{"id":6511,"date":"2025-11-15T11:28:30","date_gmt":"2025-11-15T03:28:30","guid":{"rendered":"https:\/\/nullthought.net\/?p=6511"},"modified":"2025-11-15T11:49:25","modified_gmt":"2025-11-15T03:49:25","slug":"roofline%ef%bc%9a%e8%a7%a3%e5%86%b3%e4%b8%8d%e5%90%8c%e6%9e%b6%e6%9e%84%e8%be%b9%e7%bc%98%e4%be%a7%ef%bc%88edge%ef%bc%89%e4%b8%8a%e5%a4%9a%e6%a8%a1%e6%80%81-llm%e7%9a%84%e7%bc%96%e8%af%91%e5%92%8c","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6511","title":{"rendered":"Roofline\uff1a\u89e3\u51b3\u4e0d\u540c\u67b6\u6784\u8fb9\u7f18\u4fa7\uff08Edge\uff09\u4e0a\u6a21\u578b\u7684\u7f16\u8bd1\u548c\u90e8\u7f72"},"content":{"rendered":"\n<p><strong><a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">Roofline <\/a><\/strong>\u76ee\u524d\u672c\u8d28\u4e0a\u662f\u4e00\u5bb6 <strong>\u505a\u201c\u4e0b\u4e00\u4ee3 AI \u7f16\u8bd1\u5668 + SDK\u201d\u7684\u8fb9\u7f18\u7b97\u529b\u5e73\u53f0\u516c\u53f8<\/strong>\uff0c\u4ea7\u54c1\u5f62\u6001\u53ef\u4ee5\u7406\u89e3\u4e3a\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-small-font-size\">\u201c\u628a\u4efb\u610f\u6846\u67b6\u91cc\u7684\u6a21\u578b\uff0c\u4e00\u884c Python \u8c03\u7528\uff0c\u7f16\u8bd1\u5e76\u90e8\u7f72\u5230\u5404\u79cd CPU \/ GPU \/ NPU \/ SoC \u4e0a\u7684\u5de5\u5177\u94fe\u548c\u8fd0\u884c\u65f6\u201d\u3002<\/p>\n<\/blockquote>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e00\u3001\u6838\u5fc3\u4ea7\u54c1\uff1aRoofline SDK\uff08\u57fa\u4e8e MLIR \u7684<a href=\"https:\/\/nullthought.net\/?p=3085\" target=\"_blank\" rel=\"noreferrer noopener\">\u8fb9\u7f18 AI<\/a> \u7f16\u8bd1\u5e73\u53f0\uff09<\/h4>\n\n\n\n<p>\u5b98\u65b9\u5728 Arm \u7684\u89e3\u51b3\u65b9\u6848\u624b\u518c\u91cc\uff0c\u628a\u81ea\u5bb6\u4ea7\u54c1\u603b\u4f53\u79f0\u4e3a <strong>Roofline SDK<\/strong>\uff0c\u6838\u5fc3\u662f\u4e00\u4e2a\u201c\u53ef\u91cd\u5b9a\u5411\uff08retargetable\uff09\u7684 AI \u7f16\u8bd1\u5668 + SDK\u201d\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/p>\n\n\n\n<p>\u5b83\u7684\u76ee\u6807\u5f88\u660e\u786e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8ba9 <strong>AI \u7528\u6237 \/ \u5e94\u7528\u5f00\u53d1\u8005<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u53ef\u4ee5\u628a <strong>PyTorch \/ TensorFlow \/ \u5176\u4ed6\u6846\u67b6<\/strong>\u91cc\u7684\u6a21\u578b\uff0c\u5feb\u901f\u8dd1\u5728\u5404\u79cd edge SoC \u4e0a\uff0c\u800c\u4e0d\u7528\u81ea\u5df1\u5199\u4e00\u5806\u7b97\u5b50\u3001\u9a71\u52a8\u548c\u624b\u5de5\u4f18\u5316\u3002<a href=\"https:\/\/community.intel.com\/t5\/Blogs\/Tech-Innovation\/Artificial-Intelligence-AI\/Roofline-AI-s-Role-in-Advancing-Compiler-Technology-with-oneAPI\/post\/1642798?