[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-ARM-software--ComputeLibrary":3,"tool-ARM-software--ComputeLibrary":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",138956,2,"2026-04-05T11:33:21",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":76,"owner_avatar_url":77,"owner_bio":78,"owner_company":79,"owner_location":79,"owner_email":79,"owner_twitter":79,"owner_website":80,"owner_url":81,"languages":82,"stars":112,"forks":113,"last_commit_at":114,"license":79,"difficulty_score":115,"env_os":116,"env_gpu":117,"env_ram":118,"env_deps":119,"category_tags":128,"github_topics":129,"view_count":23,"oss_zip_url":79,"oss_zip_packed_at":79,"status":16,"created_at":144,"updated_at":145,"faqs":146,"releases":175},2645,"ARM-software\u002FComputeLibrary","ComputeLibrary","The Compute Library is a set of computer vision and machine learning functions optimised for both Arm CPUs and GPUs using SIMD technologies.","ComputeLibrary 是一套专为 Arm 架构打造的开源计算机视觉与机器学习函数库。它针对 Arm Cortex-A、Neoverse 处理器以及 Mali GPU 进行了深度优化，利用 SIMD 技术显著提升了底层算子的执行效率，解决了在嵌入式及移动设备上运行深度学习模型时常见的性能瓶颈问题。\n\n无论是需要部署高效推理引擎的嵌入式开发者，还是致力于算法优化的研究人员，都能从中受益。它提供了超过 100 种机器学习函数，涵盖多种卷积算法（如 Winograd、FFT、GeMM 等），并支持 FP32、FP16、INT8 等多种数据精度，能够灵活适配从智能手机到边缘服务器的各类场景。\n\n其核心亮点在于对 Arm 最新技术（如 SVE2）的即时支持，以及通过内核融合、快速数学运算和纹理利用等高级优化手段，实现了超越其他开源替代方案的性能表现。此外，ComputeLibrary 采用宽松的 MIT 许可证，构建选项高度可配置，允许开发者生成轻量级二进制文件，并支持针对特定设备进行微调。作为一个成熟且活跃的开源项目，它为在 Arm 生态中构建高性能 AI 应用提供了坚实的底层基础。","> [!IMPORTANT]\n> **Static Library Name Change**\n>\n> - libarm_compute-static.a will be renamed to libarm_compute.a in distributed pre-built binaries.\n> - This change will be in effect starting from the first release in or after Jan 2026.\n>\n\n\u003Cbr>\n\u003Cdiv align=\"center\">\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FARM-software_ComputeLibrary_readme_686aa3674122.png\"\u002F>\u003Cbr>\u003Cbr>\n\u003C\u002Fdiv>\n\n# Compute Library [![acl-release-shield]][acl-release]\n\n[acl-release-shield]: https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002FARM-software\u002FComputeLibrary?label=latest&color=green\n[acl-release]: https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Freleases\u002Flatest\n\nThe Compute Library is a collection of low-level machine learning functions optimized for Arm® Cortex®-A, Arm® Neoverse™ and Arm® Mali™ GPUs architectures.\u003Cbr>\n\nThe library provides superior performance to other open source alternatives and immediate support for new Arm® technologies e.g. SVE2.\n\nKey Features:\n\n- Open source software available under a permissive MIT license\n- Over 100 machine learning functions for CPU and GPU\n- Multiple convolution algorithms (GeMM, Winograd, FFT, Direct and indirect-GeMM)\n- Support for multiple data types: FP32, FP16, INT8, UINT8, BFLOAT16\n- Micro-architecture optimization for key ML primitives\n- Highly configurable build options enabling lightweight binaries\n- Advanced optimization techniques such as kernel fusion, Fast math enablement and texture utilization\n- Device and workload specific tuning using OpenCL tuner and GeMM optimized heuristics\n\n## Documentation\n[![docs-shield]][docs-index]\n\n[docs-shield]: https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdocumentation-latest-green\n[docs-index]: https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Findex.xhtml\n\n> Note: The documentation includes the reference API, changelogs, build guide, contribution guide, errata, etc.\n\n\u003Cbr>\n\n## Pre-built binaries\nAll the binaries can be downloaded from [here](https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Freleases).\n\n\u003Cbr>\n\nPre-build binaries are generated with the following security \u002F good coding practices related flags:\n> -Wall, -Wextra, -Wformat=2, -Winit-self, -Wstrict-overflow=2, -Wswitch-default, -Woverloaded-virtual, -Wformat-security, -Wctor-dtor-privacy, -Wsign-promo, -Weffc++, -pedantic, -fstack-protector-strong\n\n## Supported Architectures\u002FTechnologies\n\n- Arm® CPUs:\n    - Arm® Cortex®-A processor family using Arm® Neon™ technology\n    - Arm® Neoverse™ processor family\n    - Arm® Cortex®-R processor family with Armv8-R AArch64 architecture using Arm® Neon™ technology\n    - Arm® Cortex®-X1 processor using Arm® Neon™ technology\n\n- Arm® Mali™ GPUs:\n    - Arm® Mali™-G processor family\n    - Arm® Mali™-T processor family\n\n- x86\n\n\u003Cbr>\n\n## Supported Systems\n\n- Android™\n- Bare Metal\n- Linux®\n- OpenBSD®\n- macOS®\n- Tizen™\n- QNX® (Experimental)\n- FreeBSD® (Experimental)\n\n\u003Cbr>\n\n## Resources\n- [Tutorial: Running AlexNet on Raspberry Pi with Compute