consistent size and index types about gtensor

[bd4]

FWIW, the SYCL spec landed on using size_t for all index types, and int for dimensions: https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#subsec:item.class

The fact that the gpu blas APIs typically use int indexes makes things fun here. Bottom line, we don't want silent failures if overflow happens, even in release builds.

[germasch]

I guess I'll agree that while there was some intention behind the different size for multi-d index vs 1-d index, it's really not possible to not do this in a way that doesn't cause potential subtle overflow bugs.

So probably what we should do is to use <class>::size_type consistently everywhere, get rid of int, with the former being defaulted to gt::size_type which in itself should probably default to std::size_t (but one could argue about uint).

Maybe actually ssize_t would be better, since there are times where negative indices may make some sense. I think this need some more thinking/investigation...

[bd4]

yes this is subtle. As a start, I am using size_t for index calculations inside kernels and for SYCL and whenever doing interations over container.size() or calc_size(shape). This involves no change to the external interface.

[bd4]

There are potentially 3 types at play here:

1. size_type - for addressing in to memory
2. index_type, backend - what the underlying GPU lib uses, e.g. for dim3 in CUDA (uint32_t)
3. index_type, gt high level - currently int, as part of shape_type, not convinced this should be different from (2)

I'm inclined to make gt::index_type a standard that should be used e.g. for GT_LAMBDA (gt::index_type i), and then we have knobs to change this and aren't stuck with a bunch of code with hard coded ints. I am also inclined to change the default to uint32_t on HIP/CUDA and to size_t for SYCL, to match the backend.

[bd4]

Having both size_type and index_type is a bit confusing here. One is 1d, the other is N-d, but of course 1d is just a special case of N-d, so it's wierd.

[germasch]

Having both size_type and index_type is a bit confusing here. One is 1d, the other is N-d, but of course 1d is just a special case of N-d, so it's wierd.

Right, I think I just wrote something along those same lines in another thread.

[bd4]

I think an interesting experiment here, would be to change all indexes to size_type and change default size_type to uint32_t, and see if (a) tests, benchmarks etc pass, (b) gene passes tests, and (c) performance of benchmarks and/or GENE improves significantly. If the answer is yet, then I think it's worth doing. Other than some reverse iteration direction loops in bandsolver which can be rewritten, I don't see negative indexes being used anywhere. It would also be interesting to see if using uint32_t consistently works around the ROCm compiler bug with -O2.