Supported NumPy features¶
One objective of Numba is having a seamless integration with NumPy. NumPy arrays provide an efficient storage method for homogeneous sets of data. NumPy dtypes provide type information useful when compiling, and the regular, structured storage of potentially large amounts of data in memory provides an ideal memory layout for code generation. Numba excels at generating code that executes on top of NumPy arrays.
NumPy support in Numba comes in many forms:
Numba understands calls to NumPy ufuncs and is able to generate equivalent native code for many of them.
NumPy arrays are directly supported in Numba. Access to Numpy arrays is very efficient, as indexing is lowered to direct memory accesses when possible.
Numba is able to generate ufuncs and gufuncs. This means that it is possible to implement ufuncs and gufuncs within Python, getting speeds comparable to that of ufuncs/gufuncs implemented in C extension modules using the NumPy C API.
The following sections focus on the Numpy features supported in nopython mode, unless otherwise stated.
Scalar types¶
Numba supports the following Numpy scalar types:
Integers: all integers of either signedness, and any width up to 64 bits
Booleans
Real numbers: single-precision (32-bit) and double-precision (64-bit) reals
Complex numbers: single-precision (2x32-bit) and double-precision (2x64-bit) complex numbers
Datetimes and timestamps: of any unit
Character sequences (but no operations are available on them)
Structured scalars: structured scalars made of any of the types above and arrays of the types above
The following scalar types and features are not supported:
Arbitrary Python objects
Half-precision and extended-precision real and complex numbers
Nested structured scalars the fields of structured scalars may not contain other structured scalars
The operations supported on NumPy scalars are almost the same as on the
equivalent built-in types such as int
or float
. You can use a type’s
constructor to convert from a different type or width. In addition you can use
the view(np.<dtype>)
method to bitcast all int
and float
types
within the same width. However, you must define the scalar using a NumPy
constructor within a jitted function. For example, the following will work:
>>> import numpy as np
>>> from numba import njit
>>> @njit
... def bitcast():
... i = np.int64(-1)
... print(i.view(np.uint64))
...
>>> bitcast()
18446744073709551615
Whereas the following will not work:
>>> import numpy as np
>>> from numba import njit
>>> @njit
... def bitcast(i):
... print(i.view(np.uint64))
...
>>> bitcast(np.int64(-1))
---------------------------------------------------------------------------
TypingError Traceback (most recent call last)
...
TypingError: Failed in nopython mode pipeline (step: ensure IR is legal prior to lowering)
'view' can only be called on NumPy dtypes, try wrapping the variable with 'np.<dtype>()'
File "<ipython-input-3-fc40aaab84c4>", line 3:
def bitcast(i):
print(i.view(np.uint64))
Structured scalars support attribute getting and setting, as well as member lookup using constant strings. Strings stored in a local or global tuple are considered constant strings and can be used for member lookup.
import numpy as np
from numba import njit
arr = np.array([(1, 2)], dtype=[('a1', 'f8'), ('a2', 'f8')])
fields_gl = ('a1', 'a2')
@njit
def get_field_sum(rec):
fields_lc = ('a1', 'a2')
field_name1 = fields_lc[0]
field_name2 = fields_gl[1]
return rec[field_name1] + rec[field_name2]
get_field_sum(arr[0]) # returns 3
It is also possible to use local or global tuples together with literal_unroll
:
import numpy as np
from numba import njit, literal_unroll
arr = np.array([(1, 2)], dtype=[('a1', 'f8'), ('a2', 'f8')])
fields_gl = ('a1', 'a2')
@njit
def get_field_sum(rec):
out = 0
for f in literal_unroll(fields_gl):
out += rec[f]
return out
get_field_sum(arr[0]) # returns 3
Record subtyping¶
Warning
This is an experimental feature.
Numba allows width subtyping of structured scalars.
