[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-92":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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BwVzWBmdtkq8YPZLqXna4E1FT3M%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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E67XHLL7db9GOlGHdVrBAdjqtXU%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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZDQtqBSR0N65mJrj3FTJxyAPi6M%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_92_fm.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7hykzZj%2F4sHPwJD4o7r8CGHLDvg%3D","OpenCL,算子优化,GEMM,Subgroup,量化推理","详解 OpenCL Argmax、Softmax、GEMM\u002FGEMV 算子开发，介绍子组、向量化、合并访存、GGUF 量化等核心优化手段，覆盖大模型推理高频算子落地方案。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 OpenCL 编程 3：《OpenCL 算子的实现与优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw01.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eGqpJeWzZjxnNXxe%2Fs2Bfs5jYok%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">8 月 4 日，2026 夏季 InfiniTensor 训练营 OpenCL 编程第三课《OpenCL 算子的实现与优化》开讲。\r\n上一课讲解 OpenCL 四大抽象模型、NDRange 索引、局部内存同步、树状归约等设备端基础语法，为本节课各类算子落地与性能调优打下底层基础。\r\n\r\n本节课围绕大模型推理高频算子展开，完整讲解 Argmax、Softmax、GEMV \u002F GEMM 标准实现，同时系统介绍子组、向量化、量化、访存合并等全套性能优化手段。\u003C\u002Fspan>\r\n\r\n## 核心归约算子实现\r\n### Argmax 最大值索引算子\r\n#### 1. 算子作用与业务场景\r\n输入一组数值，输出全局最大值对应的索引；常用于分类输出、大模型 Token 采样筛选。\r\n\r\n#### 2. 两种实现方案\r\n(1). 单工作组树状归约\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bp62vdvT2QgwUBUSEOnorIsAGuw%3D)\r\n\r\n(2). 多工作组分阶段归约\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw03.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xix4hTibw5aoqSjaCJGEGzfTuJY%3D)\r\n\r\n- 多工作组并行计算各组局部最大值+索引，写回全局内存；\r\n- Host 端判断局部结果数量，少量则主机完成最终比较，量大则新开 Kernel 二次归约；\r\n- 优势：充分利用多 CU 计算单元，适配超大数组场景。\r\n\r\n### Softmax 归一化算子\r\n原始公式易指数溢出，统一先减去维度内最大值再计算指数，缩小数值区间。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw04.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FyPafcTRw14ztJyMsTdKsCjuiSY%3D)\r\n\r\n#### 1. 一维 Softmax 实现\r\n(1). 小规模（单工作组可容纳）：\r\n- 工作组内先归约求全局最大值 → 各线程减最大值计算指数 → 归约求指数总和 → 每个元素除以总和输出。\r\n\r\n(2). 大规模多段实现：\r\n- 多工作组分段求局部最大值、局部指数和，Host 汇总全局最值与总和后二次 Kernel 完成归一化。\r\n\r\n#### 2. 二维 Softmax 两种轴实现\r\n- Axis = 0（按列归一）：每列独立执行 Softmax 流程；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw05.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EQO%2FVHbGZ9BvgI1%2BBPrTjFxD7o4%3D)\r\n\r\n- Axis = 1（按行归一）：每行分配一个工作组，每行内完成归约与归一，实现更简洁。\r\n\r\n## 基础矩阵乘法 GEMM\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw06.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eBwf4B1GxFIcOzDa0Dj2iEbLgxM%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw07.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=nOoeG6obk9N0viHcjdU7q%2B6CzA8%3D)\r\n\r\n### 基础实现\r\n1.朴素方法：一个线程计算 C 矩阵一个元素。\r\n\r\n2.性能痛点：朴素 GEMM 中，每个线程独立计算一个输出元素，会导致输入矩阵元素被大量重复从全局内存加载，数据复用率低；同时矩阵列方向访问可能造成非连续访存，降低带宽利用率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw08.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uyBBl1Qg3FaQI1YZxAWgAu2SlKc%3D)\r\n\r\n\r\n## GEMV 矩阵向量乘性能优化\r\n### 子组优化（Subgroup）\r\n1.**概念**：硬件级线程束（NVIDIA Warp \u002F AMD Wavefront \u002F Intel Subgroup），支持寄存器级低延迟通信。\r\n\r\n2.**优势**：使用 `sub_group_reduce_add` 等内置指令替代 Local Memory + Barrier，显著降低同步开销。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw09.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=zPsyA46UvgDJMsNegzVLJmBjviw%3D)\r\n\r\n3.**合并内存访问**（Coalesced Access）：设计让相邻线程访问相邻地址，将多次内存事务合并为单次缓存行读取，对各种 GPU 架构均具有重要优化价值，尤其适用于带宽受限 Kernel。