{"id":487,"date":"2026-07-08T03:00:13","date_gmt":"2026-07-07T19:00:13","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=487"},"modified":"2026-07-08T17:27:00","modified_gmt":"2026-07-08T09:27:00","slug":"fusednormalization%ef%bc%9a%e6%8a%8a%e6%95%b0%e6%8d%ae%e5%a2%9e%e5%bc%ba%e7%9a%84%e6%9c%80%e5%90%8e%e4%b8%80%e6%ad%a5%e8%9e%8d%e8%bf%9b%e4%b8%80%e6%ac%a1%e5%86%85%e5%ad%98%e9%81%8d%e5%8e%86","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/487\/","title":{"rendered":"(11) FusedNormalization\uff1a\u628a\u6570\u636e\u589e\u5f3a\u7684\u6700\u540e\u4e00\u6b65\u878d\u8fdb\u4e00\u6b21\u5185\u5b58\u904d\u5386"},"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\u5341\u4e00<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u524d\u9762\u4e09\u7bc7\u6587\u7ae0\u6211\u4eec\u628a\u6570\u636e\u5f15\u64ce\u8bb2\u4e86\u4e2a\u5927\u6982\uff1a\u7b2c 7 \u7bc7\u4ecb\u7ecd\u4e86\u7edf\u4e00\u7684\u6570\u636e\u52a0\u8f7d\u62bd\u8c61\uff0c\u7b2c 8 \u7bc7\u6df1\u5165\u4e86 DTS \u81ea\u5b9a\u4e49\u6570\u636e\u683c\u5f0f\uff0c\u7b2c 9 \u7bc7\u8bb2\u4e86\u591a\u7ebf\u7a0b\u9884\u5904\u7406\u3001NUMA \u611f\u77e5\u4e0e\u5f02\u6b65\u53cc\u7f13\u51b2\u3002\u5982\u679c\u628a\u8bad\u7ec3\u7ba1\u7ebf\u6bd4\u4f5c\u4e00\u6761\u9ad8\u901f\u516c\u8def\uff0cDTS \u89e3\u51b3\u7684\u662f&#8221;\u6e90\u5934\u6536\u8d39\u7ad9&#8221;\u7684\u95ee\u9898\uff0c\u591a\u7ebf\u7a0b\u548c\u53cc\u7f13\u51b2\u89e3\u51b3\u7684\u662f&#8221;\u8f66\u9053\u6570\u91cf\u548c\u5e76\u884c\u8c03\u5ea6&#8221;\u7684\u95ee\u9898\uff0c\u800c\u4eca\u5929\u8981\u8bb2\u7684 <strong>FusedNormalization<\/strong>\uff0c\u89e3\u51b3\u7684\u662f<strong>\u9884\u5904\u7406\u6d41\u6c34\u7ebf\u5185\u90e8\u6700\u540e\u4e00\u4e2a\u62e5\u5835\u70b9<\/strong>\u2014\u2014\u6570\u636e\u589e\u5f3a\u7684\u6700\u540e\u51e0\u6b65\u80fd\u4e0d\u80fd\u4e0d\u8981\u53cd\u590d\u8bfb\u5199\u5185\u5b58\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u95ee\u9898\u770b\u4f3c\u4e0d\u8d77\u773c\uff0c\u4f46\u5bf9\u4e8e ImageNet \u8fd9\u79cd\u6bcf\u4e2a epoch \u8981\u5904\u7406 128 \u4e07\u5f20\u56fe\u7247\u7684\u8bad\u7ec3\u4efb\u52a1\u6765\u8bf4\uff0cToTensor\u3001Normalize\u3001RandomHorizontalFlip\u3001RandomErasing \u8fd9\u4e9b\u64cd\u4f5c\u7684\u6392\u5217\u7ec4\u5408\u65b9\u5f0f\uff0c\u76f4\u63a5\u51b3\u5b9a\u4e86 CPU \u9884\u5904\u7406\u5230\u5e95\u80fd\u4e0d\u80fd\u5582\u9971 GPU\u3002Tech-Renaissance \u7684\u505a\u6cd5\u662f\u628a\u5b83\u4eec\u878d\u8fdb\u4e00\u6b21\u5185\u5b58\u904d\u5386\uff1a\u4e00\u5f20 uint8 \u56fe\u7247\u8fdb\u53bb\uff0c\u4e00\u5f20\u5df2\u7ecf\u5f52\u4e00\u5316\u597d\u3001\u53ef\u80fd\u5df2\u7ecf\u6c34\u5e73\u7ffb\u8f6c\u3001\u53ef\u80fd\u5df2\u7ecf\u968f\u673a\u64e6\u9664\u7684 float\/FP16 \u56fe\u7247\u51fa\u6765\u3002\u6ca1\u6709\u4e2d\u95f4\u5f20\u91cf\uff0c\u6ca1\u6709\u53cd\u590d\u62f7\u8d1d\uff0c\u6ca1\u6709 Python \u5c42\u7684\u51fd\u6570\u8c03\u7528\u5f00\u9500\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6570\u636e\u589e\u5f3a\uff1a\u8bad\u7ec3\u524d\u7684&#8221;\u6807\u51c6\u5316\u52a0\u5de5&#8221;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8fdb\u5165\u5b9e\u73b0\u7ec6\u8282\u4e4b\u524d\uff0c\u5148\u56de\u987e\u4e00\u4e0b\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u91cc\u5e38\u89c1\u7684\u9884\u5904\u7406\u6d41\u7a0b\u3002\u4ee5 ImageNet \u8bad\u7ec3\u4e3a\u4f8b\uff0c\u4e00\u5f20\u539f\u59cb JPEG \u56fe\u7247\u4ece DTS \u6587\u4ef6\u4e2d\u88ab\u8bfb\u53d6\u51fa\u6765\u540e\uff0c\u901a\u5e38\u4f1a\u7ecf\u5386\u4ee5\u4e0b\u6b65\u9aa4\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u89e3\u7801<\/strong>\uff1a\u628a JPEG \u5b57\u8282\u6d41\u89e3\u538b\u6210 uint8 \u7684\u50cf\u7d20\u6570\u7ec4\uff08H\u00d7W\u00d7C\uff09\u3002<\/li>\n\n\n\n<li><strong>\u968f\u673a\u88c1\u526a\/\u7f29\u653e<\/strong>\uff1a<code>RandomResizedCrop<\/code> \u628a\u56fe\u7247\u88c1\u5230 224\u00d7224\u3002<\/li>\n\n\n\n<li><strong>\u968f\u673a\u6c34\u5e73\u7ffb\u8f6c<\/strong>\uff1a\u4ee5 50% \u6982\u7387\u5de6\u53f3\u7ffb\u8f6c\u3002<\/li>\n\n\n\n<li><strong>\u7c7b\u578b\u8f6c\u6362<\/strong>\uff1a<code>ToTensor<\/code> \u628a uint8 \u7684 [0,255] \u6620\u5c04\u5230 float \u7684 [0,1]\u3002<\/li>\n\n\n\n<li><strong>\u5f52\u4e00\u5316<\/strong>\uff1a<code>Normalize<\/code> \u6309\u6bcf\u4e2a\u901a\u9053\u7684 mean\/std \u505a <code>(x - mean) \/ std<\/code>\u3002<\/li>\n\n\n\n<li><strong>\u968f\u673a\u64e6\u9664<\/strong>\uff1a\u6309\u4e00\u5b9a\u6982\u7387\u628a\u56fe\u7247\u67d0\u4e2a\u77e9\u5f62\u533a\u57df\u6e05\u96f6\uff08RandomErasing\uff09\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u6b65\u9aa4\u5728 PyTorch \u91cc\u901a\u5e38\u5199\u6210\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"python\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">transforms.Compose([\n    RandomResizedCrop(224),\n    RandomHorizontalFlip(),\n    ToTensor(),\n    Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),\n    RandomErasing(p=0.5)\n])<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u5199\u6cd5\u6e05\u6670\u3001\u6a21\u5757\u5316\uff0c\u6bcf\u4e2a\u53d8\u6362\u8d1f\u8d23\u4e00\u4ef6\u4e8b\u3002\u4f46\u4ece\u6027\u80fd\u89d2\u5ea6\u770b\uff0c\u5b83\u6709\u4e00\u4e2a\u9690\u5f62\u6210\u672c\uff1a<strong>\u6bcf\u7ecf\u8fc7\u4e00\u4e2a\u53d8\u6362\uff0c\u5c31\u8981\u4ea7\u751f\u4e00\u4efd\u4e2d\u95f4\u7ed3\u679c<\/strong>\u3002<code>ToTensor<\/code> \u8f93\u51fa\u4e00\u4efd float \u5f20\u91cf\uff0c<code>Normalize<\/code> \u5728\u4e0a\u9762\u518d\u505a\u4e00\u4efd\uff0c<code>RandomErasing<\/code> \u53c8\u8981\u518d\u8bfb\u5199\u4e00\u904d\u3002\u5bf9\u4e8e 224\u00d7224\u00d73 \u7684\u56fe\u7247\uff0c\u4e00\u5f20\u56fe\u5c31\u662f 602 KB \u7684 float \u6570\u636e\uff1b\u4e00\u4e2a batch 256 \u5f20\u5c31\u662f 150 MB\uff1b\u4e00\u4e2a epoch 128 \u4e07\u5f20\u5c31\u662f\u7ea6 750 GB \u7684\u4e2d\u95f4\u5185\u5b58\u6d41\u91cf\u3002\u8fd9\u8fd8\u6ca1\u7b97\u6c34\u5e73\u7ffb\u8f6c\u3001\u64e6\u9664\u7b49\u64cd\u4f5c\u5bf9\u540c\u4e00\u4efd\u6570\u636e\u7684\u53cd\u590d\u8bbf\u95ee\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u66f4\u5fae\u5999\u7684\u5730\u65b9\u5728\u4e8e\u7f13\u5b58\u3002\u6bcf\u4e2a\u64cd\u4f5c\u72ec\u7acb\u5730\u628a\u6574\u5f20\u56fe\u4ece\u5185\u5b58\u52a0\u8f7d\u5230 CPU \u7f13\u5b58\uff0c\u5904\u7406\u5b8c\u518d\u5199\u56de\uff1b\u4e0b\u4e00\u4e2a\u64cd\u4f5c\u53c8\u4ece\u5185\u5b58\u91cd\u65b0\u52a0\u8f7d\u3002\u5982\u679c\u6570\u636e\u91cf\u8d85\u51fa L2\/L3 \u5bb9\u91cf\uff0c\u540c\u4e00\u4e2a\u50cf\u7d20\u4f1a\u88ab\u53cd\u590d\u9a71\u9010\u548c\u91cd\u8f7d\u3002\u5bf9\u4e8e CPU \u4fa7\u9884\u5904\u7406\u6765\u8bf4\uff0c\u5185\u5b58\u5e26\u5bbd\u548c\u7f13\u5b58\u5229\u7528\u7387\u5f80\u5f80\u662f\u771f\u6b63\u7684\u74f6\u9888\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u4e3b\u6d41\u6846\u67b6\u5982\u4f55\u5904\u7406\u8fd9\u4e2a\u95ee\u9898<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 PyTorch torchvision\uff1a\u7075\u6d3b\u4f46\u72ec\u7acb\u904d\u5386<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">torchvision \u7684 v1 transforms \u628a\u6bcf\u4e2a\u64cd\u4f5c\u5c01\u88c5\u6210\u72ec\u7acb\u7684 Python callable\uff0c\u4e3b\u8981\u9762\u5411 PIL Image \u8bbe\u8ba1\uff0c\u5bf9 <code>torch.Tensor<\/code> \u7684\u652f\u6301\u5e76\u4e0d\u7edf\u4e00\u5b8c\u6574\u3002\u597d\u5904\u662f\u63a5\u53e3\u7b80\u5355\u3001\u6613\u4e8e\u7ec4\u5408\u3001\u8c03\u8bd5\u65b9\u4fbf\uff1b\u574f\u5904\u662f\u6bcf\u4e2a transform \u90fd\u662f\u4e00\u6b21\u72ec\u7acb\u7684\u5185\u5b58\u904d\u5386\uff0c\u6bcf\u6b21\u90fd\u8981\u7ecf\u8fc7 Python \u6d3e\u53d1\u3001\u5f20\u91cf\u521b\u5efa\u548c\u5185\u5b58\u5206\u914d\uff0c\u53d8\u6362\u4e4b\u95f4\u5f88\u96be\u505a\u8de8\u6b65\u9aa4\u4f18\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">v2 transforms \u771f\u6b63\u6539\u4e3a\u4ee5 <code>torch.Tensor<\/code> \u4e3a\u6838\u5fc3\u5b9e\u73b0\uff0c\u5e76\u6269\u5c55\u652f\u6301\u4e86 bbox\u3001mask \u7b49\u7ed3\u6784\u5316\u6570\u636e\uff0c\u4e5f\u5f15\u5165\u4e86\u4e00\u4e9b\u5185\u90e8\u878d\u5408\uff08\u4f8b\u5982\u5bf9\u8fb9\u754c\u6846\u7684 <code>SanitizeBoundingBox<\/code>\u3001<code>ConvertBoundingBox<\/code> \u7b49\uff09\uff0c\u5bf9\u6027\u80fd\u6709\u4e00\u5b9a\u6539\u5584\u3002\u4f46\u53ea\u8981\u4ecd\u7136\u4ee5\u72ec\u7acb\u7684 transform \u5bf9\u8c61\u7ec4\u5408\uff0c\u5c31\u96be\u4ee5\u5728\u8bed\u4e49\u4e0a\u628a ToTensor\u3001Normalize\u3001Flip\u3001Erase \u5408\u5e76\u6210\u4e00\u6b21\u50cf\u7d20\u7ea7\u904d\u5386\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 NVIDIA DALI\uff1a\u628a\u9884\u5904\u7406\u642c\u5230 GPU<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">NVIDIA DALI \u662f\u53e6\u4e00\u6761\u8def\u7ebf\uff1aJPEG \u89e3\u7801\u3001resize\u3001normalize\u3001flip \u5168\u90e8\u653e\u5230 GPU \u4e0a\u505a\u3002\u5b83\u7684 <code>fn.crop_mirror_normalize<\/code> \u53ef\u4ee5\u628a\u7c7b\u578b\u8f6c\u6362\u3001\u5f52\u4e00\u5316\u3001\u88c1\u526a\u3001\u7ffb\u8f6c\u5408\u5e76\u5728\u4e00\u6b21 CUDA kernel \u91cc\u5b8c\u6210\uff0c\u6027\u80fd\u975e\u5e38\u51fa\u8272\u3002\u4f46 DALI \u662f NVIDIA \u4e13\u7528\u65b9\u6848\uff0c\u4e3b\u8981\u9762\u5411 Linux\uff0c\u4e14\u5bf9\u8bad\u7ec3\u6846\u67b6\u7684\u4fb5\u5165\u6027\u8f83\u5f3a\u2014\u2014\u4f60\u9700\u8981\u7528 DALI \u7684 pipeline \u66ff\u4ee3\u539f\u6709\u6570\u636e\u52a0\u8f7d\u5668\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 TensorFlow tf.data\uff1a\u56fe\u7ea7\u4f18\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">TensorFlow \u7684 <code>tf.data<\/code> \u7ba1\u7ebf\u4f1a\u628a\u591a\u4e2a <code>map<\/code> \u64cd\u4f5c\u5c3d\u91cf\u878d\u5408\u3002\u5b83\u4e3b\u8981\u89e3\u51b3\u7684\u662f Python \u51fd\u6570\u8c03\u7528\u5f00\u9500\u548c\u5e76\u884c\u5ea6\u95ee\u9898\uff0c\u5bf9\u4e8e\u5177\u4f53\u7684 ToTensor+Normalize \u8fd9\u79cd\u6570\u503c\u7ea7\u878d\u5408\uff0c\u4f18\u52bf\u4e0d\u5982\u624b\u5199 SIMD kernel \u660e\u663e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.4 \u878d\u5408\u7684\u901a\u7528\u4ef7\u503c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u65e0\u8bba\u54ea\u79cd\u6846\u67b6\uff0c\u5927\u5bb6\u5176\u5b9e\u90fd\u5728\u505a\u540c\u4e00\u4ef6\u4e8b\uff1a<strong>\u51cf\u5c11\u4e2d\u95f4\u5f20\u91cf\u3001\u51cf\u5c11\u5185\u5b58\u904d\u5386\u3001\u51cf\u5c11\u6570\u636e\u642c\u79fb<\/strong>\u3002\u5728 CPU \u4fa7\uff0c\u5185\u5b58\u5e26\u5bbd\u5f80\u5f80\u662f\u74f6\u9888\uff1b\u5728 GPU \u4fa7\uff0ckernel launch \u548c\u663e\u5b58\u8bfb\u5199\u662f\u74f6\u9888\u3002\u628a\u591a\u4e2a\u5c0f\u64cd\u4f5c\u5408\u5e76\u6210\u4e00\u6b21\u904d\u5386\uff0c\u662f\u7cfb\u7edf\u4f18\u5316\u91cc\u6700\u57fa\u672c\u4e5f\u6700\u6709\u6548\u7684\u624b\u6bb5\u4e4b\u4e00\u3002Tech-Renaissance \u7684 FusedNormalization \u8d70\u7684\u5c31\u662f\u8fd9\u6761\u8def\uff0c\u53ea\u4e0d\u8fc7\u5b83\u53d1\u751f\u5728 CPU \u9884\u5904\u7406\u9636\u6bb5\uff0c\u4e3a\u540e\u7eed\u7684 H2D \u4f20\u8f93\u548c GPU \u8ba1\u7b97\u51c6\u5907\u6570\u636e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001FusedNormalization \u5728 Tech-Renaissance \u4e2d\u7684\u5b9a\u4f4d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u9884\u5904\u7406\u7ba1\u7ebf\u7531 <code>Preprocessor<\/code> \u7edf\u4e00\u7ba1\u7406\u3002\u7528\u6237\u5728 <code>Setup<\/code> \u91cc\u914d\u7f6e\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\u7684 transforms\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=\"\">PREPROCESSOR_SETTING\n    .dataset(\"IMAGENET\", data_path)\n    .normalization(NormMode::IMAGENET)\n    .train_transforms(\n        RandomResizedCrop(224),\n        RandomHorizontalFlip(),\n        RandomErasing(0.5f)\n    )\n    .val_transforms(\n        Resize(256),\n        CenterCrop(224)\n    )\n    .commit();<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u8fd9\u91cc <code>RandomHorizontalFlip<\/code>\u3001<code>RandomErasing<\/code> \u548c <code>Normalize<\/code> \u7684\u8bed\u4e49\u88ab<strong>\u6846\u67b6\u63a5\u7ba1\u4e86<\/strong>\u3002<code>Normalize<\/code> \u751a\u81f3\u4e0d\u5141\u8bb8\u76f4\u63a5\u51fa\u73b0\u5728 <code>.train_transforms()<\/code> \u91cc\u2014\u2014\u7528\u6237\u5fc5\u987b\u901a\u8fc7 <code>.normalization(NormMode::...)<\/code> \u6765\u914d\u7f6e\u5f52\u4e00\u5316\u53c2\u6570\u3002\u771f\u6b63\u5e72\u6d3b\u7684\u662f\u4e00\u4e2a\u5728 PO\uff08PreprocessOperation\uff09\u94fe\u672b\u5c3e<strong>\u81ea\u52a8\u6ce8\u5165<\/strong>\u7684 <code>FusedNormalization<\/code> \u5bf9\u8c61\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8bbe\u8ba1\u6709\u4e24\u4e2a\u5173\u952e\u542b\u4e49\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\uff0c<strong>FusedNormalization \u662f\u6846\u67b6\u5185\u90e8\u7c7b\uff0c\u7528\u6237\u4e0d\u53ef\u76f4\u63a5\u6784\u9020<\/strong>\u3002\u5b83\u88ab\u786c\u7f16\u7801\u4e3a PO \u94fe\u7684\u6700\u540e\u4e00\u4e2a\u64cd\u4f5c\uff0c\u56e0\u4e3a\u81f3\u5c11 ToTensor\uff08uint8 \u2192 float\/FP16\uff09\u8fd9\u4e00\u6b65\u65e0\u6cd5\u7701\u7565\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\uff0c<strong>RandomHorizontalFlip \u548c RandomErasing \u88ab\u964d\u7ea7\u4e3a&#8221;\u53c2\u6570\u8bb0\u5f55\u7c7b&#8221;<\/strong>\u3002\u5b83\u4eec\u53ea\u8d1f\u8d23\u628a\u53c2\u6570\u4f20\u7ed9 <code>FusedNormalization<\/code>\uff0c\u81ea\u5df1\u7684 <code>execute()<\/code> \u4e0d\u4f1a\u88ab\u8c03\u7528\u3002\u8fd9\u907f\u514d\u4e86\u72ec\u7acb\u6267\u884c\u5e26\u6765\u7684\u989d\u5916\u5185\u5b58\u62f7\u8d1d\u3002<code>RandomHorizontalFlip::execute()<\/code> \u548c <code>RandomErasing::execute()<\/code> \u90fd\u76f4\u63a5\u629b\u51fa <code>TR_NOT_IMPLEMENTED<\/code>\uff0c\u660e\u786e\u544a\u77e5\u5f00\u53d1\u8005\u8fd9\u4e9b\u64cd\u4f5c\u5df2\u7ecf\u88ab\u878d\u5408\u8fdb FusedNormalization\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u4ee3\u7801\u5c42\u9762\u7684\u81ea\u52a8\u6ce8\u5165\u673a\u5236<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">PO \u94fe\u7684\u6784\u9020\u53d1\u751f\u5728 <code>Preprocessor::set_train_transforms()<\/code> \u4e2d\u3002\u6846\u67b6\u904d\u5386\u7528\u6237\u4f20\u5165\u7684\u6240\u6709 transforms\uff0c\u628a\u4e09\u4e2a\u8bb0\u5f55\u7c7b\uff08<code>RandomErasing<\/code>\u3001<code>RandomHorizontalFlip<\/code>\u3001<code>Normalize<\/code>\uff09\u7684\u53c2\u6570\u63d0\u53d6\u51fa\u6765\uff0c\u5176\u4ed6\u771f\u5b9e PO\uff08\u5982 <code>RandomResizedCrop<\/code>\u3001<code>ColorJitter<\/code>\uff09\u4fdd\u7559\u5728 <code>filtered_ops<\/code> \u4e2d\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\/data\/preprocessor.cpp\nbool flip_enabled = false;\nbool erase_enabled = false;\nfloat erase_p = 0.0f;\nfloat erase_scale_min = 0.02f;\nfloat erase_scale_max = 0.33f;\nfloat erase_ratio_min = 0.3f;\nfloat erase_ratio_max = 3.3f;\nNormMode norm_mode = NormMode::NO_NORM;\n\nstd::vector&lt;std::unique_ptr&lt;PreprocessOperation>> filtered_ops;\nfor (auto&amp; op : temp_ops) {\n    if (auto* re = dynamic_cast&lt;RandomErasing*>(op.get())) {\n        erase_enabled = true;\n        erase_p = re->get_p();\n        erase_scale_min = re->scale_min();\n        erase_scale_max = re->scale_max();\n        erase_ratio_min = re->ratio_min();\n        erase_ratio_max = re->ratio_max();\n    } else if (auto* rhf = dynamic_cast&lt;RandomHorizontalFlip*>(op.get())) {\n        flip_enabled = true;\n    } else if (auto* norm = dynamic_cast&lt;Normalize*>(op.get())) {\n        norm_mode = norm->mode();\n    } else {\n        filtered_ops.push_back(std::move(op));\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6846\u67b6\u63d0\u53d6\u51fa\u4e09\u4e2a\u8bb0\u5f55\u7c7b\u7684\u53c2\u6570\u540e\uff0c\u628a\u771f\u5b9e\u7684 resize\/crop \u7b49\u64cd\u4f5c\u4fdd\u7559\u5728 <code>filtered_ops<\/code> \u4e2d\uff0c\u7136\u540e\u5728 PO \u94fe\u672b\u5c3e\u6ce8\u5165 <code>FusedNormalization<\/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=\"\">\/\/ src\/data\/preprocessor.cpp\nNormalizePreset preset = NormalizePreset::NO_NORM;\nswitch (norm_mode) {\n    case NormMode::NO_NORM:  preset = NormalizePreset::NO_NORM;  break;\n    case NormMode::MNIST:    preset = NormalizePreset::MNIST;    break;\n    case NormMode::CIFAR:    preset = NormalizePreset::CIFAR;    break;\n    case NormMode::IMAGENET: preset = NormalizePreset::IMAGENET; break;\n    case NormMode::MLPERF:   preset = NormalizePreset::MLPERF;   break;\n}\n\nbool use_amp = GlobalRegistry::instance().using_amp();\n\nauto fused_norm = std::make_unique&lt;FusedNormalization>(\n    