[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-91":104},{"enabled":4,"title":5,"statusText":5,"buttonText":5,"campSlug":5,"stage":6,"classroomId":7,"hasClassroom":4},false,"",0,null,{"total":9,"list":10,"pageNum":44,"pageSize":45,"size":9,"startRow":44,"endRow":9,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":48,"navigateFirstPage":44,"navigateLastPage":44},3,[11,25,35],{"slug":12,"title":13,"logo":7,"liveUrl":14,"prize":15,"sponsor":16,"description":17,"organizer":18,"startTime":19,"endTime":20,"registrationDeadline":21,"status":22,"dashboardVisibility":23,"createdAt":24},"spring2026","2026春季启元人工智能大赛","https:\u002F\u002Fwww.bilibili.com\u002Fblackboard\u002Flive\u002Flive-activity-player.html?cid=1822703816&quality=0","200,000","麒麟软件有限公司","“2026春季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。\n各参赛选手需在规定的比赛周期内，依据赛事规则选定赛题并完成相应挑战，优胜者将共同角逐总额20万元的赛事奖金池，并且还将有机会获得实习和知名企业优质岗位内推资格。","启元实验室","2026-04-19 16:00:00","2026-07-13 16:00:00","2026-06-16 16:00:00","ongoing","private","2026-04-16 08:00:00",{"slug":26,"title":27,"logo":7,"liveUrl":14,"prize":28,"sponsor":7,"description":29,"organizer":18,"startTime":30,"endTime":31,"registrationDeadline":32,"status":33,"dashboardVisibility":23,"createdAt":34},"autumn2025","2025秋季启元人工智能大赛","150,000","“2025秋季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。","2025-11-08 08:00:00","2026-01-13 08:00:00","2025-12-30 08:00:00","ended","2025-11-03 08:00:00",{"slug":36,"title":37,"logo":7,"liveUrl":14,"prize":38,"sponsor":7,"description":39,"organizer":18,"startTime":40,"endTime":41,"registrationDeadline":42,"status":33,"dashboardVisibility":23,"createdAt":43},"summer2025","启元人工智能大赛(系统方向)","80,000","启元人工智能大赛由启元实验室主办，其中系统赛道基于启元实验室九源软件栈的相关开源项目，通过拉动开源社区力量，一方面促进国产智能计算生态的发展，另一方面也为人工智能系统方向的学生及从业者提供一个实践平台。","2025-07-08 08:00:00","2025-08-25 08:00:00","2025-07-22 08:00:00","2025-06-22 08:00:00",1,5,true,8,[44],{"total":45,"list":50,"pageNum":44,"pageSize":45,"size":45,"startRow":44,"endRow":45,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":103,"navigateFirstPage":44,"navigateLastPage":44},[51,62,73,84,94],{"id":52,"slug":53,"title":54,"logo":55,"publicLiveEmbedUrl":14,"description":56,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":16,"startTime":58,"endTime":59,"status":22,"dashboardVisibility":60,"createdAt":58,"updatedAt":61},"019eeeff-6b00-7d55-a629-efafded3e232","summer2026","2026年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2026-06\u002F019eef7f-7816-7828-9309-79c2c61dd568.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=U5uEBYHvOr5t1PntxcixNNUXfd0%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。","...","2026-06-22 08:00:00","2026-10-19 08:00:00","public","2026-10-12 08:00:00",{"id":63,"slug":64,"title":65,"logo":66,"publicLiveEmbedUrl":7,"description":67,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":68,"startTime":69,"endTime":70,"status":33,"dashboardVisibility":23,"createdAt":71,"updatedAt":72},"019af3f3-e06a-7245-aa8d-37dc99a69c73","winter2025","2025年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-12\u002F019afc69-b7ca-7805-91fb-605429671c01.png?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IMYJid%2Fjp0BG%2FDLv%2BBmsvMXo4Q8%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。 \n","新华三技术有限公司","2025-12-25 08:00:00","2026-03-27 08:00:00","2025-12-06 08:00:00","2026-10-22 20:24:30",{"id":74,"slug":36,"title":75,"logo":76,"publicLiveEmbedUrl":7,"description":77,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":78,"startTime":79,"endTime":80,"status":81,"dashboardVisibility":23,"createdAt":82,"updatedAt":83},"01980342-208c-7803-b3f3-19e1d31960ba","2025年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-07\u002F0198045e-28fe-78f0-95fe-23835ce58395.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BAb9SulTr2AaX%2FcpwGmSPY%2BghM0%3D","InfiniTensor大模型与人工智能系统训练营由启元实验室及清华大学 AI 系统方向团队发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营联合开源项目开发者、学术界大牛、工业界资深工程师参与授课与项目指导。课程采用「线上直播授课 + 在线答疑 + 项目实践辅导 」的方式进行，确保每位同学都能获得充分成长。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","北京积算科技有限公司","2025-07-14 08:00:00","2025-09-30 08:00:00","upcoming","2025-07-13 08:00:00","2026-10-22 20:24:58",{"id":85,"slug":86,"title":87,"logo":88,"publicLiveEmbedUrl":7,"description":89,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":90,"endTime":91,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":93},"01980cf9-a7d8-7796-ae86-0bbe05334add","winter2024","2024年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-ce23-7624-a287-9f07dd0d85f6.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。报名学员请阅读“导学”课程页的训练营介绍。","2025-01-05 08:00:00","2025-03-22 08:00:00","2025-07-15 08:00:00","2026-10-22 20:24:39",{"id":95,"slug":96,"title":97,"logo":98,"publicLiveEmbedUrl":7,"description":99,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":100,"endTime":101,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":102},"01980c20-363a-7dc3-8f5a-58a68e08dcc4","summer2024","2024年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-fbfa-788a-afcd-17fa9be74ff9.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区和高校，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过课程、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","2024-07-20 08:00:00","2024-09-29 08:00:00","2026-10-22 20:24:49",[44],{"currentItem":105,"processedData":125},{"img":106,"keywords":107,"treeType":44,"description":108,"showStatus":44,"updateTime":109,"type":110,"title":111,"content":112,"nextArticle":113,"createTime":109,"sortOrder":117,"canonicalPath":118,"id":119,"contentType":120,"prevArticle":121},"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_fm.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=a0olcL%2BS0mW0GeTOQDXvhOcqfKU%3D","AI 编译器,后端优化,AutoTuning,Triton,显存复用","讲解 AI 编译器后端全套优化方案，包含显存生命周期内存池、多层循环变换、FlashAttention、Auto Tuning 自动调优、Triton 与九齿 DSL 代码生成全流程。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 AI 编译器 3：《AI 编译器中的后端优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw01.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j%2F%2FFNlL28%2Bq%2BAxzetQOsNIXCSKo%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">8 月 4 日，2026 夏季 InfiniTensor 训练营 AI 编译器方向第三课《AI 编译器中的后端优化》开讲。\r\n前置课程《AI 编译器中的前端优化》讲解了四类核心等价图变换、子图替换工程实践与 CUDA Graph 底层加速方案等内容。\u003C\u002Fspan>\r\n\r\n本节课将结合计算图算子实现与代码生成流程，系统拆解后端优化的四大代表性方法。\r\n\r\n\r\n## 内存管理优化\r\n\r\n### 优化目标与方向\r\n1. **降低运行时开销**：调整分配\u002F释放时机，减少 `cudaMalloc`\u002F`cudaFree` 等高频调用。\r\n2. **降低显存占用**：优化内存规划，减少峰值显存\r\n\r\n### Naive 版本 vs 优化版本\r\n1.**Naive 版本痛点**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw02.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5yS%2FAVVvPgEhgCd%2BfjHN8RlXhHk%3D)\r\n\r\n- 推理前一次性为所有 Tensor 分配内存，直到结束才释放。\r\n- 大量非活跃 Tensor 持续占用显存，造成浪费。\r\n\r\n2.**优化版本**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw03.