[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-Tencent--ncnn":3,"tool-Tencent--ncnn":61},[4,18,28,37,45,53],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":17},4358,"openclaw","openclaw\u002Fopenclaw","OpenClaw 是一款专为个人打造的本地化 AI 助手，旨在让你在自己的设备上拥有完全可控的智能伙伴。它打破了传统 AI 助手局限于特定网页或应用的束缚，能够直接接入你日常使用的各类通讯渠道，包括微信、WhatsApp、Telegram、Discord、iMessage 等数十种平台。无论你在哪个聊天软件中发送消息，OpenClaw 都能即时响应，甚至支持在 macOS、iOS 和 Android 设备上进行语音交互，并提供实时的画布渲染功能供你操控。\n\n这款工具主要解决了用户对数据隐私、响应速度以及“始终在线”体验的需求。通过将 AI 部署在本地，用户无需依赖云端服务即可享受快速、私密的智能辅助，真正实现了“你的数据，你做主”。其独特的技术亮点在于强大的网关架构，将控制平面与核心助手分离，确保跨平台通信的流畅性与扩展性。\n\nOpenClaw 非常适合希望构建个性化工作流的技术爱好者、开发者，以及注重隐私保护且不愿被单一生态绑定的普通用户。只要具备基础的终端操作能力（支持 macOS、Linux 及 Windows WSL2），即可通过简单的命令行引导完成部署。如果你渴望拥有一个懂你",349277,3,"2026-04-06T06:32:30",[13,14,15,16],"Agent","开发框架","图像","数据工具","ready",{"id":19,"name":20,"github_repo":21,"description_zh":22,"stars":23,"difficulty_score":24,"last_commit_at":25,"category_tags":26,"status":17},9989,"n8n","n8n-io\u002Fn8n","n8n 是一款面向技术团队的公平代码（fair-code）工作流自动化平台，旨在让用户在享受低代码快速构建便利的同时，保留编写自定义代码的灵活性。它主要解决了传统自动化工具要么过于封闭难以扩展、要么完全依赖手写代码效率低下的痛点，帮助用户轻松连接 400 多种应用与服务，实现复杂业务流程的自动化。\n\nn8n 特别适合开发者、工程师以及具备一定技术背景的业务人员使用。其核心亮点在于“按需编码”：既可以通过直观的可视化界面拖拽节点搭建流程，也能随时插入 JavaScript 或 Python 代码、调用 npm 包来处理复杂逻辑。此外，n8n 原生集成了基于 LangChain 的 AI 能力，支持用户利用自有数据和模型构建智能体工作流。在部署方面，n8n 提供极高的自由度，支持完全自托管以保障数据隐私和控制权，也提供云端服务选项。凭借活跃的社区生态和数百个现成模板，n8n 让构建强大且可控的自动化系统变得简单高效。",184740,2,"2026-04-19T23:22:26",[16,14,13,15,27],"插件",{"id":29,"name":30,"github_repo":31,"description_zh":32,"stars":33,"difficulty_score":10,"last_commit_at":34,"category_tags":35,"status":17},10095,"AutoGPT","Significant-Gravitas\u002FAutoGPT","AutoGPT 是一个旨在让每个人都能轻松使用和构建 AI 的强大平台，核心功能是帮助用户创建、部署和管理能够自动执行复杂任务的连续型 AI 智能体。它解决了传统 AI 应用中需要频繁人工干预、难以自动化长流程工作的痛点，让用户只需设定目标，AI 即可自主规划步骤、调用工具并持续运行直至完成任务。\n\n无论是开发者、研究人员，还是希望提升工作效率的普通用户，都能从 AutoGPT 中受益。开发者可利用其低代码界面快速定制专属智能体；研究人员能基于开源架构探索多智能体协作机制；而非技术背景用户也可直接选用预置的智能体模板，立即投入实际工作场景。\n\nAutoGPT 的技术亮点在于其模块化“积木式”工作流设计——用户通过连接功能块即可构建复杂逻辑，每个块负责单一动作，灵活且易于调试。同时，平台支持本地自托管与云端部署两种模式，兼顾数据隐私与使用便捷性。配合完善的文档和一键安装脚本，即使是初次接触的用户也能在几分钟内启动自己的第一个 AI 智能体。AutoGPT 正致力于降低 AI 应用门槛，让人人都能成为 AI 的创造者与受益者。",183572,"2026-04-20T04:47:55",[13,36,27,14,15],"语言模型",{"id":38,"name":39,"github_repo":40,"description_zh":41,"stars":42,"difficulty_score":10,"last_commit_at":43,"category_tags":44,"status":17},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,"2026-04-05T11:01:52",[14,15,13],{"id":46,"name":47,"github_repo":48,"description_zh":49,"stars":50,"difficulty_score":24,"last_commit_at":51,"category_tags":52,"status":17},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",161692,"2026-04-20T11:33:57",[14,13,36],{"id":54,"name":55,"github_repo":56,"description_zh":57,"stars":58,"difficulty_score":24,"last_commit_at":59,"category_tags":60,"status":17},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。",109154,"2026-04-18T11:18:24",[14,15,13],{"id":62,"github_repo":63,"name":64,"description_en":65,"description_zh":66,"ai_summary_zh":66,"readme_en":67,"readme_zh":68,"quickstart_zh":69,"use_case_zh":70,"hero_image_url":71,"owner_login":72,"owner_name":72,"owner_avatar_url":73,"owner_bio":74,"owner_company":75,"owner_location":75,"owner_email":75,"owner_twitter":75,"owner_website":76,"owner_url":77,"languages":78,"stars":108,"forks":109,"last_commit_at":110,"license":111,"difficulty_score":112,"env_os":113,"env_gpu":114,"env_ram":115,"env_deps":116,"category_tags":120,"github_topics":121,"view_count":24,"oss_zip_url":75,"oss_zip_packed_at":75,"status":17,"created_at":141,"updated_at":142,"faqs":143,"releases":173},10191,"Tencent\u002Fncnn","ncnn","ncnn is a high-performance neural network inference framework optimized for the mobile platform","ncnn 是由腾讯开源的一款高性能神经网络前向计算框架，专为移动端设备极致优化。它的核心使命是解决深度学习模型在手机等资源受限设备上“跑不动”或“跑得慢”的难题，让开发者能够轻松将复杂的 AI 算法部署到移动端，从而打造出响应迅速的智能应用。\n\n无论是希望为 App 集成人脸识别、图像风格迁移功能的移动开发工程师，还是需要将研究成果落地的算法研究人员，ncnn 都是理想的选择。与普通用户不同，它主要服务于技术构建者，帮助他们把人工智能带到用户的指尖。目前，QQ、微信、天天 P 图等国民级应用背后都有 ncnn 的身影。\n\nncnn 的技术亮点十分突出：首先，它完全零依赖，无需安装任何第三方库即可编译运行，极大降低了集成门槛；其次，它具备卓越的跨平台能力，支持 Android、iOS、Linux、Windows 乃至 WebAssembly 等多种环境；最重要的是，其在手机 CPU 上的推理速度优于当前所有已知的开源框架，并针对主流手机芯片指令集进行了深度汇编级优化。如果你正在寻找一个轻量、快速且稳定的移动端 AI 推理方案，ncnn 值得尝试。","![ncnn](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_43541a5c1cfd.png)\n\n# ncnn\n\n[![License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-BSD_3_Clause-blue.svg?style=for-the-badge)](LICENSE.txt)\n[![Download Total Count](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fdownloads\u002FTencent\u002Fncnn\u002Ftotal.svg?style=for-the-badge)](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases)\n[![codecov](https:\u002F\u002Fimg.shields.io\u002Fcodecov\u002Fc\u002Fgithub\u002FTencent\u002Fncnn\u002Fmaster?style=for-the-badge)](https:\u002F\u002Fcodecov.io\u002Fgh\u002FTencent\u002Fncnn)\n\nncnn is a high-performance neural network inference computing framework optimized for mobile platforms.\nncnn is deeply considerate about deployment and uses on mobile phones from the beginning of design.\nncnn does not have third-party dependencies.\nIt is cross-platform and runs faster than all known open-source frameworks on mobile phone cpu.\nDevelopers can easily deploy deep learning algorithm models to the mobile platform by using efficient ncnn implementation, creating intelligent APPs, and bringing artificial intelligence to your fingertips.\nncnn is currently being used in many Tencent applications, such as QQ, Qzone, WeChat, Pitu, and so on.\n\nncnn 是一个为手机端极致优化的高性能神经网络前向计算框架。\nncnn 从设计之初深刻考虑手机端的部署和使用。\n无第三方依赖，跨平台，手机端 cpu 的速度快于目前所有已知的开源框架。\n基于 ncnn，开发者能够将深度学习算法轻松移植到手机端高效执行，\n开发出人工智能 APP，将 AI 带到你的指尖。\nncnn 目前已在腾讯多款应用中使用，如：QQ，Qzone，微信，天天 P 图等。\n\n---\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd>\n\u003Cb>技术交流 QQ 群\u003C\u002Fb>\u003Cbr \u002F>\n637093648 (超多大佬)\u003Cbr \u002F>\n答案：卷卷卷卷卷（已满）\n\u003C\u002Ftd>\n\u003Ctd rowspan=3>\n\u003Cb>Telegram Group\u003C\u002Fb>\n\n\u003Chttps:\u002F\u002Ft.me\u002Fncnnyes>\n\u003C\u002Ftd>\n\u003Ctd rowspan=3>\n\u003Cb>Discord Channel\u003C\u002Fb>\n\n\u003Chttps:\u002F\u002Fdiscord.gg\u002FYRsxgmF>\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\n\u003Cb>Pocky QQ 群（MLIR YES!）\u003C\u002Fb>\u003Cbr \u002F>\n677104663 (超多大佬)\u003Cbr \u002F>\n答案：multi-level intermediate representation\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\n\u003Cb>他们都不知道 pnnx 有多好用群\u003C\u002Fb>\u003Cbr \u002F>\n818998520 (新群！)\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n---\n\n## Download & Build status\n\nhttps:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\n\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd rowspan=2>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_b8065e48b4c3.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n **[how to build ncnn library](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build) on Linux \u002F Windows \u002F macOS \u002F Raspberry Pi3, Pi4 \u002F POWER \u002F Android \u002F NVIDIA Jetson \u002F iOS \u002F WebAssembly \u002F AllWinner D1 \u002F Loongson 2K1000**\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Source\u003C\u002Ftd>\n\u003Ctd colspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-full-source.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_62e1bde4ef8f.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for Android](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-android)\n- [Build for Termux on Android](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-termux-on-android)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Android\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fandroid.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aandroid)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Android shared\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android-vulkan-shared.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Fupload.wikimedia.org\u002Fwikipedia\u002Fcommons\u002Fthumb\u002F3\u002F37\u002FHMOS_Logo_Icon.svg\u002F240px-HMOS_Logo_Icon.svg.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for HarmonyOS with cross-compiling](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-harmonyos-with-cross-compiling)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>HarmonyOS\u003C\u002Ftd>\n\u003Ctd>\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fharmonyos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aharmonyos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>HarmonyOS shared\u003C\u002Ftd>\n\u003Ctd>\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_97a545d4bf67.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for iOS on macOS with xcode](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-ios-on-macos-with-xcode)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>iOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ios-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ios.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fios.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aios)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>iOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ios-simulator-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ios-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=10>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_5addd8d2ad40.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for macOS](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-macos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>macOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-macos-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-macos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=1>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fmacos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Amacos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Mac-Catalyst\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-mac-catalyst-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-mac-catalyst.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=1>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fmac-catalyst.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Amac-catalyst)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>watchOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-watchos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fwatchos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Awatchos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>watchOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-watchos-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>tvOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Ftvos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Atvos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>tvOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos-simulator-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>visionOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fvisionos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Avisionos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>visionOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos-simulator-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Apple xcframework\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-apple-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-apple.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=1>\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_ab9e2a35d6a5.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for Linux \u002F NVIDIA Jetson \u002F Raspberry Pi3, Pi4 \u002F POWER](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-linux)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Ubuntu 22.04\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2204.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2204-shared.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-x64-gpu-gcc.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-x64-gpu-gcc)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Ubuntu 24.04\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2404.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2404-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=5>\n \u003Cimg alt=\"windows\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_6a84e2a2daf3.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for Windows x64 using VS2017](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-windows-x64-using-visual-studio-community-2017)\n- [Build for Windows x64 using MinGW-w64](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-windows-x64-using-mingw-w64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2015\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2015.