For example, dtype([('a', 'f8'), ('b', 'i8')])
will be considered a subtype of dtype([('a', 'f8')]
, because
the second is a strict subset of the first, i.e. field a
is of the same type and is in the same position in both
types. The subtyping relationship will matter in cases where compilation for a certain input is not allowed, but the
input is a subtype of another, allowed type.
import numpy as np
from numba import njit, typeof
from numba.core import types
record1 = np.array([1], dtype=[('a', 'f8')])[0]
record2 = np.array([(2,3)], dtype=[('a', 'f8'), ('b', 'f8')])[0]
@njit(types.float64(typeof(record1)))
def foo(rec):
return rec['a']
foo(record1)
foo(record2)
Without subtyping the last line would fail. With subtyping, no new compilation will be triggered, but the
compiled function for record1
will be used for record2
.
See also
Numpy scalars reference.
Array types¶
Numpy arrays of any of the scalar types above are supported, regardless of the shape or layout.
Array access¶
Arrays support normal iteration. Full basic indexing and slicing is supported. A subset of advanced indexing is also supported: only one advanced index is allowed, and it has to be a one-dimensional array (it can be combined with an arbitrary number of basic indices as well).
See also
Numpy indexing reference.
Structured array access¶
Numba presently supports accessing fields of individual elements in structured arrays by attribute as well as by getting and setting. This goes slightly beyond the NumPy API, which only allows accessing fields by getting and setting. For example:
from numba import njit
import numpy as np
record_type = np.dtype([("ival", np.int32), ("fval", np.float64)], align=True)
def f(rec):
value = 2.5
rec[0].ival = int(value)
rec[0].fval = value
return rec
arr = np.ones(1, dtype=record_type)
cfunc = njit(f)
# Works
print(cfunc(arr))
# Does not work
print(f(arr))
The above code results in the output:
[(2, 2.5)]
Traceback (most recent call last):
File "repro.py", line 22, in <module>
print(f(arr))
File "repro.py", line 9, in f
rec[0].ival = int(value)
AttributeError: 'numpy.void' object has no attribute 'ival'
The Numba-compiled version of the function executes, but the pure Python version raises an error because of the unsupported use of attribute access.
Note
This behavior will eventually be deprecated and removed.
Attributes¶
The following attributes of Numpy arrays are supported:
The flags
object¶
The object returned by the flags
attribute supports
the contiguous
, c_contiguous
and f_contiguous
attributes.
The flat
object¶
The object returned by the flat
attribute supports
iteration and indexing, but be careful: indexing is very slow on
non-C-contiguous arrays.
The real
and imag
attributes¶
Numpy supports these attributes regardless of the dtype but Numba chooses to
limit their support to avoid potential user error. For numeric dtypes,
Numba follows Numpy’s behavior. The real
attribute
returns a view of the real part of the complex array and it behaves as an identity
function for other numeric dtypes. The imag
attribute
returns a view of the imaginary part of the complex array and it returns a zero
array with the same shape and dtype for other numeric dtypes. For non-numeric
dtypes, including all structured/record dtypes, using these attributes will
result in a compile-time (TypingError) error. This behavior differs from
Numpy’s but it is chosen to avoid the potential confusion with field names that
overlap these attributes.
Calculation¶
The following methods of Numpy arrays are supported in their basic form (without any optional arguments):
The corresponding top-level Numpy functions (such as numpy.prod()
)
are similarly supported.
Other methods¶
The following methods of Numpy arrays are supported:
argsort()
(kind
key word argument supported for values'quicksort'
and'mergesort'
)astype()
(only the 1-argument form)copy()
(without arguments)dot()
(only the 1-argument form)flatten()
(no order argument; ‘C’ order only)item()
(without arguments)itemset()
(only the 1-argument form)ptp()
(without arguments)ravel()
(no order argument; ‘C’ order only)repeat()
(no axis argument)reshape()
(only the 1-argument form)sort()
(without arguments)sum()
(with or without theaxis
and/ordtype
arguments.)axis
only supportsinteger
values.If the
axis
argument is a compile-time constant, all valid values are supported. An out-of-range value will result in aLoweringError
at compile-time.If the
axis
argument is not a compile-time constant, only values from 0 to 3 are supported. An out-of-range value will result in a runtime exception.All numeric
dtypes
are supported in thedtype
parameter.timedelta
arrays can be used as input arrays buttimedelta
is not supported asdtype
parameter.When a
dtype
is given, it determines the type of the internal accumulator. When it is not, the selection is made automatically based on the input array’sdtype
, mostly following the same rules as NumPy. However, on 64-bit Windows, Numba uses a 64-bit accumulator for integer inputs (int64
forint32
inputs anduint64
foruint32
inputs), while NumPy would use a 32-bit accumulator in those cases.
view()
(only the 1-argument form)
Warning
Sorting may be slightly slower than Numpy’s implementation.
