[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-36":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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Loo9mNcAiqrEJeoxKJGChNjsIP0%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aQ7DEIPmeZ6G2GUnnO42fZDzkoY%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rRUH7HCePjJ0x1Z1esoWYs0tjmE%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_36_fm.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e4AQ7xyE1Kx07oe6CAXuq%2B4W6Hk%3D","InfiniTensor,冬季训练营,AI编译器,DLCompiler,Triton","介绍 DLCompiler 作为 Triton 深度学习编译器拓展方向，讲解架构设计与开发实践。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营 AI 编译器第五课《拓展 Triton 深度学习编译器——DLCompiler》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TckJ%2BrUR1UqIyOXobWIoShviF5o%3D)\r\n\r\n2026 年 1 月 26 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 AI 编译器系列第五课《拓展 Triton 深度学习编译器——DLCompiler》开讲。\r\n\r\n本节课由上海人工智能实验室团队讲解其推出的 **DLCompiler**，DLCompiler 在 Triton 编程模型基础上进行了扩展，通过针对 DSA 芯片的后端适配，实现算子高效执行。\r\n\r\n## 算子开发现状\r\n\r\n### DSA架构特性\r\n国产 AI 芯片多采用 DSA 架构，与通用 GPU 存在显著差异：\r\n- **专用计算单元**：Cube（矩阵计算）、Vector（向量计算）、Scale（标量计算）\r\n- **多级缓存体系**：L0A、L0B、L0C 等多级片上缓存\r\n- **数据搬运机制**：需要显式管理数据从主机到片上缓存的搬运过程\r\n\r\n### 开发门槛高\r\n传统国产芯片算子开发面临多重挑战：\r\n- **硬件知识要求**：需要深入了解芯片架构、内存层次、计算单元特性\r\n- **代码复杂度高**：单个 Matmul 算子可能需要数百行底层代码\r\n- **调优困难**：手动优化逻辑复杂，性能与维护难以兼顾\r\n- **架构兼容性**：不同代际芯片架构差异大，代码复用性差\r\n\r\n### 生态碎片化\r\n- 各家厂商提供独立的软件栈，开发者需要针对不同厂商重复开发相同功能，造成资源浪费。\r\n\r\n## DLCompiler 整体架构和解决方案\r\n### 整体架构和流程\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xRLkGElKcjxmlX8bDdXuj7mhWKY%3D)\r\n\r\n### 安装和测试\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=65LZDDqStGwmimGNGSuirBvj5ww%3D)\r\n\r\n### DLCompiler 开发算子范式\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Bt99Z77RF0gaZsduJUVuIhEhjsU%3D)\r\n\r\n#### 编程范式差异\r\n\r\n **CUDA 模型：**\r\n- 处理单元：每个线程处理单个数据点\r\n- 编程粒度：标量级别的细粒度控制\r\n- 代码复杂度：需要手动管理内存、同步、边界条件\r\n\r\n**DLCompiler\u002FTriton 模型：**\r\n- 处理单元：每个 program 处理一个数据块（tile）\r\n- 编程粒度：块级别的粗粒度抽象\r\n- 自动优化：编译器自动处理内存布局、并行化等细节\r\n\r\n### DLCompiler 开发算子流程\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EuwYCYmAvmu4lleMASvS6ouyYjs%3D)\r\n\r\n### DLCompiler 开发 GEMM 算子\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SWdGrFPVEGZoVHMpZ3rQMyaWjcs%3D)\r\n\r\n### 跨架构的 DSL 扩展\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wQrmV158q%2BSfzbItOoSQL5KVJLs%3D)\r\n\r\n### Producer\u002FConsumer 编程模型\r\n创新性地将生产者-消费者模型引入算子开发，通过将 Cube 计算（生产者）和 Vector 计算（消费者）解耦，实现流水线并行，显著提升硬件利用率：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0MExJRSJU%2BHxdIhPr39N7jEdQUw%3D)\r\n\r\n\r\n### Tile-based 编程接口\r\n针对国产芯片的多级缓存体系，DLCompiler 实现了精细化的 Tile 管理：\r\n\r\n#### 数据搬运优化\r\n- **Global Memory → L0A\u002FL0B**：预取下一轮计算所需数据\r\n- **L0C → Global Memory**：异步写回计算结果\r\n- **重叠计算与搬运**：最大化硬件资源利用率\r\n\r\n#### 计算流水化\r\n将算子分解为多个阶段：\r\n1. **数据加载阶段**：从全局内存加载数据到片上缓存\r\n2. **计算阶段**：执行 Cube\u002FVector 计算\r\n3. **结果写回阶段**：将结果写回全局内存\r\n\r\n各阶段可并行执行，形成高效的计算流水线。