Library](https:\u002F\u002Fcommunity.arm.com\u002Fprocessors\u002Fb\u002Fblog\u002Fposts\u002Frunning-alexnet-on-raspberry-pi-with-compute-library)\n- [Gian Marco's talk on Performance Analysis for Optimizing Embedded Deep Learning Inference Software](https:\u002F\u002Fwww.embedded-vision.com\u002Fplatinum-members\u002Farm\u002Fembedded-vision-training\u002Fvideos\u002Fpages\u002Fmay-2019-embedded-vision-summit)\n- [Gian Marco's talk on optimizing CNNs with Winograd algorithms at the EVS](https:\u002F\u002Fwww.embedded-vision.com\u002Fplatinum-members\u002Farm\u002Fembedded-vision-training\u002Fvideos\u002Fpages\u002Fmay-2018-embedded-vision-summit-iodice)\n- [Gian Marco's talk on using SGEMM and FFTs to Accelerate Deep Learning](https:\u002F\u002Fwww.embedded-vision.com\u002Fplatinum-members\u002Farm\u002Fembedded-vision-training\u002Fvideos\u002Fpages\u002Fmay-2016-embedded-vision-summit-iodice)\n\n\u003Cbr>\n\n## Experimental builds\n\n**⚠ Important** Bazel and CMake builds are experimental CPU only builds, please see the [documentation][docs-howtobuild] for more details.\n\n[docs-howtobuild]: https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Fhow_to_build.xhtml\n\n\u003Cbr>\n\n## How to contribute\n\nContributions to the Compute Library are more than welcome.\nIf you are interested in contributing, please have a look at our [how to contribute guidelines][docs-contributionguidelines].\n\n[docs-contributionguidelines]: https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Fcontribution_guidelines.xhtml\n\n### Developer Certificate of Origin (DCO)\nBefore the Compute Library accepts your contribution, you need to certify its origin and give us your permission. To manage this process we use the Developer Certificate of Origin (DCO) V1.1 (https:\u002F\u002Fdevelopercertificate.org\u002F)\n\nTo indicate that you agree to the the terms of the DCO, you \"sign off\" your contribution by adding a line with your name and e-mail address to every git commit message:\n\n```Signed-off-by: John Doe \u003Cjohn.doe@example.org>```\n\nYou must use your real name, no pseudonyms or anonymous contributions are accepted.\n\n### Public mailing list\nFor technical discussion, the ComputeLibrary project has a public mailing list: acl-dev@lists.linaro.org\nThe list is open to anyone inside or outside of Arm to self subscribe.  In order to subscribe, please visit the following website:\nhttps:\u002F\u002Flists.linaro.org\u002Fmailman3\u002Flists\u002Facl-dev.lists.linaro.org\u002F\n\n\u003Cbr>\n\n## License and Contributions\n\nThe software is provided under MIT license. Contributions to this project are accepted under the same license.\n\n### Other Projects\nThis project contains code from other projects as listed below. The original license text is included in those source files.\n\n* The OpenCL header library is licensed under Apache License, Version 2.0, which is a permissive license compatible with MIT license.\n\n* The half library is licensed under MIT license.\n\n* The libnpy library is licensed under MIT license.\n\n* The stb image library is either licensed under MIT license or is in Public Domain. It is used by this project under the terms of MIT license.\n\n* The KleidiAI library is licensed under Apache License, Version 2.0.\n\n* The GoogleTest library is used by KleidiAI and is licensed under BSD-3-Clause license.\n\n* The Benchmark library is used by KleidiAI and is licensed under Apache License, Version 2.0.\n\n\u003Cbr>\n\n## Trademarks and Copyrights\n\nAndroid is a trademark of Google LLC.\n\nArm, Cortex, Mali and Neon are registered trademarks or trademarks of Arm Limited (or its subsidiaries) in the US and\u002For elsewhere.\n\nBazel is a trademark of Google LLC., registered in the U.S. and other\ncountries.\n\nCMake is a trademark of Kitware, Inc., registered in the U.S. and other\ncountries.\n\nLinux® is the registered trademark of Linus Torvalds in the U.S. and other countries.\n\nMac and macOS are trademarks of Apple Inc., registered in the U.S. and other\ncountries.\n\nTizen is a registered trademark of The Linux Foundation.\n\nWindows® is a trademark of the Microsoft group of companies.\n\nQNX® is a trademark of QNX, a division of BlackBerry.\n\nFreeBSD® is a registered trademark of The FreeBSD Foundation.\n","> [!IMPORTANT]\n> **静态库名称变更**\n>\n> - 分发的预编译二进制文件中，libarm_compute-static.a 将被重命名为 libarm_compute.a。\n> - 此更改将于 2026 年 1 月或之后的第一个版本开始生效。