u5f53K\u4e3a128\u65f6\uff0c\u53ea\u970076.8TFLOPS\u7684\u7b97\u529b\u5c31\u53ef\u4ee5\u4ea7\u751f0.6TB\u7684\u7ed3\u679c(0.6×128=76.8)\uff0c\u800c2080ti\u7684\u7406\u8bba\u7b97\u529b\u662f107.6\uff0c\u6b64\u65f6\u7684\u786c\u4ef6\u5229\u7528\u7387\u4e3a71.4%\u3002\u6570\u636e\u5e93\u7684\u6570\u636e\u89c4\u6a21N\u4e00\u822c\u90fd\u5728\u767e\u4e07\u4ee5\u4e0a\uff0cknn\u641c\u7d22\u65f6TopK\u7684\u53d6\u503c\u4e00\u822c\u57281\uff5e128\u4e4b\u95f4\uff0c\u4f7f\u5f97TopK\u8fdc\u5c0f\u4e8eN\u3002\u56e0\u4e3a\u5927\u591a\u6570\u8ddd\u79bb\u8ba1\u7b97\u7ed3\u679c\u4e0d\u4f1a\u51fa\u73b0\u5728\u6700\u7ec8\u7684\u641c\u7d22\u7ed3\u679c\u91cc\uff0c\u6240\u4ee5\u5bf9\u8ddd\u79bb\u8ba1\u7b97\u7ed3\u679c\u8fdb\u884c\u8fc7\u6ee4\uff0c\u53ef\u4ee5\u5927\u5927\u51cf\u5c0f\u8f93\u51fa\u7684\u5b57\u8282\u6570\u3002<\/p>\n

\u9608\u503c\u7684\u4ea7\u751f\u4e0e\u66f4\u65b0<\/h3>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u5047\u8bbe\u6570\u636e\u662f\u968f\u673a\u5206\u5e03\u7684\uff0c\u641c\u7d22\u7ed3\u679c\u4e5f\u4f1a\u8fd1\u4f3c\u5747\u5300\u7684\u5206\u6563\u5728N\u4e2d\u3002\u53ef\u4ee5\u5c06N\u5206\u4e3a\u591a\u6bb5\uff0c\u9010\u6bb5\u641c\u7d22\uff0c\u5e76\u4f7f\u6bcf\u4e2a\u5206\u6bb5\u5927\u5c0f\u5747\u7b49\u4e8e\u5df2\u7ecf\u641c\u7d22\u7684\u6570\u636e\u91cf\u3002\u5f53N\u4e3a131072\uff0cTopK\u4e3a128\u65f6\uff0c\u5206\u6bb5\u5927\u5c0fn\u53ef\u4ee5\u8bbe\u4e3a\u4ee5\u4e0b\u503c\uff1a<\/p>\n\n\n\n\n
\u5206\u6bb5<\/td>\n0<\/td>\n1<\/td>\n2<\/td>\n3<\/td>\n4<\/td>\n5<\/td>\n6<\/td>\n7<\/td>\n8<\/td>\n<\/tr>\n
n<\/td>\n512<\/td>\n512<\/td>\n1024<\/td>\n2048<\/td>\n4096<\/td>\n8192<\/td>\n16384<\/td>\n32768<\/td>\n65536<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u88681<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u5904\u7406\u5206\u6bb50\u65f6\uff0c\u6ca1\u6709\u8fc7\u6ee4\u9608\u503c\uff0c\u5bf9n\u6761\u6570\u636e\u8fdb\u884c\u8ddd\u79bb\u8ba1\u7b97\u3001\u6392\u5e8f\uff0c\u4fdd\u5b58\u524dTopK\u4e2a\u7ed3\u679c\uff0c\u5c06\u6700\u540e\u4e00\u4e2a\u7ed3\u679c\u7684\u8ddd\u79bb\u503c\u8bbe\u7f6e\u4e3a\u9608\u503cT\u3002\u5904\u7406\u5206\u6bb51\uff5eS\u65f6\uff0c\u4f7f\u7528\u9608\u503cT\u5bf9\u8ddd\u79bb\u8ba1\u7b97\u7ed3\u679c\u8fdb\u884c\u8fc7\u6ee4\uff0c\u5c06\u65b0\u8fc7\u6ee4\u7684\u7ed3\u679c\u4e0e\u65e7\u7684\u7ed3\u679c\u5408\u5e76\u3001\u6392\u5e8f\uff0c\u7528\u7b2cTopK\u4e2a\u7ed3\u679c\u7684\u8ddd\u79bb\u503c\u66f4\u65b0\u9608\u503cT\uff0c\u91cd\u590d\u6b64\u6b65\u9aa4\uff0c\u76f4\u5230\u6240\u6709\u6570\u636e\u5904\u7406\u5b8c\