{"id":541,"date":"2026-07-08T03:10:36","date_gmt":"2026-07-07T19:10:36","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=541"},"modified":"2026-07-08T18:58:41","modified_gmt":"2026-07-08T10:58:41","slug":"cbr%e8%9e%8d%e5%90%88%e7%ae%97%e5%ad%90%ef%bc%9aconv-batchnorm-relu%e7%9a%84%e4%b8%89%e4%bd%8d%e4%b8%80%e4%bd%93","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/541\/","title":{"rendered":"(20) CBR\u878d\u5408\u7b97\u5b50\uff1aConv + BatchNorm + ReLU\u7684\u4e09\u4f4d\u4e00\u4f53"},"content":{"rendered":"\n

\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341<\/h2>\n\n\n\n

\u5728\u51e0\u4e4e\u4efb\u4f55\u73b0\u4ee3\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u91cc\uff0c\u4f60\u90fd\u80fd\u627e\u5230\u8fd9\u6837\u4e00\u79cd\u7ed3\u6784\uff1a\u4e00\u4e2a\u5377\u79ef\u5c42\u540e\u9762\u7d27\u8ddf\u6279\u5f52\u4e00\u5316\uff08BatchNorm\uff09\uff0c\u518d\u7d27\u8ddf ReLU \u6fc0\u6d3b\u3002ResNet\u3001VGG\u3001MobileNet\u3001EfficientNet \u90fd\u662f\u5982\u6b64\u3002\u5728 PyTorch \u98ce\u683c\u7684\u4ee3\u7801\u91cc\uff0c\u5b83\u901a\u5e38\u5199\u6210\u4e09\u884c\uff1a<\/p>\n\n\n\n

nn.Conv2d(...),\nnn.BatchNorm2d(...),\nnn.ReLU(...)<\/pre>\n\n\n\n
\u8fd9\u4e09\u4e2a\u7b97\u5b50\u5728\u6570\u5b66\u4e0a\u5404\u53f8\u5176\u804c\u2014\u2014\u5377\u79ef\u63d0\u53d6\u7a7a\u95f4\u7279\u5f81\uff0cBatchNorm \u7a33\u5b9a\u8bad\u7ec3\u5206\u5e03\uff0cReLU \u5f15\u5165\u975e\u7ebf\u6027\u2014\u2014\u4f46\u5728\u786c\u4ef6\u6267\u884c\u4e0a\u5b83\u4eec\u5374\u7d27\u7d27\u76f8\u90bb\u3002\u6b63\u662f\u8fd9\u79cd\u76f8\u90bb\u5173\u7cfb\uff0c\u7ed9\u4e86\u6846\u67b6\u4e00\u4e2a\u5de8\u5927\u7684\u4f18\u5316\u673a\u4f1a\uff1a\u628a\u5b83\u4eec\u878d\u5408\u6210\u4e00\u4e2a\u7b97\u5b50<\/strong>\u6765\u6267\u884c\u3002<\/p>\n\n\n\n
\u8fd9\u5c31\u662f Tech-Renaissance \u4e2d\u7684 CBR \u878d\u5408\u7b97\u5b50\u3002\u672c\u6587\u8981\u8bb2\u6e05\u695a\u4e09\u4ef6\u4e8b\uff1aCBR \u878d\u5408\u4e3a\u4ec0\u4e48\u80fd\u5feb\u3001\u4e3a\u4ec0\u4e48\u8bad\u7ec3\u573a\u666f\u6bd4\u63a8\u7406\u573a\u666f\u66f4\u96be\u505a\u3001\u4ee5\u53ca Tech-Renaissance \u662f\u5982\u4f55\u5728 AMP \u8def\u5f84\u4e0a\u628a CBR \u6253\u9020\u6210\u4e00\u7b49\u7b97\u5b50\u7684\u3002<\/p>\n\n\n\n
\u4e00\u3001\u5206\u7acb\u7b97\u5b50\u7684\u5f00\u9500\uff1a\u5e26\u5bbd\u624d\u662f\u771f\u6b63\u7684\u74f6\u9888<\/h2>\n\n\n\n
\u5728\u6ca1\u6709\u878d\u5408\u7684\u60c5\u51b5\u4e0b\uff0c\u4e00\u4e2a CBR \u5757\u5728\u524d\u5411\u4f20\u64ad\u4e2d\u81f3\u5c11\u8981\u7ecf\u5386\u4e09\u6b21\u72ec\u7acb\u7684 CUDA kernel \u63d0\u4ea4\uff1a<\/p>\n\n\n\n
    \n
  1. Conv<\/strong>\uff1a\u8bfb\u53d6\u8f93\u5165\u7279\u5f81\u56fe X<\/code> \u548c\u5377\u79ef\u6838 W<\/code>\uff0c\u8ba1\u7b97\u5377\u79ef\u7ed3\u679c\uff0c\u628a\u4e2d\u95f4\u7ed3\u679c conv_output<\/code> \u5199\u56de\u5168\u5c40\u663e\u5b58\uff1b<\/li>\n\n\n\n
  2. BatchNorm<\/strong>\uff1a\u8bfb\u53d6 conv_output<\/code>\uff0c\u8ba1\u7b97\u6216\u5957\u7528\u5747\u503c\u3001\u65b9\u5dee\u3001\u7f29\u653e\u3001\u504f\u79fb\uff0c\u628a bn_output<\/code> \u5199\u56de\u663e\u5b58\uff1b<\/li>\n\n\n\n
  3. ReLU<\/strong>\uff1a\u8bfb\u53d6 bn_output<\/code>\uff0c\u505a max(0, x)<\/code>\uff0c\u628a\u6700\u7ec8\u7ed3\u679c\u5199\u56de\u663e\u5b58\u3002<\/li>\n<\/ol>\n\n\n\n
    \u4e09\u6b21 launch\u3001\u4e09\u6b21\u5168\u5c40\u5185\u5b58\u4e8b\u52a1\u3001\u4e24\u4efd\u5b8c\u6574\u5927\u5c0f\u7684\u4e2d\u95f4\u6fc0\u6d3b\u5f20\u91cf\u3002\u5bf9\u4e8e ResNet-50 \u8fd9\u79cd\u6709 50 \u591a\u4e2a\u5377\u79ef\u5c42\u7684\u7f51\u7edc\uff0c\u8fd9\u79cd\u5f00\u9500\u4f1a\u4e00\u5c42\u4e00\u5c42\u7d2f\u52a0\u3002<\/p>\n\n\n\n
    \u66f4\u91cd\u8981\u7684\u662f\uff0c\u73b0\u4ee3 CNN \u8bad\u7ec3\u5728 A100\u3001RTX 5090 \u8fd9\u7c7b\u9ad8\u7aef GPU \u4e0a\u5f80\u5f80\u662f\u663e\u5b58\u5e26\u5bbd\u53d7\u9650<\/strong>\u800c\u975e\u7b97\u529b\u53d7\u9650\u3002\u4ee5 A100 \u4e3a\u4f8b\uff0cFP16 Tensor Core \u7684\uff08\u7a20\u5bc6\uff09\u5cf0\u503c\u7b97\u529b\u7ea6\u4e3a 312 TFLOPS\uff1bA100 40GB \u7684 HBM \u5e26\u5bbd\u7ea6\u4e3a 1.55 TB\/s\uff0c80GB \u7248\u7ea6\u4e3a 2.0 TB\/s\u3002\u6309\u7167 roofline \u6a21\u578b\u7684\u53e3\u5f84\uff0c\u82e5\u6309 40GB \u7248\u8ba1\u7b97\uff0c\u8fd0\u7b97\u5f3a\u5ea6\u9700\u8981\u8fbe\u5230\u7ea6 200 FLOPs\/byte \u624d\u80fd\u4e0d\u88ab\u5e26\u5bbd\u62d6\u6162\uff1b\u82e5\u6309 80GB \u7248\u8ba1\u7b97\u5219\u7ea6\u4e3a 156 FLOPs\/byte\u3002\u5377\u79ef\u672c\u8eab\u7684\u8ba1\u7b97\u5bc6\u5ea6\u8f83\u9ad8\uff0c\u901a\u5e38\u80fd\u8fbe\u5230\u8fd9\u4e2a\u95e8\u69db\uff1b\u4f46 BN \u548c ReLU \u90fd\u662f\u9010\u5143\u7d20\u64cd\u4f5c\uff0c\u8ba1\u7b97\u91cf\u6781\u5c0f\u3001\u8bbf\u5b58\u91cf\u6781\u5927\uff0c\u662f\u5178\u578b\u7684”\u5185\u5b58\u5899”\u64cd\u4f5c\u3002\u5982\u679c\u80fd\u8ba9\u8fd9\u4e9b\u9010\u5143\u7d20\u64cd\u4f5c\u7d27\u6328\u7740\u5377\u79ef\u6267\u884c\uff0c\u628a\u4e2d\u95f4\u7ed3\u679c\u7559\u5728\u5bc4\u5b58\u5668\u6216\u5171\u4eab\u5185\u5b58\u91cc\uff0c\u800c\u4e0d\u662f\u6765\u56de\u642c\u8fd0\u5230\u5168\u5c40\u663e\u5b58\uff0c\u5c31\u80fd\u663e\u8457\u51cf\u5c11\u5e26\u5bbd\u538b\u529b\u3002<\/p>\n\n\n\n
    \u6240\u4ee5 CBR \u878d\u5408\u6700\u6839\u672c\u7684\u6536\u76ca\u4e0d\u662f”\u7b97\u5f97\u66f4\u5feb”\uff0c\u800c\u662f\u5c11\u8bfb\u5c11\u5199\u4e2d\u95f4\u7ed3\u679c<\/strong>\u3002\u8fd9\u4e5f\u662f\u540e\u9762\u6240\u6709\u8bbe\u8ba1\u51b3\u7b56\u7684\u51fa\u53d1\u70b9\u3002<\/p>\n\n\n\n
    \u4e3e\u4e2a\u5177\u4f53\u4f8b\u5b50\u3002\u5047\u8bbe\u4e00\u4e2a\u4e2d\u95f4\u5c42\u8f93\u51fa\u5c3a\u5bf8\u4e3a N=256, H=W=56, C=64<\/code>\uff0c\u4f7f\u7528 FP16\uff0c\u90a3\u4e48\u4e00\u5f20\u5b8c\u6574\u7279\u5f81\u56fe\u5360\u7528\uff1a<\/p>\n\n\n\n
