[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-76":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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ihPLIAuCbx%2FbInTQLbtnjbL%2FgEc%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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RFJPol03COkEgvG5I%2FXurn1xrVs%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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=3EgCBk5kjXXxH6wOy5O8vDSNjzc%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_76_fm.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xMLx1aKicifjCk4hq%2Bjh8%2B0nRjA%3D","CUDA,规约优化,共享内存,Bank 冲突,Roofline 模型","基于 Roofline 模型提升计算强度，讲解原子操作弊端、Warp \u002F 共享内存树状规约，解析 GPU 内存层级与 Bank 冲突，介绍 Cooperative Groups 跨块同步优化手段。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 CUDA 编程 3：《内存模型与规约优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw01.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=91EZXUb%2Fz4BAZxZjrrTbyBMuue8%3D)\r\n\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 23 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第三课《内存模型与规约优化》开讲。\r\n前置课程《性能模型与逐元素优化》讲了 CUDA 并行编程中的“内存墙”与 Roofline 模型，演示了向量化访存、内存对齐与合并机制等内容。\r\n\u003C\u002Fspan>\r\n\r\n本节课承接上节并行加法性能优化内容，从 Roofline 模型出发，深入探讨如何通过提高计算强度突破性能瓶颈。重点讲解并行规约中的原子操作陷阱、分治思想应用、GPU 内存层级架构，以及 Shared Memory Bank Conflict 等底层优化细节。\r\n\r\n\r\n## 回顾与性能瓶颈分析\r\n1. **上节课优化路径回顾**\r\n   - CPU → GPU FP32 → 访存向量化（Fold4）→ 半精度量化\r\n   - 数据量对性能表现的影响（小数据量 vs 大数据量）\r\n2. **Roofline 模型指导优化**\r\n   - 目标：右移 Roofline 曲线，提高计算强度 $AI = W \u002F Q$\r\n   - 策略：增加计算量 $W$（如算子融合）、减少访存量 $Q$\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw02.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=c2VxnvRyL%2F6Cq1X2j0KiZgxHF64%3D)\r\n\r\n3. **从向量加法到累加求和**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw03.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FGiIA%2BlGk57fg62SmvI6xGLhxlc%3D)\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw04.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DJjidHIKv6iuQUQbrYTPE9fcrHc%3D)\r\n\r\n\r\n   - 尴尬并行（Embarrassingly Parallel）vs 数据竞争\r\n   - 如何解决这个竞争的问题？\r\n\r\n## 原子操作与初步规约优化\r\n### 原子操作（Atomic Operations）\r\n- 概念：不可分割的最小执行单元（All-or-Nothing）\r\n- CUDA API：`atomicAdd` 的使用与正确性验证\r\n- 性能代价：NCU 分析显示指令发射严重阻塞\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw05.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9R1FPybdg6xoaVOin8cvZSp3cXM%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw06.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OeVHubGm0TT0PqApxh%2FT0VOvfHI%3D)\r\n\r\n### 分治规约\r\n- 核心思路：将全局归约分解为子问题，减少原子操作次数\r\n\r\n#### 1. warp 级规约\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw07.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kGTAFzREUZLYZ1RDVbeBriI97zg%3D)\r\n\r\n- Lane ID 获取\r\n- 实现：每个 warp 内单线程循环累加 + warp 间原子操作\r\n- 效果：相比全线程原子操作，性能提升约 **30倍**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw08.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ye5ZekG3BgE%2FVMGxyBxDEW3k7%2Bo%3D)\r\n\r\n\r\n#### 2. block 级规约的尝试与反思\r\n- 实现：每个 block 内单线程循环累加\r\n- 问题：活跃 warp 数下降，计算\u002F内存利用率降低\r\n- 结论：单纯减少原子操作数不等于最优，需平衡并行度\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw09.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ehnMhEZDsr4s5i2tuSeeP12H82g%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw10.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ijArwdKPKJ%2B%2BlsASqljdq%2BKzi2k%3D)\r\n\r\n\r\n\r\n## GPU 内存层级架构（Memory Hierarchy）\r\n### 内存墙问题与层级解决方案\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw11.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mkRNUfCDUaQ7tiJmY0BU8D1V%2BWs%3D)\r\n\r\n- 速度 vs 容量 vs 成本的权衡\r\n- 局部性原理：常用数据放上层，大规模数据放下层\r\n\r\n### GPU 内存结构\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw12.