u6bd5\u3002\u7531\u4e8e\u641c\u7d22\u7ed3\u679c\u5206\u5e03\u5747\u5300\uff0c\u524dn\u6b21\u8fc7\u6ee4\u5f97\u5230TopK\u4e2a\u7ed3\u679c\uff0c\u518d\u8fc7\u6ee4n\u6b21\uff0c\u7ed3\u679c\u4e2a\u6570\u4e5f\u5e94\u8be5\u8ddfTopK\u63a5\u8fd1\u3002\u56e0\u4e3a\u8fc7\u6ee4\u7684\u7ed3\u679c\u6570\u4e0d\u4f1a\u5b8c\u5168\u7b49\u4e8eTopK\uff0c\u53ef\u80fd\u4f1a\u5927\u4e8e\uff0c\u4e5f\u53ef\u80fd\u4f1a\u5c0f\u4e8e\uff0c\u5e76\u4e14\u9700\u8981\u4fdd\u7559\u4e0a\u6b21\u6392\u5e8f\u7684\u7ed3\u679c\uff0c\u6240\u4ee5\u7ed3\u679c\u7684\u5b58\u50a8\u7a7a\u95f4\u5e94\u8be5\u5927\u4e8e2TopK\uff0c\u5982\u8bbe\u7f6e\u4e3a4TopK\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u7531\u4e8e\u6bcf\u4e2a\u5206\u6bb5\u5747\u53ef\u8fc7\u6ee4\u51faTopK\u4e2a\u7ed3\u679c\uff0c\u4f9d\u6b21\u5904\u7406\u5b8c\u6240\u6709\u5206\u6bb5\u624d\u80fd\u5f97\u5230\u6700\u7ec8\u7ed3\u679c\uff0c\u9020\u6210\u6392\u5e8f\u8d1f\u62c5\u8f83\u91cd\u3002\u5728\u5b9e\u9645\u4f7f\u7528\u65f6\uff0c\u53ef\u4ee5\u5728\u521d\u671f\u4f7f\u7528\u8f83\u5c11\u7684\u8fc7\u6ee4\u76ee\u6807\u6570\uff0c\u5e76\u9010\u6b65\u589e\u52a0\u5230TopK\u3002\u5f53TopK\u4e3a512\u65f6\uff0c\u8fc7\u6ee4\u76ee\u6807\u6570\u3001\u83b7\u53d6\u9608\u503cT\u7684\u4f4d\u7f6e\u7b49\u53c2\u6570\u53ef\u4ee5\u4f7f\u7528\u88682\u4e2d\u7684\u503c\u3002\u5176\u4e2d\u5206\u6bb54\u30016\u30018\u300110\u4e0e\u5206\u6bb53\u30015\u30017\u30019\u7684\u76ee\u6807\u6570\u76f8\u540c\uff0c\u5c5e\u4e8e\u77eb\u6b63\u5206\u6bb5\uff0c\u662f\u4e3a\u4e86\u9632\u6b62\u4e0a\u4e00\u4e2a\u5206\u6bb5\u6ca1\u6709\u8fc7\u6ee4\u51fa\u76ee\u6807\u6570\u91cf\u7684\u7ed3\u679c\uff0c\u5bfc\u81f4\u8fc7\u6ee4\u9608\u503cT\u6301\u7eed\u4e0d\u66f4\u65b0\u3002\u4e3a\u4e86\u83b7\u5f97\u66f4\u5feb\u7684\u6392\u5e8f\u6548\u679c\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u51cf\u5c11\u77eb\u6b63\u5206\u6bb5\u7684\u6bd4\u4f8b\u3002<\/p>\n\n\n\n\n\n\n
\u5206\u6bb5<\/td>\n0<\/td>\n1<\/td>\n2<\/td>\n3<\/td>\n4<\/td>\n5<\/td>\n6<\/td>\n7<\/td>\n8<\/td>\n9<\/td>\n10<\/td>\n<\/tr>\n
n<\/td>\n128<\/td>\n128<\/td>\n256<\/td>\n512<\/td>\n1024<\/td>\n2048<\/td>\n4096<\/td>\n8192<\/td>\n16384<\/td>\n32768<\/td>\n65536<\/td>\n<\/tr>\n
\u76ee\u6807\u6570<\/td>\n—<\/td>\n32<\/td>\n32<\/td>\n64<\/td>\n64<\/td>\n128<\/td>\n128<\/td>\n256<\/td>\n256<\/td>\n512<\/td>\n512<\/td>\n<\/tr>\n
T\u4f4d\u7f6e<\/td>\n32<\/td>\n32<\/td>\n64<\/td>\n64<\/td>\n128<\/td>\n128<\/td>\n256<\/td>\n256<\/td>\n512<\/td>\n512<\/td>\n—<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u88682<\/p>\n