    bytes = N * H * W * C * sizeof(half);   \/\/ sizeof(half) == 2\n      = 256 * 56 * 56 * 64 * 2\n      \u2248 102 MB<\/pre>\n\n\n\n
    \u5206\u7acb\u5b9e\u73b0\u4e0b\uff0cConv \u5199\u4e00\u6b21 conv_output<\/code>\u3001BN \u8bfb\u5b83\u518d\u5199 bn_output<\/code>\u3001ReLU \u8bfb bn_output<\/code> \u518d\u5199\u6700\u7ec8\u7ed3\u679c\uff0c\u4ec5\u8fd9\u4e00\u5c42\u5c31\u591a\u4ea7\u751f\u7ea6 200 MB \u7684\u663e\u5b58\u8bfb\u5199\u3002ResNet-50 \u6709 50 \u591a\u5c42\u5377\u79ef\uff0c\u800c VGG16BN \u867d\u7136\u53ea\u6709 13 \u4e2a\u5377\u79ef\u5c42\uff0c\u4f46\u7279\u5f81\u56fe\u5206\u8fa8\u7387\u66f4\u9ad8\u3001\u901a\u9053\u6570\u66f4\u5927\uff0c\u5355\u5c42\u7279\u5f81\u56fe\u7684\u6570\u636e\u91cf\u66f4\u5927\u3002CBR \u878d\u5408\u628a\u5176\u4e2d\u4e00\u6b21\u5b8c\u6574\u7279\u5f81\u56fe\u7684\u5199\u548c\u4e00\u6b21\u8bfb\u7701\u6389\uff0c\u6536\u76ca\u5c31\u8fd9\u6837\u4e00\u5c42\u5c42\u653e\u5927\u3002<\/p>\n\n\n\n
    \u4e8c\u3001\u8bad\u7ec3\u6bd4\u63a8\u7406\u66f4\u96be\u878d\u5408\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u76f4\u63a5\u6298\u53e0 BN<\/h2>\n\n\n\n
    \u5728\u63a8\u7406\u573a\u666f\uff0cCBR \u878d\u5408\u6709\u4e00\u4e2a\u975e\u5e38\u6f02\u4eae\u7684\u6570\u5b66\u6377\u5f84\uff1a\u628a BatchNorm \u7684\u53c2\u6570\u6298\u53e0\u8fdb\u5377\u79ef\u7684\u6743\u91cd\u91cc<\/strong>\u3002<\/p>\n\n\n\n
    \u5047\u8bbe\u5377\u79ef\u8f93\u51fa\u4e3a\uff1a<\/p>\n\n\n\n
    y = conv(x, W, b);<\/pre>\n\n\n\n
    BN \u7684\u53d8\u6362\u4e3a\uff1a<\/p>\n\n\n\n
    bn(y) = gamma * (y - mean) \/ sqrt(var + eps) + beta;<\/pre>\n\n\n\n
    \u628a\u4e24\u5f0f\u5408\u5e76\uff0c\u53ef\u5f97\u5230\u4e00\u4e2a\u7b49\u4ef7\u7684\u5377\u79ef\uff1a<\/p>\n\n\n\n
    W_fold = W * gamma \/ sqrt(var + eps);\nb_fold = (b - mean) * gamma \/ sqrt(var + eps) + beta;\nout    = conv(x, W_fold, b_fold);\nout    = relu(out);<\/pre>\n\n\n\n
    \u53ea\u8981\u63a8\u7406\u65f6\u4f7f\u7528\u56fa\u5b9a\u7684 running_mean<\/code> \u548c running_var<\/code>\uff0c\u8fd9\u4e2a\u6298\u53e0\u662f\u6570\u5b66\u7cbe\u786e<\/strong>\u7684\u3002TensorRT\u3001ONNX Runtime\u3001TensorFlow Lite \u7b49\u63a8\u7406\u5f15\u64ce\u666e\u904d\u91c7\u7528\u8fd9\u79cd\u7b56\u7565\uff0c\u6536\u76ca\u975e\u5e38\u76f4\u63a5\u3002<\/p>\n\n\n\n
    \u4f46\u8bad\u7ec3\u573a\u666f\u4e0d\u884c\u3002\u8bad\u7ec3\u65f6 BN \u7684\u5747\u503c\u548c\u65b9\u5dee\u6765\u81ea\u5f53\u524d batch\uff0c\u8fd8\u8981\u66f4\u65b0 running statistics\uff0c\u5e76\u4e14\u53cd\u5411\u4f20\u64ad\u9700\u8981\u4fdd\u5b58 saved_mean<\/code> \u548c saved_inv_var<\/code> \u6765\u8ba1\u7b97 d_scale<\/code> \u548c d_bias<\/code>\u3002\u4f60\u4e0d\u80fd\u7b80\u5355\u5730\u628a BN \u6298\u53e0\u8fdb\u5377\u79ef\u6743\u91cd\uff0c\u56e0\u4e3a\u6bcf\u6b21\u8fed\u4ee3\u7684\u7edf\u8ba1\u91cf\u90fd\u5728\u53d8\u3002\u56e0\u6b64\uff0c\u8bad\u7ec3\u9636\u6bb5\u7684 CBR \u878d\u5408\u5fc5\u987b\u4fdd\u7559 BN \u7684\u5b8c\u6574\u8bed\u4e49\uff0c\u540c\u65f6\u5c3d\u91cf\u628a\u5377\u79ef\u3001\u7edf\u8ba1\u91cf\u751f\u6210\u3001\u5f52\u4e00\u5316\u3001\u6fc0\u6d3b\u5728\u7269\u7406\u4e0a\u9760\u5f97\u66f4\u8fd1\u3002<\/p>\n\n\n\n
    \u8bad\u7ec3\u65f6 BN \u7684\u53cd\u5411\u4f20\u64ad\u6d89\u53ca\u4e09\u4e2a\u68af\u5ea6\uff1a\u4e0a\u6e38\u68af\u5ea6 dY<\/code>\u3001\u8f93\u5165 X<\/code> \u7684\u68af\u5ea6 dX<\/code>\u3001\u4ee5\u53ca\u53c2\u6570 gamma<\/code>\/beta<\/code> \u7684\u68af\u5ea6\u3002\u8981\u7b97 dX<\/code>\uff0c\u5fc5\u987b\u77e5\u9053\u5f53\u524d batch \u7684 saved_mean<\/code> \u548c saved_inv_var<\/code>\uff1b\u8981\u66f4\u65b0 running statistics<\/code>\uff0c\u53c8\u9700\u8981 batch \u7684 mean \u548c var\u3002\u8fd9\u4e9b\u91cf\u90fd\u65e0\u6cd5\u50cf\u63a8\u7406\u90a3\u6837\u9884\u5148\u6298\u53e0\u8fdb\u5377\u79ef\u6743\u91cd\u3002\u66f4\u8fdb\u4e00\u6b65\uff0cBN \u7684\u53cd\u5411\u901a\u5e38\u9700\u8981\u5148\u505a\u4e00\u6b21 dY * mask<\/code> \u6765\u5904\u7406 ReLU \u7684\u6b7b\u4ea1\u795e\u7ecf\u5143\uff0c\u7136\u540e\u624d\u80fd\u8fdb\u5165 batchnorm_backward<\/code>\u3002\u6240\u4ee5\u8bad\u7ec3\u8def\u5f84\u7684 CBR \u878d\u5408\u4e0d\u662f\u6d88\u53bb BN\uff0c\u800c\u662f\u628a”ReLU \u63a9\u7801\u5e94\u7528 + BN \u53cd\u5411 + \u5377\u79ef\u68af\u5ea6”\u7684\u6d41\u6c34\u7ebf\u5c3d\u91cf\u7d27\u51d1\u5730\u7ec4\u7ec7\u8d77\u6765\u3002<\/p>\n\n\n\n
    \u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u4e3b\u6d41\u8bad\u7ec3\u6846\u67b6\u5bf9 CBR \u878d\u5408\u76f8\u5bf9\u4fdd\u5b88\u3002PyTorch \u7684 torch.compile<\/code> \/ TorchInductor \u53ef\u4ee5\u5728 FX \u56fe\u7ea7\u522b\u505a\u7b97\u5b50\u878d\u5408\uff0c\u4f46\u6548\u679c\u53d7\u52a8\u6001\u56fe\u3001shape \u53d8\u5316\u3001CUDA Graph \u7a33\u5b9a\u6027\u7b49\u56e0\u7d20\u5f71\u54cd\uff1bTensorFlow \u7684 XLA \u4e5f\u53d6\u51b3\u4e8e\u56fe\u80fd\u5426\u88ab\u5b8c\u6574\u6355\u83b7\u548c\u7f16\u8bd1\u3002Tech-Renaissance \u7684\u9009\u62e9\u662f\uff1a\u5728 AMP \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u8def\u5f84\u4e0a\uff0c\u628a CBR \u4f5c\u4e3a\u4e00\u4e2a\u4e00\u7b49\u7b97\u5b50\u5b8c\u6574\u5b9e\u73b0<\/strong>\uff0c\u5305\u62ec\u524d\u5411\u3001\u53cd\u5411\u3001\u9996\u5c42\u7279\u5316\u548c\u63a8\u7406\u8def\u5f84\u3002<\/p>\n\n\n\n
    \u4e09\u3001\u4ece BluePrint \u5230\u8ba1\u7b97\u56fe\uff1a\u8ba9\u6846\u67b6\u81ea\u52a8\u8ba4\u51fa CBR<\/h2>\n\n\n\n
    \u5728 Tech-Renaissance \u7684\u9ad8\u5c42 DSL BluePrint<\/code> \u91cc\uff0c\u7528\u6237\u53ef\u4ee5\u76f4\u63a5\u5199\uff1a<\/p>\n\n\n\n
    auto layer = cbr(64, 3, 1, 1);        \/\/ out_ch=64, kernel=3, stride=1, padding=1\n\/\/ \u6216\u8005\u7b49\u4ef7\u5199\u6cd5\nauto layer = conv_bn_relu(64, 3, 1, 1);<\/pre>\n\n\n\n
