Replace thrust::complex by libcu++

[bernhardmgruber]

No description provided.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[bernhardmgruber]

Discussed with @miscco, apparently something similar was tried with #260 but abandoned. Later #454 tried to replace thrust complex partially, but it was reverted in #1497, because the numerical results did not work out. It seems cuda::std::complex still needs fixes.

[miscco]

The issue is more with the math functions than with anything else

[davebayer]

I think the biggest problem is that thrust::complex is trivially default constructible, but cuda::std::complex is not.

The replacement could lead to some unexpected behaviour. An example:

thrust::complex<float> alloc(size_t n)
{
  return new thrust::complex<float>[n];
}

The old implementation does not zero initialize the allocated array, but when implemented by cuda::std::complex, it does.

This can have some serious performance consequences, e. g. for NUMA systems where the memory will be first touched by a single thread.

[miscco]

I think the biggest problem is that thrust::complex is trivially default constructible, but cuda::std::complex is not.

That is actually a quite important problem, as certain libraries require trivially_copyable types, for example rapids device_uvector

The long term solution that we want is a cuda::complex which effectively has all the functionality of cuda::std::complex but is just a plain aggregate. The issue there is that we would then need to duplicate all math functions because they need to take a cuda::std::complex and return a cuda::std::complex and cannot rely on e.g implicit conversion through inheritance