Ncnn

pnnx转换的ncnn模型文件使用fp16保存
pnnx在linux上链接pthread，修复windows minmax编译问题
pnnx新增静态msvc crt cmake选项
修正pnnx hardtanh 参数的ncnn转换
修复pnnx macos动态库加载路径的问题

New Contributors
* MouriNaruto made their first contribution in https://github.com/Tencent/ncnn/pull/3591
* YoungSx made their first contribution in https://github.com/Tencent/ncnn/pull/3655
* hariag made their first contribution in https://github.com/Tencent/ncnn/pull/3656
* EdVince made their first contribution in https://github.com/Tencent/ncnn/pull/3667
* mirrorsysu made their first contribution in https://github.com/Tencent/ncnn/pull/3696
* jasonZhang892 made their first contribution in https://github.com/Tencent/ncnn/pull/3710
* UNeedCryDear made their first contribution in https://github.com/Tencent/ncnn/pull/3649

**Full Changelog**: https://github.com/Tencent/ncnn/compare/20220216...20220420

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
| file | content | arch |
|---|---|---|
|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | |
|ncnn-android.zip | android 静态库/动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |
|ncnn-android-vulkan.zip | android 静态库/动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |
|ncnn-ios.zip | ios 静态库，with and w/o bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |
|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w/o bitcode | arm64 + arm64e + x86_64 |
|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |
|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-ubuntu.zip | ubuntu linux 静态库/动态库，支持 GPU，模型转换工具 | x86_64 |
|ncnn-windows.zip | windows 静态库/动态库，支持 GPU，模型转换工具 | x86 + x86_64 |
|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |

conv sgemm pack4/pack1to4/pack4to1 x86 sse2/avx优化
conv3x3s1 winograd pack4/pack4to1 x86 sse2/avx优化
conv int8 gemm pack8to4/pack8to1/pack1to8 x86 xop/avx2/avx512-vnni/avx-vnni优化
conv3x3s1 int8 winograd pack8to4/pack8to1 x86 xop/avx2/avx512-vnni/avx-vnni优化
scale x86 avx优化(Yoh-Z)
interp x86 avx优化(Yoh-Z)
conv pack arm neon优化
x86 avx512基础架构
默认启用x86 avx512编译和运行时检测
解耦合x86 fma和avx2
不依赖libgcc的x86 cpu指令集探测
支持动态权重的卷积
修正可能因Mat成员函数没有内联导致的非法指令问题
修正可能因函数对象实例没有内联导致的非法指令问题
修正单元测试比较函数错误(yyuzhong)
binaryop/unaryop/reduction支持4维输入
新增Tile层和torch.repeat的转换
新增MatMul层和torch.matmul的转换
armv8.2 dot编译为运行时可选
支持sw_64平台(wzyforgit)
增加c-api的cmake开关
c-api增加默认mat构造函数(tpoisonooo)
简化binaryop的函数对象代码(tpoisonooo)
修正interp nearest在有非常规scale_factor参数计算错误的问题
简化c-api自定义层forward_n参数类型
删除非avx2编译时退化sse2的警告(kagurazakakotori)
在64位编译时使用_mm_cvtsi128_si64降低内存访问(kagurazakakotori)
修正low-level op api文档错误(FeiGeChuanShu)
修正crop test缺失的doffset参数(xh-liu-tech)
修正arm convolution pack1to4 int8权重重排(cmdbug)
简化get_current_time平台相关宏(cmdbug)
修正armv7无neon编译时计算错误的问题
增加c906 v223工具链(zchrissirhcz)
添加第二个qq技术交流群答案(LJoson)
python ci禁用tools和examples构建
ci动态库编译禁用LTO
ci更新swiftshader-20220211
删除travis ci和readme相关条目(proydakov)
新增yolo-fastest模型benchmark(dog-qiuqiu)
更新来自Q-engineering树莓派/jetson-nano等benchmark数据
benchmark增加zynq-7020/z8350/n5105
pnnx支持转换torch dequantize/quantize_per_tensor/quantized.linearrelu/argmax/argmin/clone/normal/expand/var/amax/amin/logsumexp/prod/sum/arange/matmul/zeros_like/expand_like/deformconv2d/roialign/norm/stack/repeat/zeros/roll/remainder
pnnx自动删除dropout算子
pnnx自动删除无pads的pad和noop算术表达式
pnnx常量折叠
pnnx转换4维常量数据
pnnx支持half数据类型导出的模型
pnnx转ncnn时删除尾部的reshape/permute
pnnx合并conv1d-bn convtranspose1d-bn
pnnx合并单一维度全select为unbind
pnnx确保算子名唯一性
修正pnnx转ncnn时遇到无法展开的表达式发生崩溃的问题
pnnx转ncnn支持负数pads的F.pad
pnnx转ncnn合并transpose-matmul
pnnx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
pnnx命令行参数的shape指定输入类型
pnnx自动寻找pytorch安装目录(Yutyrannus)
pnnx ci自动拷贝dll文件(Yutyrannus)
添加pnnx命令行工具用法说明(ling0322)

