Backends

Backends

Verbs and registered functions can be implemented for different backends. The dispatching is generally distinguishing different implementations. However, sometimes we want to use different implementations for the same type. In this case, we can use the backend argument to register the function for a specific backend.

For example, we want to implement a function rep for different backends.

from pipda import register_func

@register_func(dispatchable="args")
def rep(x, y):
    raise NotImplementedError

@rep.register(cls=int, backend="python")
def _(x, y):
    return [x] * y

@rep.register(cls=int, backend="numpy")
def _(x, y):
    import numpy as np
    return np.repeat(x, y)

# Later registered backend has higher priority
# But a warning will be shown since we have two implementations for int
rep(1, 3)  # np.array([1, 1, 1])
# Use __backend argument to specify the backend
rep(1, 3, __backend="python")  # [1, 1, 1]

To eliminate the warning, we can also use the favored implementation for a backend:

from pipda import register_func

@register_func(dispatchable="args")
def rep(x, y):
    raise NotImplementedError

@rep.register(cls=int, backend="python")
def _(x, y):
    return [x] * y

@rep.register(cls=int, backend="numpy", favored=True)
def _(x, y):
    import numpy as np
    return np.repeat(x, y)

# Later registered backend has higher priority
# No warnings anymore
rep(1, 3)  # np.array([1, 1, 1])
rep(1, 3, __backend="python")  # [1, 1, 1]

Verbs apply the same rule for backends.