[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-31":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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uiV9kOffa9Xq8DT%2BbErZU0JyQ98%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8u%2BaQaeliN4QzF6MZWiB4jHYDCE%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8pXXbnSlNEtVJCJ1e4RzUlqmRn8%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_31_fm.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qGL1eO3fPIgdVnIcYjMWfp0iOoo%3D","InfiniTensor,冬季训练营,CUDA,分块优化,不规则访存","讲解 CUDA 分块与不规则访存优化方法，帮助定位访存瓶颈并提升并行计算效率。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营 CUDA 编程第四课《分块与不规则访存》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw01.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=iYLMm1FHVCkjCGEJ3zUaTcI3RBE%3D)\r\n\r\n\r\n2026 年 1 月 22 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第四课《分块与不规则访存》开讲。\r\n\r\n### 矩阵乘法（GEMM）基础\r\n1. 矩阵乘法实现方式\r\n2. 矩阵乘法计算复杂度分析：\r\nO(n³) 计算 vs O(n²) 访存\r\n3. 行优先（Row Major）与列优先（Column Major）存储格式\r\n4. GEMM 通用矩阵乘法规范（转置选项、缩放因子Alpha\u002FBeta）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw02.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hROrUk4oP0BXmqfttPA1vvm2F78%3D)\r\n\r\n### 基础矩阵乘法实现与性能分析\r\n1. 简单矩阵乘法 Kernel 实现\r\n2. 内存合并访问的重要性\r\n3. NCU 性能分析工具使用\r\n4. Roofline 模型分析计算密集型特征\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw03.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=onRFgRMsRz%2FbX8Uc7sk4rCbmWBg%3D)\r\n\r\n\r\n### 共享内存优化技术\r\n1. 数据复用与共享内存缓存\r\n2. 分块（Tiling）技术原理\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw04.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ct55odp4xrNvfRVd4BTjWCFVOWs%3D)\r\n\r\n3. Tile 大小选择策略（32 的合理性）\r\n4. 同步机制保证数据完整性\r\n\r\n### 分级分块优化策略\r\n1. 1D Block Tiling优化原理\r\n\r\n把单纯使用共享内存（SMEM Caching）进阶到一维分块（1D Blocktiling）后的性能变化作对比，主要包含以下关键发现：\r\n\r\n* **性能提升指标**：通过 Nsight Compute (NCU) 的 Profiling 结果可以看出，采用 1D Blocktiling 后，**L1 缓存的命中率（Hit Rate）增加**。\r\n* **带宽利用优化**：缓存命中率的提升直接带动了内存带宽有效利用率的增加，从而提升了整体计算吞吐量。\r\n* **进一步优化的思考**：图片底部提出了从 1D 向 2D 演进的核心问题 —— 既然在列方向（Column）进行拓展（1D）能带来收益，那么是否可以继续在行方向（Row）进行拓展？这也引出了后续 **2D Block Tiling** 的优化思路。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw05.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ujjOMX9aSBgY8rA556rn5ch3XAg%3D)\r\n\r\n2. 2D Block Tiling实现方法\r\n访存数量优化分析（9次→5次\u002Fentry）\r\n\r\n下图详细对比了 **SMEM Caching** 与 **2D Block Tiling** 在访存效率上的显著差异，直观展示了二维分块如何通过增加寄存器层面的数据复用来降低全局内存访问压力：\r\n\r\n* **SMEM Caching 模式**：在仅使用共享内存缓存的情况下，计算每个结果元素（Entry）平均需要 **9 次 Load 操作**（源于对 A 和 B 的重复读取以及 C 的读写）。\r\n* **2D Block Tiling 模式**：通过在寄存器层面引入二维分块（如图中 Tile A 和 Tile B 的配合，一次计算 Tile C 中的多个元素），数据复用率大幅提升。图示表明，该策略将计算每个 Entry 的平均访存开销显著降低至 **5 次 Load 操作**。\r\n* **核心收益**：这种策略的核心在于“分摊开销”。通过一次性加载 A 的一小块和 B 的一小块到寄存器，并计算出 C 的对应子块（如 $2\times2$），使得加载到寄存器的数据能在多次计算指令中被重复使用，从而突破了带宽瓶颈。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw06.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Bvbuyrk1bcL2nGn29cAz3NHSVbU%3D)\r\n\r\n3. Bank Conflict 在矩阵乘法中的出现与解决\r\nSwizzling 技术\r\n\r\n### 高级优化技术\r\n1. 向量化访存优化\r\n2. WarpTiling 技术\r\n3. 与 cuBLAS 性能对比（达到 82.6%）\r\n4. 不同矩阵尺寸下的性能变化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw07.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Jvl2ZvmqWZxqz1rqZc6V6rxefpU%3D)\r\n\r\n### 不规则访存与高级主题\r\n1. Z-order Curve（Morton码）原理\r\n2. 空间局部性保证技术\r\n3. 线性包围体层次（LBVH）算法\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw08.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=l6IvsmmmQLYc6owV1SoF%2F08D0%2FI%3D)\r\n\r\n4. 光线追踪中的空间数据结构\r\n5. 不规则访存优化策略\r\n6. 