stellar-group / hpx-kokkos Goto Github PK

The executor_test fails for me when building hpx-kokkos with CUDA and running ctest.

To be more specific, the test fails on line

I've checked, and it is during the tests with the cuda executor (calling the same test method with the default host executor later in the test works fine). From what I can see, the 42 in line is not copied to the device properly, causing passthrough in the lambda to be 0 instead of 42 (verified by a printf inside the lambda).

I quickly tested having a local passthrough2 = 42 variable (declared at the beginning of the test method) which is then captured by the lambda and correctly copied to the device. Using this instead of the argument passthrough makes the test pass (simply done to verify that this is the only thing going wrong in the test).

The underlying issue seems to be that the pass-through to the device fails (at least for my build).

Utilized versions:

• hpx-kokkos master
• HPX 1.6.0
• Kokkos 3.3.01
• nvcc (with gcc 9.3.0 and CUDA 11.2)