2.2. Just-in-Time compilation

2.2.1. JIT functions

@numba.jit(signature=None, nopython=False, nogil=False, cache=False, forceobj=False, locals={})

Compile the decorated function on-the-fly to produce efficient machine code. All parameters all optional.

If present, the signature is either a single signature or a list of signatures representing the expected Types and signatures of function arguments and return values. Each signature can be given in several forms:

  • A tuple of Types and signatures arguments (for example (numba.int32, numba.double)) representing the types of the function’s arguments; Numba will then infer an appropriate return type from the arguments.
  • A call signature using Types and signatures, specifying both return type and argument types. This can be given in intuitive form (for example numba.void(numba.int32, numba.double)).
  • A string representation of one of the above, for example "void(int32, double)". All type names used in the string are assumed to be defined in the numba.types module.

nopython and nogil are boolean flags. locals is a mapping of local variable names to Types and signatures.

This decorator has several modes of operation:

  • If one or more signatures are given in signature, a specialization is compiled for each of them. Calling the decorated function will then try to choose the best matching signature, and raise a TypeError if no appropriate conversion is available for the function arguments. If converting succeeds, the compiled machine code is executed with the converted arguments and the return value is converted back according to the signature.
  • If no signature is given, the decorated function implements lazy compilation. Each call to the decorated function will try to re-use an existing specialization if it exists (for example, a call with two integer arguments may re-use a specialization for argument types (numba.int64, numba.int64)). If no suitable specialization exists, a new specialization is compiled on-the-fly, stored for later use, and executed with the converted arguments.

If true, nopython forces the function to be compiled in nopython mode. If not possible, compilation will raise an error.

If true, forceobj forces the function to be compiled in object mode. Since object mode is slower than nopython mode, this is mostly useful for testing purposes.

If true, nogil tries to release the global interpreter lock inside the compiled function. The GIL will only be released if Numba can compile the function in nopython mode, otherwise a compilation warning will be printed.

If true, cache enables a file-based cache to shorten compilation times when the function was already compiled in a previous invocation. The cache is maintained in the __pycache__ subdirectory of the directory containing the source file; if the current user is not allowed to write to it, though, it falls back to a platform-specific user-wide cache directory (such as $HOME/.cache/numba on Unix platforms).

Not all functions can be cached, since some functionality cannot be always persisted to disk. When a function cannot be cached, a warning is emitted; use NUMBA_WARNINGS to see it.

The locals dictionary may be used to force the Types and signatures of particular local variables, for example if you want to force the use of single precision floats at some point. In general, we recommend you let Numba’s compiler infer the types of local variables by itself.

Here is an example with two signatures:

@jit(["int32(int32)", "float32(float32)"], nopython=True)
def f(x): ...

Not putting any parentheses after the decorator is equivalent to calling the decorator without any arguments, i.e.:

def f(x): ...

is equivalent to:

def f(x): ...

The decorator returns a Dispatcher object.


If no signature is given, compilation errors will be raised when the actual compilation occurs, i.e. when the function is first called with some given argument types.


Compilation can be influenced by some dedicated Environment variables.

2.2.2. Generated JIT functions

@numba.generated_jit(nopython=False, nogil=False, cache=False, forceobj=False, locals={})

Like the jit() decorator, but calls the decorated function at compile-time, passing the types of the function’s arguments. The decorated function must return a callable which will be compiled as the function’s implementation for those types, allowing flexible kinds of specialization.

The generated_jit() decorator returns a Dispatcher object.

2.2.3. Dispatcher objects

class Dispatcher

The class of objects created by calling jit() or generated_jit(). You shouldn’t try to create such an object in any other way. Calling a Dispatcher object calls the compiled specialization for the arguments with which it is called, letting it act as an accelerated replacement for the Python function which was compiled.

In addition, Dispatcher objects have the following methods and attributes:


The pure Python function which was compiled.


Print out a listing of the function source code annotated line-by-line with the corresponding Numba IR, and the inferred types of the various variables. If file is specified, printing is done to that file object, otherwise to sys.stdout.


Return a dictionary keying compiled function signatures to the human readable LLVM IR generated for the function. If the signature keyword is specified a string corresponding to that individual signature is returned.


Return a dictionary keying compiled function signatures to the human-readable native assembler code for the function. If the signature keyword is specified a string corresponding to that individual signature is returned.


Recompile all existing signatures. This can be useful for example if a global or closure variable was frozen by your function and its value in Python has changed. Since compiling isn’t cheap, this is mainly for testing and interactive use.