    \u8fd9\u4f1a\u751f\u6210\u4e00\u4e2a NodeKind::CBR<\/code> \u8282\u70b9\u3002\u5728 ArchPlan::step9_merge_triple()<\/code> \u4e2d\uff0c\u5f53 AMP \u5f00\u542f\u65f6\uff0c\u6846\u67b6\u4f1a\u81ea\u52a8\u626b\u63cf\u8fde\u7eed\u7684 Conv \u2192 Bn2d \u2192 ReLU<\/code> \u4e09\u5c42\uff0c\u628a\u5b83\u4eec\u5408\u5e76\u6210\u4e00\u4e2a LayerKind::CBR<\/code>\uff1a<\/p>\n\n\n\n
    merge_pattern_triple(LayerKind::Conv, LayerKind::Bn2d, LayerKind::ReLU, LayerKind::CBR, build_cbr);<\/pre>\n\n\n\n
    \u4e5f\u5c31\u662f\u8bf4\uff0c\u5373\u4f7f\u7528\u6237\u5199\u7684\u662f\u4e09\u4e2a\u72ec\u7acb\u5c42\uff0c\u53ea\u8981\u5f00\u542f\u4e86 AMP\uff0c\u7f16\u8bd1\u5668\u4e5f\u4f1a\u5c1d\u8bd5\u628a\u5b83\u4eec\u7194\u94f8\u6210\u4e00\u4e2a\u878d\u5408\u7b97\u5b50\u3002\u8fd9\u79cd”\u56fe\u7ea7\u6a21\u5f0f\u5339\u914d + \u7b97\u5b50\u7ea7\u7279\u5316\u5b9e\u73b0”\u7684\u5206\u5c42\u8bbe\u8ba1\uff0c\u662f\u9759\u6001\u56fe\u6846\u67b6\u7684\u4f18\u52bf\uff1a\u6211\u4eec\u5728\u7f16\u8bd1\u671f\u5c31\u77e5\u9053\u6574\u4e2a\u7f51\u7edc\u957f\u4ec0\u4e48\u6837\uff0c\u53ef\u4ee5\u5927\u80c6\u5730\u505a\u5168\u5c40\u91cd\u5199\uff0c\u800c\u4e0d\u5fc5\u5728\u8fd0\u884c\u65f6\u52a8\u6001\u6355\u83b7\u56fe\u3002<\/p>\n\n\n\n
    \u5728 Compiler<\/code> \u6784\u5efa\u8ba1\u7b97\u56fe\u65f6\uff0cCBR \u5c42\u88ab\u5206\u914d\u5b8c\u6574\u7684\u53c2\u6570\u7ec4\uff1a\u5377\u79ef\u6743\u91cd\u3001BN \u7684 gamma\/beta\u3001BN \u7684 running mean\/var\uff0c\u4ee5\u53ca\u7528\u4e8e\u53cd\u5411\u7684\u68af\u5ea6\u5360\u4f4d\u3002src\/graph\/layer_descriptor_registry.cpp<\/code> \u4e2d\u4e3a CBR \u5b9a\u4e49\u4e86 23 \u4e2a\u5f20\u91cf<\/strong>\uff08Conv \u90e8\u5206 8 \u4e2a\u3001BN \u90e8\u5206 13 \u4e2a\u3001ReLU \u90e8\u5206 2 \u4e2a\uff09\uff0c\u8fd9\u4e9b\u5f20\u91cf\u7684\u5e03\u5c40\u3001\u6570\u636e\u7c7b\u578b\u3001\u5185\u5b58\u533a\u57df\u5728 MemoryPlan<\/code> \u9636\u6bb5\u5c31\u5168\u90e8\u786e\u5b9a\u4e0b\u6765\u3002\u8fd0\u884c\u65f6\uff0c\u524d\u5411\u7b97\u5b50 CBR_AMP_FWD<\/code> \u7684\u5b9e\u9645\u8f93\u5165\u8f93\u51fa\u4e3a\uff1a<\/p>\n\n\n\n
      \n
    • \u8f93\u5165\uff088 \u4e2a\uff09<\/strong>\uff1aX<\/code>\uff08FP16 \u7279\u5f81\u56fe\uff09\u3001amp_w<\/code>\uff08FP16 \u5377\u79ef\u6838\uff09\u3001bn_w<\/code>\uff08FP32 gamma\uff09\u3001bn_b<\/code>\uff08FP32 beta\uff09\u3001prev_mean<\/code>\uff08FP32 running mean\uff09\u3001prev_var<\/code>\uff08FP32 running var\uff09\u3001eps<\/code>\uff08FP32 \u6807\u91cf\uff09\u3001mom<\/code>\uff08FP32 \u6807\u91cf\uff09\uff1b<\/li>\n\n\n\n
    • \u8f93\u51fa\uff0810 \u4e2a\uff09<\/strong>\uff1aconv_output<\/code>\u3001bn_sum<\/code>\u3001bn_sq_sum<\/code>\u3001bn_output<\/code>\u3001saved_mean<\/code>\u3001saved_inv_var<\/code>\u3001relu_output<\/code>\u3001relu_mask<\/code>\u3001next_mean<\/code>\u3001next_var<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n
      \u5176\u4e2d X<\/code> \u7531\u7f16\u8bd1\u5668\u81ea\u52a8\u524d\u7f6e\uff0ceps<\/code> \u548c mom<\/code> \u662f\u5168\u5c40\u6807\u91cf\u5f20\u91cf ID\uff1bbn_output<\/code> \u5728\u8bad\u7ec3\u524d\u5411\u4e2d\u5e76\u4e0d\u5b58\u653e\u5b8c\u6574\u7279\u5f81\u56fe\uff0c\u800c\u662f\u88ab\u590d\u7528\u6765\u653e\u7f6e eq_scale<\/code>\/eq_bias<\/code> \u8fd9\u4e24\u4e2a\u5c0f\u5411\u91cf\u3002\u8fd9\u79cd\u663e\u5b58\u590d\u7528\u53ea\u6709\u5728 MemoryPlan \u7684\u9759\u6001\u89c4\u5212\u4e0b\u624d\u662f\u5b89\u5168\u7684\u3002<\/p>\n\n\n\n
      \u56db\u3001\u524d\u5411\uff1a\u4e09\u6bb5\u5f0f cuDNN Graph\uff0c\u4e2d\u95f4\u7ed3\u679c\u80fd\u4e0d\u5199\u5c31\u4e0d\u5199<\/h2>\n\n\n\n
      CBR \u524d\u5411\u7b97\u5b50 CBR_AMP_FWD<\/code> \u7684\u5b9e\u73b0\u4f4d\u4e8e src\/backend\/ops\/dtensor\/cbr_op.cpp<\/code>\u3002\u5b83\u7684\u6838\u5fc3\u4e0d\u662f\u5199\u4e00\u4e2a\u5de8\u5927\u7684\u81ea\u5b9a\u4e49 CUDA kernel\uff0c\u800c\u662f\u628a\u95ee\u9898\u62c6\u6210\u4e09\u5f20 cuDNN Frontend Graph<\/strong>\uff0c\u5206\u522b\u8dd1\u5728\u4e09\u6761\u8ba1\u7b97\u6d41\u4e0a\uff0c\u7528 CUDA event \u505a\u8de8\u6d41\u540c\u6b65\u3002<\/p>\n\n\n\n
      4.1 Conv + GenStats\uff08COMP_1\uff09<\/h3>\n\n\n\n
      \u7b2c\u4e00\u5f20\u56fe\u53ea\u505a\u4e24\u4ef6\u4e8b\uff1a\u7528 conv_fprop<\/code> \u8ba1\u7b97\u5377\u79ef\u8f93\u51fa\uff1b\u7528 genstats<\/code> \u4ece\u5377\u79ef\u8f93\u51fa\u76f4\u63a5\u751f\u6210\u901a\u9053\u7ea7\u7684 sum<\/code> \u548c sq_sum<\/code>\u3002<\/p>\n\n\n\n
      auto conv_out = graph->conv_fprop(X, W, conv_opts);\nauto [sum, sq_sum] = graph->genstats(conv_out, genstats_opts);<\/pre>\n\n\n\n
      \u8fd9\u4e00\u6b65\u7684\u5173\u952e\u5728\u4e8e genstats<\/code>\u3002\u5728\u5355\u72ec\u5b9e\u73b0 BN \u65f6\uff0c\u6846\u67b6\u9700\u8981\u5148\u628a conv_output<\/code> \u5199\u56de\u5168\u5c40\u663e\u5b58\uff0c\u518d\u8bfb\u53d6\u4e00\u6b21\u6765\u8ba1\u7b97\u5747\u503c\u548c\u65b9\u5dee\uff1b\u800c cuDNN \u7684 genstats<\/code> \u53ef\u4ee5\u5728\u5377\u79ef forward \u7684\u540c\u65f6\uff0c\u4ece\u5bc4\u5b58\u5668\u6216\u5171\u4eab\u5185\u5b58\u91cc\u76f4\u63a5\u5f52\u7ea6\u51fa\u7edf\u8ba1\u91cf\u3002\u8fd9\u610f\u5473\u7740\u6211\u4eec\u7701\u6389\u4e86”\u4e3a\u4e86\u7b97 BN \u800c\u518d\u8bfb\u4e00\u904d\u6574\u5f20\u7279\u5f81\u56fe”\u7684\u5e26\u5bbd\u5f00\u9500\u3002\u5bf9\u4e8e 256\u00d756\u00d756\u00d764<\/code> \u7684 FP16 \u7279\u5f81\u56fe\uff0c\u8fd9\u4e00\u6b21\u5c31\u907f\u514d\u4e86\u7ea6 102 MB \u7684\u5168\u5c40\u663e\u5b58\u8bfb\u53d6\u3002<\/p>\n\n\n\n
      4.2 BN Finalize\uff08COMP_2\uff09<\/h3>\n\n\n\n
      \u62ff\u5230 sum<\/code> \u548c sq_sum<\/code> \u540e\uff0c\u7b2c\u4e8c\u5f20\u56fe bn_finalize<\/code> \u5728\u7eaf FP32 \u7cbe\u5ea6\u4e0b\u5b8c\u6210\uff1a<\/p>\n\n\n\n
        \n
      • \u8ba1\u7b97 saved_mean<\/code>\u3001saved_inv_var<\/code>\uff1b<\/li>\n\n\n\n
