4.4. Extending the Numba Frontend

Warning

The Numba APIs described in this document are not guaranteed to be stable. External packages that rely on these APIs may break with new Numba releases. Their description is mostly useful in the context of extending Numba withing the Numba codebase.

4.4.1. Overview

The frontend of Numba first analyzes the control and data flow of a function. It then performs type inference to deduce the types of all intermediate values and identify points where types must be coerced. Type inference attempts to determine a single specific type for variable. When a variable’s type cannot be deduced, or it is determined to take on multiple specific types depending on the control flow, its type falls back to pyobject.

The frontend must succeed in typing all variables unambiguously (i.e. they must not be typed as pyobject) in order for the backend to generate code in nopython mode, because the backend uses type information to match appropriate code generators with the values they operate on. Extending the frontend primarily consists of adding support for new types, to allow variables that hold instances of these types to be typed unambiguously.

4.4.1.1. Numba Types

All Numba types are instances of classes that inherit from numba.types.Type. Numba types can be parameterized (for example, arrays and records), in which case their Type classes will take constructor arguments defining the parameters. Different instances of a parameterized type usually denote distinct types and can trigger different, specialized code generation in the backend.

Note

In the rest of this document, when we refer to a “type”, we mean the Numba type unless we explicitly write “Python type”.

4.4.1.2. Type Inference Mechanism

Type inference is performed for variables in three cases:

  • When automatic JIT compilation is used: the types of function arguments must be deduced from the values passed in.
  • When global variables are accessed: the Numba types of those globals is deduced from their values at the time of compilation.
  • Intermediate values: within a function, the type of every intermediate variable must be deduced.

The types of intermediate values are determined by iteratively propagating type information through the data flow graph (DFG). Each iteration propagates type information along the edges of the DFG, until convergence is reached. When two edges flow into the same node and differing type information is propagated, the type of the node is resolved as pyobject.

In order for the propagation to proceed through functions, operators, and attributes, Numba needs to make use of Type Signatures, which map input types to output types. Numba needs type signatures for:

  • Object attributes: This can include the attributes of instances of Python classes, or modules.
  • Global values: Objects (such as functions) accessed from the global namespace.
  • Operators and other “implicit” functions: Certain Python syntax (like a + b, or iter(o)) triggers special function calls. To overload these operations, a type signature for the appropriate function must be registered.
  • Other entities not described in this document, such as builtin functions.

All type inference happens with a Typing Context. Each target has its own Typing Context - presently there are two, for the CPU and CUDA backends. The majority of type signatures are common between these contexts, but the creation of a context for each target allows specialisation based on intrinsics or other specialised operations and types that a target may support.

4.4.2. Tutorial

We will extend the Numba frontend to support typing a class that it does not currently support by:

  • Adding a Numba Type corresponding to the class,
  • Adding the relevant type signatures for a function and an attribute of the class, and
  • Adding a type signature for overloading an elementary operation.

The example will add support for a module named interval, which is assumed to be external to Numba and contains the following:

class Interval(object):
    '''A half-open interval on the real number line.'''
    def __init__(self, lo, hi):
        self.lo = lo
        self.hi = hi

    def __repr__(self):
        return 'Interval(%f, %f)' % (self.lo, self.hi)

def valid_interval(interval):
    '''Return True if interval.lo <= interval.hi'''
    return interval.lo <= interval.hi

4.4.2.1. Creating a New Numba Type

Types are defined in the numba.types module. To create a new Numba type, subclass numba.types.Type and make a single instance of it:

class IntervalType(numba.types.Type):
    def __init__(self):
        super(IntervalType, self).__init__(name='Interval')

interval_type = IntervalType()

This enables interval_type to be used to declare argument and return types in @jit decorations. For example:

@jit(numba.types.bool_(interval_type, numba.types.float32))
def inside(interval, x):
    return interval.lo <= x < interval.hi

4.4.2.2. Organizing Type Signatures with a Registry

Numba uses a Registry (class numba.typing.templates.Registry) to hold collections of related type signatures for attributes, globals and operators.

Examples of the use of a Registry can be found in numba.typing.cmathdecl, numba.typing.npydecl, and some other modules in numba.typing.

For our interval example, we will create a new Registry. This is overkill for a small set of type signatures, but is representative of what would be required when adding type signatures for more complicated classes and modules.

We will create the numba.typing.intervaldecl module and add the following:

from numba.typing.templates import Registry

registry = Registry()
register = registry.register
register_attr = registry.register_attr
register_global = registry.register_global

register, register_attr, and register_global may now be used later in the module as decorators to record functions that compute the type signatures of functions, attributes, and globals, respectively.

