Function Types#

Firstly, we explain the difference between primary, compositional, mixed and standalone functions. These four function categorizations are all mutually exclusive, and combined they constitute the set of all functions in Ivy, as outlined in the simple Venn diagram below.

Primary Functions#

Primary functions are essentially the lowest level building blocks in Ivy. Each primary function has a unique backend-specific implementation for each backend specified in ivy/functional/backends/backend_name/ These are generally implemented as light wrapping around an existing function in the backend framework, which serves a near-identical purpose.

Primary functions must both be specified in ivy/functional/ivy/ and also in each of the backend files ivy/functional/backends/backend_name/

The function in ivy/functional/ivy/ includes the type hints, docstring, and docstring examples (explained in more detail in the subsequent sections), but does not include an actual implementation.

Instead, in ivy/functional/ivy/, primary functions simply defer to the backend-specific implementation.

For example, the code for ivy.tan() in ivy/functional/ivy/ (with decorators and docstrings removed) is given below:

def tan(
    x: Union[ivy.Array, ivy.NativeArray],
    out: Optional[ivy.Array] = None,
) -> ivy.Array:
    return ivy.current_backend(x).tan(x, out=out)

The backend-specific implementation of ivy.tan() for PyTorch in ivy/functional/backends/torch/ is given below:

def tan(
    x: torch.Tensor,
    out: Optional[torch.Tensor] = None
) -> torch.Tensor:
    x = _cast_for_unary_op(x)
    return torch.tan(x, out=out)

The reason that the Ivy implementation has type hint Union[ivy.Array, ivy.NativeArray] but PyTorch implementation has torch.Tensor is explained in the Arrays section. Likewise, the reason that the out argument in the Ivy implementation has array type hint ivy.Array whereas x has Union[ivy.Array, ivy.NativeArray] is also explained in the Arrays section.

Compositional Functions#

Compositional functions on the other hand do not have backend-specific implementations. They are implemented as a composition of other Ivy functions, which themselves can be either compositional, primary, or mixed (explained below).

Therefore, compositional functions are only implemented in ivy/functional/ivy/, and there are no implementations in any of the backend files ivy/functional/backends/backend_name/

For example, the implementation of ivy.cross_entropy() in ivy/functional/ivy/ (with docstrings and decorators removed) is given below:

def cross_entropy(
    true: Union[ivy.Array, ivy.NativeArray],
    pred: Union[ivy.Array, ivy.NativeArray],
    axis: int = -1,
    epsilon: float = 1e-7,
    reduction: str = "mean",
    out: Optional[ivy.Array] = None
) -> ivy.Array:
    ivy.utils.assertions.check_elem_in_list(reduction, ["none", "sum", "mean"])
    pred = ivy.clip(pred, epsilon, 1 - epsilon)
    log_pred = ivy.log(pred)
    return _reduce_loss(reduction, log_pred * true, axis, out)

Mixed Functions#

Sometimes, a function may only be provided by some of the supported backends. In this case, we have to take a mixed approach. We should always have a backend-specific implementation if there is a similar function provided by a certain backend. This maximises runtime efficiency, as the function in the backend will be implemented directly in C or C++. Such functions have some backend-specific implementations in ivy/functional/backends/backend_name/, but not for all backends. To support backends that do not have a backend-specific implementation, a compositional implementation is also provided in ivy/functional/ivy/ Compositional functions should only be used when there is no similar function to wrap in the backend.

Because these functions include both a compositional implementation and also at least one backend-specific implementation, these functions are referred to as mixed.

When using ivy without a backend set explicitly (for example ivy.set_backend() has not been called), then the function called is always the one implemented in ivy/functional/ivy/ For primary functions, then ivy.current_backend(array_arg).func_name(...) will call the backend-specific implementation in ivy/functional/backends/backend_name/ directly. However, as just explained, mixed functions implement a compositional approach in ivy/functional/ivy/, without deferring to the backend. Therefore, when no backend is explicitly set, then the compositional implementation is always used for mixed functions, even for backends that have a more efficient backend-specific implementation. Typically the backend should always be set explicitly though (using ivy.set_backend() for example), and in this case the efficient backend-specific implementation will always be used if it exists.

