add Array template index and use long long for ProjDataInMemory

[KrisThielemans]

Drastic revision that adds a indexT to VectorWithOffset, Array and IndexRange to be able to use something else than int.
Then uses that to allow for more bins in ProjDataInMemory.

fixes #1505

[KrisThielemans]

C++ tests are fine. Weird things in the Python tests.

@z-k-li in principle you could check this with SIRF and "in memory" acquisition data. The STIR Python stuff will likely fail as per the tests.

[KrisThielemans]

The tests currently fail for unsigned indices (although I thought I had fixed that). However, this is currently not uesed, except in thest. I also still expect problems with SWIG as above.

@ChristianHinge @z-k-li could you give this a go with parallelproj (either CPU or GPU). It doesn't have #1674 yet, but that fix doesn't affect parallelproj. (You'd have to disable building of STIR Python.)

For GPU, no doubt you will have to change

STIR/examples/samples/forward_projector_parallelproj.par

Lines 6 to 8 in 25b3656

	; this keyword allows increasing the number of chunks that the projector uses
	; increase if you run out of GPU memory
	num_gpu_chunks:=1

[ChristianHinge]

@KrisThielemans I will try this out next week since we are just finalizing the MICCAI submission :-)! So just make sure I understand correctly - I will recompile the PR without python bindings and benchmark RAM + clock against the master branch. Is that correct?

[KrisThielemans]

That's correct, but of course after installing CUDA drivers/toolkit, and switching the projector to parallelproj.

[ChristianHinge]

@KrisThielemans I ran the recon using parallelproj using 4i5s, zoom 0.5, segment 4 (same recon everything as the benchmark I ran for #1674, but using parallelproj)

Time: 118min (TOF CPU: 16 min)
RAM: 167GB (TOF CPU: 19GB)

I ran it on 4xA40 GPUs. When monitoring, nvidia-smi, I can see the work being offloaded to the GPUs, albeit the utilization is quite low since the GPU is iddle most of the time. The resulting images look virtually identical to the master branch, but the voxel values differ a tiny bit.

[KrisThielemans]

Thanks @ChristianHinge at least we know it works now (it used to crash), which is great. I do expect differences between parallelproj (Joseph projector) and the ray-tracing matrix (Siddon with a few lines), but images should look overall quite similar.

Indeed, most of the computation time sits in copying and reordering data. The GPU probably flies through it all. This will be work for our imminent hackathons.

[ChristianHinge]

It is awesome to see the recon work on GPU! And quite cool that it distributes the workload evenly between the four A40s.