We are going to Abandon Cuda without Mingw support on windows

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey there. As an NVidia Partner our company has been developing an AI product for use with Cuda. At the moment however we are not happy with the current support situation with Cuda development on windows.

Seems that nvidia is not supporting the Mingw toolchain to build code on windows. This is basically unacceptable and despite weeks of man hours trying to work around the issue we have not had any success.

Our product previously was only supporting Linux/Mac and is developed with C++/Cmake. We were able to port to windows using mingw and managed to get our complicated application to build just fine on windows with that toolchain.

Now that we have added Cuda to our product, it seems this complicates things because our entire toolchain is basically not supported by nvidia.

I spent weeks attempting to use visual studio as nvidia requires for Cuda only to fail over and over. It’s to the point now that I am second-guessing my thought to support Cuda on windows in the first place.

I think I am going to have to tell my customers that we will only support the competing OpenCL on windows due to nvidia’s unfaltering insistence on using visual studio as the build toolchain which basically is a massive failure.

We have tried using gvsbuild to get parts of our gtk/cairo/gtkmm working but it fails to build other dependencies. We tried using vcpkg to install our dependencies but it lacks such basic features as specifying the version of a package to install and they just moved to gtk4 making it hard to install gtk3 packages…

Mixing/Matching pkg-config from those toolsets doesn’t seem to work. CMake lacks support for finding things built this way…

Basically the way we were building with mingw was working fine. We are only doing this because despite supporting unix gcc-based toolchains on arm for instance, you just won’t support mingw on windows which is confusing.

So I spent the time to create an account and this post to basically say I have had enough and I’m going to just blame Nvidia when my customers ask why only Linux/Mac platforms support Cuda. We spent time to become an Inception Partner but everytime we make a suggestion they just tell us to post on the forum…

So this is that post. Support Mingw or we move on.

1 Like
2

You might be interested in WSL2 CUDA support. Another method to get your voice heard is to file a bug. This would essentially be an RFE or “feature request”. The bug system is appropriate for that usage.

3

I am not authorized to file a bug. The link to join the developer program also says I am not authorized.

What are the next steps?

4

The link has been updated.

Please join the developer program, then you will be able to file a bug. For example click here:

https://developer.nvidia.com/

Then you will see a “join” green button in the upper right hand corner. Click that. Follow the steps to join the developer program. (I’m actually not sure how you are posting here if you are not already part of the developer program.)

After you have joined the developer program, you will be able to file a bug.

1 Like
5

  1. You need to be quite optimistic to believe filing a bug about this would make it happen; I wouldn’t hold my breath.

  2. I’d still consider OpenCL, despite it being a tragic, doomed, standard.

  3. You wrote that:

    It’s to the point now that I am second-guessing my thought to support CUDA on windows in the first place.

    well, if it’s a matter of choice - why support Windows at all? It’s a dead-end closed-source operating system. I’d ditch Windows support and make my life easier.

  4. Just to make it clear - it is quite inappropriate and anti-social of NVIDIA to keep its basic toolchain closed-source, so people can’t port it to other platforms, while also not adapting/porting it themselves to all popular platforms - like mingw on Windows.