New Contributors
* wzyforgit made their first contribution in https://github.com/Tencent/ncnn/pull/3421
* dog-qiuqiu made their first contribution in https://github.com/Tencent/ncnn/pull/3470
* xh-liu-tech made their first contribution in https://github.com/Tencent/ncnn/pull/3475
* ling0322 made their first contribution in https://github.com/Tencent/ncnn/pull/3487
* kagurazakakotori made their first contribution in https://github.com/Tencent/ncnn/pull/3527
* LJoson made their first contribution in https://github.com/Tencent/ncnn/pull/3532
* Yoh-Z made their first contribution in https://github.com/Tencent/ncnn/pull/3540
* yyuzhong made their first contribution in https://github.com/Tencent/ncnn/pull/3556

**Full Changelog**: https://github.com/Tencent/ncnn/compare/20211208...20220216

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
| file | content | arch |
|---|---|---|
|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | |
|ncnn-android.zip | android 静态库/动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |
|ncnn-android-vulkan.zip | android 静态库/动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |
|ncnn-ios.zip | ios 静态库，with and w/o bitcode | armv7 + arm64 + arm64e + i386 + x86_64 |
|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU，with and w/o bitcode | arm64 + arm64e + x86_64 |
|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |
|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-ubuntu.zip | ubuntu linux 静态库/动态库，支持 GPU，模型转换工具 | x86_64 |
|ncnn-windows.zip | windows 静态库/动态库，支持 GPU，模型转换工具 | x86 + x86_64 |
|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |

Mat数据结构支持4维
新增Convolution3D, Pooling3D和对应的pnnx算子转换
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
C api增加4维mat
Convolution1D常规的simd优化(sse/avx/neon/rvv/msa)
降低gpu推理时的cpu占用
降低单元测试cpu占用
改进pnnx转ncnn的batch轴识别
更新operators文档
修复开启simpleocv时仍然寻找系统opencv的问题(zchrissirhcz)
修正p2pnet例子绘图bug(FeiGeChuanShu)

Deep AI ncnn colorization siggraph17

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
| file | content | arch |
|---|---|---|
|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | |
|ncnn-android.zip | android 静态库/动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |
|ncnn-android-vulkan.zip | android 静态库/动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 + riscv64 |
|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |
|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 |
|ncnn-ios.zip | ios 静态库 | arm64 |
|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |
|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |
|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |
|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |
|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |
|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |
|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |
|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |
|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-visionos.zip | visionos 静态库 | arm64 |
|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |
|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |
|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-ubuntu.zip | ubuntu linux 静态库/动态库，支持 GPU，模型转换工具 | x86_64 |
|ncnn-windows.zip | windows 静态库/动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |
|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |

