[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-23":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=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mI4adTQzTPRyYrPmMuEsZWwmTNo%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=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8YWKiH9EYPM3GQM%2F7JdCBPynFmw%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=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JY%2B%2B2oyU0ZlhO2Io1fpbwb5rDUc%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_23_fm.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hLH04YuI1xIPn0Ci68%2BKMme6Yqo%3D","InfiniTensor,冬季训练营,CUDA,内存模型,规约优化","讲解 CUDA 内存模型与规约优化，覆盖共享内存使用、规约设计与吞吐提升策略。","2026-10-22 20:38:11",2,"# InfiniTensor 冬季训练营 CUDA 编程第三课《内存模型与规约优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw01.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Yt1MhiqWDoMMyCtNOVpgXz1bokk%3D)\r\n\r\n\r\n2026 年 1 月 19 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第三课《内存模型与规约优化》开讲。\r\n\r\n本节课介绍如何通过 Roofline 模型分析内存和计算瓶颈，并详细讲解如何利用向量化、低精度计算、内存对齐、访存合并等技术优化 CUDA 程序的性能。\r\n\r\n### 性能瓶颈分析与优化方法论\r\n#### 1.瓶颈分析基础与 Roofline 模型\r\n- 内存瓶颈与计算瓶颈的识别\r\n- 如何定位性能天花板\r\n\r\n#### 2.性能优化方法论\r\n- 右移优化：提升计算强度（AI）\r\n- 算子融合与计算密度提升方法\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw02.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hi3DtT%2BiPEn4H4aou8LSAo8EVfo%3D)\r\n\r\n### 向量化与低精度优化\r\n#### 1. 向量化计算基础\r\n- 介绍向量化计算的概念\r\n- 内存对齐与访存合并的重要性\r\n#### 2. 低精度计算优化\r\n- 半精度（Half）计算的优势\r\n- 混合精度计算策略（如 half2）\r\n#### 3. 性能对比分析\r\n- 不同精度与向量化级别下的性能表现\r\n\r\n### 并行规约与原子操作\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw03.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xl0Q62tFrbgeJabmhUrMiaZ1QBA%3D)\r\n\r\n#### 1. 并行规约问题引入\r\n- 从向量加法到累加求和的并行化挑战\r\n#### 2. 原子操作基础\r\n- 原子操作的定义与用途\r\n- atomicAdd 的使用与性能影响\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw04.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CDR8%2BuKQGgYn%2F21P6yZ9AzYces0%3D)\r\n\r\n#### 3. 规约策略优化\r\n- Warp 级规约：\r\n左侧的代码 reduce_warp_global_kernel 采用了 Grid-Stride Loop（网格跨步循环） 模式，使其能够处理超出网格尺寸的数据量；为了提高效率，逻辑中加入了 Warp 层级的控制，仅让每个 Warp 中的第一个线程（Lane 0）负责执行循环累加，计算出该 Warp 的局部和（partial sum）。右侧文字总结了关键点：通过这种方式减少了不必要的线程工作，最后利用 atomicAdd（原子加） 操作将各个 Warp 计算出的局部和安全地汇总到全局输出中，从而避免了数据竞争并完成了最终的规约。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw05.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GtBcTwpr5Uy24pfpR8ZwUtWrRyM%3D)\r\n\r\n- Block 级规约\r\n- 树状规约算法\r\n#### 4. 性能对比与瓶颈分析\r\n- 不同规约策略的性能差异\r\n- 原子操作导致的串行化问题\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw06.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mgP5ZNZe61e3k2MBT6n7UdiWNnk%3D)\r\n\r\n### 内存层次结构与访存优化\r\n#### 1. GPU 内存架构概述\r\n- 物理内存层次：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw07.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=94HIpRpByFSPmXTdR8ii4bZRIzc%3D)\r\n\r\n- 逻辑内存模型：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw08.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aDOECzpG35HUrTHn5nwIus0iPzU%3D)\r\n\r\n#### 2. 共享内存与访存优化\r\n- 共享内存的分配与使用（静态\u002F动态）\r\n- 同步操作：`__syncthreads()`\r\n#### 3. 树状归约的优化实现\r\n- 共享内存中的规约实现\r\n- 访存模式优化避免 Bank Conflict\r\n\r\n### Warp 级优化\r\n#### 1. Warp 调度与线程组织\r\n- Warp 与 Lane 的概念\r\n- Warp 级同步与通信\r\n#### 2. Warp Shuffle 操作\r\n- `__shfl_down_sync()`的使用与优势\r\n- 基于寄存器的数据交换\r\n#### 3. Warp 级规约优化\r\n- 结合 Warp Shuffle 与共享内存的规约策略：在并行累加中一种基于 **Warp Shuffle（线程束洗牌）** 的进阶优化方案，核心在于引入了 **`__shfl_down_sync`** 函数。