I can't tell where the problem is. Why does the version of Thrust that I get with stdpar=gpu work, whereas the version from GitHub doesn't?

Thanks

MCVE

#include <thrust/universal_vector.h>

thrust::universal_vector<float> m_x ;

void AllocateNodePersistent(int numNode)
{
  m_x.resize(numNode);
}

It works with stdpar=gpu

$ nvc++ -std=c++17 -stdpar=gpu -c bug.

#969 introduced some support for C++ arrays. So far, we know this works for fields of structs. We should add more tests to find out what happens if we allow arrays as actual function parameters, etc. Pretty sure it won't work because the support for arrays in TypeConverter is incomplete. But, we should find out!

I would like to be able to determine my device-wide primitive's temp_storage_bytes before I have all of the primitive's arguments ready. The interface for obtaining it ostensibly requires everything to be ready for the actual run - and the documentation does not make it clear what arguments it actually needs and what it's going to do with them (e.g. will it look at the input at all, in any way? Pr

Introduced in #128.

The original reason for that is purely technical - it is required to call unshare(CLONE_NEWUSER) in the main thread of a program. Since cachepot-dist is multi-threaded and the build is executed on a new thread, we used a hack in which we fork() and call unshare in the forked child (fork() forks the calling thread into a main thread of a forked child process).

Us