Universal functions (ufunc)

A universal function (or ufunc for short) is a function that operates on ndarrays in an element-by-element fashion, supporting array broadcasting, type casting, and several other standard features. That is, a ufunc is a “vectorized” wrapper for a function that takes a fixed number of specific inputs and produces a fixed number of specific outputs.

In NumPy, universal functions are instances of the numpy.ufunc class. Many of the built-in functions are implemented in compiled C code. The basic ufuncs operate on scalars, but there is also a generalized kind for which the basic elements are sub-arrays (vectors, matrices, etc.), and broadcasting is done over other dimensions. One can also produce custom ufunc instances using the frompyfunc factory function.

Broadcasting

Each universal function takes array inputs and produces array outputs by performing the core function element-wise on the inputs (where an element is generally a scalar, but can be a vector or higher-order sub-array for generalized ufuncs). Standard broadcasting rules are applied so that inputs not sharing exactly the same shapes can still be usefully operated on. Broadcasting can be understood by four rules:

1. All input arrays with ndim smaller than the input array of largest ndim, have 1’s prepended to their shapes.
2. The size in each dimension of the output shape is the maximum of all the input sizes in that dimension.
3. An input can be used in the calculation if its size in a particular dimension either matches the output size in that dimension, or has value exactly 1.
4. If an input has a dimension size of 1 in its shape, the first data entry in that dimension will be used for all calculations along that dimension. In other words, the stepping machinery of the ufunc will simply not step along that dimension (the stride will be 0 for that dimension).

Broadcasting is used throughout NumPy to decide how to handle disparately shaped arrays; for example, all arithmetic operations (+, -, *, …) between ndarrays broadcast the arrays before operation.

A set of arrays is called “broadcastable” to the same shape if the above rules produce a valid result, i.e., one of the following is true:

1. The arrays all have exactly the same shape.
2. The arrays all have the same number of dimensions and the length of each dimensions is either a common length or 1.
3. The arrays that have too few dimensions can have their shapes prepended with a dimension of length 1 to satisfy property 2.

Example

If a.shape is (5,1), b.shape is (1,6), c.shape is (6,) and d.shape is () so that d is a scalar, then a, b, c, and d are all broadcastable to dimension (5,6); and

• a acts like a (5,6) array where a[:,0] is broadcast to the other columns,
• b acts like a (5,6) array where b[0,:] is broadcast to the other rows,
• c acts like a (1,6) array and therefore like a (5,6) array where c[:] is broadcast to every row, and finally,
• d acts like a (5,6) array where the single value is repeated.

Output type determination

The output of the ufunc (and its methods) is not necessarily an ndarray, if all input arguments are not ndarrays. Indeed, if any input defines an __array_ufunc__ method, control will be passed completely to that function, i.e., the ufunc is overridden.

If none of the inputs overrides the ufunc, then all output arrays will be passed to the __array_prepare__ and __array_wrap__ methods of the input (besides ndarrays, and scalars) that defines it and has the highest __array_priority__ of any other input to the universal function. The default __array_priority__ of the ndarray is 0.0, and the default __array_priority__ of a subtype is 1.0. Matrices have __array_priority__ equal to 10.0.

All ufuncs can also take output arguments. If necessary, output will be cast to the data-type(s) of the provided output array(s). If a class with an __array__ method is used for the output, results will be written to the object returned by __array__. Then, if the class also has an __array_prepare__ method, it is called so metadata may be determined based on the context of the ufunc (the context consisting of the ufunc itself, the arguments passed to the ufunc, and the ufunc domain.) The array object returned by __array_prepare__ is passed to the ufunc for computation. Finally, if the class also has an __array_wrap__ method, the returned ndarray result will be passed to that method just before passing control back to the caller.

Use of internal buffers

Internally, buffers are used for misaligned data, swapped data, and data that has to be converted from one data type to another. The size of internal buffers is settable on a per-thread basis. There can be up to buffers of the specified size created to handle the data from all the inputs and outputs of a ufunc. The default size of a buffer is 10,000 elements. Whenever buffer-based calculation would be needed, but all input arrays are smaller than the buffer size, those misbehaved or incorrectly-typed arrays will be copied before the calculation proceeds. Adjusting the size of the buffer may therefore alter the speed at which ufunc calculations of various sorts are completed. A simple interface for setting this variable is accessible using the function

Error handling

Universal functions can trip special floating-point status registers in your hardware (such as divide-by-zero). If available on your platform, these registers will be regularly checked during calculation. Error handling is controlled on a per-thread basis, and can be configured using the functions

Casting Rules

At the core of every ufunc is a one-dimensional strided loop that implements the actual function for a specific type combination. When a ufunc is created, it is given a static list of inner loops and a corresponding list of type signatures over which the ufunc operates. The ufunc machinery uses this list to determine which inner loop to use for a particular case. You can inspect the .types attribute for a particular ufunc to see which type combinations have a defined inner loop and which output type they produce (character codes are used in said output for brevity).