embed 支持int8量化
gemm 支持int8量化
multiheadattention 支持int8量化
新增spectrogram和inverse spectrogram实现
arm rmsnorm neon优化
arm layernorm neon fp32/bf16s/fp16s优化
x86 rmsnorm sse2/avx/avx512优化
x86 layernorm sse2/avx/avx512优化
x86 gemm int8 sse2/xop/avx/avx512/vnni/vnniint8优化
更新riscv vector标准到1.0，重写全部ncnn riscv优化代码，自动探测rvv/zfh/zvfh/xtheadvector并分发
riscv gemm rvv优化支持128bit/256bit vlen
禁用x86倒数优化避免可能的精度损失
改善harmonyos cpu拓扑结构abi兼容性
暂时禁用mesa驱动的vulkan矩阵扩展支持
兼容ndk-21编译asimdfhm目标的错误导致的问题
兼容clang-18编译avx512bf16时编译器崩溃的问题
禁用msvc对windows arm平台exp/tanh的svml优化以解决计算错误
探测avxvnniint8/avxvnniint16/avxneconvert指令集
runtime cpu开启时仅使用ncnn cmake内置的编译参数
删除windows arm32支持(Shironana817)
android默认启用16kb pagesize编译，android-api升级到21
vkCreateDevice失败时不直接崩溃(Upliner)
为powerpc架构跳过0.5附近数值的unaryop round测试用例
pnnx更新到torch-2.5
pnnx支持从traced inputs自动设定inputshape
pnnx编译不再输出来自torch头文件的警告
pnnx重排pass level2内的全部顺序，并复用pattern
pnnx不再保存debug中间模型(LJoson)
pnnx输出python脚本的onnx导出代码更新到export(whyb)
pnnx合并t5-layernorm为rmsnorm
pnnx不再折叠具有动态shape的tensor
pnnx在输出的python脚本中使用隐含的int转换避免trace时常数化
pnnx转换Tensor.select为ncnn crop+squeeze
pnnx转换onnx constantofshape为torch.zeros/ones
pnnx修正onnx clip在可选min/max缺失时的转换问题
ci更新riscv64工具链
ci添加c908/spacemit-x60
ci webassembly兼容node>20
ci android添加riscv64目标并打包
添加vim3 vulkan跑分数据(GIBEREZ)


New Contributors
* ankushgoel27 made their first contribution in https://github.com/Tencent/ncnn/pull/5709
* Shironana817 made their first contribution in https://github.com/Tencent/ncnn/pull/5811
* GIBEREZ made their first contribution in https://github.com/Tencent/ncnn/pull/5821

**Full Changelog**: https://github.com/Tencent/ncnn/compare/20240820...20241226

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
| file | content | arch |
|---|---|---|
|ncnn-full-source.zip |包含全部 submodule 代码的完整源码 | |
|ncnn-android.zip | android 静态库/动态库 | armeabi-v7a + arm64-v8a + x86 + x86_64 |
|ncnn-android-vulkan.zip | android 静态库/动态库，支持 GPU | armeabi-v7a + arm64-v8a + x86 + x86_64 |
|ncnn-apple.zip | apple xcframework，ios + ios-simulator + macos + mac-catalyst + watchos + watchos-simulator + tvos + tvos-simulator + visionos + visionos-simulator | arm64 + arm64e + x86_64 |
|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 |
|ncnn-ios.zip | ios 静态库 | arm64 |
|ncnn-ios-vulkan.zip | ios 静态库，支持 GPU | arm64 |
|ncnn-ios-simulator.zip | ios simulator 静态库 | x86_64 + arm64 |
|ncnn-ios-simulator-vulkan.zip | ios simulator 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-macos.zip | macos 静态库 | x86_64 + arm64 |
|ncnn-macos-vulkan.zip | macos 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-mac-catalyst.zip | mac catalyst 静态库 | x86_64 + arm64 |
|ncnn-mac-catalyst-vulkan.zip | mac catalyst 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-watchos.zip | watchos 静态库 | armv7k + arm64_32 |
|ncnn-watchos-simulator.zip | watchos simulator 静态库 | x86_64 + arm64 |
|ncnn-tvos.zip | tvos 静态库 | x86_64 + arm64 |
|ncnn-tvos-vulkan.zip | tvos 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-tvos-simulator.zip | tvos simulator 静态库 | x86_64 + arm64 |
|ncnn-tvos-simulator-vulkan.zip | tvos simulator 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-visionos.zip | visionos 静态库 | arm64 |
|ncnn-visionos-vulkan.zip | visionos 静态库，支持 GPU | arm64 |
|ncnn-visionos-simulator.zip | visionos simulator 静态库 | x86_64 + arm64 |
|ncnn-visionos-simulator-vulkan.zip | visionos simulator 静态库，支持 GPU | x86_64 + arm64 |
|ncnn-ubuntu.zip | ubuntu linux 静态库/动态库，支持 GPU，模型转换工具 | x86_64 |
|ncnn-windows.zip | windows 静态库/动态库，支持 GPU，模型转换工具 | x86 + x64 + arm + arm64 |
|ncnn-webassembly.zip | webassembly 静态库 | wasm32 + simd + threads + simd-threads |