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZBqGfA25m73vhkzkVWXTOhPX218%3D)\r\n\r\n#### 1. 物理硬件单元\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw13.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5UpuM2neOYhbFtv7U74oBiODxEc%3D)\r\n\r\n#### 2. 逻辑内存模型\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw14.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=k4u7dohVQoV1AiNVaavP8WC%2B4f8%3D)\r\n\r\n\r\n## 高效规约算法演进\r\n### 基于 Shared Memory 的树状规约\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw15.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A6ZkpKmq8hqZBWaSBekJSZwjan0%3D)\r\n\r\n- 动态共享内存分配：`extern __shared__` + Kernel 启动参数配置\r\n- 块内同步：`__syncthreads()` 的作用与必要性\r\n- 性能：比 warp 级规约快 **5.8倍**，比 block 级规约快 **6.4倍**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw16.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZE7KoaukkoTY854V4ECITUV3Ifk%3D)\r\n\r\n### Warp 规约优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw17.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=UFGzmCvaEPj%2BTE53kQwpM%2FjM5Xk%3D)\r\n\r\n- `__shfl_down_sync`：寄存器级直接通信，无需 Shared Memory 和同步\r\n- 优势：更高计算强度（AI 提升至 1.5），更低延迟\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw18.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kGrHUoWMUl2WguMczp8nv%2FsuQ1Y%3D)\r\n\r\n\r\n### 混合规约策略（Warp Shuffle + Tree Reduce）\r\n- 步骤：Warp 内 Shuffle 规约 → 结果写入 Shared Memory → 块内树状规约\r\n- 编译器提示：`#pragma unroll` 循环展开\r\n- 性能进一步提升\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw19.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tX9PE9juTdFRidH1KKNtoD38V8o%3D)\r\n\r\n\r\n### 跨块规约优化\r\n#### 1. Two-Pass 方法：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw20.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fgcfD%2BIzL0CZxDz%2FMl937yJUKWk%3D)\r\n\r\n- Kernel1 块内规约写临时数组 → Kernel2 块间规约\r\n- 缺点：额外 Kernel 启动开销、隐式同步、额外访存\r\n\r\n#### 2. Cooperative Groups（CUDA 9.0+）\r\n- `grid.sync()`：跨 Block 同步\r\n- 启动要求：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw21.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JY6lAYaHjapM4PAr5vHWnQ221X8%3D)\r\n\r\n- 优势：可以减少多 Kernel 调度开销，在满足硬件资源约束且同步成本较低的场景下具有性能优势\r\n- 注意：死锁风险、硬件\u002F驱动版本限制\r\n\r\n#### 3. Two-Pass vs. Grid Sync\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw22.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E8%2BiXFFH05lMGEdwi3e%2BFKY72ys%3D)\r\n\r\n\r\n## Shared Memory Bank Conflict 深度解析\r\n### Bank Conflict 原理\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw23.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GHf25KSM3G%2FEmVV5VoZuh36RAb0%3D)\r\n\r\n- 在当前主流 NVIDIA GPU 架构中，共享内存划分为 32 Banks，每 Bank 4 Bytes\r\n- 冲突条件：同一 Warp 内多线程访问 **同一 Bank 的不同地址** → 会增加共享内存访问 transaction 数量，使访问被拆分执行，导致吞吐下降\r\n- 非冲突特例：同一 Bank 同一地址 → 广播（Broadcast）\r\n\r\n### 树状规约分析\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw24.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pgvL3Pae%2BloLfKcS%2FztPNwcN%2FHs%3D)\r\n\r\n\r\n### NCU 诊断 Bank Conflict\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw25.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZeGD21oigSutdVyMnr7GSLF5qus%3D)\r\n\r\n- Source 页面指标\r\n- Memory Workload Analysis\r\n\r\n### Padding 优化技巧（预告）\r\n- 通过内存填充改变 Bank 映射关系\r\n- 案例：矩阵转置中 Padding 带来 **2倍** 性能提升（下节课详解）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw26.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=726%2B9KkL70kHAqCDphh3dwmAmcs%3D)\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**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![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DxTm5MDZexUaY7e6ZSV7bC5jKbA%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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=T3zAwjNQevtf8RDwKi3412WjCPc%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-75","study-75","2026 夏季 InfiniTensor 训练营 AI 编译器 2：《AI 编译器中的前端优化》",33,"\u002Fstudy-material\u002Fstudy-76","study-76","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-77","study-77","2026 夏季 InfiniTensor 训练营 AI 编程与智能体应用 3：《Agent 在科研场景的应用实践》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw01.