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw10.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PBp7GfXKSr%2FKZkUpFLAHPreGlUg%3D)\r\n\r\n### 向量化加载与计算\r\n1.**技术**：使用 `vloadN` 及向量类型（如 half8, float4）一次性加载\u002F计算多个元素。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw11.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LzLQ2cyZi9dA9MXqaWfUTMYfXuY%3D)\r\n\r\n2.**收益**：向量化加载可以减少访存操作调用次数，并提高数据吞吐。\r\n\r\n3.**注意**：必须保证内存地址对齐，否则会退化为多次访问；需关注寄存器压力，避免溢出到全局内存。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw12.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Vocmxafv5ZB3rE70qe6EEaZDZa8%3D)\r\n\r\n\r\n### 单工作项处理多行数据\r\n1.**适用场景**：大规模数据下硬件资源已饱和时。\r\n\r\n2.**原理**：减少线程调度开销，利用缓存局部性，提高指令级并行度。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw13.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uX67yuEwV%2BQm5F1tngFLOj5eE%2Fc%3D)\r\n\r\n3.**权衡**：行数过多会导致寄存器消耗线性增长，需防止寄存器溢出导致的性能断崖。\r\n\r\n\r\n## Gemv 量化\r\n### 量化基础\r\n1.**动机**：量化通过降低权重表示精度减少存储空间和内存带宽压力。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw14.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cgb2hyy7YY0cj7kz9y5tPQ0eiO4%3D)\r\n\r\n2.**GGUF 量化**：\r\n- **传统量化**：QX_Y（X=位宽，0=对称，1=非对称）。Q8-0 为 32 元素一组 + 1个缩放因子。\r\n\r\n- **K 系列**：层次化量化，对缩放因子再量化，进一步压缩元数据。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw15.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=98Iw1Ru0f8ezKvqckB54UV%2FVKiE%3D)\r\n\r\n- **IQ 系列**：引入重要性矩阵（Importance Matrix）指导量化。\r\n\r\n### GEMV Q8-0 量化实现\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw16.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wrGTl3NVq4iUJdUzrQDhRIrMXaU%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw17.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kMplHSZWX2fFSDbXuhjQEwHq8eY%3D)\r\n\r\n\r\n### Gemv 量化实现——应用纹理内存\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw18.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2F6WoFFGfj8%2BKpgHL7UpaeNNcbe4%3D)\r\n\r\n- 纹理内存（Image Object）：端侧 GPU 专用只读缓存，对二维空间局部性访问有硬件级优化，可以作为 Buffer 之外的一种存储方式。\r\n\r\n## 课程课后作业\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw19.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ESJ9aIuU%2FPt4lN8VD1U8v0Us1TM%3D)\r\n\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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BXH4UNzWFOrSPQr6JCB10ba6f1k%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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2FSVGJ8eOraf6k1nlkjH0x9NZjY%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-91","study-91","2026 夏季 InfiniTensor 训练营 AI 编译器 3：《AI 编译器中的后端优化》",18,"\u002Fstudy-material\u002Fstudy-92","study-92","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-93","study-93","2026 夏季 InfiniTensor 训练营大咖课 1：天数智芯 AI 与加速计算负责人单天逸——《AI and AI Chip》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw01.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eGqpJeWzZjxnNXxe%2Fs2Bfs5jYok%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">8 月 4 日，2026 夏季 InfiniTensor 训练营 OpenCL 编程第三课《OpenCL 算子的实现与优化》开讲。\u003Cbr>上一课讲解 OpenCL 四大抽象模型、NDRange 索引、局部内存同步、树状归约等设备端基础语法，为本节课各类算子落地与性能调优打下底层基础。\u003C\u002Fp>\n\u003Cp>本节课围绕大模型推理高频算子展开，完整讲解 Argmax、Softmax、GEMV \u002F GEMM 标准实现，同时系统介绍子组、向量化、量化、访存合并等全套性能优化手段。