Bound checks in numpy array accessors

[SylvainCorlay]

The current behavior of data access methods in py::array and py::array_t is to check the dimension bounds every time the accessor is called, via the check_dimensions function, event when using it in a loop.

I agree that when exposing a feature to Python, it must be written in such a way that regardless of the values passed to it, it cannot crash the interpreter, but these accessors are not directly exposed to the Python users. Instead, I think that one of the main reasons people would right c++ extensions with pybind is because they want speed, and they might be disappointed with this behavior. #500 already improved the situation, but as we discussed there, I would prefer to completely remove the bound checking.

[aldanor]

Which access methods are you most interested in, do you have a minimal example? In a loop, you have unchecked version of byte_offset, plus you can prefetch the data pointer once before the loop starts?

[SylvainCorlay]

I think that all the access methods should be like this on the c++ side (except for at which according to the STL idioms (e.g. std::vector) is checked).

I opened this in a separate issue from the original PR hoping that you would reconsider.

The unchecked version are only a few assembly instructions which makes looping worth considering.

[wjakob]

Closing this since unchecked access is now provided.