\n>\n\n\u003Cbr>\n\u003Cdiv align=\"center\">\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FARM-software_ComputeLibrary_readme_686aa3674122.png\"\u002F>\u003Cbr>\u003Cbr>\n\u003C\u002Fdiv>\n\n# 计算库 [![acl-release-shield]][acl-release]\n\n[acl-release-shield]: https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002FARM-software\u002FComputeLibrary?label=latest&color=green\n[acl-release]: https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Freleases\u002Flatest\n\n计算库是一组针对 Arm® Cortex®-A、Arm® Neoverse™ 和 Arm® Mali™ GPU 架构优化的低层机器学习函数集合。\u003Cbr>\n\n该库在性能上优于其他开源替代方案，并能立即支持新的 Arm® 技术，例如 SVE2。\n\n主要特性：\n\n- 开源软件，采用宽松的 MIT 许可证\n- 针对 CPU 和 GPU 提供超过 100 种机器学习函数\n- 多种卷积算法（GeMM、Winograd、FFT、直接和间接 GeMM）\n- 支持多种数据类型：FP32、FP16、INT8、UINT8、BFLOAT16\n- 针对关键 ML 原语进行微架构优化\n- 可高度配置的构建选项，生成轻量级二进制文件\n- 先进的优化技术，如内核融合、快速数学启用和纹理利用\n- 使用 OpenCL 调优器和 GeMM 优化启发式方法进行设备和工作负载特定调优\n\n## 文档\n[![docs-shield]][docs-index]\n\n[docs-shield]: https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdocumentation-latest-green\n[docs-index]: https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Findex.xhtml\n\n> 注：文档包括参考 API、变更日志、构建指南、贡献指南、勘误表等。\n\n\u003Cbr>\n\n## 预编译二进制文件\n所有二进制文件均可从 [这里](https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Freleases) 下载。\n\n\u003Cbr>\n\n预编译二进制文件在构建时启用了以下安全和良好编码实践相关的编译标志：\n> -Wall, -Wextra, -Wformat=2, -Winit-self, -Wstrict-overflow=2, -Wswitch-default, -Woverloaded-virtual, -Wformat-security, -Wctor-dtor-privacy, -Wsign-promo, -Weffc++, -pedantic, -fstack-protector-strong\n\n## 支持的架构\u002F技术\n\n- Arm® CPU：\n    - 使用 Arm® Neon™ 技术的 Arm® Cortex®-A 系列处理器\n    - Arm® Neoverse™ 系列处理器\n    - 搭载 Armv8-R AArch64 架构并使用 Arm® Neon™ 技术的 Arm® Cortex®-R 系列处理器\n    - 使用 Arm® Neon™ 技术的 Arm® Cortex®-X1 处理器\n\n- Arm® Mali™ GPU：\n    - Arm® Mali™-G 系列处理器\n    - Arm® Mali™-T 系列处理器\n\n- x86\n\n\u003Cbr>\n\n## 支持的系统\n\n- Android™\n- 裸机\n- Linux®\n- OpenBSD®\n- macOS®\n- Tizen™\n- QNX®（实验性）\n- FreeBSD®（实验性）\n\n\u003Cbr>\n\n## 资源\n- [教程：使用计算库在树莓派上运行 AlexNet](https:\u002F\u002Fcommunity.arm.com\u002Fprocessors\u002Fb\u002Fblog\u002Fposts\u002Frunning-alexnet-on-raspberry-pi-with-compute-library)\n- [詹马可关于优化嵌入式深度学习推理软件性能分析的演讲](https:\u002F\u002Fwww.embedded-vision.com\u002Fplatinum-members\u002Farm\u002Fembedded-vision-training\u002Fvideos\u002Fpages\u002Fmay-2019-embedded-vision-summit)\n- [詹马可在 EVS 上关于使用 Winograd 算法优化 CNN 的演讲](https:\u002F\u002Fwww.embedded-vision.com\u002Fplatinum-members\u002Farm\u002Fembedded-vision-training\u002Fvideos\u002Fpages\u002Fmay-2018-embedded-vision-summit-iodice)\n- [詹马可关于使用 SGEMM 和 FFT 加速深度学习的演讲](https:\u002F\u002Fwww.embedded-vision.com\u002Fplatinum-members\u002Farm\u002Fembedded-vision-training\u002Fvideos\u002Fpages\u002Fmay-2016-embedded-vision-summit-iodice)\n\n\u003Cbr>\n\n## 实验性构建\n\n**⚠ 重要提示** Bazel 和 CMake 构建是仅限 CPU 的实验性构建，请参阅 [文档][docs-howtobuild] 获取更多详细信息。\n\n[docs-howtobuild]: https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Fhow_to_build.xhtml\n\n\u003Cbr>\n\n## 如何贡献\n\n我们非常欢迎对计算库的贡献。如果您有兴趣参与贡献，请查看我们的 [贡献指南][docs-contributionguidelines]。\n\n[docs-contributionguidelines]: https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Fcontribution_guidelines.xhtml\n\n### 开发者来源证书 (DCO)\n在计算库接受您的贡献之前，您需要证明其来源并向我们授权。为此，我们使用开发者来源证书 (DCO) V1.1（https:\u002F\u002Fdevelopercertificate.org\u002F）。\n\n要表示您同意 DCO 条款，您需要在每个 Git 提交信息中添加一行包含您的姓名和电子邮件地址的内容来“签署”您的贡献：\n\n```Signed-off-by: John Doe \u003Cjohn.doe@example.org>```\n\n您必须使用真实姓名，不接受假名或匿名贡献。\n\n### 公开邮件列表\n对于技术讨论，计算库项目有一个公开的邮件列表：acl-dev@lists.linaro.org\n该列表向 Arm 内部及外部的所有人开放，任何人都可以自行订阅。要订阅，请访问以下网站：\nhttps:\u002F\u002Flists.linaro.org\u002Fmailman3\u002Flists\u002Facl-dev.lists.linaro.org\u002F\n\n\u003Cbr>\n\n## 许可证与贡献\n\n本软件根据 MIT 许可证提供。对该项目的贡献也遵循相同的许可证。\n\n### 其他项目\n本项目包含来自以下列出的其他项目的代码。原始许可证文本已包含在这些源文件中。\n\n* OpenCL 头文件库根据 Apache License, Version 2.0 许可，这是一种与 MIT 许可证兼容的宽松许可证。\n\n* half 库根据 MIT 许可证许可。\n\n* libnpy 库根据 MIT 许可证许可。\n\n* stb image 库要么根据 MIT 许可证许可，要么属于公共领域。本项目在 MIT 许可证条款下使用了它。\n\n* KleidiAI 库根据 Apache License, Version 2.0 许可。\n\n* GoogleTest 库由 KleidiAI 使用，并根据 BSD-3-Clause 许可证许可。\n\n* Benchmark 库由 KleidiAI 使用，并根据 Apache License, Version 2.0 许可。\n\n\u003Cbr>\n\n## 商标与版权\n\nAndroid 是 Google LLC 的注册商标。\n\nArm、Cortex、Mali 和 Neon 是 Arm Limited（或其子公司）在美国及其他地区的注册商标或普通商标。\n\nBazel 是 Google LLC 在美国及其他国家注册的商标。\n\nCMake 是 Kitware, Inc. 在美国及其他国家注册的商标。\n\nLinux® 是 Linus Torvalds 在美国及其他国家注册的商标。\n\nMac 和 macOS 是 Apple Inc. 在美国及其他国家注册的商标。\n\nTizen 是 The Linux Foundation 注册的商标。\n\nWindows® 是微软公司集团的注册商标。\n\nQNX® 是黑莓公司旗下 QNX 部门的注册商标。\n\nFreeBSD® 是 The FreeBSD Foundation 注册的商标。","# Compute Library 快速上手指南\n\nCompute Library 是专为 Arm® Cortex®-A、Arm® Neoverse™ 和 Arm® Mali™ GPU 架构优化的底层机器学习函数库。它提供了超过 100 种针对 CPU 和 GPU 的机器学习原语，支持多种卷积算法和数据类型（FP32, FP16, INT8 等），旨在为嵌入式深度学习推理提供卓越性能。\n\n## 环境准备\n\n### 系统要求\nCompute Library 支持多种操作系统，主要包括：\n- **Linux®** (推荐)\n- **Android™**\n- **macOS®**\n- Bare Metal, OpenBSD®, Tizen™ 等\n- *注：QNX® 和 FreeBSD® 目前处于实验性支持阶段。