2.2.4. Vectorized functions (ufuncs and DUFuncs)

@numba.vectorize(*, signatures=[], identity=None, nopython=True, target='cpu', forceobj=False, locals={})

Compile the decorated function and wrap it either as a Numpy ufunc or a Numba DUFunc. The optional nopython, forceobj and locals arguments have the same meaning as in numba.jit().

signatures is an optional list of signatures expressed in the same form as in the numba.jit() signature argument. If signatures is non-empty, then the decorator will compile the user Python function into a Numpy ufunc. If no signatures are given, then the decorator will wrap the user Python function in a DUFunc instance, which will compile the user function at call time whenever Numpy can not find a matching loop for the input arguments.

identity is the identity (or unit) value of the function being implemented. Possible values are 0, 1, None, and the string "reorderable". The default is None. Both None and "reorderable" mean the function has no identity value; "reorderable" additionally specifies that reductions along multiple axes can be reordered.

If there are several signatures, they must be ordered from the more specific to the least specific. Otherwise, Numpy’s type-based dispatching may not work as expected. For example, the following is wrong:

@vectorize(["float64(float64)", "float32(float32)"])
def f(x): ...

as running it over a single-precision array will choose the float64 version of the compiled function, leading to much less efficient execution. The correct invocation is:

@vectorize(["float32(float32)", "float64(float64)"])
def f(x): ...

target is a string for backend target; Available values are “cpu”, “parallel”, and “cuda”. To use a multithreaded version, change the target to “parallel”:

@vectorize(["float64(float64)", "float32(float32)"], target='parallel')
def f(x): ...

For the CUDA target, use “cuda”:

@vectorize(["float64(float64)", "float32(float32)"], target='cuda')
def f(x): ...
@numba.guvectorize(signatures, layout, *, identity=None, nopython=True, target='cpu', forceobj=False, locals={})

Generalized version of numba.vectorize(). While numba.vectorize() will produce a simple ufunc whose core functionality (the function you are decorating) operates on scalar operands and returns a scalar value, numba.guvectorize() allows you to create a Numpy ufunc whose core function takes array arguments of various dimensions.

The additional argument layout is a string specifying, in symbolic form, the dimensionality and size relationship of the argument types and return types. For example, a matrix multiplication will have a layout string of "(m,n),(n,p)->(m,p)". Its definition might be (function body omitted):

@guvectorize(["void(float64[:,:], float64[:,:], float64[:,:])"],
def f(a, b, result):
    """Fill-in *result* matrix such as result := a * b"""

If one of the arguments should be a scalar, the corresponding layout specification is () and the argument will really be given to you as a zero-dimension array (you have to dereference it to get the scalar value). For example, a one-dimension moving average with a parameterable window width may have a layout string of "(n),()->(n)".

Note that any output will be given to you preallocated as an additional function argument: your code has to fill it with the appropriate values for the function you are implementing.

If your function doesn’t take an output array, you should omit the “arrow” in the layout string (e.g. "(n),(n)").

See also

Specification of the layout string as supported by Numpy. Note that Numpy uses the term “signature”, which we unfortunately use for something else.

class numba.DUFunc

The class of objects created by calling numba.vectorize() with no signatures.

DUFunc instances should behave similarly to Numpy ufunc objects with one important difference: call-time loop generation. When calling a ufunc, Numpy looks at the existing loops registered for that ufunc, and will raise a TypeError if it cannot find a loop that it cannot safely cast the inputs to suit. When calling a DUFunc, Numba delegates the call to Numpy. If the Numpy ufunc call fails, then Numba attempts to build a new loop for the given input types, and calls the ufunc again. If this second call attempt fails or a compilation error occurs, then DUFunc passes along the exception to the caller.

See also

The “Dynamic universal functions” section in the user’s guide demonstrates the call-time behavior of DUFunc, and discusses the impact of call order on how Numba generates the underlying ufunc.


The actual Numpy ufunc object being built by the DUFunc instance. Note that the DUFunc object maintains several important data structures required for proper ufunc functionality (specifically the dynamically compiled loops). Users should not pass the ufunc value around without ensuring the underlying DUFunc will not be garbage collected.


The number of DUFunc (ufunc) inputs. See ufunc.nin.


The number of DUFunc outputs. See ufunc.nout.


The total number of possible DUFunc arguments (should be nin + nout). See ufunc.nargs.


The number of input types supported by the DUFunc. See ufunc.ntypes.


A list of the supported types given as strings. See ufunc.types.


The identity value when using the ufunc as a reduction. See ufunc.identity.

reduce(A, *, axis, dtype, out, keepdims)

Reduces A‘s dimension by one by applying the DUFunc along one axis. See ufunc.reduce.

accumulate(A, *, axis, dtype, out)

Accumulate the result of applying the operator to all elements. See ufunc.accumulate.

reduceat(A, indices, *, axis, dtype, out)

Performs a (local) reduce with specified slices over a single axis. See ufunc.reduceat.

outer(A, B)

Apply the ufunc to all pairs (a, b) with a in A, and b in B. See ufunc.outer.

at(A, indices, *, B)

Performs unbuffered in place operation on operand A for elements specified by indices. If you are using Numpy 1.7 or earlier, this method will not be present. See ufunc.at.