1.5. Creating Numpy universal functions¶
1.5.1. The @vectorize decorator¶
Numba’s vectorize allows Python functions taking scalar input arguments to be used as NumPy ufuncs. Creating a traditional NumPy ufunc is not not the most straightforward process and involves writing some C code. Numba makes this easy. Using the vectorize() decorator, Numba can compile a pure Python function into a ufunc that operates over NumPy arrays as fast as traditional ufuncs written in C.
Using vectorize(), you write your function as operating over input scalars, rather than arrays. Numba will generate the surrounding loop (or kernel) allowing efficient iteration over the actual inputs.
The vectorize() decorator has two modes of operation:
- Eager, or decoration-time, compilation: If you pass one or more type signatures to the decorator, you will be building a Numpy universal function (ufunc). The rest of this subsection describes building ufuncs using decoration-time compilation.
- Lazy, or call-time, compilation: When not given any signatures, the decorator will give you a Numba dynamic universal function (DUFunc) that dynamically compiles a new kernel when called with a previously unsupported input type. A later subsection, “Dynamic universal functions”, describes this mode in more depth.
As described above, if you pass a list of signatures to the vectorize() decorator, your function will be compiled into a Numpy ufunc. In the basic case, only one signature will be passed:
from numba import vectorize, float64
@vectorize([float64(float64, float64)])
def f(x, y):
return x + y
If you pass several signatures, beware that you have to pass most specific signatures before least specific ones (e.g., single-precision floats before double-precision floats), otherwise type-based dispatching will not work as expected:
@vectorize([int32(int32, int32),
int64(int64, int64),
float32(float32, float32),
float64(float64, float64)])
def f(x, y):
return x + y
The function will work as expected over the specified array types:
>>> a = np.arange(6)
>>> f(a, a)
array([ 0, 2, 4, 6, 8, 10])
>>> a = np.linspace(0, 1, 6)
>>> f(a, a)
array([ 0. , 0.4, 0.8, 1.2, 1.6, 2. ])
but it will fail working on other types:
>>> a = np.linspace(0, 1+1j, 6)
>>> f(a, a)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: ufunc 'ufunc' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''
You might ask yourself, “why would I go through this instead of compiling a simple iteration loop using the @jit decorator?”. The answer is that NumPy ufuncs automatically get other features such as reduction, accumulation or broadcasting. Using the example above:
>>> a = np.arange(12).reshape(3, 4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> f.reduce(a, axis=0)
array([12, 15, 18, 21])
>>> f.reduce(a, axis=1)
array([ 6, 22, 38])
>>> f.accumulate(a)
array([[ 0, 1, 2, 3],
[ 4, 6, 8, 10],
[12, 15, 18, 21]])
>>> f.accumulate(a, axis=1)
array([[ 0, 1, 3, 6],
[ 4, 9, 15, 22],
[ 8, 17, 27, 38]])
See also
Standard features of ufuncs (NumPy documentation).
The vectorize() decorator supports multiple ufunc targets:
Target | Description |
---|---|
cpu | Single-threaded CPU |
parallel | Multi-core CPU |
cuda | CUDA GPU Note This creates an ufunc-like object. See documentation for CUDA ufunc for detail. |
A general guideline is to choose different targets for different data sizes and algorithms. The “cpu” target works well for small data sizes (approx. less than 1KB) and low compute intensity algorithms. It has the least amount of overhead. The “parallel” target works well for medium data sizes (approx. less than 1MB). Threading adds a small delay. The “cuda” target works well for big data sizes (approx. greater than 1MB) and high compute intensity algorithms. Transfering memory to and from the GPU adds significant overhead.
1.5.2. The @guvectorize decorator¶
While vectorize() allows you to write ufuncs that work on one element at a time, the guvectorize() decorator takes the concept one step further and allows you to write ufuncs that will work on an arbitrary number of elements of input arrays, and take and return arrays of differing dimensions. The typical example is a running median or a convolution filter.
Contrary to vectorize() functions, guvectorize() functions don’t return their result value: they take it as an array argument, which must be filled in by the function. This is because the array is actually allocated by NumPy’s dispatch mechanism, which calls into the Numba-generated code.