\r\n它改变了传统利用“共享内存”进行线程间通信的方式，允许线程直接通过**寄存器**高效地读取同一 Warp 内下游线程（`lane_id + offset`）的数据；代码展示了如何通过一个从 16 到 1 的循环步长，在不进行显式同步（无 `__syncthreads`）且无需处理复杂边界条件的情况下，极快地完成 Warp 内部的数值归约，最终仅由 0 号线程将结果原子加到全局内存中，这显著降低了访存延迟并提升了 Kernel 的执行效率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw09.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hdNJmIfbyynRtrhzn7RxPtEwrLk%3D)\r\n\r\n### 跨 Block 规约\r\n#### 1. 跨 Block 规约的挑战\r\n- 原子操作 vs 多核函数规约\r\n#### 2. Two-Pass 规约策略\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw10.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mJkQr0%2FAdyojgFwmeu%2FDYlDT3Xg%3D)\r\n\r\n#### 3.跨块规约优化\r\n一次 Kernel 发射完成全局求和，所必须采用的 Cooperative Groups（协作组） 及其 Host 端启动配置。\r\n由于标准的 CUDA 编程模型中不同 Thread Block 之间无法直接同步，要实现全局规约通常需要多次 Kernel 启动。而图中展示的方案通过 **`cudaLaunchCooperativeKernel 实现了`** Grid 级别的全局同步，但这要求开发者必须进行严格的资源管理：首先使用 **`cudaOccupancyMaxActiveBlocksPerMultiprocessor`** 精确计算出硬件能同时承载的最大活跃线程块数量（因为所有 Block 必须同时驻留在 GPU 上才能避免死锁），接着验证设备是否支持 Cooperative Launch 属性，最后通过特定的 API 启动核心，从而在单次发射中完成从局部到全局的高效累加。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw11.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xn3%2B38TjDEG2kr5I4p77nhaAOyg%3D)\r\n\r\n\r\n### Bank Conflict 分析与优化\r\n#### 1. Bank Conflict 的概念与影响\r\n- 共享内存的 Bank 组织与访问模式\r\n- 冲突类型：共享内存的读\u002F写 Bank Conflict\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw12.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qtIm7m5VjnI4kOSU2z8JpFO9Wts%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw13.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HstEI%2F9QHS2u41wBTpRojRHszi8%3D)\r\n\r\n\r\n#### 2. Bank Conflict 的检测工具\r\n- NCU 工具中的相关指标分析\r\n#### 3. 优化策略\r\n- 访存模式优化（如树状规约的调整）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw14.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=19HXM%2BjC2aaI%2BWptjKkp4MJTIXk%3D)\r\n\r\n- Padding（内存填充）技术\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw15.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j%2Ff5ZY5%2BXOzX4pDzsZIKp7cAJC8%3D)\r\n\r\n\r\n### 总结\r\n本节课重点掌握使用 CUDA 进行并行计算的核心概念，并学会使用各种工具和技术分析和优化程序性能，特别是解决常见的访存瓶颈和并行规约问题。下节课将解决 Bank conflict 的另一技巧！\r\n\r\n\r\n\r\n![2025冬季训练营课程表](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Q30J6aqu5J%2BBnh7mHX9KqjpwEAg%3D)\r\n\r\n### 💡 学习贴士\r\n更多细节及完整课程内容，请观看直播或查看课程回放。\r\n\r\n链接：https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\r\n\r\n### 📚 训练营报名持续进行中\r\n扫码报名与加入群聊，或点击 **「阅读原文」** 直达官网报名\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw7.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=x27RQsUfb%2FJ2%2FOIRcLgioJ0jILM%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-22","study-22","InfiniTensor 冬季训练营通信与并行第二课《大模型并行策略与通信优化》",86,"\u002Fstudy-material\u002Fstudy-23","study-23","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-24","study-24","InfiniTensor 冬季训练营 Triton & 九齿第二课《九齿一重：初窥》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw01.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Yt1MhiqWDoMMyCtNOVpgXz1bokk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 19 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第三课《内存模型与规约优化》开讲。\u003C\u002Fp>\n\u003Cp>本节课介绍如何通过 Roofline 模型分析内存和计算瓶颈，并详细讲解如何利用向量化、低精度计算、内存对齐、访存合并等技术优化 CUDA 程序的性能。\u003C\u002Fp>\n\u003Ch3 id=\"性能瓶颈分析与优化方法论\">\u003Ca class=\"anchor\" href=\"#性能瓶颈分析与优化方法论\" aria-label=\"跳转到标题：性能瓶颈分析与优化方法论\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>性能瓶颈分析与优化方法论\u003C\u002Fh3>\u003Ch4 id=\"1-瓶颈分析基础与-roofline-模型\">\u003Ca class=\"anchor\" href=\"#1-瓶颈分析基础与-roofline-模型\" aria-label=\"跳转到标题：1.