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=91EZXUb%2Fz4BAZxZjrrTbyBMuue8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 23 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第三课《内存模型与规约优化》开讲。\u003Cbr>前置课程《性能模型与逐元素优化》讲了 CUDA 并行编程中的“内存墙”与 Roofline 模型，演示了向量化访存、内存对齐与合并机制等内容。\u003Cbr>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课承接上节并行加法性能优化内容，从 Roofline 模型出发，深入探讨如何通过提高计算强度突破性能瓶颈。重点讲解并行规约中的原子操作陷阱、分治思想应用、GPU 内存层级架构，以及 Shared Memory Bank Conflict 等底层优化细节。\u003C\u002Fp>\n\u003Ch2 id=\"回顾与性能瓶颈分析\">\u003Ca class=\"anchor\" href=\"#回顾与性能瓶颈分析\" aria-label=\"跳转到标题：回顾与性能瓶颈分析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>回顾与性能瓶颈分析\u003C\u002Fh2>\u003Col>\n\u003Cli>\u003Cstrong>上节课优化路径回顾\u003C\u002Fstrong>\u003Cul>\n\u003Cli>CPU → GPU FP32 → 访存向量化（Fold4）→ 半精度量化\u003C\u002Fli>\n\u003Cli>数据量对性能表现的影响（小数据量 vs 大数据量）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Roofline 模型指导优化\u003C\u002Fstrong>\u003Cul>\n\u003Cli>目标：右移 Roofline 曲线，提高计算强度 $AI = W \u002F Q$\u003C\u002Fli>\n\u003Cli>策略：增加计算量 $W$（如算子融合）、减少访存量 $Q$\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw02.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=c2VxnvRyL%2F6Cq1X2j0KiZgxHF64%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>\u003Cstrong>从向量加法到累加求和\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw03.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FGiIA%2BlGk57fg62SmvI6xGLhxlc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw04.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DJjidHIKv6iuQUQbrYTPE9fcrHc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>尴尬并行（Embarrassingly Parallel）vs 数据竞争\u003C\u002Fli>\n\u003Cli>如何解决这个竞争的问题？\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"原子操作与初步规约优化\">\u003Ca class=\"anchor\" href=\"#原子操作与初步规约优化\" aria-label=\"跳转到标题：原子操作与初步规约优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>原子操作与初步规约优化\u003C\u002Fh2>\u003Ch3 id=\"原子操作-atomic-operations\">\u003Ca class=\"anchor\" href=\"#原子操作-atomic-operations\" aria-label=\"跳转到标题：原子操作（Atomic Operations）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>原子操作（Atomic Operations）\u003C\u002Fh3>\u003Cul>\n\u003Cli>概念：不可分割的最小执行单元（All-or-Nothing）\u003C\u002Fli>\n\u003Cli>CUDA API：\u003Ccode>atomicAdd\u003C\u002Fcode> 的使用与正确性验证\u003C\u002Fli>\n\u003Cli>性能代价：NCU 分析显示指令发射严重阻塞\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw05.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9R1FPybdg6xoaVOin8cvZSp3cXM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw06.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OeVHubGm0TT0PqApxh%2FT0VOvfHI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"分治规约\">\u003Ca class=\"anchor\" href=\"#分治规约\" aria-label=\"跳转到标题：分治规约\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>分治规约\u003C\u002Fh3>\u003Cul>\n\u003Cli>核心思路：将全局归约分解为子问题，减少原子操作次数\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"1-warp-级规约\">\u003Ca class=\"anchor\" href=\"#1-warp-级规约\" aria-label=\"跳转到标题：1. warp 级规约\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. warp 级规约\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw07.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kGTAFzREUZLYZ1RDVbeBriI97zg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Lane ID 获取\u003C\u002Fli>\n\u003Cli>实现：每个 warp 内单线程循环累加 + warp 间原子操作\u003C\u002Fli>\n\u003Cli>效果：相比全线程原子操作，性能提升约 \u003Cstrong>30倍\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw08.