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2015-shared.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=4>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fwindows.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Awindows)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2017\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2017.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2017-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2019\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2019.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2019-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2022\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2022.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2022-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=2>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_41d6fdcceaeb.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for WebAssembly](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-webassembly)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>WebAssembly\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-webassembly.zip)\n\n\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fweb-assembly.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aweb-assembly)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=8>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_342b3a50c54d.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [Build for ARM Cortex-A family with cross-compiling](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-arm-cortex-a-family-with-cross-compiling)\n- [Build for Hisilicon platform with cross-compiling](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-hisilicon-platform-with-cross-compiling)\n- [Build for AllWinner D1](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-allwinner-d1)\n- [Build for Loongson 2K1000](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-loongson-2k1000)\n- [Build for QNX](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-qnx)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (arm)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-arm.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-arm)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (aarch64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-aarch64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-aarch64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (mips)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-mips.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-mips)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (mips64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-mips64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-mips64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (ppc64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-ppc64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-ppc64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (riscv64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-riscv64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-riscv64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (loongarch64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-loongarch64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-loongarch64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003C\u002Ftable>\n\n\n---\n\n## Support most commonly used CNN network\n\n## 支持大部分常用的 CNN 网络\n\n- Classical CNN:\n [VGG](https:\u002F\u002Fgithub.com\u002FBVLC\u002Fcaffe\u002Fwiki\u002FModel-Zoo#models-used-by-the-vgg-team-in-ilsvrc-2014)\n [AlexNet](https:\u002F\u002Fgithub.com\u002FBVLC\u002Fcaffe\u002Ftree\u002F9b891540183ddc834a02b2bd81b31afae71b2153\u002Fmodels\u002Fbvlc_alexnet)\n [GoogleNet](https:\u002F\u002Fgithub.com\u002FBVLC\u002Fcaffe\u002Ftree\u002F9b891540183ddc834a02b2bd81b31afae71b2153\u002Fmodels\u002Fbvlc_googlenet)\n Inception\n ...\n- Practical CNN:\n [ResNet](https:\u002F\u002Fgithub.com\u002Ftornadomeet\u002FResNet)\n [DenseNet](https:\u002F\u002Fgithub.com\u002Fliuzhuang13\u002FDenseNet)\n [SENet](https:\u002F\u002Fgithub.com\u002Fhujie-frank\u002FSENet)\n [FPN](https:\u002F\u002Fgithub.com\u002Funsky\u002FFPN)\n ...\n- Light-weight CNN:\n [SqueezeNet](https:\u002F\u002Fgithub.com\u002Fforresti\u002FSqueezeNet)\n [MobileNetV1](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fmodels\u002Fblob\u002Fmaster\u002Fresearch\u002Fslim\u002Fnets\u002Fmobilenet_v1.md)\n [MobileNetV2\u002FV3](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fmodels\u002Fblob\u002Fmaster\u002Fresearch\u002Fslim\u002Fnets\u002Fmobilenet\u002FREADME.md)\n [ShuffleNetV1](https:\u002F\u002Fgithub.com\u002Ffarmingyard\u002FShuffleNet)\n [ShuffleNetV2](https:\u002F\u002Fgithub.com\u002Fopconty\u002Fkeras-shufflenetV2)\n [MNasNet](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fmodels\u002Ftree\u002Fmaster\u002Fresearch\u002Fslim\u002Fnets\u002Fnasnet)\n ...\n- Face Detection:\n [MTCNN](https:\u002F\u002Fgithub.com\u002Fipazc\u002Fmtcnn)\n [RetinaFace](https:\u002F\u002Fgithub.com\u002Fbiubug6\u002FPytorch_Retinaface)\n [scrfd](https:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-scrfd)\n ...\n- Detection:\n [VGG-SSD](https:\u002F\u002Fgithub.com\u002Flzx1413\u002FCAFFE_SSD)\n [MobileNet-SSD](https:\u002F\u002Fgithub.com\u002Fchuanqi305\u002FMobileNet-SSD)\n [SqueezeNet-SSD](https:\u002F\u002Fgithub.com\u002Fchuanqi305\u002FSqueezeNet-SSD)\n [MobileNetV2-SSDLite](https:\u002F\u002Fgithub.com\u002Fchuanqi305\u002FMobileNetv2-SSDLite)\n [MobileNetV3-SSDLite](https:\u002F\u002Fgithub.com\u002FXiaoyuHuang96\u002FMobilenetV3SSDLite-tfkeras)\n ...\n- Detection:\n [Faster-RCNN](https:\u002F\u002Fgithub.com\u002Frbgirshick\u002Fpy-faster-rcnn)\n [R-FCN](https:\u002F\u002Fgithub.com\u002Fdaijifeng001\u002FR-FCN)\n ...\n- Detection:\n [YOLOv2](https:\u002F\u002Fgithub.com\u002Flongcw\u002Fyolo2-pytorch)\n [YOLOv3](https:\u002F\u002Fgithub.com\u002Fultralytics\u002Fyolov3)\n [MobileNet-YOLOv3](https:\u002F\u002Fgithub.com\u002Feric612\u002FMobileNet-YOLO)\n [YOLOv4](https:\u002F\u002Fgithub.com\u002FTianxiaomo\u002Fpytorch-YOLOv4)\n [YOLOv5](https:\u002F\u002Fgithub.com\u002Fultralytics\u002Fyolov5)\n [YOLOv7](https:\u002F\u002Fgithub.com\u002FWongKinYiu\u002Fyolov7)\n [YOLOX](https:\u002F\u002Fgithub.com\u002FMegvii-BaseDetection\u002FYOLOX)\n [YOLOv8](https:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-yolov8)\n ...\n- Detection:\n [NanoDet](https:\u002F\u002Fgithub.com\u002FRangiLyu\u002Fnanodet)\n- Segmentation:\n [FCN](https:\u002F\u002Fgithub.com\u002Funsky\u002FFPN)\n [PSPNet](https:\u002F\u002Fgithub.com\u002Fhszhao\u002FPSPNet)\n [UNet](https:\u002F\u002Fgithub.com\u002Fzhixuhao\u002Funet)\n [YOLACT](https:\u002F\u002Fgithub.com\u002Fdbolya\u002Fyolact)\n ...\n- Pose Estimation:\n [SimplePose](https:\u002F\u002Fgithub.com\u002Fdog-qiuqiu\u002FUltralight-SimplePose)\n ...\n\n---\n\n## HowTo\n\n**[use ncnn with alexnet](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-alexnet) with detailed steps, recommended for beginners :)**\n\n**[ncnn 组件使用指北 alexnet](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-alexnet.zh) 附带详细步骤，新人强烈推荐 :)**\n\n**[use netron for ncnn model visualization](https:\u002F\u002Fnetron.app)**\n\n**[use ncnn with pytorch or onnx](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-pytorch-or-onnx)**\n\n[ncnn low-level operation api](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Flow-level-operation-api)\n\n[ncnn param and model file spec](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fparam-and-model-file-structure)\n\n[ncnn operation param weight table](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Foperation-param-weight-table)\n\n[how to implement custom layer step by step](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-implement-custom-layer-step-by-step)\n\n---\n\n## FAQ\n\n**[ncnn deepwiki](https:\u002F\u002Fdeepwiki.com\u002FTencent\u002Fncnn) LLM Answering Questions ;)** \n\n**[ncnn throw error](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002FFAQ-ncnn-throw-error)**\n\n**[ncnn produce wrong result](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002FFAQ-ncnn-produce-wrong-result)**\n\n**[ncnn vulkan](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002FFAQ-ncnn-vulkan)**\n\n---\n\n## Features\n\n- Supports convolutional neural networks, supports multiple input and multi-branch structure, can calculate part of the branch\n- No third-party library dependencies, does not rely on BLAS \u002F NNPACK or any other computing framework\n- Pure C++ implementation, cross-platform, supports Android, iOS and so on\n- ARM NEON assembly level of careful optimization, calculation speed is extremely high\n- Sophisticated memory management and data structure design, very low memory footprint\n- Supports multi-core parallel computing acceleration, ARM big.LITTLE CPU scheduling optimization\n- Supports GPU acceleration via the next-generation low-overhead Vulkan API\n- Extensible model design, supports 8bit [quantization](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fquantized-int8-inference) and half-precision floating point storage, can import caffe\u002Fpytorch\u002Fmxnet\u002Fonnx\u002Fdarknet\u002Fkeras\u002Ftensorflow(mlir) models\n- Support direct memory zero copy reference load network model\n- Can be registered with custom layer implementation and extended\n- Well, it is strong, not afraid of being stuffed with 卷 QvQ\n\n## 功能概述\n\n- 支持卷积神经网络，支持多输入和多分支结构，可计算部分分支\n- 无任何第三方库依赖，不依赖 BLAS\u002FNNPACK 等计算框架\n- 纯 C++ 实现，跨平台，支持 Android \u002F iOS 等\n- ARM Neon 汇编级良心优化，计算速度极快\n- 精细的内存管理和数据结构设计，内存占用极低\n- 支持多核并行计算加速，ARM big.LITTLE CPU 调度优化\n- 支持基于全新低消耗的 Vulkan API GPU 加速\n- 可扩展的模型设计，支持 8bit [量化](tools\u002Fquantize) 和半精度浮点存储，可导入 caffe\u002Fpytorch\u002Fmxnet\u002Fonnx\u002Fdarknet\u002Fkeras\u002Ftensorflow(mlir) 模型\n- 支持直接内存零拷贝引用加载网络模型\n- 可注册自定义层实现并扩展\n- 恩，很强就是了，不怕被塞卷 QvQ\n\n---\n\n## supported platform matrix\n\n- ✅ = known work and runs fast with good optimization\n- ✔️ = known work, but speed may not be fast enough\n- ❔ = shall work, not confirmed\n- \u002F = not applied\n\n| | Windows | Linux | Android | macOS | iOS |\n| ---------- | ------- | ----- | ------- | ----- | --- |\n| intel-cpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| intel-gpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| amd-cpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| amd-gpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| nvidia-gpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| qcom-cpu | ✅ | ✅ | ✅ | \u002F | \u002F |\n| qcom-gpu | ✔️ | ✔️ | ✔️ | \u002F | \u002F |\n| arm-cpu | ✅ | ✅ | ✅ | \u002F | \u002F |\n| arm-gpu | ❔ | ✔️ | ✔️ | \u002F | \u002F |\n| apple-cpu | \u002F | \u002F | \u002F | ✔️ | ✅ |\n| apple-gpu | \u002F | \u002F | \u002F | ✔️ | ✔️ |\n| ibm-cpu | \u002F | ✔️ | \u002F | \u002F | \u002F |\n\n---\n\n## Project examples\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-squeezenet>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-styletransfer>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-mobilenetssd>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fmoli232777144\u002Fmtcnn_ncnn>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-yolov5>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fxiang-wuu\u002Fncnn-android-yolov7>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-scrfd> 🤩\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fshaoshengsong\u002Fqt_android_ncnn_lib_encrypt_example>\n\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_778fc6f8b7aa.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_40a634e57935.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_5011f635891e.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_a8389a070e98.png\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_64bff644b350.