{"id":463,"date":"2026-07-08T02:55:16","date_gmt":"2026-07-07T18:55:16","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=463"},"modified":"2026-07-08T16:19:48","modified_gmt":"2026-07-08T08:19:48","slug":"%e5%bc%a0%e9%87%8f%e3%80%81nhwc%e5%b8%83%e5%b1%80%e4%b8%8e256%e5%ad%97%e8%8a%82%e5%af%b9%e9%bd%90%ef%bc%9a%e6%a1%86%e6%9e%b6%e9%87%8c%e7%9a%84%e7%a0%96%e5%9d%97","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/463\/","title":{"rendered":"(7) \u5f20\u91cf\u3001NHWC\u5e03\u5c40\u4e0e256\u5b57\u8282\u5bf9\u9f50\uff1a\u6846\u67b6\u91cc\u7684&#8221;\u7816\u5757&#8221;"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\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\u4e03<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0a\u4e00\u7bc7\u6211\u4eec\u628a Tech-Renaissance \u7684\u4e03\u5927\u6a21\u5757\u4e32\u6210\u4e86\u4e00\u5f20\u5168\u5c40\u5730\u56fe\uff0c\u4ece DTS \u6587\u4ef6\u8bfb\u53d6\u3001CPU \u9884\u5904\u7406\u3001H2D \u5f02\u6b65\u4f20\u8f93\uff0c\u5230\u524d\u5411\u3001\u53cd\u5411\u3001\u68af\u5ea6\u901a\u4fe1\u3001\u4f18\u5316\u5668\u66f4\u65b0\uff0c\u518d\u5230\u6307\u6807\u8f93\u51fa\uff0c\u6574\u6761\u6d41\u6c34\u7ebf\u5df2\u7ecf\u5728\u4f60\u8111\u5b50\u91cc\u6709\u4e86\u4e00\u6839\u4e3b\u7ebf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u5730\u56fe\u518d\u6f02\u4eae\uff0c\u4e5f\u5f97\u4e00\u5757\u7816\u4e00\u5757\u7816\u5730\u780c\u3002\u4ece\u4eca\u5929\u5f00\u59cb\uff0c\u6211\u4eec\u628a\u8fd9\u4e9b&#8221;\u7816\u5757&#8221;\u9010\u4e00\u62ff\u8d77\u6765\u7ec6\u770b\u3002\u800c\u6240\u6709\u7816\u5757\u91cc\u6700\u57fa\u7840\u3001\u6700\u4e0d\u8d77\u773c\u3001\u5374\u53c8\u65e0\u5904\u4e0d\u5728\u7684\uff0c\u5c31\u662f<strong>\u5f20\u91cf\uff08Tensor\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f20\u91cf\u8fd9\u4e2a\u4e1c\u897f\uff0c\u521d\u5b66\u8005\u7b2c\u4e00\u6b21\u89c1\u5f80\u5f80\u89c9\u5f97\u5b83\u4e0d\u8fc7\u662f\u4e2a&#8221;\u591a\u7ef4\u6570\u7ec4&#8221;\uff1a\u6807\u91cf\u662f\u96f6\u7ef4\uff0c\u5411\u91cf\u662f\u4e00\u7ef4\uff0c\u77e9\u9635\u662f\u4e8c\u7ef4\uff0cbatch \u56fe\u50cf\u662f\u56db\u7ef4\u3002\u4f46\u5982\u679c\u4f60\u771f\u53bb\u505a\u4e00\u4e2a\u9ad8\u6027\u80fd\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u4f60\u4f1a\u53d1\u73b0\u4e8b\u60c5\u8fdc\u6ca1\u6709\u8fd9\u4e48\u7b80\u5355\u3002\u5f20\u91cf\u600e\u4e48\u5728\u5185\u5b58\u91cc\u6392\u5e03\u3001\u901a\u9053\u653e\u5728\u54ea\u4e2a\u7ef4\u5ea6\u3001\u9996\u5730\u5740\u8981\u4e0d\u8981\u5bf9\u9f50\u3001stride \u600e\u4e48\u8ba1\u7b97\u2014\u2014\u8fd9\u4e9b\u770b\u8d77\u6765\u7ec6\u679d\u672b\u8282\u7684\u95ee\u9898\uff0c\u6700\u7ec8\u4f1a\u76f4\u63a5\u51b3\u5b9a\u4f60\u5377\u79ef\u80fd\u8dd1\u591a\u5feb\u3001\u663e\u5b58\u80fd\u7701\u591a\u5c11\u3001\u4ee5\u53ca CUDA Graph \u80fd\u4e0d\u80fd\u7a33\u5b9a\u6355\u83b7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7bc7\u5c31\u6765\u8bb2\u8bb2 Tech-Renaissance \u91cc\u5f20\u91cf\u7684\u8bbe\u8ba1\u54f2\u5b66\uff1a<strong>\u4e3a\u4ec0\u4e48\u7edf\u4e00 NHWC\u3001\u4e3a\u4ec0\u4e48\u9996\u5730\u5740 256 \u5b57\u8282\u5bf9\u9f50\u3001\u4e3a\u4ec0\u4e48 DTensor \u662f\u4e00\u4e2a\u4e0d\u6301\u6709\u5185\u5b58\u7684\u7eaf\u63cf\u8ff0\u7b26\uff0c\u4ee5\u53ca CPU \u548c CUDA \u8def\u5f84\u4e3a\u4ec0\u4e48\u4f1a\u6709\u4e24\u5957 stride\u3002<\/strong> \u8fd9\u4e9b\u90fd\u662f\u540e\u9762 MemoryPlan\u3001CUDA Graph\u3001\u591a\u6d41\u5e76\u53d1\u7b49\u7bc7\u7ae0\u7684\u516c\u5171\u5e95\u5ea7\uff0c\u5fc5\u987b\u5148\u628a\u5b83\u592f\u5b9e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u5f20\u91cf\u4e0d\u53ea\u662f\u4e00\u5757\u5185\u5b58<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u4ece\u6807\u91cf\u5230\u5f20\u91cf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6df1\u5ea6\u5b66\u4e60\u7684\u8bed\u5883\u91cc\uff0c<strong>\u5f20\u91cf\uff08Tensor\uff09<\/strong>\u662f\u6570\u636e\u7684\u901a\u7528\u5bb9\u5668\u3002\u4e00\u4e2a\u6807\u91cf\u53ef\u4ee5\u770b\u4f5c\u96f6\u9636\u5f20\u91cf\uff0c\u4e00\u4e2a\u5411\u91cf\u662f\u4e00\u9636\u5f20\u91cf\uff0c\u4e00\u4e2a\u77e9\u9635\u662f\u4e8c\u9636\u5f20\u91cf\u3002\u800c\u6df1\u5ea6\u5b66\u4e60\u91cc\u6700\u5e38\u89c1\u7684\u6279\u91cf\u56fe\u50cf\u6570\u636e\u2014\u2014\u6bd4\u5982\u4e00\u4e2a batch \u7684 224\u00d7224 RGB \u56fe\u7247\u2014\u2014\u5c31\u662f\u4e00\u4e2a\u56db\u9636\u5f20\u91cf\uff0c\u5f62\u72b6\u901a\u5e38\u5199\u4f5c <code>[N, H, W, C]<\/code>\uff1aN \u662f batch \u5927\u5c0f\uff0cH \u548c W \u662f\u9ad8\u548c\u5bbd\uff0cC \u662f\u901a\u9053\u6570\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46&#8221;\u5f62\u72b6&#8221;\u53ea\u662f\u5f20\u91cf\u7684\u4e00\u4e2a\u5c5e\u6027\u3002\u4e00\u4e2a\u8bad\u7ec3\u6846\u67b6\u91cc\u7684\u5f20\u91cf\uff0c\u81f3\u5c11\u8981\u643a\u5e26\u4ee5\u4e0b\u4fe1\u606f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5f62\u72b6\uff08Shape\uff09<\/strong>\uff1a\u6bcf\u4e2a\u7ef4\u5ea6\u7684\u5927\u5c0f\uff1b<\/li>\n\n\n\n<li><strong>\u6570\u636e\u7c7b\u578b\uff08DType\uff09<\/strong>\uff1aFP32\u3001FP16\u3001INT8\u3001INT32 \u7b49\uff1b<\/li>\n\n\n\n<li><strong>\u5185\u5b58\u5e03\u5c40\uff08Layout\uff09<\/strong>\uff1aNCHW \u8fd8\u662f NHWC\uff0c\u901a\u9053\u5728\u6700\u5185\u5c42\u8fd8\u662f\u5916\u5c42\uff1b<\/li>\n\n\n\n<li><strong>Stride<\/strong>\uff1a\u6cbf\u6bcf\u4e2a\u7ef4\u5ea6\u8d70\u4e00\u683c\u9700\u8981\u8df3\u8fc7\u591a\u5c11\u5143\u7d20\uff1b<\/li>\n\n\n\n<li><strong>\u6240\u5728\u8bbe\u5907<\/strong>\uff1aCPU \u5185\u5b58\u8fd8\u662f GPU \u663e\u5b58\uff1b<\/li>\n\n\n\n<li><strong>\u662f\u5426\u9700\u8981\u68af\u5ea6<\/strong>\uff1a\u662f\u5426\u53c2\u4e0e\u53cd\u5411\u4f20\u64ad\uff1b<\/li>\n\n\n\n<li><strong>\u6307\u5411\u7684\u8ba1\u7b97\u56fe\u8282\u70b9<\/strong>\uff1a\u52a8\u6001\u56fe\u91cc\u8fd8\u8981\u8bb0\u5f55\u5b83\u662f\u54ea\u4e2a\u7b97\u5b50\u4ea7\u751f\u7684\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u4fe1\u606f\u91cc\uff0cShape\u3001DType \u548c Layout \u662f\u6700\u57fa\u7840\u7684\uff0c\u5b83\u4eec\u5171\u540c\u51b3\u5b9a\u4e86&#8221;\u8fd9\u5757\u5185\u5b58\u91cc\u5230\u5e95\u88c5\u7740\u4ec0\u4e48&#8221;\u3002\u800c Stride \u5219\u544a\u8bc9\u4f60&#8221;\u600e\u4e48\u6309\u903b\u8f91\u987a\u5e8f\u8bbf\u95ee\u5b83&#8221;\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 NCHW \u4e0e NHWC\uff1a\u4e24\u79cd\u4e16\u754c\u89c2<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u540c\u6837\u4e00\u5f20 <code>[N, H, W, C]<\/code> \u7684\u56fe\u50cf batch\uff0c\u5728\u5185\u5b58\u91cc\u53ef\u4ee5\u6709\u4e24\u79cd\u5b8c\u5168\u4e0d\u540c\u7684\u6392\u6cd5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>NCHW<\/strong> \u628a\u901a\u9053\u653e\u5728\u9ad8\u5bbd\u4e4b\u524d\u3002\u903b\u8f91\u4e0a\u4f60\u5148\u904d\u5386 batch N\uff0c\u518d\u904d\u5386\u901a\u9053 C\uff0c\u7136\u540e\u624d\u662f\u9ad8 H\u3001\u5bbd W\u3002\u5185\u5b58\u91cc\u7684\u987a\u5e8f\u5927\u81f4\u662f\uff1a\u5148\u628a\u7b2c\u4e00\u5f20\u56fe\u7684\u6240\u6709 R \u901a\u9053\u50cf\u7d20\uff08\u6574\u5f20 H\u00d7W \u5e73\u9762\uff09\u653e\u5b8c\uff0c\u518d\u653e\u6240\u6709 G \u901a\u9053\u50cf\u7d20\uff0c\u7136\u540e B \u901a\u9053\uff0c\u63a5\u7740\u7b2c\u4e8c\u5f20\u56fe\u2026\u2026 \u56e0\u6b64\uff0c<strong>\u540c\u4e00\u901a\u9053\u7684\u6570\u636e\u5728\u5185\u5b58\u91cc\u662f\u8fde\u7eed\u5b58\u653e\u7684<\/strong>\uff0c\u4f46\u540c\u4e00\u4e2a\u50cf\u7d20\u4f4d\u7f6e\u7684\u4e0d\u540c\u901a\u9053\u503c\u4e4b\u95f4\u9694\u4e86\u4e00\u6574\u5f20 H\u00d7W \u5e73\u9762\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>NHWC<\/strong> \u5219\u628a\u901a\u9053\u653e\u5728\u6700\u5185\u5c42\u3002\u5185\u5b58\u987a\u5e8f\u662f\uff1a\u5148\u653e\u7b2c\u4e00\u4e2a\u7a7a\u95f4\u4f4d\u7f6e\uff08H=0, W=0\uff09\u7684\u6240\u6709\u901a\u9053\u503c\uff08R\u3001G\u3001B\uff09\uff0c\u518d\u653e\u4e0b\u4e00\u4e2a\u7a7a\u95f4\u4f4d\u7f6e\u7684\u6240\u6709\u901a\u9053\u503c\u2026\u2026 \u56e0\u6b64\uff0c<strong>\u540c\u4e00\u4e2a\u50cf\u7d20\u4f4d\u7f6e\u7684\u4e0d\u540c\u901a\u9053\u503c\u5728\u5185\u5b58\u91cc\u662f\u6328\u7740\u7684<\/strong>\uff0c\u4f46\u76f8\u90bb\u7a7a\u95f4\u4f4d\u7f6e\u7684\u540c\u4e00\u901a\u9053\u6570\u636e\u5728\u5185\u5b58\u91cc\u5e76\u4e0d\u8fde\u7eed\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e24\u79cd\u5e03\u5c40\u6ca1\u6709\u7edd\u5bf9\u7684\u6570\u5b66\u4f18\u52a3\uff0c\u4f46<strong>\u8bbf\u5b58\u7279\u6027\u5dee\u5f02\u5de8\u5927<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">NCHW \u9002\u5408\u505a&#8221;\u901a\u9053\u7ea7&#8221;\u64cd\u4f5c\uff1a\u6bd4\u5982\u4f60\u8981\u5bf9\u6574\u5f20\u7279\u5f81\u56fe\u505a\u5168\u5c40\u6c60\u5316\uff0c\u6216\u8005\u6309\u901a\u9053\u505a\u5f52\u4e00\u5316\uff0cNCHW \u4e0b\u540c\u4e00\u901a\u9053\u7684\u6570\u636e\u662f\u8fde\u7eed\u7684\uff0c\u8bbf\u95ee\u5c40\u90e8\u6027\u597d\u3002NHWC \u5219\u66f4\u9002\u5408&#8221;\u7a7a\u95f4\u7ea7&#8221;\u64cd\u4f5c\uff1a\u6bd4\u5982\u5377\u79ef\u6838\u5728 H-W \u5e73\u9762\u4e0a\u6ed1\u52a8\u65f6\uff0c\u6bcf\u6ed1\u5230\u4e00\u4e2a\u4f4d\u7f6e\u90fd\u8981\u540c\u65f6\u8bfb\u591a\u4e2a\u901a\u9053\uff0cNHWC \u4e0b\u8fd9\u4e9b\u901a\u9053\u503c\u5728\u5185\u5b58\u91cc\u7d27\u6328\u7740\uff0c\u4e00\u6b21\u8bfb\u53d6\u5c31\u80fd\u62ff\u5230\u4e00\u4e2a\u50cf\u7d20\u4f4d\u7f6e\u7684\u6240\u6709\u901a\u9053\u3002\u800c\u5728 NCHW \u4e0b\uff0c\u540c\u4e00\u4e2a\u50cf\u7d20\u7684 C \u4e2a\u901a\u9053\u8de8\u8d8a\u4e86\u6574\u4e2a H\u00d7W \u5e73\u9762\uff0c\u9700\u8981\u591a\u6b21\u5206\u6563\u8bbf\u95ee\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3b\u6d41\u6846\u67b6\u7684\u9009\u62e9\u4e5f\u5f88\u6709\u8da3\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>PyTorch<\/strong> \u9ed8\u8ba4\u4f7f\u7528 NCHW\u3002\u8fd9\u4e3b\u8981\u4e0e cuDNN \u65e9\u671f\u5bf9 NCHW \u8def\u5f84\u652f\u6301\u6700\u6210\u719f\u7684\u751f\u6001\u6709\u5173\u3002cuDNN \u6700\u521d\u7684\u8bb8\u591a\u7b97\u5b50\u5bf9 NCHW \u652f\u6301\u6700\u6210\u719f\uff0cPyTorch \u6cbf\u7528\u4e86\u8fd9\u4e00\u4f20\u7edf\u3002<\/li>\n\n\n\n<li><strong>TensorFlow<\/strong> \u65e9\u671f\u9ed8\u8ba4\u4f7f\u7528 NHWC\uff0c\u539f\u56e0\u4e4b\u4e00\u662f TensorFlow \u6700\u521d\u5728 CPU \u4e0a\u5f00\u53d1\uff0cNHWC \u5bf9 CPU cache \u66f4\u53cb\u597d\u3002\u4f46 TensorFlow \u4e5f\u652f\u6301 <code>data_format='NCHW'<\/code>\uff0c\u5728 GPU \u4e0a\u5e38\u5e38\u66f4\u5feb\u3002<\/li>\n\n\n\n<li><strong>ONNX Runtime\u3001TensorRT<\/strong> \u7b49\u63a8\u7406\u6846\u67b6\u5219\u4f1a\u6839\u636e\u540e\u7aef\u548c\u786c\u4ef6\u81ea\u52a8\u505a\u5e03\u5c40\u8f6c\u6362\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u957f\u671f\u4ee5\u6765\uff0c\u793e\u533a\u91cc\u6709\u4e00\u79cd\u8bf4\u6cd5\uff1a&#8221;GPU \u7528 NCHW\uff0cCPU \u7528 NHWC\u3002&#8221;\u8fd9\u53e5\u8bdd\u5728\u51e0\u5e74\u524d\u5927\u4f53\u6210\u7acb\uff0c\u4f46\u8fd1\u4e9b\u5e74\u60c5\u51b5\u5df2\u7ecf\u53d1\u751f\u53d8\u5316\u2014\u2014cuDNN \u5bf9 NHWC \u7684\u652f\u6301\u8d8a\u6765\u8d8a\u597d\uff0cTensor Core \u5728\u5f88\u591a\u573a\u666f\u4e0b\u53cd\u800c\u66f4\u504f\u597d NHWC\u3002NVIDIA \u5b98\u65b9\u6587\u6863\u548c\u4e0d\u5c11\u8bc4\u6d4b\u90fd\u8868\u660e\uff0c\u5bf9\u4e8e ResNet\u3001VGG \u8fd9\u7c7b\u5178\u578b CNN\uff0cNHWC \u914d\u5408 FP16 Tensor Core \u5e38\u5e38\u80fd\u83b7\u5f97\u6bd4 NCHW \u66f4\u9ad8\u7684\u541e\u5410\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.3 \u5185\u5b58\u5bf9\u9f50\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u662f&#8221;\u968f\u4fbf\u653e&#8221;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u9664\u4e86\u5e03\u5c40\uff0c\u53e6\u4e00\u4e2a\u5bb9\u6613\u88ab\u5ffd\u89c6\u7684\u95ee\u9898\u662f<strong>\u5bf9\u9f50\uff08Alignment\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GPU \u8bbf\u95ee\u5168\u5c40\u5185\u5b58\u65f6\uff0c\u662f\u4ee5 