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">community.intel.com+1<\/a><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u8ba9 <strong>\u82af\u7247\u5382 \/ IP \u5382\u5546<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u4e0d\u7528\u4ece\u96f6\u5199\u4e00\u5957\u7f16\u8bd1\u5668\u548c SDK\uff0c\u800c\u662f\u628a\u81ea\u8eab\u5df2\u6709\u4f18\u5316\u96c6\u6210\u8fdb Roofline \u7684 SDK\uff0c\u8ba9\u5ba2\u6237\u201c\u62ff\u5230\u82af\u7247\u5c31\u80fd\u8dd1\u6a21\u578b\u201d\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5728\u4ea7\u54c1\u5f62\u6001\u4e0a\uff0cRoofline SDK \u4e3b\u8981\u7531\u4e24\u5757\u7ec4\u6210\uff08\u8fd9\u662f\u4f60\u53ef\u4ee5\u7406\u89e3\u4e3a\u4e24\u4e2a\u201c\u5b50\u4ea7\u54c1\u201d\uff09<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net<\/a>\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Attic\uff1aAI \u5e94\u7528\u5c42\uff08\u4e00\u7ad9\u5f0f\u5e94\u7528\u5c42\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u8d1f\u8d23\u201c<strong>\u4e0a\u5c42\u6a21\u578b\u548c\u5e94\u7528<\/strong>\u201d\uff1a\n<ul class=\"wp-block-list\">\n<li>\u63a5\u6536\u6765\u81ea <strong>\u4efb\u610f\u6846\u67b6<\/strong> \u7684\u7f51\u7edc\uff08\u6bd4\u5982 PyTorch \/ TF \/ ONNX\uff09\u3002<\/li>\n\n\n\n<li>\u505a\u6a21\u578b\u56fe\u5c42\u9762\u7684\u5904\u7406\u3001\u7ba1\u7406\u3001\u7248\u672c\u63a7\u5236\u7b49\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u9762\u5411\u7684\u662f\uff1a<strong>\u505a\u5e94\u7528 \/ \u505a\u4ea7\u54c1\u7684\u5f00\u53d1\u56e2\u961f<\/strong>\uff0c\u4ed6\u4eec\u5173\u5fc3\u7684\u662f\u201c\u8fd9\u4e2a\u6a21\u578b\u600e\u4e48\u8dd1\u8d77\u6765\u3001\u600e\u4e48\u96c6\u6210\u5230\u4e1a\u52a1\u91cc\u201d\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>AI \u786c\u4ef6\u5c42\uff08\u4e00\u7ad9\u5f0f\u786c\u4ef6\u5c42\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u8d1f\u8d23\u201c<strong>\u4e0b\u5c42\u786c\u4ef6\u548c\u7b97\u529b\u8d44\u6e90<\/strong>\u201d\uff1a\n<ul class=\"wp-block-list\">\n<li>\u7edf\u4e00\u96c6\u6210 <strong>CPU\u3001GPU\u3001NPU\u3001DSP<\/strong> \u7b49\u5404\u79cd\u7b97\u529b\u5355\u5143\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/li>\n\n\n\n<li>\u5728 Arm \u751f\u6001\u91cc\u91cd\u70b9\u652f\u6301 <strong>Cortex-A \/ Cortex-M \/ Cortex-R \/ Ethos NPU \/ Mali GPU \/ Mali ISP \u7b49 IP<\/strong>\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u9762\u5411\u7684\u662f\uff1a<strong>\u82af\u7247\u5382 \/ \u6a21\u7ec4\u5382 \/ \u786c\u4ef6\u5e73\u53f0\u63d0\u4f9b\u65b9<\/strong>\uff0c\u628a\u786c\u4ef6\u7279\u5b9a\u4f18\u5316\u5d4c\u8fdb\u8fd9\u4e2a\u5c42\uff0c\u6700\u7ec8\u5bf9\u4e0a\u5c42\u5e94\u7528\u906e\u853d\u590d\u6742\u5ea6\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u4e24\u5757\u52a0\u5728\u4e00\u8d77\uff0c\u5b98\u65b9\u5b9a\u4f4d\u4e3a\uff1a <strong>\u201c\u552f\u4e00\u4e00\u4e2a\u53ef\u4ee5\u5728\u8fb9\u7f18\u4fa7\u5bf9\u591a\u6a21\u6001 LLM \u505a\u5168\u7f16\u8bd1\uff08fully compiled\uff09\u90e8\u7f72\u3001\u5e76\u80fd\u5bf9\u6700\u65b0\u6a21\u578b\u63d0\u4f9b day-0 \u652f\u6301\u7684\u65b9\u6848\u201d<\/strong>\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e8c\u3001\u5e95\u5c42\u6280\u672f\uff1a\u4e0b\u4e00\u4ee3 AI \u7f16\u8bd1\u5668\uff08MLIR \u751f\u6001 + IREE \/ Torch-MLIR\uff09<\/h4>\n\n\n\n<p>\u4ece\u5b98\u7f51\u548c\u6280\u672f\u8d44\u6599\u6765\u770b\uff0cRoofline \u7684\u6838\u5fc3\u662f\u4e00\u4e2a <strong>\u57fa\u4e8e MLIR\uff08Multi-Level Intermediate Representation\uff0c\u591a\u5c42\u4e2d\u95f4\u8868\u793a\uff09\u7684 AI \u7f16\u8bd1\u5668\u6808<\/strong>\uff1a<a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai+2GitHub+2<\/a><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GitHub \u4e0a\u4ed6\u4eec\u5927\u91cf fork \/ \u53c2\u4e0e\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>iree<\/strong>\uff1aMLIR-based \u673a\u5668\u5b66\u4e60\u7f16\u8bd1\u5668\u548c runtime \u5de5\u5177\u5305\uff1b<\/li>\n\n\n\n<li><strong>torch-mlir<\/strong>\uff1a\u628a PyTorch \u6a21\u578b\u5bfc\u5165\u5230 MLIR \u751f\u6001\uff1b<\/li>\n\n\n\n<li><strong>stablehlo \/ llvm-project<\/strong> \u7b49\u3002<a href=\"https:\/\/github.com\/RooflineAI?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">GitHub<\/a><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5728 Intel oneAPI \/ Level Zero \u76f8\u5173\u7684\u5206\u4eab\u4e2d\uff0c\u4ed6\u4eec\u5c55\u793a\u4e86\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5982\u4f55\u628a <strong>PyTorch \/ TensorFlow<\/strong> \u6a21\u578b\u7f16\u8bd1\u6210 <strong>SPIR-V<\/strong> \u7b49\u4e2d\u95f4\u8868\u793a\uff0c\u4ee5\u9002\u914d\u4e0d\u540c\u786c\u4ef6\u540e\u7aef\u3002<a href=\"https:\/\/community.intel.com\/t5\/Blogs\/Tech-Innovation\/Artificial-Intelligence-AI\/Roofline-AI-s-Role-in-Advancing-Compiler-Technology-with-oneAPI\/post\/1642798?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">community.intel.com+1<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u4f60\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u4e3a\u4e00\u4e2a <strong>\u9762\u5411\u8fb9\u7f18\u8bbe\u5907\u7684\u201c\u4e13\u4e1a\u7248 TVM \/ XLA + \u4e00\u7ad9\u5f0f SDK\u201d<\/strong>\uff0c\u4f46\u7279\u522b\u5f3a\u8c03\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u53ef\u91cd\u5b9a\u5411<\/strong>\uff1a\u6362\u4e00\u9897 SoC \u6216 NPU\uff0c\u4e0d\u7528\u91cd\u6765\u4e00\u5957\uff0c\u4e3b\u8981\u505a\u914d\u7f6e\u548c\u96c6\u6210\u5de5\u4f5c\uff1b<\/li>\n\n\n\n<li><strong>\u591a\u8bbe\u5907\u5f02\u6784<\/strong>\uff1a\u540c\u4e00 SoC \u4e0a CPU + GPU + NPU \u7684\u4efb\u52a1\u5212\u5206\u548c\u8c03\u5ea6\u4e5f\u8003\u8651\u5728\u5185\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e09\u3001\u4ea7\u54c1\u4f7f\u7528\u6d41\u7a0b\uff1a\u4ece\u201c\u4efb\u610f\u6a21\u578b\u201d\u5230\u201cEdge \u8bbe\u5907\u56fa\u4ef6\u201d<\/h4>\n\n\n\n<p>\u5b98\u7f51\u7684 \u201cHow it works\u201d \u628a\u6d41\u7a0b\u62c6\u6210 6 \u6b65\uff0c\u672c\u8d28\u5c31\u662f\u4e00\u4e2a <strong>\u5b8c\u6574\u7684\u7f16\u8bd1\u90e8\u7f72\u7ba1\u7ebf<\/strong>\uff1a<a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai+1<\/a><\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u52a0\u8f7d\u8bad\u7ec3\u597d\u7684\u6a21\u578b\uff08Load trained AI model\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u63a5\u6536\u6765\u81ea PyTorch \/ TF \/ \u5176\u4ed6\u6846\u67b6\u7684\u5df2\u8bad\u7ec3\u7f51\u7edc\u3002<\/li>\n\n\n\n<li>\u7528\u6237\u89c6\u89d2\uff1a<strong>\u7528\u4e00\u4e2a Python API \u628a\u6a21\u578b\u201c\u4ea4\u7ed9 Roofline\u201d<\/strong>\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u4f18\u5316\u4e0e\u91cf\u5316\uff08Optimization &amp; quantization\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u81ea\u52a8\u6267\u884c\u56fe\u4f18\u5316\u3001\u7b97\u5b50\u878d\u5408\u3001\u5e38\u91cf\u6298\u53e0\u7b49\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u591a\u79cd\u7cbe\u5ea6\uff1aFP32 \/ FP16 \/ INT8 \u7b49\uff0c\u9762\u5411 edge \u8bbe\u5907\u505a\u5185\u5b58\u5360\u7528\u548c\u65f6\u5ef6\u4f18\u5316\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u6570\u5b66\u62bd\u8c61\uff08Mathematical abstraction\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u628a\u7f51\u7edc\u8868\u793a\u62bd\u8c61\u4e3a\u7edf\u4e00\u7684\u4e2d\u95f4\u8868\u793a\uff08MLIR IR \/ StableHLO \u7b49\uff09\uff0c<\/li>\n\n\n\n<li>\u4fbf\u4e8e\u5728\u4e0d\u540c\u786c\u4ef6\u95f4\u91cd\u7528\u540c\u4e00\u5957\u4e0a\u5c42\u4f18\u5316\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u9488\u5bf9\u5177\u4f53\u786c\u4ef6\u7684\u4ee3\u7801\u751f\u6210\uff08Code generation for specific hardware\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u751f\u6210\u5bf9\u5e94\u540e\u7aef\u7684\u4ee3\u7801\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bd4\u5982\u9488\u5bf9 CPU \u751f\u6210 LLVM IR \/ \u673a\u5668\u7801\uff1b<\/li>\n\n\n\n<li>\u9488\u5bf9 GPU\/\u52a0\u901f\u5668\u751f\u6210 SPIR-V \/ nvvm \/ rocdl \u7b49\u3002<a href=\"https:\/\/dvcon-proceedings.org\/wp-content\/uploads\/134-Compiling-AI-Workloads-for-On-Device-Inference-on-Heterogeneous-Systems-using-MLIR.pdf?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">dvcon-proceedings.org<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u6807\u51c6\u5316\u7684\u8f6f\u786c\u4ef6\u63a5\u53e3\uff08Standardized HW-SW interfaces\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u63d0\u4f9b\u7edf\u4e00\u7684\u8c03\u7528\u63a5\u53e3\u3001\u7f13\u51b2\u7ba1\u7406\u3001\u6267\u884c\u8c03\u5ea6\u7b49\u3002<\/li>\n\n\n\n<li>\u5bf9\u4e1a\u52a1\u4fa7\u5f00\u53d1\u8005\u6765\u8bf4\uff0c<strong>\u770b\u5230\u7684\u662f\u4e00\u4e2a\u5e72\u51c0\u7684 API\uff0c\u800c\u4e0d\u662f\u5404\u5bb6 NPU \u7684\u788e\u7247\u5316 SDK<\/strong>\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u56fa\u4ef6\u4e0e\u8fd0\u884c\u65f6\uff08Firmware and runtime\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u6253\u5305\u4e3a\u53ef\u5728\u76ee\u6807\u8bbe\u5907\u4e0a\u76f4\u63a5\u8fd0\u884c\u7684 