动态并行（CUDA Dynamic Parallelism）概念与应用场景\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw09.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bFhnGhXI0UnE7KGR2I7UAvt5BzY%3D)\r\n\r\n\r\n### 总结\r\n本节课主要讲解了 CUDA 矩阵乘法（GEMM）的优化技术，从基础实现逐步深入到高级优化方法，下节课将讲解进阶的时空优化技术和系统级优化方法。\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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DEth0OhERyx68xufLNiv%2BIt7foE%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![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw7.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OXZH4IYHRMHbFPG7EYCBYHhSMto%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-30","study-30","InfiniTensor 冬季训练营 CUDA 编程第四课《CUDA 程序迁移至天数智芯 GPU》",78,"\u002Fstudy-material\u002Fstudy-31","study-31","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-32","study-32","InfiniTensor 冬季训练营 OpenCL 编程第二课《OpenCL 编程抽象与语法》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw01.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=iYLMm1FHVCkjCGEJ3zUaTcI3RBE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 22 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第四课《分块与不规则访存》开讲。\u003C\u002Fp>\n\u003Ch3 id=\"矩阵乘法-gemm-基础\">\u003Ca class=\"anchor\" href=\"#矩阵乘法-gemm-基础\" aria-label=\"跳转到标题：矩阵乘法（GEMM）基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>矩阵乘法（GEMM）基础\u003C\u002Fh3>\u003Col>\n\u003Cli>矩阵乘法实现方式\u003C\u002Fli>\n\u003Cli>矩阵乘法计算复杂度分析：\u003Cbr>O(n³) 计算 vs O(n²) 访存\u003C\u002Fli>\n\u003Cli>行优先（Row Major）与列优先（Column Major）存储格式\u003C\u002Fli>\n\u003Cli>GEMM 通用矩阵乘法规范（转置选项、缩放因子Alpha\u002FBeta）\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw02.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hROrUk4oP0BXmqfttPA1vvm2F78%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"基础矩阵乘法实现与性能分析\">\u003Ca class=\"anchor\" href=\"#基础矩阵乘法实现与性能分析\" aria-label=\"跳转到标题：基础矩阵乘法实现与性能分析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>基础矩阵乘法实现与性能分析\u003C\u002Fh3>\u003Col>\n\u003Cli>简单矩阵乘法 Kernel 实现\u003C\u002Fli>\n\u003Cli>内存合并访问的重要性\u003C\u002Fli>\n\u003Cli>NCU 性能分析工具使用\u003C\u002Fli>\n\u003Cli>Roofline 模型分析计算密集型特征\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw03.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=onRFgRMsRz%2FbX8Uc7sk4rCbmWBg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"共享内存优化技术\">\u003Ca class=\"anchor\" href=\"#共享内存优化技术\" aria-label=\"跳转到标题：共享内存优化技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>共享内存优化技术\u003C\u002Fh3>\u003Col>\n\u003Cli>数据复用与共享内存缓存\u003C\u002Fli>\n\u003Cli>分块（Tiling）技术原理\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw04.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ct55odp4xrNvfRVd4BTjWCFVOWs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>Tile 大小选择策略（32 的合理性）\u003C\u002Fli>\n\u003Cli>同步机制保证数据完整性\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"分级分块优化策略\">\u003Ca class=\"anchor\" href=\"#分级分块优化策略\" aria-label=\"跳转到标题：分级分块优化策略\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>分级分块优化策略\u003C\u002Fh3>\u003Col>\n\u003Cli>1D Block Tiling优化原理\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>把单纯使用共享内存（SMEM Caching）进阶到一维分块（1D Blocktiling）后的性能变化作对比，主要包含以下关键发现：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>性能提升指标\u003C\u002Fstrong>：通过 Nsight Compute (NCU) 的 Profiling 结果可以看出，采用 1D Blocktiling 后，\u003Cstrong>L1 缓存的命中率（Hit Rate）增加\u003C\u002Fstrong>。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>带宽利用优化\u003C\u002Fstrong>：缓存命中率的提升直接带动了内存带宽有效利用率的增加，从而提升了整体计算吞吐量。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>进一步优化的思考\u003C\u002Fstrong>：图片底部提出了从 1D 向 2D 演进的核心问题 —— 既然在列方向（Column）进行拓展（1D）能带来收益，那么是否可以继续在行方向（Row）进行拓展？