{"id":660,"date":"2023-04-13T14:13:59","date_gmt":"2023-04-13T06:13:59","guid":{"rendered":"https:\/\/scutvk.cn\/?p=660"},"modified":"2023-04-13T14:20:52","modified_gmt":"2023-04-13T06:20:52","slug":"3-%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e8%bd%af%e4%bb%b6%e4%bc%98%e5%8c%96%e6%8a%80%e6%9c%af","status":"publish","type":"post","link":"https:\/\/scutvk.cn\/?p=660","title":{"rendered":"3 \u6df1\u5ea6\u5b66\u4e60\u8f6f\u4ef6\u4f18\u5316\u6280\u672f"},"content":{"rendered":"\n
\n\n\n\n

\u6ce8\uff1a\u672c\u535a\u5ba2\u4e3a\u672c\u4eba\uff08scutvk\uff09\u7eaf\u539f\u521b\u5185\u5bb9\uff0c\u7981\u6b62\u79c1\u81ea\u8f6c\u8f7d\u3002<\/p>\n\n\n\n

\n\n\n\n

\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u5e94\u7528\u7684\u5e7f\u6cdb\u4f20\u64ad\uff0c\u5982\u4f55\u5728\u8f6f\u4ef6\u5c42\u9762\u4e0a\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u8ba1\u7b97\u548c\u964d\u4f4e\u5185\u5b58\u4e5f\u662f\u6a21\u578b\u8f7b\u91cf\u5316\u9700\u8981\u8003\u8651\u7684\u95ee\u9898\u3002\u8fd9\u4e9b\u4f18\u5316\u6280\u672f\u5305\u62ec\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u3001\u7f16\u8bd1\u5668\u548c\u5e93\u7b49\u65b9\u9762\u7684\u6539\u8fdb\u3002\u6211\u4eec\u5c06\u4f9d\u6b21\u4ecb\u7ecd\u8fd9\u4e9b\u65b9\u9762\u7684\u5173\u952e\u6280\u672f\uff0c\u5e76\u8ba8\u8bba\u5b83\u4eec\u7684\u4e3b\u8981\u4f18\u7f3a\u70b9\u3002<\/p>\n\n\n\n

3.1 \u6846\u67b6\u4ee3\u7801\u4f18\u5316<\/h3>\n\n\n\n

\u6b64\u8282\u5305\u62ec\u4e00\u4e9b\u4f18\u5316\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4ee3\u7801\u7684tricks\u3002<\/p>\n\n\n\n

3.1.1 \u7b97\u5b50\u878d\u5408<\/h4>\n\n\n\n

\u57fa\u672c\u601d\u60f3\u662f\u5c06\u591a\u4e2a\u76f8\u90bb\u7684\u8ba1\u7b97\u64cd\u4f5c\uff08\u7b97\u5b50\uff09\u5408\u5e76\u4e3a\u4e00\u4e2a\u66f4\u5927\u7684\u64cd\u4f5c\uff0c\u4ece\u800c\u51cf\u5c11\u8ba1\u7b97\u4e2d\u7684\u5197\u4f59\u64cd\u4f5c\u3001\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u548c\u6027\u80fd\u3002<\/p>\n\n\n\n

\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u5305\u542b\u5927\u91cf\u7684\u8ba1\u7b97\u64cd\u4f5c\uff0c\u5982\u52a0\u6cd5\u3001\u4e58\u6cd5\u3001\u5377\u79ef\u7b49\u3002\u8fd9\u4e9b\u64cd\u4f5c\u901a\u5e38\u662f\u9010\u4e2a\u6267\u884c\u7684\uff0c\u4f46\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u5c06\u5b83\u4eec\u5408\u5e76\u53ef\u4ee5\u51cf\u5c11\u8ba1\u7b97\u91cf\u3001\u5185\u5b58\u8bbf\u95ee\u6b21\u6570\u4ee5\u53ca\u6570\u636e\u4f20\u8f93\u5f00\u9500\u3002<\/p>\n\n\n\n

\u7b97\u5b50\u878d\u5408\u7684\u4f18\u52bf\u5305\u62ec\uff1a<\/p>\n\n\n\n

    \n
  1. \u63d0\u9ad8\u8ba1\u7b97\u6027\u80fd\uff1a\u901a\u8fc7\u51cf\u5c11\u5197\u4f59\u64cd\u4f5c\u548c\u5185\u5b58\u8bbf\u95ee\u6b21\u6570\uff0c\u7b97\u5b50\u878d\u5408\u53ef\u4ee5\u964d\u4f4e\u8ba1\u7b97\u65f6\u95f4\u3002<\/li>\n\n\n\n
  2. \u51cf\u5c11\u5185\u5b58\u5360\u7528\uff1a\u5408\u5e76\u64cd\u4f5c\u53ef\u4ee5\u51cf\u5c11\u4e2d\u95f4\u53d8\u91cf\u7684\u4f7f\u7528\uff0c\u4ece\u800c\u964d\u4f4e\u5185\u5b58\u5360\u7528\u3002<\/li>\n\n\n\n
  3. \u964d\u4f4e\u80fd\u8017\uff1a\u51cf\u5c11\u8ba1\u7b97\u64cd\u4f5c\u548c\u5185\u5b58\u8bbf\u95ee\u53ef\u4ee5\u964d\u4f4e\u6574\u4f53\u80fd\u8017\u3002<\/li>\n<\/ol>\n\n\n\n