Casting must be done on one or more of the inputs whenever the ufunc does not have a core loop implementation for the input types provided. If an implementation for the input types cannot be found, then the algorithm searches for an implementation with a type signature to which all of the inputs can be cast “safely.” The first one it finds in its internal list of loops is selected and performed, after all necessary type casting. Recall that internal copies during ufuncs (even for casting) are limited to the size of an internal buffer (which is user settable).

By the above description, the casting rules are essentially implemented by the question of when a data type can be cast “safely” to another data type. The answer to this question can be determined in Python with a function call: can_cast(fromtype, totype). The Figure below shows the results of this call for the 24 internally supported types on the author’s 64-bit system. You can generate this table for your system with the code given in the Figure.

>>> def print_table(ntypes):
...     print 'X',
...     for char in ntypes: print char,
...     print
...     for row in ntypes:
...         print row,
...         for col in ntypes:
...             print int(np.can_cast(row, col)),
...         print
>>> print_table(np.typecodes['All'])
X ? b h i l q p B H I L Q P e f d g F D G S U V O M m
? 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
b 0 1 1 1 1 1 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 0 0
h 0 0 1 1 1 1 1 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
i 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 0 0
l 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 0 0
q 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 0 0
p 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 0 0
B 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0
H 0 0 0 1 1 1 1 0 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 0 0
I 0 0 0 0 1 1 1 0 0 1 1 1 1 0 0 1 1 0 1 1 1 1 1 1 0 0
L 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 1 1 0 1 1 1 1 1 1 0 0
Q 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 1 1 0 1 1 1 1 1 1 0 0
P 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 1 1 0 1 1 1 1 1 1 0 0
e 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 0 0
f 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
d 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 0 0
g 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 1 1 1 0 0
F 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0
D 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 0 0
G 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 0 0
S 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 0 0
U 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0
V 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0
O 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0
M 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
m 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1

You should note that, while included in the table for completeness, the ‘S’, ‘U’, and ‘V’ types cannot be operated on by ufuncs. Also, note that on a 32-bit system the integer types may have different sizes, resulting in a slightly altered table.

Mixed scalar-array operations use a different set of casting rules that ensure that a scalar cannot “upcast” an array unless the scalar is of a fundamentally different kind of data (i.e., under a different hierarchy in the data-type hierarchy) than the array. This rule enables you to use scalar constants in your code (which, as Python types, are interpreted accordingly in ufuncs) without worrying about whether the precision of the scalar constant will cause upcasting on your large (small precision) array.

Overriding Ufunc behavior

Classes (including ndarray subclasses) can override how ufuncs act on them by defining certain special methods. For details, see Standard array subclasses.

ufunc

Optional keyword arguments

All ufuncs take optional keyword arguments. Most of these represent advanced usage and will not typically be used.

out

The first output can be provided as either a positional or a keyword parameter. Keyword ‘out’ arguments are incompatible with positional ones.

The ‘out’ keyword argument is expected to be a tuple with one entry per output (which can be None for arrays to be allocated by the ufunc). For ufuncs with a single output, passing a single array (instead of a tuple holding a single array) is also valid.

Passing a single array in the ‘out’ keyword argument to a ufunc with multiple outputs is deprecated, and will raise a warning in numpy 1.10, and an error in a future release.

If ‘out’ is None (the default), a uninitialized return array is created. The output array is then filled with the results of the ufunc in the places that the broadcast ‘where’ is True. If ‘where’ is the scalar True (the default), then this corresponds to the entire output being filled. Note that outputs not explicitly filled are left with their uninitialized values.

where

Accepts a boolean array which is broadcast together with the operands. Values of True indicate to calculate the ufunc at that position, values of False indicate to leave the value in the output alone. This argument cannot be used for generalized ufuncs as those take non-scalar input.

