Why can't nvcc find my Visual C++ installation?

Question

I'm running Windows 7 Pro x64 on a Core i5 with a NVIDIA 3100m, which is CUDA compatible.

I've tried installing both the 32-bit and 64-bit CUDA toolkits from NVIDIA, unfortunately from with either of them I cannot compile anything; nvcc says "cannot find a supported cl version. Only MSVC 8.0 and MSVC 9.0 are supported".

I have the x86 and x86-64 compilers installed via the Windows 7 SDK (compiler version 15.00.30729.01 for both arches). Both compilers are operating correctly; I've built and tested C and C++ code using them. I've tried running nvcc from command shells set up for both 32 bit and 64 bit compilation, and using the -ccbin command line option to nvcc to point it at the Visual C++ install directory.

What is the right way of handling this setup? Is there some way I make nvcc be more verbose about what is going on? The -v flag isn't terrible helpful. Ideally some way to make it show what it is finding versus what it's expecting to find. Will this work better if I install Visual C++ Express instead? Or is only a commercial version of VC++ supported for use with CUDA?

Answer 1

Include this line

compiler-bindir = C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\bin

in

C:\CUDA\bin64\nvcc.profile

Answer 2

With CUDA v6.0, having an older version of Visual Studio (I'm using 2010) is still required, but the other solutions did not work for me.

To make this work, add the following to your nvcc.profile (C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v6.0/bin/nvcc.profile):

CUDA_NVCC_FLAGS += --compiler-bindir = "-IE:/PROGRA~2/MICROS~2.0/VC/bin"

I also have VS 2013 installed, and I needed this argument to prevent nvcc from finding the newer (sadly incompatible) version of cl.exe