I'm building (well, trying to build) ArrayFire 3.3.1 on a machine with no relevant GPU (but with CUDA 7.5). What I'm getting is that Compute Capabilities 20, 30 and 50 are enabled. But what if I'm targetting a device with is 35 (3.5) or 52 etc.?

I should be able to specify - say, in an environment variable - the target compute capability values I want. You'll note that in CUDA's samples Makefiles they use a string with the different SM values relevant; perhaps a CUDA_SM variable with a space-separated list of values? And an indication that instead of the default one can set this value to whatever one would like?