[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-50":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=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=o29Dy5KMbkQVCWTVaFUX1%2F9v%2FI8%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=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RF6et0tKpeNsR%2FwxsCjy7zBTZbU%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=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mArtxwAhCa0dE5Cm17SDDzrBtQ8%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_50_fm.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=H2hkeHu4Zte5bZjGCFNXY0SIjYk%3D","InfiniTensor,冬季训练营,CUDA,量化,工业级部署","聚焦 CUDA 量化与工业级调优部署，覆盖精度权衡、性能优化和线上落地流程。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营 CUDA 编程第七课《量化与工业级调优部署》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw01.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NlwRy9d3K4x6aaOJtp%2By%2BvRHZMI%3D)\r\n\r\n\r\n\r\n2026 年 2 月 2 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第七课《量化与工业级调优部署》开讲。\r\n\r\n本节课系统介绍 CUDA 程序的终极优化方法，包括低精度量化、PTX 级调优、微指令优化等关键技术，将 CUDA 程序性能推向极致。\r\n\r\n\r\n### 回顾：并行计算的基本定律\r\n\r\n#### 1. 阿姆达尔定律（Amdahl's Law）\r\n#### 2. 古斯塔夫森定律（Gustafson's Law）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw02.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EsjG5zGO1iT51%2BWPyyb0fmLyJVg%3D)\r\n\r\n\r\n### 计算强度（Arithmetic Intensity）\r\n\r\n计算强度 AI=计算量 W\u002F访存量 Q，是衡量程序性能的重要指标。\r\n\r\n**优化策略**：\r\n- **提高 W**：通过算子融合等方式增加计算量\r\n- **降低 Q**：减少访存压力，传输更少的数据\r\n- **同时优化**：既提高计算量又减少访存量\r\n\r\n### 半精度（FP16）的优势\r\n\r\n使用FP16相比FP32可以：\r\n- **减少50%数据传输**：16位 vs 32位\r\n- **提升计算强度**：相同时间内处理更多数据\r\n- **降低内存占用**：减少显存需求\r\n\r\n## 低精度数据类型详解\r\n\r\n### BF16（Brain Floating Point 16）\r\n\r\nBF16 由 Google Brain 发明，专为深度学习设计：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw03.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NC845LKo0xVMseG518JUhydG7%2BM%3D)\r\n\r\n### TF32（Tensor Float 32）\r\n\r\n#### 1. TF32 由 NVIDIA 在 Ampere 架构中引入：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw04.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EMRv2J1BAe9fIG3U6H92M9%2B274M%3D)\r\n\r\n#### 2. Tensor Core 和混合精度\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw05.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DF4v7EU4QWxDYlaNHhR%2B7F9twrI%3D)\r\n\r\n\r\n### FP8 及其变体\r\n\r\nFP8 是 8 位低精度类型，有两种主要变体：\r\n\r\n- **E4M3**：注重精度\r\n- **E5M2**：注重动态范围\r\n- **硬件支持**：Hopper 架构及后续（H100、H800等）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw06.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TmOZoKNWGahEmXyQllpHR%2Fp62Nc%3D)\r\n\r\n\r\n### INT8\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw07.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=AnMR4oY8moFqDM6nSDKV1qpBQ1Q%3D)\r\n\r\n### 量化技术详解\r\n\r\n#### 1. 量化的基本概念\r\n\r\n量化是将高精度数据映射到低精度表示的过程：\r\n\r\n- **量化**（Quantization）：高精度 → 低精度\r\n- **反量化**（Dequantization）：低精度 → 高精度\r\n\r\n#### 2. 量化分类\r\n\r\n##### (1) 线性量化 vs 非线性量化\r\n\r\n- **线性量化**：数值区间线性均匀映射\r\n  - 公式：Q(x)=round(x\u002Fscale)+zero_point\r\n- **非线性量化**：非线性映射（如对数、幂次等）\r\n  - 优势：在特定场景下能更好平衡精度和范围\r\n\r\n##### (2) 对称量化 vs 非对称量化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw08.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=INQbUyQFIPP0rLETsrm9Ez%2F32rs%3D)\r\n\r\n\r\n##### (3) 静态量化 vs 动态量化\r\n\r\n- **静态量化**：量化参数预先校准并固定\r\n  - 优势：无需运行时计算，性能更好\r\n  - 劣势：需要校准数据集\r\n- **动态量化**：根据实时输入动态计算量化参数\r\n  - 优势：无需预先校准\r\n  - 劣势：需要额外的运行时计算\r\n\r\n##### (4) 量化粒度\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw09.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hAuGIBaufwDWhxoXSEaldASpHQg%3D)\r\n\r\n- **张量级**：整个张量共用一套量化参数\r\n- **通道级**：每个通道共用一套参数\r\n- **组级**：通道的子集共用一套参数\r\n\r\n### INT8 量化流程\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw10.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JzEF%2FfzblHoW25H0kdbYjj%2FCwUQ%3D)\r\n\r\n\r\n\r\n## 性能优化\r\n\r\n### PTX\r\n\r\n#### 1. PTX（Parallel Thread Execution）是 NVIDIA 的虚拟指令集：\r\n\r\n- **抽象层**：对底层硬件指令的抽象\r\n- **跨代兼容**：同一套 PTX 代码可在不同架构上运行\r\n- **细粒度控制**：提供编译器无法提供的底层功能\r\n\r\n#### 2. PTX 的应用场景\r\n\r\n##### (1) DeepEP\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw11.