warp\uff0832 \u4e2a\u7ebf\u7a0b\uff09\u4e3a\u5355\u4f4d\u53d1\u8d77\u8bf7\u6c42\u7684\u3002\u5982\u679c warp \u91cc\u7684 32 \u4e2a\u7ebf\u7a0b\u8bbf\u95ee\u7684\u5730\u5740\u662f\u8fde\u7eed\u4e14\u5bf9\u9f50\u7684\uff0cGPU \u786c\u4ef6\u53ef\u4ee5\u628a\u8fd9\u4e9b\u8bf7\u6c42\u5408\u5e76\u6210\u4e00\u6b21\u6216\u51e0\u6b21\u5927\u7684\u5185\u5b58\u4e8b\u52a1\uff08Coalesced Access\uff09\uff0c\u6700\u5927\u9650\u5ea6\u5730\u5229\u7528\u663e\u5b58\u5e26\u5bbd\u3002\u53cd\u4e4b\uff0c\u5982\u679c\u5730\u5740\u5206\u6563\u6216\u4e0d\u5bf9\u9f50\uff0c\u540c\u4e00\u4e2a warp \u7684\u8bbf\u95ee\u53ef\u80fd\u88ab\u62c6\u6210\u51e0\u5341\u6b21\u5c0f\u4e8b\u52a1\uff0c\u5e26\u5bbd\u5229\u7528\u7387\u5927\u5e45\u4e0b\u964d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u4ee3 NVIDIA GPU \u7684\u5168\u5c40\u5185\u5b58\u8bbf\u95ee\u4ee5 32 \u5b57\u8282 sector \u4e3a\u57fa\u7840\uff0c\u82e5\u5e72 sector \u518d\u7ec4\u5408\u6210\u66f4\u5927\u7684\u5185\u5b58\u4e8b\u52a1\u3002\u5982\u679c warp \u8bbf\u95ee\u7684\u5730\u5740\u8fde\u7eed\u4e14\u5bf9\u9f50\uff0c\u786c\u4ef6\u5c31\u80fd\u628a\u8bf7\u6c42\u5408\u5e76\u6210\u5c11\u6570\u51e0\u6b21\u5927\u4e8b\u52a1\uff1b\u53cd\u4e4b\u5219\u4f1a\u88ab\u62c6\u6210\u591a\u6b21\u5c0f\u4e8b\u52a1\uff0c\u5e26\u5bbd\u5229\u7528\u7387\u5927\u5e45\u4e0b\u964d\u3002CUDA \u7684 <code>cudaMalloc<\/code> \u8fd4\u56de\u7684\u5730\u5740\u9ed8\u8ba4\u81f3\u5c11 256 \u5b57\u8282\u5bf9\u9f50\uff0c\u5f88\u591a\u9ad8\u6027\u80fd\u8ba1\u7b97\u5e93\uff08cuDNN\u3001cuBLAS\uff09\u5185\u90e8\u4e5f\u4f1a\u8981\u6c42\u8f93\u5165\u6307\u9488\u6ee1\u8db3\u4e00\u5b9a\u7684\u5bf9\u9f50\u6761\u4ef6\uff0c\u5426\u5219\u53ef\u80fd\u76f4\u63a5\u62d2\u7edd\u6216\u8005\u56de\u9000\u5230\u6162\u8def\u5f84\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5&#8221;\u9996\u5730\u5740\u5bf9\u9f50&#8221;\u4e0d\u662f\u5f3a\u8feb\u75c7\uff0c\u800c\u662f\u786c\u4ef6\u5c42\u9762\u7684\u786c\u6027\u8981\u6c42\u3002256 \u5b57\u8282\u5bf9\u9f50\u6bd4 128 \u5b57\u8282\u66f4\u4e25\u683c\uff0c\u53ef\u4ee5\u786e\u4fdd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5bf9\u9f50\u8fb9\u754c\u8db3\u591f\u5bbd\uff0c\u4e0d\u5bb9\u6613\u51fa\u73b0\u8de8\u8fb9\u754c\u7684\u5206\u88c2\u8bbf\u95ee\uff1b<\/li>\n\n\n\n<li>cuDNN \u7684\u67d0\u4e9b kernel \u5728\u62ff\u5230 256 \u5bf9\u9f50\u5730\u5740\u65f6\u53ef\u4ee5\u76f4\u63a5\u8d70\u6700\u4f18\u5b9e\u73b0\uff1b<\/li>\n\n\n\n<li>\u591a\u6d41\u3001\u591a\u5361\u573a\u666f\u4e0b\uff0c\u5730\u5740\u89c4\u5f8b\u4e00\u81f4\uff0c\u4fbf\u4e8e\u505a\u6279\u91cf\u4f20\u8f93\u548c\u901a\u4fe1\u4f18\u5316\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Tech-Renaissance \u7684\u5f20\u91cf\u5206\u5c42<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u91cc\uff0c\u5f20\u91cf\u4e0d\u662f\u4e00\u4e2a\u7c7b\u6253\u5929\u4e0b\uff0c\u800c\u662f\u88ab\u660e\u786e\u5730\u5206\u6210\u4e86\u4e09\u4e2a\u5c42\u6b21\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong><code>Shape<\/code> \/ <code>DType<\/code><\/strong>\uff1a\u6700\u57fa\u7840\u7684\u7c7b\u578b\u7cfb\u7edf\uff0c\u5b9a\u4e49\u5728 <code>include\/renaissance\/core\/types.h<\/code>\uff1b<\/li>\n\n\n\n<li><strong><code>Tensor<\/code><\/strong>\uff1aCPU \u7aef\u6570\u636e\u5bb9\u5668\uff0c\u53ea\u7528\u4e8e\u4e3b\u673a-\u8bbe\u5907\u6570\u636e\u642c\u8fd0\uff1b<\/li>\n\n\n\n<li><strong><code>DTensor<\/code>\uff08DistributedTensor\uff09<\/strong>\uff1a\u8bbe\u5907\u7aef\u8ba1\u7b97\u7684\u4e00\u7b49\u516c\u6c11\uff0c\u662f\u4e00\u4e2a\u4e0d\u6301\u6709\u5185\u5b58\u7684\u7eaf\u63cf\u8ff0\u7b26\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u5206\u5c42\u770b\u8d77\u6765\u591a\u6b64\u4e00\u4e3e\uff0c\u5b9e\u5219\u662f\u7ecf\u8fc7\u6df1\u601d\u719f\u8651\u7684\u3002<strong>Tensor \u8d1f\u8d23&#8221;\u6570\u636e\u5728 CPU \u4e0a\u957f\u4ec0\u4e48\u6837&#8221;\uff0cDTensor \u8d1f\u8d23&#8221;\u6570\u636e\u5728 GPU \u4e0a\u600e\u4e48\u88ab\u5f15\u7528&#8221;\uff0cMemoryPlan \u8d1f\u8d23&#8221;\u6570\u636e\u5728\u663e\u5b58\u91cc\u653e\u5728\u54ea\u91cc&#8221;\uff0cDeviceContext \u8d1f\u8d23&#8221;\u628a\u63cf\u8ff0\u7b26\u7ffb\u8bd1\u6210\u771f\u5b9e\u6307\u9488&#8221;\u3002<\/strong> \u56db\u4e2a\u89d2\u8272\u5404\u53f8\u5176\u804c\uff0c\u624d\u80fd\u628a\u9759\u6001\u56fe\u3001CUDA Graph\u3001\u591a\u5361\u5e76\u884c\u4e32\u8d77\u6765\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 Shape \u4e0e DType\uff1a\u88ab\u56fa\u5b9a\u4e3a 4D NHWC<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5148\u770b\u6700\u57fa\u7840\u7684 <code>Shape<\/code>\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/core\/types.h\nstruct Shape {\n    int dims[4] = {1, 1, 1, 1};  \/\/ NHWC\u683c\u5f0f\uff1adims[0]=N, dims[1]=H, dims[2]=W, dims[3]=C\n    ...\n    int n() const { return dims[0]; }\n    int h() const { return dims[1]; }\n    int w() const { return dims[2]; }\n    int c() const { return dims[3]; }\n    int64_t numel() const noexcept { ... }\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u8fd9\u4e2a\u8bbe\u8ba1\uff1a<code>Shape<\/code> \u6c38\u8fdc\u662f 4 \u7ef4\uff0c\u8bed\u4e49\u56fa\u5b9a\u4e3a <code>[N, H, W, C]<\/code>\u3002\u5168\u8fde\u63a5\u5c42\u7684\u6743\u91cd\u5728\u8fd9\u4e2a\u6846\u67b6\u91cc\u4e5f\u4f1a\u88ab\u8868\u8fbe\u6210 <code>[O, 1, 1, I]<\/code> \u7684 4D NHWC \u5f62\u72b6\uff0c\u504f\u7f6e\u662f <code>[O, 1, 1, 1]<\/code>\u3002\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f<strong>\u6240\u6709\u7b97\u5b50\u90fd\u6309\u540c\u4e00\u5957\u7ef4\u5ea6\u7ea6\u5b9a\u5904\u7406<\/strong>\uff0c\u4e0d\u9700\u8981\u4e3a\u5168\u8fde\u63a5\u3001\u5377\u79ef\u3001\u6c60\u5316\u5206\u522b\u5199\u5f62\u72b6\u63a8\u5bfc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Shape<\/code> \u7684\u6784\u9020\u51fd\u6570\u91cc\u8fd8\u6709\u4e00\u4e2a\u503c\u5f97\u4e00\u63d0\u7684\u7ec6\u8282\uff1a\u7531\u4e8e MSVC \u5728 <code>\/O2<\/code> \u4f18\u5316\u7ea7\u522b\u4e0b\u66fe\u5bf9\u7b80\u5355 POD \u7ed3\u6784\u4f53\u7684\u521d\u59cb\u5316\u5217\u8868\u4ea7\u751f\u9519\u8bef\u4f18\u5316\uff0c\u6846\u67b6\u91cc\u4f7f\u7528 <code>TR_NOINLINE<\/code> \u914d\u5408 <code>volatile<\/code> \u5199\u5165\u6765\u5f3a\u5236\u751f\u6210\u5b9e\u9645\u5185\u5b58\u5199\u5165\u6307\u4ee4\uff0c\u89c4\u907f\u8fd9\u4e2a\u7f16\u8bd1\u5668\u5c42\u9762\u7684\u5751\u3002\u8fd9\u4e2a\u7ec6\u8282\u8bf4\u660e\uff0c\u505a\u5e95\u5c42\u57fa\u7840\u8bbe\u65bd\u8f6f\u4ef6\uff0c\u4f60\u9762\u5bf9\u7684\u4e0d\u53ea\u662f\u7b97\u6cd5\u548c\u67b6\u6784\uff0c\u8fd8\u6709\u7f16\u8bd1\u5668\u7684&#8221;\u813e\u6c14&#8221;\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>DType<\/code> \u5219\u5f88\u7b80\u5355\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/core\/types.h\nenum class DType : uint8_t { FP32, FP16, INT8, INT32 };<\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 Tensor\uff1aCPU \u7aef\u7684&#8221;\u4e8c\u7b49\u516c\u6c11&#8221;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Tensor<\/code> \u7c7b\u7684\u57fa\u672c\u60c5\u51b5\u662f\u8fd9\u6837\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/tensor.h\n\/**\n * @class Tensor\n * @brief CPU \u7aef\u6570\u636e\u5bb9\u5668\uff08\u79fb\u52a8\u8bed\u4e49\uff0c\u7981\u7528\u62f7\u8d1d\uff09\n *\n * \u7528\u4e8e\u4ece\u4e3b\u673a\u5411\u8bbe\u5907\uff08H2D\uff09\u6216\u8bbe\u5907\u5411\u4e3b\u673a\uff08D2H\uff09\u642c\u8fd0\u6570\u636e\u3002\n * ...\n * V4.20\u91cd\u8981\u53d8\u66f4\uff1a\n * - Tensor\u7c7b\u5f3a\u5236\u7d27\u51d1\u5e03\u5c40\uff08\u65e0padding\uff09\n * - \u6240\u6709\u6570\u636e\u6309NHWC\u987a\u5e8f\u8fde\u7eed\u5b58\u50a8\n * - \u79fb\u9664\u6240\u6709\u884c\u6b65\u5e45\u5bf9\u9f50\u8981\u6c42\n * - \u5bf9\u9f50\u8981\u6c42\u4ec5\u4fdd\u7559\uff1a\u9996\u5730\u5740256\u5b57\u8282\u5bf9\u9f50\n * - \u4e0eDTensor\u7684non-compact\u5e03\u5c40\u8f6c\u6362\u5728transfer\u65f6\u81ea\u52a8\u5904\u7406\n *\/<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684\u5173\u952e\u8bcd\u662f\uff1a<strong>\u5f3a\u5236\u7d27\u51d1\u3001\u65e0 padding\u3001NHWC\u3001\u9996\u5730\u5740 256 \u5b57\u8282\u5bf9\u9f50<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48 Tensor \u8981\u5f3a\u5236\u7d27\u51d1\uff1f\u56e0\u4e3a\u5b83\u53ea\u8d1f\u8d23 CPU \u7aef\u7684\u6570\u636e\u642c\u8fd0\u3002\u65e0\u8bba\u662f\u4ece\u78c1\u76d8\u8bfb\u5230\u5185\u5b58\u3001\u4ece\u5185\u5b58\u62f7\u5230\u663e\u5b58\uff0c\u8fd8\u662f\u4ece\u663e\u5b58\u62f7\u56de\u5185\u5b58\uff0cCPU \u4fa7\u7684\u6570\u636e\u90fd\u5e0c\u671b\u662f&#8221;\u6240\u89c1\u5373\u6240\u5f97&#8221;\uff1aN\u00d7H\u00d7W\u00d7C \u4e2a\u5143\u7d20\uff0c\u6bcf\u4e2a\u5143\u7d20\u5360\u5bf9\u5e94 dtype \u7684\u5b57\u8282\u6570\uff0c\u603b\u5927\u5c0f\u5c31\u662f <code>numel() * sizeof(dtype)<\/code>\uff0c\u4e0d\u591a\u4e0d\u5c11\u3002\u5982\u679c CPU \u4fa7\u4e5f\u505a padding\uff0c\u90a3\u6587\u4ef6\u8bfb\u53d6\u3001CRC \u6821\u9a8c\u3001\u9884\u5904\u7406\u7ba1\u7ebf\u90fd\u4f1a\u53d8\u590d\u6742\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46 Tensor \u7684\u9996\u5730\u5740\u4ecd\u7136\u8981\u6c42 256 \u5b57\u8282\u5bf9\u9f50\u3002\u5b9e\u73b0\u4e0a\uff0cGPU \u6a21\u5f0f\u7528 <code>cudaMallocHost<\/code> \u5206\u914d\u9875\u9501\u5b9a\u5185\u5b58\uff08CUDA \u4fdd\u8bc1\u8fd4\u56de\u5730\u5740\u81f3\u5c11 256 \u5bf9\u9f50\uff09\uff0cCPU \u6a21\u5f0f\u7528 <code>mi_malloc_aligned(..., 256)<\/code>\uff0c\u5e76\u505a\u9632\u5fa1\u6027\u6821\u9a8c\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/tensor\/tensor.cpp\nTensor::Tensor(const Shape&amp; shape, DType dtype) : shape_(shape), dtype_(dtype) {\n    ...\n    nbytes_ = static_cast&lt;size_t>(num) * elem_size;\n    elem_size_ = elem_size;\n\n    constexpr size_t base_alignment = 256;\n\n#ifdef TR_USE_CUDA\n    cudaError_t err = cudaMallocHost(&amp;ptr_, nbytes_);\n#else\n    ptr_ = mi_malloc_aligned(nbytes_, base_alignment);\n#endif\n\n    if (reinterpret_cast&lt;uintptr_t>(ptr_) % base_alignment != 0) { ... }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">Tensor \u8fd8\u7981\u7528\u62f7\u8d1d\u3001\u53ea\u652f\u6301\u79fb\u52a8\u8bed\u4e49\u3002\u539f\u56e0\u5f88\u76f4\u63a5\uff1a\u5728\u8bad\u7ec3\u70ed\u8def\u5f84\u4e0a\u968f\u610f\u590d\u5236\u4e00\u5757 CPU \u5185\u5b58\u662f\u4e0d\u53ef\u63a5\u53d7\u7684\uff0c\u6846\u67b6\u9f13\u52b1\u4f60\u7528 <code>std::move<\/code> \u8f6c\u79fb\u6240\u6709\u6743\uff0c\u5b9e\u73b0\u96f6\u5f00\u9500\u4f20\u9012\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 DTensor\uff1a\u8bbe\u5907\u7aef\u7684\u7eaf\u63cf\u8ff0\u7b26<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4 Tensor \u662f&#8221;\u62ff\u7740\u4e00\u5757\u5185\u5b58\u7684\u5bb9\u5668&#8221;\uff0c\u90a3 <code>DTensor<\/code> \u5c31\u662f&#8221;\u63cf\u8ff0\u4e00\u5757\u5185\u5b58\u7684\u56fe\u7eb8&#8221;\u3002<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/distributed_tensor.h\n\/**\n * @brief \u5206\u5e03\u5f0f\u5f20\u91cf \u2014 \"\u4e00\u5f20\u56fe\u7eb8\uff0c\u516b\u5361\u5171\u4eab\"\u7684\u5185\u5b58\u62bd\u8c61\n *\n * DTensor \u662f\u4e00\u4e2a\u7eaf\u865a\u62df\u6982\u5ff5\uff1a\u53ea\u5b58\u5f62\u72b6\/\u504f\u79fb\u91cf\/stride\uff0c\u4e0d\u6301\u6709\u5185\u5b58\u3001\u4e0d\u5b58\u6307\u9488\u3002\n * \u540c\u4e00\u4e2aDTensor\u6307\u4ee3\u6240\u6709\u5361\u4e0a\u76f8\u540c\u7684\u7269\u7406\u5185\u5b58\u533a\u57df\uff0c\u5e03\u5c40\u5b8c\u5168\u76f8\u540c\uff0c\u4f46\u6570\u636e\u53ef\u4ee5\u6709\u522b\u3002\n *\n * \u804c\u8d23\u5206\u79bb\uff1a\n *   DTensor       \u2014 \u63cf\u8ff0\u5355\u5f20\u91cf\u5728\u591a\u5361\u4e0a\u7684\u7edf\u4e00\u5185\u5b58\u89c6\u56fe\uff08\u672c\u6587\u4ef6\uff09\n *   MemoryPlan    \u2014 \u5206\u914d DTensor\u3001\u8ba1\u7b97\u504f\u79fb\u3001\u786e\u4fdd\u65e0\u95f4\u9699\u548c 256B \u5bf9\u9f50\n *   DeviceContext \u2014 \u5c06 DTensor.offset \u89e3\u6790\u4e3a\u771f\u5b9e GPU\/CPU \u6307\u9488\n *\n * \u5173\u952e\u7ea6\u5b9a\uff1a\n *   - \u5168\u90e8 4D NHWC \u7269\u7406\u5e03\u5c40\uff0cShape \u5b58\u903b\u8f91\u7ef4\u5ea6\uff08padding\u524d\uff09\n *   - \u9996\u5730\u5740\u5fc5\u5b9a 256 \u5b57\u8282\u5bf9\u9f50\uff0cslot_bytes() \u5fc5\u5b9a\u4e3a 256 \u7684\u6574\u6570\u500d\n *   - \u521b\u5efa\u540e id\/shape\/dtype\/region \u4e0d\u53ef\u53d8\n *   - offset_ \u7531 MemoryPlan::finalize \u8d4b\u503c\uff0c\u6b64\u524d\u6052\u4e3a -1\uff08\u54e8\u5175\uff09\n *   - stride \u7531 shape + dtype \u8ba1\u7b97\uff0c\u521b\u5efa\u540e\u4e0d\u53ef\u53d8\n *\/<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u91cc\u4e0d\u5b58\u6307\u9488\uff0c\u53ea\u5b58\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>id<\/code>\uff1a\u5168\u5c40\u552f\u4e00\u7f16\u53f7\uff1b<\/li>\n\n\n\n<li><code>shape \/ dtype \/ region<\/code>\uff1a\u5f62\u72b6\u3001\u7c7b\u578b\u3001\u6240\u5c5e\u663e\u5b58\u8bed\u4e49\u533a\uff1b<\/li>\n\n\n\n<li><code>offset_<\/code>\uff1a\u5728 MemoryPlan \u91cc\u7684\u5b57\u8282\u504f\u79fb\uff1b<\/li>\n\n\n\n<li><code>slot_bytes_<\/code>\uff1a\u5728\u663e\u5b58\u91cc\u5360\u7528\u7684\u69fd\u4f4d\u5927\u5c0f\uff1b<\/li>\n\n\n\n<li>\u4e24\u5957 stride\uff1aCUDA \u5bf9\u9f50 stride \u548c CPU \u7d27\u51d1 stride\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd0\u884c\u65f6\u9700\u8981\u771f\u5b9e\u6307\u9488\u600e\u4e48\u529e\uff1f\u901a\u8fc7 <code>DeviceContext::ptr_at(dtensor_id)<\/code> \u52a8\u6001\u89e3\u6790\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/backend\/device_context.cpp\nvoid* DeviceContext::ptr_at(int dtensor_id) const noexcept {\n    const DTensor&amp; dt = current_mp_->get_dtensor(dtensor_id);\n    return ArenaKeeper::instance().ptr_at(rank_for_context_, static_cast&lt;size_t>(dt.offset()));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u5a01\u529b\u5728\u4e8e\uff1a<strong>\u540c\u4e00\u5f20 MemoryPlan \u56fe\u7eb8\u53ef\u4ee5\u5728\u591a\u5f20 GPU \u4e0a\u5171\u4eab<\/strong>\u3002\u516b\u5f20\u5361\u4e0a\u540c\u4e00\u4e2a DTensor id \u62e5\u6709\u76f8\u540c\u7684 offset \u548c slot_bytes\uff0c\u53ea\u662f\u5206\u522b\u7ed1\u5b9a\u5230\u5404\u81ea Arena \u7684\u57fa\u5730\u5740\u3002\u591a\u5361\u5e76\u884c\u65f6\u4e0d\u9700\u8981\u4e3a\u6bcf\u5f20\u5361\u7ef4\u62a4\u4e00\u4efd\u72ec\u7acb\u7684\u5f20\u91cf\u63cf\u8ff0\uff0cMemoryPlan \u76f4\u63a5\u590d\u7528\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u4e3a\u4ec0\u4e48 Tech-Renaissance \u7edf\u4e00\u91c7\u7528 NHWC<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 GPU \u8bad\u7ec3\uff0c\u5e03\u5c40\u8981\u4e3a GPU \u670d\u52a1<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u5b9a\u4f4d\u662f\u9ad8\u541e\u5410\u8bad\u7ec3\u6846\u67b6\uff0c\u4e3b\u6218\u573a\u5728 NVIDIA GPU\u3002\u5728\u8fd9\u4e2a\u524d\u63d0\u4e0b\uff0c\u7edf\u4e00 NHWC \u662f\u81ea\u7136\u800c\u7136\u7684\u51b3\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u539f\u56e0\u5e76\u4e0d\u7384\u4e4e\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\uff0c<strong>cuDNN \u5bf9 NHWC \u7684\u4f18\u5316\u5df2\u7ecf\u975e\u5e38\u6210\u719f<\/strong>\u3002cuDNN Frontend \u7684 Graph API \u5141\u8bb8\u8c03\u7528\u65b9\u901a\u8fc7 stride \u663e\u5f0f\u8868\u8fbe NHWC\uff0cNVIDIA \u5728 Ampere\u3001Hopper \u53ca\u540e\u7eed\u67b6\u6784\u4e0a\u4e3a NHWC \u505a\u4e86\u5927\u91cf\u9488\u5bf9\u6027\u7684\u878d\u5408 kernel\u3002\u5bf9\u4e8e Conv + BN + ReLU \u8fd9\u79cd\u7ecf\u5178\u7ec4\u5408\uff0cNHWC \u914d\u5408 FP16 Tensor Core \u5f80\u5f80\u80fd\u8fbe\u5230\u6700\u4f18\u541e\u5410\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\uff0c<strong>NHWC \u4e0e\u6570\u636e\u7ba1\u7ebf\u5929\u7136\u5339\u914d<\/strong>\u3002\u56fe\u50cf\u9884\u5904\u7406\u672c\u8d28\u4e0a\u662f\u5728 H-W \u5e73\u9762\u4e0a\u5bf9 RGB \u50cf\u7d20\u505a\u64cd\u4f5c\uff0c\u9884\u5904\u7406\u5b8c\u540e\u76f4\u63a5\u4ea7\u51fa NHWC \u683c\u5f0f\u6700\u81ea\u7136\uff0c\u4e0d\u9700\u8981\u989d\u5916\u505a\u4e00\u6b21 NHWC \u2192 NCHW \u7684\u8f6c\u7f6e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e09\uff0c<strong>\u907f\u514d\u5e03\u5c40\u8f6c\u6362\u7684\u5f00\u9500<\/strong>\u3002PyTorch \u91cc\u5982\u679c\u4f60\u7528 <code>channels_last<\/code>\uff0c\u6846\u67b6\u4f1a\u5728 NCHW \u548c NHWC \u4e4b\u95f4\u505a\u8f6c\u6362\uff1b\u8f6c\u6362\u672c\u8eab\u662f\u4e00\u6b21\u663e\u5b58\u8bfb\u5199\uff0c\u5bf9\u4e8e\u9ad8\u5206\u8fa8\u7387\u7279\u5f81\u56fe\u5e76\u4e0d\u4fbf\u5b9c\u3002Tech-Renaissance \u4ece\u7b2c\u4e00\u5c42\u8f93\u5165\u5f00\u59cb\u5c31\u7edf\u4e00 NHWC\uff0c\u6240\u6709\u5377\u79ef\u3001\u6c60\u5316\u3001BN\u3001\u6fc0\u6d3b\u90fd\u5728\u540c\u4e00\u79cd\u5e03\u5c40\u4e0b\u5b8c\u6210\uff0c\u6d88\u9664\u4e86\u8fd9\u79cd\u6765\u56de\u5012\u817e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u7136\uff0c\u7edf\u4e00 NHWC \u4e0d\u662f\u6ca1\u6709\u4ee3\u4ef7\u3002CPU \u7aef\u67d0\u4e9b\u901a\u9053\u7ea7\u64cd\u4f5c\u7684\u5c40\u90e8\u6027\u4f1a\u53d8\u5dee\uff0c\u6240\u4ee5\u6211\u4eec\u5728 CPU \u7b97\u5b50\u91cc\u7528 SIMD \u548c\u7d27\u51d1 stride \u505a\u8865\u507f\u3002\u4f46\u8fd9\u662f\u4e00\u4e2a\u6e05\u9192\u7684 trade-off\uff1aGPU \u8bad\u7ec3\u541e\u5410\u624d\u662f\u6838\u5fc3\u6307\u6807\uff0cCPU \u53ea\u662f\u8f85\u52a9\u642c\u8fd0\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 cuDNN \u6865\u63a5\uff1a\u7528 NCHW \u987a\u5e8f\u4f20 dim\uff0c\u7528 stride \u8868\u8fbe NHWC<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u4e2a\u6709\u8da3\u7684\u7ec6\u8282\u662f\uff0cTech-Renaissance \u8c03\u7528 cuDNN \u65f6\uff0c<strong>\u7ef4\u5ea6\u6570\u7ec4\u4ecd\u7136\u6309 NCHW \u987a\u5e8f\u4f20\u5165<\/strong>\uff0c\u4f46 stride \u7ed9\u51fa <code>c_stride = 1<\/code>\uff0c\u8fd9\u6837 cuDNN \u5c31\u77e5\u9053\u7269\u7406\u5e03\u5c40\u5176\u5b9e\u662f