4.4.2.3. Adding an Attribute Value Type Signature

We can add type signatures for attributes of instances of Interval, so that lo and hi are recognized as returning float32 types. This requires creating a subclass of numba.typing.templates.AttributeTemplate (add the following to numba.typing.intervaldecl):

from numba.types import float32
from numba.typing.templates import AttributeTemplate

@register_attr
class IntervalAttributes(AttributeTemplate):
    key = interval_type

    # We will store the interval bounds as 32-bit floats
    _attributes = dict(lo=float32, hi=float32)

    def generic_resolve(self, value, attr):
        return self._attributes[attr]

The key attribute of the template contains the Numba type that needs to be matched to use this template. It can either be an instance of a Type subclass, or the subclass itself, for parametric types.

The AttributeTemplate will first look for a method of the form resolve_<attribute name> to get the type of a specific attribute, otherwise it will delegate to the generic_resolve() method. This call takes both the Numba type instance (useful for parametric types) of the value being accessed, and the name of the attribute. The return value from generic_resolve() is the type of the value returned by the attribute access.

4.4.2.4. Adding a Function Type Signature

In order for the Numba type inference engine to recognize the valid_interval global function, we need to provide a function type signature for it. This is done using a numba.typing.templates.ConcreteTemplate. Add the following to numba.typing.intervaldecl:

from numba.types import bool_, Function
from numba.typing.templates import ConcreteTemplate, signature
from interval import valid_interval

class ValidIntervalSignature(ConcreteTemplate):
    key = valid_interval
    cases = [
        signature(bool_, interval_type)
    ]

register_global(valid_interval, Function(ValidIntervalSignature))

The key for looking up the function type is the Python function itself, valid_interval in this example. The cases attribute lists all of the supported function signature combinations. The first argument to signature is the return type, and the remaining arguments are the types of the function arguments. Only positional arguments are supported for function types (i.e. no keyword arguments).

4.4.2.5. Overloading Elementary Operations

Next, suppose we want to add support for a + operation between two intervals. We need to make a ConcreteTemplate where the key is the string "+". Add to numba.typing.intervaldecl:

from numba.typing.templates import ConcreteTemplate

@builtin
class AdditionSignature(ConcreteTemplate):
    key = '+'
    cases = [
        signature(interval_type, interval_type, interval_type)
    ]

Several templates with the same key can be inserted, and each will be checked for a matching function signatures in the order of insertion. This allows the same key to be overloaded with different numbers of arguments and different argument types.

The list of special function keys includes:

Todo

correct this list

Key Description
+ Addition (2 args) and unary positive (1 arg)
- Subtraction (2 args) and unary negative (1 arg)
* Multiplication
/? Divide (only Python 2)
/ True divide
// Floor divide
% Modulo
** Power
<< Left shift
>> Right shift
& Bitwise AND
| Bitwise OR
^ Bitwise XOR
getiter Get an iterator (equivalent to __iter__())
iternext Return the next element from an iterator (equivalent to __next__())
getitem Get an item (equivalent to __getitem__())

These keys come directly from operations in the Numba IR (see Stage 2: Generate the Numba IR).

In-place operations (like a += b) are assumed to have the same signature as the right-hand side of the expanded form (a = a + b).

4.4.2.6. Installing the Registry in a Typing Context

Once all required type signatures have been added to a Registry, it can then be installed into a typing context. In this example, we will make the registry that we have created available to all typing contexts, so we will make sure that it is installed by modifying numba.typing.context.BaseContext:

class Context(BaseContext):
    def init(self):
        self.install(cmathdecl.registry)
        self.install(intervaldecl.registry)
        self.install(mathdecl.registry)
        self.install(npydecl.registry)
        self.install(operatordecl.registry)

Note the addition of the installation of intervaldecl.registry.

4.4.2.7. Enabling Type Inference for Function Arguments and Globals

Numba is infers the types of arguments and global variables, using the BaseContext.resolve_data_type method. In order to add support for the Interval class, we must first create a function that detects Interval instances. Create a new module, numba.interval_support, containing:

import interval

def is_interval(typ):
    return isinstance(typ, interval.Interval)

Then modify the BaseContext.get_data_type function in numba.typing.context so that just before the final return statement, the following check is added:

if interval_support.is_interval(val):
    return types.interval_type

and add an import for numba.interval_support to the top of the file.

4.4.3. Conclusion

So far we have added support for typing for an attribute, a function, an elementary operator, and have added type inference for function arguments and globals. However, this does not yet enable any change in the code generated by Numba, which requires the addition of backend support for the Interval class, described in the next section.