What is the recommended way to substitute default_construct_range in Thrust 2.8

[correaa]

I upgraded to CUDA 12.9 and I see that Thrust doesn't have default_construct_range.
What is the recommended way to replace it?

https://gitlab.com/correaa/boost-multi/-/jobs/10277394080#L406

[bernhardmgruber]

We renamed this member to value_initialize_range in #1990. Note that this is an internal entity (it's inside thrust::detail) and we reserve the right to break those at any time. You can write this functionality yourself using e.g. thrust::uninitialized_fill_n or thrust::for_each_n.

[bernhardmgruber]

Alright, then this should do what you need:

thrust::for_each_n(first, n, [](auto& elem) { new (&elem) T; });

[correaa]

well, I have pass the allocator as well.

The full thing is

cccl/thrust/thrust/detail/allocator/value_initialize_range.inl

Line 78 in 87f7246

thrust::for_each_n(allocator_system<Allocator>::get(a), p, n, construct1_via_allocator<Allocator>(a));

as you pointed out earlier.

with the caveat that, if don't rely on ::detail namespace, I will have to replicate a bit of more code (that, or simply rely on detail::value_initialize_range)

[bernhardmgruber]

So, lets pass the allocator to the lambda then and use what construct1_via_allocator does:

thrust::for_each_n(allocator_system<Allocator>::get(a), first, n, [a](auto& x) { thrust::detail::allocator_traits<Allocator>::construct(a, &x); });

And then if you look up what thrust::detail::allocator_traits<Allocator>::construct(a, &x); does, it should just boil down to:

  ::new (static_cast<void*>(&x)) T();

See:

cccl/thrust/thrust/detail/allocator/allocator_traits.inl

Line 176 in 1c9f6c4

::new (static_cast<void*>(p)) T();

And if you want to skip initialization, then:

  ::new (static_cast<void*>(&x)) T;

[correaa]

mmm, now I worry that the lambda must be __host__ __device__.

In any case, this makes me realize that thurst::detail::...::construct ignores the allocator argument, I though that such thing would be use as part of the scoped allocator protocol, so I was wary to jump to ::new too early. But I guess that since there is no no-detail construct, I will have to be content with that anyway.

[bernhardmgruber]

In any case, this makes me realize that thurst::detail::...::construct ignores the allocator argument

Thrust checks whether the allocator has a construct member function, but AFAIK no Thrust allocator has one. If that is important to you, you can implement this of course.