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ye5ZekG3BgE%2FVMGxyBxDEW3k7%2Bo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-block-级规约的尝试与反思\">\u003Ca class=\"anchor\" href=\"#2-block-级规约的尝试与反思\" aria-label=\"跳转到标题：2. block 级规约的尝试与反思\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. block 级规约的尝试与反思\u003C\u002Fh4>\u003Cul>\n\u003Cli>实现：每个 block 内单线程循环累加\u003C\u002Fli>\n\u003Cli>问题：活跃 warp 数下降，计算\u002F内存利用率降低\u003C\u002Fli>\n\u003Cli>结论：单纯减少原子操作数不等于最优，需平衡并行度\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw09.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ehnMhEZDsr4s5i2tuSeeP12H82g%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw10.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ijArwdKPKJ%2B%2BlsASqljdq%2BKzi2k%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"gpu-内存层级架构-memory-hierarchy\">\u003Ca class=\"anchor\" href=\"#gpu-内存层级架构-memory-hierarchy\" aria-label=\"跳转到标题：GPU 内存层级架构（Memory Hierarchy）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GPU 内存层级架构（Memory Hierarchy）\u003C\u002Fh2>\u003Ch3 id=\"内存墙问题与层级解决方案\">\u003Ca class=\"anchor\" href=\"#内存墙问题与层级解决方案\" aria-label=\"跳转到标题：内存墙问题与层级解决方案\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>内存墙问题与层级解决方案\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw11.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mkRNUfCDUaQ7tiJmY0BU8D1V%2BWs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>速度 vs 容量 vs 成本的权衡\u003C\u002Fli>\n\u003Cli>局部性原理：常用数据放上层，大规模数据放下层\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"gpu-内存结构\">\u003Ca class=\"anchor\" href=\"#gpu-内存结构\" aria-label=\"跳转到标题：GPU 内存结构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GPU 内存结构\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw12.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZBqGfA25m73vhkzkVWXTOhPX218%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-物理硬件单元\">\u003Ca class=\"anchor\" href=\"#1-物理硬件单元\" aria-label=\"跳转到标题：1. 物理硬件单元\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 物理硬件单元\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw13.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5UpuM2neOYhbFtv7U74oBiODxEc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-逻辑内存模型\">\u003Ca class=\"anchor\" href=\"#2-逻辑内存模型\" aria-label=\"跳转到标题：2. 逻辑内存模型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 逻辑内存模型\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw14.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=k4u7dohVQoV1AiNVaavP8WC%2B4f8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"高效规约算法演进\">\u003Ca class=\"anchor\" href=\"#高效规约算法演进\" aria-label=\"跳转到标题：高效规约算法演进\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>高效规约算法演进\u003C\u002Fh2>\u003Ch3 id=\"基于-shared-memory-的树状规约\">\u003Ca class=\"anchor\" href=\"#基于-shared-memory-的树状规约\" aria-label=\"跳转到标题：基于 Shared Memory 的树状规约\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>基于 Shared Memory 的树状规约\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw15.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A6ZkpKmq8hqZBWaSBekJSZwjan0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>动态共享内存分配：\u003Ccode>extern __shared__\u003C\u002Fcode> + Kernel 启动参数配置\u003C\u002Fli>\n\u003Cli>块内同步：\u003Ccode>__syncthreads()\u003C\u002Fcode> 的作用与必要性\u003C\u002Fli>\n\u003Cli>性能：比 warp 级规约快 \u003Cstrong>5.8倍\u003C\u002Fstrong>，比 block 级规约快 \u003Cstrong>6.4倍\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw16.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZE7KoaukkoTY854V4ECITUV3Ifk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"warp-规约优化\">\u003Ca class=\"anchor\" href=\"#warp-规约优化\" aria-label=\"跳转到标题：Warp 规约优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Warp 规约优化\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw17.