Partial Mixed Functions#

There may be instances wherein the native backend function does not encompass the full range of possible cases that ivy wants to support. One example of this is ivy.linear for which the torch native function torch.nn.functional.linear only supports the weight argument to be a 2 dimensional tensor while as ivy also allows the weight argument to be 3 dimensional. While achieving the objective of having superset behaviour across the backends, the native functionality of frameworks should be made use of as much as possible. Even if a framework-specific function doesn’t provide complete superset behaviour, we should still make use of the partial behaviour that it provides and then add more logic for the remaining part. This is explained in detail in the Maximizing Usage of Native Functionality section. Ivy allows this partial support with the help of the partial_mixed_handler attribute which should be added to the backend implementation with a boolean function that specifies some condition on the inputs to switch between the compositional and primary implementations. For example, the torch backend implementation of linear` looks like:

def linear(
    x: torch.Tensor,
    weight: torch.Tensor,
    bias: Optional[torch.Tensor] = None,
    out: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    return torch.nn.functional.linear(x, weight, bias)

linear.partial_mixed_handler = lambda x, weight, **kwargs: weight.ndim == 2

And to the compositional implementation, we must add the handle_partial_mixed_function decorator. When the backend is set, the handle_partial_mixed_function decorator is added to the primary implementation inside the _wrap_function according to the order in the FN_DECORATORS list. When the function is executed, the handle_partial_mixed_function decorator first evaluates the boolean function using the given inputs, and we use the backend-specific implementation if the result is True and the compositional implementation otherwise.

For further information on decorators, please refer to the Function Wrapping section.

For all mixed functions, we must add the mixed_backend_wrappers attribute to the compositional implementation of mixed functions to specify which additional wrappers need to be applied to the primary implementation and which ones from the compositional implementation should be skipped. We do this by creating a dictionary of two keys, to_add and to_skip, each containing the tuple of wrappers to be added or skipped respectively. In general, handle_out_argument, inputs_to_native_arrays and outputs_to_ivy_arrays should always be added to the primary implementation and inputs_to_ivy_arrays should be skipped. For the linear function, mixed_backend_wrappers was added in the following manner.

linear.mixed_backend_wrappers = {
   "to_add": (
   "to_skip": ("inputs_to_ivy_arrays", "handle_partial_mixed_function"),

When the backend is set, we handle these wrappers for the primary implementation inside the _wrap_function.

Standalone Functions#

Standalone functions are functions which do not reference any other primary, compositional or mixed functions whatsoever.

By definition, standalone functions can only reference themselves or other standalone functions. Most commonly, these functions are convenience functions (see below).

As a first example, every function in the module is a standalone function. All of these either: (a) reference no other function at all, (b) only reference themselves recursively, or (c) reference other standalone functions.

A few other examples outside of the module are: ivy.default which simply returns x if it exists else the default value, ivy.cache_fn which wraps a function such that when cache=True is passed, then a previously cached output is returned, and ivy.stable_divide which simply adds a small constant to the denominator of the division.

Nestable Functions#

Nestable functions are functions which can accept ivy.Container instances in place of any of the arguments. Multiple containers can also be passed in for multiple arguments at the same time, provided that the containers share a common nested structure. If an ivy.Container is passed, then the function is applied to all of the leaves of the container, with the container leaf values passed into the function at the corresponding arguments. In this case, the function will return an ivy.Container in the output. Primary, compositional, mixed, and standalone functions can all also be nestable. This categorization is not mutually exclusive, as outlined by the Venn diagram below:

The nestable property makes it very easy to write a single piece of code that can deal either with individual arguments or arbitrary batches of nested arguments. This is very useful in machine learning, where batches of different training data often need to be processed concurrently. Another example is when the same operation must be performed on each weight in a network. This nestable property of Ivy functions means that the same function can be used for any of these use cases without modification.

This added support for handling ivy.Container instances is all handled automatically when _wrap_function is applied to every function in the ivy module during backend setting. This will add the handle_nestable wrapping to the function if it has the @handle_nestable decorator. This function wrapping process is covered in a bit more detail in the Function Wrapping section.

Nestable functions are explained in more detail in the Containers <containers.rst> section.

Convenience Functions#

A final group of functions are the convenience functions (briefly mentioned above). Convenience functions do not form part of the computation graph directly, and they do not directly modify arrays. However, they can be used to organize and improve the code for other functions which do modify the arrays. Convenience functions can be primary, compositional, mixed, or standalone functions. Many are also nestable. This is another categorization which is not mutually exclusive, as outlined by the Venn diagram below:

Primary convenience functions include: ivy.can_cast which determines if one data type can be cast to another data type according to type-promotion rules, ivy.dtype which gets the data type for the input array, and which gets the device for the input array.

Compositional convenience functions include: ivy.set_default_dtype which sets the global default data dtype, ivy.default_dtype which returns the correct data type to use, considering both the inputs and the globally set default data type, and ivy.get_all_arrays_on_dev which gets all arrays which are currently on the specified device.

Standalone convenience functions include: ivy.get_backend which returns a local Ivy module with the associated backend framework. ivy.nested_map which enables an arbitrary function to be mapped across the leaves of an arbitrary nest, and ivy.index_nest which enables an arbitrary nest to be recursively indexed.

There are many other examples. The convenience functions are not grouped by file or folder. Feel free to have a look through all of the submodules, you should be able to spot quite a few!

Round Up

This should have hopefully given you a good feel for the different function types.

If you have any questions, please feel free to reach out on discord in the function types channel!