Add grblas.op namespace to put all operations together.
#67
Conversation
Also, `grblas.op.numpy`, which remains lazy. We now have `grblas.ops` and `grblas.op`. Perhaps we should rename `grblas.ops`.
|
Some methods like Should we always convert monoids to binaryop when appropriate? This should always be safe and possible to do, right? This is similar to upgrading builtin binaryops to monoids when possible in ewise_add. |
|
Given users can And yes, auto-converting from monoid to binary or binary to monoid as required by a method seems okay to me. |
We do this by adding unsurprising attributes to all our operators.
All this made this PR much more substantial. Question: can we not allow using a semiring where a binaryop or monoid is expected? For example, with ewise_add. min_plus(A | B) # bad. does this do min or plus?
min_plus.binary(A | B) # better
plus(A | B) # best
min_plus(A & B) # bad
min_plus.monoid(A & B) # better
min(A & B) # best |
| class TypedOpBase: | ||
| def __init__(self, name, type_, return_type, gb_obj, gb_name): | ||
| def __init__(self, parent, name, type_, return_type, gb_obj, gb_name): | ||
| self.parent = parent |
There was a problem hiding this comment.
Is .parent the best attribute name for the untyped version?
binary.plus[int].parent is binary.plus
|
I should have said I plan to disallow using semiring where it grabs the binary/monoid from it such as ewise_add. Let me know if you prefer otherwise. |
Although this is part of the spec, it's pretty weird to do in Python. It's also easy to do e.g. `my_semiring.binaryop` if you want the binarop.
| self.name = name | ||
| self._typed_ops = {} | ||
| self.types = {} | ||
| self._anonymous = anonymous |
There was a problem hiding this comment.
Should this be private or not?
|
Note that we no longer include Monoids in |
Also, complex monoids still aren't working.
|
@jim22k one more question for you. When users register new operators, we assign them to e.g. |
This is mostly because SuiteSparse doesn't like user-defined binops here. If the user also defined a monoid from the binop, then we can use the monoid instead.
Also, use `GrB` objects when possible over `GxB` objects. I think this may be nicer for the recorder.
| # np.array(values2, dtype=np.complex128), | ||
| # ], | ||
| # ] | ||
| # ) |
There was a problem hiding this comment.
Btw @jim22k complex monoids still don't work. I apparently added a check in "operators.py" to avoid the segfault.
|
Well, this "small" PR became quite large. Merging! |
2 participants
Also,
grblas.op.numpy, which remains lazy.We now have
grblas.opsandgrblas.op. Perhaps we should renamegrblas.ops.This was @jim22k's idea to help make
grblaseasier to use.