preset, use_amp, flip_enabled, erase_enabled,\n    erase_p, erase_scale_min, erase_scale_max,\n    erase_ratio_min, erase_ratio_max, 0\n);\ntrain_ops_template_.push_back(std::move(fused_norm));<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u9a8c\u8bc1\u96c6\u7684\u5904\u7406\u7c7b\u4f3c\uff0c\u4f46 <code>erase_enabled<\/code> \u5f3a\u5236\u4e3a <code>false<\/code>\uff0c\u56e0\u4e3a RandomErasing \u53ea\u5141\u8bb8\u7528\u4e8e\u8bad\u7ec3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001FusedNormalization \u878d\u5408\u4e86\u4ec0\u4e48<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>FusedNormalization<\/code> \u8981\u505a\u7684\u4e8b\u60c5\u662f\u8fd9\u6837\u7684\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\/data\/fused_normalization.h\n\/**\n * @class FusedNormalization\n * @brief \u878d\u5408\u5f52\u4e00\u5316\u64cd\u4f5c\uff1aToTensor + RandomHorizontalFlip + Normalize + RandomErasing\n *\n * \u8fd9\u662fTR4\u6846\u67b6\u6700\u91cd\u8981\u7684\u9884\u5904\u7406\u64cd\u4f5c\uff0c\u5b83\u5c06PyTorch\u4e2d\u56db\u4e2a\u72ec\u7acb\u7684\u9884\u5904\u7406\u6b65\u9aa4\u878d\u5408\u4e3a\u4e00\u6b21\u5185\u5b58\u904d\u5386\uff1a\n * - ToTensor\uff1auint8_t[0,255] \u2192 float[0,1]\uff08\u9664\u4ee5255\uff09\n * - RandomHorizontalFlip\uff1a50%\u6982\u7387\u6c34\u5e73\u7ffb\u8f6c\uff08\u53ef\u9009\uff0cflip_enabled_\u63a7\u5236\uff09\n * - Normalize\uff1a(x - mean) \/ stddev\uff0c\u652f\u6301ImageNet\/MNIST\/CIFAR\/MLPerf\u56db\u79cd\u9884\u8bbe\n * - RandomErasing\uff1a\u968f\u673a\u77e9\u5f62\u533a\u57df\u64e6\u9664\uff08\u53ef\u9009\uff0cerase_enabled_\u63a7\u5236\uff09\n *\/<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\uff0c<strong>\u4e00\u5f20 uint8 \u56fe\u7247\u8fdb\u5165 <code>execute()<\/code> \u540e\uff0c\u7ecf\u8fc7\u4e00\u6b21\u7edf\u4e00\u7684 <code>execute()<\/code> \u8c03\u7528\uff0c\u5c31\u5f97\u5230\u4e86\u6700\u7ec8\u8981\u4f20\u7ed9 GPU \u7684 float \u6216 FP16 \u6570\u636e<\/strong>\u3002\u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>ToTensor + RandomHorizontalFlip + Normalize \u5728\u540c\u4e00\u6b21\u50cf\u7d20\u904d\u5386\u4e2d\u5b8c\u6210<\/strong>\uff1a\u6bcf\u8bfb\u5165\u4e00\u4e2a\u50cf\u7d20\uff0c\u5148\u51b3\u5b9a\u4ece\u955c\u50cf\u4f4d\u7f6e\u8fd8\u662f\u539f\u4f4d\u7f6e\u8bfb\u53d6\uff0c\u7136\u540e\u505a ToTensor \u548c Normalize\uff0c\u6700\u540e\u76f4\u63a5\u5199\u5165\u8f93\u51fa\u7f13\u51b2\u533a\u3002<\/li>\n\n\n\n<li><strong>RandomErasing \u4f5c\u4e3a\u540c\u4e00\u6b21 <code>execute()<\/code> \u8c03\u7528\u5185\u7684\u8f7b\u91cf\u540e\u7eed\u6b65\u9aa4\u5b8c\u6210<\/strong>\uff1a\u7531\u4e8e\u64e6\u9664\u533a\u57df\u53ea\u5360\u56fe\u50cf\u7684\u4e00\u5c0f\u90e8\u5206\uff0cFusedNormalization \u5728\u4e3b\u904d\u5386\u7ed3\u675f\u540e\u5bf9\u8f93\u51fa\u7f13\u51b2\u533a\u7684\u5bf9\u5e94\u533a\u57df\u505a\u4e00\u6b21 <code>memset<\/code> \u6e05\u96f6\uff0c\u800c\u4e0d\u662f\u518d\u5bf9\u6574\u5f20\u56fe\u505a\u4e00\u6b21\u5b8c\u6574\u904d\u5386\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u9700\u8981\u5148\u8f6c float \u518d normalize \u518d flip \u518d erase\uff0c\u6240\u6709\u51b3\u7b56\u548c\u8ba1\u7b97\u90fd\u5728\u540c\u4e00\u4e2a\u51fd\u6570\u8c03\u7528\u5185\u5b8c\u6210\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6267\u884c\u5165\u53e3\u662f <code>execute()<\/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=\"\">\/\/ src\/data\/fused_normalization.cpp\nvoid FusedNormalization::execute(\n    const uint8_t* input_ptr,\n    int32_t input_width, int32_t input_height,\n    size_t input_stride,\n    uint8_t* output_ptr,\n    int32_t&amp; output_width, int32_t&amp; output_height,\n    size_t&amp; output_stride,\n    Generator* rng, ...\n) {\n    output_width = input_width;\n    output_height = input_height;\n\n    TR_CHECK(output_stride == 0, ValueError,\n             \"FusedNormalization does not support external output_stride setting. \"\n             \"Its output stride is determined solely by output_size, num_channels, and AMP mode.\");\n    output_stride = compact_output_stride_;\n\n    bool do_flip = false;\n    if (flip_enabled_) {\n        do_flip = (uniform(0.0f, 1.0f, rng) &lt; 0.5f);\n    }\n\n    EraseRect erase_rect;\n    if (erase_enabled_) {\n        erase_rect = generate_erase_rect(input_height, input_width, rng);\n    }\n    \/\/ ... \u4e3b\u904d\u5386\u5b8c\u6210 ToTensor + Flip + Normalize\uff0c\u968f\u540e\u6309\u9700 apply_erase\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u4e24\u4e2a\u968f\u673a\u51b3\u7b56\u70b9\uff1a\u662f\u5426\u6c34\u5e73\u7ffb\u8f6c\u3001\u662f\u5426\u4ee5\u53ca\u5728\u54ea\u91cc\u64e6\u9664\u3002\u4e24\u4e2a\u51b3\u7b56\u90fd\u5728\u904d\u5386\u5f00\u59cb\u524d\u5b8c\u6210\uff0c\u968f\u540e\u8fdb\u5165\u7edf\u4e00\u7684\u50cf\u7d20\u5904\u7406\u5faa\u73af\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001FP32 \u4e0e FP16 \u53cc\u7cbe\u5ea6\u8f93\u51fa<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">FusedNormalization \u652f\u6301\u4e24\u79cd\u8f93\u51fa\u683c\u5f0f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>FP32<\/strong>\uff1a\u6bcf\u4e2a\u50cf\u7d20 <code>num_channels<\/code> \u4e2a float\uff0c\u65e0\u901a\u9053 padding\u3002<\/li>\n\n\n\n<li><strong>FP16\/AMP<\/strong>\uff1a\u6bcf\u4e2a\u50cf\u7d20\u56fa\u5b9a 4 \u4e2a <code>uint16_t<\/code>\uff0c\u6709\u6548\u901a\u9053\u4e0d\u8db3 4 \u65f6\u7528 0 \u586b\u5145\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">FP16 \u8def\u5f84\u662f\u4e3a\u4e86\u914d\u5408\u6846\u67b6\u7684\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u3002AMP \u6a21\u5f0f\u4e0b\uff0cGPU \u62ff\u5230\u7684\u5c31\u662f\u5df2\u7ecf\u538b\u7f29\u597d\u7684 FP16 \u6570\u636e\uff0cH2D \u4f20\u8f93\u91cf\u76f4\u63a5\u51cf\u534a\u3002\u800c\u4e14 FusedNormalization \u5728 CPU \u7aef\u505a FP32\u2192FP16 \u7684\u8f6c\u6362\uff0c\u8c03\u7528\u7684\u662f <strong>F16C<\/strong> \u6269\u5c55\u7684 <code>_mm_cvtps_ph<\/code> \u6307\u4ee4\uff08\u5e38\u89c1\u652f\u6301 AVX2 \u7684 CPU \u5747\u652f\u6301 F16C\uff09\uff0c\u6bd4 GPU \u7aef\u505a\u8f6c\u6362\u66f4\u7701\u663e\u5b58\u5e26\u5bbd\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=\"\">\/\/ src\/data\/fused_normalization.cpp\ninline std::uint16_t fp32_to_half(float f) noexcept {\n    __m128 v32 = _mm_set_ss(f);\n    __m128i v16 = _mm_cvtps_ph(v32, 0);\n    return static_cast&lt;std::uint16_t>(_mm_cvtsi128_si32(v16));\n}<\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001SIMD \u4e0e F16C\uff1a\u4e00\u6b21\u5904\u7406\u591a\u4e2a\u50cf\u7d20<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">FusedNormalization \u7684 FP16 \u8def\u5f84\u501f\u52a9 <strong>SSE \u5411\u91cf\u5316\u4e0e F16C \u786c\u4ef6\u6269\u5c55<\/strong>\u505a\u4e86\u4e13\u95e8\u4f18\u5316\u3002\u4ee5 RGB \u4e09\u901a\u9053\u4e3a\u4f8b\uff0c\u6838\u5fc3\u51fd\u6570 <code>simd_process_2pixels_c3<\/code> \u4e00\u6b21\u8bfb\u53d6 8 \u5b57\u8282\uff08\u4e24\u4e2a\u50cf\u7d20\uff0c6 \u4e2a\u6709\u6548\u901a\u9053 + 2 \u5b57\u8282\u586b\u5145\uff09\uff0c\u628a\u5b83\u6269\u5c55\u6210 8 \u4e2a 32 \u4f4d\u6574\u6570\uff0c\u518d\u8f6c\u6210 8 \u4e2a float\uff0c\u4e58\u4e0a\u9884\u8ba1\u7b97\u7684 <code>mul = 1\/(255*std)<\/code>\u3001\u51cf\u53bb <code>sub = mean\/std<\/code>\uff0c\u6700\u540e\u7528 <code>_mm_cvtps_ph<\/code> \u6253\u5305\u6210 8 \u4e2a FP16\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\/data\/fused_normalization.cpp\ninline void simd_process_2pixels_c3(const std::uint8_t* p, std::uint16_t* dst,\n                                     __m128 mul_v, __m128 sub_v) noexcept {\n    __m128i u8x8 = _mm_loadl_epi64(reinterpret_cast&lt;const __m128i*>(p));\n    __m128i i32_0 = _mm_cvtepu8_epi32(u8x8);\n    __m128 f0 = _mm_cvtepi32_ps(i32_0);\n    f0 = _mm_mul_ps(f0, mul_v);      \/\/ \u666e\u901a\u6807\u91cf\u5e7f\u64ad\u4e58\u6cd5\n    f0 = _mm_sub_ps(f0, sub_v);\n    __m128i h0 = _mm_cvtps_ph(f0, 0);\n\n    __m128i shifted = _mm_srli_si128(u8x8, 3);\n    __m128i i32_1 = _mm_cvtepu8_epi32(shifted);\n    __m128 f1 = _mm_cvtepi32_ps(i32_1);\n    f1 = _mm_mul_ps(f1, mul_v);\n    f1 = _mm_sub_ps(f1, sub_v);\n    __m128i h1 = _mm_cvtps_ph(f1, 0);\n\n    __m128i h01 = _mm_unpacklo_epi64(h0, h1);\n    _mm_storeu_si128(reinterpret_cast&lt;__m128i*>(dst), h01);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ffb\u8f6c\u8def\u5f84 <code>simd_process_2pixels_c3_flip<\/code> \u5728\u6b64\u57fa\u7840\u4e0a\u628a\u7ed3\u679c\u5199\u5230\u76ee\u6807\u884c\u7684\u955c\u50cf\u4f4d\u7f6e\uff0c\u5b9e\u73b0&#8221;\u8fb9\u8f6c\u6362\u8fb9\u7ffb\u8f6c&#8221;\uff0c\u4e0d\u9700\u8981\u5148 flip \u518d normalize\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\/data\/fused_normalization.cpp\ninline void simd_process_2pixels_c3_flip(...) {\n    \/\/ ... \u4e0e noflip \u76f8\u540c\u7684\u8f6c\u6362\u903b\u8f91 ...\n    __m128i h01_swapped = _mm_shuffle_epi32(h01, _MM_SHUFFLE(1, 0, 3, 2));\n    std::uint16_t* pos = dst + (W - 1 - (w + 1)) * 4;\n    _mm_storeu_si128(reinterpret_cast&lt;__m128i*>(pos), h01_swapped);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5355\u901a\u9053 MNIST \u5219\u6709\u4e13\u95e8\u7684 <code>simd_process_4pixels_c1<\/code>\uff0c\u4e00\u6b21\u5904\u7406 4 \u4e2a\u7070\u5ea6\u50cf\u7d20\uff0c\u8f93\u51fa 4 \u4e2a FP16 \u503c\u5e76 padding \u5230 4 \u901a\u9053\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\/data\/fused_normalization.cpp\ninline void simd_process_4pixels_c1(...) {\n    int v; std::memcpy(&amp;v, p, sizeof(v));\n    __m128i u8x4 = _mm_cvtsi32_si128(v);\n    __m128i i32 = _mm_cvtepu8_epi32(u8x4);\n    __m128 f = _mm_cvtepi32_ps(i32);\n    f = _mm_mul_ps(f, mul_v);\n    f = _mm_sub_ps(f, sub_v);\n    __m128i h = _mm_cvtps_ph(f, 0);\n\n    __m128i zero = _mm_setzero_si128();\n    __m128i t0 = _mm_unpacklo_epi16(h, zero);\n    _mm_storeu_si128(reinterpret_cast&lt;__m128i*>(dst),      _mm_unpacklo_epi32(t0, zero));\n    _mm_storeu_si128(reinterpret_cast&lt;__m128i*>(dst + 8),  _mm_unpackhi_epi32(t0, zero));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u975e AVX2 \u5e73\u53f0\u5219\u56de\u9000\u5230 Eigen3 \u5b9e\u73b0 FP32 \u8def\u5f84\uff0cAMP \u5728\u975e AVX2 \u73af\u5883\u4e0b\u76f4\u63a5\u629b\u51fa <code>TR_NOT_IMPLEMENTED<\/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=\"\">\/\/ src\/data\/fused_normalization.cpp\n#if !defined(__AVX2__)\n    if (use_amp_) {\n        TR_NOT_IMPLEMENTED(\n            \"AMP=ON (FP16 output) is not supported in non-AVX2 mode. \"\n            \"FusedNormalization requires AVX2 for AMP\/FP16 support.