CUDA-gcc incompatibility workaround

[Dubslow]

So I have a new OS install, with gcc 4.7.2, which as yet is unsupported by CUDA 5. I have no intentions of installing a second compiler, so does anyone know of another work around?

I suppose all I really need is a way to compile mfaktc, CUDALucas, and Msieve (or someone to do it for me).

PS Does anyone know of a free/libre replacement for nvcc? (And I suppose you'd need to duplicate the libs as well )

[Ken_g6]

See if your package manager has a CUDA dev package. Mine does, but it requires uninstalling the proprietary driver, which means I couldn't play games. But if you don't want to play games, it might work for you.

Edit: P.S. I do plan to install a second compiler - in a virtual machine.

[Dubslow]

Bleh.

On the other hand, based on something I read a few hours ago, is it possible to mount my old installation (as I have already done) and then chroot into it? Does chroot change the entire environment, so that I would be using the new kernel with the old file system and programs? I think this bears research, and any comments/tips/advice are greatly appreciated.

[owftheevil]

Its a bit flaky, but this worked for me, at least to get CUDALucus built on Ubuntu12.10.

Copy all of cuda-5.0 to /usr/local/ or wherever you want to put it, make a link cuda linking to cuda 5.0 in the same place, then in cuda5.0/include/ edit line 80 of host_config.h, changing the 6 to a 7.

gcc4.7 really isn't supported by cuda5.0, so some things won't build. CUDALucas does and seems to work ok, but I haven't tried mfaktc. Some of the examples in the SDK do not work.

Now a warning. On my box, with any 300+ driver on Ubuntu12.04 or 12.10, CUDALucas2.04 and later occasionally hang on the cudaMemcpy in lucas_square. Toolkit version 4.2 or 5.0, built with gcc4.6 or 4.7 don't make any difference.

[chalsall]

Originally Posted by Dubslow

Does chroot change the entire environment, so that I would be using the new kernel with the old file system and programs? I think this bears research, and any comments/tips/advice are greatly appreciated.

chroot changes the root of the file system. So the "host" kernel will be running the chroot'ed programs. Keep in mind that any dynamically linked libraries are pulled from the chroot'ed file system as well.

[Dubslow]

Originally Posted by owftheevil

Its a bit flaky, but this worked for me, at least to get CUDALucus built on Ubuntu12.10.

Copy all of cuda-5.0 to /usr/local/ or wherever you want to put it, make a link cuda linking to cuda 5.0 in the same place, then in cuda5.0/include/ edit line 80 of host_config.h, changing the 6 to a 7.

gcc4.7 really isn't supported by cuda5.0, so some things won't build. CUDALucas does and seems to work ok, but I haven't tried mfaktc. Some of the examples in the SDK do not work.

Now a warning. On my box, with any 300+ driver on Ubuntu12.04 or 12.10, CUDALucas2.04 and later occasionally hang on the cudaMemcpy in lucas_square. Toolkit version 4.2 or 5.0, built with gcc4.6 or 4.7 don't make any difference.

Well, the 5.0 installer wouldn't even install properly, so I don't have anything in the cuda-5.0 folder.

I was however able to copy my 4.<something> environment over to the new system with gcc 4.7, edit out the the thingy in the header file and compile CUDALucas; I was not, however, able to test it, due to driver issues. Hopefully that will be resolved soon.

Originally Posted by chalsall

chroot changes the root of the file system. So the "host" kernel will be running the chroot'ed programs. Keep in mind that any dynamically linked libraries are pulled from the chroot'ed file system as well.

Okay, that sounds about like what I had described. I might give it a go then.