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kYO0LHHXqT7yT4t%2F5eoS7w1G8ek%3D)\r\n\r\n- **生命周期分析**：基于拓扑排序模拟执行过程，按 Tensor 实际使用周期规划空间。\r\n- **引用计数机制**：维护后续 OP 引用计数，归零时立即回收空间供复用。\r\n- **Lazy Allocator 实现**：\r\n   - 模拟阶段：使用红黑树维护空闲块，计算 Offset 和 Peak，不调用真实设备接口。\r\n   - 执行阶段：仅申请一次峰值大小的连续内存，通过 `Base + Offset` 获取实际地址。\r\n\r\n3.**实战收益**：以 Inception-v3 为例，优化后支持 Batch Size 从 256 提升至 2048。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw04.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BPnNyJSJ01iNqSkYh4hfyWBcToY%3D)\r\n\r\n> 💡 **核心思想**：静态分析确定生命周期 + 统一内存池实现空间复用。InfiniTensor 与 Refactor Graph 等项目均采用此思路。\r\n\r\n\r\n## 算子层面优化\r\n### 性能瓶颈判断\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw05.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wLbRJ3e06d4N8iSc%2Fk9VJ%2F4OiDs%3D)\r\n\r\n1.**Roofline 模型**：综合峰值算力与内存带宽，判断任务是 **访存密集型** 还是 **计算密集型**。\r\n\r\n2.**优化侧重**：\r\n- **访存密集**：减少数据搬运，提高复用率（如 FlashAttention）。\r\n- **计算密集**：提高计算单元利用率（如 Tensor Core、向量化）。\r\n\r\n### 循环优化\r\n1.**循环分块 (Loop Blocking)**：切分迭代空间，使工作集驻留 Cache。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw06.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jwt118CuKFprV58WdNDcuV8b9sw%3D)\r\n\r\n2.**循环展开 (Unrolling)**：减少控制开销，提高指令级并行度。\r\n\r\n3.**循环重排 (Reordering)**：匹配内存布局（行\u002F列优先），变跳跃访问为连续访问。\r\n- *案例*：矩阵乘法 IJK → IKJ，提升 B 矩阵缓存命中率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw07.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OgTsftnvn4WEzmo0K8ThEt%2F2Xio%3D)\r\n\r\n4.**循环融合 (Fusion)**：合并同范围循环，减少开销并促进数据复用。\r\n\r\n5.**循环拆分 (Splitting)**：消除依赖与条件分支，为并行化创造条件。\r\n\r\n### Online Softmax 与 FlashAttention\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw08.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uhy8ZQYQTd9gX0cvrvcUk7c7Kn4%3D)\r\n\r\n- **Safe Softmax**：减去最大值避免指数上溢。\r\n- **Online Softmax**：流式更新最大值与指数和，支持分块计算。\r\n- **核心价值**：避免存储完整 N×N 注意力矩阵，显著降低长序列推理显存。\r\n\r\n### 指令优化\r\n1.**CPU 向量化**：SSE\u002FAVX\u002FNeon 指令，配合数据对齐与连续访存。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw09.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QgoLRKsYFzjwjnVZeZ6BSr%2FxuZY%3D)\r\n\r\n2.**GPU 张量计算**：Tensor Core 针对矩阵乘累加设计。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw10.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FmOlkcdjb9fZXzIneHB1xUMFYYA%3D)\r\n\r\n3.**PTX 指令解析**：理解 `mma.sync.align` 等指令的形状、精度与寄存器约束。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw11.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=O15rEcP87sCTvPUKhX%2Ff1LSNrQY%3D)\r\n\r\n\r\n### 内存优化\r\n1.**预取 (Prefetch)**：提前搬运数据至高速缓存。\r\n\r\n2.**双缓冲 (Double Buffering)**：计算与数据传输流水线化，隐藏 IO 延迟。\r\n\r\n\r\n## Auto Tuning\r\n### 为什么需要 Auto Tuning？\r\n- 同一算子存在多种算法实现（如 `cuDNN` 卷积算法），最优解依赖于输入形状、数据类型及硬件型号。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw12.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qOA3Y331l2LHMs8pBkJL7qvET44%3D)\r\n\r\n- 调度参数空间巨大（Tile 大小、线程数、流水线级数等），人工试错成本高。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw13.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WQ4rRESWXGvQGYOf%2Bm%2FbDcwcRbk%3D)\r\n\r\n\r\n1.  **查询缓存**：命中性能数据库则直接使用最优配置。\r\n2.  **参数搜索**：未命中时，通过网格搜索\u002F贝叶斯优化\u002FRL 等策略生成候选。\r\n3.  **实测评估**：在目标硬件上运行并测量性能。\r\n4.  **反馈入库**：将最优结果缓存，避免重复调优。\r\n\r\n\r\n## 代码生成\r\n### Triton 语言与编译器\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw14.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VD9mOtb6pvVafas%2Bg82MDPD2BZY%3D)\r\n\r\n1.**定位**：：Triton 是一种基于 Python 语法设计的 GPU Kernel DSL，通过编译器自动生成高性能 GPU 代码，降低 CUDA Kernel 开发复杂度。\r\n\r\n2.