nollvm option on windows for curand_device

I am trying to compile a cuda fortran code under windows that uses curand_device.
In particular, the small code snippet on page 38 of
http://www.pgroup.com/doc/pgi17cudaint.pdf

Under linux it works perfectly by compiling it as
pgfortran -Mcuda=nollvm t.cuf

Under windows however I cannot get it working.
Other codes using host curand, cublas, etc. work under windows.

from the pgice prompt 17.4 with compile command
pgfortran -Mcuda=nollvm,cuda8.0 t.cuf
I get error messages:

PGF90-F-0155-Compiler failed to translate accelerator region (see -Minfo messages): Device compiler exited with error status code (t.cuf: 1)
PGF90/x86-64 Windows 17.4-0: compilation aborted

-Minfo does not provide any additional message.
without nollvm there are undefined references to ‘__pgicudalib…’
-Mcudalib=curand does not help

I am lost.

Hi abalogh,

This is a known issue on Windows and a problem with the front-end CUDA compiler. Unfortunately because of this, you won’t be able call cuRand from device code when building on Windows. The only workaround is to use Linux.

-Mat