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ApP%2BNcljodKnzpy7pLzsZk6IB7Y%3D)\r\n\r\n\r\n##### (2) DeepGEMM\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw12.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2F0wu5I4LgQPpjvpXeKRlVGVFmM%3D)\r\n\r\n##### （3）NVIDIA Fuser\r\n##### （4）flashllm\r\n\r\n\r\n### 微指令调优技术\r\n#### 1. 循环展开（#pragma unroll）\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw13.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hrGCBoHbhLcr47MaoXeCslQmURQ%3D)\r\n\r\n\r\n#### 2. builtin_assume 和builtin_assume_aligned\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw14.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=g0QLYGJ4JCVZFFXE8KxuxIHzY1c%3D)\r\n\r\n\r\n\r\n#### 3. restrict 限定符\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw15.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1xCH8EoxcEWbRJMZOGf6txnWpG4%3D)\r\n\r\n\r\n#### 4.align\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw16.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qLbTuMWrmXp%2F7bA5joKFwPTYbAY%3D)\r\n\r\n#### 5. _ldg() 和 _const_\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw17.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MVnp3VhttkD21SqIwvK%2BVENDUb0%3D)\r\n\r\n\r\n### Warp Divergence\r\n\r\n#### 1. 线程发散问题\r\n\r\n同一个warp内的线程执行不同分支会导致：\r\n\r\n- **串行化执行**：不同分支需要串行执行\r\n- **性能下降**：较为明显的性能下降\r\n\r\n#### 2. 解决方案\r\n\r\n#### (1) 调整算法\r\n\r\n通过预处理数据或简化条件判断逻辑，完全避免发散。\r\n\r\n#### (2) 掩码技术（Masking）\r\n\r\n所有线程执行所有操作，但通过掩码控制结果生效：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw18.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aRPgQbCXX8GZqVrvtZgAd9LOzws%3D)\r\n\r\n\r\n#### (3) PTX 谓词寄存器\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw19.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BMgdT37S8wW07eF4vygaebU4oNQ%3D)\r\n\r\n\r\n### 指令级并行（ILP）优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw20.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yZ%2FWpnmNe53dAU8rWmbNDUouQeE%3D)\r\n\r\n\r\n#### 1. ILP vs TLP\r\n\r\n- **TLP**（Thread Level Parallelism）：通过增加线程数提高并行度\r\n- **ILP**（Instruction Level Parallelism）：通过指令重排和并行执行提高效率\r\n\r\n#### 2. ILP优化策略\r\n\r\n##### (1) 双发射（Dual Issue）\r\n\r\n将无冲突的指令并行执行：\r\n\r\n- **算术指令 + 内存指令**：通常可以无冲突并行\r\n- **性能提升**：依赖于指令结构和瓶颈类型，在某些内核中可带来可观收益\r\n\r\n##### (2) AoS vs SoA\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw21.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cXdemz%2FtA3G6bda9TNlTfyJqKdE%3D)\r\n\r\n- **AoS**（Array of Structures）\r\n- **SoA**（Structure of Arrays）\r\n\r\n**性能对比**（6M 数据）：\r\n- AoS：访存无法合并，性能较差\r\n- SoA：访存完全合并，性能提升近 2 倍\r\n\r\n**选择建议**\r\n\r\n- 一般情况：优先使用 SoA\r\n- 特殊访问模式：根据实际访问模式选择\r\n- 稀疏数据：特别注意访存合并问题\r\n\r\n\r\n## CUDA官方库介绍\r\n\r\n### 1. cuDNN\r\n\r\n深度神经网络GPU加速库：\r\n\r\n- **功能**：卷积、池化、归一化、激活等\r\n- **优化**：利用Tensor Core进行混合精度计算\r\n- **自动选择**：根据输入自动选择最优算法\r\n\r\n### 2. cuBLAS\r\n\r\n线性代数GPU加速库：\r\n\r\n- **层级**：Level 1（向量）、Level 2（矩阵-向量）、Level 3（矩阵-矩阵）\r\n- **扩展**：cuBLASXT（多GPU）、cuBLASLT（轻量级）\r\n\r\n### 3. CUB\r\n\r\nCUDA Unbound并行算法原语库：\r\n\r\n- **功能**：排序、前缀和、归约、直方图等\r\n- **层级**：线程级、block级、设备级\r\n\r\n### 4. CUTLASS\r\n\r\n高性能矩阵乘法模板库：\r\n\r\n- **优势**：高度优化的GEMM实现\r\n- **支持**：多种精度类型、Tensor Core\r\n- **易用性**：相对容易写出高性能代码\r\n\r\n### 5. cuSPARSE\r\n\r\n稀疏矩阵运算加速库：\r\n\r\n- **格式支持**：CSR、COO、CSC等\r\n- **功能**：稀疏-稠密转换、稀疏求解器等\r\n- **异步执行**：支持CUDA stream\r\n\r\n## 总结\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw22.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CzmGzM1cjfSvTIcva9ToIoaHMjY%3D)\r\n\r\n\r\nCUDA 程序的性能优化是一个系统工程，需要从算法设计、数据结构、内存访问、计算精度等多个维度综合考虑。通过掌握低精度量化、PTX 级调优、微指令优化等高级技术，可以将 CUDA 程序性能推向工业级水平。\r\n\r\n\r\n\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_kcb.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=43wRrRwaJR3bNO8aPJjTK7w99GQ%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_ewm.