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_98839e5be397.jpg\" height =\"230\"\u002F>\u003Cbr>\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fmagicse\u002Fncnn-colorization-siggraph17>\u003Cbr>\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_ea067c8670c7.jpg\" width =\"700\"\u002F>\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fmizu-bai\u002Fncnn-fortran> Call ncnn from Fortran\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fk2-fsa\u002Fsherpa> Use ncnn for real-time speech\n recognition (i.e., speech-to-text); also support embedded devices and provide\n mobile Apps (e.g., Android App)\n\n---\n\n## License\n\n[BSD 3 Clause](LICENSE.txt)\n","![ncnn](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_43541a5c1cfd.png)\n\n# ncnn\n\n[![许可证](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-BSD_3_Clause-blue.svg?style=for-the-badge)](LICENSE.txt)\n[![下载总次数](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fdownloads\u002FTencent\u002Fncnn\u002Ftotal.svg?style=for-the-badge)](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases)\n[![codecov](https:\u002F\u002Fimg.shields.io\u002Fcodecov\u002Fc\u002Fgithub\u002FTencent\u002Fncnn\u002Fmaster?style=for-the-badge)](https:\u002F\u002Fcodecov.io\u002Fgh\u002FTencent\u002Fncnn)\n\nncnn 是一个为手机端极致优化的高性能神经网络前向计算框架。\nncnn 从设计之初深刻考虑手机端的部署和使用。\n无第三方依赖，跨平台，手机端 cpu 的速度快于目前所有已知的开源框架。\n基于 ncnn，开发者能够将深度学习算法轻松移植到手机端高效执行，\n开发出人工智能 APP，将 AI 带到你的指尖。\nncnn 目前已在腾讯多款应用中使用，如：QQ，Qzone，微信，天天 P 图等。\n\n---\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd>\n\u003Cb>技术交流 QQ 群\u003C\u002Fb>\u003Cbr \u002F>\n637093648 (超多大佬)\u003Cbr \u002F>\n答案：卷卷卷卷卷（已满）\n\u003C\u002Ftd>\n\u003Ctd rowspan=3>\n\u003Cb>Telegram Group\u003C\u002Fb>\n\n\u003Chttps:\u002F\u002Ft.me\u002Fncnnyes>\n\u003C\u002Ftd>\n\u003Ctd rowspan=3>\n\u003Cb>Discord Channel\u003C\u002Fb>\n\n\u003Chttps:\u002F\u002Fdiscord.gg\u002FYRsxgmF>\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\n\u003Cb>Pocky QQ 群（MLIR YES!）\u003C\u002Fb>\u003Cbr \u002F>\n677104663 (超多大佬)\u003Cbr \u002F>\n答案：multi-level intermediate representation\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\n\u003Cb>他们都不知道 pnnx 有多好用群\u003C\u002Fb>\u003Cbr \u002F>\n818998520 (新群！)\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n---\n\n## 下载与构建状态\n\nhttps:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\n\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd rowspan=2>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_b8065e48b4c3.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n **[如何构建 ncnn 库](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build) 在 Linux \u002F Windows \u002F macOS \u002F Raspberry Pi3, Pi4 \u002F POWER \u002F Android \u002F NVIDIA Jetson \u002F iOS \u002F WebAssembly \u002F AllWinner D1 \u002F Loongson 2K1000 上**\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>源代码\u003C\u002Ftd>\n\u003Ctd colspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-full-source.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_62e1bde4ef8f.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [为 Android 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-android)\n- [为 Android 上的 Termux 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-termux-on-android)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Android\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fandroid.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aandroid)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Android 共享库\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android-vulkan-shared.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-android-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Fupload.wikimedia.org\u002Fwikipedia\u002Fcommons\u002Fthumb\u002F3\u002F37\u002FHMOS_Logo_Icon.svg\u002F240px-HMOS_Logo_Icon.svg.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [通过交叉编译为 HarmonyOS 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-harmonyos-with-cross-compiling)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>HarmonyOS\u003C\u002Ftd>\n\u003Ctd>\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fharmonyos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aharmo…\n\n[\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-mac-catalyst-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-mac-catalyst.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=1>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fmac-catalyst.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Amac-catalyst)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>watchOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-watchos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fwatchos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Awatchos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>watchOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-watchos-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>tvOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Ftvos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Atvos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>tvOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos-simulator-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-tvos-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>visionOS\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fvisionos.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Avisionos)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>visionOS-Simulator\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos-simulator-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-visionos-simulator.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Apple xcframework\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-apple-vulkan.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+cpuonly-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-apple.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=1>\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=3>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_ab9e2a35d6a5.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [为 Linux \u002F NVIDIA Jetson \u002F Raspberry Pi3、Pi4 \u002F POWER 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-linux)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Ubuntu 22.04\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2204.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2204-shared.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=2>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-x64-gpu-gcc.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-x64-gpu-gcc)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Ubuntu 24.04\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2404.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2404-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr.\n\n\u003Ctr>\n\u003Ctd rowspan=5>\n \u003Cimg alt=\"windows\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_6a84e2a2daf3.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [使用 VS2017 为 Windows x64 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-windows-x64-using-visual-studio-community-2017)\n- [使用 MinGW-w64 为 Windows x64 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-windows-x64-using-mingw-w64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2015\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2015.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2015-shared.zip)\n\n\u003C\u002Ftd>\n\u003Ctd rowspan=4>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fwindows.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Awindows)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2017\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2017.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2017-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2019\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2019.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2019-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>VS2022\u003C\u002Ftd>\n\u003Ctd>\n\n[\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2022.zip)\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F+shared-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-windows-vs2022-shared.zip)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=2>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_41d6fdcceaeb.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [为 WebAssembly 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-webassembly)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>WebAssembly\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdownload-blue?style=for-the-badge\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-webassembly.zip)\n\n\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Fweb-assembly.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Aweb-assembly)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n\u003Ctd rowspan=8>\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_342b3a50c54d.png\" width=\"120\" height=\"auto\">\n\u003C\u002Ftd>\n\u003Ctd colspan=3>\n\n- [使用交叉编译为 ARM Cortex-A 系列构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-arm-cortex-a-family-with-cross-compiling)\n- [使用交叉编译为海思平台构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-hisilicon-platform-with-cross-compiling)\n- [为全志 D1 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-allwinner-d1)\n- [为龙芯 2K1000 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-loongson-2k1000)\n- [为 QNX 构建](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-build#build-for-qnx)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (arm)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-arm.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-arm)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (aarch64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-aarch64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-aarch64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (mips)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-mips.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-mips)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (mips64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-mips64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-mips64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (ppc64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-ppc64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-ppc64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (riscv64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-riscv64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-riscv64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Linux (loongarch64)\u003C\u002Ftd>\n\u003Ctd>\u003C\u002Ftd>\n\u003Ctd>\n\n [\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTencent\u002Fncnn\u002Flinux-loongarch64.yml?branch=master&style=for-the-badge&label=build\">](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Factions?query=workflow%3Alinux-loongarch64)\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003C\u002Ftable>\n\n\n---\n\n\n\n## 支持大部分常用的 CNN 网络\n\n## 支持大部分常用的 CNN 网络\n\n- 经典 CNN:\n [VGG](https:\u002F\u002Fgithub.com\u002FBVLC\u002Fcaffe\u002Fwiki\u002FModel-Zoo#models-used-by-the-vgg-team-in-ilsvrc-2014)\n [AlexNet](https:\u002F\u002Fgithub.com\u002FBVLC\u002Fcaffe\u002Ftree\u002F9b891540183ddc834a02b2bd81b31afae71b2153\u002Fmodels\u002Fbvlc_alexnet)\n [GoogleNet](https:\u002F\u002Fgithub.com\u002FBVLC\u002Fcaffe\u002Ftree\u002F9b891540183ddc834a02b2bd81b31afae71b2153\u002Fmodels\u002Fbvlc_googlenet)\n Inception\n ...\n- 实用 CNN:\n [ResNet](https:\u002F\u002Fgithub.com\u002Ftornadomeet\u002FResNet)\n [DenseNet](https:\u002F\u002Fgithub.com\u002Fliuzhuang13\u002FDenseNet)\n [SENet](https:\u002F\u002Fgithub.com\u002Fhujie-frank\u002FSENet)\n [FPN](https:\u002F\u002Fgithub.com\u002Funsky\u002FFPN)\n ...\n- 轻量级 CNN:\n [SqueezeNet](https:\u002F\u002Fgithub.com\u002Fforresti\u002FSqueezeNet)\n [MobileNetV1](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fmodels\u002Fblob\u002Fmaster\u002Fresearch\u002Fslim\u002Fnets\u002Fmobilenet_v1.md)\n [MobileNetV2\u002FV3](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fmodels\u002Fblob\u002Fmaster\u002Fresearch\u002Fslim\u002Fnets\u002Fmobilenet\u002FREADME.md)\n [ShuffleNetV1](https:\u002F\u002Fgithub.com\u002Ffarmingyard\u002FShuffleNet)\n [ShuffleNetV2](https:\u002F\u002Fgithub.com\u002Fopconty\u002Fkeras-shufflenetV2)\n [MNasNet](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fmodels\u002Ftree\u002Fmaster\u002Fresearch\u002Fslim\u002Fnets\u002Fnasnet)\n ...\n- 人脸检测:\n [MTCNN](https:\u002F\u002Fgithub.com\u002Fipazc\u002Fmtcnn)\n [RetinaFace](https:\u002F\u002Fgithub.com\u002Fbiubug6\u002FPytorch_Retinaface)\n [scrfd](https:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-scrfd)\n ...\n- 检测:\n [VGG-SSD](https:\u002F\u002Fgithub.com\u002Flzx1413\u002FCAFFE_SSD)\n [MobileNet-SSD](https:\u002F\u002Fgithub.com\u002Fchuanqi305\u002FMobileNet-SSD)\n [SqueezeNet-SSD](https:\u002F\u002Fgithub.com\u002Fchuanqi305\u002FSqueezeNet-SSD)\n [MobileNetV2-SSDLite](https:\u002F\u002Fgithub.com\u002Fchuanqi305\u002FMobileNetv2-SSDLite)\n [MobileNetV3-SSDLite](https:\u002F\u002Fgithub.com\u002FXiaoyuHuang96\u002FMobilenetV3SSDLite-tfkeras)\n ...