What arguments should be used for CMake run / flags enabled? (please share some best known configuration)

I used this CMake option to build and execute.

cmake -G "Visual Studio 16 2019" -A ARM64 -DCMAKE_SYSTEM_NAME=Windows -DCMAKE_SYSTEM_VERSION=10.0 \
      -DCMAKE_SYSTEM_PROCESSOR=ARM64 -DWITH_OPENCL=OFF -DWITH_FFMPEG=OFF -DWITH_CUDA=OFF \
      -DBUILD_EXAMPLES=ON -DBUILD_TESTS=ON C:\opencv

Thank you for update!

What are arguments used for vcvarsall.cmd call?

Did you see these lines from CMake?

-- Performing Test HAVE_CPU_NEON_SUPPORT (check file: cmake/checks/cpu_neon.cpp)
-- Performing Test HAVE_CPU_NEON_SUPPORT - Failed
-- NEON is not supported by C++ compiler

If so, please try to fix check (ARM code without NEON optimizations is slow).

Thank you for review.

What are arguments used for vcvarsall.cmd call?

Use x64_arm64 argument

I'm trying to enable NEON, but the build doesn't pass.
Even I try to build with ARM (32bit) for UWP, an error will occur in the same part, so it seems to be another problem.

Another problem here is that CMake is not running in cross-compilation mode (CMAKE_CROSSCOMPILING is not set).
Perhaps we need to prepare Windows ARM64 CMake toolchain file. Like this, but under platforms/win-arm64 path.

I tried to fix check_neon.cpp but got this error:

fatal error C1189: #error: This header is specific to ARM targets

However /p:Platform=ARM64 parameter is used and compiler is "for ARM64".

Run Build Command(s):C:/Program Files (x86)/Microsoft Visual Studio/2019/BuildTools/MSBuild/Current/Bin/MSBuild.exe cmTC_928d6.vcxproj /p:Configuration=Debug /p:Platform=ARM64 /p:VisualStudioVersion=16.0 /v:m 
Microsoft (R) Build Engine version 16.0.461+g6ff56ef63c for .NET Framework
Copyright (C) Microsoft Corporation. All rights reserved.

  Microsoft (R) C/C++ Optimizing Compiler Version 19.20.27508.1 for ARM64
  Copyright (C) Microsoft Corporation.  All rights reserved.
  
  cpu_neon.cpp
  cl /c /Zi /W3 /WX- /diagnostics:classic /Od /Ob0 /Oy- /D WIN32 /D _WINDOWS /D "CMAKE_INTDIR=\"Debug\"" /D _ARM64_WINAPI_PARTITION_DESKTOP_SDK_AVAILABLE=1 /D _MBCS /Gm- /MDd /GS /fp:precise /Zc:wchar_t /Zc:forScope /Zc:inline /Fo"cmTC_928d6.dir\Debug\\" /Fd"cmTC_928d6.dir\Debug\vc142.pdb" /Gd /TP /analyze- /errorReport:queue E:\projects\opencv\cmake\checks\cpu_neon.cpp
  
C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools\VC\Tools\MSVC\14.20.27508\include\arm_neon.h(17): fatal error C1189: #error:  This header is specific to ARM targets [E:\build\opencv_arm64\CMakeFiles\CMakeTmp\cmTC_928d6.vcxproj]

Thanks you!

Attach CMakeVars.txt when the cmake command is executed. CMAKE_CROSSCOMPILING was set.
https://gist.github.com/shibayan/eec4209524d208cdfdea6a4eae45f949
(CMAKE_CROSSCOMPILING seems to be enabled by setting CMAKE_SYSTEM_NAME and CMAKE_SYSTEM_VERSION.)

Did you encounter a problem with a previously shared cmake command?

It was a build error when NEON was enabled, but I pushed the branch I worked on halfway.
https://github.com/shibayan/opencv/tree/arm64-neon-windows10

If you build after running cmake, you should get an error in intrin_neon.hpp.

C:\opencv\modules\core\include\opencv2/core/hal/intrin_neon.hpp(108,1): 
error C2084: function 'void cv::hal_baseline::_v128_unzip(const uint8x16_t &,const uint8x16_t &,uint8x16_t &,uint8x16_t &)' already has a body [C:\opencv\build-neon\modules\core\opencv_core.vcxproj]

I added "custom" builder with ARM64 configuration. Please take a look.

Consider pushing changes with updates on source branch of this PR (arm64-windows10)

Pushed NEON changes from branch "arm64-neon-windows10" here.

This header is specific to ARM targets

Resolved by upgrading MSVS 2019.

function 'void cv::hal_baseline::_v128_unzip(const uint8x16_t &,const uint8x16_t &,uint8x16_t &,uint8x16_t &)' already has a body

Looks like a MSVC compiler bug which aliased "int8x16_t" and "uint8x16_t" and don't accept overloads.
MSVC versions: 19.23.28107.0 and 19.24.28314.0

This simple code:

void test_aliased_type(const uint8x16_t& a) { }
void test_aliased_type(const int8x16_t& a) { }

Emits these errors:

intrin_neon.hpp(71,44): error C2084: function 'void cv::hal_baseline::test_aliased_type(const uint8x16_t &)' already has a body
intrin_neon.hpp(70): message : see previous definition of 'test_aliased_type'

Any ideas how to workaround that? (without massive code refactoring)

In MSVC, uint8x16_t and int8x16_t are defined as aliases of __n128, whereas GCC is an alias of __builtin_neon_uqi and __builtin_neon_qi, so it seems that overloading can be defined.

I found same issue: https://developercommunity.visualstudio.com/content/problem/335464/arm64-vector-intrinsics-typedefs-float32x4-t-and-i.html

I think this is a very difficult problem to solve. At least I couldn't find a workaround...

Bug fixes on the Visual Studio side will take time. The performance will drop because NEON cannot be used, but I would like to make the build pass for ARM / ARM64.
(I will continue to communicate with the Visual C++ team.)