np.finfo support for quaddtype
#159
Conversation
|
@juntyr can you please take a look at the |
| static PyObject * | ||
| QuadPrecDType_finfo(QuadPrecDTypeObject *self, PyObject *args) | ||
| { | ||
| PyObject *numpy_quaddtype_module = PyImport_ImportModule("numpy_quaddtype"); |
There was a problem hiding this comment.
I don't like doing this as well, but for now it works so
| eps: float = float(epsilon) | ||
| epsneg: float = float(epsilon) | ||
| iexp: int = int(precision) | ||
| machar: object = None |
There was a problem hiding this comment.
is just an object finfo uses to fetch the finfo properties for dtypes, since we are defining them in our package hence marking it as None
| """ | ||
| bits: int = int(bits) | ||
| eps: float = float(epsilon) | ||
| epsneg: float = float(epsilon) |
There was a problem hiding this comment.
This is wrong - it should 1 - nextafter(1, 0) [similar to how exp is nextafter(1, 2) - 1], we can probably compute this in the C++ source and then expose it as a new constant
| bits: int = int(bits) | ||
| eps: float = float(epsilon) | ||
| epsneg: float = float(epsilon) | ||
| iexp: int = int(precision) |
There was a problem hiding this comment.
iexp should be 15, I think?
| epsneg: float = float(epsilon) | ||
| iexp: int = int(precision) | ||
| machar: object = None | ||
| machep: float = float(epsilon) |
There was a problem hiding this comment.
machep should be int(log2(epsilon))?
| machar: object = None | ||
| machep: float = float(epsilon) | ||
| max: float = float(max_value) | ||
| maxexp: float = float(max_value) |
There was a problem hiding this comment.
This would probably be int(ceil(max_value))?
| machep: float = float(epsilon) | ||
| max: float = float(max_value) | ||
| maxexp: float = float(max_value) | ||
| min: float = float(smallest_normal) |
| max: float = float(max_value) | ||
| maxexp: float = float(max_value) | ||
| min: float = float(smallest_normal) | ||
| minexp: float = float(smallest_normal) |
There was a problem hiding this comment.
not sure how to compute this one
| maxexp: float = float(max_value) | ||
| min: float = float(smallest_normal) | ||
| minexp: float = float(smallest_normal) | ||
| negep: float = float(epsilon) |
| min: float = float(smallest_normal) | ||
| minexp: float = float(smallest_normal) | ||
| negep: float = float(epsilon) | ||
| nexp: int = int(bits) - int(precision) - 1 |
| minexp: float = float(smallest_normal) | ||
| negep: float = float(epsilon) | ||
| nexp: int = int(bits) - int(precision) - 1 | ||
| nmant: int = int(precision) |
There was a problem hiding this comment.
Thanks @SwayamInSync for working on this!
I am really uncomfortable with all the casts to float, since at least for some of them they will undoubtedly over- or underflow. In my mind, the finfo values should have the same dtype as the type we're describing, but I know this would likely be a breaking change; but likely one we need since otherwise e.g. finfo(quad).max don't make any sense
Hey yes good point, just fixed that |
|
What's the reason for using the explicit field constructor? |
SwayamInSync
changed the title
np.finfo support for quaddtyoenp.finfo support for quaddtype
Making this PR here as well, as a demonstration to how
np.finfoand other related information can be passed to NumPy.Edit: Below is not possible as
QuadPrecDTypeis an immutable typeBut this just feels so illegal that's why registering as
.tp_methodsFrom the NumPy side it requires numpy/numpy#29771 to get merged.