Windows support

[GunnarFarneback]

It would be nice if this would run on Windows as well as Linux. I assume the biggest hurdle is setting up the cuda environment in https://github.com/JuliaPackaging/Yggdrasil/blob/master/T/Torch/build_tarballs.jl.

[GunnarFarneback]

Work in progress on Yggdrasil: https://github.com/JuliaPackaging/Yggdrasil/pull/1529.

[bjosv79]

Suggested changes to the api in #37. Built with MSVC 2019 by adding the CMake entry: CMAKE_PREFIX_PATH = C:/Users/UserName/AppData/Local/Programs/Torch/libtorch

[DrKrar]

An thing new

[stemann]

Windows support (on CPU) would require:

• Building the C wrapper for Windows, which would likely require:
  • Building Torch_jll from source for Windows.
  • Cf.
    • https://github.com/JuliaPackaging/Yggdrasil/pull/4477
    • https://github.com/JuliaPackaging/Yggdrasil/pull/4554
  • Building ONNX_jll from source for Windows.
    • Cf.
    • ONNX does not support cross-compiling for w64-mingw32 on linux - but more recent versions than ONNX v1.10 may be easier: https://github.com/conan-io/conan-center-index/pull/25469
    • https://github.com/JuliaPackaging/Yggdrasil/pull/3839
    • https://github.com/JuliaPackaging/Yggdrasil/pull/4587

Windows support with CUDA would require:

• Either being able to build the C wrapper for Windows while using pre-built libtorch binaries (similar to approach in https://github.com/JuliaPackaging/Yggdrasil/pull/1529).
• Being able to build a CUDA-project for Windows using BinaryBuilder (not possible currently - or in any foreseeable future as nvcc is not a cross-compiler)