Skip to content

Latest commit

 

History

History
176 lines (138 loc) · 7.87 KB

README.md

File metadata and controls

176 lines (138 loc) · 7.87 KB

Pass-through type for Numba

Tools to ferry arbitrary Python objects through nopython mode. This is a stand-alone version of the Numba internal implementation PR 3640.

This has two typical use-cases:

  1. ferrying data structures not currently supported by Numba into objmode blocks via PassThruContainer
  2. creating extension types that are simplified representations of the Python class and keep a link to the Python object

It's not unlikely both can be avoided completely when starting from scratch but potentially require extensive refactoring when moving Numba into an existing code base.

Ferrying objects into objmode using PassThruContainer

PassThruContainer can be used to make Python objects not supported by nopython available inside objmode blocks. Outside objmode blocks the only supported operation on PassThruContainer is ==. Two instances are equal if the wrapped objects are identical, ie a.obj is b.obj.

In the following example an object unsupported in nopython is ferried into an objmode block and mutated there.

from __future__ import annotations
from numba import jit, objmode
from numba_passthru import PassThruContainer

class Testee:
    def __init__(self, value: int, invert=False):
        self.value = value
        self.numba_will_not_like_this = compile('pass', 'N/A', 'single')
        
    def __gt__(self, other):    # will be used in next example
        return self.value > other.value

testee = Testee(1)  
container = PassThruContainer(testee)

@jit(nopython=True)
def do_something(container):
    with objmode(value='int64'):
        setattr(container.obj, 'value_2', 2)
        value = container.obj.value
        
    return container, value
    
result, value = do_something(container)

assert container is result
assert value == 1
assert result.obj.value_2 == 2

There will be no speed-up for the code inside the objmode block and container is (un)boxed twice adding further overhead. Hence, this only makes sense in rarely visited code-path and if refactoring into a more Numba friendly form is not an option.

Note that the example above already contains the most common pattern that is pretty unpleasant to refactor into a Numba friendly form in requiring object identity being preserved through nopython (assert container is result). Obviously, this is a highly artificial requirement in this toy example but might get more real if the pass-through object is part of conditional branches.

Creating custom passthrough types

PassThroughContainer does not allow attribute access on the wrapped object in nopython and there is no way to dispatch on the type of the wrapped object. To get both you can create a Numba extension type using pass_thru_type. pass_thru_type holds a pointer to the PyObject and manages references. pass_thru_type can be used like any mem-managed member on an extension type. (Some familiarity with Numba extension types is expected for the following.)

Continuing the example above let's try to get the following code working in nopython (another toy example, no speed-up expected):

def find_max(testees: List[Testee]) -> Testee:
    result = testees[0]  # testees must not be empty
    for testee in testees[1:]:
        if testee > result:
            result = testee

    return result     

PassThroughContainer will not help here as there would be no way to dispatch > to Testee.__gt__ and there would be no way to access .value from nopython inside Testee.__gt__. Still, since Testee.value is the only attribute being accessed from nopython there is a realistic chance to get this working. Indeed, assuming we already had the (un)boxer this is a straight forward Numba extension type:

from numba import jit, types
from numba.extending import overload
from numba.typing.typeof import typeof_impl
import operator

class TesteeType(PassThruType):
    def __init__(self):
        super(TesteeType, self).__init__(name='Testee')

testee_type = TesteeType()

@typeof_impl.register(Testee)
def type_testee(val, context):
    return testee_type

@overload(operator.gt)
def testee_gt(self, other):
    if self is testee_type and other is testee_type:
        return Testee.__gt__

find_max = jit(nopython=True)(find_max)

Trying to implement the (un)boxer to somehow pass the .numba_will_not_like_this attribute around nopython (sharing a dict between boxer/unboxer etc) is not straight forward to get working for find_max alone and it is impossible to get the reference counting right in the general case. The clean approach is to have the Numba runtime manage the references by putting a NRT managed reference to the original Python object onto the extension type's data model.

pass_thru_type helps with the boiler-plate of boxing/unboxing the required MemInfoPointer. The PyObject passed into the unboxer can be unboxed directly into a pass_thru_type. On the way out the original PyObject is recovered by boxing the pass_thru_type.

from numba import cgutils
from numba.datamodel import models
from numba.extending import make_attribute_wrapper, overload, overload_method, register_model
from numba.pythonapi import NativeValue, unbox, box
from numba.targets.hashing import _Py_hash_t

from numba_passthru import pass_thru_type

@register_model(TesteeType)
class TesteeModel(models.StructModel):
    def __init__(self, dmm, fe_typ):
        members = [
            ('parent', pass_thru_type),
            ('value', types.intp),
        ]
        super(TesteeModel, self).__init__(dmm, fe_typ, members)
        
make_attribute_wrapper(TesteeType, 'value', 'value')

@unbox(TesteeType)
def unbox_testee(typ, obj, context):
    testee = cgutils.create_struct_proxy(typ)(context.context, context.builder)
    
    testee.parent = context.unbox(pass_thru_type, obj).value
    
    value = context.pyapi.object_getattr_string(obj, "value")
    native_value = context.unbox(types.intp, value)
    context.pyapi.decref(value)

    testee.value = native_value.value

    is_error = cgutils.is_not_null(context.builder, context.pyapi.err_occurred())
    return NativeValue(testee._getvalue(), is_error=is_error)
    
@box(TesteeType)
def box_testee(typ, val, context):
    val = cgutils.create_struct_proxy(typ)(context.context, context.builder, value=val)
    obj = context.box(pass_thru_type, val.parent)

    return obj

Given the implementation above TesteeType is immutable from nopython (make_attribute_wrapper creates read-only attributes). If you made a pass-through type mutable from nopython you had to make sure to reflect changes back to the Python object in the boxer. However, given the experience with reflected lists and sets there are good reasons to be careful about this.

Upward compatibility notice

This is a stand-alone version of Numba PR 3640. Import of PassThruType, pass_thru_type, PassThruContainer, pass_thru_container_type from numba is attempted first hence you will get the Numba internal implementations once the PR has landed.

This package contains an overload of int(Opaque) (essentially ptrtoint) that might break future Numba versions if Numba created diverging implementations.

This was considered too unlikely to put a version constraint on the Numba dependency (which would require a new release of numba-passthru every time a new Numba versions is released)