\u9700\u8981\u6ce8\u610f\u7684\u662f\uff1a<\/strong>\u4e3a\u786e\u4fdd\u6570\u636e\u968f\u673a\u5206\u5e03\uff0c\u53ef\u80fd\u8981\u9884\u5148\u968f\u673a\u8c03\u6362\u6bcf\u6761\u6570\u636e\u7684\u5b58\u653e\u4f4d\u7f6e\uff0c\u5426\u5219\u4f1a\u5bfc\u81f4\u8fc7\u6ee4\u51fa\u7684\u7ed3\u679c\u6570\u4e0d\u53ef\u63a7\u3002<\/p>\n

\u67e5\u8be2\u6570\u636e\u5e38\u9a7b\u5bc4\u5b58\u5668<\/h3>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u5728CUTLASS\u4ecb\u7ecd[8]\u548cmma\u6307\u4ee4\u6587\u6863[11]\u4e2d\uff0c\u8be6\u7ec6\u8bf4\u660e\u4e86TensorCore\u7684\u4f7f\u7528\u65b9\u5f0f\u3002\u4e00\u6761mma\u6307\u4ee4\u9700\u8981\u6574\u4e2awarp\u4e2d\u768432\u4e2athread\u534f\u4f5c\u5b8c\u6210\uff0c\u6bcf4\u4e2athread\u6210\u4e3a\u4e00\u7ec4\uff0c\u6bcf\u7ec4thread\u7684\u6570\u636e\u7ec4\u5408\u4e3a\u4e00\u4e2am\u6216n\uff0c\u5171\u5206\u4e3a8\u7ec4\uff0c\u4f7f\u6bcf\u6761\u6307\u4ee4\u4e00\u6b21\u81f3\u5c11\u5904\u74068\u4e2am\u6216n\u3002\u8f93\u5165\u3001\u8f93\u51fa\u7684\u6570\u636e\u7c7b\u578b\u5747\u4e3ahalf\u65f6\uff0c\u6bcf\u4e2a\u5bc4\u5b58\u5668\u5b58\u50a8\u7684\u6570\u636e\u4e2a\u6570nk=2\uff0c4\u4e2athread\u603b\u5171\u5b58\u50a8\u7684k=8\u3002\u4e00\u6761mma\u6307\u4ee4\u4e00\u6b21\u5904\u74068\u4e2am\u548cn\u65f6\uff0c\u4ea7\u751f8*8=64\u4e2a\u7ed3\u679c\uff0c\u6bcf\u4e2athread\u5b58\u50a8\u7684\u7ed3\u679c\u4e2a\u6570\u4e3a2\uff0c\u4f7f\u7528\u7684\u5bc4\u5b58\u5668\u4e2a\u6570\u4e3a1\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u5f53\u4e00\u4e2awarp\u5b8c\u6210M=64\uff0cN=64\uff0cK=128\u7684\u4e58\u6cd5\u65f6\uff0c\u901a\u7528\u65b9\u6cd5\u662f\u5c06\u6b65\u957fk\u8bbe\u7f6e\u4e3a8\uff0c\u9010\u6b65\u52a0\u8f7dM\u3001N\uff0c\u8ba1\u7b97\u4e58\u79ef\uff0c\u5e76\u4fdd\u5b58M*N\u4e2a\u7ed3\u679c\u3002\u5f53\u4e00\u6761mma\u6307\u4ee4\u4e00\u6b21\u53ea\u80fd\u5904\u74068\u4e2am\u548cn\u65f6\uff0c\u9700\u8981\u5bf9M\u3001N\u8fdb\u884c\u5207\u5757\uff0cM\u548cN\u7684\u5207\u5757\u4e2a\u6570\u5206\u522b\u4e3aTileM=M\/8=8\uff0cTileN=N\/8=8\uff0c\u6bcf\u4e2athread\u4f7f\u7528\u7684\u5bc4\u5b58\u5668\u4e2a\u6570\u5206\u522b\u4e3aA=TileM*1=8\u3001B=TileN*1=8\u3001C=TileM*TileN*1=64\u3002\u5b58\u50a8\u6570\u636e\u7528\u7684\u5bc4\u5b58\u5668\u603b\u6570\u4e3a8+8+64=80\u3002\u8fd9\u662f\u4e00\u4e2a\u901a\u7528\u65b9\u6cd5\u7684\u7b80\u7565\u63cf\u8ff0\uff0c\u5b9e\u9645\u4f7f\u7528\u65f6\u4f1a\u7528\u66f4\u590d\u6742\u7684\u64cd\u4f5c\u6b65\u9aa4\uff0c\u6765\u63d0\u9ad8\u6027\u80fd\uff0c\u5177\u4f53\u7684\u5185\u5bb9\u53ef\u4ee5\u5728cuda-samples\u7684cudaTensorCoreGemm\u7b49\u793a\u4f8b\u4e2d\u67e5\u770b\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u57284\u4e2athread\u4e2d\u5b58\u50a8\u5b8c\u6574\u7684m\u3001n\u65f6\uff0c\u6bcf\u4e2athread\u8981\u4f7f\u7528\u7684\u5bc4\u5b58\u5668\u4e2a\u6570\u4e3aK\/nk\/4=16\u3002\u5f53M\u6709\u591a\u4e2a\u5207\u5757\u65f6\uff0c\u4f7f\u7528\u7684\u5bc4\u5b58\u5668\u4e2a\u6570\u4e3aK\/nk\/4*TileM=128\u3002N\u53ef\u4ee5\u4f7f\u7528\u6700\u5c0f\u503c8\u9010\u6b65\u52a0\u8f7d\uff0c\u4f7f\u7528\u7684\u5bc4\u5b58\u5668\u4e2a\u6570\u4e3aK\/nk\/4=16\uff0c\u5b58\u50a8\u7ed3\u679c\u7684\u5bc4\u5b58\u5668\u4e2a\u6570 C=TileM*1*1=8\u3002\u5b58\u50a8\u6570\u636e\u7528\u7684\u5bc4\u5b58\u5668\u603b\u6570\u4e3a128+16+8=152\u3002\u901a\u8fc7\u8c03\u8282TileM\u7684\u5927\u5c0f\uff0c\u53ef\u4ee5\u9002\u7528\u4e0d\u540c\u7684\u8f83\u5c0f\u7684K\u503c\uff0c\u4ece\u800c\u4f7f\u67e5\u8be2\u6570\u636eM\u5e38\u9a7b\u5bc4\u5b58\u5668\u3002<\/p>\n