Note that if an uninitialized return array is created, values of False will leave those values uninitialized.

axes

A list of tuples with indices of axes a generalized ufunc should operate on. For instance, for a signature of (i,j),(j,k)->(i,k) appropriate for matrix multiplication, the base elements are two-dimensional matrices and these are taken to be stored in the two last axes of each argument. The corresponding axes keyword would be [(-2, -1), (-2, -1), (-2, -1)]. For simplicity, for generalized ufuncs that operate on 1-dimensional arrays (vectors), a single integer is accepted instead of a single-element tuple, and for generalized ufuncs for which all outputs are scalars, the output tuples can be omitted.

axis

A single axis over which a generalized ufunc should operate. This is a short-cut for ufuncs that operate over a single, shared core dimension, equivalent to passing in axes with entries of (axis,) for each single-core-dimension argument and () for all others. For instance, for a signature (i),(i)->(), it is equivalent to passing in axes=[(axis,), (axis,), ()].

keepdims

If this is set to True, axes which are reduced over will be left in the result as a dimension with size one, so that the result will broadcast correctly against the inputs. This option can only be used for generalized ufuncs that operate on inputs that all have the same number of core dimensions and with outputs that have no core dimensions , i.e., with signatures like (i),(i)->() or (m,m)->(). If used, the location of the dimensions in the output can be controlled with axes and axis.

casting

May be ‘no’, ‘equiv’, ‘safe’, ‘same_kind’, or ‘unsafe’. See can_cast for explanations of the parameter values.

Provides a policy for what kind of casting is permitted. For compatibility with previous versions of NumPy, this defaults to ‘unsafe’ for numpy < 1.7. In numpy 1.7 a transition to ‘same_kind’ was begun where ufuncs produce a DeprecationWarning for calls which are allowed under the ‘unsafe’ rules, but not under the ‘same_kind’ rules. From numpy 1.10 and onwards, the default is ‘same_kind’.

order

Specifies the calculation iteration order/memory layout of the output array. Defaults to ‘K’. ‘C’ means the output should be C-contiguous, ‘F’ means F-contiguous, ‘A’ means F-contiguous if the inputs are F-contiguous and not also not C-contiguous, C-contiguous otherwise, and ‘K’ means to match the element ordering of the inputs as closely as possible.

dtype

Overrides the dtype of the calculation and output arrays. Similar to signature.

subok

Defaults to true. If set to false, the output will always be a strict array, not a subtype.

signature

extobj

Attributes

There are some informational attributes that universal functions possess. None of the attributes can be set.

Methods

All ufuncs have four methods. However, these methods only make sense on scalar ufuncs that take two input arguments and return one output argument. Attempting to call these methods on other ufuncs will cause a ValueError. The reduce-like methods all take an axis keyword, a dtype keyword, and an out keyword, and the arrays must all have dimension >= 1. The axis keyword specifies the axis of the array over which the reduction will take place (with negative values counting backwards). Generally, it is an integer, though for ufunc.reduce, it can also be a tuple of int to reduce over several axes at once, or None, to reduce over all axes. The dtype keyword allows you to manage a very common problem that arises when naively using ufunc.reduce. Sometimes you may have an array of a certain data type and wish to add up all of its elements, but the result does not fit into the data type of the array. This commonly happens if you have an array of single-byte integers. The dtype keyword allows you to alter the data type over which the reduction takes place (and therefore the type of the output). Thus, you can ensure that the output is a data type with precision large enough to handle your output. The responsibility of altering the reduce type is mostly up to you. There is one exception: if no dtype is given for a reduction on the “add” or “multiply” operations, then if the input type is an integer (or Boolean) data-type and smaller than the size of the int_ data type, it will be internally upcast to the int_ (or uint) data-type. Finally, the out keyword allows you to provide an output array (for single-output ufuncs, which are currently the only ones supported; for future extension, however, a tuple with a single argument can be passed in). If out is given, the dtype argument is ignored.

Ufuncs also have a fifth method that allows in place operations to be performed using fancy indexing. No buffering is used on the dimensions where fancy indexing is used, so the fancy index can list an item more than once and the operation will be performed on the result of the previous operation for that item.

Available ufuncs

There are currently more than 60 universal functions defined in numpy on one or more types, covering a wide variety of operations. Some of these ufuncs are called automatically on arrays when the relevant infix notation is used (e.g., add(a, b) is called internally when a + b is written and a or b is an ndarray). Nevertheless, you may still want to use the ufunc call in order to use the optional output argument(s) to place the output(s) in an object (or objects) of your choice.

Recall that each ufunc operates element-by-element. Therefore, each scalar ufunc will be described as if acting on a set of scalar inputs to return a set of scalar outputs.

Math operations

Trigonometric functions

All trigonometric functions use radians when an angle is called for. The ratio of degrees to radians is

Bit-twiddling functions

These function all require integer arguments and they manipulate the bit-pattern of those arguments.

Comparison functions

Floating functions

Recall that all of these functions work element-by-element over an array, returning an array output. The description details only a single operation.