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0W6l0LUXrbC5gaUYctdMU%2B9NQLE%3D)\r\n\r\n\r\n### 智能自动优化\r\nDLCompiler 内置多个优化 Pass：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Z5VSVZUzynomg8LQPVRAFEm4d24%3D)\r\n\r\n\r\n## 生态建设和成果展示\r\n\r\n### 上层框架对接和生态建设\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D%2FAzrwC5aFkqYyma%2FncnH7b1kRA%3D)\r\n\r\n### 基于 DLCompiler 的算子支持与性能优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pIox53F3fXmqAhlwNySRmF7wWAU%3D)\r\n\r\n\r\n## 未来发展方向\r\n\r\n当前 DLCompiler 主要优化单卡内部算子，未来将扩展至多卡场景：\r\n- 通信隐藏：在计算的同时进行通信\r\n- 流水线优化：跨设备的计算-通信流水线\r\n- 拓扑感知：根据网络拓扑优化通信模式\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D1J4dNQk%2BHHJcA2XAQi28KKQANI%3D)\r\n\r\n\r\n## 总结\r\n本节课介绍了算子开发现状，主要讲解 DLCompiler 的架构、生态建设和成果以及未来发展规划。\r\n\r\nDLCompiler 通过扩展 Triton 的编程模型，为国产 DSA 芯片提供了高效的算子开发解决方案。它不仅降低了国产芯片的开发门槛，还通过统一的编程接口促进了生态整合。随着大模型对算力需求的持续增长，DLCompiler 这类深度学习编译器将成为连接算法创新与硬件加速的重要桥梁，推动国产 AI 芯片生态的健康发展。\r\n\r\n\r\n\r\n![2025冬季训练营课程表](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=F9UYYN6GHyw0aDIvfyvnXUb6gKw%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4qCORU7P3l9%2FAWZfM%2BMlpmzjXSU%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-35","study-35","InfiniTensor 冬季训练营 AI 编译器第四课《AI 编译器中的后端优化》",73,"\u002Fstudy-material\u002Fstudy-36","study-36","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-37","study-37","InfiniTensor 冬季训练营训练系统第三课《大模型训练中的显存优化》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TckJ%2BrUR1UqIyOXobWIoShviF5o%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 26 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 AI 编译器系列第五课《拓展 Triton 深度学习编译器——DLCompiler》开讲。\u003C\u002Fp>\n\u003Cp>本节课由上海人工智能实验室团队讲解其推出的 \u003Cstrong>DLCompiler\u003C\u002Fstrong>，DLCompiler 在 Triton 编程模型基础上进行了扩展，通过针对 DSA 芯片的后端适配，实现算子高效执行。\u003C\u002Fp>\n\u003Ch2 id=\"算子开发现状\">\u003Ca class=\"anchor\" href=\"#算子开发现状\" aria-label=\"跳转到标题：算子开发现状\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>算子开发现状\u003C\u002Fh2>\u003Ch3 id=\"dsa架构特性\">\u003Ca class=\"anchor\" href=\"#dsa架构特性\" aria-label=\"跳转到标题：DSA架构特性\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DSA架构特性\u003C\u002Fh3>\u003Cp>国产 AI 芯片多采用 DSA 架构，与通用 GPU 存在显著差异：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>专用计算单元\u003C\u002Fstrong>：Cube（矩阵计算）、Vector（向量计算）、Scale（标量计算）\u003C\u002Fli>\n\u003Cli>\u003Cstrong>多级缓存体系\u003C\u002Fstrong>：L0A、L0B、L0C 等多级片上缓存\u003C\u002Fli>\n\u003Cli>\u003Cstrong>数据搬运机制\u003C\u002Fstrong>：需要显式管理数据从主机到片上缓存的搬运过程\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"开发门槛高\">\u003Ca class=\"anchor\" href=\"#开发门槛高\" aria-label=\"跳转到标题：开发门槛高\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>开发门槛高\u003C\u002Fh3>\u003Cp>传统国产芯片算子开发面临多重挑战：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>硬件知识要求\u003C\u002Fstrong>：需要深入了解芯片架构、内存层次、计算单元特性\u003C\u002Fli>\n\u003Cli>\u003Cstrong>代码复杂度高\u003C\u002Fstrong>：单个 