[SYCL][Matrix] Replace size_t with uint32_t to specify joint matrix rows and columns about llvm HOT 2 OPEN

The reason to use size_t for "joint_matrix" shape rows and cols and things like: stride and coordinates in the joint_matrix_apply, is to be consistent with what is currently using in SYCL spec (sycl::marray, sub_group_sizes) and C++ (a span extent type is "size_t").

How these are handled in SPIRV?
Adding @gmlueck to get his opinion on this.

from llvm.

Rows/Cols are explicitly set to be constant instruction with scalar 32-bit integer type, so right now if we want to ensure conformance of our implementation we need to check for 32-bit overflow and then cast size_t to uint32_t for these parameters.

Are we just talking about some extra validation in the LLVM-to-SPIRV translator? If so, this doesn't sound too onerous.

If using size_t was causing a performance slowdown, I would be more swayed. In that case, I'd suggest using int or unsigned as the type, though, rather than uint32_t. Using uint32_t just because that is the representation in SPIR-V seems like we would be leaking implementation details into the API.

from llvm.