新增RMSNorm层和对应的pnnx转换，单元测试
x86 convolution tiled gemm优化
量化工具支持 rnn/lstm/gru 动态量化
x86 lstm int8 sse2/xop/avx2/avx512/avx512vnni/avxvnni优化
arm rnn/lstm/gru int8 neon/asimdhp/asimddp优化
multiheadattention支持qdim参数与embed_dim不同
multiheadattention支持scale参数
更新pybind11到2.12支持numpy2
添加wasi支持(quink-black)
添加x86/arm convolution/slice/concat oom单元测试
onnx2ncnn工具添加警告和推荐使用pnnx的信息输出(lll143653)
修复x86 avx512 vnni指令派发失效的问题
增强x86/arm计算内核在内存不足时的错误返回
仅在windows arm平台使用ruapu指令集探测
windows mingw编译时支持大小核和SMT探测
修复powerpc vsx计算abs可能的错误
修复arm vfpv4条件下可能的fp16s/bf16s同时启用的冲突
修复aarch64架构l2-cache很小时因gemm K分块可能的越界读错误
修复riscv v tanh计算错误(zhangyang2057)
arm/convolution_3x3_pack1to8_fp16s使用ldr/str替代ld1/st1优化(quink-black)
修复c_api无参数函数声明(quink-black)
c_api添加set_vulkan_device接口(Baiyuetribe)
pyncnn添加从python bytes内存加载模型的接口(joeyballentine)
为VkAndroidHardwareBufferImageAllocator添加NCNN_PLATFORM_API宏(Xyzhao1999)
修复mingw64编译时avx崩溃和termux编译错误(TianZerL)
修复在关闭NCNN_BF16时arm riscv编译错误
修复x86-wsl编译时的无用变量警告(Tabbleman)
create_gpu_instance()中不进行destroy_gpu_instance()(Asd-g)
更新ruapu.h(lazyparser)
修复ndk-r27在cmake阶段的编译错误(Galasnow)
添加yolov8示例代码(whyb)
pnnx支持转换dynamo导出的onnx
pnnx默认编译onnx2pnnx支持，支持转换conv/convtranspose/pad/linear/softmax/relu/resize/upsample/avgpool/maxpool/batchnorm/lrn/layernorm/instancenorm/groupnorm/rnn/lstm/gru/prelu/gelu/elu/leakyrelu/relu6/celu/hardshrink/hardsigmoid/hardswish/clip/multiheadattention/reducemin/reducemax/reducemean/reducesum/reduceprod/logsoftmax/logsigmoid/mish/selu/sigmoid/silu/softmin/softplus/softshrink/softsign/tanh/tanhshrink/expand/permute/repeat/reshape/select/slice/cat/ceil/chunk/flatten/floor/maximum/minimum/split/squeeze/stack/transpose/unbind/unsqueeze
pnnx支持转换onnx指定inputshape
pnnx转换onnx遇到动态shape时尝试折叠非动态轴相关的常量
pnnx转换onnx合并简单的shape运算pattern
pnnx清除onnx中无用的cast
pnnx接受bf16的模型转换和输入输出类型
pnnx转换torch.tile/torch.where/torch.logaddexp
pnnx转换无dilation参数的F.maxpool到ncnn
pnnx转换1到2个轴参数的torch.roll到ncnn
pnnx转换有dim参数的torch.max/torch.min时返回tuple并自动删除没有用到的indice输出
pnnx合并onnx sdpa和qdim mha
pnnx识别sdpa的batch轴
pnnx支持torch-2.3和torch-2.4
pnnx不再折叠有就地操作的别名tensor为常量
pnnx转换到的ncnn模型py自动替换long为int
ci添加windows clang
ci添加harmonyos
ci添加mingw(TianZerL)
ci添加esp32和esp32编译文档(luxincn)
重构release ci脚本