Varying precision

[richardjgowers]

Currently only float32 precision is supported. Ideally float64 would also be supported.

A good solution would:

• be invisible to users (I could put 32/64 into the same functions)
• handle mixed precision inputs
• not duplicate oodles of code if possible
• have identical tests between different precisions

[kash1102]

Hi richardjgowers

I want to share my approach to fix this issue:

• We could define a variable dtype in class CellGrid(object): located in core.pyx file and equate self.dtype = coordinates.dtype in the init(self, box, max_dist, coordinates=None) function itself and wherever necessary just use dtype attribute of cell-grid object i.e (cg.dtype).

Any inputs on this will be appreciated. Thank you :)

[richardjgowers]

@kash1102 I'm not sure that really fixes anything, it might even be cleaner just to have a property which shortcuts to the array dtype, ie

@property
def dtype(self):
    return self.coordinates.dtype

With Cython the real need is to have the dtype known at compile time so you can write optimised code.

http://cython.readthedocs.io/en/latest/src/userguide/fusedtypes.html

I think what's really needed is that the maths subroutines in the library use fused types (link above). I think they didn't exist when I first ran into this issue

[kash1102]

Hi richardjgowers

Sorry for the late response, I was busy with my academic projects.I have solved the issue by declaring

@property def datatype(self): return self.coordinates.dtype in core.pyx and using method name datatype of CellGrid object wherever necessary example dist = np.empty(Nreq, dtype=cg1.datatype)

and In the cgmath.pyx file simply replaced np.float_32 with fused type floating which would take care of both np.float_32 and np.float_64 values. I have tested manually by giving both np.float_32 and np.float_64 as input and also tested on test files in test directory it works perfectly fine.

Any inputs on this will be appreciated. Thanks :)

[richardjgowers]

Fixed in #8