    PyTorch\u4e2d\u5b9e\u73b0\u7b97\u5b50\u878d\u5408\u6709\u4e24\u79cd\u65b9\u6cd5\uff1a<\/p>\n\n\n\n

      \n
    1. TorchScript\uff1aTorchScript\u662fPyTorch\u7684\u4e00\u4e2a\u5b50\u6a21\u5757\uff0c\u5b83\u53ef\u4ee5\u5c06\u6a21\u578b\u4ecePython\u4ee3\u7801\u8f6c\u6362\u4e3a\u9759\u6001\u56fe\u8868\u793a\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u53ef\u4ee5\u5728\u6ca1\u6709Python\u89e3\u91ca\u5668\u7684\u73af\u5883\u4e2d\u8fd0\u884c\uff0c\u4ece\u800c\u63d0\u9ad8\u6027\u80fd\u3002TorchScript\u63d0\u4f9b\u4e86\u4e00\u79cd\u65b9\u6cd5\uff0c\u5373\u5c06\u6a21\u578b\u8f6c\u6362\u4e3aScriptModule\uff0c\u7136\u540e\u4f7f\u7528torch.jit.trace\u6216torch.jit.script\u8fdb\u884c\u8ddf\u8e2a\u6216\u811a\u672c\u5316\u3002\u5728\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0cTorchScript\u4f1a\u81ea\u52a8\u8fdb\u884c\u4e00\u4e9b\u4f18\u5316\uff0c\u5305\u62ec\u7b97\u5b50\u878d\u5408\u3002\u8fd9\u4e9b\u4f18\u5316\u6709\u52a9\u4e8e\u63d0\u9ad8\u8ba1\u7b97\u6027\u80fd\u548c\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/li>\n\n\n\n
    2. PyTorch FX\uff08Functional Transformations\uff09\uff1aPyTorch FX\u662f\u53e6\u4e00\u4e2a\u7528\u4e8e\u6a21\u578b\u8f6c\u6362\u548c\u4f18\u5316\u7684\u5b50\u6a21\u5757\u3002\u5b83\u53ef\u4ee5\u6355\u83b7\u6a21\u578b\u7684\u8ba1\u7b97\u56fe\uff0c\u5e76\u5141\u8bb8\u5bf9\u8ba1\u7b97\u56fe\u8fdb\u884c\u8f6c\u6362\u548c\u4f18\u5316\u3002\u901a\u8fc7\u81ea\u5b9a\u4e49\u4f18\u5316\u7b56\u7565\uff0c\u53ef\u4ee5\u5728PyTorch FX\u4e2d\u5b9e\u73b0\u7b97\u5b50\u878d\u5408\u3002\u8fd9\u4f7f\u5f97\u53ef\u4ee5\u66f4\u7075\u6d3b\u5730\u9488\u5bf9\u7279\u5b9a\u786c\u4ef6\u548c\u573a\u666f\u8fdb\u884c\u4f18\u5316\u3002<\/li>\n<\/ol>\n\n\n\n

      \u4e0d\u8fc7\uff0c\u7531\u4e8e\u786c\u4ef6\u53ef\u4ee5\u5bf9\u9ed8\u5199\u8ba1\u7b97\u6709\u4f18\u5316\uff0c\u7b97\u5b50\u878d\u5408\u4e5f\u53ef\u80fd\u8d77\u5230\u8d1f\u4f18\u5316\u7684\u4f5c\u7528\u3002<\/p>\n\n\n\n

      3.1.2 \u5185\u5b58\u7ba1\u7406<\/h4>\n\n\n\n

      \u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u901a\u5e38\u5177\u6709\u6570\u767e\u4e07\u4e2a\u53c2\u6570\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e5f\u9700\u8981\u5927\u91cf\u7684\u4e2d\u95f4\u503c\u548c\u68af\u5ea6\uff0c\u9700\u8981\u5927\u91cf\u7684\u5185\u5b58\u6765\u5b58\u50a8\u8fd9\u4e9b\u53c2\u6570\u53ca\u5176\u76f8\u5173\u4fe1\u606f\u3002\u5728\u76ee\u524d\u6700\u5148\u8fdb\u7684GPU NVIDIA Tesla H100\u7684\u663e\u5b58\u4e3a80G\uff0c\u4f46\u968f\u7740\u5404\u79cd\u5927\u6a21\u578b\u7684\u53d1\u5c55\uff0c\u6a21\u578b\u53c2\u6570\u548c\u663e\u5b58\u7684\u589e\u957f\u660e\u663e\u4e0d\u5728\u4e00\u4e2a\u6570\u91cf\u7ea7\uff0c\u7531\u6b64\uff0c\u6211\u4eec\u4e5f\u9700\u8981\u5173\u6ce8\u5185\u5b58\u7ba1\u7406\u6280\u672f\u3002<\/p>\n\n\n\n

      \u5185\u5b58\u4ea4\u6362<\/h5>\n\n\n\n

      \u5185\u5b58\u4ea4\u6362\u6280\u672f\u662f\u6307\u5728\u52a0\u901f\u8bbe\u5907\u5185\u5b58\u548c\u4e3b\u5b58\u4e4b\u95f4\u4ea4\u6362\u6570\u636e\uff0c\u901a\u8fc7\u5728\u4e0d\u4f7f\u7528\u53d8\u91cf\u65f6\u5c06\u5176\u4ece\u52a0\u901f\u8bbe\u5907\u7684\u5185\u5b58\u4ea4\u6362\u5230\u4e3b\u5b58\u7684\u65b9\u5f0f\u6765\u964d\u4f4e\u52a0\u901f\u8bbe\u5907\u7684\u5185\u5b58\u6d88\u8017\uff0c\u5e76\u5728\u4e0b\u4e00\u6b21\u8bbf\u95ee\u53d8\u91cf\u4e4b\u524d\u5c06\u5176\u4ea4\u6362\u56de\u52a0\u901f\u8bbe\u5907\u5185\u5b58\u3002<\/p>\n\n\n\n

