Should we handle zero-size matirx/vector normally?

[yhmtsai]

In our test, we test the behavior when the matrix is zero-size/zero-column/zero-row.
GPU kernel does not accept grid or block = zero more restrictedly now.
They make it semi-implicit error.
They give invalid configuration error but not from the kernel call, so we can not wrap the kernel call into macro to check the return like the other API.
(Maybe I miss the standard way to validate the grid/block setup?)

Should we still make them perform normally, which might changes something but disable the kernel with zero size?
I am thinking make it explicit error in core interface and do not deal zero-size in the kernel level.

[upsj]

This was always an error, and we added checks for most of those cases (cuda-memcheck also used to fail for that reason)