Matmul 算子可能需要数百行底层代码\u003C\u002Fli>\n\u003Cli>\u003Cstrong>调优困难\u003C\u002Fstrong>：手动优化逻辑复杂，性能与维护难以兼顾\u003C\u002Fli>\n\u003Cli>\u003Cstrong>架构兼容性\u003C\u002Fstrong>：不同代际芯片架构差异大，代码复用性差\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"生态碎片化\">\u003Ca class=\"anchor\" href=\"#生态碎片化\" aria-label=\"跳转到标题：生态碎片化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>生态碎片化\u003C\u002Fh3>\u003Cul>\n\u003Cli>各家厂商提供独立的软件栈，开发者需要针对不同厂商重复开发相同功能，造成资源浪费。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"dlcompiler-整体架构和解决方案\">\u003Ca class=\"anchor\" href=\"#dlcompiler-整体架构和解决方案\" aria-label=\"跳转到标题：DLCompiler 整体架构和解决方案\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DLCompiler 整体架构和解决方案\u003C\u002Fh2>\u003Ch3 id=\"整体架构和流程\">\u003Ca class=\"anchor\" href=\"#整体架构和流程\" aria-label=\"跳转到标题：整体架构和流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>整体架构和流程\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xRLkGElKcjxmlX8bDdXuj7mhWKY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"安装和测试\">\u003Ca class=\"anchor\" href=\"#安装和测试\" aria-label=\"跳转到标题：安装和测试\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>安装和测试\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=65LZDDqStGwmimGNGSuirBvj5ww%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"dlcompiler-开发算子范式\">\u003Ca class=\"anchor\" href=\"#dlcompiler-开发算子范式\" aria-label=\"跳转到标题：DLCompiler 开发算子范式\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DLCompiler 开发算子范式\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Bt99Z77RF0gaZsduJUVuIhEhjsU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"编程范式差异\">\u003Ca class=\"anchor\" href=\"#编程范式差异\" aria-label=\"跳转到标题：编程范式差异\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>编程范式差异\u003C\u002Fh4>\u003Cp> \u003Cstrong>CUDA 模型：\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>处理单元：每个线程处理单个数据点\u003C\u002Fli>\n\u003Cli>编程粒度：标量级别的细粒度控制\u003C\u002Fli>\n\u003Cli>代码复杂度：需要手动管理内存、同步、边界条件\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>DLCompiler\u002FTriton 模型：\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>处理单元：每个 program 处理一个数据块（tile）\u003C\u002Fli>\n\u003Cli>编程粒度：块级别的粗粒度抽象\u003C\u002Fli>\n\u003Cli>自动优化：编译器自动处理内存布局、并行化等细节\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"dlcompiler-开发算子流程\">\u003Ca class=\"anchor\" href=\"#dlcompiler-开发算子流程\" aria-label=\"跳转到标题：DLCompiler 开发算子流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DLCompiler 开发算子流程\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EuwYCYmAvmu4lleMASvS6ouyYjs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"dlcompiler-开发-gemm-算子\">\u003Ca class=\"anchor\" href=\"#dlcompiler-开发-gemm-算子\" aria-label=\"跳转到标题：DLCompiler 开发 GEMM 算子\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DLCompiler 开发 GEMM 