\u4f7f\u67e5\u8be2\u6570\u636eM\u5e38\u9a7b\u5bc4\u5b58\u5668\u6709\u5982\u4e0b\u4f18\u70b9\uff1a<\/strong>
\u4f7f\u6570\u636e\u52a0\u8f7d\u6307\u4ee4\u6570\u51cf\u534a\uff0c\u6bcf\u6b21\u53ea\u9700\u8981\u52a0\u8f7d\u6570\u636e\u5e93\u7684\u6570\u636eN\u5373\u53ef\uff1b
\u63d0\u9ad8\u5404\u7ea7cache\u7684\u5229\u7528\u7387\uff0ccache\u4e2d\u53ea\u9700\u5b58\u50a8N\u548c\u7ed3\u679c\u76f8\u5173\u7684\u6570\u636e\uff1b
\u4e0d\u9700\u8981\u540c\u6b65\u64cd\u4f5c\uff0c\u53ef\u4ee5\u8ba9\u4e0d\u540cwarp\u7684\u8ba1\u7b97\u6307\u4ee4\u4e0e\u8fc7\u6ee4\u6307\u4ee4\u5e76\u53d1\u6267\u884c\uff0c\u73b0\u6709\u7684\u901a\u7528\u65b9\u6cd5\u5728\u52a0\u8f7dM\u3001N\u6570\u636e\u65f6\uff0c\u9700\u8981\u8fdb\u884c\u540c\u6b65\uff0c\u624d\u80fd\u6709\u6bd4\u8f83\u597d\u7684\u52a0\u8f7d\u6548\u7387\u3002<\/p>\n

3.\u7ed3\u679c\u4e0e\u8ba8\u8bba<\/h2>\n

\u00a0 \u00a0 \u00a0 \u00a0 RTX 2080ti\u662fNvidia Turning\u67b6\u6784\u7684GPU\uff0cRTX 3060\u662fNvidia Ampere\u67b6\u6784\u7684GPU\uff0c\u8fd9\u4e24\u6b3eGPU\u5747\u53ef\u4ee5\u4f7f\u7528TensorCore\u8fdb\u884c\u77e9\u9635\u8ba1\u7b97\u52a0\u901f\u3002cuBLAS\u662fNvidia\u63d0\u4f9b\u7684\u52a0\u901f\u77e9\u9635\u8ba1\u7b97\u7684\u51fd\u6570\u5e93\uff0c\u5305\u542bcublasHgemm\u3001cublasSgemmEx\u3001cublasGemmEx\u7b49\u51fd\u6570\uff0c\u8fd9\u4e9b\u51fd\u6570\u7684\u8f93\u5165\u3001\u8f93\u51fa\u7c7b\u578b\uff0c\u4ee5\u53ca\u5728RTX 2080ti\u3001RTX 3060\u4e0a\u7684\u7406\u8bba\u7b97\u529b\uff08\u5355\u4f4d\uff1aTFLOPS\uff09\u5982\u4e0b\u8868\uff1a<\/p>\n\n\n\n\n\n\n
\u51fd\u6570\u540d\u79f0<\/td>\n\u8f93\u5165<\/td>\n\u8f93\u51fa<\/td>\nRTX 2080ti \u7b97\u529b<\/td>\nRTX 3060 \u7b97\u529b<\/td>\n<\/tr>\n
cublasSgemmEx<\/td>\nhalf<\/td>\nfloat<\/td>\n53.8<\/td>\n25.6<\/td>\n<\/tr>\n
cublasHgemm<\/td>\nhalf<\/td>\nhalf<\/td>\n107.6<\/td>\n51.2<\/td>\n<\/tr>\n
cublasGemmEx<\/td>\nint8<\/td>\nint<\/td>\n215.2<\/td>\n102.4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u88683<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 Faiss\u662fFacebook\u5f00\u6e90\u7684\u7528\u4e8e\u9ad8\u6548\u76f8\u4f3c\u6027\u641c\u7d22\u548c\u5bc6\u96c6\u5411\u91cf\u805a\u7c7b\u7684\u5e93\uff0c\u6211\u4eec\u6d4b\u8bd5\u7684\u7248\u672c\u662fv1.7.1\u3002\u5176\u4e2d\u7684Flat\u7b97\u6cd5\u4f7f\u7528cublasSgemmEx\u51fd\u6570\u8fdb\u884c\u77e9\u9635\u8ba1\u7b97\u52a0\u901f\uff0c\u5e76\u901a\u8fc7\u4e13\u95e8\u4f18\u5316\u8fc7\u7684\u6392\u5e8f\u51fd\u6570\u6392\u5e8f\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u6211\u4eec\u6d4b\u8bd5\u4e86\u4ee5\u4e0a\u51fd\u6570\uff0c\u4ee5\u53ca\u6211\u4eec\u5b9e\u73b0\u7684knn\u641c\u7d22\u51fd\u6570\uff08\u8f93\u5165\u3001\u8f93\u51fa\u4e0ecublasHgemm\u3001cublasSgemmEx\u76f8\u540c\uff09\uff0c\u5728K\u503c\u4e3a128\uff5e1024\u65f6\u7684\u6548\u7387\u3002\u8ddd\u79bb\u8ba1\u7b97\u65b9\u5f0f\u4e3a\u5185\u79ef\u65f6\uff0c\u7b97\u529b=2MNK\/\u8017\u65f6\uff0c\u5176\u4e2dknn\u641c\u7d22\u7684\u8017\u65f6\u5305\u542b\u6392\u5e8f\u7684\u65f6\u95f4\u3002knn\u641c\u7d22\u9ed8\u8ba4\u4f7f\u7528\u7684TopK\u4e3a16\uff0cFaiss\u7684Flat\u7b97\u6cd5\u4f7f\u7528\u7684TopK\u4e3a1\uff0c\u5e76\u4e14\u4e13\u95e8\u6d4b\u8bd5\u4e86TopK\u4e3a1\uff5e1024\u65f6\u4e24\u4e2a\u7b97\u6cd5\u7684\u6548\u7387\u3002\u6211\u4eec\u7684knn\u641c\u7d22\u7b97\u6cd5\u4f7f\u7528\u4e0e\u88682\u76f8\u540c\u7684\u65b9\u5f0f\u589e\u52a0\u8fc7\u6ee4\u76ee\u6807\u6570\u3001\u66f4\u65b0\u9608\u503c\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u6211\u4eec\u4f7f\u7528\u7684\u6392\u5e8f\u7b97\u6cd5\u662f\u53cc\u8c03\u6392\u5e8f\uff0c\u4ee3\u7801\u53c2\u8003cuda-samples\u7684bitonicSort\u7b97\u6cd5\u3002\u6211\u4eec\u5bf9\u6392\u5e8f\u7b97\u6cd5\u505a\u4e862\u5904\u6539\u8fdb\uff1a\u6392\u5e8f\u8fc7\u7a0b\u5b9e\u9645\u662f\u8fc7\u6ee4\u51fa\u7684\u65b0\u6570\u636e\u548c\u65e7\u6570\u636e\u7684\u5408\u5e76\u8fc7\u7a0b\uff0c\u65e7\u6570\u636e\u662f\u6392\u5e8f\u8fc7\u7684\u6570\u636e\uff0c\u6240\u4ee5\u4e0d\u9700\u8981\u91cd\u590d\u6392\u5e8f\uff1b\u5f53\u8fc7\u6ee4\u51fa\u7684\u6570\u636e\u591a\u4e8e\u76ee\u6807\u6570\u636e\u65f6\uff0c\u53c2\u4e0e\u6392\u5e8f\u7684\u6570\u636e\u91cf\u5927\u4e8e\u8981\u6c42\u7684\u6570\u91cf\uff0c\u6b64\u65f6\u53ea\u9700\u4fdd\u8bc1\u8981\u6c42\u6570\u91cf\u7684\u6392\u5e8f\u7ed3\u679c\u51c6\u786e\uff0c\u8d85\u51fa\u90e8\u5206\u4e0d\u9700\u8fdb\u884c\u5b8c\u6574\u7684\u6392\u5e8f\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 Deep1B[5]\u662fbig ann benchmarks\u4f7f\u7528\u7684\u6570\u636e\u96c6\u4e4b\u4e00\uff0c\u6211\u4eec\u540c\u6837\u4f7f\u7528\u5b83\u4f5c\u4e3a\u4e3b\u6570\u636e\u96c6\uff0c\u5b83\u7684\u9ed8\u8ba4\u7ef4\u5ea6\u662f96\uff0c\u6570\u636e\u7c7b\u578bfloat\u3002\u6211\u4eec\u5c06\u5b83\u4f5c\u4e3a\u5404\u79cd\u7ef4\u5ea6\u7684\u6570\u636e\u6765\u6e90\uff0c\u5982\u8fde\u7eed\u8bfb\u53d6128\u3001256\u4e2afloat\u6570\u636e\uff0c\u8f6c\u6362\u4e3ahalf\u7c7b\u578b\u6216\u8005\u91cf\u5316\u4e3aint8\u7c7b\u578b\uff0c\u7136\u540e\u4f5c\u4e3a128\u3001256\u7ef4\u5ea6\u7684\u6570\u636e\u4f7f\u7528\u3002\u8fd9\u6837\u5904\u7406\u4f1a\u5bfc\u81f4\u8fd9\u4e9b\u6570\u636e\u5931\u53bb\u672c\u8eab\u7684\u542b\u4e49\uff0c\u4f46\u8fd9\u6837\u505a\u5bf9\u5355\u7eaf\u7684\u6027\u80fd\u6d4b\u8bd5\u6ca1\u6709\u5f71\u54cd\u3002\u6211\u4eec\u540c\u6837\u6d4b\u8bd5\u4e86\u5176\u5b83\u7684\u4e00\u4e9b\u6570\u636e\u96c6\uff0c\u5982gist\u3001glove\u7b49\u3002<\/p>\n