\"\n        );\n    }\n#endif<\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u30014 \u901a\u9053 padding\uff1a\u4e3a\u4e86 GPU \u53cb\u597d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">FP16 \u8def\u5f84\u4e0b\uff0c\u5373\u4f7f\u539f\u59cb\u56fe\u7247\u662f 3 \u901a\u9053 RGB\uff0cFusedNormalization \u4e5f\u4f1a\u628a\u6bcf\u4e2a\u50cf\u7d20 padding \u5230 4 \u901a\u9053\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\/data\/fused_normalization.cpp\nif (use_amp_) {\n    compact_output_stride_ = static_cast&lt;size_t>(output_size_) * 4 * sizeof(uint16_t);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">224\u00d7224 \u7684 RGB \u56fe\u7247\uff0cFP32 \u8f93\u51fa stride \u662f <code>224 \u00d7 3 \u00d7 4 = 2688<\/code> \u5b57\u8282\uff1bFP16 \u8f93\u51fa stride \u662f <code>224 \u00d7 4 \u00d7 2 = 1792<\/code> \u5b57\u8282\u3002\u8fd9\u4e2a 4 \u901a\u9053\u5bf9\u9f50\u6709\u51e0\u4e2a\u597d\u5904\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6bcf\u4e2a\u50cf\u7d20 8 \u5b57\u8282<\/strong>\uff0c\u5929\u7136 64 \u4f4d\u5bf9\u9f50\uff0c\u7b26\u5408 GPU \u5168\u5c40\u5185\u5b58\u5408\u5e76\u8bbf\u95ee\u7684\u504f\u597d\u3002<\/li>\n\n\n\n<li><strong>H2D \u62f7\u8d1d\u548c\u540e\u7eed CUDA kernel \u8bfb\u53d6\u66f4\u7b80\u5355<\/strong>\uff0c\u53ef\u4ee5\u7528\u7edf\u4e00\u7684 <code>(h, w, 4)<\/code> \u7d22\u5f15\u3002<\/li>\n\n\n\n<li><strong>\u4e0e Tensor Core \u7684\u8f93\u5165\u683c\u5f0f\u66f4\u5951\u5408<\/strong>\uff1aAMP \u8bad\u7ec3\u901a\u5e38\u9700\u8981 FP16 \u8f93\u5165\uff0c4 \u901a\u9053\u5bf9\u9f50\u6709\u52a9\u4e8e\u540e\u7eed\u5377\u79ef\u7b49\u7b97\u5b50\u9ad8\u6548\u8bfb\u53d6\u3002<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001\u4e94\u79cd\u5f52\u4e00\u5316\u9884\u8bbe<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">FusedNormalization \u5185\u7f6e\u4e86\u4e94\u79cd\u6570\u636e\u96c6\u9884\u8bbe\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\/data\/fused_normalization.cpp\nParams get_params(NormalizePreset preset) noexcept {\n    switch (preset) {\n        case NormalizePreset::NO_NORM:\n            return { {0.0f, 0.0f, 0.0f}, {1.0f, 1.0f, 1.0f}, 3 };\n        case NormalizePreset::MNIST:\n            return { {0.1307f, 0.0f, 0.0f}, {0.3081f, 1.0f, 1.0f}, 1 };\n        case NormalizePreset::CIFAR:\n            return { {0.4914f, 0.4822f, 0.4465f}, {0.2470f, 0.2435f, 0.2616f}, 3 };\n        case NormalizePreset::IMAGENET:\n            return { {0.485f, 0.456f, 0.406f}, {0.229f, 0.224f, 0.225f}, 3 };\n        case NormalizePreset::MLPERF:\n            return { {123.68f\/255.0f, 116.78f\/255.0f, 103.94f\/255.0f},\n                     {1.0f\/255.0f, 1.0f\/255.0f, 1.0f\/255.0f}, 3 };\n    }\n    return {};\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>NO_NORM<\/code> \u7b49\u4ef7\u4e8e\u53ea\u9664\u4ee5 255\uff0c\u4e0d\u505a mean\/std \u8c03\u6574\uff1b<code>MLPERF<\/code> \u662f ImageNet \u7684\u53e6\u4e00\u79cd\u5e38\u7528\u8868\u8fbe\u5f62\u5f0f\uff0c\u628a mean \u548c std \u90fd\u5148\u9664\u4ee5 255\uff0c\u8fd9\u6837 ToTensor \u548c Normalize \u53ef\u4ee5\u5408\u5e76\u6210\u4e00\u6b21\u4e58\u51cf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u3001\u53ef\u590d\u73b0\u7684\u968f\u673a\u51b3\u7b56<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">FusedNormalization \u91cc\u7684\u968f\u673a\u6027\u6765\u81ea\u4e24\u4e2a\u5730\u65b9\uff1a\u6c34\u5e73\u7ffb\u8f6c\u548c\u968f\u673a\u64e6\u9664\u3002\u4e24\u8005\u90fd\u901a\u8fc7 <code>Generator<\/code> \u83b7\u53d6 Philox \u8ba1\u6570\u5668\u968f\u673a\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=\"\">\/\/ src\/data\/fused_normalization.cpp\nfloat FusedNormalization::uniform(float min_val, float max_val, Generator* rng) const {\n    uint64_t offset = rng->next_offset(1);\n    float rand_val = detail::philox_uniform_float(rng->seed(), offset);\n    return min_val + rand_val * (max_val - min_val);\n}\n\nint FusedNormalization::randint(int min_val, int max_val, Generator* rng) const {\n    return rng->random_int(min_val, max_val);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u662f\u4e00\u79cd\u57fa\u4e8e\u8ba1\u6570\u5668\u7684\u4f2a\u968f\u673a\u6570\u751f\u6210\u5668\uff1a\u7ed9\u5b9a (seed, offset) \u5c31\u80fd\u552f\u4e00\u786e\u5b9a\u4e00\u4e2a\u968f\u673a\u6570\uff0c\u4e0d\u9700\u8981\u7ef4\u62a4\u72b6\u6001\u5411\u91cf\u3002\u8fd9\u610f\u5473\u7740\u5373\u4f7f 200 \u4e2a\u9884\u5904\u7406\u7ebf\u7a0b\u5e76\u53d1\u6267\u884c\uff0c\u53ea\u8981 seed \u548c offset \u5e8f\u5217\u662f\u786e\u5b9a\u7684\uff0c\u6700\u7ec8\u7ed3\u679c\u5c31\u662f\u53ef\u590d\u73b0\u7684\u3002\u6211\u4eec\u540e\u9762\u7b2c 22 \u7bc7\u4f1a\u4e13\u95e8\u8bb2 Philox \u548c\u786e\u5b9a\u6027\u8bad\u7ec3\uff0c\u8fd9\u91cc\u53ea\u9700\u8981\u77e5\u9053\uff1a<strong>\u878d\u5408\u6ca1\u6709\u727a\u7272\u53ef\u590d\u73b0\u6027<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e00\u3001\u4e0e PreprocessWorker \u548c TransferStation \u7684\u8854\u63a5<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>PreprocessWorker<\/code> \u5728\u8fd0\u884c PO \u94fe\u65f6\uff0c\u4f1a\u628a FusedNormalization \u4f5c\u4e3a\u6700\u540e\u4e00\u4e2a\u64cd\u4f5c\uff0c\u8f93\u51fa\u76f4\u63a5\u5199\u5230 <code>TransferStation<\/code> \u7684\u9501\u9875\u5185\u5b58\u69fd\u4f4d\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\/data\/preprocess_worker.cpp\nif (op_id == num_ops - 1) {  \/\/ \u6700\u540e\u4e00\u4e2a\u64cd\u4f5c(FusedNormalization)\uff0c\u8f93\u51fa\u5230\u6700\u7ec8\u4f4d\u7f6e\n    dest_ptr = final_output_ptr;\n    output_stride = 0;  \/\/ \u89e6\u53d1\u81ea\u52a8\u8ba1\u7b97\uff08\u7d27\u51d1\u5e03\u5c40\uff09\n    compact_layout = true;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>final_output_ptr<\/code> \u6765\u81ea <code>request_transfer_station_slot()<\/code>\uff0c\u6307\u5411\u7684\u5c31\u662f