*\n\n### 硬件架构\n- **CPU**: Arm® Cortex®-A, Arm® Neoverse™, Arm® Cortex®-R (Armv8-R AArch64), Arm® Cortex®-X1 (需支持 Arm® Neon™ 技术)。\n- **GPU**: Arm® Mali™-G 系列, Arm® Mali™-T 系列 (通过 OpenCL)。\n- **其他**: 也提供 x86 架构支持。\n\n### 前置依赖\n在从源码编译前，请确保安装以下工具链和依赖库（以 Ubuntu\u002FDebian 为例）：\n\n```bash\nsudo apt-get update\nsudo apt-get install -y build-essential scons python3 git libopencl-dev libblas-dev libboost-dev\n```\n\n*   **SCons**: 主要的构建工具。\n*   **Python3**: SCons 运行依赖。\n*   **libopencl-dev**: 若需启用 GPU (Mali) 支持必须安装。\n*   **libblas-dev**: 用于部分后端优化。\n\n> **注意**：如果你计划使用实验性的 CMake 或 Bazel 构建方式，请额外安装对应的 `cmake` 或 `bazel` 工具，但官方推荐生产环境使用 SCons。\n\n## 安装步骤\n\n虽然项目提供预编译二进制文件（Pre-built binaries），但为了针对特定硬件微架构进行极致优化，通常建议从源码编译。\n\n### 1. 克隆仓库\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary.git\ncd ComputeLibrary\n```\n\n### 2. 编译库\n使用 `scons` 命令进行编译。以下是一个典型的构建命令，启用了 OpenCL (GPU)、Neon (CPU) 并开启优化：\n\n```bash\nscons Werror=1 -j8 debug=0 asserts=0 neon=1 opencl=1 embed_kernels=1 extra_cxx_flags=\"-fPIC\"\n```\n\n**参数说明：**\n- `-j8`: 使用 8 个线程并行编译（根据 CPU 核心数调整）。\n- `debug=0 asserts=0`: 发布模式，关闭调试和断言以提升性能。\n- `neon=1`: 启用 Arm Neon 指令集优化。\n- `opencl=1`: 启用 OpenCL 后端以支持 Mali GPU。\n- `embed_kernels=1`: 将 OpenCL 内核嵌入二进制文件，避免运行时加载文件的开销（推荐用于嵌入式部署）。\n- `extra_cxx_flags=\"-fPIC\"`: 生成位置无关代码，便于链接到共享库。\n\n> **重要提示**：从 2026 年 1 月起的版本中，静态库名称将从 `libarm_compute-static.a` 变更为 `libarm_compute.a`。\n\n编译完成后，生成的库文件位于 `build\u002F` 目录下：\n- 头文件：`include\u002F`\n- 静态库：`build\u002Flibarm_compute.a` (或 `-static.a`)\n- 测试示例：`build\u002Ftests\u002F`\n\n### 3. (可选) 下载预编译包\n如果不需要自定义编译选项，可直接从 [Releases 页面](https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Freleases) 下载对应平台的预编译二进制包。这些包已包含安全编码实践相关的编译标志（如 `-fstack-protector-strong` 等）。\n\n## 基本使用\n\nCompute Library 主要作为底层加速库被其他框架（如 TensorFlow Lite, PyTorch Mobile）调用，也可直接用于开发自定义推理引擎。以下是一个最简单的 C++ 使用流程概念示例。\n\n### 1. 包含头文件与链接\n在您的 `CMakeLists.txt` 或 Makefile 中，需包含 `include` 路径并链接生成的库：\n\n```cmake\ninclude_directories(${COMPUTE_LIBRARY_PATH}\u002Finclude)\nlink_directories(${COMPUTE_LIBRARY_PATH}\u002Fbuild)\ntarget_link_libraries(your_app arm_compute)\n# 若使用 GPU，可能还需链接 OpenCL\ntarget_link_libraries(your_app OpenCL)\n```\n\n### 2. 代码示例：初始化与执行\n以下伪代码展示了如何初始化运行时并执行一个简单的张量操作（具体 API 请参考官方文档中的 Reference API）：\n\n```cpp\n#include \"arm_compute\u002Fcore\u002FTypes.h\"\n#include \"arm_compute\u002Fruntime\u002FNEON\u002FNEFunctions.h\"\n#include \"arm_compute\u002Fruntime\u002FTensor.h\"\n#include \"arm_compute\u002Fruntime\u002FTensorAllocator.h\"\n\nusing namespace arm_compute;\n\nint main() {\n    \u002F\u002F 1. 定义张量形状和数据类型\n    const TensorShape shape(256, 256); \n    const DataType data_type = DataType::F32;\n\n    \u002F\u002F 2. 创建并分配张量内存\n    Tensor src;\n    src.allocator().init(TensorInfo(shape, 1, data_type));\n    src.allocator().allocate();\n\n    Tensor dst;\n    dst.allocator().init(TensorInfo(shape, 1, data_type));\n    dst.allocator().allocate();\n\n    \u002F\u002F 3. 配置并执行函数 (例如：简单的像素转换或矩阵运算)\n    \u002F\u002F 此处以 NEFillBorder 或其他算子为例，实际需根据具体算子配置\n    \u002F\u002F NECopy copy_func; \n    \u002F\u002F copy_func.configure(&src, &dst);\n    \n    \u002F\u002F 对于大多数算子，配置后需执行\n    \u002F\u002F copy_func.run();\n\n    \u002F\u002F 4. 释放内存\n    src.allocator().free();\n    dst.allocator().free();\n\n    return 0;\n}\n```\n\n### 3. 运行测试用例\n编译后的 `build\u002Ftests` 目录包含大量单元测试和功能验证示例，可用于验证当前环境的兼容性：\n\n```bash\n# 运行所有单元测试 (可能需要较长时间)\n.\u002Fbuild\u002Ftests\u002Funit_tests\n\n# 运行特定的验证测试 (例如验证卷积实现)\n.\u002Fbuild\u002Ftests\u002Fvalidation\u002FCL\u002FConvolutionLayer\n```\n\n更多详细用法、算子列表及高级调优技巧（如内核融合、OpenCL Tuner 使用），请参阅 [官方文档](https:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Flatest\u002Findex.xhtml)。","某嵌入式开发团队正在为基于 Arm Cortex-A 处理器的智能安防摄像头部署实时人脸检测算法，需在低功耗下保证高帧率运行。\n\n### 没有 ComputeLibrary 时\n- **推理延迟高**：使用通用开源库无法充分利用 Arm Neon SIMD 指令集，单帧处理耗时超过 80ms，导致视频流卡顿，无法达到实时检测标准。\n- **功耗失控**：由于缺乏针对特定微架构的底层优化，CPU 长期处于高负载状态，设备发热严重且电池续航大幅缩短。\n- **部署适配难**：面对 FP16 或 INT8 等量化模型需求，需手动编写大量汇编代码进行算子适配，开发周期长达数周且极易出错。\n- **硬件利用率低**：无法自动选择最优卷积算法（如 Winograd 或 FFT），导致 Mali GPU 算力闲置，整体吞吐量远低于硬件理论上限。\n\n### 使用 ComputeLibrary 后\n- **实时性能达标**：借助深度优化的 Neon 内核与 SVE2 技术，单帧处理时间降至 25ms 以内，轻松实现 30FPS 流畅检测。\n- **能效显著提升**：通过微架构级调优和算子融合技术，在同等算力下功耗降低 40%，有效解决设备过热问题。\n- **开发效率飞跃**：直接调用支持 FP16\u002FINT8 的百余个现成函数，无需底层手写优化，模型移植时间从数周压缩至数天。\n- **算力充分释放**：内置的 GeMM 启发式策略自动匹配最佳卷积算法，使 CPU 与 GPU 协同工作，吞吐量提升 3 倍以上。\n\nComputeLibrary 通过将底层硬件特性转化为开箱即用的高性能算子，让开发者在 Arm 生态中实现了极致能效与实时响应的完美平衡。