CUDA Fortran Equivalent to nvcc --default-stream per-thread

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I’m trying to replicate the results shown here in Fortran 90.

However, I haven’t found an option equivalent to “–default-stream per-thread” for the PGF90 compiler.

Does anyone have any ideas on how i can get the default stream per thread behavior, and the non-blocking default stream? Thanks.

2

Hi Loregard,

No, sorry. CUDA Fortran does not have an equivalent flag. There is an API Call you can use, “cudaSetStreamDefault” to change the default stream, but this is not on a per-thread basis.

See: CUDA Fortran Programming Guide Version 19.7 for x86 and NVIDIA Processors

Note that for OpenMP programming (which would be the most equivalent method to Mark’s pthread example), each OpenMP thread does use different default streams by default.

-Mat