search

EPCCed / gpu-directives

Contains material for a course using GPU directives
0 stars 0 forks source link

Streams with openmp #3

Open lucaparisi91 opened 2 months ago

lucaparisi91 commented 2 months ago

It is not possible to have two teams construct in the same target region.

However it is possible to launch two different kernels from two different CPU threads and those would end up in the same stream. I have checked launching two different openmp parallel loops offloaded to GPU from two different parallel sections. The generated kernels ended up running cuncurrently on two different streams. I used the nvidia compiler from hpc-sdk 24.5.

image

lucaparisi91 commented 1 month ago

interop will give more control on different cuda streams, but does not seem to be supported on either nvidia nvhpc 24.5 or clang 18.1.8 .