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=UFGzmCvaEPj%2BTE53kQwpM%2FjM5Xk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Ccode>__shfl_down_sync\u003C\u002Fcode>：寄存器级直接通信，无需 Shared Memory 和同步\u003C\u002Fli>\n\u003Cli>优势：更高计算强度（AI 提升至 1.5），更低延迟\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw18.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kGrHUoWMUl2WguMczp8nv%2FsuQ1Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"混合规约策略-warp-shuffle-tree-reduce\">\u003Ca class=\"anchor\" href=\"#混合规约策略-warp-shuffle-tree-reduce\" aria-label=\"跳转到标题：混合规约策略（Warp Shuffle + Tree Reduce）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>混合规约策略（Warp Shuffle + Tree Reduce）\u003C\u002Fh3>\u003Cul>\n\u003Cli>步骤：Warp 内 Shuffle 规约 → 结果写入 Shared Memory → 块内树状规约\u003C\u002Fli>\n\u003Cli>编译器提示：\u003Ccode>#pragma unroll\u003C\u002Fcode> 循环展开\u003C\u002Fli>\n\u003Cli>性能进一步提升\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw19.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tX9PE9juTdFRidH1KKNtoD38V8o%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"跨块规约优化\">\u003Ca class=\"anchor\" href=\"#跨块规约优化\" aria-label=\"跳转到标题：跨块规约优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>跨块规约优化\u003C\u002Fh3>\u003Ch4 id=\"1-two-pass-方法\">\u003Ca class=\"anchor\" href=\"#1-two-pass-方法\" aria-label=\"跳转到标题：1. Two-Pass 方法：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. Two-Pass 方法：\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw20.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fgcfD%2BIzL0CZxDz%2FMl937yJUKWk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Kernel1 块内规约写临时数组 → Kernel2 块间规约\u003C\u002Fli>\n\u003Cli>缺点：额外 Kernel 启动开销、隐式同步、额外访存\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-cooperative-groups-cuda-9-0\">\u003Ca class=\"anchor\" href=\"#2-cooperative-groups-cuda-9-0\" aria-label=\"跳转到标题：2. Cooperative Groups（CUDA 9.0+）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Cooperative Groups（CUDA 9.0+）\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Ccode>grid.sync()\u003C\u002Fcode>：跨 Block 同步\u003C\u002Fli>\n\u003Cli>启动要求：\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw21.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JY6lAYaHjapM4PAr5vHWnQ221X8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>优势：可以减少多 Kernel 调度开销，在满足硬件资源约束且同步成本较低的场景下具有性能优势\u003C\u002Fli>\n\u003Cli>注意：死锁风险、硬件\u002F驱动版本限制\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-two-pass-vs-grid-sync\">\u003Ca class=\"anchor\" href=\"#3-two-pass-vs-grid-sync\" aria-label=\"跳转到标题：3. Two-Pass vs. Grid Sync\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. Two-Pass vs. Grid Sync\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw22.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E8%2BiXFFH05lMGEdwi3e%2BFKY72ys%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"shared-memory-bank-conflict-深度解析\">\u003Ca class=\"anchor\" href=\"#shared-memory-bank-conflict-深度解析\" aria-label=\"跳转到标题：Shared Memory Bank Conflict 深度解析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Shared Memory Bank Conflict 深度解析\u003C\u002Fh2>\u003Ch3 id=\"bank-conflict-原理\">\u003Ca class=\"anchor\" href=\"#bank-conflict-原理\" aria-label=\"跳转到标题：Bank Conflict 原理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Bank Conflict 原理\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw23.