      • \u66f4\u65b0 next_mean<\/code>\u3001next_var<\/code>\uff1b<\/li>\n\n\n\n
      • \u8ba1\u7b97\u672c\u6b21\u524d\u5411\u771f\u6b63\u7528\u4e8e\u7f29\u653e\u504f\u79fb\u7684 eq_scale<\/code> \u548c eq_bias<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n
        \u5b83\u4eec\u7684\u8ba1\u7b97\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n
        N_ac        = N * H * W;                              \/\/ \u6bcf\u4e2a\u901a\u9053\u7d2f\u79ef\u7684\u6837\u672c\u6570\nmean        = sum   \/ N_ac;\nvar         = sq_sum \/ N_ac - mean * mean;\ninv_std     = 1.0f \/ sqrtf(var + eps);\nsaved_mean  = mean;                                   \/\/ \u4fdd\u5b58\u7ed9\u53cd\u5411\nsaved_inv_var = inv_std;                              \/\/ \u4fdd\u5b58\u7ed9\u53cd\u5411\neq_scale    = gamma * inv_std;                        \/\/ \u7528\u4e8e Apply \u7684\u7b49\u4ef7\u7f29\u653e\neq_bias     = beta - gamma * mean * inv_std;          \/\/ \u7528\u4e8e Apply \u7684\u7b49\u4ef7\u504f\u79fb\nnext_mean   = (1 - momentum) * prev_mean + momentum * mean;\nnext_var    = (1 - momentum) * prev_var  + momentum * var;<\/pre>\n\n\n\n
        \u8fd9\u91cc sum<\/code>\u3001sq_sum<\/code> \u662f\u4e24\u4e2a\u957f\u5ea6\u4e3a C<\/code> \u7684 FP32 \u5c0f\u5411\u91cf\uff0c\u8ba1\u7b97\u91cf\u6781\u5c0f\uff0c\u51e0\u4e4e\u53ef\u4ee5\u5b8c\u5168\u9690\u85cf\u5728 Conv+GenStats \u7684\u6267\u884c\u8fc7\u7a0b\u4e2d\u3002eq_scale<\/code> \u548c eq_bias<\/code> \u88ab\u5199\u5165 bn_output<\/code> \u7f13\u51b2\u533a\u7684\u524d 2 * K * sizeof(float)<\/code> \u5b57\u8282\uff0csaved_mean<\/code>\u3001saved_inv_var<\/code>\u3001next_mean<\/code>\u3001next_var<\/code> \u5219\u5404\u81ea\u5199\u5165\u72ec\u7acb\u7684 FP32 \u7f13\u51b2\u533a\u3002\u5176\u4e2d momentum<\/code> \u662f cuDNN \u98ce\u683c\u7684 exponential average factor\uff08\u65b0 batch \u7edf\u8ba1\u91cf\u7684\u6743\u91cd\uff0c\u9ed8\u8ba4 0.1\uff09\u3002<\/p>\n\n\n\n
        4.3 BN Apply + ReLU\uff08COMP_3\uff09<\/h3>\n\n\n\n
        \u7b2c\u4e09\u5f20\u56fe\u63a5\u6536\u5377\u79ef\u8f93\u51fa\u3001eq_scale<\/code>\/eq_bias<\/code>\uff0c\u628a BN \u4eff\u5c04\u53d8\u6362\u548c ReLU \u6fc0\u6d3b\u878d\u5408\u5728\u4e00\u8d77\uff1a<\/p>\n\n\n\n
        auto scaled  = graph->pointwise(conv_out, eq_scale_ta, MUL);\nauto shifted = graph->pointwise(scaled,  eq_bias_ta,  ADD);\nauto relu    = graph->pointwise(shifted, RELU_FWD);\nauto mask    = graph->pointwise(shifted, zero, CMP_GT);   \/\/ shifted > 0<\/pre>\n\n\n\n
        \u540c\u65f6\u7528 CMP_GT<\/code> \u751f\u6210\u4e00\u4e2a\u4f4d\u538b\u7f29\u7684\u5e03\u5c14\u63a9\u7801<\/strong>\uff08cuDNN \u7684 BOOLEAN<\/code> \u7c7b\u578b\uff0c\u6bcf\u4e2a\u5143\u7d20 1 bit\uff09\uff0c\u4fdd\u5b58\u4e0b\u6765\u7ed9\u53cd\u5411\u4f20\u64ad\u7528\u3002\u8fd9\u4e0e\u5206\u7acb ReLU \u7b97\u5b50\u5e38\u7528\u7684\u6bcf\u5143\u7d20 1 \u5b57\u8282 INT8 mask \u4e0d\u540c\uff1aCBR \u5185\u90e8\u4f7f\u7528\u4f4d\u538b\u7f29\u683c\u5f0f\u6765\u8282\u7701\u663e\u5b58\uff0c\u6d4b\u8bd5\u4ee3\u7801\u91cc\u4e13\u95e8\u505a\u4e86\u4e24\u79cd mask \u683c\u5f0f\u4e4b\u95f4\u7684\u8f6c\u6362\uff0c\u4ee5\u4fdd\u8bc1\u7b49\u4ef7\u6027\u6821\u9a8c\u901a\u8fc7\u3002<\/p>\n\n\n\n
        4.4 \u591a\u6d41\u540c\u6b65<\/h3>\n\n\n\n
        \u4e09\u4e2a\u9636\u6bb5\u5206\u522b\u8dd1\u5728 COMP_1<\/code>\u3001COMP_2<\/code>\u3001COMP_3<\/code> \u4e0a\u3002MultiStreamCaptureState<\/code> \u8d1f\u8d23\u6ce8\u518c\u6d41\u5e76\u5728\u9636\u6bb5\u4e4b\u95f4\u63d2\u5165 cudaStreamWaitEvent<\/code>\uff1aFinalize \u7b49 Conv+GenStats \u5b8c\u6210\uff0cApply+ReLU \u7b49 Finalize \u5b8c\u6210\u3002\u8fd9\u5957\u673a\u5236\u4e0e\u6211\u4eec\u5728\u7b2c 17 \u7bc7 CUDA Graph \u5168\u6355\u83b7\u4e2d\u8bb2\u7684\u591a\u6d41\u4f9d\u8d56\u7ba1\u7406\u662f\u540c\u4e00\u5957\u57fa\u7840\u8bbe\u65bd\uff0c\u4fdd\u8bc1\u4e86\u5373\u4f7f\u7b97\u5b50\u5185\u90e8\u62c6\u6210\u591a\u4e2a\u6d41\uff0c\u5bf9\u5916\u4ecd\u7136\u5448\u73b0\u4e3a\u4e00\u4e2a\u539f\u5b50\u7684 CBR_AMP_FWD<\/code> \u8282\u70b9\u3002<\/p>\n\n\n\n
        \u4f60\u53ef\u80fd\u4f1a\u95ee\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u76f4\u63a5\u7528\u4e00\u5f20\u5de8\u5927\u7684\u56fe\u628a Conv+BN+ReLU \u5168\u5305\u8fdb\u53bb\uff1f\u539f\u56e0\u5728\u4e8e cuDNN \u5f53\u524d\u5bf9\u8bad\u7ec3\u8def\u5f84\u7684 Conv+GenStats+BN finalize+Apply+ReLU \u8fd9\u4e94\u8fde\u64cd\u4f5c\uff0c\u5e76\u4e0d\u603b\u80fd\u6210\u529f\u5b9e\u4f8b\u5316\u4e3a\u4e00\u5f20\u53ef\u6267\u884c\u56fe\uff1b\u800c\u62c6\u6210\u4e09\u6bb5\u540e\uff0c\u6bcf\u4e00\u6bb5\u90fd\u66f4\u5c0f\u3001\u66f4\u7a33\u5b9a\uff0c\u540c\u65f6\u901a\u8fc7\u591a\u6d41\u91cd\u53e0\u8ba9\u4e09\u6bb5\u4e4b\u95f4\u7684\u7b49\u5f85\u65f6\u95f4\u88ab\u5176\u4ed6\u8ba1\u7b97\u9690\u85cf\u3002\u53e6\u5916\uff0c\u5206\u6bb5\u4e5f\u8ba9\u6211\u4eec\u80fd\u5728 BN finalize \u9636\u6bb5\u590d\u7528 bn_output<\/code> \u7f13\u51b2\u533a\u653e eq_scale<\/code>\/eq_bias<\/code>\uff0c\u8fd9\u5bf9\u663e\u5b58\u7ba1\u7406\u5f88\u5173\u952e\u3002<\/p>\n\n\n\n
        \u6bcf\u5f20\u56fe\u9996\u6b21\u6784\u5efa\u540e\u90fd\u4f1a\u88ab\u7f13\u5b58\u5230\u4ee5 cuDNN handle\u3001\u5f62\u72b6\u3001pad\/stride \u4e3a\u952e\u7684 unordered_map<\/code> \u4e2d\uff0c\u540e\u7eed\u540c\u5f62\u72b6 batch \u76f4\u63a5\u590d\u7528\uff0c\u4e0d\u9700\u8981\u91cd\u590d build\u3002<\/p>\n\n\n\n
        \u4e94\u3001\u53cd\u5411\uff1a\u628a BN \u548c ReLU \u7684\u68af\u5ea6\u4e5f\u7b97\u8fdb\u4e00\u5f20\u56fe<\/h2>\n\n\n\n
        CBR \u7684\u53cd\u5411 CBR_AMP_BWD<\/code> \u540c\u6837\u4e0d\u662f\u7b80\u5355\u5730”\u5206\u522b\u8c03\u7528 Conv BWD\u3001BN BWD\u3001ReLU BWD”\u3002\u5b83\u7684\u6267\u884c\u4e5f\u5206\u5e03\u5728\u4e09\u6761\u6d41\u4e0a\uff1a<\/p>\n\n\n\n