算子\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SWdGrFPVEGZoVHMpZ3rQMyaWjcs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"跨架构的-dsl-扩展\">\u003Ca class=\"anchor\" href=\"#跨架构的-dsl-扩展\" aria-label=\"跳转到标题：跨架构的 DSL 扩展\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>跨架构的 DSL 扩展\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wQrmV158q%2BSfzbItOoSQL5KVJLs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"producer-consumer-编程模型\">\u003Ca class=\"anchor\" href=\"#producer-consumer-编程模型\" aria-label=\"跳转到标题：Producer\u002FConsumer 编程模型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Producer\u002FConsumer 编程模型\u003C\u002Fh3>\u003Cp>创新性地将生产者-消费者模型引入算子开发，通过将 Cube 计算（生产者）和 Vector 计算（消费者）解耦，实现流水线并行，显著提升硬件利用率：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0MExJRSJU%2BHxdIhPr39N7jEdQUw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"tile-based-编程接口\">\u003Ca class=\"anchor\" href=\"#tile-based-编程接口\" aria-label=\"跳转到标题：Tile-based 编程接口\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Tile-based 编程接口\u003C\u002Fh3>\u003Cp>针对国产芯片的多级缓存体系，DLCompiler 实现了精细化的 Tile 管理：\u003C\u002Fp>\n\u003Ch4 id=\"数据搬运优化\">\u003Ca class=\"anchor\" href=\"#数据搬运优化\" aria-label=\"跳转到标题：数据搬运优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>数据搬运优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>Global Memory → L0A\u002FL0B\u003C\u002Fstrong>：预取下一轮计算所需数据\u003C\u002Fli>\n\u003Cli>\u003Cstrong>L0C → Global Memory\u003C\u002Fstrong>：异步写回计算结果\u003C\u002Fli>\n\u003Cli>\u003Cstrong>重叠计算与搬运\u003C\u002Fstrong>：最大化硬件资源利用率\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"计算流水化\">\u003Ca class=\"anchor\" href=\"#计算流水化\" aria-label=\"跳转到标题：计算流水化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>计算流水化\u003C\u002Fh4>\u003Cp>将算子分解为多个阶段：\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>数据加载阶段\u003C\u002Fstrong>：从全局内存加载数据到片上缓存\u003C\u002Fli>\n\u003Cli>\u003Cstrong>计算阶段\u003C\u002Fstrong>：执行 Cube\u002FVector 计算\u003C\u002Fli>\n\u003Cli>\u003Cstrong>结果写回阶段\u003C\u002Fstrong>：将结果写回全局内存\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>各阶段可并行执行，形成高效的计算流水线。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0W6l0LUXrbC5gaUYctdMU%2B9NQLE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"智能自动优化\">\u003Ca class=\"anchor\" href=\"#智能自动优化\" aria-label=\"跳转到标题：智能自动优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>智能自动优化\u003C\u002Fh3>\u003Cp>DLCompiler 内置多个优化 Pass：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Z5VSVZUzynomg8LQPVRAFEm4d24%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"生态建设和成果展示\">\u003Ca class=\"anchor\" href=\"#生态建设和成果展示\" aria-label=\"跳转到标题：生态建设和成果展示\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>生态建设和成果展示\u003C\u002Fh2>\u003Ch3 id=\"上层框架对接和生态建设\">\u003Ca class=\"anchor\" href=\"#上层框架对接和生态建设\" aria-label=\"跳转到标题：上层框架对接和生态建设\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>上层框架对接和生态建设\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D%2FAzrwC5aFkqYyma%2FncnH7b1kRA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"基于-dlcompiler-的算子支持与性能优化\">\u003Ca