**核心抽象 - Layout**：描述元素到线程\u002FWarp\u002F内存的映射关系，处理不同硬件布局转换（如 NCHW 与 NHWC）。\r\n\r\n3.**自动优化 Pass**：\r\n- Coalescing：合并连续内存访问。\r\n- Pipeline & Prefetch：组织计算与搬运。\r\n- Swizzling：减少 Shared Memory 访问 Bank-conflict。\r\n\r\n### 编译下降流程\r\n`Python AST` → `Triton IR` → `Triton GPU IR` → `LLVM IR` → `PTX` → `GPU Execution`\r\n\r\n### 九齿框架\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw15.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7TVqsbEot7%2F3HEragoFg9K1xtU8%3D)\r\n\r\n- 进一步封装线程块索引、Program ID 等底层细节。\r\n- 让开发者聚焦算子主体逻辑，降低 CUDA 编程门槛。\r\n\r\n\r\n\r\n\r\n## 加入训练营\r\n#### 📺 完整课程内容，请观看直播或查看课程回放：\r\n\r\n**InfiniTensor 官网**： https:\u002F\u002Fwww.infinitensor.com\r\n\r\n**B 站直播**： InfiniTensor 官方直播间\r\n\r\n**视频号直播**： InfiniTensor 视频号预约\r\n\r\n**答疑交流**： 训练营官方社群 ⬇️\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=amG0zsmi5ug%2BjSjomTzuvIj1k8Y%3D)\r\n\r\n## 关注与交流\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=b1NpkMXCjFYgZ9lW3RNT0HCHt5Y%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-90","study-90","2026 夏季 InfiniTensor 训练营 CUDA 编程 7：《寒武纪软件栈 NeuWare 和 BANG 异构计算平台 2》",19,"\u002Fstudy-material\u002Fstudy-91","study-91","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-92","study-92","2026 夏季 InfiniTensor 训练营 OpenCL 编程 3：《OpenCL 算子的实现与优化》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw01.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j%2F%2FFNlL28%2Bq%2BAxzetQOsNIXCSKo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">8 月 4 日，2026 夏季 InfiniTensor 训练营 AI 编译器方向第三课《AI 编译器中的后端优化》开讲。\u003Cbr>前置课程《AI 编译器中的前端优化》讲解了四类核心等价图变换、子图替换工程实践与 CUDA Graph 底层加速方案等内容。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课将结合计算图算子实现与代码生成流程，系统拆解后端优化的四大代表性方法。\u003C\u002Fp>\n\u003Ch2 id=\"内存管理优化\">\u003Ca class=\"anchor\" href=\"#内存管理优化\" aria-label=\"跳转到标题：内存管理优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>内存管理优化\u003C\u002Fh2>\u003Ch3 id=\"优化目标与方向\">\u003Ca class=\"anchor\" href=\"#优化目标与方向\" aria-label=\"跳转到标题：优化目标与方向\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化目标与方向\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>降低运行时开销\u003C\u002Fstrong>：调整分配\u002F释放时机，减少 \u003Ccode>cudaMalloc\u003C\u002Fcode>\u002F\u003Ccode>cudaFree\u003C\u002Fcode> 等高频调用。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>降低显存占用\u003C\u002Fstrong>：优化内存规划，减少峰值显存\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"naive-版本-vs-优化版本\">\u003Ca class=\"anchor\" href=\"#naive-版本-vs-优化版本\" aria-label=\"跳转到标题：Naive 版本 vs 优化版本\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Naive 版本 vs 优化版本\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>Naive 版本痛点\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw02.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5yS%2FAVVvPgEhgCd%2BfjHN8RlXhHk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>推理前一次性为所有 Tensor 分配内存，直到结束才释放。\u003C\u002Fli>\n\u003Cli>大量非活跃 Tensor 持续占用显存，造成浪费。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>优化版本\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw03.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kYO0LHHXqT7yT4t%2F5eoS7w1G8ek%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>生命周期分析\u003C\u002Fstrong>：基于拓扑排序模拟执行过程，按 Tensor 实际使用周期规划空间。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>引用计数机制\u003C\u002Fstrong>：维护后续 OP 引用计数，归零时立即回收空间供复用。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Lazy Allocator 实现\u003C\u002Fstrong>：\u003Cul>\n\u003Cli>模拟阶段：使用红黑树维护空闲块，计算 Offset 和 Peak，不调用真实设备接口。