NumPy

NEP 41 — First step towards a new Datatype System

title

Improved Datatype Support

Author

Sebastian Berg

Author

Stéfan van der Walt

Author

Matti Picus

Status

Draft

Type

Standard Track

Created

2020-02-03

Note

This NEP is part of a series of NEPs encompassing first information about the previous dtype implementation and issues with it in NEP 40. NEP 41 (this document) then provides an overview and generic design choices for the refactor. Further NEPs 42 and 43 go into the technical details of the datatype and universal function related internal and external API changes. In some cases it may be necessary to consult the other NEPs for a full picture of the desired changes and why these changes are necessary.

Abstract

Datatypes <data-type-objects-dtype> in NumPy describe how to interpret each element in arrays. NumPy provides int, float, and complex numerical types, as well as string, datetime, and structured datatype capabilities. The growing Python community, however, has need for more diverse datatypes. Examples are datatypes with unit information attached (such as meters) or categorical datatypes (fixed set of possible values). However, the current NumPy datatype API is too limited to allow the creation of these.

This NEP is the first step to enable such growth; it will lead to a simpler development path for new datatypes. In the long run the new datatype system will also support the creation of datatypes directly from Python rather than C. Refactoring the datatype API will improve maintainability and facilitate development of both user-defined external datatypes, as well as new features for existing datatypes internal to NumPy.

Motivation and Scope

See also

The user impact section includes examples of what kind of new datatypes will be enabled by the proposed changes in the long run. It may thus help to read these section out of order.

Motivation

One of the main issues with the current API is the definition of typical functions such as addition and multiplication for parametric datatypes (see also NEP 40) which require additional steps to determine the output type. For example when adding two strings of length 4, the result is a string of length 8, which is different from the input. Similarly, a datatype which embeds a physical unit must calculate the new unit information: dividing a distance by a time results in a speed. A related difficulty is that the current casting rules – the conversion between different datatypes – cannot describe casting for such parametric datatypes implemented outside of NumPy.

This additional functionality for supporting parametric datatypes introduces increased complexity within NumPy itself, and furthermore is not available to external user-defined datatypes. In general the concerns of different datatypes are not well well-encapsulated. This burden is exacerbated by the exposure of internal C structures, limiting the addition of new fields (for example to support new sorting methods [new_sort]).

Currently there are many factors which limit the creation of new user-defined datatypes:

  • Creating casting rules for parametric user-defined dtypes is either impossible or so complex that it has never been attempted.

  • Type promotion, e.g. the operation deciding that adding float and integer values should return a float value, is very valuable for numeric datatypes but is limited in scope for user-defined and especially parametric datatypes.

  • Much of the logic (e.g. promotion) is written in single functions instead of being split as methods on the datatype itself.

  • In the current design datatypes cannot have methods that do not generalize to other datatypes. For example a unit datatype cannot have a .to_si() method to easily find the datatype which would represent the same values in SI units.

The large need to solve these issues has driven the scientific community to create work-arounds in multiple projects implementing physical units as an array-like class instead of a datatype, which would generalize better across multiple array-likes (Dask, pandas, etc.). Already, Pandas has made a push into the same direction with its extension arrays [pandas_extension_arrays] and undoubtedly the community would be best served if such new features could be common between NumPy, Pandas, and other projects.

Scope

The proposed refactoring of the datatype system is a large undertaking and thus is proposed to be split into various phases, roughly:

  • Phase I: Restructure and extend the datatype infrastructure (This NEP 41)

  • Phase II: Incrementally define or rework API (Detailed largely in NEPs 42/43)

  • Phase III: Growth of NumPy and Scientific Python Ecosystem capabilities.