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FARM-software_ComputeLibrary_094bc23e.png","ARM-software","Arm Software","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002FARM-software_8286f64a.png","",null,"www.arm.com","https:\u002F\u002Fgithub.com\u002FARM-software",[83,87,91,95,99,102,106,109],{"name":84,"color":85,"percentage":86},"C++","#f34b7d",92.7,{"name":88,"color":89,"percentage":90},"C","#555555",6.4,{"name":92,"color":93,"percentage":94},"Python","#3572A5",0.5,{"name":96,"color":97,"percentage":98},"CMake","#DA3434",0.2,{"name":100,"color":101,"percentage":98},"Starlark","#76d275",{"name":103,"color":104,"percentage":105},"Shell","#89e051",0,{"name":107,"color":108,"percentage":105},"Go","#00ADD8",{"name":110,"color":111,"percentage":105},"Ruby","#701516",3126,817,"2026-04-02T22:14:48",4,"Linux, macOS, Android, Bare Metal, OpenBSD, Tizen, QNX (实验性), FreeBSD (实验性)","非必需。若使用 GPU，需 Arm Mali-G 或 Mali-T 系列显卡（通过 OpenCL），不支持 NVIDIA CUDA。","未说明",{"notes":120,"python":121,"dependencies":122},"该库主要针对 Arm 架构（Cortex-A, Neoverse, Mali GPU）优化，支持 x86 但核心优势在 Arm。实验性的 Bazel 和 CMake 构建仅支持 CPU。静态库名称将在 2026 年 1 月后变更。贡献代码需签署 DCO 协议并使用真实姓名。","未说明 (主要为 C++ 库，提供实验性构建支持)",[123,124,125,126,127],"OpenCL headers (Apache 2.0)","half library (MIT)","libnpy (MIT)","stb image (MIT\u002FPublic Domain)","KleidiAI (Apache 2.0)",[14,13,15],[130,131,132,133,134,135,136,137,138,139,140,141,142,143],"neon","opencl","computer-vision","arm","armv7","armv8","aarch64","machine-learning","simd","android","cpp","neural-network","sve","linux","2026-03-27T02:49:30.150509","2026-04-06T05:19:35.450608",[147,152,157,162,167,171],{"id":148,"question_zh":149,"answer_zh":150,"source_url":151},12249,"为什么在 AArch64 设备上运行程序时报错 \"No such file or directory\" 或无法执行？","这通常是因为操作系统架构与二进制文件不匹配。如果 `uname -a` 显示 `armv8l`，说明你正在 AArch64 兼容核心（如 Cortex A53）上运行 32 位操作系统。这种情况下，系统只能执行 32 位二进制文件（构建时使用 arm7a）。如果你尝试运行 64 位（AArch64）二进制文件或链接 64 位库，即使硬件支持，也会因为 OS 限制而失败。解决方法是确认你的 OS 位数，并重新构建对应架构（32 位或 64 位）的 Arm Compute Library 和应用程序。","https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Fissues\u002F842",{"id":153,"question_zh":154,"answer_zh":155,"source_url":156},12250,"升级版本后（如 v24.11），发现 NEDeconvolutionLayer 性能下降怎么办？","这是一个已知的回归问题，维护者已经提供了修复补丁。性能下降是由于特定层的实现变更导致的。你需要应用以下两个修复补丁来恢复性能：\n1. 针对 NEDeconvLayer 的修复：https:\u002F\u002Freview.mlplatform.org\u002Fc\u002Fml\u002FComputeLibrary\u002F+\u002F13408\n2. 针对 NEConvLayer 的修复：https:\u002F\u002Freview.mlplatform.org\u002Fc\u002Fml\u002FComputeLibrary\u002F+\u002F13418\n这些修复将包含在 v25.01 及以后的版本中。如果是源码编译用户，可以手动 cherry-pick 这些提交或等待新版本发布。","https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Fissues\u002F1150",{"id":158,"question_zh":159,"answer_zh":160,"source_url":161},12251,"如何在 openSUSE Tumbleweed 或 Leap 15.0+ 上解决构建时的 'TypeError: cannot use a string pattern on a bytes-like object' 错误？","这是由 Python 3 中字符串和字节处理方式的变更引起的构建脚本兼容性问题，特别是在较新的 openSUSE 版本上。该问题已被确认为 Bug 并修复。解决方案是更新到包含修复补丁的最新主分支代码，或者等待包含该修复的下一个正式版本发布。补丁已合并到 main 分支，重新拉取最新代码即可解决。","https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Fissues\u002F619",{"id":163,"question_zh":164,"answer_zh":165,"source_url":166},12252,"如何在 Graph API 中集成自定义的卷积层或使用外部硬件加速器？","Arm Compute Library 的 Graph API 主要用于编排现有的 ACL 层。若要集成自定义算子或使用外部加速器（如 FPGA 逻辑），目前并没有直接的“插入”接口。通常的做法是：\n1. 理解 Graph 执行流程：`graph.run()` 会调用 `GraphManager::execute_graph`，进而执行所有任务。\n2. 数据流转：每一层的输出 Tensor 会自动作为下一层的输入。\n3. 自定义集成策略：你可能需要在图构建前后手动管理 Tensor 数据。例如，在 `prepare()` 阶段或将数据从 ACL Tensor 复制到加速器可访问的内存区域，执行加速计算后，再将结果写回 Tensor 供后续层使用。这需要深入理解 ACL 的内存管理和生命周期，目前官方建议参考源码中的 `call_all_tasks` 逻辑进行定制开发。","https:\u002F\u002Fgithub.com\u002FARM-software\u002FComputeLibrary\u002Fissues\u002F653",{"id":168,"question_zh":169,"answer_zh":170,"source_url":151},12253,"构建 Arm Compute Library 时如何指定正确的架构和选项以避免运行时错误？","构建时必须确保 `scons` 参数与目标设备的实际硬件和操作系统匹配。常见配置示例如下：\n- 对于 64 位 ARM (AArch64) 设备：`scons arch=arm64-v8a ...`\n- 对于 32 位 ARM 设备：`scons arch=armv7a ...`\n- 启用 OpenCL：`opencl=1` (需设备支持)\n- 启用 NEON：`neon=1`\n如果在 32 位 OS 上构建了 64 位库（或反之），会导致运行时无法加载库或报错 \"No such file or directory\"。务必使用 `uname -m` 检查机器架构，并结合 `getconf LONG_BIT` 或 `file` 命令确认 OS 位数，再选择对应的 `arch` 参数进行构建。",{"id":172,"question_zh":173,"answer_zh":174,"source_url":156},12254,"如何验证 NEDeconvolutionLayer 的配置是否正确？","在配置层之前，应先使用 `NEDeconvolutionLayer::validate` 函数验证参数合法性。示例代码如下：\n```cpp\nauto status = NEDeconvolutionLayer::validate(&srcTensorInfo, &weiTensorInfo, nullptr, &dstTensorInfo, deconvInfo, fastMath);\nif(status.error_code() != ErrorCode::OK) {\n  std::cout \u003C\u003C \"ERROR: \" \u003C\u003C status.error_description().c_str() \u003C\u003C std::endl;\n  exit(1);\n}\nstd::cout \u003C\u003C \"PASSED VALIDATION\" \u003C\u003C std::endl;\n```\n只有验证通过后，才调用 `configure` 进行实际配置。这可以避免因形状、数据类型或步长不匹配导致的运行时崩溃。",[176,181,186,191,196,201,206,211,216,221,226,231,236,241,246,251,256,261,266,271],{"id":177,"version":178,"summary_zh":179,"released_at":180},62622,"v52.8.0","## v52.8.0 公开小版本发布\n### 功能\n- 添加 SME GEMM 和 GEMV 内核。