\u003C\u002Fli>\n\u003Cli>执行阶段：仅申请一次峰值大小的连续内存，通过 \u003Ccode>Base + Offset\u003C\u002Fcode> 获取实际地址。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>3.\u003Cstrong>实战收益\u003C\u002Fstrong>：以 Inception-v3 为例，优化后支持 Batch Size 从 256 提升至 2048。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw04.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BPnNyJSJ01iNqSkYh4hfyWBcToY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>💡 \u003Cstrong>核心思想\u003C\u002Fstrong>：静态分析确定生命周期 + 统一内存池实现空间复用。InfiniTensor 与 Refactor Graph 等项目均采用此思路。\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Ch2 id=\"算子层面优化\">\u003Ca class=\"anchor\" href=\"#算子层面优化\" aria-label=\"跳转到标题：算子层面优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>算子层面优化\u003C\u002Fh2>\u003Ch3 id=\"性能瓶颈判断\">\u003Ca class=\"anchor\" href=\"#性能瓶颈判断\" aria-label=\"跳转到标题：性能瓶颈判断\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>性能瓶颈判断\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw05.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wLbRJ3e06d4N8iSc%2Fk9VJ%2F4OiDs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>Roofline 模型\u003C\u002Fstrong>：综合峰值算力与内存带宽，判断任务是 \u003Cstrong>访存密集型\u003C\u002Fstrong> 还是 \u003Cstrong>计算密集型\u003C\u002Fstrong>。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>优化侧重\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>访存密集\u003C\u002Fstrong>：减少数据搬运，提高复用率（如 FlashAttention）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>计算密集\u003C\u002Fstrong>：提高计算单元利用率（如 Tensor Core、向量化）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"循环优化\">\u003Ca class=\"anchor\" href=\"#循环优化\" aria-label=\"跳转到标题：循环优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>循环优化\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>循环分块 (Loop Blocking)\u003C\u002Fstrong>：切分迭代空间，使工作集驻留 Cache。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw06.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jwt118CuKFprV58WdNDcuV8b9sw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>循环展开 (Unrolling)\u003C\u002Fstrong>：减少控制开销，提高指令级并行度。\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>循环重排 (Reordering)\u003C\u002Fstrong>：匹配内存布局（行\u002F列优先），变跳跃访问为连续访问。\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cem>案例\u003C\u002Fem>：矩阵乘法 IJK → IKJ，提升 B 矩阵缓存命中率。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw07.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OgTsftnvn4WEzmo0K8ThEt%2F2Xio%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>4.\u003Cstrong>循环融合 (Fusion)\u003C\u002Fstrong>：合并同范围循环，减少开销并促进数据复用。\u003C\u002Fp>\n\u003Cp>5.\u003Cstrong>循环拆分 (Splitting)\u003C\u002Fstrong>：消除依赖与条件分支，为并行化创造条件。\u003C\u002Fp>\n\u003Ch3 id=\"online-softmax-与-flashattention\">\u003Ca class=\"anchor\" href=\"#online-softmax-与-flashattention\" aria-label=\"跳转到标题：Online Softmax 与 FlashAttention\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Online Softmax 与 FlashAttention\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw08.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uhy8ZQYQTd9gX0cvrvcUk7c7Kn4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Safe Softmax\u003C\u002Fstrong>：减去最大值避免指数上溢。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Online Softmax\u003C\u002Fstrong>：流式更新最大值与指数和，支持分块计算。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>核心价值\u003C\u002Fstrong>：避免存储完整 N×N 注意力矩阵，显著降低长序列推理显存。