\"\"<\/p>\n

Figure 1\uff1a\u901a\u8fc7\u4ee5\u4e0a\u6d4b\u8bd5\u7ed3\u679c\u53ef\u4ee5\u770b\u5230\uff0c\u5728K\u503c\u8f83\u4f4e\u65f6\u6211\u4eec\u7684\u7b97\u6cd5\u4e5f\u6709\u5f88\u9ad8\u7684\u786c\u4ef6\u5229\u7528\u7387\uff0c\u5f53\u67e5\u8be2\u6570\u636e\u53ef\u4ee5\u5e38\u9a7b\u5bc4\u5b58\u5668\u65f6\u5229\u7528\u7387\u66f4\u9ad8\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u4f7f\u7528Nsight Compute\u5de5\u5177\u5bf9\u8ba1\u7b97\u4e0e\u8fc7\u6ee4\u51fd\u6570\u8fdb\u884c\u6027\u80fd\u5206\u6790\uff0c2080ti\u7684\u786c\u4ef6\u5229\u7528\u7387\u6700\u9ad8\u4e3a91.53%\uff0c\u6b64\u65f6TensorCore\u5229\u7528\u7387\u4e3a86.94%\uff0c3060\u7684\u786c\u4ef6\u5229\u7528\u7387\u4e0eTensorCore\u5229\u7528\u7387\u76f8\u540c\uff0c\u6700\u9ad8\u4e3a96%\u3002\u5728\u5206\u6790\u5de5\u5177\u7684Pipe Shared Cycles Active\u7684\u63d0\u793a\u4e2d\u53ef\u4ee5\u770b\u5230\uff0c2080ti\u7684fp16\u4e0etensor\u5171\u4eab\u4e00\u4e2apipe\uff0c\u5bfc\u81f4\u8fc7\u6ee4\u6307\u4ee4\u548c\u8ba1\u7b97\u6307\u4ee4\u4ea7\u751f\u7ade\u4e89\uff0c\u8fd9\u53ef\u80fd\u662fTensorCore\u7684\u5229\u7528\u7387\u4f4e\u4e8e\u786c\u4ef6\u5229\u7528\u7387\u7684\u539f\u56e0\u3002\u7531\u4e8e\u6b27\u6c0f\u8ddd\u79bb\u9700\u8981\u5bf9\u77e9\u9635\u4e58\u6cd5\u7684\u7ed3\u679c\u505a\u8fdb\u4e00\u6b65\u7684\u8ba1\u7b97\uff0c\u624d\u80fd\u8f6c\u6362\u4e3a\u8ddd\u79bb\uff0c\u56e0\u6b64\u7f16\u8bd1\u51fa\u7684\u7a0b\u5e8f\u548c\u5185\u79ef\u8ddd\u79bb\u6709\u8f83\u5927\u5dee\u5f02\uff0c\u8fd9\u4e9b\u5dee\u5f02\u5bfc\u81f4\u90e8\u5206\u7ef4\u5ea6\u65f6\u6b27\u5f0f\u8ddd\u79bb\u7684\u8ba1\u7b97\u6548\u7387\u66f4\u9ad8\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u5728\u5bf9M\u4e2a\u67e5\u8be2\u6570\u636e\u8fdb\u884c\u641c\u7d22\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6bcf\u4e2a\u5206\u6bb5\u5747\u4f1a\u4ea7\u751fM\u7ec4\u8fc7\u6ee4\u7ed3\u679c\uff0c\u6bcf\u7ec4\u7ed3\u679c\u7684\u4e2a\u6570\u4e3ax\uff0cM\u4e2ax\u7ec4\u5408\u6210\u96c6\u5408X\uff0cx\u7684\u5e73\u5747\u503c\u8fd1\u4f3c\u7b49\u4e8e\u8fc7\u6ee4\u76ee\u6807\u6570\u3002\u5bf9\u96c6\u5408X\u8fdb\u884c\u5206\u6790\u540e\u5f97\u5230\u5982\u4e0b\u7ed3\u679c\uff1a\u968f\u673a\u53d8\u91cfx\u670d\u4ece\u81ea\u7531\u5ea6\u4e3atk\u7684\u5361\u65b9\u5206\u5e03\uff0c\u5176\u4e2dtk\u8fd1\u4f3c\u7b49\u4e8e\u8fc7\u6ee4\u76ee\u6807\u6570\uff0c\u4e0e\u6570\u636e\u7ef4\u5ea6\u7b49\u53c2\u6570\u65e0\u5173\u3002<\/strong>\u5361\u65b9\u5206\u5e03\u662f\u4e00\u79cd\u7279\u6b8a\u7684Gamma\u5206\u5e03\uff0c\u6b64\u65f6Gamma\u5206\u5e03\u7684\u5f62\u72b6\u53c2\u6570\u03b1=tk\/2\u3001\u5c3a\u5ea6\u53c2\u6570\u03b2=2\u3002<\/p>\n

\"\"<\/p>\n

\"\"Figure 2\uff1a\u4e0a\u90e8\u622a\u56fe\u4e3a\u641c\u7d22\u7b97\u6cd5\u7684\u4e00\u5f20\u6267\u884c\u622a\u56fe\uff0c\u6570\u636e\u5e93\u5927\u5c0f209\u4e07\uff0c\u67e5\u8be2\u91cf21504\uff0c\u6570\u636e\u7ef4\u5ea696\uff0c\u6570\u636e\u7c7b\u578bhalf\uff0cTopK\u4e3a128\uff0c\u5faa\u73af16\u6b21\u7ed3\u675f\u3002\u7b2c1\u30012\u5217\u662f\u8ddd\u79bb\u8ba1\u7b97\u4e0e\u8fc7\u6ee4\u51fd\u6570\u7684\u6267\u884c\u65f6\u95f4\uff08\u5355\u4f4d\uff1ams\uff09\u548c\u76f8\u5e94\u7b97\u529b\uff08\u5355\u4f4d\uff1aTFLOPS\uff09\uff1b\u7b2c3\u30014\u30015\u5217\u5206\u522b\u662f\u7ed3\u679c\u89e3\u6790\u51fd\u6570\u3001\u6392\u5e8f\u51fd\u6570\u3001\u9608\u503c\u66f4\u65b0\u51fd\u6570\u7684\u6267\u884c\u65f6\u95f4\uff1b\u7b2c6\u5217\u662f\u5206\u6bb5\u5927\u5c0f\uff1b\u7b2c7\u30018\u5217\u662f\u8fc7\u6ee4\u7ed3\u679c\u6570\u7684\u5747\u503c\u548c\u6807\u51c6\u5dee\uff1b\u7b2c9\u300110\u5217\u662f\u7531\u5747\u503c\u3001\u6807\u51c6\u5dee\u8ba1\u7b97\u51fa\u7684Gamma\u5206\u5e03\u7684\u5f62\u72b6\u53c2\u6570\u03b1\u3001\u5c3a\u5ea6\u53c2\u6570\u03b2\uff1b\u7b2c11\u300112\u5217\u662f\u8fc7\u6ee4\u7ed3\u679c\u6570\u7684\u6700\u5927\u503c\u3001\u6700\u5c0f\u503c\uff1b\u7b2c13\u5217\u662f\u8fc7\u6ee4\u76ee\u6807\u6570\u3002\u6700\u540e\u4e00\u884c\u662f\u6574\u4e2a\u67e5\u8be2\u7684\u8017\u65f6\u548c\u76f8\u5e94\u7b97\u529b\u3002<\/p>\n

