Dict in njitted function

Minimum example:

import numba as nb
from numba.core import types
from numba.typed import Dict

d_test = Dict.empty(
    key_type=types.unicode_type,
    value_type=types.float64,
)

d_test['a'] = 0.1
d_test['b'] = 0.2


@nb.njit()
def nb_dict_test():
    v1 = d_test['a']
    v2 = d_test['b']
    summed_v = v1 + v2
    return summed_v

nb_dict_test()

or

d_test = Dict.empty(
    key_type=types.float64,
    value_type=types.float64,
)

d_test[0.1] = 0.1
d_test[0.2] = 0.2


@nb.njit()
def nb_dict_test():
    v1 = d_test[0.1]
    v2 = d_test[0.2]
    summed_v = v1 + v2
    return summed_v

nb_dict_test()

do not work with the error message

numba.core.errors.TypingError: Failed in nopython mode pipeline (step: ensure IR is legal prior to lowering)
The use of a DictType[unicode_type,float64]<iv=None> type, assigned to variable 'd_test' in globals, is not supported as globals are considered compile-time constants and there is no known way to compile a DictType[unicode_type,float64]<iv=None> type as a constant.
File "<string>", line 12:
<source missing, REPL/exec in use?>

Putting the dict into the function does work, but this is not the desired solution as it is slow for my case:

@nb.njit()
def nb_dict_test():
    d_test = Dict.empty(
        key_type=types.float64,
        value_type=types.float64,
    )

    d_test[0.1] = 0.1
    d_test[0.2] = 0.2

    v1 = d_test[0.1]
    v2 = d_test[0.2]
    summed_v = v1 + v2
    return summed_v

nb_dict_test()

Not sure if this is indeed not implemented or if there is an error in my implementation? The goal is to have a lookup table inside of a @nb.njit() function - is there any recommended way to do it?

[mix-protocol]

Hello! Just seconding this feature request and checking in if you have found a workaround to this? I’m currently facing the same issue (using a fixed dictionary external to njitted function) and this feature would be extremely useful.

[pengchaoandy]

It is a deviation from Python semantics in Numba that globals are treated as constants: https://numba.readthedocs.io/en/stable/reference/pysemantics.html?highlight=global#global-and-closure-variables - however, I think it ought to be possible to support your use cases, since they don’t mutate the Dict inside the jitted function.

I’ll labelling this as a feature request, to support read-only access to global typed dicts in nopython mode.

If there is a workaround to support this feature in numba 0.52? I tried to hack lower_assign by return llvmlite.ir.values.Constant if value.value is a typed.Dict, but not working.