\n- 检测:\n [Faster-RCNN](https:\u002F\u002Fgithub.com\u002Frbgirshick\u002Fpy-faster-rcnn)\n [R-FCN](https:\u002F\u002Fgithub.com\u002Fdaijifeng001\u002FR-FCN)\n ...\n- 检测:\n [YOLOv2](https:\u002F\u002Fgithub.com\u002Flongcw\u002Fyolo2-pytorch)\n [YOLOv3](https:\u002F\u002Fgithub.com\u002Fultralytics\u002Fyolov3)\n [MobileNet-YOLOv3](https:\u002F\u002Fgithub.com\u002Feric612\u002FMobileNet-YOLO)\n [YOLOv4](https:\u002F\u002Fgithub.com\u002FTianxiaomo\u002Fpytorch-YOLOv4)\n [YOLOv5](https:\u002F\u002Fgithub.com\u002Fultralytics\u002Fyolov5)\n [YOLOv7](https:\u002F\u002Fgithub.com\u002FWongKinYiu\u002Fyolov7)\n [YOLOX](https:\u002F\u002Fgithub.com\u002FMegvii-BaseDetection\u002FYOLOX)\n [YOLOv8](https:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-yolov8)\n ...\n- 检测:\n [NanoDet](https:\u002F\u002Fgithub.com\u002FRangiLyu\u002Fnanodet)\n- 分割:\n [FCN](https:\u002F\u002Fgithub.com\u002Funsky\u002FFPN)\n [PSPNet](https:\u002F\u002Fgithub.com\u002Fhszhao\u002FPSPNet)\n [UNet](https:\u002F\u002Fgithub.com\u002Fzhixuhao\u002Funet)\n [YOLACT](https:\u002F\u002Fgithub.com\u002Fdbolya\u002Fyolact)\n ...\n- 姿势估计:\n [SimplePose](https:\u002F\u002Fgithub.com\u002Fdog-qiuqiu\u002FUltralight-SimplePose)\n ...\n\n---\n\n## 操作指南\n\n**[使用 ncnn 与 AlexNet](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-alexnet) 附详细步骤，强烈推荐给初学者 :)**\n\n**[ncnn 组件使用指北 alexnet](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-alexnet.zh) 附带详细步骤，新人强烈推荐 :)**\n\n**[使用 Netron 可视化 ncnn 模型](https:\u002F\u002Fnetron.app)**\n\n**[使用 ncnn 与 PyTorch 或 ONNX](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-pytorch-or-onnx)**\n\n[ncnn 低级操作 API](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Flow-level-operation-api)\n\n[ncnn 参数和模型文件规范](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fparam-and-model-file-structure)\n\n[ncnn 操作参数权重表](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Foperation-param-weight-table)\n\n[如何逐步实现自定义层](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fhow-to-implement-custom-layer-step-by-step)\n\n---\n\n## 常见问题解答\n\n**[ncnn 深度维基](https:\u002F\u002Fdeepwiki.com\u002FTencent\u002Fncnn) 大语言模型答疑；)** \n\n**[ncnn 抛出错误](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002FFAQ-ncnn-throw-error)**\n\n**[ncnn 产生错误结果](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002FFAQ-ncnn-produce-wrong-result)**\n\n**[ncnn Vulkan](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002FFAQ-ncnn-vulkan)**\n\n---\n\n## 功能特性\n\n- 支持卷积神经网络，支持多输入和多分支结构，可计算部分分支\n- 无任何第三方库依赖，不依赖 BLAS\u002FNNPACK 等计算框架\n- 纯 C++ 实现，跨平台，支持 Android \u002F iOS 等\n- ARM Neon 汇编级良心优化，计算速度极快\n- 精细的内存管理和数据结构设计，内存占用极低\n- 支持多核并行计算加速，ARM big.LITTLE CPU 调度优化\n- 支持基于全新低消耗的 Vulkan API GPU 加速\n- 可扩展的模型设计，支持 8bit [量化](tools\u002Fquantize) 和半精度浮点存储，可导入 caffe\u002Fpytorch\u002Fmxnet\u002Fonnx\u002Fdarknet\u002Fkeras\u002Ftensorflow(mlir) 模型\n- 支持直接内存零拷贝引用加载网络模型\n- 可注册自定义层实现并扩展\n- 恩，很强就是了，不怕被塞卷 QvQ\n\n## 功能概述\n\n- 支持卷积神经网络，支持多输入和多分支结构，可计算部分分支\n- 无任何第三方库依赖，不依赖 BLAS\u002FNNPACK 等计算框架\n- 纯 C++ 实现，跨平台，支持 Android \u002F iOS 等\n- ARM Neon 汇编级良心优化，计算速度极快\n- 精细的内存管理和数据结构设计，内存占用极低\n- 支持多核并行计算加速，ARM big.LITTLE CPU 调度优化\n- 支持基于全新低消耗的 Vulkan API GPU 加速\n- 可扩展的模型设计，支持 8bit [量化](tools\u002Fquantize) 和半精度浮点存储，可导入 caffe\u002Fpytorch\u002Fmxnet\u002Fonnx\u002Fdarknet\u002Fkeras\u002Ftensorflow(mlir) 模型\n- 支持直接内存零拷贝引用加载网络模型\n- 可注册自定义层实现并扩展\n- 恩，很强就是了，不怕被塞卷 QvQ\n\n---\n\n## 支持平台矩阵\n\n- ✅ = 已知可行且运行速度快，优化良好\n- ✔️ = 已知可行，但速度可能不够快\n- ❔ = 应该可行，但尚未确认\n- \u002F = 不适用\n\n| | Windows | Linux | Android | macOS | iOS |\n| ---------- | ------- | ----- | ------- | ----- | --- |\n| intel-cpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| intel-gpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| amd-cpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| amd-gpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| nvidia-gpu | ✔️ | ✔️ | ✔️ | ✔️ | \u002F |\n| qcom-cpu | ✅ | ✅ | ✅ | \u002F | \u002F |\n| qcom-gpu | ✔️ | ✔️ | ✔️ | \u002F | \u002F |\n| arm-cpu | ✅ | ✅ | ✅ | \u002F | \u002F |\n| arm-gpu | ❔ | ✔️ | ✔️ | \u002F | \u002F |\n| apple-cpu | \u002F | \u002F | \u002F | ✔️ | ✅ |\n| apple-gpu | \u002F | \u002F | \u002F | ✔️ | ✔️ |\n| ibm-cpu | \u002F | ✔️ | \u002F | \u002F | \u002F |\n\n---\n\n## 项目示例\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-squeezenet>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-styletransfer>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-mobilenetssd>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fmoli232777144\u002Fmtcnn_ncnn>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-yolov5>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fxiang-wuu\u002Fncnn-android-yolov7>\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fnihui\u002Fncnn-android-scrfd> 🤩\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fshaoshengsong\u002Fqt_android_ncnn_lib_encrypt_example>\n\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_778fc6f8b7aa.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_40a634e57935.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_5011f635891e.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_a8389a070e98.png\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_64bff644b350.jpg\" height =\"230\"\u002F>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_98839e5be397.jpg\" height =\"230\"\u002F>\u003Cbr>\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fmagicse\u002Fncnn-colorization-siggraph17>\u003Cbr>\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_readme_ea067c8670c7.jpg\" width =\"700\"\u002F>\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fmizu-bai\u002Fncnn-fortran> 从 Fortran 调用 ncnn\n\n- \u003Chttps:\u002F\u002Fgithub.com\u002Fk2-fsa\u002Fsherpa> 使用 ncnn 进行实时语音识别（即语音转文字）；同时支持嵌入式设备，并提供移动应用（例如 Android 应用）\n\n---\n\n## 许可证\n\n[BSD 3 条款](LICENSE.txt)","# ncnn 快速上手指南\n\nncnn 是腾讯开源的高性能神经网络前向计算框架，专为移动端极致优化。它无第三方依赖、跨平台，在手机 CPU 上的推理速度优于已知的所有开源框架，已广泛应用于微信、QQ 等应用中。\n\n## 1. 环境准备\n\n### 系统要求\nncnn 支持主流操作系统及架构：\n- **移动端**: Android, iOS, HarmonyOS\n- **桌面端**: Windows, macOS, Linux (Ubuntu\u002FDebian\u002FCentOS)\n- **嵌入式**: Raspberry Pi, NVIDIA Jetson, AllWinner D1, Loongson\n- **其他**: WebAssembly\n\n### 前置依赖\nncnn 核心库**无第三方依赖**。若需自行编译或转换模型，建议安装以下工具：\n- **CMake**: 3.14 或更高版本\n- **编译器**: GCC \u002F Clang \u002F MSVC\n- **模型转换工具 (可选)**: \n - `onnx2ncnn`: 用于将 ONNX 模型转换为 ncnn 格式\n - `pnnx`: 推荐使用的 PyTorch 直接导出工具（更友好）\n\n> **国内开发者提示**：安装依赖时如遇网络问题，可配置国内镜像源（如清华源、阿里源）加速 `apt`、`pip` 或 `git clone` 操作。\n\n## 2. 安装步骤\n\n推荐直接使用预编译包，也可选择源码编译。\n\n### 方案 A：下载预编译包（推荐）\n访问 [GitHub Releases](https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest) 下载对应平台的二进制包。\n\n**Linux 示例 (Ubuntu):**\n```bash\n# 下载并解压 (以 Ubuntu 22.04 为例)\nwget https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Freleases\u002Flatest\u002Fdownload\u002Fncnn-20260113-ubuntu-2204.zip\nunzip ncnn-20260113-ubuntu-2204.zip\n\n# 设置环境变量 (可选，方便调用)\nexport NCNN_DIR=$(pwd)\u002Fncnn-20260113-ubuntu-2204\nexport LD_LIBRARY_PATH=$NCNN_DIR\u002Flib:$LD_LIBRARY_PATH\n```\n\n**Android 开发:**\n在 `build.gradle` 中引入下载的 `.aar` 文件或通过 CMake 链接下载的 `.so` 库。\n\n### 方案 B：源码编译 (Linux 示例)\n如需自定义构建或贡献代码，可源码编译：\n\n```bash\n# 1. 克隆仓库\ngit clone https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn.git\ncd ncnn\n\n# 2. 创建构建目录\nmkdir build && cd build\n\n# 3. 配置 CMake (开启 Vulkan 加速可选 -DNCNN_VULKAN=ON)\ncmake -DCMAKE_BUILD_TYPE=Release ..\n\n# 4. 编译并安装\nmake -j$(nproc)\nsudo make install\n```\n\n## 3. 基本使用\n\nncnn 的使用流程通常为：**模型转换 -> 加载模型 -> 输入数据 -> 推理 -> 输出结果**。\n\n### 第一步：准备模型\n假设你有一个 PyTorch 模型，推荐使用 `pnnx` 工具直接导出为 ncnn 格式（生成 `.param` 和 `.bin` 文件）：\n```bash\n# 需先安装 pnnx (pip install pnnx)\npnnx model.pt inputshape=[1,3,224,224]\n# 输出：model.pnnx.param, model.pnnx.bin\n```\n\n### 第二步：编写推理代码 (C++ 示例)\n以下是一个最简单的分类网络推理示例：\n\n```cpp\n#include \"net.h\"\n#include \"mat.h\"\n#include \u003Cstdio.h>\n\nint main()\n{\n \u002F\u002F 1. 初始化网络\n ncnn::Net net;\n \n \u002F\u002F 加载模型文件 (param 和 bin)\n \u002F\u002F 注意：路径需根据实际情况修改\n net.load_param(\"model.pnnx.param\");\n net.load_model(\"model.pnnx.bin\");\n\n \u002F\u002F 2. 准备输入数据\n \u002F\u002F 创建一个 3 通道 224x224 的图片矩阵\n ncnn::Mat in = ncnn::Mat::from_pixels(image_data, ncnn::Mat::PIXEL_RGB, 224, 224);\n \n \u002F\u002F 归一化处理 (均值和标准差根据模型训练时的设置调整)\n const float mean_vals[3] = {123.68f, 116.78f, 103.94f};\n const float norm_vals[3] = {0.017125f, 0.017507f, 0.017420f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n \u002F\u002F 3. 创建提取器并执行推理\n ncnn::Extractor ex = net.create_extractor();\n ex.input(\"input\", in); \u002F\u002F \"input\" 替换为模型实际的输入节点名\n\n ncnn::Mat out;\n ex.extract(\"output\", out); \u002F\u002F \"output\" 替换为模型实际的输出节点名\n\n \u002F\u002F 4. 处理输出结果\n printf(\"Output channel count: %d\\n\", out.c);\n \u002F\u002F 此处可添加代码寻找最大概率类别等逻辑\n\n return 0;\n}\n```\n\n### 编译运行示例\n假设上述代码保存为 `main.cpp`，且 ncnn 已安装在 `\u002Fusr\u002Flocal`：\n\n```bash\ng++ main.cpp -o demo -lncnn -std=c++11\n.\u002Fdemo\n```\n\n> **提示**：具体输入\u002F输出节点名称可通过查看 `.param` 文件或使用时序调试工具获取。对于 Android\u002FiOS 开发，逻辑相同，只需将 C++ 代码嵌入 respective 的工程结构中。","某初创团队开发一款离线人脸识别门禁 APP，需在低端安卓手机上实现毫秒级响应且保护用户隐私。\n\n### 没有 ncnn 时\n- **推理延迟高**：直接移植通用框架导致在旧款手机 CPU 上单次识别耗时超过 2 秒，用户开门体验极差。\n- **安装包臃肿**：依赖庞大的第三方库（如 TensorFlow Lite 完整包），使 APK 体积增加 15MB 以上，严重影响下载转化率。\n- **发热耗电严重**：低效的计算逻辑让手机在处理视频流时迅速发烫，电池电量在半小时内骤降 30%。\n- **部署复杂**：不同品牌手机的指令集差异导致适配困难，经常出现在 A 手机正常、B 手机崩溃的兼容性问题。\n\n### 使用 ncnn 后\n- **极速响应**：ncnn 针对移动端 CPU 极致优化，利用汇编级加速将识别耗时压缩至 80 毫秒内，实现“无感”开门。\n- **轻量集成**：零第三方依赖的特性使模型集成仅增加不到 2MB 体积，显著降低了用户的安装门槛。\n- **冷静运行**：高效的内存管理和计算调度大幅降低功耗，连续工作一小时手机仅微温，电量消耗控制在 5% 以内。\n- **跨平台稳定**：一套代码完美覆盖从高通到联发科等各类芯片架构，彻底解决了碎片化设备的兼容性难题。\n\nncnn 通过极致的底层优化，让复杂的深度学习模型在资源受限的移动设备上也能跑出桌面级的流畅体验。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FTencent_ncnn_b8065e48.png","Tencent","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002FTencent_f7e55588.png","",null,"https:\u002F\u002Fopensource.tencent.com","https:\u002F\u002Fgithub.com\u002FTencent",[79,83,87,91,95,99,103,106],{"name":80,"color":81,"percentage":82},"C++","#f34b7d",52.6,{"name":84,"color":85,"percentage":86},"C","#555555",37.9,{"name":88,"color":89,"percentage":90},"Python","#3572A5",4.7,{"name":92,"color":93,"percentage":94},"GLSL","#5686a5",3.8,{"name":96,"color":97,"percentage":98},"CMake","#DA3434",1,{"name":100,"color":101,"percentage":102},"Shell","#89e051",0,{"name":104,"color":105,"percentage":102},"Batchfile","#C1F12E",{"name":107,"color":75,"percentage":102},"SWIG",23122,4417,"2026-04-20T08:29:38","NOASSERTION",4,"Linux, macOS, Windows, Android, iOS, HarmonyOS","非必需。支持 Vulkan API 进行 GPU 加速（兼容 NVIDIA、AMD、Intel 及移动端 GPU），也可纯 CPU 运行。未指定具体显存大小或 CUDA 版本要求（因主要基于 Vulkan 而非 CUDA）。","未说明（针对移动端优化，通常占用较低，具体取决于模型大小）",{"notes":117,"python":118,"dependencies":119},"该框架无第三方依赖。支持跨平台编译（包括 Linux, Windows, macOS, Android, iOS, HarmonyOS, WebAssembly 等）。提供预编译包分为'Vulkan 版'（含 GPU 加速）和'CPU only 版'（仅 CPU）。构建需使用 CMake。主要针对手机端 CPU 极致优化，推理速度快于其他已知开源框架。","未说明（核心为 C++ 框架，Python 绑定为可选组件）",[],[14,15],[122,123,124,125,126,127,128,129,64,130,131,132,133,134,135,136,137,138,139,140],"inference","high-preformance","simd","arm-neon","deep-learning","artificial-intelligence","android","ios","vulkan","neural-network","caffe","mxnet","pytorch","onnx","darknet","tensorflow","mlir","keras","riscv","2026-03-27T02:49:30.150509","2026-04-20T21:07:42.603388",[144,149,154,159,164,169],{"id":145,"question_zh":146,"answer_zh":147,"source_url":148},45731,"Android JNI 开发中，同时加载两个基于 ncnn 的动态库时出现 Fatal signal 6 (SIGABRT) 崩溃，如何解决？","该问题通常由多个动态库链接同一静态库（libncnn.a）导致的符号冲突引起。维护者已通过提交修复了此问题。解决方案是更新到包含修复的最新版本 ncnn。相关修复提交：https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcommit\u002F81c7df7056013819c757664cce46e68283399f57。如果仍遇到类似问题，请确保所有依赖 ncnn 的模块使用相同版本的 ncnn 库，并检查构建配置是否正确。","https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fissues\u002F976",{"id":150,"question_zh":151,"answer_zh":152,"source_url":153},45732,"在 Android 端进行压力测试时，发现偶发性推理结果不一致，可能与什么因素有关？","经用户测试分析，该问题与特定 CPU 架构（如 QCM2290、MSM8953 等 A53 系列）对 NEON 指令集的支持差异有关。当编译选项开启 `-DANDROID_ARM_NEON=ON` 时可能出现不一致，关闭该选项（`-DANDROID_ARM_NEON=OFF`）后问题消失。此外，多线程推理（`num_threads > 1`）会加剧该问题，设置为单线程（`num_threads = 1`）可避免。建议在这些设备上禁用 NEON 或强制单线程运行以规避问题。","https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fissues\u002F3143",{"id":155,"question_zh":156,"answer_zh":157,"source_url":158},45733,"PyTorch 模型转 ONNX 再转 ncnn 后，分类任务的输出概率值精度有偏差，如何处理？","推荐使用最新的 pnnx 工具直接将 PyTorch 模型转换为 ncnn 格式，以避免 ONNX 转换过程中的精度损失。安装和使用方法如下：\n1. 安装：`pip install pnnx`\n2. 