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Ch2 id=\"核心归约算子实现\">\u003Ca class=\"anchor\" href=\"#核心归约算子实现\" aria-label=\"跳转到标题：核心归约算子实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心归约算子实现\u003C\u002Fh2>\u003Ch3 id=\"argmax-最大值索引算子\">\u003Ca class=\"anchor\" href=\"#argmax-最大值索引算子\" aria-label=\"跳转到标题：Argmax 最大值索引算子\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Argmax 最大值索引算子\u003C\u002Fh3>\u003Ch4 id=\"1-算子作用与业务场景\">\u003Ca class=\"anchor\" href=\"#1-算子作用与业务场景\" aria-label=\"跳转到标题：1. 算子作用与业务场景\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 算子作用与业务场景\u003C\u002Fh4>\u003Cp>输入一组数值，输出全局最大值对应的索引；常用于分类输出、大模型 Token 采样筛选。\u003C\u002Fp>\n\u003Ch4 id=\"2-两种实现方案\">\u003Ca class=\"anchor\" href=\"#2-两种实现方案\" aria-label=\"跳转到标题：2. 两种实现方案\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 两种实现方案\u003C\u002Fh4>\u003Cp>(1). 单工作组树状归约\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bp62vdvT2QgwUBUSEOnorIsAGuw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>(2). 多工作组分阶段归约\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw03.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xix4hTibw5aoqSjaCJGEGzfTuJY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>多工作组并行计算各组局部最大值+索引，写回全局内存；\u003C\u002Fli>\n\u003Cli>Host 端判断局部结果数量，少量则主机完成最终比较，量大则新开 Kernel 二次归约；\u003C\u002Fli>\n\u003Cli>优势：充分利用多 CU 计算单元，适配超大数组场景。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"softmax-归一化算子\">\u003Ca class=\"anchor\" href=\"#softmax-归一化算子\" aria-label=\"跳转到标题：Softmax 归一化算子\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Softmax 归一化算子\u003C\u002Fh3>\u003Cp>原始公式易指数溢出，统一先减去维度内最大值再计算指数，缩小数值区间。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw04.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FyPafcTRw14ztJyMsTdKsCjuiSY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-一维-softmax-实现\">\u003Ca class=\"anchor\" href=\"#1-一维-softmax-实现\" aria-label=\"跳转到标题：1. 一维 Softmax 实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 一维 Softmax 实现\u003C\u002Fh4>\u003Cp>(1). 小规模（单工作组可容纳）：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>工作组内先归约求全局最大值 → 各线程减最大值计算指数 → 归约求指数总和 → 每个元素除以总和输出。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>(2). 大规模多段实现：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>多工作组分段求局部最大值、局部指数和，Host 汇总全局最值与总和后二次 Kernel 完成归一化。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-二维-softmax-两种轴实现\">\u003Ca class=\"anchor\" href=\"#2-二维-softmax-两种轴实现\" aria-label=\"跳转到标题：2. 二维 Softmax 两种轴实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 二维 Softmax 两种轴实现\u003C\u002Fh4>\u003Cul>\n\u003Cli>Axis = 0（按列归一）：每列独立执行 Softmax 流程；\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw05.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EQO%2FVHbGZ9BvgI1%2BBPrTjFxD7o4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Axis = 1（按行归一）：每行分配一个工作组，每行内完成归约与归一，实现更简洁。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"基础矩阵乘法-gemm\">\u003Ca class=\"anchor\" href=\"#基础矩阵乘法-gemm\" aria-label=\"跳转到标题：基础矩阵乘法 GEMM\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>基础矩阵乘法 GEMM\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw06.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eBwf4B1GxFIcOzDa0Dj2iEbLgxM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw07.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=nOoeG6obk9N0viHcjdU7q%2B6CzA8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"基础实现\">\u003Ca class=\"anchor\" href=\"#基础实现\" aria-label=\"跳转到标题：基础实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>基础实现\u003C\u002Fh3>\u003Cp>1.朴素方法：一个线程计算 C 矩阵一个元素。\u003C\u002Fp>\n\u003Cp>2.性能痛点：朴素 GEMM 中，每个线程独立计算一个输出元素，会导致输入矩阵元素被大量重复从全局内存加载，数据复用率低；同时矩阵列方向访问可能造成非连续访存，降低带宽利用率。