Functions¶
Linear algebra¶
Basic linear algebra is supported on 1-D and 2-D contiguous arrays of floating-point and complex numbers:
numpy.kron()
(‘C’ and ‘F’ order only)numpy.trace()
(only the first argument).On Python 3.5 and above, the matrix multiplication operator from PEP 465 (i.e.
a @ b
wherea
andb
are 1-D or 2-D arrays).numpy.linalg.cond()
(only non string values inp
).numpy.linalg.eig()
(only running with data that does not cause a domain change is supported e.g. real input -> real output, complex input -> complex output).numpy.linalg.eigh()
(only the first argument).numpy.linalg.eigvals()
(only running with data that does not cause a domain change is supported e.g. real input -> real output, complex input -> complex output).numpy.linalg.eigvalsh()
(only the first argument).numpy.linalg.norm()
(only the 2 first arguments and only non string values inord
).numpy.linalg.qr()
(only the first argument).numpy.linalg.svd()
(only the 2 first arguments).
Note
The implementation of these functions needs SciPy to be installed.
Reductions¶
The following reduction functions are supported:
numpy.diff()
(only the 2 first arguments)numpy.median()
(only the first argument)numpy.nancumprod()
(only the first argument, requires NumPy >= 1.12))numpy.nancumsum()
(only the first argument, requires NumPy >= 1.12))numpy.nanmax()
(only the first argument)numpy.nanmean()
(only the first argument)numpy.nanmedian()
(only the first argument)numpy.nanmin()
(only the first argument)numpy.nanpercentile()
(only the 2 first arguments, requires NumPy >= 1.11, complex dtypes unsupported)numpy.nanquantile()
(only the 2 first arguments, requires NumPy >= 1.15, complex dtypes unsupported)numpy.nanprod()
(only the first argument)numpy.nanstd()
(only the first argument)numpy.nansum()
(only the first argument)numpy.nanvar()
(only the first argument)numpy.percentile()
(only the 2 first arguments, requires NumPy >= 1.10, complex dtypes unsupported)numpy.quantile()
(only the 2 first arguments, requires NumPy >= 1.15, complex dtypes unsupported)
Other functions¶
The following top-level functions are supported:
numpy.argsort()
(kind
key word argument supported for values'quicksort'
and'mergesort'
)numpy.array()
(only the 2 first arguments)numpy.asarray()
(only the 2 first arguments)numpy.asarray_chkfinite()
(only the 2 first arguments)numpy.asfortranarray()
(only the first argument)numpy.convolve()
(only the 2 first arguments)numpy.copy()
(only the first argument)numpy.corrcoef()
(only the 3 first arguments, requires SciPy)numpy.correlate()
(only the 2 first arguments)numpy.count_nonzero()
(axis only supports scalar values)numpy.cov()
(only the 5 first arguments)numpy.cross()
(only the 2 first arguments; at least one of the input arrays should haveshape[-1] == 3
)If
shape[-1] == 2
for both inputs, please replace yournumpy.cross()
call withnumba.np.extensions.cross2d()
.