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vCZaUGGZDDvRIe3O2f7mI%2BWDJLk%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-49","study-49","InfiniTensor 冬季训练营推理引擎第六课《vLLM-Ascend 大模型推理》",59,"\u002Fstudy-material\u002Fstudy-50","study-50","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-51","study-51","InfiniTensor 冬季训练营训练系统第五课《强化学习及 DeepSeek》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw01.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NlwRy9d3K4x6aaOJtp%2By%2BvRHZMI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 2 月 2 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第七课《量化与工业级调优部署》开讲。\u003C\u002Fp>\n\u003Cp>本节课系统介绍 CUDA 程序的终极优化方法，包括低精度量化、PTX 级调优、微指令优化等关键技术，将 CUDA 程序性能推向极致。\u003C\u002Fp>\n\u003Ch3 id=\"回顾-并行计算的基本定律\">\u003Ca class=\"anchor\" href=\"#回顾-并行计算的基本定律\" aria-label=\"跳转到标题：回顾：并行计算的基本定律\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>回顾：并行计算的基本定律\u003C\u002Fh3>\u003Ch4 id=\"1-阿姆达尔定律-amdahl-39-s-law\">\u003Ca class=\"anchor\" href=\"#1-阿姆达尔定律-amdahl-39-s-law\" aria-label=\"跳转到标题：1. 阿姆达尔定律（Amdahl's Law）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 阿姆达尔定律（Amdahl's Law）\u003C\u002Fh4>\u003Ch4 id=\"2-古斯塔夫森定律-gustafson-39-s-law\">\u003Ca class=\"anchor\" href=\"#2-古斯塔夫森定律-gustafson-39-s-law\" aria-label=\"跳转到标题：2. 古斯塔夫森定律（Gustafson's Law）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 古斯塔夫森定律（Gustafson's Law）\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw02.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EsjG5zGO1iT51%2BWPyyb0fmLyJVg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"计算强度-arithmetic-intensity\">\u003Ca class=\"anchor\" href=\"#计算强度-arithmetic-intensity\" aria-label=\"跳转到标题：计算强度（Arithmetic Intensity）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>计算强度（Arithmetic Intensity）\u003C\u002Fh3>\u003Cp>计算强度 AI=计算量 W\u002F访存量 Q，是衡量程序性能的重要指标。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>优化策略\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>提高 W\u003C\u002Fstrong>：通过算子融合等方式增加计算量\u003C\u002Fli>\n\u003Cli>\u003Cstrong>降低 Q\u003C\u002Fstrong>：减少访存压力，传输更少的数据\u003C\u002Fli>\n\u003Cli>\u003Cstrong>同时优化\u003C\u002Fstrong>：既提高计算量又减少访存量\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"半精度-fp16-的优势\">\u003Ca class=\"anchor\" href=\"#半精度-fp16-的优势\" aria-label=\"跳转到标题：半精度（FP16）的优势\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>半精度（FP16）的优势\u003C\u002Fh3>\u003Cp>使用FP16相比FP32可以：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>减少50%数据传输\u003C\u002Fstrong>：16位 vs 32位\u003C\u002Fli>\n\u003Cli>\u003Cstrong>提升计算强度\u003C\u002Fstrong>：相同时间内处理更多数据\u003C\u002Fli>\n\u003Cli>\u003Cstrong>降低内存占用\u003C\u002Fstrong>：减少显存需求\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"低精度数据类型详解\">\u003Ca class=\"anchor\" href=\"#低精度数据类型详解\" aria-label=\"跳转到标题：低精度数据类型详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>低精度数据类型详解\u003C\u002Fh2>\u003Ch3 id=\"bf16-brain-floating-point-16\">\u003Ca class=\"anchor\" href=\"#bf16-brain-floating-point-16\" aria-label=\"跳转到标题：BF16（Brain Floating Point 16）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>BF16（Brain Floating Point 16）\u003C\u002Fh3>\u003Cp>BF16 由 Google Brain 发明，专为深度学习设计：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw03.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NC845LKo0xVMseG518JUhydG7%2BM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"tf32-tensor-float-32\">\u003Ca class=\"anchor\" href=\"#tf32-tensor-float-32\" aria-label=\"跳转到标题：TF32（Tensor Float 32）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>TF32（Tensor Float 32）\u003C\u002Fh3>\u003Ch4 id=\"1-tf32-由-nvidia-在-ampere-架构中引入\">\u003Ca class=\"anchor\" href=\"#1-tf32-由-nvidia-在-ampere-架构中引入\" aria-label=\"跳转到标题：1. TF32 由 NVIDIA 在 Ampere 架构中引入：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. TF32 由 NVIDIA 在 Ampere 架构中引入：\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw04.