      \u91cd\u8ba1\u7b97<\/h5>\n\n\n\n

      \u91cd\u8ba1\u7b97\u6280\u672f\u7684\u601d\u60f3\u662f\u5c06\u7279\u5f81\u6620\u5c04\u8fd9\u6837\u7684\u4e2d\u95f4\u7ed3\u679c\u5728\u6b63\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u53ca\u65f6\u5730\u91ca\u653e\uff0c\u5728\u53cd\u5411\u4f20\u64ad\u7684\u8ba1\u7b97\u9700\u8981\u7528\u5230\u7279\u5f81\u6620\u5c04\u65f6\uff0c\u518d\u901a\u8fc7\u91cd\u65b0\u8ba1\u7b97\u7684\u65b9\u5f0f\u751f\u6210\uff0c\u8fdb\u800c\u53c2\u4e0e\u5230\u5f53\u524d\u8ba1\u7b97\u4e2d\u3002<\/p>\n\n\n\n

      \u5185\u5b58\u5171\u4eab<\/h5>\n\n\n\n

      \u5185\u5b58\u5171\u4eab\u6280\u672f\u6307\u7684\u662f\u901a\u8fc7\u5bf9\u4e0d\u540c\u53d8\u91cf\u751f\u547d\u5468\u671f\u7684\u5206\u6790\uff0c\u5728\u4e0d\u540c\u53d8\u91cf\u4e4b\u95f4\u91cd\u590d\u4f7f\u7528\u540c\u4e00\u5757\u5185\u5b58\u7a7a\u95f4\u3002<\/p>\n\n\n\n

      \"\"\/<\/figure>\n\n\n\n

      \u53c2\u8003\uff1a\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u5185\u5b58\u7ba1\u7406\u95ee\u9898\u7814\u7a76\u7efc\u8ff0_\u5510\u540d\u5a01\u7684\u535a\u5ba2-CSDN\u535a\u5ba2<\/a><\/p>\n\n\n\n

      3.1.3 \u5e76\u884c\u8ba1\u7b97<\/h4>\n\n\n\n

      \u7531\u4e8e\u5927\u91cf\u7684\u8ba1\u7b97\u548c\u6570\u636e\u5904\u7406\u9700\u6c42\uff0c\u91c7\u7528\u5e76\u884c\u8ba1\u7b97\u65b9\u6cd5\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6\u548c\u6548\u7387\u3002<\/p>\n\n\n\n

      \u5e76\u884c\u52a0\u8f7d<\/h5>\n\n\n\n

      \u6570\u636e\u5e76\u884c\u52a0\u8f7d\u662f\u6307\u5728\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u8bad\u7ec3\u65f6\uff0c\u5c06\u6570\u636e\u5206\u6210\u591a\u4e2a\u90e8\u5206\uff0c\u5e76\u540c\u65f6\u5728\u591a\u4e2aCPU\u6216\u8ba1\u7b97\u8bbe\u5907\u4e0a\u5e76\u884c\u52a0\u8f7d\u6570\u636e\uff0c\u4ee5\u63d0\u9ad8\u6570\u636e\u52a0\u8f7d\u901f\u5ea6\u548c\u6548\u7387\u3002\u5728PyTorch\u4e2d\uff0c\u4f7f\u7528torch.utils.data.distributed.DistributedSampler<\/code>\u6a21\u5757\u53ef\u4ee5\u5b9e\u73b0\u6570\u636e\u5e76\u884c\u52a0\u8f7d\u3002\u8fd9\u4e2a\u6a21\u5757\u63a5\u6536\u4e00\u4e2a\u6570\u636e\u96c6\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u5c06\u5176\u5212\u5206\u6210\u591a\u4e2a\u90e8\u5206\u3002\u6bcf\u4e2a\u90e8\u5206\u90fd\u7531\u4e00\u4e2a\u8fdb\u7a0b\u8d1f\u8d23\u52a0\u8f7d\u548c\u5904\u7406\uff0c\u63d0\u9ad8\u6570\u636e\u52a0\u8f7d\u901f\u5ea6\u548c\u6548\u7387\uff0c\u540c\u65f6\u4e5f\u4f1a\u51cf\u5c11\u6bcf\u4e2a\u8fdb\u7a0b\u7684\u8d1f\u8f7d\u3002<\/p>\n\n\n\n