numpy.delete()
(only the 2 first arguments)numpy.dtype()
(only the first argument)numpy.empty()
(only the 2 first arguments)numpy.empty_like()
(only the 2 first arguments)numpy.flatten()
(no order argument; ‘C’ order only)numpy.flip()
(no axis argument)numpy.frombuffer()
(only the 2 first arguments)numpy.full()
(only the 3 first arguments)numpy.full_like()
(only the 3 first arguments)numpy.histogram()
(only the 3 first arguments)numpy.interp()
(only the 3 first arguments; requires NumPy >= 1.10)numpy.linspace()
(only the 3-argument form)numpy.nditer
(only the first argument)numpy.ones()
(only the 2 first arguments)numpy.ones_like()
(only the 2 first arguments)numpy.partition()
(only the 2 first arguments)numpy.ptp()
(only the first argument)numpy.ravel()
(no order argument; ‘C’ order only)numpy.repeat()
(no axis argument)numpy.reshape()
(no order argument; ‘C’ order only)numpy.roll()
(only the 2 first arguments; second argumentshift
must be an integer)numpy.searchsorted()
(only the 3 first arguments)numpy.select()
(only using homogeneous lists or tuples for the first two arguments, condlist and choicelist). Additionally, these two arguments can only contain arrays (unlike Numpy that also accepts tuples).numpy.sort()
(no optional arguments)numpy.take()
(only the 2 first arguments)numpy.trapz()
(only the 3 first arguments)numpy.tri()
(only the 3 first arguments; third argumentk
must be an integer)numpy.tril()
(second argumentk
must be an integer)numpy.tril_indices()
(all arguments must be integer)numpy.tril_indices_from()
(second argumentk
must be an integer)numpy.triu()
(second argumentk
must be an integer)numpy.triu_indices()
(all arguments must be integer)numpy.triu_indices_from()
(second argumentk
must be an integer)numpy.unique()
(only the first argument)numpy.zeros()
(only the 2 first arguments)numpy.zeros_like()
(only the 2 first arguments)
The following constructors are supported, both with a numeric input (to construct a scalar) or a sequence (to construct an array):
numpy.bool_
numpy.complex64
numpy.complex128
numpy.float32
numpy.float64
numpy.int8
numpy.int16
numpy.int32
numpy.int64
numpy.intc
numpy.intp
numpy.uint8
numpy.uint16
numpy.uint32
numpy.uint64
numpy.uintc
numpy.uintp
The following machine parameter classes are supported, with all purely numerical attributes:
numpy.finfo
(machar
attribute not supported)numpy.MachAr
(with no arguments to the constructor)
Literal arrays¶
Neither Python nor Numba has actual array literals, but you can construct
arbitrary arrays by calling numpy.array()
on a nested tuple:
a = numpy.array(((a, b, c), (d, e, f)))
(nested lists are not yet supported by Numba)
Modules¶
random
¶
Numba supports top-level functions from the numpy.random module, but does not allow you to create individual RandomState instances. The same algorithms are used as for the standard random module (and therefore the same notes apply), but with an independent internal state: seeding or drawing numbers from one generator won’t affect the other.
The following functions are supported.
Initialization¶
numpy.random.seed()
: with an integer argument only
Simple random data¶
Permutations¶
numpy.random.choice()
: the optional p argument (probabilities array) is not supportednumpy.random.shuffle()
: the sequence argument must be a one-dimension Numpy array or buffer-providing object (such as abytearray
orarray.array
)
Distributions¶
Warning
The size argument is not supported in the following functions.
Note
Calling numpy.random.seed()
from non-Numba code (or from
object mode code) will seed the Numpy random generator, not the
Numba random generator.
Note
Since version 0.28.0, the generator is thread-safe and fork-safe. Each thread and each process will produce independent streams of random numbers.
stride_tricks
¶
The following function from the numpy.lib.stride_tricks
module
is supported:
as_strided()
(the strides argument is mandatory, the subok argument is not supported)
Standard ufuncs¶
One objective of Numba is having all the standard ufuncs in NumPy understood by Numba. When a supported ufunc is found when compiling a function, Numba maps the ufunc to equivalent native code. This allows the use of those ufuncs in Numba code that gets compiled in nopython mode.
Limitations¶
Right now, only a selection of the standard ufuncs work in nopython mode. Following is a list of the different standard ufuncs that Numba is aware of, sorted in the same way as in the NumPy documentation.