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"指令优化\">\u003Ca class=\"anchor\" href=\"#指令优化\" aria-label=\"跳转到标题：指令优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>指令优化\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>CPU 向量化\u003C\u002Fstrong>：SSE\u002FAVX\u002FNeon 指令，配合数据对齐与连续访存。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw09.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QgoLRKsYFzjwjnVZeZ6BSr%2FxuZY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>GPU 张量计算\u003C\u002Fstrong>：Tensor Core 针对矩阵乘累加设计。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw10.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FmOlkcdjb9fZXzIneHB1xUMFYYA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>PTX 指令解析\u003C\u002Fstrong>：理解 \u003Ccode>mma.sync.align\u003C\u002Fcode> 等指令的形状、精度与寄存器约束。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw11.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=O15rEcP87sCTvPUKhX%2Ff1LSNrQY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"内存优化\">\u003Ca class=\"anchor\" href=\"#内存优化\" aria-label=\"跳转到标题：内存优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>内存优化\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>预取 (Prefetch)\u003C\u002Fstrong>：提前搬运数据至高速缓存。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>双缓冲 (Double Buffering)\u003C\u002Fstrong>：计算与数据传输流水线化，隐藏 IO 延迟。\u003C\u002Fp>\n\u003Ch2 id=\"auto-tuning\">\u003Ca class=\"anchor\" href=\"#auto-tuning\" aria-label=\"跳转到标题：Auto Tuning\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Auto Tuning\u003C\u002Fh2>\u003Ch3 id=\"为什么需要-auto-tuning\">\u003Ca class=\"anchor\" href=\"#为什么需要-auto-tuning\" aria-label=\"跳转到标题：为什么需要 Auto Tuning？\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>为什么需要 Auto Tuning？\u003C\u002Fh3>\u003Cul>\n\u003Cli>同一算子存在多种算法实现（如 \u003Ccode>cuDNN\u003C\u002Fcode> 卷积算法），最优解依赖于输入形状、数据类型及硬件型号。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw12.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qOA3Y331l2LHMs8pBkJL7qvET44%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>调度参数空间巨大（Tile 大小、线程数、流水线级数等），人工试错成本高。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw13.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WQ4rRESWXGvQGYOf%2Bm%2FbDcwcRbk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>查询缓存\u003C\u002Fstrong>：命中性能数据库则直接使用最优配置。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>参数搜索\u003C\u002Fstrong>：未命中时，通过网格搜索\u002F贝叶斯优化\u002FRL 等策略生成候选。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>实测评估\u003C\u002Fstrong>：在目标硬件上运行并测量性能。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>反馈入库\u003C\u002Fstrong>：将最优结果缓存，避免重复调优。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2 id=\"代码生成\">\u003Ca class=\"anchor\" href=\"#代码生成\" aria-label=\"跳转到标题：代码生成\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>代码生成\u003C\u002Fh2>\u003Ch3 id=\"triton-语言与编译器\">\u003Ca class=\"anchor\" href=\"#triton-语言与编译器\" aria-label=\"跳转到标题：Triton 语言与编译器\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Triton 语言与编译器\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw14.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VD9mOtb6pvVafas%2Bg82MDPD2BZY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>定位\u003C\u002Fstrong>：：Triton 是一种基于 Python 语法设计的 GPU Kernel DSL，通过编译器自动生成高性能 GPU 代码，降低 CUDA Kernel 开发复杂度。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>核心抽象 - Layout\u003C\u002Fstrong>：描述元素到线程\u002FWarp\u002F内存的映射关系，处理不同硬件布局转换（如 NCHW 与 NHWC）。