H2D \u5f02\u6b65\u62f7\u8d1d\u7684\u6e90\u5185\u5b58\u3002\u56e0\u6b64 FusedNormalization \u7684\u8f93\u51fa layout \u5fc5\u987b\u4e0e TransferStation \u7684\u671f\u671b\u5b8c\u5168\u4e00\u81f4\u3002<code>calculate_stride()<\/code> \u5728\u8fd9\u91cc\u8d77\u5230\u4e86\u5173\u952e\u4f5c\u7528\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\/data\/fused_normalization.cpp\nsize_t FusedNormalization::calculate_stride() {\n    if (use_amp_) {\n        compact_output_stride_ = static_cast&lt;size_t>(output_size_) * 4 * sizeof(uint16_t);\n    } else {\n        compact_output_stride_ = static_cast&lt;size_t>(output_size_)\n                               * static_cast&lt;size_t>(num_channels_) * sizeof(float);\n    }\n    output_stride_ = compact_output_stride_;\n    return output_stride_;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 FusedNormalization \u62d2\u7edd\u5916\u90e8\u4f20\u5165 <code>output_stride<\/code>\u2014\u2014\u5b83\u7684 stride \u7531\u6570\u636e\u7c7b\u578b\u4e25\u683c\u51b3\u5b9a\uff0c\u4e0d\u5141\u8bb8\u8c03\u7528\u65b9\u7528 uint8 \u7684\u601d\u7ef4\u4f20\u5165\u4e00\u4e2a\u9519\u8bef\u7684\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\uff0c\u6b64\u5904 <code>output_stride_<\/code> \u662f\u4e3a\u5e95\u5c42 <code>TransferStation<\/code> \u5f02\u6b65\u62f7\u8d1d\u51c6\u5907\u7684<strong>\u884c\u8de8\u5ea6\u5b57\u8282\u6570\uff08bytes per row\uff09<\/strong>\uff0c\u8fd9\u533a\u522b\u4e8e <code>DTensor<\/code> \u7c7b\u4e2d\u7528\u4e8e\u903b\u8f91\u7d22\u5f15\u7684\u4ee5<strong>\u5143\u7d20\u4e2a\u6570<\/strong>\u4e3a\u5355\u4f4d\u7684 stride\u3002\u8fd9\u79cd\u89e3\u8026\u8ba9\u4f20\u8f93\u5c42\u53ea\u9700\u5173\u5fc3\u7269\u7406\u5185\u5b58\u8fb9\u754c\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e8c\u3001\u968f\u673a\u64e6\u9664\u7684\u5b9e\u73b0<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">RandomErasing \u5728\u8bad\u7ec3\u65f6\u6309\u6982\u7387\u628a\u56fe\u7247\u67d0\u4e2a\u77e9\u5f62\u533a\u57df\u6e05\u96f6\u3002FusedNormalization \u91cc\u7684\u5b9e\u73b0\u5206\u4e24\u6b65\uff1a\u5148\u6309\u5bf9\u6570\u5747\u5300\u5206\u5e03\u91c7\u6837\u64e6\u9664\u533a\u57df\u7684\u9ad8\u548c\u5bbd\uff0c\u518d\u5728\u6574\u4e2a\u50cf\u7d20\u904d\u5386\u7ed3\u675f\u540e\u5bf9\u8f93\u51fa\u505a memset\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\/data\/fused_normalization.cpp\nFusedNormalization::EraseRect FusedNormalization::generate_erase_rect(int H, int W, Generator* rng) const {\n    if (!erase_enabled_ || erase_p_ &lt;= 0.0f || uniform(0.0f, 1.0f, rng) >= erase_p_) {\n        return {};\n    }\n    const float img_area = static_cast&lt;float>(H) * static_cast&lt;float>(W);\n    for (int attempt = 0; attempt &lt; 10; ++attempt) {\n        float target_area_ratio = erase_scale_min_ + uniform(0.0f, 1.0f, rng)\n                                  * (erase_scale_max_ - erase_scale_min_);\n        float erase_area = target_area_ratio * img_area;\n\n        float log_aspect_ratio = log_erase_ratio_min_\n                               + uniform(0.0f, 1.0f, rng)\n                               * (log_erase_ratio_max_ - log_erase_ratio_min_);\n        float aspect_ratio = std::exp(log_aspect_ratio);\n\n        int eh = static_cast&lt;int>(std::round(std::sqrt(erase_area * aspect_ratio)));\n        int ew = static_cast&lt;int>(std::round(std::sqrt(erase_area \/ aspect_ratio)));\n\n        if (eh > 0 &amp;&amp; ew > 0 &amp;&amp; eh &lt; H &amp;&amp; ew &lt; W) {\n            int i = randint(0, H - eh, rng);\n            int j = randint(0, W - ew, rng);\n            return {i, j, eh, ew, true};\n        }\n    }\n    return {};\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u91c7\u6837\u903b\u8f91\u4e0e torchvision \u7684 <code>RandomErasing<\/code> \u4fdd\u6301\u4e00\u81f4\uff1a\u9762\u79ef\u6bd4\u4f8b\u5728 <code>[0.02, 0.33]<\/code> \u4e4b\u95f4\u5747\u5300\u91c7\u6837\uff0c\u5bbd\u9ad8\u6bd4\u5728\u5bf9\u6570\u7a7a\u95f4 <code>[0.3, 3.3]<\/code> \u4e4b\u95f4\u5747\u5300\u91c7\u6837\uff0c\u6700\u591a\u5c1d\u8bd5 10 \u6b21\u627e\u5230\u5408\u6cd5\u533a\u57df\u3002\u7136\u540e <code>apply_erase_fp32<\/code> \u6216 <code>apply_erase_fp16<\/code> \u628a\u5bf9\u5e94\u533a\u57df\u6e05\u96f6\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\/data\/fused_normalization.cpp\nvoid FusedNormalization::apply_erase_fp32(float* data, int W, int C, const EraseRect&amp; rect) const {\n    if (!rect.enabled) return;\n    for (int y = 0; y &lt; rect.eh; ++y) {\n        float* row_start = data + (static_cast&lt;size_t>(rect.i + y) * W + rect.j) * C;\n        std::memset(row_start, 0, static_cast&lt;size_t>(rect.ew) * C * sizeof(float));\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u64e6\u9664\u662f\u5728 normalize \u4e4b\u540e\u8fdb\u884c\u7684\uff0c\u6240\u4ee5\u64e6\u9664\u503c\u662f 0\uff08\u5373 normalize \u540e\u7684&#8221;\u96f6\u503c&#8221;\uff09\uff0c\u4e0e PyTorch \u884c\u4e3a\u4e00\u81f4\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e09\u3001\u6570\u503c\u6b63\u786e\u6027\u9a8c\u8bc1<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u4fdd\u8bc1 FP32 \u548c FP16 \u4e24\u6761\u8def\u5f84\u7684\u6570\u503c\u4e00\u81f4\u6027\uff0c\u6846\u67b6\u91cc\u6709\u4e00\u4e2a\u4e13\u95e8\u7684\u77eb\u6b63\u6d4b\u8bd5 <code>tests\/correction\/test_fused_normalization.cpp<\/code>\u3002\u5b83\u5bf9 MNIST \u5927\u5c0f\u7684 28\u00d728 \u7070\u5ea6\u56fe\u5206\u522b\u8dd1 FP32 \u548c AMP \u7248\u672c\uff0c\u7136\u540e\u8ba1\u7b97 MSE \u548c\u6700\u5927\u76f8\u5bf9\u8bef\u5dee\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=\"\">\/\/ tests\/correction\/test_fused_normalization.cpp\ndouble max_relative_error = 0.0;\nint relative_error_count = 0;\nconst double epsilon = 1e-6f;\n\nfor (int i = 0; i &lt; total_pixels; ++i) {\n    double abs_val = std::abs(static_cast&lt;double>(output_fp32[i]));\n    if (abs_val > epsilon) {\n        double rel_error = std::abs(static_cast&lt;double>(output_fp32[i] - output_amp[i])) \/ abs_val;\n        if (rel_error > max_relative_error) {\n            max_relative_error = rel_error;\n        }\n        relative_error_count++;\n    }\n}\n\nconst double mse_tolerance = 1e-5;\nconst double rel_error_tolerance = 1e-3;\nbool pass = (mse &lt; mse_tolerance) &amp;&amp; (max_relative_error &lt; rel_error_tolerance);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u6d4b\u8bd5\u786e\u4fdd FP16 \u5f15\u5165\u7684\u91cf\u5316\u8bef\u5dee\u5728\u53ef\u63a5\u53d7\u8303\u56f4\u5185\uff0c\u4e0d\u4f1a\u56e0\u4e3a\u878d\u5408\u800c\u727a\u7272\u8bad\u7ec3\u6b63\u786e\u6027\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u56db\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">FusedNormalization \u662f Tech-Renaissance \u6570\u636e\u5f15\u64ce\u91cc&#8221;\u5c11\u5373\u662f\u591a&#8221;\u7684\u5178\u578b\u6848\u4f8b\u3002\u5b83\u6ca1\u6709\u5f15\u5165\u65b0\u7684\u7b97\u6cd5\uff0c\u4e5f\u6ca1\u6709\u6539\u53d8\u6570\u636e\u589e\u5f3a\u7684\u8bed\u4e49\uff0c\u53ea\u662f\u628a PyTorch \u4e2d\u56db\u4e2a\u72ec\u7acb\u7684\u9884\u5904\u7406\u6b65\u9aa4\u5408\u5e76\u6210\u4e86\u4e00\u6b21\u5185\u5b58\u904d\u5386\u3002\u4f46\u6b63\u662f\u8fd9\u6b21\u5408\u5e76\uff0c\u6d88\u9664\u4e86\u4e2d\u95f4\u5f20\u91cf\u7684\u53cd\u590d\u8bfb\u5199\uff0c\u8ba9 CPU \u9884\u5904\u7406\u9636\u6bb5\u7684\u5185\u5b58\u5e26\u5bbd\u5f97\u5230\u4e86\u66f4\u9ad8\u6548\u7684\u5229\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\u8981\u70b9\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>PO \u94fe\u81ea\u52a8\u6ce8\u5165<\/strong>\uff1a\u7528\u6237\u5728 <code>Setup<\/code> \u4e2d\u914d\u7f6e <code>.normalization()<\/code>\u3001<code>RandomHorizontalFlip()<\/code>\u3001<code>RandomErasing()<\/code>\uff0c\u6846\u67b6\u5728 PO \u94fe\u672b\u5c3e\u81ea\u52a8\u6784\u9020 <code>FusedNormalization<\/code>\u3002<\/li>\n\n\n\n<li><strong>\u5355\u6b21\u50cf\u7d20\u904d\u5386\u5b8c\u6210\u6838\u5fc3\u8f6c\u6362<\/strong>\uff1aToTensor\u3001RandomHorizontalFlip\u3001Normalize \u5728\u540c\u4e00\u6b21\u904d\u5386\u4e2d\u5b8c\u6210\uff1bRandomErasing \u4f5c\u4e3a\u540c\u4e00\u6b21 <code>execute()<\/code> \u8c03\u7528\u5185\u7684\u8f7b\u91cf memset \u540e\u7eed\u6b65\u9aa4\u5b8c\u6210\uff0c\u65e0\u9700\u989d\u5916\u7684\u5b8c\u6574\u904d\u5386\u6216\u4e2d\u95f4\u5f20\u91cf\u3002<\/li>\n\n\n\n<li><strong>\u53cc\u7cbe\u5ea6\u8f93\u51fa<\/strong>\uff1aFP32 \u4fdd\u6301\u539f\u59cb\u901a\u9053\u6570\uff0cFP16\/AMP \u56fa\u5b9a 4 \u901a\u9053 padding\uff0c\u76f4\u63a5\u670d\u52a1\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u3002<\/li>\n\n\n\n<li><strong>SSE\/F16C SIMD \u52a0\u901f<\/strong>\uff1aFP16 \u8def\u5f84\u7528 SSE4.1 + F16C \u6307\u4ee4\u4e00\u6b21\u5904\u7406\u591a\u4e2a\u50cf\u7d20\uff0c\u7ffb\u8f6c\u4e5f\u5728 SIMD \u5185\u5b8c\u6210\uff08\u4ee5 <code>__AVX2__<\/code> \u7f16\u8bd1\u5b8f\u4f5c\u4e3a\u542f\u7528\u5f00\u5173\uff0c\u56e0\u4e3a\u652f\u6301 AVX2 \u7684 CPU \u5747\u652f\u6301 F16C\uff09\u3002<\/li>\n\n\n\n<li><strong>\u53ef\u590d\u73b0\u968f\u673a\u6027<\/strong>\uff1aPhilox \u8ba1\u6570\u5668\u968f\u673a\u6570\u4fdd\u8bc1\u9ad8\u5e76\u53d1\u4e0b\u7ed3\u679c\u53ef\u590d\u73b0\u3002<\/li>\n\n\n\n<li><strong>\u6570\u503c\u9a8c\u8bc1<\/strong>\uff1a\u4e13\u95e8\u7684\u77eb\u6b63\u6d4b\u8bd5\u786e\u4fdd FP32 \u4e0e FP16 \u8def\u5f84\u7684\u8bef\u5dee\u5728\u53ef\u63a7\u8303\u56f4\u5185\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u9884\u5904\u7406\u5c42\u9762\u7684\u878d\u5408\u505a\u5b8c\u4e86\uff0c\u6570\u636e\u5df2\u7ecf\u4ee5\u6700\u7d27\u51d1\u7684\u683c\u5f0f\u8eba\u5728 TransferStation \u7684\u9501\u9875\u5185\u5b58\u91cc\uff0c\u7b49\u5f85 H2D \u4f20\u8f93\u3002\u63a5\u4e0b\u6765\u6211\u4eec\u5c31\u8981\u56de\u5230\u6846\u67b6\u7684\u6838\u5fc3\u2014\u2014\u56fe\u7f16\u8bd1\u4e0e\u5185\u5b58\u89c4\u5212\uff0c\u770b\u770b Tech-Renaissance \u662f\u5982\u4f55\u7528\u9759\u6001\u56fe\u548c\u663e\u5b58\u5206\u533a\u628a GPU \u4fa7\u7684\u6548\u7387\u4e5f\u63a8\u5230\u6781\u81f4\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/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\u5341\u4e00 \u524d\u9762\u4e09\u7bc7\u6587\u7ae0\u6211\u4eec\u628a\u6570\u636e\u5f15\u64ce\u8bb2 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":688,"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-487","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\/487","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=487"}],"version-history":[{"count":4,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/487\/revisions"}],"predecessor-version":[{"id":689,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/487\/revisions\/689"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/688"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=487"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=487"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=487"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}