转换命令：`pnnx model.onnx inputshape=[1,3,224,224]`\n详细文档参考：https:\u002F\u002Fgithub.com\u002Fpnnx\u002Fpnnx 和 https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fwiki\u002Fuse-ncnn-with-pytorch-or-onnx#how-to-use-pnnx","https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fissues\u002F1145",{"id":160,"question_zh":161,"answer_zh":162,"source_url":163},45734,"MTCNN 模型在 ncnn 上的检测结果与标准 Caffe 版本差异很大，可能是什么原因？","早期版本中曾存在膨胀卷积（dilated convolution）的 bug 导致结果不一致，该问题已在后续版本中修复（见 issue #75）。如果当前版本仍出现较大差异，请检查：1. 输入图像的预处理（mean\u002Fnorm 值、像素格式）是否与原版一致；2. 模型参数文件是否正确转换；3. 是否使用了最新版的 ncnn 库。建议对比每一层的中间输出以定位差异来源。","https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fissues\u002F52",{"id":165,"question_zh":166,"answer_zh":167,"source_url":168},45735,"Linux 环境下运行 ncnn 推理程序时出现 Segmentation fault (SIGSEGV)，常见原因有哪些？","虽然具体案例中评论未提供有效技术解答，但此类错误通常由以下原因引起：1. 模型文件损坏或路径错误；2. 输入数据维度与模型定义不匹配；3. 内存越界访问（如未正确初始化 Mat 对象）；4. 多线程竞争问题。建议启用 gdb 调试，查看崩溃堆栈中具体的层名称（如 `forward_layer`），并检查对应层的输入输出形状及参数设置。确保使用的是稳定版 ncnn 并复现最小代码示例进行排查。","https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fissues\u002F3390",{"id":170,"question_zh":171,"answer_zh":172,"source_url":153},45736,"如何在 ncnn 中正确设置归一化参数（mean_vals 和 norm_vals）以匹配原训练框架？","归一化参数需严格对应原模型的预处理方式。例如，若原模型使用 ImageNet 均值 [0.485, 0.456, 0.406] 和标准差 [0.229, 0.224, 0.225]，则 ncnn 中应设置为：\n```cpp\nconst float mean_vals[3] = {0.485f*255.f, 0.456f*255.f, 0.406f*255.f};\nconst float norm_vals[3] = {1\u002F0.229f\u002F255.f, 1\u002F0.224f\u002F255.f, 1\u002F0.225f\u002F255.f};\nncnn_img.substract_mean_normalize(mean_vals, norm_vals);\n```\n注意：若原模型输入为 0-1 浮点数，则无需乘 255；若为 0-255 整数，则需乘以 255。务必与原训练代码保持一致。",[174,179,184,189,194,199,204,209,214,219,224,229,234,239,244,249,254,259,264,269],{"id":175,"version":176,"summary_zh":177,"released_at":178},360709,"20260113","编译版本，默认配置，android-ndk-r29，xcode 16.4，ubuntu-22.04，ubuntu-24.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator，支持 GPU | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库 | arm64 |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |\n|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |\n|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |\n|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-visionos.zip | visionos 静态库 | arm64 |\n|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |\n|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |\n|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n新增sdpa layer和pnnx torch.scaled_dot_product_attention的转换，支持gqa合并\n新增rotaryembed layer\nsdpa支持kvcache\nmultiheadattention支持kvcache\nlayer可选实现support_vulkan_packing\nlayer可选实现support_vulkan_any_packing\nvulkan支持bf16开关，支持旧显卡模拟转换bf16\nrmsnorm vulkan优化(@futz12)\nselu vulkan优化(@futz12)\nvulkan eltwise统一elempack shader\n简化vulkan cast\n改善M较小时在N上切块的多线程调度\ngemm x86 avx512采用N维度16切块优化\nsdpa x86使用gemm和softmax优化(@futz12)\narm neon数学函数优先使用fma指令优化(@Abandon-ht)\nunaryop tan rvv优化(@ihb2032 @lyd1992)\n新增cmake NCNN_WINXP开关，不再主动定义_WIN32_WINNT宏\nc-api新增ncnn_version_number()接口返回数值\nc-api新增更多option setter getter接口\nnet加载模型接口新增wchar_t参数类型\n新增float8和bfloat8转换函数(@chloeee99)\n格式化glsl文件\n删除shader注释和额外的空格\n不再编译onnx2ncnn\nbenchncnn内置模型param，运行时不再需要param文件\n修复modelwriter访问空bias数据崩溃问题(@csukuangfj)\n修复param解析时尝试对已读取数据再次读取的逻辑错误(@futz12)\n修复softmax多线程尾部余数的错误，优化倒数计算(@futz12)","2026-01-13T03:28:01",{"id":180,"version":181,"summary_zh":182,"released_at":183},360710,"20250916","编译版本，默认配置，android-ndk-r28c，xcode 15.2，ubuntu-22.04，ubuntu-24.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator，支持 GPU | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库 | arm64 |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |\n|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |\n|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |\n|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-visionos.zip | visionos 静态库 | arm64 |\n|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |\n|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |\n|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n新增flip算子和pnnx torch.flip的转换\nclip x86 avx512循环剩余优化\ntanh和unaryop x86 avx512循环剩余优化(@lfalive)\nsigmoid x86 avx512循环剩余优化(@futz12)\ninstancenorm x86优化(@futz12)\ngroupnorm x86 sse2\u002Favx\u002Favx512优化\ngroupnorm arm neon优化(@mmyyy22)\nsigmoid和部分数学函数 loongarch lsx\u002Flasx 优化(@AtomAlpaca)\nshufflechannel riscv rvv\u002Fzfh\u002Fzvfh\u002Fxtheadvector优化(@AtomAlpaca)\nlayernorm riscv rvv\u002Fzfh\u002Fzvfh\u002Fxtheadvector优化(@Deepdive543443)\nlayernorm vulkan优化(@futz12)\n使用size_t类型改善超大尺寸tensor的支持\n修复x86 convolution int8 在启用avx512vnni时崩溃\n修复android asset datareader在新android系统和部分手机上崩溃的问题\n初始化layer featmask为空\n简化layernorm naive c实现\n修复convdw int8 dequantize pack8\n使用putenv和平台相关api修复llvm-mingw编译问题(@zhuzeitou)\n使用combine_x用于sse\u002Favx vector拼接\n修复rnn\u002Flstm\u002Fgru int8测试因rounding导致的差异\n更新ruapu探测risc-v zfh zvfh xtheadvector和动态分发\n删除已废弃的 Extractor::set_num_threads\u002Fset_vulkan_compute api\n修复cmake时编译器支","2025-09-16T02:38:18",{"id":185,"version":186,"summary_zh":187,"released_at":188},360711,"20250503","自 20250428 起无新功能\n修复 AMD radv coopmat 的黑名单问题\n针对 Qualcomm Adreno 的 Turnip 问题的 workaround","2025-05-03T09:58:44",{"id":190,"version":191,"summary_zh":192,"released_at":193},360712,"20250428","编译版本，默认配置，android-ndk-r28b，xcode 15.2，ubuntu-22.04，ubuntu-24.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator，支持 GPU | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库 | arm64 |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |\n|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |\n|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |\n|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-visionos.zip | visionos 静态库 | arm64 |\n|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |\n|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |\n|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\nx86 卷积 int8 gemm xop\u002Favx2\u002Favx512\u002Favx512vnni\u002Favxvnni\u002Favxvnniint8优化\nrisc-v eltwise rvv优化(@xfan1024)\nrisc-v bias rvv优化(@AtomAlpaca)\nrisc-v bnll rvv优化(@AtomAlpaca)\nrisc-v celu rvv优化(@AtomAlpaca)\n重构reduction\u002Fquantize\u002Fdequantize\u002Frequantize减小二进制体积\nsoftmax支持4维输入计算，优化支持任意elempack\nparam模型文件支持字符串类型的参数值，支持自然写法的数组\nreshape支持表达式动态shape，不再支持reshape内permute参数\ncrop支持表达式动态slice\ninterp支持表达式动态output size\n设置openmp环境变量解决多个openmp冲突问题和绑核失败问题\n防止cmake编译器检测的优化导致的误识别\n改善windows7+系统中的cpu大小核检测(@futz12)\n修复HarmonyOS NEXT get_elf_hwcap返回空的问题(@peerless2012)\n添加apple a18和m4系列的cpu型号识别\n修复在cpu l2\u003C1M时可能的convolution int8 gemm计算错误\n修复android编译可能的重定义VK_USE_PLATFORM_ANDROID_KHR问题\n修复mips\u002Floongarch\u002Frisc-v架构开启simplemath编译问题\nextractor clear()重置local allocator为0，修复多次clear的警告问题\n绕过nvidia新驱动下padding可能的程序卡死问题\nncnn2table量化校准工具支持读取npy数据(@wxqwinner)\nncnn2int8量化工具默认","2025-04-28T12:38:49",{"id":195,"version":196,"summary_zh":197,"released_at":198},360713,"20241226","编译版本，默认配置，android-ndk-r27c，xcode 15.2，ubuntu-20.04，ubuntu-22.04，ubuntu-24.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator，支持 GPU | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库 | arm64 |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |\n|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |\n|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |\n|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-visionos.zip | visionos 静态库 | arm64 |\n|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |\n|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |\n|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n嵌入支持int8量化\ngemm支持int8量化\nmultiheadattention支持int8量化\n新增spectrogram和inverse spectrogram实现\narm rmsnorm neon优化\narm layernorm neon fp32\u002Fbf16s\u002Ffp16s优化\nx86 rmsnorm sse2\u002Favx\u002Favx512优化\nx86 layernorm sse2\u002Favx\u002Favx512优化\nx86 gemm int8 sse2\u002Fxop\u002Favx\u002Favx512\u002Fvnni\u002Fvnniint8优化\n更新riscv vector标准到1.0，重写全部ncnn riscv优化代码，自动探测rvv\u002Fzfh\u002Fzvfh\u002Fxtheadvector并分发\nriscv gemm rvv优化支持128bit\u002F256bit vlen\n禁用x86倒数优化避免可能的精度损失\n改善harmonyos cpu拓扑结构abi兼容性\n暂时禁用mesa驱动的vulkan矩阵扩展支持\n兼容ndk-21编译asimdfhm目标的错误导致的问题\n兼容clang-18编译avx512bf16时编译器崩溃的问题\n禁用msvc对windows arm平台exp\u002Ftanh的svml优化以解决计算错误\n探测avxvnniint8\u002Favxvnniint16\u002Favxneconvert指令集\nruntime cpu开启时仅使用ncnn cmake内置的编译参数\n删除windows arm32支持(@Shironana817)\nandroid默认启用16kb pagesize编译，android-api升级到21\nvkCreateDevice失败时不直接崩溃(@Upliner)\n为powerpc架构跳过0.5附近数值的unaryop round测试用例\npnnx更新到torch-2.5\npnnx支持从traced inputs自动设定","2024-12-26T03:27:58",{"id":200,"version":201,"summary_zh":202,"released_at":203},360714,"20240820","编译版本，默认配置，android-ndk-r27，xcode 15.2，ubuntu-20.04，ubuntu-22.04，ubuntu-24.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator，支持 GPU | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库 | arm64 |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |\n|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |\n|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |\n|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-visionos.zip | visionos 静态库 | arm64 |\n|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |\n|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |\n|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n新增RMSNorm层和对应的pnnx转换，单元测试\nx86 convolution tiled gemm优化\n量化工具支持 rnn\u002Flstm\u002Fgru 动态量化\nx86 lstm int8 sse2\u002Fxop\u002Favx2\u002Favx512\u002Favx512vnni\u002Favxvnni优化\narm rnn\u002Flstm\u002Fgru int8 neon\u002Fasimdhp\u002Fasimddp优化\nmultiheadattention支持qdim参数与embed_dim不同\nmultiheadattention支持scale参数\n更新pybind11到2.12支持numpy2\n添加wasi支持(@quink-black)\n添加x86\u002Farm convolution\u002Fslice\u002Fconcat oom单元测试\nonnx2ncnn工具添加警告和推荐使用pnnx的信息输出(@lll143653)\n修复x86 avx512 vnni指令派发失效的问题\n增强x86\u002Farm计算内核在内存不足时的错误返回\n仅在windows arm平台使用ruapu指令集探测\nwindows mingw编译时支持大小核和SMT探测\n修复powerpc vsx计算abs可能的错误\n修复arm vfpv4条件下可能的fp16s\u002Fbf16s同时启用的冲突\n修复aarch64架构l2-cache很小时因gemm K分块可能的越界读错误\n修复riscv v tanh计算错误(@zhangyang2057)\narm\u002Fconvolution_3x3_pack1to8_fp16s使用ldr\u002Fstr替代ld1\u002Fst1优化(@quink-black)\n修复c_api无参数函数声明(@quink-black)\nc_api添加set_vulkan_device接口(@Baiyuetribe)\npyncnn添加从python bytes内存加载模型的接口(@joeyballentine)\n为VkAndroidHardwareBuffer","2024-08-20T08:45:24",{"id":205,"version":206,"summary_zh":207,"released_at":208},360715,"20240410","编译版本，默认配置，android-ndk-r26c，xcode 15.2，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator，支持 GPU | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库 | arm64 |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |\n|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |\n|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |\n|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |\n|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-visionos.zip | visionos 静态库 | arm64 |\n|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n解耦合layer cpu和vulkan，不再使用virtual public继承\n支持编译动态库时编译单元测试\n单层特性掩码支持禁用多线程\nextractor set_num_threads和set_vulkan_compute现在是无操作\ngpu shader增加uniform类型改善adreno上fp16兼容性\n检测vulkan矩阵扩展8x8x16配置，fp16a条件下默认使用fp16累加\n更新stb_image rvv\u002Fneon优化\nx86 mish avx512优化(@wnqn1597)\nriscv gemm fp32 rvv优化(@Xinyu302)\n加载模型上传权重时不保留无用的临时数据\nc-api新增draw rectangle\u002Ftext\u002Fcircle\u002Fline接口(@Deepdive543443)\n修复armv7平台加载fp16模型sigbus错误\n修复reduction L2norm denormal产生inf的问题\n修复arm平台pixel_resize rounding导致的数值误差\n修复softmax arm fp16计算错误\n修复risc-v rvv输出fp16没有自动转换的问题\n修复destroy_gpu_instance在驱动加载不完整时crash的问题(@shatyuka)\ndestroy_gpu_instance等待全部设备idle(@whyb)\n修复low-level api没有load_param直接create_pipeline可能的崩溃\n修复ncnnoptimize在shape推断的崩溃\nncnnoptimize支持更多新算子，修复gemm权重丢失问题\n被调试时候禁用signal指令集检测\nwindows-arm平台使用ruapu cpu指令集检测\narm vfpv4支持时启用自动转换fp16\n在arm64架构中总是报告支持neon和vfpv4\nsimplevk寻找更多已知的vulkan驱动路径\n修复旧cpp标准下risc-v rvv编译错误\n修复某些老编译器在debug模式下编译错误\n修复uwp平台编译\n修复test_reduction运行时的警告\n修复NCNN_PIXEL_DRAWING禁用时候编译错误(@shatyuka)\n支持MSVC使用LLVM openmp运行时的配合编译(@shat","2024-04-10T11:16:33",{"id":210,"version":211,"summary_zh":212,"released_at":213},360716,"20240102","编译版本，默认配置，android-ndk-r26b，xcode 13.4.1，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28 \n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip | 包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e |\n|ncnn-ios-simulator.zip | ios simulator 静态库，with and w\u002Fo bitcode | i386 + x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n内建vulkan驱动加载功能，不依赖vulkan-sdk编译gpu功能，可直接加载显卡驱动文件 \nmsvc编译启用arm neon指令加速，启用arm64 asimdhp编译 \n实现python pnnx pypi包和python调用接口\u002F文档(@Hideousmon) \narm convolution int8 直接卷积重构支持任意elempack \n优化 vulkan global pooling性能 \n优化resize bilinear性能 \n压缩字体数据减小二进制体积 \ndeconvolution支持动态权重和对应pnnx转换 \n新增跑分数据rank card(@Qengineering) \n支持big-endian架构平台，powerpc32位 \n添加woa linux ci \n添加msvc禁用exceptions\u002Frtti的编译开关 \n在macos上使用信号探测avx512指令集支持情况 \n支持寻找32位显卡驱动文件(@whyb) \n启用benchmark编译打印4维shape(@Deepdive543443) \n修复riscv-int8 