Remove out-of-date numpy.pxd; remove unused float16_t

[jbrockmendel]

• closes #xxxx
• tests added / passed
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

[jbrockmendel]

Just updated the imports of NPY_DATETIME and NPY_TIMEDELTA. Noticed there is some redundancy in the places they are used:

tslib:

cpdef object get_value_box(ndarray arr, object loc):
    [...16 lines suppressed...]
    if arr.descr.type_num == NPY_DATETIME:
        return Timestamp(util.get_value_1d(arr, i))
    elif arr.descr.type_num == NPY_TIMEDELTA:
        return Timedelta(util.get_value_1d(arr, i))
    else:
        return util.get_value_1d(arr, i)

index:

def get_value_at(ndarray arr, object loc):
    if arr.descr.type_num == NPY_DATETIME:
        return Timestamp(util.get_value_at(arr, loc))
    elif arr.descr.type_num == NPY_TIMEDELTA:
        return Timedelta(util.get_value_at(arr, loc))
    return util.get_value_at(arr, loc)

[...]
cdef class IndexEngine:
     [...]
    cpdef get_value(self, ndarray arr, object key, object tz=None):
        """
        arr : 1-dimensional ndarray
        """
        cdef:
            object loc
            void* data_ptr

        loc = self.get_loc(key)
        if PySlice_Check(loc) or cnp.PyArray_Check(loc):
            return arr[loc]
        else:
            if arr.descr.type_num == NPY_DATETIME:
                return Timestamp(util.get_value_at(arr, loc), tz=tz)
            elif arr.descr.type_num == NPY_TIMEDELTA:
                return Timedelta(util.get_value_at(arr, loc))
            return util.get_value_at(arr, loc)

Incidentally, index.get_value_at is never used.

[jreback]

pls leave float16_t around, it is a valid type (though only partially supported generally).

[jbrockmendel]

pls leave float16_t around, it is a valid type (though only partially supported generally).

Can't. This PR removes src/numpy.pxd so as to defer to the version found in setup using numpy_incl = pkg_resources.resource_filename('numpy', 'core/include') (equiv np.get_include()). The idea being to avoid conflicting/redundant includes/namespaces.

If I add float16_t back to one of the cimports here, we get a build-time error:

Error compiling Cython file:
------------------------------------------------------------
...
# cython: profile=False

from numpy cimport (ndarray, float16_t, float64_t, int32_t, int64_t, uint8_t, uint64_t)
^
------------------------------------------------------------

pandas/_libs/index.pyx:3:0: 'numpy/float16_t.pxd' not found

Besides which, float16_t is never actually used in any of the modules it is currently cimported into.

... and looking at numpy.pxd, it looks very much like float16_t is... well I don't know what the term is:

cdef extern from "numpy/arrayobject.h":
    [...]
    ctypedef float        npy_float16
    ctypedef float        npy_float32
[...]

ctypedef npy_float16    float16_t

Is this actually an alias for float32_t?

[gfyoung]

Is this actually an alias for float32_t?

Following that logic that you presented, then yes, it would be.

[jreback]

rebase

[jbrockmendel]

I have no idea what's going on with the test failures. Can anyone reproduce locally? I'd also be ok with a "this cleanup can be future-us's problem"

[jbrockmendel]

May be related to #18064.

[jbrockmendel]

No idea what's up with these failures. Closing to focus attention on #17746, #18016, #18034.