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EMRv2J1BAe9fIG3U6H92M9%2B274M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-tensor-core-和混合精度\">\u003Ca class=\"anchor\" href=\"#2-tensor-core-和混合精度\" aria-label=\"跳转到标题：2. Tensor Core 和混合精度\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Tensor Core 和混合精度\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw05.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DF4v7EU4QWxDYlaNHhR%2B7F9twrI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"fp8-及其变体\">\u003Ca class=\"anchor\" href=\"#fp8-及其变体\" aria-label=\"跳转到标题：FP8 及其变体\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>FP8 及其变体\u003C\u002Fh3>\u003Cp>FP8 是 8 位低精度类型，有两种主要变体：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>E4M3\u003C\u002Fstrong>：注重精度\u003C\u002Fli>\n\u003Cli>\u003Cstrong>E5M2\u003C\u002Fstrong>：注重动态范围\u003C\u002Fli>\n\u003Cli>\u003Cstrong>硬件支持\u003C\u002Fstrong>：Hopper 架构及后续（H100、H800等）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw06.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TmOZoKNWGahEmXyQllpHR%2Fp62Nc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"int8\">\u003Ca class=\"anchor\" href=\"#int8\" aria-label=\"跳转到标题：INT8\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>INT8\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw07.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=AnMR4oY8moFqDM6nSDKV1qpBQ1Q%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"量化技术详解\">\u003Ca class=\"anchor\" href=\"#量化技术详解\" aria-label=\"跳转到标题：量化技术详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>量化技术详解\u003C\u002Fh3>\u003Ch4 id=\"1-量化的基本概念\">\u003Ca class=\"anchor\" href=\"#1-量化的基本概念\" aria-label=\"跳转到标题：1. 量化的基本概念\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 量化的基本概念\u003C\u002Fh4>\u003Cp>量化是将高精度数据映射到低精度表示的过程：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>量化\u003C\u002Fstrong>（Quantization）：高精度 → 低精度\u003C\u002Fli>\n\u003Cli>\u003Cstrong>反量化\u003C\u002Fstrong>（Dequantization）：低精度 → 高精度\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-量化分类\">\u003Ca class=\"anchor\" href=\"#2-量化分类\" aria-label=\"跳转到标题：2. 量化分类\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 量化分类\u003C\u002Fh4>\u003Ch5 id=\"1-线性量化-vs-非线性量化\">\u003Ca class=\"anchor\" href=\"#1-线性量化-vs-非线性量化\" aria-label=\"跳转到标题：(1) 线性量化 vs 非线性量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(1) 线性量化 vs 非线性量化\u003C\u002Fh5>\u003Cul>\n\u003Cli>\u003Cstrong>线性量化\u003C\u002Fstrong>：数值区间线性均匀映射\u003Cul>\n\u003Cli>公式：Q(x)=round(x\u002Fscale)+zero_point\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>非线性量化\u003C\u002Fstrong>：非线性映射（如对数、幂次等）\u003Cul>\n\u003Cli>优势：在特定场景下能更好平衡精度和范围\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch5 id=\"2-对称量化-vs-非对称量化\">\u003Ca class=\"anchor\" href=\"#2-对称量化-vs-非对称量化\" aria-label=\"跳转到标题：(2) 对称量化 vs 非对称量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(2) 对称量化 vs 非对称量化\u003C\u002Fh5>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw08.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=INQbUyQFIPP0rLETsrm9Ez%2F32rs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch5 id=\"3-静态量化-vs-动态量化\">\u003Ca class=\"anchor\" href=\"#3-静态量化-vs-动态量化\" aria-label=\"跳转到标题：(3) 静态量化 vs 动态量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(3) 静态量化 vs 动态量化\u003C\u002Fh5>\u003Cul>\n\u003Cli>\u003Cstrong>静态量化\u003C\u002Fstrong>：量化参数预先校准并固定\u003Cul>\n\u003Cli>优势：无需运行时计算，性能更好\u003C\u002Fli>\n\u003Cli>劣势：需要校准数据集\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>动态量化\u003C\u002Fstrong>：根据实时输入动态计算量化参数\u003Cul>\n\u003Cli>优势：无需预先校准\u003C\u002Fli>\n\u003Cli>劣势：需要额外的运行时计算\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch5 id=\"4-量化粒度\">\u003Ca class=\"anchor\" href=\"#4-量化粒度\" aria-label=\"跳转到标题：(4) 量化粒度\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(4) 量化粒度\u003C\u002Fh5>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw09.