          \n
        • COMP_1<\/strong>\uff1a\u6267\u884c BN+ReLU BWD \u878d\u5408\u5b50\u56fe\uff1b<\/li>\n\n\n\n
        • COMP_3<\/strong>\uff1a\u6267\u884c WGrad\uff08\u6743\u91cd\u68af\u5ea6\uff09\uff1b<\/li>\n\n\n\n
        • COMP_2<\/strong>\uff1a\u6267\u884c DGrad\uff08\u6570\u636e\u68af\u5ea6\uff09\u3002<\/li>\n<\/ul>\n\n\n\n
          5.1 BN + ReLU \u53cd\u5411\u878d\u5408\uff08COMP_1\uff09<\/h3>\n\n\n\n
          \u5148\u7528 MUL(dY, mask)<\/code> \u628a\u4e0a\u6e38\u68af\u5ea6\u4e0e ReLU \u524d\u5411\u63a9\u7801\u76f8\u4e58\uff0c\u5c4f\u853d\u6389 ReLU \u5173\u95ed\u7684\u795e\u7ecf\u5143\uff1b\u7136\u540e\u76f4\u63a5\u8c03\u7528 batchnorm_backward<\/code>\uff0c\u8f93\u51fa\u4e09\u6837\u4e1c\u897f\uff1a\u4f20\u5230\u5377\u79ef\u7684\u68af\u5ea6 dL\/d(conv_out)<\/code>\u3001BN gamma \u7684\u68af\u5ea6 d_scale<\/code>\u3001BN beta \u7684\u68af\u5ea6 d_bias<\/code>\u3002<\/p>\n\n\n\n
          auto dy_masked = graph->pointwise(dY_ta, mask_ta, MUL);\nauto [dx_bn, dscale, dbias] = graph->batchnorm_backward(dy_masked, x_ta, scale_ta, bn_opts);<\/pre>\n\n\n\n
          \u8fd9\u4e00\u6b65\u88ab\u5c01\u88c5\u5728\u4e00\u5f20 cuDNN Frontend Graph \u91cc\uff0cdy_masked<\/code> \u4e0d\u9700\u8981\u5199\u56de HBM\u3002\u8f93\u51fa dL\/d(conv_out)<\/code> \u88ab\u5199\u8fdb bn_output<\/code> \u7f13\u51b2\u533a\u2014\u2014\u4e5f\u5c31\u662f\u524d\u5411\u91cc\u90a3\u4e2a\u53ea\u5b58\u4e86 eq_scale<\/code>\/eq_bias<\/code> \u7684\u7f13\u51b2\u533a\u2014\u2014\u56e0\u4e3a\u8fdb\u5165\u53cd\u5411\u4e4b\u540e\uff0c\u524d\u5411\u7684 eq_scale<\/code>\/eq_bias<\/code> \u5df2\u7ecf\u4e0d\u518d\u9700\u8981\uff0c\u8fd9\u5757\u5185\u5b58\u53ef\u4ee5\u5b89\u5168\u590d\u7528\u3002<\/p>\n\n\n\n
          5.2 WGrad\uff08COMP_3\uff09\u4e0e DGrad\uff08COMP_2\uff09<\/h3>\n\n\n\n
          bn_output<\/code> \u6b64\u65f6\u4fdd\u5b58\u7684\u662f dL\/d(conv_out)<\/code>\u3002WGrad \u7528\u5b83\u548c\u539f\u59cb\u8f93\u5165 X<\/code> \u8ba1\u7b97\u5377\u79ef\u6743\u91cd\u7684\u68af\u5ea6 dW<\/code>\u3002DGrad \u7528 dL\/d(conv_out)<\/code> \u548c\u5377\u79ef\u6743\u91cd\u8ba1\u7b97\u8f93\u5165 X<\/code> \u7684\u68af\u5ea6 dX<\/code>\u3002<\/p>\n\n\n\n
          \u6e90\u7801\u91cc\u660e\u786e\u6ce8\u91ca\u4e86 DGrad \u4e3a\u4ec0\u4e48\u8981\u7b49 WGrad \u5b8c\u6210\uff1a<\/p>\n\n\n\n
          \n
          “\u7b49\u5f85 WGrad\uff0c\u786e\u4fdd WGrad \u5df2\u8bfb\u5b8c\u539f\u59cb X \u540e DGrad \u624d\u8986\u76d6 X\u3002”<\/p>\n<\/blockquote>\n\n\n\n
          \u8fd9\u4e2a\u540c\u6b65\u4e0d\u662f\u6570\u636e\u4f9d\u8d56\u2014\u2014DGrad \u4e0d\u4f9d\u8d56 WGrad \u7684\u8f93\u51fa\u2014\u2014\u800c\u662f\u56e0\u4e3a\u4e24\u8005\u90fd\u8bfb\u53d6\u539f\u59cb\u8f93\u5165 X<\/code>\uff0c\u800c DGrad \u7684\u8f93\u51fa dX<\/code> \u4f1a\u539f\u5730\u8986\u76d6 X<\/code> \u7684\u5b58\u50a8\u4f4d\u7f6e<\/strong>\u3002\u5982\u679c DGrad \u5728 WGrad \u4e4b\u524d\u542f\u52a8\uff0c\u5b83\u53ef\u80fd\u4f1a\u5728 WGrad \u8bfb\u53d6 X<\/code> \u4e4b\u524d\u5c31\u8986\u76d6\u4e86 X<\/code> \u7684\u5185\u5bb9\uff0c\u5bfc\u81f4 WGrad \u8ba1\u7b97\u51fa\u9519\u8bef\u7684\u68af\u5ea6\u3002\u8fd9\u662f\u878d\u5408\u7b97\u5b50\u4e2d\u9700\u8981\u7279\u522b\u5c0f\u5fc3\u5904\u7406\u7684\u8bfb\u5199\u51b2\u7a81\u3002<\/p>\n\n\n\n
          5.3 \u9996\u5c42\u7279\u5316<\/h3>\n\n\n\n
          \u5bf9\u4e8e\u7f51\u7edc\u7684\u7b2c\u4e00\u5c42\uff0cCBR_AMP_BWD_FIRST_LAYER<\/code> \u8df3\u8fc7\u4e86 DGrad\u2014\u2014\u56e0\u4e3a\u9996\u5c42\u7684\u8f93\u5165\u662f\u539f\u59cb\u8bad\u7ec3\u6570\u636e\uff0c\u4e0d\u662f\u4e2d\u95f4\u7279\u5f81\u56fe\uff0c\u4e0d\u9700\u8981\u8ba1\u7b97\u8f93\u5165\u6570\u636e\u7684\u68af\u5ea6\u3002\u8fd9\u8fdb\u4e00\u6b65\u8282\u7701\u4e86\u9996\u5c42\u7684\u8ba1\u7b97\u5f00\u9500\u3002\u5728\u7f16\u8bd1\u5668\u5c42\u9762\uff0c\u9996\u5c42\u548c\u975e\u9996\u5c42\u7684 BWD \u7b97\u5b50\u901a\u8fc7 to_first_layer_bwd_op<\/code> \u51fd\u6570\u8fdb\u884c\u81ea\u52a8\u8f6c\u6362\uff0c\u7528\u6237\u65e0\u9700\u5173\u5fc3\u8fd9\u4e2a\u5dee\u5f02\u3002<\/p>\n\n\n\n
          \u516d\u3001\u63a8\u7406\u8def\u5f84\uff1a\u80fd\u538b\u6210\u4e00\u56fe\u5c31\u538b\u6210\u4e00\u56fe<\/h2>\n\n\n\n
          \u8bad\u7ec3\u8def\u5f84\u9700\u8981\u751f\u6210 GenStats \u548c\u4e2d\u95f4\u7edf\u8ba1\u91cf\uff0c\u56e0\u6b64\u62c6\u6210\u4e86\u4e09\u5f20\u5b50\u56fe\u3002\u4f46\u63a8\u7406\u8def\u5f84\u4e0d\u9700\u8981\u8fd9\u4e9b\u2014\u2014\u5b83\u53ea\u9700\u8981 Conv \u2192 BN(INF) \u2192 ReLU \u7684\u7eaf\u524d\u5411\u8ba1\u7b97\uff0c\u65e2\u4e0d\u9700\u8981\u672c\u6279\u6b21\u7684\u7edf\u8ba1\u91cf\uff0c\u4e5f\u4e0d\u9700\u8981\u66f4\u65b0 running statistics\u3002<\/p>\n\n\n\n
          \u56e0\u6b64 CBR_AMP_INF<\/code> \u9996\u5148\u5c1d\u8bd5\u628a\u6574\u4e2a Conv \u2192 MUL(eq_scale) \u2192 ADD(eq_bias) \u2192 ReLU<\/code> \u8868\u8fbe\u6210\u4e00\u5f20 cuDNN Frontend Graph<\/strong>\uff1a<\/p>\n\n\n\n
          auto Y      = graph->conv_fprop(X, W, conv_opts);\nauto scaled = graph->pointwise(Y, eq_scale_ta, MUL);\nauto shifted= graph->pointwise(scaled, eq_bias_ta, ADD);\nauto relu   = graph->pointwise(shifted, RELU_FWD);<\/pre>\n\n\n\n
          \u5982\u679c cuDNN \u63a5\u53d7\u8fd9\u4e2a\u56fe\u5e76\u6210\u529f\u5b9e\u4f8b\u5316\uff0c\u5c31\u53ea\u9700\u4e00\u6b21 execute<\/code> \u8c03\u7528\uff1b\u5982\u679c\u56e0\u4e3a\u67d0\u4e9b\u5f62\u72b6\u3001\u7cbe\u5ea6\u6216\u5f15\u64ce\u9650\u5236\u5bfc\u81f4\u5355\u56fe\u6784\u5efa\u5931\u8d25\uff0c\u6846\u67b6\u4f1a\u9759\u9ed8\u56de\u9000\u5230\u4e09\u6bb5\u5f0f\u6267\u884c\uff1a\u5148 Conv INF\uff0c\u518d\u81ea\u5b9a\u4e49 BN INF kernel\uff0c\u6700\u540e ReLU INF kernel\u3002\u8fd9\u79cd”\u5148\u5c1d\u8bd5\u6700\u4f18\uff0c\u518d\u4fdd\u8bc1\u6b63\u786e”\u7684\u7b56\u7565\u5728\u4e0d\u727a\u7272\u5065\u58ee\u6027\u7684\u524d\u63d0\u4e0b\uff0c\u5c3d\u53ef\u80fd\u62ff\u5230\u878d\u5408\u6536\u76ca\u3002<\/p>\n\n\n\n