class=\"anchor\" href=\"#基于-dlcompiler-的算子支持与性能优化\" aria-label=\"跳转到标题：基于 DLCompiler 的算子支持与性能优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>基于 DLCompiler 的算子支持与性能优化\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pIox53F3fXmqAhlwNySRmF7wWAU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"未来发展方向\">\u003Ca class=\"anchor\" href=\"#未来发展方向\" aria-label=\"跳转到标题：未来发展方向\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>未来发展方向\u003C\u002Fh2>\u003Cp>当前 DLCompiler 主要优化单卡内部算子，未来将扩展至多卡场景：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>通信隐藏：在计算的同时进行通信\u003C\u002Fli>\n\u003Cli>流水线优化：跨设备的计算-通信流水线\u003C\u002Fli>\n\u003Cli>拓扑感知：根据网络拓扑优化通信模式\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_36_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D1J4dNQk%2BHHJcA2XAQi28KKQANI%3D\" alt=\"\">\u003C\u002Fp>\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>本节课介绍了算子开发现状，主要讲解 DLCompiler 的架构、生态建设和成果以及未来发展规划。\u003C\u002Fp>\n\u003Cp>DLCompiler 通过扩展 Triton 的编程模型，为国产 DSA 芯片提供了高效的算子开发解决方案。它不仅降低了国产芯片的开发门槛，还通过统一的编程接口促进了生态整合。随着大模型对算力需求的持续增长，DLCompiler 这类深度学习编译器将成为连接算法创新与硬件加速的重要桥梁，推动国产 AI 芯片生态的健康发展。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=F9UYYN6GHyw0aDIvfyvnXUb6gKw%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4qCORU7P3l9%2FAWZfM%2BMlpmzjXSU%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,133,135,137,140,142,144,147,150,153,156,159,162,165,167,169,171,173,175,178,180,182,185],{"id":129,"text":129,"level":110},"算子开发现状",{"id":131,"text":132,"level":9},"dsa架构特性","DSA架构特性",{"id":134,"text":134,"level":9},"开发门槛高",{"id":136,"text":136,"level":9},"生态碎片化",{"id":138,"text":139,"level":110},"dlcompiler-整体架构和解决方案","DLCompiler 整体架构和解决方案",{"id":141,"text":141,"level":9},"整体架构和流程",{"id":143,"text":143,"level":9},"安装和测试",{"id":145,"text":146,"level":9},"dlcompiler-开发算子范式","DLCompiler 开发算子范式",{"id":148,"text":148,"level":149},"编程范式差异",4,{"id":151,"text":152,"level":9},"dlcompiler-开发算子流程","DLCompiler 开发算子流程",{"id":154,"text":155,"level":9},"dlcompiler-开发-gemm-算子","DLCompiler 开发 GEMM 算子",{"id":157,"text":158,"level":9},"跨架构的-dsl-扩展","跨架构的 DSL 扩展",{"id":160,"text":161,"level":9},"producer-consumer-编程模型","Producer\u002FConsumer 编程模型",{"id":163,"text":164,"level":9},"tile-based-编程接口","Tile-based 编程接口",{"id":166,"text":166,"level":149},"数据搬运优化",{"id":168,"text":168,"level":149},"计算流水化",{"id":170,"text":170,"level":9},"智能自动优化",{"id":172,"text":172,"level":110},"生态建设和成果展示",{"id":174,"text":174,"level":9},"上层框架对接和生态建设",{"id":176,"text":177,"level":9},"基于-dlcompiler-的算子支持与性能优化","基于 DLCompiler 的算子支持与性能优化",{"id":179,"text":179,"level":110},"未来发展方向",{"id":181,"text":181,"level":110},"总结",{"id":183,"text":184,"level":9},"学习贴士","💡 学习贴士",{"id":186,"text":187,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]