stride units in numpy::from_data

[stefanseefeld]

Answering #142, I notice that the np::from_data functions require strides to be counted in units of bytes, rather than elements. Is there any particular reason for that choice ? Given that the function also takes a dtype argument, which provides a get_itemsize() method, it seems plausible (and a bit more convenient) for callers to not have to use sizeof(value_type) explicitly.

@TallJimbo, any thoughts ?

[TallJimbo]

NumPy's internal API uses strides in units of bytes, and I've slowly come around (even in my C++ library for multidimensional arrays, which didn't) to that being the best approach. That's because it's the only way to really support reinterpret_cast-style views into arrays, which are rare but nevertheless important (e.g. interpreting a complex<double> array as a double array, or unpacking the bytes of an uint64 that's used to represent a series of packed bits).

Needing to use bytes internally doesn't preclude having some functions that let users specify strides in units of the data type rather than bytes, but I worry that having different conventions in different places would lead to confusion.

[stefanseefeld]

Are there any places in the API where strides are passed without the dtype being known (either implicitly or explicitly) ?

[TallJimbo]

I don't think so.