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hAuGIBaufwDWhxoXSEaldASpHQg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>张量级\u003C\u002Fstrong>：整个张量共用一套量化参数\u003C\u002Fli>\n\u003Cli>\u003Cstrong>通道级\u003C\u002Fstrong>：每个通道共用一套参数\u003C\u002Fli>\n\u003Cli>\u003Cstrong>组级\u003C\u002Fstrong>：通道的子集共用一套参数\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"int8-量化流程\">\u003Ca class=\"anchor\" href=\"#int8-量化流程\" aria-label=\"跳转到标题：INT8 量化流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>INT8 量化流程\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw10.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JzEF%2FfzblHoW25H0kdbYjj%2FCwUQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"性能优化\">\u003Ca class=\"anchor\" href=\"#性能优化\" aria-label=\"跳转到标题：性能优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>性能优化\u003C\u002Fh2>\u003Ch3 id=\"ptx\">\u003Ca class=\"anchor\" href=\"#ptx\" aria-label=\"跳转到标题：PTX\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>PTX\u003C\u002Fh3>\u003Ch4 id=\"1-ptx-parallel-thread-execution-是-nvidia-的虚拟指令集\">\u003Ca class=\"anchor\" href=\"#1-ptx-parallel-thread-execution-是-nvidia-的虚拟指令集\" aria-label=\"跳转到标题：1. PTX（Parallel Thread Execution）是 NVIDIA 的虚拟指令集：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. PTX（Parallel Thread Execution）是 NVIDIA 的虚拟指令集：\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>抽象层\u003C\u002Fstrong>：对底层硬件指令的抽象\u003C\u002Fli>\n\u003Cli>\u003Cstrong>跨代兼容\u003C\u002Fstrong>：同一套 PTX 代码可在不同架构上运行\u003C\u002Fli>\n\u003Cli>\u003Cstrong>细粒度控制\u003C\u002Fstrong>：提供编译器无法提供的底层功能\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-ptx-的应用场景\">\u003Ca class=\"anchor\" href=\"#2-ptx-的应用场景\" aria-label=\"跳转到标题：2. PTX 的应用场景\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. PTX 的应用场景\u003C\u002Fh4>\u003Ch5 id=\"1-deepep\">\u003Ca class=\"anchor\" href=\"#1-deepep\" aria-label=\"跳转到标题：(1) DeepEP\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(1) DeepEP\u003C\u002Fh5>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw11.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ApP%2BNcljodKnzpy7pLzsZk6IB7Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch5 id=\"2-deepgemm\">\u003Ca class=\"anchor\" href=\"#2-deepgemm\" aria-label=\"跳转到标题：(2) DeepGEMM\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(2) DeepGEMM\u003C\u002Fh5>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw12.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2F0wu5I4LgQPpjvpXeKRlVGVFmM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch5 id=\"3-nvidia-fuser\">\u003Ca class=\"anchor\" href=\"#3-nvidia-fuser\" aria-label=\"跳转到标题：（3）NVIDIA Fuser\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（3）NVIDIA Fuser\u003C\u002Fh5>\u003Ch5 id=\"4-flashllm\">\u003Ca class=\"anchor\" href=\"#4-flashllm\" aria-label=\"跳转到标题：（4）flashllm\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（4）flashllm\u003C\u002Fh5>\u003Ch3 id=\"微指令调优技术\">\u003Ca class=\"anchor\" href=\"#微指令调优技术\" aria-label=\"跳转到标题：微指令调优技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>微指令调优技术\u003C\u002Fh3>\u003Ch4 id=\"1-循环展开-pragma-unroll\">\u003Ca class=\"anchor\" href=\"#1-循环展开-pragma-unroll\" aria-label=\"跳转到标题：1. 循环展开（#pragma unroll）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 循环展开（#pragma unroll）\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw13.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hrGCBoHbhLcr47MaoXeCslQmURQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-builtin-assume-和builtin-assume-aligned\">\u003Ca class=\"anchor\" href=\"#2-builtin-assume-和builtin-assume-aligned\" aria-label=\"跳转到标题：2. builtin_assume 和builtin_assume_aligned\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. builtin_assume 和builtin_assume_aligned\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw14.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=g0QLYGJ4JCVZFFXE8KxuxIHzY1c%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-restrict-限定符\">\u003Ca class=\"anchor\" href=\"#3-restrict-限定符\" aria-label=\"跳转到标题：3. restrict 限定符\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. restrict 限定符\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw15.