\n- 添加细粒度的 SME 特性标志。\n- 放宽所有算子的支持尺寸配置检查。\n- 在 Windows(R) 中检测除 FEAT_FHM 之外的所有可用硬件特性。\n\n### 修复\n- TensorShape::collapse 和 total_size 中对 std::accumulate 的误用。\n- CL 后端 Fp16 MMUL 内核中针对小 N 的无效内核启发式方法和验证。\n- 防止使用 clang-cl 构建时可能出现的溢出。\n- 停止在 Git 历史记录中忽略 txt 文件。\n- half.hpp 包含文件中潜在的 -Wdeprecated-literal-operator 错误。\n\n文档（API、构建指南、贡献指南、勘误表等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.8.0\u002Findex.xhtml","2026-01-23T12:25:49",{"id":182,"version":183,"summary_zh":184,"released_at":185},62623,"v52.7.0","## v52.7.0 公开小版本发布\n\n### 功能\n- 添加单ISA支持，并在多ISA CMake 构建中启用 SME。\n- 为 gemm 核心添加 SME1 FP32 支持。\n- 通过 SCons 在仅包含 Linux(R) CPU 的多 ISA 构建中提供 libarm_compute.a。\n\n### 修复\n- 对 MMUL 扩展版本低于 1.1.0 的情况，排除 FP16 MMUL Reshaped RHS 内核。\n- 从公共头文件中移除私有包含。\n- 在 Android(TM) 和 Linux(R) 中检测 FEAT_SME2。\n\n### 重构\n- 简化验证\u002FCPP 和 UNIT 中嵌套的 concat\u002Fcombine\u002Fzip 使用。\n- 简化验证\u002FCL 和 CPP 中嵌套的 concat 使用。\n- 简化验证\u002FNEON 中嵌套的 concat 使用。\n\n文档（API、构建指南、贡献指南、勘误表等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.7.0\u002Findex.xhtml","2025-12-05T12:08:46",{"id":187,"version":188,"summary_zh":189,"released_at":190},62624,"v52.6.0","## v52.6.0 公开小版本发布\n\n### 功能\n- 在量化版 CpuGemmConv2d 中启用 F32 输出\n\n### 修复\n- 如果张量尺寸较大，则使某些 CPU 操作失效\n- CpuGemmDirectConv2d 中缺少输出类型验证\n- 在 CpuActivation 中处理 configure() 之后的填充更新\n\n### 重构\n- 简化 validation\u002FNEON 目录中嵌套的 zip 使用\n- 简化 validation\u002FCL 目录中嵌套的 combine 和 zip 使用\n- 简化 validation\u002FNEON 目录中嵌套的 combine 使用\n\n### 性能\n- 对 FP16 反演仅执行一次细化迭代\n\n文档（API、构建指南、贡献指南、勘误表等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.6.0\u002Findex.xhtml","2025-10-31T16:12:16",{"id":192,"version":193,"summary_zh":194,"released_at":195},62625,"v52.5.0","## v52.5.0 公开小版本发布\n### 功能\n- 在 CPU 和 GPU 平台上添加性能剖析跟踪点\n- 将 Perfetto 性能分析器设为默认后端\n- 进一步现代化 CMake 构建系统\n- 添加 CMakePresets.json\n\n### 修复\n- 处理 CpuActivation 中 configure() 后的填充更新问题\n- 渲染后的未发布 README.md 中存在损坏的 URL\n- 使用 CMake 构建时，macOS 上出现链接器错误\n\n### 性能\n- 添加 FP16 GEMM MMUL 仅重排右侧矩阵内核\n\n文档（API、构建指南、贡献指南、勘误等）请访问：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.5.0\u002Findex.xhtml","2025-10-13T21:10:20",{"id":197,"version":198,"summary_zh":199,"released_at":200},62626,"v52.4.0","## v52.4.0 公开小版本发布\n### 通知\n- 针对 macOS 和 Windows 的预编译二进制文件的生成目前正处于审核中，可能在本次发布后暂时不可用。\n\n### 功能\n- 更新了用于静态量化操作符 CpuGEMMLowp 及其相关测试。\n\n### 修复\n- 修复了 CpuFullyConnected 验证方法中可能出现的空指针访问问题。\n\n### 性能\n- 移除了激活核函数中的 switch 语句。\n\n文档（API、构建指南、贡献指南、勘误表等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.4.0\u002Findex.xhtml\n","2025-08-27T07:44:57",{"id":202,"version":203,"summary_zh":204,"released_at":205},62627,"v52.3.0","## v52.3.0 公开小版本发布\n### 功能\n- 在 NEQuantizationLayer 中支持 QSYMM8_PER_CHANNEL。\n- 为 CpuFullyConnected 添加无状态包装器。\n\n### 修复\n- 在配置后更新量化时，支持混合类型量化矩阵乘法。\n- 防止在 GEMM 计算行和时发生越界读取。\n- 解决 Dimensions::collapse() 中的越界访问问题。\n\n### 性能\n- 移除 SVE 激活函数中的 switch 语句。\n- 移除 SVE2 激活函数中的 switch 语句。\n\n文档（API、构建指南、贡献指南、勘误等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.3.0\u002Findex.xhtml","2025-07-04T14:02:55",{"id":207,"version":208,"summary_zh":209,"released_at":210},62628,"v52.2.0","## v52.2.0 公开小版本发布\n### 功能\n- 启用非转置的 BF16 重排。\n\n### 修复\n- 修复多 ISA 构建中的重排测试失败问题。\n- 修复 a64_hgemm_8x24 中操作数过早预取的问题。\n\n文档（API、构建指南、贡献指南、勘误等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.2.0\u002Findex.xhtml","2025-06-13T09:01:03",{"id":212,"version":213,"summary_zh":214,"released_at":215},62629,"v52.1.0","## v52.1.0 公开小版本发布\n### 功能\n- 将 GEMM 无状态执行限制为仅支持固定格式的内核\n- 添加包装类以公开 cpu::CpuPool2d 功能\n- 启用非转置的 F32 重新排序\n\n文档（API、构建指南、贡献指南、勘误表等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.1.0\u002Findex.xhtml","2025-06-02T09:04:18",{"id":217,"version":218,"summary_zh":219,"released_at":220},62630,"v52.0.1","## v52.0.1 公开补丁版本\n### 修复\n- 在 CpuIm2ColKernel 中用零填充填充区域\n- 公共头文件通过 -Wundef 检查\n- 将 run_parallel_pretranspose_B_array 的线程拆分限制为窗口大小\n\n文档（API、构建指南、贡献指南、勘误表等）请访问：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.0.1\u002Findex.xhtml","2025-05-15T09:10:24",{"id":222,"version":223,"summary_zh":224,"released_at":225},62631,"v52.0.0","## v52.0.0 公开重大发布\n### 修复\n- 使 NEReorderLayer 向后兼容\n- 为 Datatype::BFLOAT16 添加字符串转换\n- 为 Winograd 变换添加缺失的头文件，以更好地处理剩余数据\n- 更新 3x3 Winograd 系数，以提高数值稳定性\n文档（API、构建指南、贡献指南、勘误等）请参见：\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv52.0.0\u002Findex.xhtml","2025-05-01T15:32:14",{"id":227,"version":228,"summary_zh":229,"released_at":230},62632,"v25.04","## v25.04 Public Major Release\n### Feat\n- Add Neon(TM) and SVE hybrid FP16 matmul kernels using FP32 accumulation.\n### Fix\n- Fix BF16 CpuGemmAssembly tests.\n- SME softmax FP32 kernel failing given large inputs.