CUB and CUDA version

[Mathias Wagner]

I love to use CUB for not having to write my own reduction code etc. However the code has to run on various cluster and though I don't have control over the CUDA version installed.

I did not find anything obvious but CUB 1.4.1 fails to compile with CUDA 6.0 (which I have to support because the driver on the cluster cannot handle anything newer).

Is there a workaround for that?

The error I get when trying to compile is related to the occupancy api which was only introduce in 6.5 

cub/block/../iterator/../util_device.cuh(225): error: identifier "cudaOccupancyMaxActiveBlocksPerMultiprocessor" is undefined

[Nicolas Höft]

Yes, in cub/util_device.cuh you could simply comment the call to cudaOccupancyMaxActiveBlocksPerMultiprocessor() and uncomment the next block doing this calculation manually.

Nicolas