This Library isn't Compatible with Windows about cuda

[tmcdonell]

ld and nm it should find from your mingw install (which ghc is using). Do the notes in the windows read me help at all, or are they just completely outdated now?

You are right, I don’t regularly test on windows, so I rely on people sending feedback/patches when things break.

[MilesLitteral]

Yes, I believe the instructions are currently outdated, I've made hotfixes which make the library build on Windows, I can open a pull request if you'd like to maybe address the issue fully?

Oh Mingw, hmmm my Haskell build on Win64 is using Msys but I doubt thats the full issue??? All in all it will take some ground work which I don't mind helping out with at all

The issue seems to be that CUDA is difficult to link with the MinGW environment unless one builds the CUDA libraries with MinGW's GCC (or conversely MSYS's gcc) and manually installs them; yet this is equally difficult based on my research

https://forums.developer.nvidia.com/t/cuda-with-mingw-how-to-get-cuda-running-under-mingw/415/7

[tmcdonell]

If you do get it working yes please open a PR, that would be fantastic!

Windows support was added a long time ago via a contributor, but I believe they no longer need it thus it hasn't been kept up to date ):

[MilesLitteral]

Hey @tmcdonell I have made progress with the Library and can report the following:

The issue seems two fold:
First, the function defaultCUDAInstallPath (in Setup.exe) was acting wonky, in my branch I extrapolated the function to have platform specific behavior (it works very well)

NOTE: even though the Setup.hs finds the CUDA_PATH it seems to default to this function no matter what

Second the generated cuda-0.11.0.0-inplace.conf gives the wrong path for the dlls:

library-dirs:
    ...\cuda-master\dist-newstyle\build\x86_64-windows\ghc-8.10.7\cuda-0.11.0.0\build
    C:\NVIDIA\Toolkit\CUDA\v9.0\lib\x64 C:\ghcup\msys64\mingw64\lib

dynamic-library-dirs:
    ...\cuda-master\dist-newstyle\build\x86_64-windows\ghc-8.10.7\cuda-0.11.0.0\build
    C:\NVIDIA\Toolkit\CUDA\v9.0\lib\x64 C:\ghcup\msys64\mingw64\lib

it seems to think the .lib and .dll files are in the same place based on the inplace, when you put the .dll it's looking for in the same place as the lib path and update the defautCUDAInstallPath function this library builds again on Windows no problem

I will open a pull request soon, if you know what function is setting dynamic-library-dirs I'd appreciate the pointer :)