sigmoid激活的测试失败问题(@MollySophia) \n修复deconvolution x86 bias非对齐访问的问题 \n修复prelu x86 sse指令非对齐访问的问题(@AIOa) \n修复windows上openmp设置线程数为0的警告 \n修复在支持16bit\u002F8bit的gpu上有关fp16sa shader使用fp16 shared变量的警告 \n修复nvidia vulkan驱动在程序退出的crash \n修复vkimagemat from_android_hardware_buffer缺失的elemsize参数错误 \n修复simpleocv Mat模板ptr的偏移错误 \n添加更过的gpu相关python绑定接口(@joeyballentine) \nandroid vulkan包的api版本降低到14\u002F21 \npnnx支持转换recompute_scale_factor=True的nn.Upsample \n新增nn.Identity测试 \n修复pnnx路径切分的问题 \n修复pnnx生成ncnn py空格对齐(@cmdbug) \npnnx生成的py可以直接执行推理 \npython pnnx返回优化后的torch模型 \n删除无用的代码(@ningjiang233) \n改善cmake toolchain文件(@zchrissirhcz) \n新增watchos和tvos ci \n修复linux sde ci的运行错误 \n更新POWER clang版本信息的文档(@JeremyRand) \n更新有关vulkan\u002Flibomp-dev依赖的文档(@JeremyRand) \n更新有关编译python模块CMAKE_TOOLCHAIN_FILE环境变量的文档(@JeremyRand) \n修复Rasberry拼写错误(@JeremyRand) \nFAQ新增有关pyncnn数据连续性的文档(@lll143653) \n更新readme下载页表格 \n添加Nintendo 3DS编译信息(@Deepdive543443) \n添加oncloud amlogic s805跑分数据(@mizu-bai) \n添加树莓派5 gpu跑分数据(@FantasyGmm) \n添加Jetson TX2跑分数据(@FantasyGmm) \n添加8gen2跑分数据(@m","2024-01-02T04:06:31",{"id":215,"version":216,"summary_zh":217,"released_at":218},360717,"20231027","编译版本，默认配置，android-ndk-r25c，xcode 13.4.1，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28 \n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip | 包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e |\n|ncnn-ios-simulator.zip | ios simulator 静态库，with and w\u002Fo bitcode | i386 + x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\nx86 卷积 int8 gemm 重构支持任意 elempack \nx86 卷积 int8 winograd 重构支持任意 elempack \narm 卷积 int8 gemm 重构支持任意 elempack \narm 卷积 int8 winograd 重构支持任意 elempack \ngelu Vulkan 优化（@FhqTreap） \nconvolution1d Vulkan 优化（@FhqTreap） \ngridsample x86 优化（@Yoh-Z） \nriscv gemm fp32 优化（@Xinyu302） \n新增 erf\u002Fshrink 和 ONNX 转换（@brightening-eyes） \n新增 diag 和 PNNX 转换（@wnqn1597） \n新增 celu 和 PNNX 转换（@wnqn1597） \n新增 simplemath，允许不依赖 libm 编译使用数学函数（@HonestDeng） \npooling adaptive 支持动态的输出尺寸和 PNNX 转换 \nelu selu 支持 4 维输入输出 \nslice 支持 indices 参数 \nmemorydata 支持 tag 参数和 fp16 存储 \nx86 selu shufflechannel 优化（@wnqn1597） \n修复卷积 Vulkan 在固定 shape 时的结果错误 \n修复权重 tag 潜在的溢出（@lrw04） \n按层加载模型减少内存占用（@daquexian） \n修复老版本 GCC 编译 AVX2 gather 的错误（@chainsx） \n修复老版本 GCC 编译 _mm256_set_m128 的错误（@whyb） \n修复新版本 Protobuf 编译问题 \n修复老版本 glibc round 编译问题 \n修复 c906 工具链编译错误 \npyncnn 启用 Vulkan 支持（@Hideousmon） \npyncnn 添加 load_param_mem 接口（@JeremyRand @theflyingzamboni） \nPNNX 支持 PyTorch 2.1 \nPNNX 消除 moduleop 的输出 unpack \nPNNX moduleop 将权重 shape 作为参数写入 param，内部权重顺序为使用顺序 \nPNNX 改善 reflect replicated pad 匹配 \nPNNX 合并 conv3d-bn 和 deconv3d-bn \nPNNX 转换 PyTorch narrow（@zyt1024） \nPNNX 转换 PyTorch lgamma（@shudorcl） \nPNNX 转换 PyTorch positive（@nicochen1118） \nPNNX 转换 PyTorch cumprod（@Jiang-Weibo） \nPNNX 转换 PyTorch mv\u002Fnn.ReplicationPad3d（@ShuRaymond） \nPNNX 转换 F.pairwise_distance（@Marsyule） \nPNNX 转换 PyTorch view_as_real\u002Fpytorch.view_as_complex（@Baiyuetribe） \n修复 PNNX 与新版本 Protobuf 编译问题（@HuPengsheet） \n修复 PNNX 改变目录下划线的错误 \nONNX2NCNN 支持 celu 转换（@brightening-eyes） \n自动为 pull request 添加 label \n修复 OHOS 工具链编译错误 \n改进 codeformat 脚本使用函数（@xiezheng-XD） \n添加 rk3566","2023-10-27T06:17:10",{"id":220,"version":221,"summary_zh":222,"released_at":223},360718,"20230816","编译版本，默认配置，android-ndk-r25c，xcode 13.4.1，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28 \n| 文件 | 内容 | 架构 |\n|---|---|---|\n|ncnn-full-source.zip | 包含全部 submodule 代码的完整源码 | | \n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e |\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e |\n|ncnn-ios-simulator.zip | ios simulator 静态库，with and w\u002Fo bitcode | i386 + x86_64 + arm64 |\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\n|ncnn-mac-catalyst.zip | mac catalyst 静态库，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\n\n实现全部的binaryop explicit广播规则类型 \nx86直接卷积权重变换的avx2\u002Favx512优化 \nx86 int8直接卷积支持任意elempack和sse2\u002Fxop\u002Favx2\u002Favx512\u002Fvnni优化 \nppc64 power8\u002Fpower9 vsx工具链支持，编译器检查和intrinsic翻译优化(@JeremyRand) \n更新glslang并启用VK_KHR_cooperative_matrix扩展和优化 \n修复pyncnn自定义layer模型权重加载 \nc_api新增Mat border\u002Flayer_to_index api(@Mek101) \nVkCompute::submit_and_wait现在能返回错误值(@Upliner) \n修复老版本clang编译时too many microtasks问题 \n修复clang-cl cpuid函数兼容性(@charlescao460) \n修复新版本protobuf c++17编译问题 \n修复老版本编辑器sleep递归调用错误(@whyb) \n编译时检查loongarch lasx扩展支持并自动启用 \n清理multiheadattention arm优化代码 \nbinaryop支持一维outer axis广播规则，保持旧的兼容行为 \nbenchncnn支持从命令行参数中指定自定义模型和输入(@tpoisonooo) \nmacos平台静态编译链接需要的系统库(@Baiyuetribe) \n更改amd集显上的显存分配策略为仅设备优先，修复在bios设置大显存时分配失败问题 \nonnx2ncnn遇到不支持transpose类型输出错误信息(@huoshuai-dot) \npnnx支持多算子到多算子的图变换 \npnnx新增转换torch.round\u002Ftrunc\u002Ffill\u002Findex_put\u002Fto\u002Ftype_as\u002Ftopk\u002Ffmod\u002Fcross\u002Ft\u002Fmaximum\u002Fminimum \npnnx合并chinese-clip\u002Fsam-iamge-encoder attention结构 \npnnx合并F.scaled_dot_product_attention \npnnx消除无用的expand\u002Fexpand_as\u002Ftype_as \npnnx修正fp16模型在优化时的权重变换错误 \npnnx修正负数shape索引越界问题(@Justin62628) \npnnx修复转换后py文件执行时权限错误(@zhenjiaguo) \npnnx转换ncnn global pooling后自动添加reshape \npnnx转换非zero padding模式的卷积到ncnn \npnnx转换2维nn.Linear为ncnn gemm \npnnx转换torch.stack为ncnn concat+reshape \npnnx转换torch.t到ncnn permute(@XiaBing992) \npnnx转换logsigmoid\u002Flog_softmax为ncnn sigmoid\u002Fsoftmax+log(@lrw04) \npnnx修复slice_copy输出的类型信息 \npnnx修复表达式中int64转换溢出问题 \npnnx修复reshape表达式消除后的ghost结点 \npnnx合并表达式时折叠shape为1的类似标量的权重 \npnnx合并表达式支持max\u002Fmin \npnnx改善图中有inplace操作时的输出结点连接探测，带来更多的常量折叠 \n添加ncnn","2023-08-16T05:54:44",{"id":225,"version":226,"summary_zh":227,"released_at":228},360719,"20230517","编译版本，默认配置，android-ndk-r25c，xcode 13.4.1，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-apple-vulkan.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e |\r\n|ncnn-ios-simulator.zip | ios simulator 静态库，with and w\u002Fo bitcode | i386 + x86_64 + arm64 |\r\n|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-mac-catalyst.zip | mac catalyst 静态库，with and w\u002Fo bitcode | x86_64 + arm64 |\r\n|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU，with and w\u002Fo bitcode | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\narm convolution winograd重构支持任意elempack\r\narm convolution sgemm重构支持任意elempack\r\narm convolution直接卷积重构支持任意elempack\r\narm deconvolution\u002Fmatmul 调用 gemm 完成计算\r\narm softmax支持任意elempack和bf16\u002Ffp16优化\r\narm multiheadattention fp16sa softmax优化\r\narm\u002Fx86 convolution1d直接卷积重构支持任意elempack和优化\r\n粗糙的vulkan gemm和multiheadattention优化\r\nmultiheadattention支持输入attention mask\r\nsigmoid\u002Fswish\u002Fclip\u002Fgelu\u002Fmish\u002Ftanh支持4d输入\r\n减少double类型的使用(@zhiliu6)\r\narm a53\u002Fa55架构检测和流水线优化\r\n允许注册自定义层替代内置实现\r\nx86 asin\u002Facos\u002Fatan\u002Fatan2 sse2\u002Favx\u002Favx512优化(@MouriNaruto)\r\nsse_mathfunc迁移floor\u002Fceil(@Yoh-Z)\r\nx86 mathfun迁移abs(@Yoh-Z)\r\nsimpleocv新增cv::imdecode内存加载图片(@AlOa)\r\n新增配合vulkan vma使用的三种扩展支持(@whyb)\r\n新增获取vkinstance的接口(@whyb)\r\n新增通用的sleep接口(@whyb)\r\ninnerproduct允许2维高度1的输入输出\r\n修复multiheadattention分配内存存在的多线程竞争问题\r\n修复在获取不到cache信息时的除0错误\r\n修复scale avx512计算错误\r\n修复exynos9810非法指令错误\r\n老旧adreno驱动中禁用fp16a以解决计算错误\r\n绕过n卡padding shader编译错误\r\n移除platform.h中无用的aarch64判断(@dreamcmi)\r\n修正modelwriter squeeze层参数id错误(@irexyc)\r\n修复gcc-13编译错误(@hillwoodroc)\r\n修复gcc-5.2 aarch64编译错误\r\n修复aosp编译错误(@caofx0418)\r\n修复n卡上benchmark退出时的crash(@triple-Mu)\r\n修复获取cpu cache信息潜在的fd泄漏\r\n优化lightmode循环条件(@MambaWong)\r\n绕过新版moltenvk的兼容性问题\r\n绕过n卡在multiheadattention softmax结果偶发nan的兼容性问题\r\n调用cpu.h接口时强制初始化全局cpu信息\r\npnnx支持torch-2.0\r\npnnx支持complex数据类型\r\npnnx转换torch.baddbmm\u002Ftorch.mm\u002Ftorch.stft家族\u002Ftorch.std\u002FF.scaled_dot_product_attention\r\npnnx支持fp16权重的torchscript\r\npnnx支持非forward的其他函数入口\r\npnnx当只有一个动态维时候折叠reshape的shape表达式\r\npnnx识别常数常量和表达式中的折叠\r\npnnx自动删除maxpool无indices输出项\r\npnnx总是删除convtransposed output_size参数\r\npnnx合并gelu表达式\r\npnnx合并vit\u002Fclip\u002Fdiffusers attention\r\n修正pnnx的RNN\u002FGRU省略输出项的python代码生成\r\n修正pnnx转换ir时潜在的负INT_MAX下溢问题\r\n修正pnnx fprintf类型不匹配(@kernelbin)\r\n修复pnnx windows编译错误(@Yoh-Z)\r\npyncnn model zoo添加yolov7-tiny(@kennybradley)\r\npyncnn model zoo添加yolov8s(@triple-Mu)\r\nmacos pypi包使用完整版本号\r\n改善wasm ci编译效率\r\n更新ci swiftshader版本\r\n更新cmake ios toolchain，新增ios-simulator arm64和mac catalyst ci\r\n添加qnx toolchain和编译步骤(@zchrissirhcz)\r\n删除ubuntu-18.04的ci\r\n更新3A5000 benchmark数据(@wzyforgit)\r\n新增2K1000LA benchmark数据(@lrzlin)\r\n新增icpc icc benchmark数据(@mizu-bai)\r\n新增Hyper-V Linux Guest benchmark数据(@MouriNaruto)\r\n新增和更新op4lts\u002Fop5\u002FVF2\u002FFT2000\u002F3A4000 benchmark数据(@MobtgZhang)\r\n更新centos编译文档(@inisis)\r\n更新windows msvc编译文档(@kernelbin)\r\nfaq新增关于cmake版本升级的内容(@inisis)\r\nfaq新增关于显卡节能模式的内容(@whyb)\r\n修正citation和benchmark文档中的拼写错误(@zchrissirhcz)\r\n修正pnnx代码和readme中的拼写错误(@jsyzdej @zchrissirhcz)\r\n\r\n## New Contributors\r\n* @whyb made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4614\r\n* @AlOa made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4557\r\n* @caofx0418 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4622\r\n* @dreamcmi made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4628\r\n* @kernelbin made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4647\r\n* @lrzlin made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4656\r\n* @irexyc made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4658\r\n* @jsyzdej made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4665\r\n* @hillwoodroc made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4684\r\n* @MobtgZhang made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4687\r\n* @kennybradley made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4693\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20230223...20230517","2023-05-17T08:48:03",{"id":230,"version":231,"summary_zh":232,"released_at":233},360720,"20230223","编译版本，默认配置，android-ndk-r25c，xcode 14.0.1，ubuntu-18.04，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-3.1.28\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\n扩充binaryop broadcast规则\r\n新增copyto算子，对应于torch inplace slice copy操作\r\nx86 gemm优化，新增 transpose_output 参数\r\nx86 multiheadattention优化\r\nx86 groupnorm优化(@EdVince)\r\narm gemm优化，包括fp16s\u002Ffp16sa\r\narm gelu优化(@EdVince)\r\narm multiheadattention优化(@EdVince)\r\n新增获取cpu l2\u002Fl3 cache大小接口，通过sysconf\u002Fwin32-api和linux sysfs\r\nx86 gemm 依据l2 cache分块的优化\r\nx86 convolution\u002Fdeconvolution\u002Fdeformableconv2d\u002Fmatmul 调用 gemm 完成计算\r\nx86 convolution winograd重构支持任意elempack\r\nx86 convolution直接卷积重构支持任意elempack\r\nx86公共的bfloat转换函数\r\nslice\u002Feltwise\u002Fconcat支持4d输入\r\nc api新增获取output indexes names接口\r\n改善vulkan winograd f43 fp16计算数值稳定性\r\n修复gpu信息bug bliz初始化问题(@weirdseed)\r\n修正arm bfloat2float和float2bfloat命名相反的问题\r\n更新riscv winograd f32系数，修复一些警告\r\n更好的riscv rvv tanh实现\r\n为ncnnoptimize\u002Fncnn2int8添加新加的算子和参数\r\n修复musl libc编译问题\r\n更新stb image和image write，启用arm neon优化\r\n更新emsdk版本到3.1.28，开启SIMPLEOCV(@ncnnnnn)\r\npnnx新增torch.cumsum转换(@csukuangfj)\r\npnnx新增torch.atan2\u002Flog10转换\r\npnnx自动替换pow(x,2)为square(x)\r\n修正pnnx windows slice end参数问题(@Yoh-Z)\r\npnnx自动删除无用的Tensor.clone(@Yoh-Z)\r\npnnx自动展开模型输入tuple和list类型\r\npnnx转ncnn时分析binaryop broadcast规则并插入适当的reshape\r\npnnx折叠常数常量，修复常数转换MemoryData兼容性问题\r\npnnx合并pixel unshuffle(@Yoh-Z)\r\n去除pnnx readme多余空行(@inisis)\r\n去除pnnx无用的include(@XiangYyang)\r\n修正pyncnn output_indexes接口错误(@wyushun)\r\n修复最新macos vulkan sdk兼容性问题(@w1ndseeker)\r\n删除python代码无用的import(@dianjiaogit)\r\n修复macos ci的xcode版本和vulkan sdk安装问题\r\n更新ci中已废弃的create release步骤\r\n添加CITATION.cff(@tpoisonooo)\r\n更新cpu benchmark数据(@wzyforgit)\r\n修复README编译状态badge(@tpoisonooo)\r\n修复README编译链接(@tuduweb)\r\n修正拼写错误(@hwdef @hiteshhedwig)\r\n添加ncnn-fortran例子(@mizu-bai)\r\n添加sherpa-ncnn实时语音识别例子(@csukuangfj)\r\n\r\n## New Contributors\r\n* @mizu-bai made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4423\r\n* @inisis made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4428\r\n* @dianjiaogit made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4378\r\n* @wyushun made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4453\r\n* @w1ndseeker made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4472\r\n* @hiteshhedwig made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4486\r\n* @weirdseed made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4493\r\n* @XiangYyang made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4497\r\n* @tuduweb made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4530\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20221128...20230223","2023-02-23T04:56:12",{"id":235,"version":236,"summary_zh":237,"released_at":238},360721,"20221128","编译版本，默认配置，android-ndk-r25b，xcode 12.4，ubuntu-18.04，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\n新增loongarch64 