runtime \/ \u56fa\u4ef6\u6620\u50cf\u3002<\/li>\n\n\n\n<li>\u5728\u771f\u5b9e\u6d4b\u8bd5\u4e2d\uff0c\u4ed6\u4eec\u5728 <strong>Pixel 8 Pro\u3001Raspberry Pi 5<\/strong> \u7b49\u8bbe\u5907\u4e0a\u5b9e\u6d4b\u8fc7\u591a\u4e2a CNN \/ Transformer \u5de5\u4f5c\u8d1f\u8f7d\u3002<a href=\"https:\/\/dvcon-proceedings.org\/wp-content\/uploads\/134-Compiling-AI-Workloads-for-On-Device-Inference-on-Heterogeneous-Systems-using-MLIR.pdf?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">dvcon-proceedings.org+1<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>\u5b98\u7f51\u5f3a\u8c03\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-small-font-size\">\u201c\u6211\u4eec\u8ba9\u90e8\u7f72\u5c31\u50cf\u4e00\u6b21 Python \u8c03\u7528\u90a3\u4e48\u7b80\u5355\uff08We make deployment as easy as a Python call\uff09\u201d<a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai<\/a><\/p>\n<\/blockquote>\n\n\n\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c<strong>\u5bf9\u5e94\u7528\u5f00\u53d1\u8005\uff0c\u4ea7\u54c1\u5448\u73b0\u4e3a\u4e00\u4e2a Python SDK\uff1b\u5bf9\u786c\u4ef6\u5382\u5546\uff0c\u4ea7\u54c1\u5448\u73b0\u4e3a\u4e00\u5957\u53ef\u96c6\u6210\u7684\u7f16\u8bd1 + \u8fd0\u884c\u65f6\u6808<\/strong>\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u56db\u3001\u6027\u80fd\u4e0e\u4f18\u52bf\uff1a\u7075\u6d3b\u3001\u6548\u7387\u9ad8\u3001\u5bf9\u5f00\u53d1\u8005\u53cb\u597d<\/h4>\n\n\n\n<p>\u5b98\u7f51 &amp; \u6280\u672f\u8d44\u6599\u91cc\u7ed9\u51fa\u7684\u51e0\u4e2a\u5173\u952e\u5356\u70b9\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Flexible \u2013 \u4efb\u610f\u6a21\u578b \u00d7 \u4efb\u610f\u786c\u4ef6<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u201cDeploy any model from any framework on different target hardware.\u201d <a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai+1<\/a><\/li>\n\n\n\n<li>\u5bf9\u5e94\u7528\u56e2\u961f\uff1a\u4e0d\u7528\u88ab\u5355\u4e00\u6846\u67b6\u6216\u786c\u4ef6\u6808\u9501\u6b7b\uff1b<\/li>\n\n\n\n<li>\u5bf9\u82af\u7247\u5382\uff1a\u53ea\u8981\u63a5\u597d\u540e\u7aef\uff0c\u5c31\u80fd\u627f\u63a5\u6574\u4e2a\u4e0a\u5c42\u751f\u6001\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Efficient \u2013 \u663e\u8457\u964d\u4f4e\u5185\u5b58\u548c\u65f6\u5ef6<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u5b98\u7f51\u9875\u9762\u63d0\u5230\uff1a<strong>\u53ef\u505a\u5230 4\u00d7 \u66f4\u4f4e\u7684\u5185\u5b58\u5360\u7528\u30011.5\u00d7 \u66f4\u4f4e\u7684\u65f6\u5ef6<\/strong>\uff0c\u4e14\u5728\u7cbe\u5ea6\u4e0d\u635f\u5931\u7684\u524d\u63d0\u4e0b\u3002<a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai<\/a><\/li>\n\n\n\n<li>\u5728 Arm \u8054\u5408\u65b9\u6848\u6587\u6863\u4e2d\uff0c\u5bf9\u67d0\u4e9b\u5bf9\u6bd4\u57fa\u7ebf\uff08\u6bd4\u5982 TorchInductor \/ \u4f20\u7edf\u65b9\u6848\uff09\uff0c\u7ed9\u51fa\u7684\u6570\u636e\u751a\u81f3\u662f\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u6700\u9ad8 35\u00d7 \u66f4\u597d\u7684\u5185\u5b58\u4f7f\u7528\uff0c15\u00d7 \u5ef6\u8fdf\u6539\u5584<\/strong>\uff08\u5177\u4f53\u53d6\u51b3\u4e8e\u6a21\u578b\u548c\u5e73\u53f0\uff09\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+1<\/a><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>DVCon \u8bba\u6587\u91cc\u7684\u5b9e\u9a8c\u663e\u793a\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 Pixel 8 Pro\uff08\u5bf9\u6bd4 TFLite\uff09\u548c Raspberry Pi 5\uff08\u5bf9\u6bd4 TorchInductor\uff09\u4e0a\uff0c\u5bf9 CNN + Transformer \u4efb\u52a1\uff0c\u4ed6\u4eec\u5728\u901f\u5ea6\u548c\u5185\u5b58\u4e0a\u90fd\u660e\u663e\u5360\u4f18\u3002<a href=\"https:\/\/dvcon-proceedings.org\/wp-content\/uploads\/134-Compiling-AI-Workloads-for-On-Device-Inference-on-Heterogeneous-Systems-using-MLIR.pdf?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">dvcon-proceedings.org<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Easy \u2013 \u5bf9\u5f00\u53d1\u8005\u7684\u201c\u5168\u6258\u7ba1\u201d\u4f53\u9a8c<\/strong>\n<ul class=\"wp-block-list\">\n<li>SDK \u8986\u76d6\u4ece <strong>\u91cf\u5316\u3001\u8c03\u4f18\u3001\u8c03\u8bd5\u5230\u90e8\u7f72<\/strong> \u7684\u5b8c\u6574\u6d41\u7a0b\uff0c\u4f60\u53ef\u4ee5\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8981\u4e48\u201c\u6253\u5f00\u6240\u6709\u7ec6\u8282\u201d\u505a\u624b\u52a8\u8c03\u4f18\uff1b<\/li>\n\n\n\n<li>\u8981\u4e48\u201c\u5b8c\u5168\u4ea4\u7ed9 Roofline \u81ea\u52a8\u5904\u7406\u201d\u3002<a href=\"https:\/\/www.roofline.ai\/\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai+1<\/a><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u7279\u522b\u5f3a\u8c03\u5bf9 <strong>\u6700\u65b0 GenAI \/ LLM \/ \u591a\u6a21\u6001\u6a21\u578b\u7684 day-0 \u652f\u6301<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bd4\u5982 DeepSeek R1 Distill Qwen 1.5B\uff0c\u4ed6\u4eec\u5728\u535a\u5ba2\u4e2d\u5c55\u793a\u4e86\uff1a\n<ul class=\"wp-block-list\">\n<li>\u901a\u8fc7 Roofline pipeline \u201c\u5f00\u7bb1\u5373\u8dd1\u201d\uff0c<\/li>\n\n\n\n<li>tokens\/s \u6027\u80fd\u63a5\u8fd1 llama.cpp\uff0c\u540c\u65f6\u65f6\u5185\u5b58\u5360\u7528\u6bd4 TorchInductor \u4f4e\u7ea6 4\u00d7\u3002<a href=\"https:\/\/www.roofline.ai\/news?5759fe65_page=2&amp;utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">roofline.ai+1<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e94\u3001\u9762\u5411\u7684\u5ba2\u6237 &amp; \u5178\u578b\u573a\u666f<\/h4>\n\n\n\n<p>\u4ece\u516c\u5f00\u4fe1\u606f\u6765\u770b\uff0cRoofline \u7684\u4ea7\u54c1\u4e3b\u8981\u670d\u52a1\u4e8e\u4e09\u7c7b\u7528\u6237\uff1a<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+2 roofline.ai+2<\/a><\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u505a\u8fb9\u7f18 AI \u4ea7\u54c1\u7684\u8bbe\u5907\u5382 \/ \u89e3\u51b3\u65b9\u6848\u5546<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u6bd4\u5982\u667a\u80fd\u6444\u50cf\u5934\u3001\u5de5\u4e1a\u7ec8\u7aef\u3001\u673a\u5668\u4eba\u3001\u8f66\u8f7d\u6444\u50cf\u5934\u3001\u53ef\u7a7f\u6234\u8bbe\u5907\u7b49\uff1b<\/li>\n\n\n\n<li>\u5e0c\u671b\uff1a<strong>\u7528\u6700\u65b0\u7684 LLM \/ \u591a\u6a21\u6001\u6a21\u578b\uff0c\u4f46\u786c\u4ef6\u8d44\u6e90\u6709\u9650<\/strong>\uff08\u529f\u8017 \/ \u5185\u5b58 \/ \u5e26\u5bbd\u521a\u521a\u591f\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u82af\u7247\u5382 \/ IP \u5382\u5546\uff08\u7279\u522b\u662f Arm \u751f\u6001\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u9700\u8981\u7ed9\u5ba2\u6237\u201c\u73b0\u6210\u80fd\u8dd1 AI \u7684 SDK\u201d\uff0c\u800c\u4e0d\u662f\u53ea\u7ed9\u4e2a\u88f8\u82af\u7247\uff1b<\/li>\n\n\n\n<li>Roofline \u5e2e\u4ed6\u4eec\u628a\uff1a\n<ul class=\"wp-block-list\">\n<li>\u81ea\u5df1\u7684 NPU \/ GPU \/ CPU \u4f18\u5316\uff0c<\/li>\n\n\n\n<li>\u96c6\u6210\u8fdb\u4e0a\u5c42\u901a\u7528\u7f16\u8bd1\u5668\u548c SDK \u91cc\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5b66\u672f \/ \u7814\u7a76\u56e2\u961f<\/strong>\n<ul class=\"wp-block-list\">\n<li>HiPEAC \u62a5\u9053\u91cc\u63d0\u5230\uff1a\u5bf9\u5b66\u672f\u548c\u6559\u5b66\u7528\u9014\uff0c\u6709\u7279\u6b8a\u6388\u6743\u6a21\u5f0f\u3002<a href=\"https:\/\/www.hipeac.net\/news\/7069\/aachen-spinoff-roofline-enables-agile-adaptation-to-new-edge-ai-models-and-hardware\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">HiPEAC<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">\u516d\u3001\u548c\u4f20\u7edf\u65b9\u6848\uff08TFLite\u3001TVM\u3001\u5382\u5546\u81ea\u5e26 SDK\uff09\u7684\u5dee\u5f02<\/h4>\n\n\n\n<p>\u7ed3\u5408\u4ed6\u4eec\u516c\u5f00\u7684\u5b9a\u4f4d\uff0c\u53ef\u4ee5\u7b80\u5355\u5bf9\u6bd4\u4e00\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u548c TFLite \/ ONNX Runtime \u7b49\u901a\u7528\u63a8\u7406\u5f15\u64ce\u76f8\u6bd4<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>Roofline \u66f4\u50cf\u4e00\u4e2a <strong>\u201c\u53ef\u91cd\u5b9a\u5411\u7684\u7f16\u8bd1\u5668\u6808 + SDK\u201d<\/strong>\uff0c\u800c\u4e0d\u662f\u5355\u4e00\u5f15\u64ce\uff1b<\/li>\n\n\n\n<li>\u66f4\u5f3a\u8c03\uff1a\n<ul class=\"wp-block-list\">\n<li>\u9762\u5411 <strong>\u591a SoC\u3001\u591a\u4e2a\u8ba1\u7b97\u5355\u5143\u7684\u5f02\u6784\u8c03\u5ea6<\/strong>\uff1b<\/li>\n\n\n\n<li>\u5bf9 <strong>LLM \/ \u591a\u6a21\u6001\u6a21\u578b<\/strong> \u7684\u652f\u6301\u548c day-0 \u9002\u914d\u80fd\u529b\u3002<a href=\"https:\/\/armkeil.blob.core.windows.net\/developer\/Files\/pdf\/ai-ecosystem-catalogue\/arm-Roofline-AI-Solution-Brief-vUpload.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">armkeil.blob.core.windows.net+2 dvcon-proceedings.org+2<\/a><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u548c TVM \/ XLA \/ IREE \u7b49\u7f16\u8bd1\u5668\u9879\u76ee\u76f8\u6bd4<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>Roofline \u5728\u6280\u672f\u4e0a\u5927\u91cf\u57fa\u4e8e\u8fd9\u4e9b\u5f00\u6e90\u9879\u76ee\uff08MLIR \/ IREE \/ Torch-MLIR \u7b49\uff09\uff0c<a href=\"https:\/\/github.com\/RooflineAI?