\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>自动优化 Pass\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Coalescing：合并连续内存访问。\u003C\u002Fli>\n\u003Cli>Pipeline & Prefetch：组织计算与搬运。\u003C\u002Fli>\n\u003Cli>Swizzling：减少 Shared Memory 访问 Bank-conflict。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"编译下降流程\">\u003Ca class=\"anchor\" href=\"#编译下降流程\" aria-label=\"跳转到标题：编译下降流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>编译下降流程\u003C\u002Fh3>\u003Cp>\u003Ccode>Python AST\u003C\u002Fcode> → \u003Ccode>Triton IR\u003C\u002Fcode> → \u003Ccode>Triton GPU IR\u003C\u002Fcode> → \u003Ccode>LLVM IR\u003C\u002Fcode> → \u003Ccode>PTX\u003C\u002Fcode> → \u003Ccode>GPU Execution\u003C\u002Fcode>\u003C\u002Fp>\n\u003Ch3 id=\"九齿框架\">\u003Ca class=\"anchor\" href=\"#九齿框架\" aria-label=\"跳转到标题：九齿框架\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>九齿框架\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_91_zw15.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7TVqsbEot7%2F3HEragoFg9K1xtU8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>进一步封装线程块索引、Program ID 等底层细节。\u003C\u002Fli>\n\u003Cli>让开发者聚焦算子主体逻辑，降低 CUDA 编程门槛。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"加入训练营\">\u003Ca class=\"anchor\" href=\"#加入训练营\" aria-label=\"跳转到标题：加入训练营\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>加入训练营\u003C\u002Fh2>\u003Ch4 id=\"完整课程内容-请观看直播或查看课程回放\">\u003Ca class=\"anchor\" href=\"#完整课程内容-请观看直播或查看课程回放\" aria-label=\"跳转到标题：📺 完整课程内容，请观看直播或查看课程回放：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📺 完整课程内容，请观看直播或查看课程回放：\u003C\u002Fh4>\u003Cp>\u003Cstrong>InfiniTensor 官网\u003C\u002Fstrong>： \u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\">https:\u002F\u002Fwww.infinitensor.com\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Cstrong>B 站直播\u003C\u002Fstrong>： InfiniTensor 官方直播间\u003C\u002Fp>\n\u003Cp>\u003Cstrong>视频号直播\u003C\u002Fstrong>： InfiniTensor 视频号预约\u003C\u002Fp>\n\u003Cp>\u003Cstrong>答疑交流\u003C\u002Fstrong>： 训练营官方社群 ⬇️\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=amG0zsmi5ug%2BjSjomTzuvIj1k8Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"关注与交流\">\u003Ca class=\"anchor\" href=\"#关注与交流\" aria-label=\"跳转到标题：关注与交流\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>关注与交流\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752561&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=b1NpkMXCjFYgZ9lW3RNT0HCHt5Y%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,132,135,137,139,141,144,146,148,151,154,156,159,161,163,165,169],{"id":129,"text":129,"level":110},"内存管理优化",{"id":131,"text":131,"level":9},"优化目标与方向",{"id":133,"text":134,"level":9},"naive-版本-vs-优化版本","Naive 版本 vs 优化版本",{"id":136,"text":136,"level":110},"算子层面优化",{"id":138,"text":138,"level":9},"性能瓶颈判断",{"id":140,"text":140,"level":9},"循环优化",{"id":142,"text":143,"level":9},"online-softmax-与-flashattention","Online Softmax 与 FlashAttention",{"id":145,"text":145,"level":9},"指令优化",{"id":147,"text":147,"level":9},"内存优化",{"id":149,"text":150,"level":110},"auto-tuning","Auto Tuning",{"id":152,"text":153,"level":9},"为什么需要-auto-tuning","为什么需要 Auto Tuning？",{"id":155,"text":155,"level":110},"代码生成",{"id":157,"text":158,"level":9},"triton-语言与编译器","Triton 语言与编译器",{"id":160,"text":160,"level":9},"编译下降流程",{"id":162,"text":162,"level":9},"九齿框架",{"id":164,"text":164,"level":110},"加入训练营",{"id":166,"text":167,"level":168},"完整课程内容-请观看直播或查看课程回放","📺 完整课程内容，请观看直播或查看课程回放：",4,{"id":170,"text":170,"level":110},"关注与交流"]