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1xCH8EoxcEWbRJMZOGf6txnWpG4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"4-align\">\u003Ca class=\"anchor\" href=\"#4-align\" aria-label=\"跳转到标题：4.align\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4.align\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw16.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qLbTuMWrmXp%2F7bA5joKFwPTYbAY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"5-ldg-和-const\">\u003Ca class=\"anchor\" href=\"#5-ldg-和-const\" aria-label=\"跳转到标题：5. _ldg() 和 const\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>5. _ldg() 和 \u003Cem>const\u003C\u002Fem>\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw17.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MVnp3VhttkD21SqIwvK%2BVENDUb0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"warp-divergence\">\u003Ca class=\"anchor\" href=\"#warp-divergence\" aria-label=\"跳转到标题：Warp Divergence\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Warp Divergence\u003C\u002Fh3>\u003Ch4 id=\"1-线程发散问题\">\u003Ca class=\"anchor\" href=\"#1-线程发散问题\" aria-label=\"跳转到标题：1. 线程发散问题\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 线程发散问题\u003C\u002Fh4>\u003Cp>同一个warp内的线程执行不同分支会导致：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>串行化执行\u003C\u002Fstrong>：不同分支需要串行执行\u003C\u002Fli>\n\u003Cli>\u003Cstrong>性能下降\u003C\u002Fstrong>：较为明显的性能下降\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-解决方案\">\u003Ca class=\"anchor\" href=\"#2-解决方案\" aria-label=\"跳转到标题：2. 解决方案\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 解决方案\u003C\u002Fh4>\u003Ch4 id=\"1-调整算法\">\u003Ca class=\"anchor\" href=\"#1-调整算法\" aria-label=\"跳转到标题：(1) 调整算法\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(1) 调整算法\u003C\u002Fh4>\u003Cp>通过预处理数据或简化条件判断逻辑，完全避免发散。\u003C\u002Fp>\n\u003Ch4 id=\"2-掩码技术-masking\">\u003Ca class=\"anchor\" href=\"#2-掩码技术-masking\" aria-label=\"跳转到标题：(2) 掩码技术（Masking）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(2) 掩码技术（Masking）\u003C\u002Fh4>\u003Cp>所有线程执行所有操作，但通过掩码控制结果生效：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw18.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aRPgQbCXX8GZqVrvtZgAd9LOzws%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-ptx-谓词寄存器\">\u003Ca class=\"anchor\" href=\"#3-ptx-谓词寄存器\" aria-label=\"跳转到标题：(3) PTX 谓词寄存器\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(3) PTX 谓词寄存器\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw19.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BMgdT37S8wW07eF4vygaebU4oNQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"指令级并行-ilp-优化\">\u003Ca class=\"anchor\" href=\"#指令级并行-ilp-优化\" aria-label=\"跳转到标题：指令级并行（ILP）优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>指令级并行（ILP）优化\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw20.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yZ%2FWpnmNe53dAU8rWmbNDUouQeE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-ilp-vs-tlp\">\u003Ca class=\"anchor\" href=\"#1-ilp-vs-tlp\" aria-label=\"跳转到标题：1. ILP vs TLP\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. ILP vs TLP\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>TLP\u003C\u002Fstrong>（Thread Level Parallelism）：通过增加线程数提高并行度\u003C\u002Fli>\n\u003Cli>\u003Cstrong>ILP\u003C\u002Fstrong>（Instruction Level Parallelism）：通过指令重排和并行执行提高效率\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-ilp优化策略\">\u003Ca class=\"anchor\" href=\"#2-ilp优化策略\" aria-label=\"跳转到标题：2. ILP优化策略\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. ILP优化策略\u003C\u002Fh4>\u003Ch5 id=\"1-双发射-dual-issue\">\u003Ca class=\"anchor\" href=\"#1-双发射-dual-issue\" aria-label=\"跳转到标题：(1) 双发射（Dual Issue）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(1) 双发射（Dual Issue）\u003C\u002Fh5>\u003Cp>将无冲突的指令并行执行：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>算术指令 + 内存指令\u003C\u002Fstrong>：通常可以无冲突并行\u003C\u002Fli>\n\u003Cli>\u003Cstrong>性能提升\u003C\u002Fstrong>：依赖于指令结构和瓶颈类型，在某些内核中可带来可观收益\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch5 id=\"2-aos-vs-soa\">\u003Ca class=\"anchor\" href=\"#2-aos-vs-soa\" aria-label=\"跳转到标题：(2) AoS vs SoA\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(2) AoS vs SoA\u003C\u002Fh5>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw21.