          \u4e03\u3001Mode C \u7ecf\u9a8c\u641c\u7d22\uff1a\u4e3a\u6bcf\u4e00\u5c42\u5339\u914d\u6700\u4f18\u5f15\u64ce<\/h2>\n\n\n\n
          cuDNN \u4e3a\u540c\u4e00\u4e2a\u5377\u79ef\u64cd\u4f5c\u63d0\u4f9b\u4e86\u591a\u79cd\u5f15\u64ce\u5b9e\u73b0\uff0c\u4e0d\u540c\u7684\u5f15\u64ce\u5728\u4e0d\u540c\u7684\u5f20\u91cf\u5f62\u72b6\u3001\u5185\u5b58\u5e03\u5c40\u548c\u786c\u4ef6\u67b6\u6784\u4e0b\uff0c\u6027\u80fd\u5dee\u5f02\u53ef\u80fd\u975e\u5e38\u663e\u8457\u3002\u5e38\u89c4\u505a\u6cd5\u662f\u4f7f\u7528 cuDNN \u7684\u542f\u53d1\u5f0f\u5f15\u64ce\u9009\u62e9\uff0c\u8ba9 cuDNN \u6839\u636e\u5c11\u6570\u51e0\u4e2a\u53c2\u6570\u505a\u4e00\u4e2a\u7c97\u7565\u4f30\u8ba1\uff0c\u9009\u4e00\u4e2a”\u5e94\u8be5\u8fd8\u884c”\u7684\u5f15\u64ce\u3002\u4f46\u542f\u53d1\u5f0f\u7b97\u6cd5\u65e0\u6cd5\u4fdd\u8bc1\u9009\u5230\u6700\u4f18\u3002<\/p>\n\n\n\n
          Tech-Renaissance \u7684 CBR \u878d\u5408\u7b97\u5b50\u5f15\u5165\u4e86 Mode C \u7ecf\u9a8c\u641c\u7d22<\/strong>\u673a\u5236\uff1a\u9488\u5bf9 A100 \u548c RTX 5090 \u7b49\u76ee\u6807 GPU\uff0c\u9884\u5148\u5bf9\u6240\u6709\u53ef\u80fd\u7684\u5f62\u72b6\u7ec4\u5408\u8fdb\u884c\u7a77\u4e3e\u5f0f\u57fa\u51c6\u6d4b\u8bd5\uff0c\u5c06\u6700\u4f18\u5f15\u64ce\u7684 tag \u8bb0\u5f55\u4e3a\u7ecf\u9a8c\u6570\u636e\uff0c\u4ee5 C++ constexpr \u6570\u7ec4\u7684\u5f62\u5f0f\u7f16\u8bd1\u8fdb\u6846\u67b6\u4e8c\u8fdb\u5236\u4e2d\u3002\u4ee5 A100 FP16 \u7ecf\u9a8c\u8868\u4e3a\u4f8b\uff0c\u5b83\u5305\u542b 92 \u6761\u8bb0\u5f55\uff0c\u8986\u76d6\u4e86 ResNet-50 \u548c VGG16BN \u7b49\u7ecf\u5178\u6a21\u578b\u7684\u5377\u79ef\u5c42\u5f62\u72b6\u7ec4\u5408\uff0c\u4ee5\u53ca conv_fprop<\/code>\u3001conv_genstats<\/code>\u3001conv_wgrad<\/code>\u3001conv_dgrad<\/code> \u7b49\u64cd\u4f5c\u7c7b\u578b\u3002<\/p>\n\n\n\n
          \u6bcf\u6761\u7ecf\u9a8c\u8bb0\u5f55\u5927\u81f4\u5305\u542b\uff1a<\/p>\n\n\n\n
            \n
          • shape_key<\/strong>\uff1a\u67e5\u8be2\u952e\uff0c\u7cbe\u786e\u5230 GPU \u578b\u53f7\u3001CUDA\/cuDNN \u7248\u672c\u3001\u7b97\u5b50\u7c7b\u578b\u3001\u5f20\u91cf\u5404\u7ef4\u5ea6\u3001stride\u3001padding\u3001\u5e03\u5c40\u7b49\uff1b<\/li>\n\n\n\n
          • winner_tag<\/strong>\uff1a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u6700\u4f18\u7684\u5f15\u64ce\u6807\u7b7e\uff1b<\/li>\n\n\n\n
          • backup1_tag \/ backup2_tag<\/strong>\uff1a\u4e24\u4e2a\u5907\u9009\u6807\u7b7e\uff0c\u7528\u4e8e\u6700\u4f18\u5f15\u64ce\u4e0d\u53ef\u7528\u65f6\u56de\u9000\uff1b<\/li>\n\n\n\n
          • workspace_bytes \/ benchmark_time_ms<\/strong>\uff1a\u5de5\u4f5c\u7a7a\u95f4\u9700\u6c42\u548c\u5b9e\u6d4b\u8017\u65f6\u3002<\/li>\n<\/ul>\n\n\n\n
            \u8fd0\u884c\u65f6\uff0c\u6846\u67b6\u6839\u636e\u5f53\u524d\u5f20\u91cf\u5f62\u72b6\u6784\u9020\u67e5\u8be2\u952e\uff0c\u5728 constexpr \u8868\u4e2d\u4e8c\u5206\u67e5\u627e\u5339\u914d\u7684\u8bb0\u5f55\uff0c\u7136\u540e\u6309\u4e09\u7ea7\u4f18\u5148\u7ea7\u4f9d\u6b21\u5c1d\u8bd5\u6784\u5efa\u5bf9\u5e94\u5f15\u64ce\uff1awinner \u2192 backup1 \u2192 backup2\u3002\u5982\u679c\u5168\u90e8\u5931\u8d25\uff0c\u5219\u4f18\u96c5\u56de\u9000\u5230\u6807\u51c6\u7684\u542f\u53d1\u5f0f\u9009\u62e9\u3002<\/p>\n\n\n\n
            \u8fd9\u79cd\u8bbe\u8ba1\u7684\u4f18\u52bf\u5728\u4e8e\u96f6\u8fd0\u884c\u65f6\u641c\u7d22\u5f00\u9500<\/strong>\uff1a\u7ecf\u9a8c\u8868\u662f\u7f16\u8bd1\u671f\u5e38\u91cf\uff0c\u4e8c\u5206\u67e5\u627e\u5728\u7eb3\u79d2\u7ea7\u5b8c\u6210\uff0c\u5339\u914d\u5230\u6807\u7b7e\u540e\u76f4\u63a5\u8c03\u7528 build_plan_at_index<\/code>\uff0c\u4e0d\u9700\u8981\u904d\u5386\u548c\u8bc4\u4f30\u6240\u6709\u5019\u9009\u5f15\u64ce\u3002\u8fd9\u5bf9 Conv+GenStats\u3001WGrad\u3001DGrad \u4e09\u4e2a\u5b50\u56fe\u90fd\u72ec\u7acb\u9002\u7528\u3002<\/p>\n\n\n\n
            \u7ecf\u9a8c\u6570\u636e\u901a\u8fc7\u72ec\u7acb\u7684 Python \u811a\u672c\u751f\u6210\uff1a\u811a\u672c\u904d\u5386\u76ee\u6807\u6a21\u578b\u7684\u6240\u6709\u5377\u79ef\u5c42\uff0c\u63d0\u53d6\u5f62\u72b6\u53c2\u6570\uff0c\u7528 Mode A \u679a\u4e3e\u6240\u6709\u53ef\u7528\u5f15\u64ce\uff0c\u9010\u4e2a warmup \u540e\u8ba1\u65f6\uff0c\u9009\u51fa\u6700\u5feb\u7684\u5f15\u64ce\u4f5c\u4e3a winner\uff0c\u518d\u9009\u4e24\u4e2a\u6b21\u4f18\u4f5c\u4e3a backup\uff0c\u6700\u7ec8\u5199\u5165 include\/generated\/cbr_experience_*.hpp<\/code>\u3002\u76ee\u524d\u6846\u67b6\u4e3a A100 \u548c RTX 5090 \u5206\u522b\u751f\u6210\u4e86 FP16 \u7ecf\u9a8c\u8868\u3002<\/p>\n\n\n\n
            \u516b\u3001\u7ea6\u675f\u3001\u9650\u5236\u4e0e\u6b63\u786e\u6027\u9a8c\u8bc1<\/h2>\n\n\n\n
            CBR \u878d\u5408\u7b97\u5b50\u76ee\u524d\u6709\u4e00\u4e9b\u660e\u786e\u7684\u7ea6\u675f\uff0c\u5b83\u4eec\u5927\u591a\u6765\u81ea\u5e95\u5c42\u5e93\u80fd\u529b\u8fb9\u754c\u6216\u5de5\u7a0b\u53d6\u820d\uff1a<\/p>\n\n\n\n
              \n
            1. \u4ec5 CUDA \u8def\u5f84<\/strong>\uff1aCPU \u8def\u5f84\u76f4\u63a5\u629b\u51fa NotImplementedError<\/code>\u3002\u8fd9\u4e0d\u662f\u8bbe\u8ba1\u7f3a\u9677\uff0c\u800c\u662f CBR \u7684\u6536\u76ca\u4e3b\u8981\u6765\u81ea GPU \u5f20\u91cf\u6838\u5fc3\u548c\u5168\u5c40\u663e\u5b58\u5e26\u5bbd\uff0cCPU \u7aef\u505a\u878d\u5408\u610f\u4e49\u6709\u9650\u3002<\/li>\n\n\n\n
            2. \u4ec5 AMP\/FP16 \u6a21\u5f0f<\/strong>\uff1acuDNN \u7684 Conv + GenStats<\/code> \u7ec4\u5408\u76ee\u524d\u5bf9 FP16 \u652f\u6301\u6700\u5b8c\u6574\uff1bFP32 \u6a21\u5f0f\u4e0b\u5373\u4f7f\u5f00\u542f\u878d\u5408\uff0c\u5e26\u5bbd\u8282\u7701\u5e26\u6765\u7684\u6536\u76ca\u4e5f\u4e0d\u8db3\u4ee5\u62b5\u6d88\u5de5\u7a0b\u590d\u6742\u5ea6\uff0c\u4e14\u6574\u4f53\u901f\u5ea6\u5f80\u5f80\u4ecd\u4e0d\u53ca AMP \u8def\u5f84\u3002\u56e0\u6b64 Tech-Renaissance \u7684 FP32 \u8def\u5f84\u4ecd\u4f7f\u7528\u72ec\u7acb\u7684 Conv\u3001BN2D\u3001ReLU \u7b97\u5b50\u3002<\/li>\n\n\n\n