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw08.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uyBBl1Qg3FaQI1YZxAWgAu2SlKc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"gemv-矩阵向量乘性能优化\">\u003Ca class=\"anchor\" href=\"#gemv-矩阵向量乘性能优化\" aria-label=\"跳转到标题：GEMV 矩阵向量乘性能优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GEMV 矩阵向量乘性能优化\u003C\u002Fh2>\u003Ch3 id=\"子组优化-subgroup\">\u003Ca class=\"anchor\" href=\"#子组优化-subgroup\" aria-label=\"跳转到标题：子组优化（Subgroup）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>子组优化（Subgroup）\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>概念\u003C\u002Fstrong>：硬件级线程束（NVIDIA Warp \u002F AMD Wavefront \u002F Intel Subgroup），支持寄存器级低延迟通信。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>优势\u003C\u002Fstrong>：使用 \u003Ccode>sub_group_reduce_add\u003C\u002Fcode> 等内置指令替代 Local Memory + Barrier，显著降低同步开销。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw09.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=zPsyA46UvgDJMsNegzVLJmBjviw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>合并内存访问\u003C\u002Fstrong>（Coalesced Access）：设计让相邻线程访问相邻地址，将多次内存事务合并为单次缓存行读取，对各种 GPU 架构均具有重要优化价值，尤其适用于带宽受限 Kernel。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw10.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PBp7GfXKSr%2FKZkUpFLAHPreGlUg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"向量化加载与计算\">\u003Ca class=\"anchor\" href=\"#向量化加载与计算\" aria-label=\"跳转到标题：向量化加载与计算\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>向量化加载与计算\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>技术\u003C\u002Fstrong>：使用 \u003Ccode>vloadN\u003C\u002Fcode> 及向量类型（如 half8, float4）一次性加载\u002F计算多个元素。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw11.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LzLQ2cyZi9dA9MXqaWfUTMYfXuY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>收益\u003C\u002Fstrong>：向量化加载可以减少访存操作调用次数，并提高数据吞吐。\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>注意\u003C\u002Fstrong>：必须保证内存地址对齐，否则会退化为多次访问；需关注寄存器压力，避免溢出到全局内存。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw12.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Vocmxafv5ZB3rE70qe6EEaZDZa8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"单工作项处理多行数据\">\u003Ca class=\"anchor\" href=\"#单工作项处理多行数据\" aria-label=\"跳转到标题：单工作项处理多行数据\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>单工作项处理多行数据\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>适用场景\u003C\u002Fstrong>：大规模数据下硬件资源已饱和时。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>原理\u003C\u002Fstrong>：减少线程调度开销，利用缓存局部性，提高指令级并行度。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw13.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uX67yuEwV%2BQm5F1tngFLOj5eE%2Fc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>权衡\u003C\u002Fstrong>：行数过多会导致寄存器消耗线性增长，需防止寄存器溢出导致的性能断崖。\u003C\u002Fp>\n\u003Ch2 id=\"gemv-量化\">\u003Ca class=\"anchor\" href=\"#gemv-量化\" aria-label=\"跳转到标题：Gemv 量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Gemv 量化\u003C\u002Fh2>\u003Ch3 id=\"量化基础\">\u003Ca class=\"anchor\" href=\"#量化基础\" aria-label=\"跳转到标题：量化基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>量化基础\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>动机\u003C\u002Fstrong>：量化通过降低权重表示精度减少存储空间和内存带宽压力。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw14.