Figure 3\uff1a\u5de6\u4fa7\u84dd\u8272\u5b9e\u7ebf\u662ftk=32\u7684\u6982\u7387\u5bc6\u5ea6\u66f2\u7ebf\uff0c\u84dd\u8272\u5706\u70b9\u662f\u76ee\u6807\u6570\u4e3a32\u7684\u8fc7\u6ee4\u7ed3\u679c\u6570\u7684\u7edf\u8ba1\u503c\uff1b\u6a59\u8272\u5b9e\u7ebf\u662ftk=64\u7684\u6982\u7387\u5bc6\u5ea6\u66f2\u7ebf\uff0c\u6a59\u8272\u5706\u70b9\u662f\u76ee\u6807\u6570\u4e3a64\u7684\u8fc7\u6ee4\u7ed3\u679c\u6570\u7684\u7edf\u8ba1\u503c\uff1b\u7eff\u8272\u5b9e\u7ebf\u662ftk=128\u7684\u6982\u7387\u5bc6\u5ea6\u66f2\u7ebf\uff0c\u7eff\u8272\u5706\u70b9\u662f\u76ee\u6807\u6570\u4e3a128\u7684\u8fc7\u6ee4\u7ed3\u679c\u6570\u7684\u7edf\u8ba1\u503c\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u7531\u4e8e\u9996\u4e2a\u5206\u6bb5\u662f\u5b8c\u6574\u7684\u8f93\u51fa\u3001\u6392\u5e8f\uff0c\u5e76\u5b58\u50a8\u5230\u7ed3\u679c\u961f\u5217\uff0c\u56e0\u6b64\u7ed3\u679c\u961f\u5217\u6c38\u4e0d\u4e3a\u7a7a\uff0c\u6240\u4ee5\u521d\u59cb\u8fc7\u6ee4\u76ee\u6807\u6570\u8303\u56f4\u5185\u7684\u53ec\u56de\u7387\u4e3a1\u3002\u5f53\u4f7f\u7528\u88682\u7684\u65b9\u5f0f\u589e\u52a0\u8fc7\u6ee4\u76ee\u6807\u6570\u65f6\uff0c\u7531\u4e8e\u591a\u4e2a\u77eb\u6b63\u5206\u6bb5\u7684\u5b58\u5728\uff0c\u4f7fTopK\u7684\u53ec\u56de\u7387\u4e5f\u63a5\u8fd1\u4e8e1\u3002\u53ef\u4ee5\u901a\u8fc7\u7ed3\u679c\u961f\u5217\u7684\u586b\u5145\u60c5\u51b5\u63a8\u7b97\u53ec\u56de\u7387\uff0c\u5f53\u586b\u5145\u4f4d\u7f6e\u8fbe\u5230\u6216\u8d85\u8fc7TopK\u65f6\uff0c\u53ec\u56de\u7387\u4e3a1\u3002\u901a\u8fc7\u63a7\u5236\u8fc7\u6ee4\u7684\u7ed3\u679c\u6570\uff0c\u53ef\u4ee5\u51cf\u5c0f\u8f93\u51fa\u5e26\u5bbd\uff0c\u540c\u65f6\u4e5f\u51cf\u5c0f\u4e86\u6392\u5e8f\u7684\u6b21\u6570\u548c\u6bcf\u6b21\u6392\u5e8f\u7684\u8ba1\u7b97\u91cf\uff0c\u4f7f\u786c\u4ef6\u7684\u6027\u80fd\u5f97\u5230\u5145\u5206\u5229\u7528\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 Felix Chern\u7b49\u4eba\u5728Google TPU\u4e0a\u5b9e\u73b0\u4e86\u5cf0\u503c\u6027\u80fd\u7684knn\u641c\u7d22\u65b9\u6cd5[7]\uff0c\u6838\u5fc3\u65b9\u6cd5\u4e5f\u662f\u5bf9\u7ed3\u679c\u8fdb\u884c\u8fc7\u6ee4\uff0c\u4e0d\u8fc7\u4ed6\u4eec\u6ca1\u6709\u4f7f\u7528\u8fed\u4ee3\u66f4\u65b0\u9608\u503c\u7684\u7b56\u7565\uff0c\u800c\u662f\u901a\u8fc7\u53ec\u56de\u7387\u6765\u63a7\u5236\u8fc7\u6ee4\u7ed3\u679c\u7684\u6570\u91cf\uff0c\u6d88\u9664\u4e86\u8fed\u4ee3\u3001\u6392\u5e8f\u7684\u5f00\u9500\u3002\u4ed6\u4eec\u5bf9\u5185\u5b58\u74f6\u9888\u3001\u6307\u4ee4\u74f6\u9888\u505a\u4e86\u5f88\u8be6\u7ec6\u7684\u5206\u6790\u3002<\/p>\n