      train_dataset = <\u52a0\u8f7d\u6570\u636e\u96c6>\ntrain_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) # \u521b\u5efa\u91c7\u6837\u5668\ntrain_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size,\n                                          shuffle=False, num_workers=num_workers,\n                                          pin_memory=True, sampler=train_sampler) # \u521b\u5efa\u6570\u636e\u52a0\u8f7d\u5668<\/code><\/pre>\n\n\n\n
      \u5e76\u884c\u5904\u7406<\/h5>\n\n\n\n
      \u5728\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u8bad\u7ec3\u65f6\uff0c\u5c06\u6570\u636e\u5206\u6210\u591a\u4e2a\u90e8\u5206\uff0c\u540c\u65f6\u5728\u591a\u4e2aGPU\u6216\u8ba1\u7b97\u8bbe\u5907\u4e0a\u5bf9\u8fd9\u4e9b\u6570\u636e\u8fdb\u884c\u5e76\u884c\u5904\u7406\uff0c\u4ee5\u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6\u548c\u6548\u7387\u3002\u5728PyTorch\u4e2d\uff0c\u4f7f\u7528torch.nn.DataParallel<\/code>\u6a21\u5757\u53ef\u4ee5\u5b9e\u73b0\u6570\u636e\u5e76\u884c\u5904\u7406\u3002\u8fd9\u4e2a\u6a21\u5757\u63a5\u6536\u4e00\u4e2a\u6a21\u578b\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u81ea\u52a8\u5c06\u6a21\u578b\u590d\u5236\u5230\u591a\u4e2aGPU\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n
      3.2 \u7f16\u8bd1\u5668<\/h3>\n\n\n\n
      \u5728\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u4e2d\uff0c\u7f16\u8bd1\u5668\u626e\u6f14\u7740\u81f3\u5173\u91cd\u8981\u7684\u89d2\u8272\uff0c\u5b83\u5c06\u9ad8\u7ea7\u8bed\u8a00\u7f16\u5199\u7684\u6a21\u578b\u8f6c\u6362\u4e3a\u9488\u5bf9\u7279\u5b9a\u786c\u4ef6\u5e73\u53f0\u7684\u4f4e\u7ea7\u6307\u4ee4\u3002\u4f18\u5316\u7f16\u8bd1\u5668\u80fd\u591f\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\uff0c\u964d\u4f4e\u5ef6\u8fdf\uff0c\u4ece\u800c\u4f7f\u5f97\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u80fd\u591f\u5728\u5404\u79cd\u786c\u4ef6\u8bbe\u5907\u4e0a\u5b9e\u73b0\u66f4\u9ad8\u6027\u80fd\u3002\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u8bb8\u591a\u73b0\u6210\u7684\u3001\u4e13\u95e8\u4e3a\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u8bbe\u8ba1\u7684\u7f16\u8bd1\u5668\u3002<\/p>\n\n\n\n
        \n
      1. TVM<\/p>
        TVM \u662f\u4e00\u4e2a\u5f00\u6e90\u7684\u6df1\u5ea6\u5b66\u4e60\u7f16\u8bd1\u5668\u6846\u67b6\uff0c\u65e8\u5728\u4e3a\u5404\u79cd\u786c\u4ef6\u5e73\u53f0\u4f18\u5316\u548c\u90e8\u7f72\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u3002\u5b83\u8d77\u6e90\u4e8e\u534e\u76db\u987f\u5927\u5b66\u7684\u7cfb\u7edf\u7814\u7a76\u9879\u76ee\u3002TVM \u63d0\u4f9b\u4e86\u4e00\u4e2a\u7075\u6d3b\u3001\u53ef\u6269\u5c55\u7684\u7f16\u8bd1\u67b6\u6784\uff0c\u53ef\u4ee5\u5b9e\u73b0\u9ad8\u6548\u7684\u6a21\u578b\u4f18\u5316\u548c\u4ee3\u7801\u751f\u6210\uff0c\u4ece\u800c\u5728\u5404\u79cd\u786c\u4ef6\u8bbe\u5907\u4e0a\u5b9e\u73b0\u9ad8\u6027\u80fd\u6df1\u5ea6\u5b66\u4e60\u63a8\u7406\u3002<\/p>
        \u4e3b\u8981\u7279\u70b9\uff1a<\/p>\n
          \n
        1. \u786c\u4ef6\u65e0\u5173\uff1aTVM \u652f\u6301\u591a\u79cd\u786c\u4ef6\u5e73\u53f0\uff0c\u5305\u62ec CPU\u3001GPU\u3001FPGA \u548c\u4e13\u7528 AI \u82af\u7247\u7b49\u3002\u8fd9\u4f7f\u5f97\u5f00\u53d1\u8005\u65e0\u9700\u5bf9\u7279\u5b9a\u786c\u4ef6\u5e73\u53f0\u8fdb\u884c\u624b\u52a8\u4f18\u5316\uff0c\u80fd\u591f\u81ea\u52a8\u9002\u5e94\u4e0d\u540c\u7684\u786c\u4ef6\u8bbe\u5907\u3002<\/li>\n\n\n\n