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cgb2hyy7YY0cj7kz9y5tPQ0eiO4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>GGUF 量化\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cp>\u003Cstrong>传统量化\u003C\u002Fstrong>：QX_Y（X=位宽，0=对称，1=非对称）。Q8-0 为 32 元素一组 + 1个缩放因子。\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>K 系列\u003C\u002Fstrong>：层次化量化，对缩放因子再量化，进一步压缩元数据。\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw15.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=98Iw1Ru0f8ezKvqckB54UV%2FVKiE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>IQ 系列\u003C\u002Fstrong>：引入重要性矩阵（Importance Matrix）指导量化。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"gemv-q8-0-量化实现\">\u003Ca class=\"anchor\" href=\"#gemv-q8-0-量化实现\" aria-label=\"跳转到标题：GEMV Q8-0 量化实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GEMV Q8-0 量化实现\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw16.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wrGTl3NVq4iUJdUzrQDhRIrMXaU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw17.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kMplHSZWX2fFSDbXuhjQEwHq8eY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"gemv-量化实现-应用纹理内存\">\u003Ca class=\"anchor\" href=\"#gemv-量化实现-应用纹理内存\" aria-label=\"跳转到标题：Gemv 量化实现——应用纹理内存\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Gemv 量化实现——应用纹理内存\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw18.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2F6WoFFGfj8%2BKpgHL7UpaeNNcbe4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>纹理内存（Image Object）：端侧 GPU 专用只读缓存，对二维空间局部性访问有硬件级优化，可以作为 Buffer 之外的一种存储方式。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"课程课后作业\">\u003Ca class=\"anchor\" href=\"#课程课后作业\" aria-label=\"跳转到标题：课程课后作业\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>课程课后作业\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_92_zw19.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ESJ9aIuU%2FPt4lN8VD1U8v0Us1TM%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\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BXH4UNzWFOrSPQr6JCB10ba6f1k%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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2FSVGJ8eOraf6k1nlkjH0x9NZjY%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,133,137,140,143,146,149,152,154,157,160,162,164,167,169,172,175,177,179,182],{"id":129,"text":129,"level":110},"核心归约算子实现",{"id":131,"text":132,"level":9},"argmax-最大值索引算子","Argmax 最大值索引算子",{"id":134,"text":135,"level":136},"1-算子作用与业务场景","1. 算子作用与业务场景",4,{"id":138,"text":139,"level":136},"2-两种实现方案","2. 两种实现方案",{"id":141,"text":142,"level":9},"softmax-归一化算子","Softmax 归一化算子",{"id":144,"text":145,"level":136},"1-一维-softmax-实现","1. 一维 Softmax 实现",{"id":147,"text":148,"level":136},"2-二维-softmax-两种轴实现","2. 二维 Softmax 两种轴实现",{"id":150,"text":151,"level":110},"基础矩阵乘法-gemm","基础矩阵乘法 GEMM",{"id":153,"text":153,"level":9},"基础实现",{"id":155,"text":156,"level":110},"gemv-矩阵向量乘性能优化","GEMV 矩阵向量乘性能优化",{"id":158,"text":159,"level":9},"子组优化-subgroup","子组优化（Subgroup）",{"id":161,"text":161,"level":9},"向量化加载与计算",{"id":163,"text":163,"level":9},"单工作项处理多行数据",{"id":165,"text":166,"level":110},"gemv-量化","Gemv 量化",{"id":168,"text":168,"level":9},"量化基础",{"id":170,"text":171,"level":9},"gemv-q8-0-量化实现","GEMV Q8-0 量化实现",{"id":173,"text":174,"level":9},"gemv-量化实现-应用纹理内存","Gemv 量化实现——应用纹理内存",{"id":176,"text":176,"level":110},"课程课后作业",{"id":178,"text":178,"level":110},"加入训练营",{"id":180,"text":181,"level":136},"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":183,"text":183,"level":110},"关注与交流"]