NHWC\u3002\u8fd9\u79cd&#8221;\u8868\u91cc\u4e0d\u4e00&#8221;\u662f cuDNN API \u5386\u53f2\u5305\u88b1\u9020\u6210\u7684\u3002<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/distributed_tensor.h\n\/**\n * @brief \u4ee5 cuDNN \u7684 NCHW API \u987a\u5e8f\u8fd4\u56de\u903b\u8f91\u7ef4\u5ea6\n *\n * dim_index: 0\u2192N, 1\u2192C, 2\u2192H, 3\u2192W\n *\n * \u5173\u952e\uff1acuDNN API \u671f\u671b NCHW \u987a\u5e8f\uff0c\u4f46\u7269\u7406\u5e03\u5c40\u662f NHWC\u3002\n * \u6b64\u65b9\u6cd5\u586b\u6b63\u786e\u503c\u5373\u53ef\u2014\u2014cuDNN \u901a\u8fc7 cudnn_stride(1)==1 \u81ea\u52a8\u8bc6\u522b NHWC\u3002\n *\/\nint64_t cudnn_dim(int dim_index) const noexcept { ... }\n\nint64_t cudnn_stride(int dim_index) const noexcept {\n    switch (dim_index) {\n        case 0: return n_stride_cuda_;\n        case 1: return c_stride_cuda_;  \/\/ = 1\n        case 2: return h_stride_cuda_;\n        case 3: return w_stride_cuda_;\n        default: return 0;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>cudnn_stride(1) == 1<\/code> \u8fd9\u4e2a\u4fe1\u53f7\u4e00\u7ed9\uff0ccuDNN \u5c31\u660e\u767d&#8221;C \u7ef4\u5ea6\u662f\u6700\u5185\u5c42&#8221;\uff0c\u4e8e\u662f\u6309 NHWC \u7269\u7406\u5e03\u5c40\u8c03\u5ea6\u5bf9\u5e94\u7684 kernel\u3002\u8fd9\u662f\u4e00\u79cd\u4f18\u96c5\u7684&#8221;\u9690\u5f0f\u534f\u5546&#8221;\u2014\u2014\u6846\u67b6\u4e0d\u9700\u8981\u663e\u5f0f\u5730\u544a\u8bc9 cuDNN&#8221;\u6211\u7528\u7684\u662f NHWC&#8221;\uff0ccuDNN \u81ea\u5df1\u4ece stride \u4fe1\u606f\u4e2d\u63a8\u65ad\u51fa\u6765\uff0c\u5e76\u9009\u62e9\u6700\u4f18\u7684\u6267\u884c\u8ba1\u5212\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001256 \u5b57\u8282\u5bf9\u9f50\uff1a\u4ece\u5de5\u5177\u51fd\u6570\u5230\u5168\u5c40\u5951\u7ea6<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u4e00\u4e2a\u4e0d\u8d77\u773c\u7684\u5de5\u5177\u51fd\u6570<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a\u6846\u67b6\u7684 256 \u5b57\u8282\u5bf9\u9f50\u90fd\u57fa\u4e8e\u8fd9\u4e2a\u51fd\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/core\/types.h\nconstexpr inline size_t align_up_256(size_t size) noexcept {\n    return (size + 255) &amp; ~static_cast&lt;size_t>(255);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u672c\u8eab\u53ea\u662f\u4e00\u884c\u4f4d\u8fd0\u7b97\uff0c\u4f46\u5b83\u80cc\u540e\u662f\u4e00\u5957\u8d2f\u7a7f\u5168\u5c40\u7684\u5951\u7ea6\uff1aTensor \u5206\u914d\u8981 256 \u5bf9\u9f50\u3001DTensor \u7684 slot \u8981 256 \u5bf9\u9f50\u3001MemoryPlan \u7684 offset \u8981 256 \u5bf9\u9f50\u3001workspace \u8981 256 \u5bf9\u9f50\u3001TSR \u6587\u4ef6 RAW \u6a21\u5f0f\u7684\u6570\u636e\u5757\u4e5f\u8981 256 \u5bf9\u9f50\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 DTensor \u7684 slot_bytes\uff1a\u5bf9\u9f50 + \u9884\u7559<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a DTensor \u5728\u663e\u5b58\u91cc\u5360\u7528\u7684\u7269\u7406\u69fd\u4f4d\u4e0d\u662f\u7b80\u5355\u7684 <code>numel * sizeof(dtype)<\/code>\uff0c\u800c\u662f\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/distributed_tensor.h\nstatic uint64_t compute_slot_bytes(const Shape&amp; shape, DType dtype, Region region) noexcept {\n    ...\n    int64_t pc = align_up(static_cast&lt;int64_t>(shape.c()),\n                          static_cast&lt;int64_t>(alignment));\n    uint64_t elems = static_cast&lt;uint64_t>(shape.n()) * shape.h() * shape.w() * pc;\n\n    if (dtype == DType::FP16) {\n        return utils::align_up_256(elems * 2 + 16);  \/\/ FP16: 2 bytes per element\n    } else if (dtype == DType::INT8) {\n        return utils::align_up_256(elems * 1 + 16);  \/\/ INT8: 1 byte per element\n    } else {\n        return 2 * utils::align_up_256(elems * 2 + 16);  \/\/ FP32 \u57fa\u51c6 = 2\u00d7FP16\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u51e0\u4e2a\u7ec6\u8282\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5148\u6309 region \u51b3\u5b9a\u7684 <code>alignment<\/code> \u5bf9 C \u901a\u9053\u505a padding\uff1b<\/li>\n\n\n\n<li>\u8ba1\u7b97 padded \u5143\u7d20\u6570\u540e\uff0c\u518d\u6309 dtype \u6362\u7b97\u5b57\u8282\uff1b<\/li>\n\n\n\n<li>\u6700\u540e <code>+16<\/code> \u9884\u7559\uff0c\u518d <code>align_up_256<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u90a3\u4e2a <code>+16<\/code> \u4e0d\u662f\u62cd\u8111\u888b\u3002\u5b83\u4e3a\u67d0\u4e9b\u7b97\u5b50\u5728\u8fb9\u754c\u5904\u7406\u3001\u5411\u91cf\u5316\u8bfb\u5199\u6216 CPU \u4fa7 XNNPACK \u7b49\u8def\u5f84\u4e0a\u63d0\u4f9b\u4e86\u5b89\u5168\u8fb9\u9645\u3002\u800c 256 \u5bf9\u9f50\u5219\u4fdd\u8bc1\u4e86 MemoryPlan \u7ebf\u6027\u7d2f\u52a0\u65f6\uff0c\u4e0b\u4e00\u4e2a DTensor \u7684 offset \u4ecd\u7136\u843d\u5728 256 \u8fb9\u754c\u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u5730\u65b9\u662f <strong>FP32 \u69fd\u4f4d = 2 \u00d7 FP16 \u7b49\u6548\u69fd\u4f4d<\/strong>\u3002\u8fd9\u4e2a\u500d\u7387\u5173\u7cfb\u4e0d\u662f\u5de7\u5408\uff1a\u5b83\u8ba9\u540c\u4e00\u5f20\u91cf\u5728\u4e0d\u540c\u7cbe\u5ea6\u4e0b\u7684\u663e\u5b58\u4f4d\u7f6e\u4fdd\u6301\u5b57\u8282\u7ea7\u5bf9\u5e94\uff0c\u540e\u7eed\u5c31\u53ef\u4ee5\u7528\u5355\u4e2a range op \u5b8c\u6210\u6574\u533a FP16\u2194FP32 \u7684\u8f6c\u6362\uff0c\u800c\u4e0d\u5fc5\u9010\u5c42\u3001\u9010\u5f20\u91cf\u5730\u5199\u5faa\u73af\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 MemoryPlan \u5982\u4f55\u4fdd\u8bc1\u6bcf\u4e2a offset \u90fd 256 \u5bf9\u9f50<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>MemoryPlan::finalize()<\/code> \u505a\u4e00\u904d\u7ebf\u6027\u7d2f\u52a0\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/graph\/memory_plan.cpp\nvoid MemoryPlan::finalize() {\n    uint64_t cursor = 0;\n\n    for (size_t ri = 0; ri &lt; static_cast&lt;size_t>(Region::NUM_REGIONS); ++ri) {\n        auto&amp; info = region_infos_[ri];\n        info.base_offset = cursor;\n\n        