        2. \u591a\u6846\u67b6\u652f\u6301\uff1aTVM \u652f\u6301\u591a\u79cd\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u5982 TensorFlow\u3001PyTorch\u3001MXNet\u3001ONNX \u7b49\u3002\u5f00\u53d1\u8005\u53ef\u4ee5\u5c06\u8fd9\u4e9b\u6846\u67b6\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\u5bfc\u5165\u5230 TVM \u4e2d\u8fdb\u884c\u4f18\u5316\u548c\u90e8\u7f72\u3002<\/li>\n\n\n\n
        3. \u7aef\u5230\u7aef\u4f18\u5316\uff1aTVM \u91c7\u7528\u4e86\u4e00\u79cd\u7aef\u5230\u7aef\u7684\u4f18\u5316\u7b56\u7565\uff0c\u4ece\u7b97\u5b50\u5c42\u9762\uff08\u5982\u5377\u79ef\u3001\u77e9\u9635\u4e58\u6cd5\u7b49\u57fa\u672c\u8ba1\u7b97\u5355\u5143\uff09\u5230\u6574\u4e2a\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u8fdb\u884c\u4f18\u5316\u3002\u8fd9\u79cd\u4f18\u5316\u7b56\u7565\u53ef\u4ee5\u6709\u6548\u5730\u63d0\u9ad8\u8ba1\u7b97\u6027\u80fd\uff0c\u964d\u4f4e\u5185\u5b58\u5360\u7528\u548c\u5ef6\u8fdf\u3002<\/li>\n\n\n\n
        4. \u81ea\u52a8\u8c03\u5ea6\uff1aTVM \u5229\u7528\u81ea\u52a8\u8c03\u5ea6\u7b97\u6cd5\uff08AutoTVM \u548c AutoScheduler\uff09\u641c\u7d22\u6700\u4f18\u7684\u8ba1\u7b97\u7b56\u7565\u548c\u6570\u636e\u5e03\u5c40\uff0c\u4ece\u800c\u81ea\u52a8\u5730\u4e3a\u5404\u79cd\u786c\u4ef6\u5e73\u53f0\u751f\u6210\u9ad8\u6027\u80fd\u7684\u4ee3\u7801\u3002\u8fd9\u5927\u5927\u964d\u4f4e\u4e86\u4f18\u5316\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u95e8\u69db\u548c\u6210\u672c\u3002<\/li>\n<\/ol>\n<\/li>\n\n\n\n
        5. TC<\/p>
          Tensor Comprehensions\uff08TC\uff09\u662f\u4e00\u4e2a\u5f00\u6e90\u7684\u3001\u7528\u4e8e\u9ad8\u6027\u80fd\u6df1\u5ea6\u5b66\u4e60\u8ba1\u7b97\u7684\u7f16\u7a0b\u6846\u67b6\u3002TC \u65e8\u5728\u8ba9\u5f00\u53d1\u8005\u4ee5\u4e00\u79cd\u66f4\u7b80\u6d01\u3001\u66f4\u76f4\u89c2\u7684\u65b9\u5f0f\u7f16\u5199\u6df1\u5ea6\u5b66\u4e60\u7b97\u5b50\uff08\u4f8b\u5982\u5377\u79ef\u3001\u77e9\u9635\u4e58\u6cd5\u7b49\uff09\u3002TC \u662f\u7531 Facebook AI Research (FAIR) \u5f00\u53d1\u7684\uff0c\u53ef\u4ee5\u81ea\u52a8\u751f\u6210\u9488\u5bf9\u7279\u5b9a\u786c\u4ef6\u7684\u9ad8\u6548\u4ee3\u7801\uff0c\u5c24\u5176\u662f\u9488\u5bf9 GPU \u8fdb\u884c\u4f18\u5316<\/strong>\u3002<\/p>
          \u4e3b\u8981\u7279\u70b9\uff1a<\/p>\n
            \n
          1. \u57fa\u4e8e Polyhedral \u6a21\u578b\uff1aTC \u4f7f\u7528\u4e86\u4e00\u79cd\u79f0\u4e3a Polyhedral \u6a21\u578b\u7684\u6570\u5b66\u8868\u793a\u65b9\u6cd5\uff0c\u901a\u8fc7\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u81ea\u52a8\u5730\u4e3a\u7ed9\u5b9a\u7684\u8ba1\u7b97\u4efb\u52a1\u627e\u5230\u9ad8\u6548\u7684\u5b9e\u73b0\u65b9\u5f0f\u3002Polyhedral \u6a21\u578b\u53ef\u4ee5\u5bf9\u5faa\u73af\u5d4c\u5957\u7ed3\u6784\u8fdb\u884c\u6df1\u5ea6\u5206\u6790\uff0c\u4ece\u800c\u4e3a\u4f18\u5316\u63d0\u4f9b\u5f3a\u5927\u7684\u652f\u6301\u3002<\/li>\n\n\n\n
          2. \u57fa\u4e8e Halide \u8bed\u8a00\uff1aTC \u4f7f\u7528\u4e86\u7c7b\u4f3c\u4e8e Halide \u7684\u9886\u57df\u7279\u5b9a\u8bed\u8a00\uff08DSL\uff09\uff0c\u5141\u8bb8\u5f00\u53d1\u8005\u7528\u7b80\u6d01\u7684\u8bed\u6cd5\u7f16\u5199\u6df1\u5ea6\u5b66\u4e60\u7b97\u5b50\u3002\u901a\u8fc7\u8fd9\u79cd\u8bed\u8a00\uff0c\u5f00\u53d1\u8005\u53ef\u4ee5\u8f7b\u677e\u5730\u63cf\u8ff0\u8ba1\u7b97\u8fc7\u7a0b\u548c\u6570\u636e\u8bbf\u95ee\u6a21\u5f0f\uff0c\u800c\u65e0\u9700\u5173\u5fc3\u5e95\u5c42\u786c\u4ef6\u5b9e\u73b0\u7684\u7ec6\u8282\u3002<\/li>\n\n\n\n