lsx向量指令集优化，包括absval\u002Fbatchnorm\u002Fbias\u002Fbinaryop\u002Fcast\u002Fclip\u002Fconcat\u002Fconvolution1d\u002Fconvolutiondepthwise\u002Fconvolution\u002Fcrop\u002Fdeconvolutiondepthwise\u002Fdeconvolution\u002Fdequantize\u002Fdropout\u002Feltwise\u002Fflatten\u002Fhardsigmoid\u002Fhardswish\u002Finnerproduct\u002Finterp\u002Fmish\u002Fpacking\u002Fpadding\u002Fpooling\u002Fprelu\u002Fquantize\u002Frelu\u002Frequantize\u002Fsigmoid\u002Fslice\u002Fsoftmax\u002Fswish\u002Ftanh\u002Funaryop算子(@junchao-loongson)\r\nlayernorm x86优化(@LinHeLurking @LRY89757)\r\nbatchnorm\u002Felu\u002Fprelu\u002Fgelu x86优化(@LRY89757)\r\nsoftmax arm neon优化(@luqiang-guo)\r\nbatchnorm\u002Finstancenorm riscv vector优化(@thelastlin)\r\ndeformableconv2d x86优化(@miemie2013)\r\nelu vulkan优化(@Yoh-Z)\r\nconvolution int8 x86 sse2\u002Favx2优化\r\n更新riscv vector segment load\u002Fstore(@thelastlin)\r\n改善内存池回收机制(@LinHeLurking)\r\n新增获取cpu物理核心数量api，默认线程数设为物理大核心数量\r\n实现控制单层运算特性是否启用的参数\r\n更通用的macos\u002Fios cpu特性探测过程，a15\u002Fa16\u002Fm2启用bf16和i8mm指令集\r\n统一innerproduct x86 fp32\u002Ffp16s内核代码\r\n修复在android省电模式cpu离线导致openmp崩溃的问题\r\n实现glu算子与对应的pnnx转换(@csukuangfj)\r\n新增fold和unfold算子\r\n新增gridsample算子与对应的pnnx转换(@LRY89757)\r\nlstm支持proj_size参数\r\ngroupnorm支持1d\u002F2d\u002F4d输入计算\r\nsqueeze\u002Fexpanddims支持4d输入输出\r\nmultiheadattention支持kdim vdim参数\r\n修复convolutiondepthwise allocator的错误设置(@w8501)\r\n修正windows arm环境中convolution权重为空的问题\r\n修复onnx2ncnn blob名字超出255长度的问题(@ZhangGe6)\r\n修正expanddims axes参数id错误的问题(@LiuYi-Up)\r\n修正c api allocator无法工作的问题(@qiqikit)\r\n更严格的编译器armv7 fp16功能检查和兼容\r\n修复老版本gcc编译avx512代码的编译错误(@bestpower)\r\n修复windows-arm64编译(@zchrissirhcz)\r\n修复在老版本ndk引用ncnn链接atomic内置函数失败的问题\r\n修复新版本pybind11编译错误(@tpoisonooo)\r\npython模块支持mat.numpy()(@csukuangfj)\r\n更新pybind11和glslang子模块\r\npyncnn发布python 3.11包和windows arm版本\r\npnnx支持pytorch 1.13\r\npnnx现已支持在cpu上加载gpu导出的torchscript\r\npnnx保存onnx-zero模型文件\r\npnnx转换时将常量存储在临时文件减少内存占用\r\npnnx新增命令行参数fp16=0\u002F1控制是否用fp16保存onnx-zero\u002Fncnn模型\r\npnnx支持大部分数学函数转换，新增nn.Softmax2d\u002Fnn.Fold\u002Fnn.Unfold\u002FF.fold\u002FF.unfold\u002Fbitwise_left_shift\u002Fbitwise_right_shift转换\r\npnnx改善和匹配inplace slice copy操作\r\n融合更多静态的F.convND\u002FF.linear为nn module\r\n合并临接的reshape\r\n合并pad到conv中\r\n改善pnnx F.softmax转换对dtype兼容性(@EdVince)\r\n修正pnnx softmax\u002Fnormalize\u002Fslice负数axis转换错误的问题\r\n修正pnnx slice end下标错误问题\r\n修正pnnx转ncnn保存fp16权重没考虑对齐的问题\r\npnnx遇到动态size时不再折叠为常量\r\npnnx自动折叠new_full\u002Ffull_like\r\nyolov5示例支持yolov5 6.2(@shaoshengsong)\r\n修复编译警告(@tpoisonooo @veahow)\r\n删除无用空行(@MollySophia @Menci)\r\n修正空格对齐(@tonori)\r\n修正拼写错误(@LRY89757 @Zepan @eltociear)\r\n忽略.xmake目录，CMakeSettings.json，Visual Studio CMake文件(@zchrissirhcz)\r\n重构README(@septs)\r\n改善README布局(@magicse)\r\n添加一些示例项目链接(@magicse @shaoshengsong)\r\nfaq新增有关禁用fp16设置的内容(@MisakaBit)\r\n更新riscv rvv ci\r\n新增c906 ci\r\n新增loongarch64 lsx ci\r\n迁移部分github action ci到腾讯ci\r\n新增TH1520 cmake toolchain(@luyanaa)\r\n切分大型单元测试加快多进程测试速度\r\n新增Intel Celeron M 420跑分(@MouriNaruto)\r\n新增T-Head TH1520跑分(@YuzukiTsuru)\r\n新增rock5b rk3588跑分(@hwdef)\r\n\r\n## New Contributors\r\n* @LinHeLurking made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4065\r\n* @septs made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4114\r\n* @w8501 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4173\r\n* @MollySophia made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4187\r\n* @Menci made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4188\r\n* @magicse made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4193\r\n* @tonori made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4217\r\n* @YuzukiTsuru made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4240\r\n* @ZhangGe6 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4236\r\n* @MisakaBit made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4248\r\n* @LiuYi-Up made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4259\r\n* @veahow made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4274\r\n* @csukuangfj made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4283\r\n* @Zepan made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4287\r\n* @bestpower made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4294\r\n* @shaoshengsong made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4328\r\n* @junchao-loongson made their first contri","2022-11-28T05:45:49",{"id":240,"version":241,"summary_zh":242,"released_at":243},360722,"20220729","编译版本，默认配置，android-ndk-r24，xcode 12.4，ubuntu-18.04，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\nbatchnorm avx512 优化(@LRY89757)\r\n新增DeformableConv2d层和单元测试(@miemie2013)\r\n修复conv3x3 winograd tensorcore权重数据错乱导致结果出错的问题\r\n修复memorydata 4维数据转换的问题\r\npnnx转换torchvision.ops.DeformConv2d到ncnn\r\npnnx自动删除无用的 mul + torch.ones 和 add + torch.zeros\r\npnnx修复动态shape时删除无用pad可能的崩溃问题\r\npnnx修复动态shape时错误删除upsample的问题\r\n添加sse优化文档(@DC-Zhou)\r\n更加严格的编译器riscv vector支持检查，删除rvv-0.7.1编译支持\r\n更新ci中android ndk路径，使用android-ndk-r24打包\r\n\r\n## New Contributors\r\n* @DC-Zhou made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4053\r\n* @miemie2013 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4070\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20220721...20220729","2022-07-29T03:04:38",{"id":245,"version":246,"summary_zh":247,"released_at":248},360723,"20220721","编译版本，默认配置，android-ndk-r23c，xcode 12.4，ubuntu-18.04，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\narmv5 convolution gemm int8优化\r\narmv6 dsp convolution gemm int8优化\r\narmv6 dsp convolution int8 winograd优化\r\nmips msa\u002Floongson mmi convolution int8 winograd优化\r\narmv8.4 i8mm convolution gemm int8优化\r\n探测编译器armv8.4\u002Farmv8.6的支持情况\r\n优化innerproduct fp16s权重转换的内存消耗\r\n统一arm eltwise不同elempack的分支\r\n修复多线程下arm rnn\u002Fgru\u002Flstm计算结果错误的问题\r\n修复android-ndk-r16b编译多线程运行报错的问题\r\nloongarch架构强制识别为mips以提升性能(@HougeLangley)\r\n修复非常老版本的gcc编译错误\r\nMat创建时检查OOM\r\n修复在android api 26编译找不到vkGetAndroidHardwareBufferPropertiesANDROID符号的问题\r\n修复x86 fp32转fp16可能存在的内存泄漏\r\npnnx支持torch 1.12\r\npnnx识别torchscript文件格式并输出报错\r\npnnx转换torch.tensor_split\r\npnnx合并多次同轴slice为tensor_split，修正插入位置\r\npnnx去除无用的一倍upsample\r\npnnx转ncnn时合并多个BinaryOp为加权求和Eltwise\r\npnnx合并megvii风格的shufflechannel+slice\r\n添加pkgconfig(@djdisodo)\r\n优化检测示例后处理nms(@jedi007)\r\nexample检查加载模型返回值(@zchrissirhcz @jedi007)\r\n添加Loongson2F toolchain(@luyanaa)\r\n添加君正x2000 toolchain\r\n添加ncnn svg图标(@ArchieMeng)\r\n改善protobuf FAQ文档(@tpoisonooo)\r\nREADME添加ncnn-android-yolov7(@xiang-wuu)\r\n添加yolov7示例(@cmdbug)\r\n添加yolov7_pnnx示例(@hariag)\r\nbenchmark新增fastestdet模型(@dog-qiuqiu)\r\n新增armv8.6 ci和coverage\r\n新增x86无sse ci\r\n新增x86 address sanitizer ci\r\n\r\n## New Contributors\r\n* @djdisodo made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3984\r\n* @jedi007 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4001\r\n* @xiang-wuu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4038\r\n* @ArchieMeng made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4037\r\n* @HougeLangley made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F4044\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20220701...20220721","2022-07-21T03:48:58",{"id":250,"version":251,"summary_zh":252,"released_at":253},360724,"20220701","编译版本，默认配置，android-ndk-r23c，xcode 12.4，ubuntu-18.04，ubuntu-20.04，ubuntu-22.04，vs2015，vs2017，vs2019，vs2022，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x86_64 + arm + arm64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\nx86\u002Farm\u002Fmips\u002Frisc-v\u002Fvulkan 去除无用的权重内存占用\r\n改善x86\u002Farm\u002Fmips\u002Frisc-v winograd卷积选择策略，独立出dot函数\r\n合并逐元素运算算子不同elempack的实现\r\nx86 sse\u002Favx bnll\u002Ftan优化(@jasonZhang892)\r\nx86 avx512 tanh优化(@jasonZhang892)\r\nx86 winograd 输入变换函数优化\r\nx86 sse\u002Favx convolution winograd23\u002F42 pack1优化\r\nx86 f16c innerproduct fp16s优化\r\narm neon tan\u002Farcsin\u002Farcos优化(@jasonZhang892)\r\n改善arm sgemm卷积的选择策略\r\narm neon pooling bf16s优化\r\narm neon innerproduct汇编优化\r\narmv7 vfpv4编译探测和运行时检测\r\narmv7 vfpv4 cast fp16优化\r\narmv7 vfpv4 innerproduct fp16s优化\r\narmv5 sgemm卷积优化\r\nmips msa cast fp16优化\r\nmips sgemm卷积\u002Fconvdw3x3优化\r\nmips msa innerproduct fp16s优化\r\nloongson mmi convolution gemm int8优化\r\nrisc-v vector erfc和gelu优化(@thelastlin)\r\n优化sgemm和winograd尾尺寸寄存器布局\r\nrisc-v sgemm\u002Fwinograd卷积\u002Finnerproduct\u002Fconvdw3x3优化\r\navx512bf16\u002Favx512fp16编译探测和运行时检测\r\navx512bf16\u002Favx512fp16 cast bf16\u002Ffp16优化\r\narmv8.2 asimdfhm，armv8.4 bf16 i8mm，armv8.6 sve sve2编译探测和运行时检测\r\n新增einsum层实现和对应的pnnx转换\r\nsimpleomp支持libgomp abi\r\nlayernorm支持一维输入和沿w做norm\r\nrnn\u002Flstm\u002Fgru支持openmp多线程加速\r\nmali-t760启用fp16运算\r\n更多的binaryop arm\u002Fmips\u002Friscv\u002Fx86特化实现\r\n修复unaryop_x86 gcc-4.4编译问题(@Yoh-Z)\r\n修复Mat fill gcc-4.4编译问题(@Yoh-Z)\r\n添加Power层单元测试(@proydakov)\r\nyolov5_pnnx示例自动适应不同的num_class数量(@FeiGeChuanShu)\r\n修正yolox输入shape w!=h的情况(@FeiGeChuanShu)\r\n修复armv7中fp16转fp32发生bus error的问题(@cugxchen)\r\n修复convdw\u002Fdeconvdw的avx512代码路径(@Yoh-Z)\r\n修复imreadwrite中total可能的溢出问题(@Z841973620)\r\n去除scale_x86中无效的死代码(@luyanaa)\r\n去除pnnx ir中无效的死代码(@moozae)\r\n修正test_deconvolutiondepthwise3d中printf参数错误(@Nlzy)\r\n修复mips架构上cmake寻找thread的错误\r\n修复在不支持cooperative matrix扩展测试崩溃的问题\r\n改善risc-v vector vfredsum\u002Fvfredusum编译器兼容(@thelastlin)\r\n修复某些arm编译器循环优化劣化的问题\r\n更新glslang，修复在使用系统glslang的include路径问题\r\n拆分arm82源码到单独文件，减小编译体积和内存占用\r\n修复ios universal arm82编译开关启用的问题\r\n统一winograd函数命名\r\n修复padding arm编译警告\r\n修复ios\u002Ftools\u002Farm82\u002Fnon-int8编译警告(@proydakov)\r\n修复LGTM警告(@proydakov)\r\npnnx支持转换torch bmm\u002Fmin\u002Fmax\u002Feinsum\u002Farange\u002Fbitwise_and\u002Fbitwise_not\u002Fbitwise_or\u002Fbitwise_xor\u002Feq\u002Fgather\u002Fge\u002Fgt\u002Fle\u002Flt\u002Fne\u002Fnorm\u002Findex_select\u002Fscatter_add\u002Fcomplex\u002Fimag\u002Freal\u002Ffft\u002Ffft2\u002Ffftn\u002Fhfft\u002Fhfft2\u002Fhfftn\u002Fifft\u002Fifft2\u002Fifftn\u002Fihfft\u002Fihfft2\u002Fihfftn\u002Firfft\u002Firfft2\u002Firfftn\u002Frfft\u002Frfft2\u002Frfftn，Tensor new_ones\u002Fnew_zeros\u002Fmasked_fill，F.normalize一维情况\r\npnnx ir支持复数数据类型\r\npnnx支持转换Tensor select到ncnn\r\n新增pnnx导出为onnx函数\r\npnnx导出ncnn fp16存储设计为一个pass\r\npnnx添加更多hardsigmoid合并模式\r\npnnx合并multiheadattention的尾部unpack\r\npnnx在静态shape输入时能有效的折叠常量\r\npnnx合并静态权重的卷积F.convND为nn.ConvND\r\n修复pnnx生成slice表达式遇到动态参数崩溃的问题\r\npnnx支持dict作为模型输出的转换\r\npnnx转换ncnn模型遇到4d\u002F5d输入nn.Linear自动添加reshape\r\npnnx去除单输入的cat算子\r\npnnx在合并表达式时跳过可折叠的常量\r\npnnx兼容更多inplace风格的算子，改善子图匹配浮点和整数比较\r\npnnx导出moduleop时存出所有内部权重\r\npnnx添加vit_b_32和convnext端到端模型测试\r\npnnx添加swin_transformer模型测试\r\ngitignore添加python生成的文件(@triple-Mu)\r\n添加c906和c910 v240 toolchain\r\n迁移pnnx，loongarch和gpu的ci到自建服务器\r\n修复loongarch ci\r\n添加avx512 spr cpu ci\r\n更新ci的qemu版本\r\ncmake安装目标路径采用gnuinstalldirs\r\n限制github action配置的权限(@nathannaveen)\r\nci的swiftshader使用单线程\r\n修复vs2022 ci中protobuf兼容问题\r\n添加关于关闭android界面和设置cpu\u002Fgpu性能模式的信息\r\n更新README中QQ群信息(@zchrissirhcz)\r\nREADME中的YOLOV改为YOLOv(@zhiqwang)\r\n更新树莓派和d1的编译文档\r\n更新添加自定义算子文档的过时信息(@LRY89757)\r\n添加英文版faq文档(@Jianbo-Ning)\r\n添加英文版build-for-visualstudio文档(@dankernel)\r\n添加vision_transformer benchmark(@tpoisonooo)\r\n更新rk3399 rk3288 gpu benchmark数据\r\n更新qcom810 qcom855plus benchmark数据\r\n添加Jetson AGX Orin\u002FJetson AGX Xavier\u002FAX620A benchmark数据(@BUG1989)\r\n添加loongson和sunway benchmark数据(@wzyforgit)\r\n添加RK3588 benchmark数据(@FeiGeChuanShu)\r\n添加amd 5700g benchmark数据(@hwdef)\r\nrelease添加ubuntu 22.04预编译包\r\nrelease android采用ndk-r23c编译\r\nrelease预编译包保护软链接\r\n\r\n## New Contributors\r\n* @LRY89757 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3741\r\n* @nathannaveen made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3758\r\n* @Z841973620 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3757\r\n* @Nlzy made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3774\r\n* @cugxchen made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3779\r\n* @luyanaa made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3821\r\n* @triple-Mu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3824\r\n* @Jianbo-Ning made their first contribution in https:\u002F","2022-07-01T05:50:56",{"id":255,"version":256,"summary_zh":257,"released_at":258},360725,"20220420","编译版本，默认配置，android-ndk-r21d，xcode 12.4，ubuntu-18.04，ubuntu-20.04，vs2015，vs2017，vs2019，vs2022，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x86_64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\nconv vulkan im2col+sgemm优化\r\nconv vulkan winograd43优化\r\nconv vulkan implicit gemm优化\r\ndeconv vulkan sgemm+col2im优化\r\nconv\u002Fdeconv vulkan local memory优化\r\nconv vulkan 直接卷积unroll优化\r\n改善conv vulkan winograd23\u002Fwinograd43选择策略\r\n融合conv vulkan winograd 前后的pad\u002Fcrop到transform中\r\ninnerproduct vulkan 