瓶颈分析基础与 Roofline 模型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.瓶颈分析基础与 Roofline 模型\u003C\u002Fh4>\u003Cul>\n\u003Cli>内存瓶颈与计算瓶颈的识别\u003C\u002Fli>\n\u003Cli>如何定位性能天花板\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-性能优化方法论\">\u003Ca class=\"anchor\" href=\"#2-性能优化方法论\" aria-label=\"跳转到标题：2.性能优化方法论\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.性能优化方法论\u003C\u002Fh4>\u003Cul>\n\u003Cli>右移优化：提升计算强度（AI）\u003C\u002Fli>\n\u003Cli>算子融合与计算密度提升方法\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw02.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hi3DtT%2BiPEn4H4aou8LSAo8EVfo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"向量化与低精度优化\">\u003Ca class=\"anchor\" href=\"#向量化与低精度优化\" aria-label=\"跳转到标题：向量化与低精度优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>向量化与低精度优化\u003C\u002Fh3>\u003Ch4 id=\"1-向量化计算基础\">\u003Ca class=\"anchor\" href=\"#1-向量化计算基础\" aria-label=\"跳转到标题：1. 向量化计算基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 向量化计算基础\u003C\u002Fh4>\u003Cul>\n\u003Cli>介绍向量化计算的概念\u003C\u002Fli>\n\u003Cli>内存对齐与访存合并的重要性\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-低精度计算优化\">\u003Ca class=\"anchor\" href=\"#2-低精度计算优化\" aria-label=\"跳转到标题：2. 低精度计算优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 低精度计算优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>半精度（Half）计算的优势\u003C\u002Fli>\n\u003Cli>混合精度计算策略（如 half2）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-性能对比分析\">\u003Ca class=\"anchor\" href=\"#3-性能对比分析\" aria-label=\"跳转到标题：3. 性能对比分析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 性能对比分析\u003C\u002Fh4>\u003Cul>\n\u003Cli>不同精度与向量化级别下的性能表现\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"并行规约与原子操作\">\u003Ca class=\"anchor\" href=\"#并行规约与原子操作\" aria-label=\"跳转到标题：并行规约与原子操作\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>并行规约与原子操作\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw03.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xl0Q62tFrbgeJabmhUrMiaZ1QBA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-并行规约问题引入\">\u003Ca class=\"anchor\" href=\"#1-并行规约问题引入\" aria-label=\"跳转到标题：1. 并行规约问题引入\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 并行规约问题引入\u003C\u002Fh4>\u003Cul>\n\u003Cli>从向量加法到累加求和的并行化挑战\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-原子操作基础\">\u003Ca class=\"anchor\" href=\"#2-原子操作基础\" aria-label=\"跳转到标题：2. 原子操作基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 原子操作基础\u003C\u002Fh4>\u003Cul>\n\u003Cli>原子操作的定义与用途\u003C\u002Fli>\n\u003Cli>atomicAdd 的使用与性能影响\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw04.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CDR8%2BuKQGgYn%2F21P6yZ9AzYces0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-规约策略优化\">\u003Ca class=\"anchor\" href=\"#3-规约策略优化\" aria-label=\"跳转到标题：3. 规约策略优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 规约策略优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>Warp 级规约：\u003Cbr>左侧的代码 reduce_warp_global_kernel 采用了 Grid-Stride Loop（网格跨步循环） 模式，使其能够处理超出网格尺寸的数据量；为了提高效率，逻辑中加入了 Warp 层级的控制，仅让每个 Warp 中的第一个线程（Lane 0）负责执行循环累加，计算出该 Warp 的局部和（partial sum）。右侧文字总结了关键点：通过这种方式减少了不必要的线程工作，最后利用 atomicAdd（原子加） 操作将各个 Warp 计算出的局部和安全地汇总到全局输出中，从而避免了数据竞争并完成了最终的规约。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw05.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GtBcTwpr5Uy24pfpR8ZwUtWrRyM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Block 级规约\u003C\u002Fli>\n\u003Cli>树状规约算法\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"4-性能对比与瓶颈分析\">\u003Ca class=\"anchor\" href=\"#4-性能对比与瓶颈分析\" aria-label=\"跳转到标题：4. 