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GHf25KSM3G%2FEmVV5VoZuh36RAb0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>在当前主流 NVIDIA GPU 架构中，共享内存划分为 32 Banks，每 Bank 4 Bytes\u003C\u002Fli>\n\u003Cli>冲突条件：同一 Warp 内多线程访问 \u003Cstrong>同一 Bank 的不同地址\u003C\u002Fstrong> → 会增加共享内存访问 transaction 数量，使访问被拆分执行，导致吞吐下降\u003C\u002Fli>\n\u003Cli>非冲突特例：同一 Bank 同一地址 → 广播（Broadcast）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"树状规约分析\">\u003Ca class=\"anchor\" href=\"#树状规约分析\" aria-label=\"跳转到标题：树状规约分析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>树状规约分析\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw24.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pgvL3Pae%2BloLfKcS%2FztPNwcN%2FHs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"ncu-诊断-bank-conflict\">\u003Ca class=\"anchor\" href=\"#ncu-诊断-bank-conflict\" aria-label=\"跳转到标题：NCU 诊断 Bank Conflict\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>NCU 诊断 Bank Conflict\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw25.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZeGD21oigSutdVyMnr7GSLF5qus%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Source 页面指标\u003C\u002Fli>\n\u003Cli>Memory Workload Analysis\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"padding-优化技巧-预告\">\u003Ca class=\"anchor\" href=\"#padding-优化技巧-预告\" aria-label=\"跳转到标题：Padding 优化技巧（预告）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Padding 优化技巧（预告）\u003C\u002Fh3>\u003Cul>\n\u003Cli>通过内存填充改变 Bank 映射关系\u003C\u002Fli>\n\u003Cli>案例：矩阵转置中 Padding 带来 \u003Cstrong>2倍\u003C\u002Fstrong> 性能提升（下节课详解）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_76_zw26.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=726%2B9KkL70kHAqCDphh3dwmAmcs%3D\" alt=\"\">\u003C\u002Fp>\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\u002Fportal_news\u002F2026-xjq.png?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DxTm5MDZexUaY7e6ZSV7bC5jKbA%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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=T3zAwjNQevtf8RDwKi3412WjCPc%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,132,135,137,141,144,147,149,152,155,158,160,163,166,169,171,174,177,180,183,186,188,191,194,196,199],{"id":129,"text":129,"level":110},"回顾与性能瓶颈分析",{"id":131,"text":131,"level":110},"原子操作与初步规约优化",{"id":133,"text":134,"level":9},"原子操作-atomic-operations","原子操作（Atomic Operations）",{"id":136,"text":136,"level":9},"分治规约",{"id":138,"text":139,"level":140},"1-warp-级规约","1. warp 级规约",4,{"id":142,"text":143,"level":140},"2-block-级规约的尝试与反思","2. block 级规约的尝试与反思",{"id":145,"text":146,"level":110},"gpu-内存层级架构-memory-hierarchy","GPU 内存层级架构（Memory Hierarchy）",{"id":148,"text":148,"level":9},"内存墙问题与层级解决方案",{"id":150,"text":151,"level":9},"gpu-内存结构","GPU 内存结构",{"id":153,"text":154,"level":140},"1-物理硬件单元","1. 物理硬件单元",{"id":156,"text":157,"level":140},"2-逻辑内存模型","2. 逻辑内存模型",{"id":159,"text":159,"level":110},"高效规约算法演进",{"id":161,"text":162,"level":9},"基于-shared-memory-的树状规约","基于 Shared Memory 的树状规约",{"id":164,"text":165,"level":9},"warp-规约优化","Warp 规约优化",{"id":167,"text":168,"level":9},"混合规约策略-warp-shuffle-tree-reduce","混合规约策略（Warp Shuffle + Tree Reduce）",{"id":170,"text":170,"level":9},"跨块规约优化",{"id":172,"text":173,"level":140},"1-two-pass-方法","1. Two-Pass 方法：",{"id":175,"text":176,"level":140},"2-cooperative-groups-cuda-9-0","2. Cooperative Groups（CUDA 9.0+）",{"id":178,"text":179,"level":140},"3-two-pass-vs-grid-sync","3. Two-Pass vs. Grid Sync",{"id":181,"text":182,"level":110},"shared-memory-bank-conflict-深度解析","Shared Memory Bank Conflict 深度解析",{"id":184,"text":185,"level":9},"bank-conflict-原理","Bank Conflict 原理",{"id":187,"text":187,"level":9},"树状规约分析",{"id":189,"text":190,"level":9},"ncu-诊断-bank-conflict","NCU 诊断 Bank Conflict",{"id":192,"text":193,"level":9},"padding-优化技巧-预告","Padding 优化技巧（预告）",{"id":195,"text":195,"level":110},"加入训练营",{"id":197,"text":198,"level":140},"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：",{"id":200,"text":200,"level":110},"关注与交流"]