            3. \u8f93\u51fa\u901a\u9053\u9700\u4e3a 8 \u7684\u500d\u6570<\/strong>\uff1a\u8fd9\u662f cuDNN TensorCore BN \u5185\u90e8\u7ea6\u675f\u3002\u5982\u679c\u7528\u6237\u6a21\u578b\u4e0d\u6ee1\u8db3\uff0c\u7f16\u8bd1\u5668\u4f1a\u62a5\u9519\u5e76\u63d0\u793a\u63d2\u5165 channel_padding<\/code> \u6216\u8c03\u6574\u901a\u9053\u6570\u3002<\/li>\n\n\n\n
            4. \u7edf\u4e00\u4f7f\u7528 NHWC \u7279\u5f81\u56fe\u5e03\u5c40\u548c KRSC \u5377\u79ef\u6838\u5e03\u5c40<\/strong>\uff1a\u8fd9\u4e0e PyTorch \u9ed8\u8ba4\u7684 NCHW \u4e0d\u540c\uff0c\u4f46\u548c cuDNN \u5185\u90e8\u7684\u504f\u597d\u4e00\u81f4\uff0c\u80fd\u8ba9\u5377\u79ef\u548c BN \u66f4\u5bb9\u6613\u88ab\u878d\u5408\u6267\u884c\u3002\u6846\u67b6\u5728\u7f16\u8bd1\u671f\u5c31\u5b8c\u6210\u5e03\u5c40\u63a8\u5bfc\uff0c\u8fd0\u884c\u65f6\u4e0d\u9700\u8981\u989d\u5916\u7684 transpose\u3002<\/li>\n<\/ol>\n\n\n\n
              \u4e3a\u4e86\u786e\u4fdd\u878d\u5408\u4e0d\u53d8\u6210”\u5077\u6362\u6570\u5b66”\uff0c\u6211\u4eec\u5728 tests\/op\/<\/code> \u4e2d\u4e3a CBR \u5199\u4e86\u56db\u7ec4\u7b49\u4ef7\u6027\u6d4b\u8bd5\uff1a<\/p>\n\n\n\n
                \n
              • test_cbr_amp_fwd.cpp<\/code><\/li>\n\n\n\n
              • test_cbr_amp_bwd.cpp<\/code><\/li>\n\n\n\n
              • test_cbr_amp_bwd_first_layer.cpp<\/code><\/li>\n\n\n\n
              • test_cbr_amp_inf.cpp<\/code><\/li>\n<\/ul>\n\n\n\n
                \u6bcf\u7ec4\u90fd\u628a CBR \u8def\u5f84\u548c\u72ec\u7acb\u7684 Conv+BN2D+ReLU<\/code> \u8def\u5f84\u505a\u6570\u503c\u5bf9\u6bd4\u3002\u7531\u4e8e CBR \u4f7f\u7528 cuDNN \u7684\u4f4d\u538b\u7f29 BOOLEAN<\/code> \u63a9\u7801\uff0c\u800c\u72ec\u7acb ReLU \u4f7f\u7528\u6bcf\u5143\u7d20 INT8 \u63a9\u7801\uff0c\u6d4b\u8bd5\u91cc\u8fd8\u4e13\u95e8\u505a\u4e86 mask \u8f6c\u7801\u5bf9\u9f50\u3002<\/p>\n\n\n\n
                \u4e5d\u3001CBR \u4e0e\u5206\u7acb\u7b97\u5b50\u7684\u6839\u672c\u5dee\u5f02<\/h2>\n\n\n\n
                \u4e0b\u8868\u628a CBR \u4e0e\u72ec\u7acb Conv+BN2D+ReLU \u5728\u5b9e\u73b0\u5c42\u9762\u505a\u4e00\u4e2a\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n
                \u7ef4\u5ea6<\/th>\u5206\u7acb Conv+BN2D+ReLU<\/th>CBR \u878d\u5408\u7b97\u5b50<\/th><\/tr><\/thead>
                \u8ba1\u7b97\u56fe\u8282\u70b9<\/td>3 \u4e2a\u8282\u70b9<\/td>1 \u4e2a CBR \u8282\u70b9<\/td><\/tr>
                \u9ed8\u8ba4\u8c03\u5ea6\u6d41<\/td>\u5404\u6309 op_stream_policy<\/code> \u5206\u914d<\/td>\u5165\u53e3\u5728 COMP_1<\/code>\uff0c\u5185\u90e8 fan out \u5230 COMP_1\/2\/3<\/code><\/td><\/tr>
                \u524d\u5411 kernel\/\u56fe\u63d0\u4ea4<\/td>Conv\u3001BN\u3001ReLU \u5404\u4e00\u6b21<\/td>3 \u5f20 cuDNN FE Graph\uff0c\u5185\u90e8\u5c3d\u53ef\u80fd\u878d\u5408<\/td><\/tr>
                BN \u7edf\u8ba1\u91cf<\/td>BN \u5355\u72ec\u8bfb\u53d6 conv_output<\/code> \u7b97 mean\/var<\/td>Conv \u540c\u65f6\u751f\u6210 sum<\/code>\/sq_sum<\/code><\/td><\/tr>
                \u4e2d\u95f4\u6fc0\u6d3b<\/td>conv_output<\/code>\u3001bn_output<\/code> \u90fd\u4f5c\u4e3a\u5b8c\u6574\u5f20\u91cf<\/td>bn_output<\/code> \u590d\u7528\u4e3a eq_scale<\/code>\/eq_bias<\/code> \u6216\u53cd\u5411 dL\/d(conv_out)<\/code><\/td><\/tr>
                ReLU \u63a9\u7801<\/td>\u6bcf\u5143\u7d20 INT8<\/td>cuDNN BOOLEAN<\/code> \u4f4d\u538b\u7f29<\/td><\/tr>
                \u901a\u9053\u7ea6\u675f<\/td>\u4e00\u822c\u65e0\u7279\u6b8a\u8981\u6c42<\/td>\u8f93\u51fa\u901a\u9053\u9700\u4e3a 8 \u7684\u500d\u6570<\/td><\/tr>
                \u7cbe\u5ea6\u652f\u6301<\/td>FP32 \/ AMP<\/td>\u4ec5 AMP<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                \u4ece\u663e\u5b58\u4e8b\u52a1\u89d2\u5ea6\u770b\uff0c\u524d\u5411\u4f20\u64ad\u4e2d\u5206\u7acb\u8def\u5f84\u7684 conv_output<\/code> \u8981\u7ecf\u5386 1 \u6b21\u5199\u5165\u548c 2 \u6b21\u8bfb\u53d6\uff08\u7edf\u8ba1\u91cf\u8ba1\u7b97 + BN Apply\uff09\uff0c\u800c CBR \u878d\u5408\u4e2d conv_output<\/code> \u53ea\u6709 1 \u6b21\u5199\u5165\u548c 1 \u6b21\u8bfb\u53d6\uff08BN Apply\uff09\uff0c\u4e14\u8fd9\u6b21\u8bfb\u53d6\u7684\u7ed3\u679c\u5728\u5bc4\u5b58\u5668\u4e2d\u76f4\u63a5\u6d41\u5411 ReLU\uff0c\u4e0d\u518d\u4ea7\u751f bn_output<\/code> \u7684\u4e2d\u95f4\u5199\u5165\u3002\u53cd\u5411\u4f20\u64ad\u4e2d\uff0c\u5206\u7acb\u8def\u5f84\u7684 ReLU_BWD \u548c BN_BWD \u5404\u9700\u8981\u4e00\u6b21 HBM \u5199\u5165\u548c\u8bfb\u53d6\u6765\u4f20\u9012\u4e2d\u95f4\u68af\u5ea6\uff1bCBR \u878d\u5408\u4e2d\u8fd9\u4e24\u4e2a\u64cd\u4f5c\u88ab\u5408\u5e76\uff0cdy_masked<\/code> \u76f4\u63a5\u901a\u8fc7\u5bc4\u5b58\u5668\u4f20\u9012\u3002<\/p>\n\n\n\n
                \u5341\u3001CBR \u5728\u6574\u4e2a\u7cfb\u7edf\u91cc\u7684\u4f4d\u7f6e<\/h2>\n\n\n\n
                \u5355\u72ec\u770b\u4e00\u4e2a CBR \u7b97\u5b50\uff0c\u5b83\u8282\u7701\u7684\u4e0d\u8fc7\u662f\u51e0\u6b21\u4e2d\u95f4\u7279\u5f81\u56fe\u7684\u8bfb\u5199\uff1b\u4f46\u628a\u5b83\u653e\u5728\u6574\u4e2a\u7f51\u7edc\u91cc\u770b\uff0c\u6536\u76ca\u4f1a\u88ab\u5c42\u6570\u653e\u5927\u3002\u5728 README \u62a5\u544a\u7684 VGG16BN A100\u00d78 \u8bad\u7ec3\u4e2d\uff0cTech-Renaissance \u8fbe\u5230\u4e86 9,310.13 images\/sec<\/strong>\uff0c\u6bd4 PyTorch torch.compile<\/code> \u7684 7,351.20 images\/sec \u9ad8\u51fa 26.65%<\/strong>\u3002\u8fd9\u4e2a\u5dee\u8ddd\u4e0d\u662f\u67d0\u4e00\u4e2a\u5355\u70b9\u9020\u6210\u7684\uff0c\u800c\u662f\u9759\u6001\u56fe\u7f16\u8bd1\u3001CUDA Graph \u5168\u6355\u83b7\u3001MemoryPlan\u3001\u591a\u6d41\u8c03\u5ea6\u548c CBR \u8fd9\u7c7b\u878d\u5408\u7b97\u5b50\u53e0\u52a0\u8d77\u6765\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n