性能对比与瓶颈分析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. 性能对比与瓶颈分析\u003C\u002Fh4>\u003Cul>\n\u003Cli>不同规约策略的性能差异\u003C\u002Fli>\n\u003Cli>原子操作导致的串行化问题\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw06.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mgP5ZNZe61e3k2MBT6n7UdiWNnk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"内存层次结构与访存优化\">\u003Ca class=\"anchor\" href=\"#内存层次结构与访存优化\" aria-label=\"跳转到标题：内存层次结构与访存优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>内存层次结构与访存优化\u003C\u002Fh3>\u003Ch4 id=\"1-gpu-内存架构概述\">\u003Ca class=\"anchor\" href=\"#1-gpu-内存架构概述\" aria-label=\"跳转到标题：1. GPU 内存架构概述\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. GPU 内存架构概述\u003C\u002Fh4>\u003Cul>\n\u003Cli>物理内存层次：\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw07.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=94HIpRpByFSPmXTdR8ii4bZRIzc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>逻辑内存模型：\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw08.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aDOECzpG35HUrTHn5nwIus0iPzU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-共享内存与访存优化\">\u003Ca class=\"anchor\" href=\"#2-共享内存与访存优化\" aria-label=\"跳转到标题：2. 共享内存与访存优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 共享内存与访存优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>共享内存的分配与使用（静态\u002F动态）\u003C\u002Fli>\n\u003Cli>同步操作：\u003Ccode>__syncthreads()\u003C\u002Fcode>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-树状归约的优化实现\">\u003Ca class=\"anchor\" href=\"#3-树状归约的优化实现\" aria-label=\"跳转到标题：3. 树状归约的优化实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 树状归约的优化实现\u003C\u002Fh4>\u003Cul>\n\u003Cli>共享内存中的规约实现\u003C\u002Fli>\n\u003Cli>访存模式优化避免 Bank Conflict\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"warp-级优化\">\u003Ca class=\"anchor\" href=\"#warp-级优化\" aria-label=\"跳转到标题：Warp 级优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Warp 级优化\u003C\u002Fh3>\u003Ch4 id=\"1-warp-调度与线程组织\">\u003Ca class=\"anchor\" href=\"#1-warp-调度与线程组织\" aria-label=\"跳转到标题：1. Warp 调度与线程组织\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. Warp 调度与线程组织\u003C\u002Fh4>\u003Cul>\n\u003Cli>Warp 与 Lane 的概念\u003C\u002Fli>\n\u003Cli>Warp 级同步与通信\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-warp-shuffle-操作\">\u003Ca class=\"anchor\" href=\"#2-warp-shuffle-操作\" aria-label=\"跳转到标题：2. Warp Shuffle 操作\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Warp Shuffle 操作\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Ccode>__shfl_down_sync()\u003C\u002Fcode>的使用与优势\u003C\u002Fli>\n\u003Cli>基于寄存器的数据交换\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-warp-级规约优化\">\u003Ca class=\"anchor\" href=\"#3-warp-级规约优化\" aria-label=\"跳转到标题：3. Warp 级规约优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. Warp 级规约优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>结合 Warp Shuffle 与共享内存的规约策略：在并行累加中一种基于 \u003Cstrong>Warp Shuffle（线程束洗牌）\u003C\u002Fstrong> 的进阶优化方案，核心在于引入了 \u003Cstrong>\u003Ccode>__shfl_down_sync\u003C\u002Fcode>\u003C\u002Fstrong> 函数。\u003Cbr>它改变了传统利用“共享内存”进行线程间通信的方式，允许线程直接通过\u003Cstrong>寄存器\u003C\u002Fstrong>高效地读取同一 Warp 内下游线程（\u003Ccode>lane_id + offset\u003C\u002Fcode>）的数据；代码展示了如何通过一个从 16 到 1 的循环步长，在不进行显式同步（无 \u003Ccode>__syncthreads\u003C\u002Fcode>）且无需处理复杂边界条件的情况下，极快地完成 Warp 内部的数值归约，最终仅由 0 号线程将结果原子加到全局内存中，这显著降低了访存延迟并提升了 Kernel 的执行效率。