\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv25.04\u002Findex.xhtml","2025-04-17T13:01:15",{"id":232,"version":233,"summary_zh":234,"released_at":235},62633,"v25.03.1","## v25.03.1 Public Major Release\n### Feat\n- Add experimental QNX(R) support.\n- Add matmul fp16->fp32 kernels to enable fp16 PyTorch attention through ACL.\n### Fix\n- Replace .word with .inst when encoding instructions.\n- Neon(TM) detection for Bare Metal.\n### Refactor\n- Refactor reorder kernel and layer.\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv25.03.1\u002Findex.xhtml","2025-04-04T14:05:10",{"id":237,"version":238,"summary_zh":239,"released_at":240},62634,"v25.03","## v25.03 Public Major Release\n### Feat\n- Notice: Migration to Semantic Versioning will take place by the end of April\n- Modernize ACL CMake build\n- Add a wrapper class for CpuPRelu operators\n### Fix\n- Validation in Cpu Deconv for negative padded cases\n- Reserved register list in [U]Int8 SME2 Softmax kernels\n- Register allocation in [U]Int8 SME2 Softmax kernels\n- C and C++ build flags assigned to proper SCons flags\n- Don't pass filenames to the check-bad-style pre-commit hook\n- Apply -fPIC flag both to C and C++ code\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv25.03\u002Findex.xhtml","2025-03-21T11:00:17",{"id":242,"version":243,"summary_zh":244,"released_at":245},62635,"v25.02.1","## v25.02.1 Public Major Release\n### Feat\n- Add stateless support for GEMM kernels that need working_space\n- Add extra_cc_flags flag to SCons\n### Fix\n- Enable wrapper tests\n- Refactor format_code.py and pre-commit config\n- Adjust tolerance in CPP\u002FDFT\u002FDFT1D\u002FComplex test\n### Refactor\n- Remove dynamic fusion and compute kernel writer files and mentions\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv25.02.1\u002Findex.xhtml","2025-03-07T10:02:12",{"id":247,"version":248,"summary_zh":249,"released_at":250},62636,"v25.02","## v25.02 Public Major Release\n### Feat\n- Detect number of CPU cores in OpenBSD\n- Support tensors with dynamic shapes in NEGEMM\n- Support FP16 dequantization in NEGEMMLowpMatrixMultiplyCore\n- Add a public API for CpuMeanStdDevNormalization\n- Enable BF16 inputs in CpuFullyConnected\n### Fix\n- Linking errors in C++17 while compiling with clang\n- False positive compiler warning stringop-overflow\n- Redundant declaration warning of constexpr static data member (in C++17)\n- Make GemmLowp return an error in validate when F16 is not supported\n- Reorder interleave_by in CpuGemmAssemblyDispatch test code\n- Gemm_hybrid_quantized.hpp was passing incorrect K size to the kernel\n- Wrong kernel choice in CpuMul when build does not have SME2\n- Incorrect scheduling hint heuristic for GEMMs\n- Incorrect trademark usage in Readme for Arm(R)-Neoverse(TM) core\n### Refactor\n- Use operator API inside NEMeanstdDevNormalizationLayer\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv25.02\u002Findex.xhtml","2025-02-17T16:40:38",{"id":252,"version":253,"summary_zh":254,"released_at":255},62637,"v25.01","## v25.01 Public Major Release\n### Feat\n- Add KleidiAI as third_party module\n- Add NHWC FP16 kernels in CpuDirectConv\n- Add support of all non-quantized data types for NEScatter\n- Implement NEScatter for FP32 for all size configurations for Add\u002FSub\u002FMin\u002FMax\u002FUpdate\n- Add option to print time used by each iteration in the validation suite\n- Support multi ISA build for macOS\n### Fix\n- Performance regression in NEDeconvolutionLayer\n- Performance regression in NEConvolutionLayer\n- Usages of dynamic shapes in the library\n- Use separate build flags for C and C++ for CMake\n- Compiler error with gcc14 in 3rd party header stb_image\n- Werror=noexcept compilation issue in NEScatter\n- Unused tolerance_f16 in non-F16 builds\n- SegFault in SME Softmax Int8 tests\n- Disable pre-commit copyright validation for outside contributions\n- SME2 interleaved s8 x s8 = f32 kernel mismatches\n- Invalidate Bf16 Softmax when FEAT_SVE is not present and fix the tests\n- Illegal instruction caused by SVE instruction outside streaming mode\n- SME Winograd output transform 4x4_3x3 kernel\n- Misspell in SConstruct:301: 'estate' to 'arch'\n### Refactor\n- Removed deprecated NCHW kernels from CpuDirectConv2d\n- Check pre-commit copyright, Android.bp and formatting separately\n### Perf\n- Choose latest Gpu if Gpu name is not recognized and alter GEMM heuristics\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv25.01\u002Findex.xhtml","2025-01-30T17:05:10",{"id":257,"version":258,"summary_zh":259,"released_at":260},62638,"v24.12","## v24.12 Public Major Release\n### Feat\n- Add a build flag to make scheduler object