Here is a very simple example:
@guvectorize([(int64[:], int64[:], int64[:])], '(n),()->(n)')
def g(x, y, res):
for i in range(x.shape[0]):
res[i] = x[i] + y[0]
The underlying Python function simply adds a given scalar (y) to all elements of a 1-dimension array. What’s more interesting is the declaration. There are two things there:
- the declaration of input and output layouts, in symbolic form: (n),()->(n) tells NumPy that the function takes a n-element one-dimension array, a scalar (symbolically denoted by the empty tuple ()) and returns a n-element one-dimension array;
- the list of supported concrete signatures as in @vectorize; here we only support int64 arrays.
Note
The concrete signature does not allow for scalar values, even though the layout may mention them. In this example, the second argument is declared as int64[:], not int64. This is why it must be dereferenced by fetching y[0].
We can now check what the compiled ufunc does, over a simple example:
>>> a = np.arange(5)
>>> a
array([0, 1, 2, 3, 4])
>>> g(a, 2)
array([2, 3, 4, 5, 6])
The nice thing is that NumPy will automatically dispatch over more complicated inputs, depending on their shapes:
>>> a = np.arange(6).reshape(2, 3)
>>> a
array([[0, 1, 2],
[3, 4, 5]])
>>> g(a, 10)
array([[10, 11, 12],
[13, 14, 15]])
>>> g(a, np.array([10, 20]))
array([[10, 11, 12],
[23, 24, 25]])
Note
Both vectorize() and guvectorize() support passing nopython=True as in the @jit decorator. Use it to ensure the generated code does not fallback to object mode.
1.5.3. Dynamic universal functions¶
As described above, if you do not pass any signatures to the vectorize() decorator, your Python function will be used to build a dynamic universal function, or DUFunc. For example:
from numba import vectorize
@vectorize
def f(x, y):
return x * y
The resulting f() is a DUFunc instance that starts with no supported input types. As you make calls to f(), Numba generates new kernels whenever you pass a previously unsupported input type. Given the example above, the following set of interpreter interactions illustrate how dynamic compilation works:
>>> f
<numba._DUFunc 'f'>
>>> f.ufunc
<ufunc 'f'>
>>> f.ufunc.types
[]
The example above shows that DUFunc instances are not ufuncs. Rather than subclass ufunc’s, DUFunc instances work by keeping a ufunc member, and then delegating ufunc property reads and method calls to this member (also known as type aggregation). When we look at the initial types supported by the ufunc, we can verify there are none.
Let’s try to make a call to f():
>>> f(3,4)
12
>>> f.types # shorthand for f.ufunc.types
['ll->l']
If this was a normal Numpy ufunc, we would have seen an exception complaining that the ufunc couldn’t handle the input types. When we call f() with integer arguments, not only do we receive an answer, but we can verify that Numba created a loop supporting C long integers.
We can add additional loops by calling f() with different inputs:
>>> f(1.,2.)
2.0
>>> f.types
['ll->l', 'dd->d']
We can now verify that Numba added a second loop for dealing with floating-point inputs, "dd->d".
If we mix input types to f(), we can verify that Numpy ufunc casting rules are still in effect:
>>> f(1,2.)
2.0
>>> f.types
['ll->l', 'dd->d']
This example demonstrates that calling f() with mixed types caused Numpy to select the floating-point loop, and cast the integer argument to a floating-point value. Thus, Numba did not create a special "dl->d" kernel.
This DUFunc behavior leads us to a point similar to the warning given above in “The @vectorize decorator” subsection, but instead of signature declaration order in the decorator, call order matters. If we had passed in floating-point arguments first, any calls with integer arguments would be cast to double-precision floating-point values. For example:
>>> @vectorize
... def g(a, b): return a / b
...
>>> g(2.,3.)
0.66666666666666663
>>> g(2,3)
0.66666666666666663
>>> g.types
['dd->d']
If you require precise support for various type signatures, you should specify them in the vectorize() decorator, and not rely on dynamic compilation.