Numba compiles Python code with LLVM to code which can be natively executed at runtime. This happens by decorating Python functions, which allows users to create native functions for different input types, or to create them on the fly:
@jit('f8(f8[:])')
def sum1d(my_double_array):
sum = 0.0
for i in range(my_double_array.shape[0]):
sum += my_double_array[i]
return sum
To make the above example work for any compatile input types automatically, we can create a function that specializes automatically:
@autojit()
def sum1d(my_array):
...
User elementary types are summarized in the table below and can be found in the numba namespace. These types can be further used to specify arrays in a similar manner to Cython’s memoryviews.
| Type Name | Result Type |
|---|---|
| float_ | float32 |
| double | float64 |
| longdouble | float128 |
| char | signed char |
| int8 | int8 (char) |
| int16 | int16 |
| int32 | int32 |
| int64 | int64 |
| complex64 | float complex |
| complex128 | double complex |
| complex256 | long double complex |
Unsigned integer counterparts are available under the name uint8 etc. Also, short-names are available with the style ‘<char>N’ where char is ‘b’, ‘i’, ‘u’, ‘f’, and ‘c’ for boolean, integer, unsigned, float and complex types respectively with ‘N’ indicating the number of bytes in the type. Thus, f8 is equivalent to float32, and c16 is equivalent to double complex.
Native platform-dependent types are also available under names such as int_, short, ulonglong, etc.
Types are names that can be imported from the numba namespace. Alternatively, they can be specified in strings in the jit decorator.
The jit decorator can take keyword arguments: restype, and argtypes to specify the function signature. Alternatively, the signature can be expressed by passing a single argument to jit either as a string as shown above or directly (assuming the type names have been imported from the numba module):
from numba import f8, jit
@jit(f8(f8[:]))
def sum(arr):
...
Notice how the argument types are passed in as arguments to the return type treated as a python function. Previously, this same syntax was used but embedded in a string which avoids having to import f8 from numba directly.
Arrays may be specified strided or C or Fortran contiguous. For instance, float[:, :] specifies a strided 2D array of floats. float[:, ::1] specifies that the array is C contiguous (row-major), and float[::1, :] specifies that the array is Fortran contiguous (column-major).
Then autojit decorator takes backend as an optional argument, which may be set to bytecode or ast. Both backends currently have different capabilities, but the next release plans for the ast backend to supersede the bytecode backend. The default is ast and the bytecode backend has been deprecated and will be removed in the next release.