for (int32_t dt_id : region_dt_ids_[ri]) {\n            auto&amp; entry = entries_[id_to_idx_.at(dt_id)];\n            entry.dt.offset_ = cursor;\n            cursor += entry.dt.slot_bytes();\n        }\n\n        info.total_bytes = cursor - info.base_offset;\n    }\n\n    total_bytes_ = cursor;\n\n    validate_region_order();\n    validate_contiguity();\n    validate_layer_correspondence();\n    validate_alignment();\n    finalized_ = true;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u6bcf\u4e2a <code>slot_bytes()<\/code> \u90fd\u662f 256 \u7684\u500d\u6570\uff0c\u6240\u4ee5 cursor \u6c38\u8fdc\u505c\u7559\u5728 256 \u8fb9\u754c\u4e0a\uff0c\u6bcf\u4e2a DTensor \u7684 <code>offset_<\/code> \u81ea\u7136\u6ee1\u8db3 256 \u5bf9\u9f50\u3002\u6700\u540e <code>validate_alignment()<\/code> \u518d\u505a\u4e00\u6b21\u9632\u5fa1\u6027\u6821\u9a8c\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/graph\/memory_plan.cpp\nvoid MemoryPlan::validate_alignment() const {\n    for (const auto&amp; entry : entries_) {\n        TR_CHECK(entry.dt.offset() % 256 == 0, ValueError,\n             \"Alignment violated: DTensor \" &lt;&lt; entry.dt.id &lt;&lt; \" offset=\" &lt;&lt; entry.dt.offset());\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd&#8221;\u6784\u9020\u65f6\u4fdd\u8bc1 + \u7ed3\u675f\u65f6\u6821\u9a8c&#8221;\u7684\u53cc\u91cd\u673a\u5236\uff0c\u8ba9 CUDA Graph \u5168\u6355\u83b7\u6240\u9700\u7684\u5730\u5740\u7a33\u5b9a\u6027\u6709\u4e86\u4e00\u4e2a\u575a\u5b9e\u57fa\u7840\u3002\u540e\u9762\u8bb2 CUDA Graph \u65f6\u4f1a\u770b\u5230\uff0c\u5730\u5740\u7a33\u5b9a\u662f\u591a\u4e48\u91cd\u8981\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u53cc\u8f68 stride\uff1aCUDA \u5bf9\u9f50\u4e0e CPU \u7d27\u51d1<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f Tech-Renaissance \u5f20\u91cf\u7cfb\u7edf\u91cc\u6700\u6709\u4e2a\u6027\u7684\u8bbe\u8ba1\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u524d\u9762\u8bf4\u8fc7\uff0cCPU \u7aef\u7684 Tensor \u5f3a\u5236\u7d27\u51d1\u65e0 padding\uff1b\u4f46 GPU \u7aef\u4e3a\u4e86\u9002\u5e94 cuDNN \u548c Tensor Core \u7684\u5411\u91cf\u5316\u8981\u6c42\uff0c\u67d0\u4e9b Region \u7684 FP16 \u5f20\u91cf\u9700\u8981\u5728 C \u901a\u9053\u505a padding\u3002\u6bd4\u5982\u8f93\u5165\u7f13\u51b2\u533a\u6309 4 \u5bf9\u9f50\uff0c\u7279\u5f81\u56fe\u533a\u6309 8 \u5bf9\u9f50\u3002\u8fd9\u5c31\u5bfc\u81f4\u540c\u4e00\u4e2a\u903b\u8f91\u5f62\u72b6\u5728 CPU \u548c CUDA \u4e24\u6761\u8def\u5f84\u4e0a\u7684 stride \u4e0d\u4e00\u6837\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f DTensor \u76f4\u63a5\u63d0\u4f9b\u4e86\u4e24\u5957 stride\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/distributed_tensor.h\n\/**\n * @brief CUDA \u5bf9\u9f50 stride \u2014 \u7531 padded_c()\uff08cuda_alignment\uff09\u63a8\u5bfc\n *\/\nint64_t n_stride_cuda() const noexcept { return n_stride_cuda_; }\nint64_t h_stride_cuda() const noexcept { return h_stride_cuda_; }\nint64_t w_stride_cuda() const noexcept { return w_stride_cuda_; }\nint64_t c_stride_cuda() const noexcept { return c_stride_cuda_; }\n\n\/**\n * @brief CPU \u7d27\u51d1 stride \u2014 \u6846\u67b6\u4fdd\u8bc1 CPU \u4e0a\u6240\u6709 DTensor \u5fc5\u5b9a\u7d27\u51d1\n *\/\nint64_t n_stride_cpu()  const noexcept { return n_stride_cpu_; }\nint64_t h_stride_cpu()  const noexcept { return h_stride_cpu_; }\nint64_t w_stride_cpu()  const noexcept { return w_stride_cpu_; }\nint64_t c_stride_cpu()  const noexcept { return c_stride_cpu_; }<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CUDA \u8def\u5f84\u4e0a\uff0cC \u901a\u9053\u53ef\u80fd padding\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/distributed_tensor.h\nuint8_t cuda_alignment() const noexcept {\n    if (dtype == DType::FP16) {\n        switch (region) {\n            case Region::I_A_DATA:\n            case Region::I_B_DATA:  return 4;\n            case Region::F_FEATURE_FP16:\n            case Region::F_GRAD_SLOT_FP16:\n#ifdef TR_USE_CUDA\n                return 8;\n#else\n                return 1;\n#endif\n            default: return 1;\n        }\n    }\n    ...\n    return 1;\n}\n\nint64_t padded_c() const noexcept {\n    return align_up(static_cast&lt;int64_t>(shape.c()),\n                    static_cast&lt;int64_t>(cuda_alignment()));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6784\u9020\u65f6\u4e24\u5957 stride \u5206\u522b\u8ba1\u7b97\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/tensor\/distributed_tensor.h\nDistributedTensor(int32_t i, Shape s, DType d, Region r)\n    : id(i), shape(s), dtype(d), region(r), ... {\n    slot_bytes_ = compute_slot_bytes(s, d, r);\n    int64_t ac = padded_c();\n    c_stride_cuda_ = 1;\n    w_stride_cuda_ = ac;\n    h_stride_cuda_ = ac * s.w();\n    n_stride_cuda_ = ac * s.w() * s.h();\n\n    c_stride_cpu_ = 1;\n    w_stride_cpu_ = s.c();\n    h_stride_cpu_ = s.w() * s.c();\n    n_stride_cpu_ = s.h() * s.w() * s.c();\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CPU \u8def\u5f84\u59cb\u7ec8\u7d27\u51d1\uff0c<code>w_stride_cpu = C<\/code>\uff1bCUDA \u8def\u5f84\u5728\u9700\u8981\u65f6 <code>w_stride_cuda = padded_c<\/code>\u3002\u4e24\u8005\u8c01\u8be5\u7528\u54ea\u5957\uff0c\u7531 capture \u548c\u7b97\u5b50\u5b9e\u73b0\u81ea\u884c\u9009\u62e9\u3002\u4f8b\u5982 CPU capture \u7528\u7d27\u51d1 stride\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/graph\/capture_cpu.cpp\nop_ctx->n_stride = dt_in.n_stride_cpu();\nop_ctx->h_stride = dt_in.h_stride_cpu();\nop_ctx->w_stride = dt_in.w_stride_cpu();\nop_ctx->c_stride = dt_in.c_stride_cpu();<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CUDA \u5377\u79ef\u7b97\u5b50\u7528\u5bf9\u9f50 stride\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/backend\/ops\/dtensor\/conv_op_impl.cpp\ninline std::vector&lt;int64_t> make_nhwc_stride(const DTensor&amp; dt) {\n    