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cXdemz%2FtA3G6bda9TNlTfyJqKdE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>AoS\u003C\u002Fstrong>（Array of Structures）\u003C\u002Fli>\n\u003Cli>\u003Cstrong>SoA\u003C\u002Fstrong>（Structure of Arrays）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>性能对比\u003C\u002Fstrong>（6M 数据）：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>AoS：访存无法合并，性能较差\u003C\u002Fli>\n\u003Cli>SoA：访存完全合并，性能提升近 2 倍\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>选择建议\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>一般情况：优先使用 SoA\u003C\u002Fli>\n\u003Cli>特殊访问模式：根据实际访问模式选择\u003C\u002Fli>\n\u003Cli>稀疏数据：特别注意访存合并问题\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"cuda官方库介绍\">\u003Ca class=\"anchor\" href=\"#cuda官方库介绍\" aria-label=\"跳转到标题：CUDA官方库介绍\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA官方库介绍\u003C\u002Fh2>\u003Ch3 id=\"1-cudnn\">\u003Ca class=\"anchor\" href=\"#1-cudnn\" aria-label=\"跳转到标题：1. cuDNN\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. cuDNN\u003C\u002Fh3>\u003Cp>深度神经网络GPU加速库：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>功能\u003C\u002Fstrong>：卷积、池化、归一化、激活等\u003C\u002Fli>\n\u003Cli>\u003Cstrong>优化\u003C\u002Fstrong>：利用Tensor Core进行混合精度计算\u003C\u002Fli>\n\u003Cli>\u003Cstrong>自动选择\u003C\u002Fstrong>：根据输入自动选择最优算法\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"2-cublas\">\u003Ca class=\"anchor\" href=\"#2-cublas\" aria-label=\"跳转到标题：2. cuBLAS\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. cuBLAS\u003C\u002Fh3>\u003Cp>线性代数GPU加速库：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>层级\u003C\u002Fstrong>：Level 1（向量）、Level 2（矩阵-向量）、Level 3（矩阵-矩阵）\u003C\u002Fli>\n\u003Cli>\u003Cstrong>扩展\u003C\u002Fstrong>：cuBLASXT（多GPU）、cuBLASLT（轻量级）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"3-cub\">\u003Ca class=\"anchor\" href=\"#3-cub\" aria-label=\"跳转到标题：3. CUB\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. CUB\u003C\u002Fh3>\u003Cp>CUDA Unbound并行算法原语库：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>功能\u003C\u002Fstrong>：排序、前缀和、归约、直方图等\u003C\u002Fli>\n\u003Cli>\u003Cstrong>层级\u003C\u002Fstrong>：线程级、block级、设备级\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"4-cutlass\">\u003Ca class=\"anchor\" href=\"#4-cutlass\" aria-label=\"跳转到标题：4. CUTLASS\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. CUTLASS\u003C\u002Fh3>\u003Cp>高性能矩阵乘法模板库：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>优势\u003C\u002Fstrong>：高度优化的GEMM实现\u003C\u002Fli>\n\u003Cli>\u003Cstrong>支持\u003C\u002Fstrong>：多种精度类型、Tensor Core\u003C\u002Fli>\n\u003Cli>\u003Cstrong>易用性\u003C\u002Fstrong>：相对容易写出高性能代码\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"5-cusparse\">\u003Ca class=\"anchor\" href=\"#5-cusparse\" aria-label=\"跳转到标题：5. cuSPARSE\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>5. cuSPARSE\u003C\u002Fh3>\u003Cp>稀疏矩阵运算加速库：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>格式支持\u003C\u002Fstrong>：CSR、COO、CSC等\u003C\u002Fli>\n\u003Cli>\u003Cstrong>功能\u003C\u002Fstrong>：稀疏-稠密转换、稀疏求解器等\u003C\u002Fli>\n\u003Cli>\u003Cstrong>异步执行\u003C\u002Fstrong>：支持CUDA stream\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 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\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_50_zw22.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CzmGzM1cjfSvTIcva9ToIoaHMjY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>CUDA 程序的性能优化是一个系统工程，需要从算法设计、数据结构、内存访问、计算精度等多个维度综合考虑。通过掌握低精度量化、PTX 级调优、微指令优化等高级技术，可以将 CUDA 程序性能推向工业级水平。