拆分两阶段优化\r\n补全conv 1x1 vulkan任意packing\r\n补全conv 3x3 winograd vulkan任意packing\r\nconv\u002Fdeconv vulkan pack4 nvidia tensorcore优化\r\nx86 sse\u002Favx 数学函数优化(@Yoh-Z)\r\nunaryop x86 优化(@Yoh-Z)\r\nfloor\u002Fceil\u002Fabs x86 sse优化(@MouriNaruto)\r\nconvoluition\u002Fconvoluitiondepthwise\u002Finnerproduct\u002Fpadding\u002Fpooling\u002Finterp\u002Feltwise\u002Fcrop\u002Freshape\u002Fslice\u002Fhardsigmoid\u002Fswish\u002Fbinaryop\u002Fclip\u002Frelu\u002Fsigmoid\u002Funaryop x86 avx512优化\r\nconv sgemm avx512优化\r\nconv3x3 winograd avx512优化\r\ndeconvolution\u002Fdeconvolutiondepthwise x86直接反卷积实现\r\nsoftmax x86 sse\u002Favx\u002Favx512优化\r\nquantize\u002Fdequantize\u002Frequantize mips msa优化\r\nconv int8\u002Fconvdw int8\u002Finnerproduct int8 mips msa优化\r\nmultiheadattention arm neon优化(@EdVince)\r\nsoftmax arm neon优化\r\nconv3x3 winograd transform部分提出为可复用函数\r\nx86 f16c指令集检测和分发\r\n删除没什么用的avx2-fp16相关代码\r\nsimpleomp允许最多32个microtask参数\r\n添加loongson mmi头文件和编译支持\r\n新增deconv1d，deconv3d和对应的pnnx转换\r\n修正老版本gcc的avx512编译参数问题\r\n修正sigmoid x86在很大数值输入返回nan的问题\r\n修正gpu推理convdw发生unlocked pool allocator destoryed too early的问题\r\n避免mips msa推理时可能发生浮点数异常\r\nbatchnorm加载参数时避免除0异常\r\n为新算子更新modelwriter\r\ncopy_make_border添加reflect类型\r\nmali g31\u002Fg52启用fp16\r\n修复armhf工具链编译问题\r\nglobal pooling强制使用fp32累加避免nan问题\r\n修复某些android系统无法dlsym getauxval的问题\r\n修正新版本moltenvk tanh兼容问题\r\n提出vulkan激活函数，glsl中实现include\r\n修复armv7编译单元测试失败的问题(@jasonZhang892)\r\n修正conv3x3 winograd矩阵注释(@MouriNaruto)\r\n修正how-to-build拼写错误，更新jetson-nano编译文档(@tpoisonooo)\r\n更新ios编译文档(@mirrorsysu)\r\n一些注释和代码清理和修复编译警告(@tpoisonooo)\r\n修正readme中的单词大小写(@YoungSx)\r\n更新use-ncnn-with-own-project中的glslang的库列表\r\nci新增msvc arm\u002Farm64目标\r\nci新增linux loongarch目标\r\nci更新windows matrix和vs2022目标\r\n修复vs2019打包\r\n新增yolov5_pnnx例子\r\n新增nanodetplus_pnnx例子\r\n减少yolov5例子中后处理耗时(@UNeedCryDear)\r\n修复yolov5.py框位置问题(@hariag)\r\n更新ls2k1000的benchmark数据\r\npnnx支持转换torch unbind\u002Fones\u002Fones_like\u002Ffull\u002Ffull_like\u002Frandn_like\u002Fempty\u002Fempty_like\u002Faddmm\r\npnnx支持torch 1.11.0版本\r\npnnx转换的ncnn模型文件使用fp16保存\r\npnnx在linux上链接pthread，修复windows minmax编译问题\r\npnnx新增静态msvc crt cmake选项\r\n修正pnnx hardtanh 参数的ncnn转换\r\n修复pnnx macos动态库加载路径的问题\r\n\r\n## New Contributors\r\n* @MouriNaruto made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3591\r\n* @YoungSx made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3655\r\n* @hariag made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3656\r\n* @EdVince made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3667\r\n* @mirrorsysu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3696\r\n* @jasonZhang892 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3710\r\n* @UNeedCryDear made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3649\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20220216...20220420","2022-04-20T04:02:56",{"id":260,"version":261,"summary_zh":262,"released_at":263},360726,"20220216","编译版本，默认配置，android-ndk-r21d，xcode 12.4，ubuntu-18.04，ubuntu-20.04，vs2015，vs2017，vs2019，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x86_64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\nconv sgemm pack4\u002Fpack1to4\u002Fpack4to1 x86 sse2\u002Favx优化\r\nconv3x3s1 winograd pack4\u002Fpack4to1 x86 sse2\u002Favx优化\r\nconv int8 gemm pack8to4\u002Fpack8to1\u002Fpack1to8 x86 xop\u002Favx2\u002Favx512-vnni\u002Favx-vnni优化\r\nconv3x3s1 int8 winograd pack8to4\u002Fpack8to1 x86 xop\u002Favx2\u002Favx512-vnni\u002Favx-vnni优化\r\nscale x86 avx优化(Yoh-Z)\r\ninterp x86 avx优化(Yoh-Z)\r\nconv pack arm neon优化\r\nx86 avx512基础架构\r\n默认启用x86 avx512编译和运行时检测\r\n解耦合x86 fma和avx2\r\n不依赖libgcc的x86 cpu指令集探测\r\n支持动态权重的卷积\r\n修正可能因Mat成员函数没有内联导致的非法指令问题\r\n修正可能因函数对象实例没有内联导致的非法指令问题\r\n修正单元测试比较函数错误(yyuzhong)\r\nbinaryop\u002Funaryop\u002Freduction支持4维输入\r\n新增Tile层和torch.repeat的转换\r\n新增MatMul层和torch.matmul的转换\r\narmv8.2 dot编译为运行时可选\r\n支持sw_64平台(wzyforgit)\r\n增加c-api的cmake开关\r\nc-api增加默认mat构造函数(tpoisonooo)\r\n简化binaryop的函数对象代码(tpoisonooo)\r\n修正interp nearest在有非常规scale_factor参数计算错误的问题\r\n简化c-api自定义层forward_n参数类型\r\n删除非avx2编译时退化sse2的警告(kagurazakakotori)\r\n在64位编译时使用_mm_cvtsi128_si64降低内存访问(kagurazakakotori)\r\n修正low-level op api文档错误(FeiGeChuanShu)\r\n修正crop test缺失的doffset参数(xh-liu-tech)\r\n修正arm convolution pack1to4 int8权重重排(cmdbug)\r\n简化get_current_time平台相关宏(cmdbug)\r\n修正armv7无neon编译时计算错误的问题\r\n增加c906 v223工具链(zchrissirhcz)\r\n添加第二个qq技术交流群答案(LJoson)\r\npython ci禁用tools和examples构建\r\nci动态库编译禁用LTO\r\nci更新swiftshader-20220211\r\n删除travis ci和readme相关条目(proydakov)\r\n新增yolo-fastest模型benchmark(dog-qiuqiu)\r\n更新来自Q-engineering树莓派\u002Fjetson-nano等benchmark数据\r\nbenchmark增加zynq-7020\u002Fz8350\u002Fn5105\r\npnnx支持转换torch dequantize\u002Fquantize_per_tensor\u002Fquantized.linearrelu\u002Fargmax\u002Fargmin\u002Fclone\u002Fnormal\u002Fexpand\u002Fvar\u002Famax\u002Famin\u002Flogsumexp\u002Fprod\u002Fsum\u002Farange\u002Fmatmul\u002Fzeros_like\u002Fexpand_like\u002Fdeformconv2d\u002Froialign\u002Fnorm\u002Fstack\u002Frepeat\u002Fzeros\u002Froll\u002Fremainder\r\npnnx自动删除dropout算子\r\npnnx自动删除无pads的pad和noop算术表达式\r\npnnx常量折叠\r\npnnx转换4维常量数据\r\npnnx支持half数据类型导出的模型\r\npnnx转ncnn时删除尾部的reshape\u002Fpermute\r\npnnx合并conv1d-bn convtranspose1d-bn\r\npnnx合并单一维度全select为unbind\r\npnnx确保算子名唯一性\r\n修正pnnx转ncnn时遇到无法展开的表达式发生崩溃的问题\r\npnnx转ncnn支持负数pads的F.pad\r\npnnx转ncnn合并transpose-matmul\r\npnnx转ncnn在pooling123d前后增加升维和降维的reshape模拟nn.MaxPool123d处理无batch维数据的行为\r\npnnx命令行参数的shape指定输入类型\r\npnnx自动寻找pytorch安装目录(Yutyrannus)\r\npnnx ci自动拷贝dll文件(Yutyrannus)\r\n添加pnnx命令行工具用法说明(ling0322)\r\n\r\n## New Contributors\r\n* @wzyforgit made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3421\r\n* @dog-qiuqiu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3470\r\n* @xh-liu-tech made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3475\r\n* @ling0322 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3487\r\n* @kagurazakakotori made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3527\r\n* @LJoson made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3532\r\n* @Yoh-Z made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3540\r\n* @yyuzhong made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3556\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20211208...20220216","2022-02-16T03:13:41",{"id":265,"version":266,"summary_zh":267,"released_at":268},360727,"20211208","编译版本，默认配置，android-ndk-r21d，xcode 12.4，ubuntu-18.04，ubuntu-20.04，vs2015，vs2017，vs2019，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x86_64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\nMat数据结构支持4维\r\n新增Convolution3D, Pooling3D和对应的pnnx算子转换\r\n这些算子支持4维输入输出(Cast, Packing, ReLU, BatchNorm, Reshape, Flatten, Permute, Crop)和对应的pnnx算子转换\r\nC api增加4维mat\r\nConvolution1D常规的simd优化(sse\u002Favx\u002Fneon\u002Frvv\u002Fmsa)\r\n降低gpu推理时的cpu占用\r\n降低单元测试cpu占用\r\n改进pnnx转ncnn的batch轴识别\r\n更新operators文档\r\n修复开启simpleocv时仍然寻找系统opencv的问题(zchrissirhcz)\r\n修正p2pnet例子绘图bug(FeiGeChuanShu)\r\n支持c906 v2.2.2新工具链\r\n更可靠的ci任务取消机制\r\nci新增avx512和nvidia t4\r\n修复python wheel发布脚本\r\n更新ci lavapipe版本(ljtjerry)\r\n更新ci webassembly支持nodejs v16\r\n更新FAQ(zhaqu, Bright476, Rinfair-CSP-A016)\r\n修正拼写错误(cmdbug)\r\n\r\n## New Contributors\r\n* @zhaqu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3374\r\n* @ljtjerry made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3387\r\n* @Rinfair-CSP-A016 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3399\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fcompare\u002F20211122...20211208","2021-12-08T03:18:59",{"id":270,"version":271,"summary_zh":272,"released_at":273},360728,"20211122","编译版本，默认配置，android-ndk-r21d，xcode 12.4，ubuntu-18.04，ubuntu-20.04，vs2015，vs2017，vs2019，emscripten-2.0.8\r\n| file | content | arch |\r\n|---|---|---|\r\n|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | | \r\n|ncnn-android.zip | android 静态库\u002F动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-android-vulkan.zip | android 静态库\u002F动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |\r\n|ncnn-ios.zip | ios 静态库，with and w\u002Fo bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |\r\n|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w\u002Fo bitcode | arm64 + arm64e + x86_64 |\r\n|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |\r\n|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |\r\n|ncnn-ubuntu.zip | ubuntu linux 静态库\u002F动态库，支持 GPU，模型转换工具 | x86_64 |\r\n|ncnn-windows.zip | windows 静态库\u002F动态库，支持 GPU，模型转换工具 | x86 + x86_64 |\r\n|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |\r\n\r\nPNNX(PyTorch Neural Network Exchange)是PyTorch模型部署的新方式，可以避开ONNX中间商，导出比较干净的高层OP\r\nrisc-v v binaryop, hardswish, hardsigmoid, prelu, selu, dropout, gru, softmax优化(thelastlin)\r\nrisc-v v conv1x1 fc优化\r\narm neon requantize leakyrelu优化\r\narm neon innerproduct gemm int8优化\r\n针对c906 sgemm pack优化(yaobyPerfxlab, xianyi)\r\nx86 avx 卷积激活优化(zhiliu6)\r\nx86 sse convolution, convolutiondepthwise, pooling优化(Timen)\r\n修正layernorm affine计算错误\r\n修正pooling adaptive计算错误\r\n修正deconvolution output padding在有bias时的计算错误\r\ninterp支持cubic aligncorner插值\r\ninterp支持对2维数据w方向拉伸\r\n新增convolutiondepthwise1d和pnnx转换\r\nrnn\u002Flstm\u002Fgru支持不相等的输入输出个数\r\n修正squeeze和expanddims层axes的处理\r\n使用整数计算pooling adaptive参数上下界(Yutyrannus)\r\n修复armv7 neon round模式差异\r\n修复x86 sse\u002Favx round模式差异\r\n修复int8输入单元测试可能的越界读\r\n修复在某些android平台无法获得auxv变量的问题\r\n修正apple a11 a12检测armv8.2 dot扩展指令错误的问题\r\n内存引用加载模型时不再拷贝到内存\r\n修复pyncnn numpy转Mat时非对齐拷贝出错的问题\r\n正确检测和支持apple a15和m1(zchrissirhcz)\r\n修复AVX-only代码和用户提供opt时的单元测试逻辑(Timen)\r\nhardswish激活合并入convolution和innerproduct(zhiliu6)\r\n自动解耦extract的Mat数据与Net实例的内存池\r\nNet的custom_layer_to_index移到public(Timen)\r\n删除代码中的无用变量(Sinky-Yan)\r\ncmake检测esp32的xtensa架构\r\ncmake install安装ncnn工具(jinmingyi1998)\r\n修正hardswish test beta参数(zhiliu6)\r\n修复ncnnoptimize无法生成合理int8权重的问题\r\nncnnoptimize支持embd层\r\n修正onnx2ncnn concat算子负数axis转换的问题\r\n修复onnx2ncnn合并expand算子(grimoire)\r\n修复某些arm kernel越界读数据的问题\r\n修复NCNN_STDIO=OFF的编译问题\r\n新增YOLOX例子, 更新预处理逻辑(FateScript)\r\n新增RobustVideoMatting例子(FeiGeChuanShu)\r\n新增scrfd croudhuman例子(MarsTechHAN)\r\n新增YOLOv5 v6.0例子(zhiliu6)\r\n新增CrowdCounting-P2PNet例子(FeiGeChuanShu)\r\nreadme添加yolox(Sinky-Yan)\r\n更新readme文档(fzyzcjy)\r\n修复msvc编译器警告(TianZerL)\r\n一些拼写错误修正(cmdbug, huoshuai-dot)\r\n更新faq文档(ncnnnnn, luqiang-guo, zhiqwang, cmdbug, CharlesHuan, Shiro-Nana, zmq175)\r\n更新operators算子文档(soragotosann)\r\n更新d1和ls2k编译文档\r\n新增termux编译文档(Sinky-Yan)\r\n更新msvc编译文档(ncnnnnn)\r\n更新编译文档(dankernel, mlbo, xiguadong)\r\n更新macos openmp安装方法(zhiqwang)\r\n更新量化文档中的链接(ShiquanYu)\r\n修正python编译文档路径错误(nixondutt)\r\nbenchmark新增m1数据(zhiqwang)\r\nbenchmark新增mbp数据(AnnYellow)\r\nbenchmark新增khadas vim3 amlogic a311d数据(elejke, FeiGeChuanShu)\r\nbenchmark新增Phytium FT-2000+\u002F64数据\r\nbenchmark新增RK3568数据(BowShotDS)\r\nbenchmark新增RK3328数据(Liuyufanlyf)\r\nbenchmark新增Ingenic X2000和T40数据(MarsTechHAN)\r\nci更新swiftshader\r\nci新增基于lavapipe的gpu测试\r\nci删除travis arm32(Richuanwu)\r\nci更新xcode版本\r\n\r\n## New Contributors\r\n* @SinKy-Yan made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3124\r\n* @FateScript made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3110\r\n* @BowShotDS made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3145\r\n* @Liuyufanlyf made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3164\r\n* @yaobyPerfxlab made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3159\r\n* @TianZerL made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3188\r\n* @grimoire made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3189\r\n* @dankernel made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3248\r\n* @Richuanwu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3279\r\n* @ShiquanYu made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3283\r\n* @nixondutt made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3293\r\n* @mlbo made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3314\r\n* @luqiang-guo made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3332\r\n* @Yutyrannus made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3333\r\n* @xiguadong made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3344\r\n* @soragotosann made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3345\r\n* @huoshuai-dot made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3348\r\n* @fzyzcjy made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3358\r\n* @CharlesHuan made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3361\r\n* @Shiro-Nana made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fncnn\u002Fpull\u002F3368\r\n* @zmq175 made their first contribution in https:\u002F\u002Fgithub.com\u002FTencent\u002Fnc","2021-11-22T10:01:12"]