enable CUDA managed pointers

[KrisThielemans]

To avoid copying data to/from GPU all the time, CUDA managed pointers might come to the rescue. This would become more attractive once supported by parallelproj (WIP by @gschramm ). A possible strategy would then be

1. change these to CUDA managed pointers to avoid them being copied every time.
   

   STIR/src/include/stir/recon_buildblock/Parallelproj_projector/ParallelprojHelper.h

   Lines 50 to 51 in b67c9f4

   	std::vector<float> xstart;
   	std::vector<float> xend;

2. somehow add optional argument to Array etc to allocate with CUDA managed pointer (could probably be done by templating the allocator/destructor)
3. make images by default CUDA-managed
4. add CUDA numerical operations (either as member or global)
5. replace some iterator-based loops for numerical operations of images with numerical operations
6. investigate what happens with projection data.

[KrisThielemans]

Links from @casperdcl

Slides from 2025-04-07 hackathon
NumCu: minimal Python/C++/CUDA library which uses CuVec's CPython buffer protocol
gschramm/square_array: misc experiments with Python ctypes & CUDA
__cuda_array_interface__: standard for exposing memory pointer addresses in Python
__dlpack__: same as __cuda_array_interface__ but also handling ownership & multiple devices/streams
DLPack reference implementations
C++17/20 & CUDA stdpar: even has a saxpy example
C++26 execution control

shared_ptr and CUDA

https://forums.developer.nvidia.com/t/can-we-use-cudamallocmanaged-to-allocate-memory-for-smart-pointer/283092/2
Comment from @wjp:

Something to be aware of with std::allocate_shared is that it will allocate the metadata (ref count and such) in the same block as the managed object, which could put the metadata in CUDA managed memory. That would potentially make moving the shared_ptr around slow. See the Notes from https://en.cppreference.com/w/cpp/memory/shared_ptr/allocate_shared .

[gschramm]

I have release the first release candidate of libparallelproj v2.0.0
https://github.com/KUL-recon-lab/libparallelproj/releases/tag/v2.0.0-rc1

official documentation of the C-API is here:
https://libparallelproj.readthedocs.io/en/latest/c_api.html

See my post in the discord STIR channel

[KrisThielemans]

@Dimitra-Kyriakopoulou proposes some small helpers to wrap CuVec to stir::Array in Dimitra-Kyriakopoulou@3b1ea8f. It uses the following constructor

STIR/src/include/stir/Array.h

Line 151 in 46da4ef

inline Array(const IndexRange<num_dimensions>& range, shared_ptr<elemT[]> data_sptr);

This approach seems quite interesting.

See also #1689 (comment)

@casperdcl any comments?

[casperdcl]

@Dimitra-Kyriakopoulou proposes some small helpers to wrap CuVec to stir::Array

lgtm