Math operations¶
UFUNC |
MODE |
|
---|---|---|
name |
object mode |
nopython mode |
add |
Yes |
Yes |
subtract |
Yes |
Yes |
multiply |
Yes |
Yes |
divide |
Yes |
Yes |
logaddexp |
Yes |
Yes |
logaddexp2 |
Yes |
Yes |
true_divide |
Yes |
Yes |
floor_divide |
Yes |
Yes |
negative |
Yes |
Yes |
power |
Yes |
Yes |
remainder |
Yes |
Yes |
mod |
Yes |
Yes |
fmod |
Yes |
Yes |
divmod (*) |
Yes |
Yes |
abs |
Yes |
Yes |
absolute |
Yes |
Yes |
fabs |
Yes |
Yes |
rint |
Yes |
Yes |
sign |
Yes |
Yes |
conj |
Yes |
Yes |
exp |
Yes |
Yes |
exp2 |
Yes |
Yes |
log |
Yes |
Yes |
log2 |
Yes |
Yes |
log10 |
Yes |
Yes |
expm1 |
Yes |
Yes |
log1p |
Yes |
Yes |
sqrt |
Yes |
Yes |
square |
Yes |
Yes |
reciprocal |
Yes |
Yes |
conjugate |
Yes |
Yes |
gcd |
Yes |
Yes |
lcm |
Yes |
Yes |
(*) not supported on timedelta types
Trigonometric functions¶
UFUNC |
MODE |
|
---|---|---|
name |
object mode |
nopython mode |
sin |
Yes |
Yes |
cos |
Yes |
Yes |
tan |
Yes |
Yes |
arcsin |
Yes |
Yes |
arccos |
Yes |
Yes |
arctan |
Yes |
Yes |
arctan2 |
Yes |
Yes |
hypot |
Yes |
Yes |
sinh |
Yes |
Yes |
cosh |
Yes |
Yes |
tanh |
Yes |
Yes |
arcsinh |
Yes |
Yes |
arccosh |
Yes |
Yes |
arctanh |
Yes |
Yes |
deg2rad |
Yes |
Yes |
rad2deg |
Yes |
Yes |
degrees |
Yes |
Yes |
radians |
Yes |
Yes |
Bit-twiddling functions¶
UFUNC |
MODE |
|
---|---|---|
name |
object mode |
nopython mode |
bitwise_and |
Yes |
Yes |
bitwise_or |
Yes |
Yes |
bitwise_xor |
Yes |
Yes |
bitwise_not |
Yes |
Yes |
invert |
Yes |
Yes |
left_shift |
Yes |
Yes |
right_shift |
Yes |
Yes |
Comparison functions¶
UFUNC |
MODE |
|
---|---|---|
name |
object mode |
nopython mode |
greater |
Yes |
Yes |
greater_equal |
Yes |
Yes |
less |
Yes |
Yes |
less_equal |
Yes |
Yes |
not_equal |
Yes |
Yes |
equal |
Yes |
Yes |
logical_and |
Yes |
Yes |
logical_or |
Yes |
Yes |
logical_xor |
Yes |
Yes |
logical_not |
Yes |
Yes |
maximum |
Yes |
Yes |
minimum |
Yes |
Yes |
fmax |
Yes |
Yes |
fmin |
Yes |
Yes |
Floating functions¶
UFUNC |
MODE |
|
---|---|---|
name |
object mode |
nopython mode |
isfinite |
Yes |
Yes |
isinf |
Yes |
Yes |
isnan |
Yes |
Yes |
signbit |
Yes |
Yes |
copysign |
Yes |
Yes |
nextafter |
Yes |
Yes |
modf |
Yes |
No |
ldexp |
Yes (*) |
Yes |
frexp |
Yes |
No |
floor |
Yes |
Yes |
ceil |
Yes |
Yes |
trunc |
Yes |
Yes |
spacing |
Yes |
Yes |
(*) not supported on windows 32 bit
Datetime functions¶
UFUNC |
MODE |
|
---|---|---|
name |
object mode |
nopython mode |
isnat |
Yes |
Yes |