这也引出了后续 \u003Cstrong>2D Block Tiling\u003C\u002Fstrong> 的优化思路。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw05.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ujjOMX9aSBgY8rA556rn5ch3XAg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>2D Block Tiling实现方法\u003Cbr>访存数量优化分析（9次→5次\u002Fentry）\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>下图详细对比了 \u003Cstrong>SMEM Caching\u003C\u002Fstrong> 与 \u003Cstrong>2D Block Tiling\u003C\u002Fstrong> 在访存效率上的显著差异，直观展示了二维分块如何通过增加寄存器层面的数据复用来降低全局内存访问压力：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>SMEM Caching 模式\u003C\u002Fstrong>：在仅使用共享内存缓存的情况下，计算每个结果元素（Entry）平均需要 \u003Cstrong>9 次 Load 操作\u003C\u002Fstrong>（源于对 A 和 B 的重复读取以及 C 的读写）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>2D Block Tiling 模式\u003C\u002Fstrong>：通过在寄存器层面引入二维分块（如图中 Tile A 和 Tile B 的配合，一次计算 Tile C 中的多个元素），数据复用率大幅提升。图示表明，该策略将计算每个 Entry 的平均访存开销显著降低至 \u003Cstrong>5 次 Load 操作\u003C\u002Fstrong>。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>核心收益\u003C\u002Fstrong>：这种策略的核心在于“分摊开销”。通过一次性加载 A 的一小块和 B 的一小块到寄存器，并计算出 C 的对应子块（如 $2 imes2$），使得加载到寄存器的数据能在多次计算指令中被重复使用，从而突破了带宽瓶颈。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw06.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Bvbuyrk1bcL2nGn29cAz3NHSVbU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>Bank Conflict 在矩阵乘法中的出现与解决\u003Cbr>Swizzling 技术\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"高级优化技术\">\u003Ca class=\"anchor\" href=\"#高级优化技术\" aria-label=\"跳转到标题：高级优化技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>高级优化技术\u003C\u002Fh3>\u003Col>\n\u003Cli>向量化访存优化\u003C\u002Fli>\n\u003Cli>WarpTiling 技术\u003C\u002Fli>\n\u003Cli>与 cuBLAS 性能对比（达到 82.6%）\u003C\u002Fli>\n\u003Cli>不同矩阵尺寸下的性能变化\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw07.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Jvl2ZvmqWZxqz1rqZc6V6rxefpU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"不规则访存与高级主题\">\u003Ca class=\"anchor\" href=\"#不规则访存与高级主题\" aria-label=\"跳转到标题：不规则访存与高级主题\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>不规则访存与高级主题\u003C\u002Fh3>\u003Col>\n\u003Cli>Z-order Curve（Morton码）原理\u003C\u002Fli>\n\u003Cli>空间局部性保证技术\u003C\u002Fli>\n\u003Cli>线性包围体层次（LBVH）算法\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw08.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=l6IvsmmmQLYc6owV1SoF%2F08D0%2FI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"4\">\n\u003Cli>光线追踪中的空间数据结构\u003C\u002Fli>\n\u003Cli>不规则访存优化策略\u003C\u002Fli>\n\u003Cli>动态并行（CUDA Dynamic Parallelism）概念与应用场景\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_31_zw09.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bFhnGhXI0UnE7KGR2I7UAvt5BzY%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>本节课主要讲解了 CUDA 矩阵乘法（GEMM）的优化技术，从基础实现逐步深入到高级优化方法，下节课将讲解进阶的时空优化技术和系统级优化方法。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DEth0OhERyx68xufLNiv%2BIt7foE%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OXZH4IYHRMHbFPG7EYCBYHhSMto%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,133,135,137,139,141,143,146],{"id":129,"text":130,"level":9},"矩阵乘法-gemm-基础","矩阵乘法（GEMM）基础",{"id":132,"text":132,"level":9},"基础矩阵乘法实现与性能分析",{"id":134,"text":134,"level":9},"共享内存优化技术",{"id":136,"text":136,"level":9},"分级分块优化策略",{"id":138,"text":138,"level":9},"高级优化技术",{"id":140,"text":140,"level":9},"不规则访存与高级主题",{"id":142,"text":142,"level":9},"总结",{"id":144,"text":145,"level":9},"学习贴士","💡 学习贴士",{"id":147,"text":148,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]