Just wanted to post this in case any others didn’t notice what happens during an install.
To review, I believe the installation instructions [for most platforms] are usually broken up into 3 steps:
-
Install the appropriate CUDA “driver” (for example, “cudadriver_3.0.1-beta1_macos.pkg”).
-
Install the appropriate CUDA “toolkit” (for example, “cudatoolkit_3.0-beta1_macos_32.pkg”)
-
Install the appropriate CUDA/GPU computing SDK and examples (for example, “gpucomputingsdk_3.0-beta1_macos_32.pkg”)
[In OS X the order is different: do step 2 before doing step 1] These steps should actually be:
-
Install CUDA driver
-
Install CUDA toolkit
-
Re-install CUDA driver from step #1
-
Install CUDA SDK
[EDIT: or simply the following (taken from the ReadMe)]
-
Install CUDA toolkit
-
Install CUDA driver
-
Install CUDA SDK
The reason for this is that, while in step 1, the following file is installed: “/usr/local/cuda/lib/libcuda.dylib”, step 2’s installer package simply overwrites any existing /usr/local/cuda/lib/ folder with its own, which doesn’t include libcuda.dylib.
Hope this helps…
[EDIT] Okay, I’m an idiot… I just read the read me (/Developer/GPU Computing/doc/CUDA_SDK_release_notes.txt) and, at least for the 3.0 version, the installation instructions I gave above are incorrect. The installation order it gives will prevent the problem I described above.