Some details here: I am running VS2005 on Vista x64, 181.22 64-bit driver, have the x64 toolkit and x64 SDK installed. Tried to compile the samples for Win32 target, VS2005 complained it did not have the cutil32d/cutil32 libs, so I went ahead and compiled the cutil solution for 32 bit target.
Even the template didn’t compile:
.\Release\template.obj : fatal error LNK1112: module machine type ‘x64’ conflicts with target machine type ‘X86’
Question:
Having a Vista x64 machine, does this mean I am stuck with compiling for x64 targets only???
If not, what are the options, recommendations here for dual targets (win32 and x64)? I guess I could install the win32 cuda sdk also, but since the toolkit is x64 this is not going to work (installing the win32 toolkit alongside the x64 toolkit is probably not a good idea… even if the env vars get set manually in VS2005).
Another thought: if I’m mostly interested in Win32 targets, can I not do the following:
uninstall the the x64 version of the toolkit (v2.1) and x64 SDK (x2.1), reinstall the respective win32 versions for vista, then using VS2005 to target win32 only.
Or do I need to clean registries etc before reinstalling the win32 versions, and what about the 181.22 x64 driver then (my assumption is the driver needs to stay x64). Or is this a dangerous thing to do? Nvidia folks, please a word of wisdom here, thanks!
I don’t think you can install both versions of the SDK, and I doubt you’d want to uninstall/reinstall/compile/uninstall/reinstall every time you wanted to compile a 32-bit version.
If you’re just compiling the 32-bit versions on your box and deploying them somewhere else, a quick (but somewhat convoluted) solution would be to use Virtual PC to run a virtual machine on your development machine with a 32-bit version of Windows/VS/CUDA SDK installed and just compile them on there (you don’t need actual CUDA hardware to compile). Or, you could just find another unused 32-bit windows machine with XP Pro or Vista Business/Ultimate on there and Remote Desktop to it when you want to compile.
Kind of a pain, I know, but I hear that the ability to target one platform or another (from either) is going to be included in one of the upcoming releases of CUDA (hopefully 2.2, but maybe the one after that).