return {dt.n_stride_cuda(), dt.c_stride_cuda(),\n            dt.h_stride_cuda(), dt.w_stride_cuda()};\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c DTensor \u662f\u975e\u7d27\u51d1\u7684\uff0cCPU \u4e0e GPU \u4e4b\u95f4\u642c\u8fd0\u65f6\u5c31\u4f1a\u89e6\u53d1\u4e00\u6b21 layout conversion\uff0c\u628a padding \u7684\u901a\u9053\u8865\u4e0a\u6216\u53bb\u6389\u3002\u8fd9\u4e2a\u8f6c\u6362\u7531 <code>transfer_to_rank<\/code> \/ <code>fetch_from_rank<\/code> \u81ea\u52a8\u5904\u7406\uff0c\u7528\u6237\u65e0\u611f\u77e5\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u4e0e PyTorch \u7684\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bf4\u5230\u8fd9\u91cc\uff0c\u53ef\u80fd\u6709\u4eba\u4f1a\u95ee\uff1aPyTorch \u4e0d\u662f\u4e5f\u6709 <code>channels_last<\/code> \u5417\uff1f\u5b83\u4e0d\u4e5f\u652f\u6301 NHWC \u5417\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u662f\u7684\uff0c\u4f46\u8def\u5f84\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u9ed8\u8ba4\u662f NCHW\u3002<code>torch.channels_last<\/code> \u662f\u5728 NCHW \u4e3b\u8def\u5f84\u4e4b\u4e0a\u63d0\u4f9b\u7684\u4e00\u79cd\u53ef\u9009\u5185\u5b58\u683c\u5f0f\uff0c\u7528\u6237\u9700\u8981\u663e\u5f0f\u8c03\u7528 <code>.to(memory_format=torch.channels_last)<\/code>\u3002\u5b83\u5e26\u6765\u4e86\u51e0\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4e0d\u662f\u6240\u6709\u7b97\u5b50\u90fd\u539f\u751f\u652f\u6301 channels_last<\/strong>\u3002\u67d0\u4e9b\u7b97\u5b50\u5728 NHWC \u8f93\u5165\u4e0b\u4f1a\u5185\u90e8\u8f6c\u56de NCHW\uff0c\u4ea7\u751f\u989d\u5916\u7684 transpose\uff1b<\/li>\n\n\n\n<li><strong>\u4e0e eager \u7684\u4ea4\u4e92\u590d\u6742<\/strong>\u3002\u52a8\u6001\u56fe\u91cc\u7b97\u5b50\u4e00\u4e2a\u63a5\u4e00\u4e2a\u5730\u8dd1\uff0c\u5e03\u5c40\u8f6c\u6362\u70b9\u4e0d\u56fa\u5b9a\uff0c\u6027\u80fd\u6ce2\u52a8\u5927\uff1b<\/li>\n\n\n\n<li><strong>\u9759\u6001\u4f18\u5316\u53d7\u9650<\/strong>\u3002<code>torch.compile<\/code> \u53ef\u4ee5\u505a\u5e03\u5c40\u4f20\u64ad\uff0c\u4f46\u56fe\u65ad\u88c2\uff08graph break\uff09\u548c\u52a8\u6001\u5f62\u72b6\u4f1a\u8ba9\u5b83\u96be\u4ee5\u4fdd\u8bc1\u7aef\u5230\u7aef NHWC\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u9009\u62e9\u662f\uff1a<strong>\u4ece\u7b2c\u4e00\u5c42\u5f00\u59cb\u5c31\u53ea\u8ba4 NHWC<\/strong>\u3002\u6ca1\u6709\u53ef\u9009\u683c\u5f0f\uff0c\u6ca1\u6709\u8fd0\u884c\u65f6\u8f6c\u6362\uff0c\u6240\u6709\u7b97\u5b50\u3001\u6240\u6709\u9884\u5904\u7406\u3001\u6240\u6709\u663e\u5b58\u5e03\u5c40\u90fd\u56f4\u7ed5 NHWC \u8bbe\u8ba1\u3002\u8fd9\u4e0d\u662f\u8bf4 NHWC \u666e\u904d\u4f18\u4e8e NCHW\uff0c\u800c\u662f\u5728\u6211\u4eec\u7684\u76ee\u6807\u573a\u666f\u2014\u2014GPU \u4e0a\u8ffd\u6c42\u6781\u81f4\u8bad\u7ec3\u541e\u5410\u2014\u2014\u4e0b\uff0c\u7edf\u4e00 NHWC \u662f\u6700\u7b80\u6d01\u3001\u6700\u53ef\u63a7\u7684\u65b9\u6848\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f20\u91cf\u770b\u8d77\u6765\u662f\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u91cc\u6700&#8221;\u65e0\u804a&#8221;\u7684\u90e8\u4ef6\uff0c\u4f46\u6b63\u662f\u8fd9\u4e9b&#8221;\u65e0\u804a&#8221;\u7684\u7ec6\u8282\uff0c\u51b3\u5b9a\u4e86\u6846\u67b6\u80fd\u4e0d\u80fd\u8dd1\u5f97\u5feb\u3001\u80fd\u4e0d\u80fd\u7a33\u3001\u80fd\u4e0d\u80fd\u6269\u5c55\u5230\u591a\u5361\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u5f20\u91cf\u8bbe\u8ba1\u53ef\u4ee5\u6982\u62ec\u4e3a\u56db\u53e5\u8bdd\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Shape \u56fa\u5b9a 4D NHWC<\/strong>\uff0c\u7edf\u4e00\u6240\u6709\u7b97\u5b50\u7684\u7ef4\u5ea6\u8bed\u4e49\uff1b<\/li>\n\n\n\n<li><strong>Tensor \u662f CPU \u7aef\u7d27\u51d1\u5bb9\u5668<\/strong>\uff0c\u9996\u5730\u5740 256 \u5bf9\u9f50\uff0c\u53ea\u8d1f\u8d23\u642c\u8fd0\uff1b<\/li>\n\n\n\n<li><strong>DTensor \u662f\u8bbe\u5907\u7aef\u7eaf\u63cf\u8ff0\u7b26<\/strong>\uff0c\u4e0d\u6301\u5185\u5b58\uff0c\u901a\u8fc7 offset \u5728\u591a\u5361\u95f4\u5171\u4eab\u540c\u4e00\u4efd\u5e03\u5c40\u56fe\u7eb8\uff1b<\/li>\n\n\n\n<li><strong>\u53cc\u8f68 stride \u5206\u79bb CUDA \u5bf9\u9f50\u4e0e CPU \u7d27\u51d1<\/strong>\uff0c\u8ba9 GPU \u7b97\u5b50\u8d70\u6700\u4f18\u8def\u5f84\uff0cCPU \u642c\u8fd0\u4fdd\u6301\u7b80\u5355\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u800c\u8fd9\u4e9b\u8bbe\u8ba1\u7684\u5171\u540c\u5e95\u5ea7\uff0c\u5c31\u662f <strong>MemoryPlan<\/strong>\u2014\u2014\u90a3\u4e2a\u628a\u663e\u5b58\u6309\u8bed\u4e49\u5206\u533a\u3001\u4e00\u904d\u7ebf\u6027\u5e03\u5c40\u3001\u4fdd\u8bc1 256 \u5b57\u8282\u5bf9\u9f50\u7684&#8221;\u663e\u5b58\u56fe\u7eb8\u7cfb\u7edf&#8221;\u3002DTensor \u7684 offset \u4ece\u54ea\u6765\u3001slot_bytes \u600e\u4e48\u5b9a\u3001\u591a\u5361\u5e03\u5c40\u5982\u4f55\u4e00\u81f4\uff0c\u6700\u7ec8\u90fd\u8981\u843d\u5230 MemoryPlan \u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u628a\u89c6\u89d2\u8f6c\u5411\u6570\u636e\u4fa7\uff0c\u770b\u770b Tech-Renaissance \u5982\u4f55\u7528\u540c\u4e00\u5957\u62bd\u8c61\u517c\u5bb9 MNIST\u3001CIFAR \u4e0e ImageNet\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\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\u4e03 \u4e0a\u4e00\u7bc7\u6211\u4eec\u628a Tech-Ren [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":673,"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-463","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\/463","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=463"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/463\/revisions"}],"predecessor-version":[{"id":672,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/463\/revisions\/672"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/673"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=463"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=463"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=463"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}