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw09.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hdNJmIfbyynRtrhzn7RxPtEwrLk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"跨-block-规约\">\u003Ca class=\"anchor\" href=\"#跨-block-规约\" aria-label=\"跳转到标题：跨 Block 规约\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>跨 Block 规约\u003C\u002Fh3>\u003Ch4 id=\"1-跨-block-规约的挑战\">\u003Ca class=\"anchor\" href=\"#1-跨-block-规约的挑战\" aria-label=\"跳转到标题：1. 跨 Block 规约的挑战\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 跨 Block 规约的挑战\u003C\u002Fh4>\u003Cul>\n\u003Cli>原子操作 vs 多核函数规约\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-two-pass-规约策略\">\u003Ca class=\"anchor\" href=\"#2-two-pass-规约策略\" aria-label=\"跳转到标题：2. Two-Pass 规约策略\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Two-Pass 规约策略\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw10.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mJkQr0%2FAdyojgFwmeu%2FDYlDT3Xg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-跨块规约优化\">\u003Ca class=\"anchor\" href=\"#3-跨块规约优化\" aria-label=\"跳转到标题：3.跨块规约优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.跨块规约优化\u003C\u002Fh4>\u003Cp>一次 Kernel 发射完成全局求和，所必须采用的 Cooperative Groups（协作组） 及其 Host 端启动配置。\u003Cbr>由于标准的 CUDA 编程模型中不同 Thread Block 之间无法直接同步，要实现全局规约通常需要多次 Kernel 启动。而图中展示的方案通过 \u003Cstrong>\u003Ccode>cudaLaunchCooperativeKernel 实现了\u003C\u002Fcode>\u003C\u002Fstrong> Grid 级别的全局同步，但这要求开发者必须进行严格的资源管理：首先使用 \u003Cstrong>\u003Ccode>cudaOccupancyMaxActiveBlocksPerMultiprocessor\u003C\u002Fcode>\u003C\u002Fstrong> 精确计算出硬件能同时承载的最大活跃线程块数量（因为所有 Block 必须同时驻留在 GPU 上才能避免死锁），接着验证设备是否支持 Cooperative Launch 属性，最后通过特定的 API 启动核心，从而在单次发射中完成从局部到全局的高效累加。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw11.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xn3%2B38TjDEG2kr5I4p77nhaAOyg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"bank-conflict-分析与优化\">\u003Ca class=\"anchor\" href=\"#bank-conflict-分析与优化\" aria-label=\"跳转到标题：Bank Conflict 分析与优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Bank Conflict 分析与优化\u003C\u002Fh3>\u003Ch4 id=\"1-bank-conflict-的概念与影响\">\u003Ca class=\"anchor\" href=\"#1-bank-conflict-的概念与影响\" aria-label=\"跳转到标题：1. Bank Conflict 的概念与影响\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. Bank Conflict 的概念与影响\u003C\u002Fh4>\u003Cul>\n\u003Cli>共享内存的 Bank 组织与访问模式\u003C\u002Fli>\n\u003Cli>冲突类型：共享内存的读\u002F写 Bank Conflict\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw12.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qtIm7m5VjnI4kOSU2z8JpFO9Wts%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw13.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HstEI%2F9QHS2u41wBTpRojRHszi8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-bank-conflict-的检测工具\">\u003Ca class=\"anchor\" href=\"#2-bank-conflict-的检测工具\" aria-label=\"跳转到标题：2. Bank Conflict 的检测工具\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Bank Conflict 的检测工具\u003C\u002Fh4>\u003Cul>\n\u003Cli>NCU 工具中的相关指标分析\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-优化策略\">\u003Ca class=\"anchor\" href=\"#3-优化策略\" aria-label=\"跳转到标题：3. 优化策略\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 优化策略\u003C\u002Fh4>\u003Cul>\n\u003Cli>访存模式优化（如树状规约的调整）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw14.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=19HXM%2BjC2aaI%2BWptjKkp4MJTIXk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Padding（内存填充）技术\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_23_zw15.