U++ forum: Welcome to the forum

Re: Experimental CUDA support [message #61460 is a reply to message #61459]

Sun, 16 February 2025 09:54

mirek
Messages: 14256
Registered: November 2005

Ultimate Member

zsolt wrote on Sat, 15 February 2025 23:08

I'm absolutely not an expert, but AFAIK nvidia-cuda-toolkit on Ubuntu can be used in Clang or G++ projects.

In Linux yes. In Win32, no way - you need to use Visual C++. At least that is Nvidia's position. https://stackoverflow.com/questions/68726422/is-it-possible- to-integrate-nvcc-with-gcc-g-in-windows

Now nvcc seems to work by separating cuda code and host code, then compiles cuda code and creates a new c++ file which combines host code and compiled cuda code (as arrays like

unsigned long long fatbinData[]= {0x00100001ba55ed50ull ....
)

Then this file is compiled by host compiler. There are explicit checks in CUDA headers for MSC++ compiler (#ifdef MSC_VER). There might also be a problem with CUDA runtime library, but I doubt that...

I believe that if I take clang-cl and give that to nvcc, with some additional trucks like -D MSC_VER it should work. Do not have time right nor energy now to test right now...

Report message to a moderator