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">GitHub+1<\/a><\/li>\n\n\n\n<li>\u4ea7\u54c1\u5c42\u9762\u591a\u4e86\u4e00\u5c42\uff1a\n<ul class=\"wp-block-list\">\n<li>\u9762\u5411\u82af\u7247\u5382\u548c\u5e94\u7528\u5f00\u53d1\u8005\u7684 <strong>\u5546\u4e1a SDK\u3001\u7edf\u4e00\u63a5\u53e3\u3001\u5de5\u5177\u94fe\u6574\u5408\u3001\u5546\u4e1a\u652f\u6301 \/ \u96c6\u6210\u670d\u52a1<\/strong>\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e03\u3001\u5c0f\u7ed3\uff1a\u4e00\u53e5\u8bdd\u6982\u62ec Roofline \u7684\u4ea7\u54c1<\/h4>\n\n\n\n<p>\u5982\u679c\u7528\u4e00\u53e5\u504f\u5de5\u7a0b\u5316\u7684\u8bdd\u5e2e\u4f60\u8bb0\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-small-font-size\"><strong>Roofline = \u4e00\u4e2a\u57fa\u4e8e MLIR \u7684\u201c\u8fb9\u7f18 AI \u7f16\u8bd1 + SDK \u5e73\u53f0\u201d\uff0c\u4e0a\u63a5\u4efb\u610f AI \u6846\u67b6\u548c\u6a21\u578b\uff0c\u4e0b\u63a5\u5404\u79cd CPU\/GPU\/NPU\/SoC\uff0c\u901a\u8fc7 Attic \u5e94\u7528\u5c42 + \u786c\u4ef6\u5c42\uff0c\u628a\u590d\u6742\u7684\u7f16\u8bd1\u3001\u91cf\u5316\u3001\u5f02\u6784\u8c03\u5ea6\u548c runtime \u6253\u5305\u6210\u201c\u4e00\u884c Python \u8c03\u7528\u5c31\u80fd\u8dd1\u201d\u7684\u4ea7\u54c1\u3002<\/strong><\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Roofline \u76ee\u524d\u672c\u8d28\u4e0a\u662f\u4e00\u5bb6 \u505a\u201c\u4e0b\u4e00\u4ee3 AI \u7f16\u8bd1\u5668 + SDK\u201d\u7684\u8fb9\u7f18\u7b97\u529b\u5e73\u53f0\u516c\u53f8\uff0c\u4ea7\u54c1\u5f62\u6001\u53ef\u4ee5\u7406\u89e3 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[3,8],"tags":[39,40],"class_list":["post-6511","post","type-post","status-publish","format-standard","hentry","category-it","category-tech","tag-ai","tag-iot"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=3\" rel=\"category\">IT<\/a> <a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"Roofline \u76ee\u524d\u672c\u8d28\u4e0a\u662f\u4e00\u5bb6 \u505a\u201c\u4e0b\u4e00\u4ee3 AI \u7f16\u8bd1\u5668 + SDK\u201d\u7684\u8fb9\u7f18\u7b97\u529b\u5e73\u53f0\u516c\u53f8\uff0c\u4ea7\u54c1\u5f62\u6001\u53ef\u4ee5\u7406\u89e3&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6511","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6511"}],"version-history":[{"count":3,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6511\/revisions"}],"predecessor-version":[{"id":6516,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6511\/revisions\/6516"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6511"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6511"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6511"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}