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j%2Ff5ZY5%2BXOzX4pDzsZIKp7cAJC8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"总结\">\u003Ca class=\"anchor\" href=\"#总结\" aria-label=\"跳转到标题：总结\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>总结\u003C\u002Fh3>\u003Cp>本节课重点掌握使用 CUDA 进行并行计算的核心概念，并学会使用各种工具和技术分析和优化程序性能，特别是解决常见的访存瓶颈和并行规约问题。下节课将解决 Bank conflict 的另一技巧！\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Q30J6aqu5J%2BBnh7mHX9KqjpwEAg%3D\" alt=\"2025冬季训练营课程表\">\u003C\u002Fp>\n\u003Ch3 id=\"学习贴士\">\u003Ca class=\"anchor\" href=\"#学习贴士\" aria-label=\"跳转到标题：💡 学习贴士\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>💡 学习贴士\u003C\u002Fh3>\u003Cp>更多细节及完整课程内容，请观看直播或查看课程回放。\u003C\u002Fp>\n\u003Cp>链接：\u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\">https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\u003C\u002Fa>\u003C\u002Fp>\n\u003Ch3 id=\"训练营报名持续进行中\">\u003Ca class=\"anchor\" href=\"#训练营报名持续进行中\" aria-label=\"跳转到标题：📚 训练营报名持续进行中\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📚 训练营报名持续进行中\u003C\u002Fh3>\u003Cp>扫码报名与加入群聊，或点击 \u003Cstrong>「阅读原文」\u003C\u002Fstrong> 直达官网报名\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw7.jpg?Expires=1792752839&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=x27RQsUfb%2FJ2%2FOIRcLgioJ0jILM%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,134,137,139,142,145,148,150,153,156,159,162,164,167,170,173,176,179,182,185,188,191,194,197,200,203,206,209,211,214],{"id":129,"text":129,"level":9},"性能瓶颈分析与优化方法论",{"id":131,"text":132,"level":133},"1-瓶颈分析基础与-roofline-模型","1.瓶颈分析基础与 Roofline 模型",4,{"id":135,"text":136,"level":133},"2-性能优化方法论","2.性能优化方法论",{"id":138,"text":138,"level":9},"向量化与低精度优化",{"id":140,"text":141,"level":133},"1-向量化计算基础","1. 向量化计算基础",{"id":143,"text":144,"level":133},"2-低精度计算优化","2. 低精度计算优化",{"id":146,"text":147,"level":133},"3-性能对比分析","3. 性能对比分析",{"id":149,"text":149,"level":9},"并行规约与原子操作",{"id":151,"text":152,"level":133},"1-并行规约问题引入","1. 并行规约问题引入",{"id":154,"text":155,"level":133},"2-原子操作基础","2. 原子操作基础",{"id":157,"text":158,"level":133},"3-规约策略优化","3. 规约策略优化",{"id":160,"text":161,"level":133},"4-性能对比与瓶颈分析","4. 性能对比与瓶颈分析",{"id":163,"text":163,"level":9},"内存层次结构与访存优化",{"id":165,"text":166,"level":133},"1-gpu-内存架构概述","1. GPU 内存架构概述",{"id":168,"text":169,"level":133},"2-共享内存与访存优化","2. 共享内存与访存优化",{"id":171,"text":172,"level":133},"3-树状归约的优化实现","3. 树状归约的优化实现",{"id":174,"text":175,"level":9},"warp-级优化","Warp 级优化",{"id":177,"text":178,"level":133},"1-warp-调度与线程组织","1. Warp 调度与线程组织",{"id":180,"text":181,"level":133},"2-warp-shuffle-操作","2. Warp Shuffle 操作",{"id":183,"text":184,"level":133},"3-warp-级规约优化","3. Warp 级规约优化",{"id":186,"text":187,"level":9},"跨-block-规约","跨 Block 规约",{"id":189,"text":190,"level":133},"1-跨-block-规约的挑战","1. 跨 Block 规约的挑战",{"id":192,"text":193,"level":133},"2-two-pass-规约策略","2. Two-Pass 规约策略",{"id":195,"text":196,"level":133},"3-跨块规约优化","3.跨块规约优化",{"id":198,"text":199,"level":9},"bank-conflict-分析与优化","Bank Conflict 分析与优化",{"id":201,"text":202,"level":133},"1-bank-conflict-的概念与影响","1. Bank Conflict 的概念与影响",{"id":204,"text":205,"level":133},"2-bank-conflict-的检测工具","2. Bank Conflict 的检测工具",{"id":207,"text":208,"level":133},"3-优化策略","3. 优化策略",{"id":210,"text":210,"level":9},"总结",{"id":212,"text":213,"level":9},"学习贴士","💡 学习贴士",{"id":215,"text":216,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]