4.\u7ed3\u8bba<\/h2>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u867d\u7136GPU\u6709\u5f88\u9ad8\u7684\u7406\u8bba\u7b97\u529b\uff0c\u4f46\u5f88\u5c11\u6709\u5e94\u7528\u80fd\u5c06\u5176\u5145\u5206\u5229\u7528\u8d77\u6765\u3002\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u5957\u4ee5\u8fc7\u6ee4\u4e3a\u57fa\u7840\u7684\u7b97\u6cd5\uff0c\u901a\u8fc7\u5bf9\u8f93\u51fa\u7ed3\u679c\u8fdb\u884c\u8fc7\u6ee4\uff0c\u51cf\u5c11\u8f93\u51fa\u5e26\u5bbd\uff0c\u63d0\u9ad8\u786c\u4ef6\u5229\u7528\u7387\uff1b\u901a\u8fc7\u5bf9\u8fc7\u6ee4\u7ed3\u679c\u6570\u91cf\u7684\u63a7\u5236\uff0c\u63d0\u9ad8\u6392\u5e8f\u901f\u5ea6\uff0c\u4f46\u51e0\u4e4e\u4e0d\u5f71\u54cd\u53ec\u56de\u7387\uff1b\u901a\u8fc7\u4e13\u95e8\u9488\u5bf9\u4f4e\u7ef4\u5ea6\u8ba1\u7b97\u7684\u4f18\u5316\uff0c\u4f7f\u65b0\u786c\u4ef6\u7684\u5229\u7528\u7387\u8fbe\u523096%\u3002<\/p>\n

\u672c\u6587\u7684\u90e8\u5206\u5185\u5bb9\u6765\u81ea\u6211\u7684blog[9]\u548c\u4e13\u5229\u3002
Ding Nan[4]\u3001\u82f1\u4f1f\u8fbe[10]\u7b49\u5bf9GPU\u7684\u5185\u5b58\u74f6\u9888\u3001\u6307\u4ee4\u74f6\u9888\u4e5f\u6709\u5f88\u8be6\u7ec6\u7684\u5206\u6790\u3002<\/p>\n

\u53c2\u8003\u6587\u732e<\/h2>\n

[1] Jia, Zhe, et al. “Dissecting the NVidia Turing T4 GPU via microbenchmarking.” arXiv preprint arXiv:1903.07486 (2019).<\/p>\n

[2] Ebrahimpour, Hossein, and Abbas Kouzani. “Face recognition using bagging KNN.” International conference on signal processing and communication systems (ICSPCS\u20192007) australia, gold coast. sn, 2007.<\/p>\n

[3] Gordo, Albert, et al. “Deep image retrieval: Learning global representations for image search.” European conference on computer vision. Springer, Cham, 2016.<\/p>\n

[4] Ding, Nan, and Samuel Williams. “An instruction roofline model for gpus.” 2019 IEEE\/ACM Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems (PMBS) (2019): 7-18.<\/p>\n

[5] Babenko, Artem, and Victor Lempitsky. “Efficient indexing of billion-scale datasets of deep descriptors.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.<\/p>\n

[6] Johnson, Jeff, Matthijs Douze, and Herv\u00e9 J\u00e9gou. “Billion-scale similarity search with gpus.” IEEE Transactions on Big Data 7.3 (2019): 535-547.<\/p>\n

[7] Chern, Felix, et al. “Tpu-knn: K nearest neighbor search at peak flop\/s.” arXiv preprint arXiv:2206.14286 (2022).<\/p>\n

[8] Kerr, Andrew, et al. “Cutlass: Fast linear algebra in cuda c++.” NVIDIA Developer Blog (2017).<\/p>\n

[9] hws000. “The fastest knn search algorithm” https:\/\/blog.simbot.net\/2022\/03\/12\/the-fastest-knn-search-algorithm\/<\/p>\n

[10] “Matrix Multiplication Background User’s Guide” NVIDIA Documentation Center https:\/\/docs.nvidia.com\/deeplearning\/performance\/dl-performance-matrix-multiplication\/index.html<\/p>\n

[11] “Matrix multiply-accumulate operation using mma instruction” NVIDIA Documentation Center https:\/\/docs.nvidia.com\/cuda\/parallel-thread-execution\/index.html#warp-level-matrix-instructions-for-mma<\/p>\n","protected":false},"excerpt":{"rendered":"

\u672c\u6587\u521b\u9020\u6027\u7684\u63d0\u51fa\u4e00\u79cd\u65b0\u7684knn\uff08k nearest neighbor\uff09\u641c\u7d22\u52a0\u901f\u7b97\u6cd5\uff0c\u53ef\u4ee5\u8ba9\u786c\u4ef6\u7684\u5229\u7528\u7387\u8fbe\u523090%\u4ee5\u4e0a\u3002\u672c\u7b97\u6cd5\u5c06knn\u641c\u7d22\u4ece\u8ba1\u7b97\u5bc6\u96c6\u3001\u5185\u5b58IO\u5bc6\u96c6\u7b97\u6cd5\uff0c\u6539\u8fdb\u4e3a\u7eaf\u8ba1\u7b97\u5bc6\u96c6\u7b97\u6cd5\uff0c\u5145\u5206\u6316\u6398\u786c\u4ef6\u80fd\u529b\uff0c\u53ea\u8981\u5806\u66f4\u591a\u7684\u786c\u4ef6\uff0c\u5c31\u80fd\u5b9e\u73b0\u66f4\u5feb\u7684\u8ba1\u7b97\u3002\u65b0\u7b97\u6cd5\u4e0d\u4f46\u53ef\u4ee5\u8fd0\u884c\u5728GPU\u4e0a\uff0c\u8fd8\u53ef\u4ee5\u9ad8\u6548\u8fd0\u884c\u4e8e\u5404\u79cd\u4e13\u4e1a\u77e9\u9635\u52a0\u901f\u5361\u4e0a\u3002<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[29],"tags":[36,39,31,37,35,34],"_links":{"self":[{"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/posts\/400"}],"collection":[{"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/comments?post=400"}],"version-history":[{"count":7,"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/posts\/400\/revisions"}],"predecessor-version":[{"id":415,"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/posts\/400\/revisions\/415"}],"wp:attachment":[{"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/media?parent=400"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/categories?post=400"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.simbot.net\/wp-json\/wp\/v2\/tags?post=400"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}