On Thu, 2024-07-11 at 19:31 -0400, Warren Weckesser wrote: > I have implemented quite a few generalized ufuncs over in ufunclab > (https://github.com/WarrenWeckesser/ufunclab), and in the process I > have accumulated a gufunc "wish list". Two items on that list are: > > (1) the ability to impose constraints on the core dimensions that are > checked when the gufunc is called. By far the most common use-case I > have is requiring that a dimension have length at least 1. To do this > currently, I check the shapes in the ufunc loop function, and if they > are not valid, raise an exception and hope that the gufunc machinery > processes it as expected when the loop function returns. (Sorry, I'm > using lingo--"loop function", "core dimension", etc--that will be > familiar to those who already know the ufunc C API, but not so > familiar to general users of NumPy.) > > (2) the ability to have the output dimension be a function of the > input dimensions, instead of being limited to one of the input > dimensions or an independent dimension. Then one could create, for > example, a 1-d convolution gufunc with shape signature that is > effectively `(m),(n)->(m + n - 1)` (corresponding to `mode='full'` in > `np.convolve`) and the gufunc code would automatically allocate the > output with the correct shape and dtype. >
Nice, thanks! I have to look at the implementation in detail, but this seems like a good idea. Have to look at the PR for bike-shedding, but I think we should just add this. (You won't be able to know these relations from reading the signature, but I doubt it's worth worrying about that.) This seems like it should cover all or at least almost all of the things that have come up about ufunc core dimension flexibility (might be nice to check briefly, but even if not I suspect the hook here is the right choice). - Sebastian > I have proposed a change in https://github.com/numpy/numpy/pull/26908 > that makes both these features possible. A field is added to the > PyUFuncObject that is an optional user-defined C function that the > gufunc author implements. When a gufunc is called, this function is > called with an array of the values of the core dimensions of the > input > and output arrays. Some or all of the output core dimensions might be > -1, meaning the arrays are to be allocated by the gufunc/iterator > machinery. The new "hook" allows the user to check the given core > dimensions and raise an exception if some constraint is not > satisfied. > The user-defined function can also replace those -1 values with sizes > that it computes based on the given input core dimensions. > > To define the 1-d convolution gufunc, the actual shape signature that > is passed to `PyUFunc_FromFuncAndDataAndSignature` is `(m),(n)->(p)`. > When a user passes arrays with shapes, say, (20,) and (30,) as the > input and with no output array specified, the user-defined function > will get the array [20, 30, -1]. It would replace -1 with m + n - 1 = > 49 and return. If the caller *does* include an output array in the > call, the core dimension of that array will be the third element of > the array passed to the user-defined function. In that case, the > function verifies that the value equals m + n - 1, and raises an > exception if it doesn't. > > Here's that 1-d convolution, called `conv1d_full` here, in action: > > ``` > In [14]: import numpy as np > > In [15]: from experiment import conv1d_full > > In [16]: type(conv1d_full) > Out[16]: numpy.ufunc > ``` > > `m = 4`, `n = 6`, so the output shape is `p = m + n - 1 = 9`: > > ``` > In [17]: conv1d_full([1, 2, 3, 4], [-1, 1, 2, 1.5, -2, 1]) > Out[17]: array([-1. , -1. , 1. , 4.5, 11. , 9.5, 2. , -5. , 4. ]) > ``` > > Standard broadcasting: > > ``` > In [18]: conv1d_full([[1, 2, 3, 4], [0.5, 0, -1, 1]], [-1, 1, 2, 1.5, > -2, 1]) > Out[18]: > array([[-1. , -1. , 1. , 4.5 , 11. , 9.5 , 2. , -5. , 4. ], > [-0.5 , 0.5 , 2. , -1.25, -2. , 1. , 3.5 , -3. , 1. ]]) > ``` > > Comments here or over in the pull request are welcome. The essential > changes to the source code are just 7 lines in `ufunc_object.c` and 7 > lines in `ufuncobject.h`. The rest of the changes in the PR create a > couple gufuncs that use the new feature, with corresponding unit > tests. > > Warren > _______________________________________________ > NumPy-Discussion mailing list -- numpy-discussion@python.org > To unsubscribe send an email to numpy-discussion-le...@python.org > https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ > Member address: sebast...@sipsolutions.net > _______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