          3. \u81ea\u52a8\u4f18\u5316\uff1aTC \u5229\u7528\u81ea\u52a8\u641c\u7d22\u7b97\u6cd5\uff0c\u4e3a\u7ed9\u5b9a\u7684\u6df1\u5ea6\u5b66\u4e60\u7b97\u5b50\u627e\u5230\u6700\u4f18\u7684\u6620\u5c04\u7b56\u7565\uff0c\u4ee5\u5b9e\u73b0\u9ad8\u6027\u80fd\u8ba1\u7b97\u3002\u8fd9\u79cd\u81ea\u52a8\u4f18\u5316\u65b9\u6cd5\u964d\u4f4e\u4e86\u624b\u52a8\u4f18\u5316\u7684\u590d\u6742\u6027\u548c\u5de5\u4f5c\u91cf\u3002<\/li>\n<\/ol>\n<\/li>\n\n\n\n
          4.  Glow<\/span> 
            Glow\uff08Graph-Lowering Compiler\uff09\u7531Facebook AI Research\uff08FAIR\uff09\u5f00\u53d1\uff0c\u65e8\u5728\u4e3a\u795e\u7ecf\u7f51\u7edc\u63d0\u4f9b\u9ad8\u6548\u7684\u786c\u4ef6\u52a0\u901f\u3002Glow\u7684\u4e3b\u8981\u76ee\u6807\u662f\u4f18\u5316\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u6027\u80fd\uff0c\u4ee5\u4fbf\u5728\u5404\u79cd\u786c\u4ef6\u5e73\u53f0\u4e0a\u5b9e\u73b0\u6700\u4f73\u6027\u80fd\u3002Glow\u7684\u8bbe\u8ba1\u7075\u6d3b\u4e14\u53ef\u6269\u5c55\uff0c\u53ef\u4ee5\u652f\u6301\u591a\u79cd\u786c\u4ef6\u8bbe\u5907\u3002\u5b83\u7684\u4e00\u4e2a\u663e\u8457\u4f18\u52bf\u662f\u5176\u80fd\u591f\u751f\u6210\u7279\u5b9a\u4e8e\u786c\u4ef6\u7684\u4f18\u5316\u4ee3\u7801\uff0c\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u5728\u4e0d\u540c\u786c\u4ef6\u8bbe\u5907\u4e0a\u7684\u6267\u884c\u901f\u5ea6\u3002<\/p> <\/li>\n\n\n\n
          5.  nGraph+PlaidML<\/span> 
            nGraph \u662f\u4e00\u4e2a\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u56fe\u7f16\u8bd1\u5668\u3002\u5b83\u7531\u82f1\u7279\u5c14\u516c\u53f8\u5f00\u53d1\uff0c\u76ee\u7684\u662f\u4e3a\u4e86\u4f18\u5316\u5404\u79cd\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff08\u5982 TensorFlow\uff09\u5728\u4e0d\u540c\u786c\u4ef6\u5e73\u53f0\u4e0a\u7684\u6027\u80fd\u3002nGraph \u7684\u6838\u5fc3\u7ec4\u4ef6\u662f\u4e00\u4e2a\u9ad8\u7ea7\u7684\u4e2d\u95f4\u8868\u793a\uff08IR\uff09\uff0c\u5b83\u53ef\u4ee5\u8868\u793a\u8ba1\u7b97\u56fe\u3002nGraph \u901a\u8fc7\u5c06\u6846\u67b6\u7684\u8ba1\u7b97\u56fe\u8f6c\u6362\u4e3a\u5176\u4e2d\u95f4\u8868\u793a\uff0c\u7136\u540e\u5bf9\u5176\u8fdb\u884c\u4f18\u5316\uff0c\u4ee5\u5728\u76ee\u6807\u786c\u4ef6\u5e73\u53f0\u4e0a\u5b9e\u73b0\u66f4\u9ad8\u7684\u6027\u80fd\u3002<\/p> 
            PlaidML \u662f\u4e00\u4e2a\u5f00\u6e90\u7684\u8bbe\u5907\u65e0\u5173\u7684\u6df1\u5ea6\u5b66\u4e60\u52a0\u901f\u5e93\uff0c\u65e8\u5728\u4e3a\u5404\u79cd\u786c\u4ef6\u5e73\u53f0\u63d0\u4f9b\u7edf\u4e00\u7684\u9ad8\u6027\u80fd\u8ba1\u7b97\u652f\u6301\u3002\u5b83\u652f\u6301\u8bb8\u591a\u4e0d\u540c\u7684\u8bbe\u5907\uff0c\u5982 NVIDIA\u3001AMD \u548c\u82f1\u7279\u5c14\u7684 GPU\uff0c\u4ee5\u53ca\u5176\u4ed6\u50cf CPU\u3001FPGA \u7b49\u786c\u4ef6\u3002PlaidML \u7684\u6838\u5fc3\u662f\u4e00\u4e2a\u9ad8\u5ea6\u4f18\u5316\u7684\u591a\u540e\u7aef\u8ba1\u7b97\u5f15\u64ce\uff0c\u5b83\u53ef\u4ee5\u4e3a\u5404\u79cd\u8bbe\u5907\u751f\u6210\u9ad8\u6548\u7684\u8ba1\u7b97\u5185\u6838\u3002<\/p> <\/li>\n\n\n\n
          6.  XLA<\/span> 
            XLA\uff08\u52a0\u901f\u7ebf\u6027\u4ee3\u6570\uff09\u662f\u4e00\u4e2a\u7f16\u8bd1\u5668\u548c\u8fd0\u884c\u65f6\u5e93\uff0c\u65e8\u5728\u4f18\u5316TensorFlow\uff08\u540e\u652f\u6301Pytorch\u3001Keras\uff09\u8ba1\u7b97\u56fe\u3002\u5b83\u7531Google\u5f00\u53d1\uff0c\u7528\u4e8e\u63d0\u9ad8TensorFlow\u5728\u5404\u79cd\u786c\u4ef6\u5e73\u53f0\u4e0a\u7684\u6027\u80fd\u3002XLA\u901a\u8fc7\u5c06\u8ba1\u7b97\u56fe\u8f6c\u6362\u4e3a\u9002\u5408\u786c\u4ef6\u7279\u6027\u7684\u9ad8\u6548\u4f4e\u7ea7\u4ee3\u7801\u6765\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\u3002<\/p><\/li>\n<\/ol>\n\n\n\n