                CBR \u5728\u8fd9\u4e2a\u7cfb\u7edf\u91cc\u7684\u89d2\u8272\uff0c\u662f\u628a”\u9ad8\u5c42\u7f51\u7edc\u4e2d\u968f\u5904\u53ef\u89c1\u7684 Conv+BN+ReLU”\u8fd9\u4e2a\u6a21\u5f0f\u538b\u69a8\u5230\u6781\u81f4\u3002\u5b83\u8ba9\u6846\u67b6\u5728\u4e0d\u505a\u4efb\u4f55\u7528\u6237\u4fb5\u5165\u5f0f\u6539\u9020\u7684\u524d\u63d0\u4e0b\uff0c\u5c31\u80fd\u66f4\u6709\u6548\u5730\u5229\u7528 GPU \u5e26\u5bbd\u3002\u5f53\u7136\uff0c\u6211\u4eec\u4e0d\u80fd\u628a 26.65% \u7684\u52a0\u901f\u5168\u90e8\u5f52\u529f\u4e8e CBR\u3002\u5b83\u662f\u4e00\u4e2a\u4e58\u6570\u56e0\u5b50\uff1aCUDA Graph \u6d88\u9664\u4e86 launch overhead\uff0cMemoryPlan \u4fdd\u8bc1\u4e86\u5730\u5740\u7a33\u5b9a\uff0c\u591a\u6d41\u8c03\u5ea6\u9690\u85cf\u4e86\u540c\u6b65\u7b49\u5f85\uff0cCBR \u5219\u628a\u6700\u70ed\u7684\u7b97\u5b50\u6a21\u5f0f\u538b\u69a8\u5230\u66f4\u63a5\u8fd1\u786c\u4ef6\u6781\u9650\u3002\u5b83\u4eec\u5f7c\u6b64\u4f9d\u8d56\uff0c\u7f3a\u5c11\u4efb\u4f55\u4e00\u4e2a\uff0c\u5176\u4ed6\u4f18\u5316\u90fd\u96be\u4ee5\u5145\u5206\u53d1\u6325\u3002<\/p>\n\n\n\n
                \u5341\u4e00\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n
                CBR \u878d\u5408\u7b97\u5b50\u662f Tech-Renaissance \u7b97\u5b50\u5c42\u4f18\u5316\u7684\u4e00\u4e2a\u7f29\u5f71\u3002\u5b83\u7684\u8bbe\u8ba1\u601d\u8def\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a<\/p>\n\n\n\n
                  \n
                1. \u56fe\u7ea7\u81ea\u52a8\u8bc6\u522b<\/strong>\uff1a\u5728 ArchPlan<\/code> \u91cc\u628a Conv+BN+ReLU<\/code> \u81ea\u52a8\u5408\u5e76\u6210 CBR<\/code>\uff1b<\/li>\n\n\n\n
                2. \u5e95\u5c42 cuDNN Graph<\/strong>\uff1a\u7528 Frontend Graph \u628a\u5377\u79ef\u3001\u7edf\u8ba1\u91cf\u751f\u6210\u3001BN \u5f52\u4e00\u5316\u3001ReLU \u5c3d\u53ef\u80fd\u4e32\u6210\u6700\u5c11\u7684\u6267\u884c\u5355\u5143\uff1b<\/li>\n\n\n\n
                3. \u591a\u6d41\u91cd\u53e0<\/strong>\uff1a\u628a\u4e0d\u540c\u9636\u6bb5\u62c6\u5230 COMP_1\/2\/3<\/code>\uff0c\u7528 event \u4fdd\u8bc1\u4f9d\u8d56\uff0c\u63d0\u9ad8\u5e76\u884c\u5ea6\uff1b<\/li>\n\n\n\n
                4. \u5185\u5b58\u590d\u7528<\/strong>\uff1abn_output<\/code> \u5728\u524d\u5411\u548c\u53cd\u5411\u626e\u6f14\u4e0d\u540c\u89d2\u8272\uff0c\u51cf\u5c11\u663e\u5b58\u5360\u7528\uff1b<\/li>\n\n\n\n
                5. \u8bad\u7ec3\u4e0e\u63a8\u7406\u517c\u987e<\/strong>\uff1a\u524d\u5411\u4fdd\u7559\u5b8c\u6574 BN \u7edf\u8ba1\u66f4\u65b0\uff0c\u63a8\u7406\u5219\u5c1d\u8bd5\u5355\u56fe\u878d\u5408\u5e76\u5e26 fallback\uff1b<\/li>\n\n\n\n
                6. \u5f15\u64ce\u7ecf\u9a8c\u641c\u7d22<\/strong>\uff1a\u7528\u9884\u7f16\u8bd1\u7684 Mode C \u7ecf\u9a8c\u8868\u8df3\u8fc7\u8fd0\u884c\u65f6\u542f\u53d1\u5f0f\u641c\u7d22\uff1b<\/li>\n\n\n\n
                7. \u4e25\u683c\u7b49\u4ef7\u6027\u6d4b\u8bd5<\/strong>\uff1a\u786e\u4fdd\u878d\u5408\u540e\u7684\u7ed3\u679c\u4e0e\u5206\u7acb\u7b97\u5b50\u4e00\u81f4\u3002<\/li>\n<\/ol>\n\n\n\n
                  \u5b83\u4e0d\u662f\u628a\u4e09\u6bb5\u4ee3\u7801\u7c97\u66b4\u5730\u7c98\u5728\u4e00\u8d77\uff0c\u800c\u662f\u5728\u4fdd\u8bc1\u6570\u5b66\u6b63\u786e\u7684\u524d\u63d0\u4e0b\uff0c\u628a”\u8be5\u7701\u7684\u5185\u5b58\u8bfb\u5199\u3001\u8be5\u5408\u5e76\u7684 kernel launch\u3001\u8be5\u5e76\u884c\u7684\u8ba1\u7b97\u9636\u6bb5”\u5168\u90e8\u505a\u5230\u6781\u81f4\u3002\u56fe\u7ea7\u81ea\u52a8\u5408\u5e76\u8ba9\u7528\u6237\u65e0\u611f\u77e5\uff0c\u5e95\u5c42 cuDNN Graph \u548c\u591a\u6d41\u8c03\u5ea6\u8ba9\u786c\u4ef6\u6ee1\u8d1f\u8377\uff0c\u9759\u6001\u5185\u5b58\u89c4\u5212\u8ba9\u590d\u7528\u5b89\u5168\u53ef\u63a7\uff0c\u7b49\u4ef7\u6027\u6d4b\u8bd5\u8ba9\u901f\u5ea6\u4e0d\u4ee5\u727a\u7272\u6b63\u786e\u6027\u4e3a\u4ee3\u4ef7\u2014\u2014\u8fd9\u6b63\u662f Tech-Renaissance \u505a\u7b97\u5b50\u878d\u5408\u7684\u57fa\u672c\u65b9\u6cd5\u8bba\u3002<\/p>\n\n\n\n
                  \u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u5c06\u7ee7\u7eed\u6cbf\u7740”\u878d\u5408”\u8fd9\u6761\u4e3b\u7ebf\uff0c\u8fdb\u5165\u4f18\u5316\u5668\u5c42\u9762\u2014\u2014\u770b\u770b SGD\u3001AdamW\u3001LARS \u7684\u6743\u91cd\u66f4\u65b0\u53c8\u662f\u5982\u4f55\u88ab\u6279\u91cf\u878d\u5408\u5230\u6574\u533a\u5185\u5b58\u4e0a\u7684\u3002<\/p>\n\n\n\n
                  <\/p>\n","protected":false},"excerpt":{"rendered":"
                  \u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341 \u5728\u51e0\u4e4e\u4efb\u4f55\u73b0\u4ee3\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u91cc […]<\/p>\n","protected":false},"author":1,"featured_media":700,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-container-style":"default","site-container-layout":"default","site-sidebar-layout":"default","disable-article-header":"default","disable-site-header":"default","disable-site-footer":"default","disable-content-area-spacing":"default","footnotes":""},"categories":[15],"tags":[],"class_list":["post-541","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-15"],"_links":{"self":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/541","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/comments?post=541"}],"version-history":[{"count":4,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/541\/revisions"}],"predecessor-version":[{"id":654,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/541\/revisions\/654"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/700"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=541"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=541"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=541"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}