thread_local and make it default in Bazel build\n### Fix\n- CPU regression in Reshape from excess threads\n- NEDeconvolutionLayer regression\n- Ensure bias type is BF16 for BF16 indirect convolutions\n### Perf\n- Disable mmul kernel selection for fp16 in GPU backend\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv24.12\u002Findex.xhtml","2024-12-19T11:46:54",{"id":262,"version":263,"summary_zh":264,"released_at":265},62639,"v24.11.1","## v24.11.1 Public Minor Release\n### Feat\n- Add stateless GEMM execution via ICPPKernel::run_op\n- TensorShape class supports dynamic shapes\n- Add skeletons for Dynamic GEMM operator\n- Convert Double rounding to Single rounding quantization behaviour in both Cpu\u002FGpu backend\n### Fix\n- Detect Advanced SIMD support on Windows®\n### Perf\n- Implement activation heuristics for Neoverse™ V1\n- Optimize PReLU on quantized datatypes\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv24.11.1\u002Findex.xhtml","2024-12-02T17:46:15",{"id":267,"version":268,"summary_zh":269,"released_at":270},62640,"v24.11","## v24.11 Public Major Release\n### Feat\n- Add SVE SoftmaxLayer kernel for BF16\n- Provide stateless API for CpuGemmLowpMatrixMultiplyCore, CpuQuantize, and DequantizationLayer\n- Extend static quantization interface for both matmul and convolution operations\n### Fix\n- Clarify Third-Party IP licenses\n- Check if CpuGemmAssemblyDispatch is configured in CpuMatMul before continue\n- Add BF16 support for CpuGemmAssemblyDispatchWrapper\n- Detect SVE support on Windows® to run the available kernels\n- Fixed missing cstdint include which occurs with GCC 15\n- Disable -O2 when building for Windows® as this crashes when certain compiler versions are used\n- Make cast on CPU truncate float to int instead of round to be consistent with other ML frameworks\n- Return error in validate() for CpuGemmLowpMatrixMultiplyCore if pretransposed A or B are true as this is not supported\n- Avoid implicit conversion from __fp16 to arm_compute::bfloat16 to avoid illegal instructions in hardware with FP16 but no BF16 support\n- Softmax SME2 kernel selection now correctly detects if SME2 is supported\n- Requantization rounding issues in CPU\u002FGPU Quantize\n- Scale normalising coefficient in GPU LogSoftmax\n- Apply consistent rounding policy in NEReduceMean\n- Revert default memory manager for NEQLSTMLayer\n- Create default memory manager when none is provided\n### Refactor\n- Turn duplicated code in the elementwise_binary kernel into templates to reduce code size\n- Move CpuSoftmaxKernel LUT to LUTManager to consolidate location of all LUTs\n### Perf\n- Use SME instead of SVE for subtractions in SoftmaxLayer for Q8 relating to LUT address calculation\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv24.11\u002Findex.xhtml","2024-11-18T11:53:52",{"id":272,"version":273,"summary_zh":274,"released_at":275},62641,"v24.09","## v24.09 Public Major Release\r\n\r\n\r\n### Feat\r\n\r\n\r\n- Provide a wrapper class to expose cpu::CpuSoftmaxGeneric\r\n\r\n- Detect number of cores in Windows®\r\n\r\n- Add Optimized SME kernel for QASYMM8_SIGNED elementwise addition operation\r\n\r\n\r\n### Fix\r\n\r\n\r\n- LogSoftmax Int8\u002FUInt8 mismatches in Cpu\r\n\r\n- Rounding of negative integers in pooling 2d\u002F3d gpu kernels\r\n\r\n- OpenMP® linker error on Windows®\r\n\r\n- Rounding of negative integers in pooling 2d\u002F3d kernels\r\n\r\n- Patches linker failure for cpu::CpuSoftmaxGeneric in partial builds\r\n\r\n- Cpu\u002FGpu Reverse data type support\r\n\r\n- QSYMM16 broadcasted subtraction failures\r\n\r\n- CpuMulKernel validation when there is x-broadcasting for some types\r\n\r\n- Data type validation in depthwise op in Cpu\r\n\r\n- Update macOS® build instructions\r\n\r\n- Validation tests compute reference and target on each iteration\r\n\r\n- Reset permuted input and weights on configure in NEDepthwiseConvolutionLayer\r\n\r\n- Selectively enable CL job chaining\r\n\r\n\r\n### Refactor\r\n\r\n\r\n- Generate only one shared library when building with CMake\r\n\r\n- Add BF16 LUT for Softmax Layer with tests\r\n\r\n- Move heuristic logic of activation kernel into separate class\r\n\r\n- Removed unused CommandBuffer.\r\n\r\n\r\n### Perf\r\n\r\n\r\n- Allocate Persistent and Prepare tensors at start of prepare()\r\n\r\n- Use mws in OMPScheduler for better thread throttling\r\n\r\n- Enable FP16 winograd in CpuConv2d for v8a multi_isa builds.\r\n\r\n\r\nDocumentation (API, build guide, contribution guide, errata, etc.) available here:\r\nhttps:\u002F\u002Fartificial-intelligence.sites.arm.com\u002Fcomputelibrary\u002Fv24.09\u002Findex.xhtml\r\n","2024-09-27T13:56:50"]