All curand() samples fail with GTX 1080 on CUDA 7.5

Other samples compile and run OK with VS2013 Update 5. However MersenneTwister and other curand samples fail with CUDA 7.5. They work fine with CUDA 8.0 but throw CURAND_STATUS_LAUNCH_FAILURE for 7.5. We tested with multiple GTX 1080 cards on multiple machines. Using both 368.25 and 368.39. There is currently no driver available for GTX 1080 cards that can run curand samples on CUDA 7.5

I confirm that neither curand64_75.dll (windows) nor libcurand.so.7.5.18 (linux) contain PTX, therefore they will only work for the architectural generations for which they have SASS (2.x, 3.x, 5.x)

By contrast, the corresponding cublas libraries do have PTX embedded.

It appears that this oversight (not including PTX) has been addressed in CUDA 8RC. The libcurand.so.8.0.27 contains PTX.

The above observations can be confirmed with the cuobjdump tool.

The workaround therefore on GTX 1080 or any pascal device is to use CUDA 8RC or CUDA 8 production (or later) when it becomes available.