            \"\"<\/figure>\n\n\n\n
            \u53c2\u8003\uff1a\u6df1\u5ea6\u5b66\u4e60\u7f16\u8bd1\u5668\u6574\u7406 – \u77e5\u4e4e (zhihu.com)<\/a><\/p>\n\n\n\n
            3.3 \u8ba1\u7b97\u5e93<\/h3>\n\n\n\n
              \n
            1.  cuDNN<\/span> 
              cuDNN\uff08CUDA Deep Neural Network library\uff09\u662f\u4e00\u4e2a\u7531NVIDIA\u5f00\u53d1\u7684\u57fa\u4e8eGPU\u7684\u6df1\u5ea6\u5b66\u4e60\u5e93\uff0c\u5b83\u63d0\u4f9b\u4e86\u9ad8\u5ea6\u4f18\u5316\u7684\u5e95\u5c42\u51fd\u6570\uff0c\u4f7f\u5f97\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u5982TensorFlow\u3001PyTorch\u3001Caffe\u7b49\u80fd\u591f\u5145\u5206\u5229\u7528NVIDIA GPU\u7684\u8ba1\u7b97\u80fd\u529b\u3002cuDNN\u6781\u5927\u5730\u52a0\u901f\u4e86\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\uff0c\u4f7f\u5f97\u5f00\u53d1\u8005\u80fd\u591f\u5feb\u901f\u5730\u6784\u5efa\u548c\u8bad\u7ec3\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u3002<\/p> <\/li>\n\n\n\n
            2.  BLAS<\/span> 
              BLAS\uff08Basic Linear Algebra Subprograms\uff09\u662f\u4e00\u7ec4\u7ebf\u6027\u4ee3\u6570\u57fa\u672c\u64cd\u4f5c\u7684\u51fd\u6570\u5e93\u3002\u63d0\u4f9b\u4e86\u4e00\u4e9b\u9ad8\u6548\u7684\u6570\u5b66\u64cd\u4f5c\u51fd\u6570\uff0c\u4f8b\u5982\u5411\u91cf\u548c\u77e9\u9635\u7684\u4e58\u6cd5\u3001\u77e9\u9635\u7684\u8f6c\u7f6e\u3001\u5411\u91cf\u7684\u5185\u79ef\u7b49\uff0c\u53ef\u4ee5\u88ab\u5e7f\u6cdb\u5730\u7528\u4e8e\u5404\u79cd\u79d1\u5b66\u8ba1\u7b97\u3001\u5de5\u7a0b\u5e94\u7528\u548c\u6570\u636e\u5206\u6790\u7b49\u9886\u57df\u3002<\/p> <\/li>\n\n\n\n
            3.  cuBLAS<\/span> 
              cuBLAS\uff08CUDA Basic Linear Algebra Subprograms\uff09\u662f\u4e00\u4e2aNVIDIA\u5f00\u53d1\u7684\u57fa\u4e8eCUDA\u7684\u7ebf\u6027\u4ee3\u6570\u5e93\u3002\u5b83\u4e3aGPU\uff08\u56fe\u5f62\u5904\u7406\u5668\uff09\u4e0a\u7684\u7ebf\u6027\u4ee3\u6570\u8ba1\u7b97\u63d0\u4f9b\u4e86\u9ad8\u6027\u80fd\u7684\u5b9e\u73b0\uff0c\u4e3b\u8981\u9488\u5bf9NVIDIA GPU\u67b6\u6784\u3002cuBLAS\u5e93\u63d0\u4f9b\u4e86BLAS\u89c4\u8303\u4e2d\u5b9a\u4e49\u7684\u5404\u79cd\u7ebf\u6027\u4ee3\u6570\u64cd\u4f5c\uff0c\u5305\u62ec\u5411\u91cf\u3001\u77e9\u9635\u4ee5\u53ca\u590d\u6742\u6570\u7684\u52a0\u6cd5\u3001\u51cf\u6cd5\u3001\u70b9\u4e58\u3001\u77e9\u9635\u4e58\u6cd5\u7b49\u3002cuBLAS\u7684\u4e3b\u8981\u4f18\u52bf\u5728\u4e8e\u5b83\u5145\u5206\u5229\u7528\u4e86GPU\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\uff0c\u4f7f\u5f97\u7ebf\u6027\u4ee3\u6570\u8fd0\u7b97\u80fd\u5728GPU\u4e0a\u9ad8\u6548\u6267\u884c\u3002<\/p> <\/li>\n\n\n\n
            4. Intel-MKL<\/p>
              Intel-MKL\uff08Intel Math Kernel Library\uff09\u662fIntel \u5f00\u53d1\u7684\u4e00\u5957\u6570\u5b66\u5e93\uff0c\u65e8\u5728\u4e3a\u9ad8\u6027\u80fd\u8ba1\u7b97\u63d0\u4f9b\u4f18\u5316\u7684\u6570\u5b66\u51fd\u6570\u3002\u5b83\u9488\u5bf9\u82f1\u7279\u5c14\u5904\u7406\u5668\u8fdb\u884c\u4e86\u9ad8\u5ea6\u4f18\u5316\uff0c\u652f\u6301C\u3001C++\u548cFortran\u7f16\u7a0b\u8bed\u8a00\uff0c\u5e76\u63d0\u4f9b\u4e86\u4e00\u5957\u6613\u4e8e\u4f7f\u7528\u7684API\u63a5\u53e3\u3002<\/p>
              Intel-MKL\u5305\u542b\u4ee5\u4e0b\u51e0\u4e2a\u4e3b\u8981\u529f\u80fd\uff1a<\/p>\n
                \n