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=43wRrRwaJR3bNO8aPJjTK7w99GQ%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_ewm.jpg?Expires=1792752820&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vCZaUGGZDDvRIe3O2f7mI%2BWDJLk%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,135,138,141,144,146,149,152,155,158,161,164,166,169,172,175,178,181,184,187,189,192,195,198,201,204,207,210,212,215,218,221,224,227,230,233,236,239,242,245,248,251,254,257,260,263,266,269,272,275,278,280,283],{"id":129,"text":130,"level":9},"回顾-并行计算的基本定律","回顾：并行计算的基本定律",{"id":132,"text":133,"level":134},"1-阿姆达尔定律-amdahl-39-s-law","1. 阿姆达尔定律（Amdahl's Law）",4,{"id":136,"text":137,"level":134},"2-古斯塔夫森定律-gustafson-39-s-law","2. 古斯塔夫森定律（Gustafson's Law）",{"id":139,"text":140,"level":9},"计算强度-arithmetic-intensity","计算强度（Arithmetic Intensity）",{"id":142,"text":143,"level":9},"半精度-fp16-的优势","半精度（FP16）的优势",{"id":145,"text":145,"level":110},"低精度数据类型详解",{"id":147,"text":148,"level":9},"bf16-brain-floating-point-16","BF16（Brain Floating Point 16）",{"id":150,"text":151,"level":9},"tf32-tensor-float-32","TF32（Tensor Float 32）",{"id":153,"text":154,"level":134},"1-tf32-由-nvidia-在-ampere-架构中引入","1. TF32 由 NVIDIA 在 Ampere 架构中引入：",{"id":156,"text":157,"level":134},"2-tensor-core-和混合精度","2. Tensor Core 和混合精度",{"id":159,"text":160,"level":9},"fp8-及其变体","FP8 及其变体",{"id":162,"text":163,"level":9},"int8","INT8",{"id":165,"text":165,"level":9},"量化技术详解",{"id":167,"text":168,"level":134},"1-量化的基本概念","1. 量化的基本概念",{"id":170,"text":171,"level":134},"2-量化分类","2. 量化分类",{"id":173,"text":174,"level":45},"1-线性量化-vs-非线性量化","(1) 线性量化 vs 非线性量化",{"id":176,"text":177,"level":45},"2-对称量化-vs-非对称量化","(2) 对称量化 vs 非对称量化",{"id":179,"text":180,"level":45},"3-静态量化-vs-动态量化","(3) 静态量化 vs 动态量化",{"id":182,"text":183,"level":45},"4-量化粒度","(4) 量化粒度",{"id":185,"text":186,"level":9},"int8-量化流程","INT8 量化流程",{"id":188,"text":188,"level":110},"性能优化",{"id":190,"text":191,"level":9},"ptx","PTX",{"id":193,"text":194,"level":134},"1-ptx-parallel-thread-execution-是-nvidia-的虚拟指令集","1. PTX（Parallel Thread Execution）是 NVIDIA 的虚拟指令集：",{"id":196,"text":197,"level":134},"2-ptx-的应用场景","2. PTX 的应用场景",{"id":199,"text":200,"level":45},"1-deepep","(1) DeepEP",{"id":202,"text":203,"level":45},"2-deepgemm","(2) DeepGEMM",{"id":205,"text":206,"level":45},"3-nvidia-fuser","（3）NVIDIA Fuser",{"id":208,"text":209,"level":45},"4-flashllm","（4）flashllm",{"id":211,"text":211,"level":9},"微指令调优技术",{"id":213,"text":214,"level":134},"1-循环展开-pragma-unroll","1. 循环展开（#pragma unroll）",{"id":216,"text":217,"level":134},"2-builtin-assume-和builtin-assume-aligned","2. builtin_assume 和builtin_assume_aligned",{"id":219,"text":220,"level":134},"3-restrict-限定符","3. restrict 限定符",{"id":222,"text":223,"level":134},"4-align","4.align",{"id":225,"text":226,"level":134},"5-ldg-和-const","5. _ldg() 和 const",{"id":228,"text":229,"level":9},"warp-divergence","Warp Divergence",{"id":231,"text":232,"level":134},"1-线程发散问题","1. 线程发散问题",{"id":234,"text":235,"level":134},"2-解决方案","2. 解决方案",{"id":237,"text":238,"level":134},"1-调整算法","(1) 调整算法",{"id":240,"text":241,"level":134},"2-掩码技术-masking","(2) 掩码技术（Masking）",{"id":243,"text":244,"level":134},"3-ptx-谓词寄存器","(3) PTX 谓词寄存器",{"id":246,"text":247,"level":9},"指令级并行-ilp-优化","指令级并行（ILP）优化",{"id":249,"text":250,"level":134},"1-ilp-vs-tlp","1. ILP vs TLP",{"id":252,"text":253,"level":134},"2-ilp优化策略","2. ILP优化策略",{"id":255,"text":256,"level":45},"1-双发射-dual-issue","(1) 双发射（Dual Issue）",{"id":258,"text":259,"level":45},"2-aos-vs-soa","(2) AoS vs SoA",{"id":261,"text":262,"level":110},"cuda官方库介绍","CUDA官方库介绍",{"id":264,"text":265,"level":9},"1-cudnn","1. cuDNN",{"id":267,"text":268,"level":9},"2-cublas","2. cuBLAS",{"id":270,"text":271,"level":9},"3-cub","3. CUB",{"id":273,"text":274,"level":9},"4-cutlass","4. CUTLASS",{"id":276,"text":277,"level":9},"5-cusparse","5. cuSPARSE",{"id":279,"text":279,"level":110},"总结",{"id":281,"text":282,"level":9},"学习贴士","💡 学习贴士",{"id":284,"text":285,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]