              1. \u7ebf\u6027\u4ee3\u6570\uff1aIntel MKL\u63d0\u4f9b\u4e86BLAS\u548cLAPACK\uff08\u7ebf\u6027\u4ee3\u6570\u5305\uff09\u5e93\u7684\u4f18\u5316\u5b9e\u73b0\u3002\u8fd9\u4e9b\u5e93\u5305\u542b\u7528\u4e8e\u5411\u91cf\u3001\u77e9\u9635\u8fd0\u7b97\u4ee5\u53ca\u7ebf\u6027\u65b9\u7a0b\u7ec4\u6c42\u89e3\u7684\u51fd\u6570\u3002<\/li>\n\n\n\n
              2. \u5feb\u901f\u5085\u91cc\u53f6\u53d8\u6362\uff08FFT\uff09\uff1aIntel MKL\u63d0\u4f9b\u4e86\u9ad8\u6548\u7684FFT\u5e93\uff0c\u7528\u4e8e\u5904\u7406\u4e00\u7ef4\u3001\u591a\u7ef4\u548c\u5b9e\u6570\u3001\u590d\u6570\u6570\u636e\u7684\u79bb\u6563\u5085\u91cc\u53f6\u53d8\u6362\u3002<\/li>\n\n\n\n
              3. \u5411\u91cf\u6570\u5b66\uff1aIntel MKL\u63d0\u4f9b\u4e86\u4e00\u7cfb\u5217\u4f18\u5316\u7684\u5411\u91cf\u6570\u5b66\u51fd\u6570\uff0c\u7528\u4e8e\u5904\u7406\u5411\u91cf\u8fd0\u7b97\uff0c\u5982\u52a0\u6cd5\u3001\u51cf\u6cd5\u3001\u4e58\u6cd5\u3001\u9664\u6cd5\u7b49\u3002<\/li>\n\n\n\n
              4. \u7a00\u758f\u77e9\u9635\u8fd0\u7b97\uff1aIntel MKL\u63d0\u4f9b\u4e86\u4e00\u7cfb\u5217\u7a00\u758f\u77e9\u9635\u8fd0\u7b97\u51fd\u6570\uff0c\u5982\u7a00\u758f\u77e9\u9635-\u5411\u91cf\u4e58\u6cd5\u3001\u7a00\u758f\u77e9\u9635-\u7a20\u5bc6\u77e9\u9635\u4e58\u6cd5\u7b49\u3002<\/li>\n\n\n\n
              5. \u975e\u7ebf\u6027\u4f18\u5316\uff1aIntel MKL\u63d0\u4f9b\u4e86\u4e00\u5957\u4f18\u5316\u6c42\u89e3\u5668\uff0c\u7528\u4e8e\u89e3\u51b3\u7ebf\u6027\u548c\u975e\u7ebf\u6027\u4f18\u5316\u95ee\u9898\u3002<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n\n\n\n
                \u53c2\u8003\uff1a<\/p>\n\n\n\n
                  \n
                1. \u6df1\u5ea6\u5b66\u4e60\u8ba1\u7b97\u5e93(1)-\u6982\u89c8BLAS\u3001DNN\u52a0\u901f\u5e93 – \u77e5\u4e4e (zhihu.com)<\/a><\/li>\n\n\n\n
                2. \u57fa\u672c\u7ebf\u6027\u4ee3\u6570\u5b50\u7a0b\u5e8f – \u7ef4\u57fa\u767e\u79d1 (wikipedia.org)<\/a><\/li>\n\n\n\n
                3. Math Kernel Library – Wikipedia<\/a><\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"
                  \u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u5e94\u7528\u7684\u5e7f\u6cdb\u4f20\u64ad\uff0c\u5982\u4f55\u5728\u8f6f\u4ef6\u5c42\u9762\u4e0a\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u8ba1\u7b97\u548c\u964d\u4f4e\u5185\u5b58\u4e5f\u662f\u6a21\u578b\u8f7b\u91cf\u5316\u9700\u8981\u8003\u8651\u7684\u95ee\u9898\u3002\u8fd9\u4e9b\u4f18\u5316\u6280\u672f\u5305\u62ec\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u3001\u7f16\u8bd1\u5668\u548c\u5e93\u7b49\u65b9\u9762\u7684\u6539\u8fdb\u3002\u6211\u4eec\u5c06\u4f9d\u6b21\u4ecb\u7ecd\u8fd9\u4e9b\u65b9\u9762\u7684\u5173\u952e\u6280\u672f\uff0c\u5e76\u8ba8\u8bba\u5b83\u4eec\u7684\u4e3b\u8981\u4f18\u7f3a\u70b9\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_publicize_message":"","jetpack_is_tweetstorm":false,"jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","enabled":false}}},"categories":[2],"tags":[],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts\/660"}],"collection":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=660"}],"version-history":[{"count":2,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts\/660\/revisions"}],"predecessor-version":[{"id":663,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts\/660\/revisions\/663"}],"wp:attachment":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=660"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=660"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=660"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}