For a more detailed accounting of the various phases, see “Plan to Approach the Full Refactor” in the Implementation section below. This NEP proposes to move ahead with the necessary creation of new dtype subclasses (Phase I), and start working on implementing current functionality. Within the context of this NEP all development will be fully private API or use preliminary underscored names which must be changed in the future. Most of the internal and public API choices are part of a second Phase and will be discussed in more detail in the following NEPs 42 and 43. The initial implementation of this NEP will have little or no effect on users, but provides the necessary ground work for incrementally addressing the full rework.

The implementation of this NEP and the following, implied large rework of how datatypes are defined in NumPy is expected to create small incompatibilities (see backward compatibility section). However, a transition requiring large code adaption is not anticipated and not within scope.

Specifically, this NEP makes the following design choices which are discussed in more details in the detailed description section:

  1. Each datatype will be an instance of a subclass of np.dtype, with most of the datatype-specific logic being implemented as special methods on the class. In the C-API, these correspond to specific slots. In short, for f = np.dtype("f8"), isinstance(f, np.dtype) will remain true, but type(f) will be a subclass of np.dtype rather than just np.dtype itself. The PyArray_ArrFuncs which are currently stored as a pointer on the instance (as PyArray_Descr->f), should instead be stored on the class as typically done in Python. In the future these may correspond to python side dunder methods. Storage information such as itemsize and byteorder can differ between different dtype instances (e.g. “S3” vs. “S8”) and will remain part of the instance. This means that in the long run the current lowlevel access to dtype methods will be removed (see PyArray_ArrFuncs in NEP 40).

  2. The current NumPy scalars will not change, they will not be instances of datatypes. This will also be true for new datatypes, scalars will not be instances of a dtype (although isinstance(scalar, dtype) may be made to return True when appropriate).

Detailed technical decisions to follow in NEP 42.

Further, the public API will be designed in a way that is extensible in the future:

  1. All new C-API functions provided to the user will hide implementation details as much as possible. The public API should be an identical, but limited, version of the C-API used for the internal NumPy datatypes.

The changes to the datatype system in Phase II must include a large refactor of the UFunc machinery, which will be further defined in NEP 43:

  1. To enable all of the desired functionality for new user-defined datatypes, the UFunc machinery will be changed to replace the current dispatching and type resolution system. The old system should be mostly supported as a legacy version for some time.

Additionally, as a general design principle, the addition of new user-defined datatypes will not change the behaviour of programs. For example common_dtype(a, b) must not be c unless a or b know that c exists.

User Impact

The current ecosystem has very few user-defined datatypes using NumPy, the two most prominent being: rational and quaternion. These represent fairly simple datatypes which are not strongly impacted by the current limitations. However, we have identified a need for datatypes such as:

  • bfloat16, used in deep learning

  • categorical types

  • physical units (such as meters)

  • datatypes for tracing/automatic differentiation

  • high, fixed precision math

  • specialized integer types such as int2, int24

  • new, better datetime representations

  • extending e.g. integer dtypes to have a sentinel NA value

  • geometrical objects [pygeos]

Some of these are partially solved; for example unit capability is provided in astropy.units, unyt, or pint, as numpy.ndarray subclasses. Most of these datatypes, however, simply cannot be reasonably defined right now. An advantage of having such datatypes in NumPy is that they should integrate seamlessly with other array or array-like packages such as Pandas, xarray [xarray_dtype_issue], or Dask.

The long term user impact of implementing this NEP will be to allow both the growth of the whole ecosystem by having such new datatypes, as well as consolidating implementation of such datatypes within NumPy to achieve better interoperability.

Examples

The following examples represent future user-defined datatypes we wish to enable. These datatypes are not part the NEP and choices (e.g. choice of casting rules) are possibilities we wish to enable and do not represent recommendations.

Simple Numerical Types

Mainly used where memory is a consideration, lower-precision numeric types such as :ref:`bfloat16 <https://en.wikipedia.org/wiki/Bfloat16_floating-point_format>` are common in other computational frameworks. For these types the definitions of things such as np.common_type and np.can_cast are some of the most important interfaces. Once they support np.common_type, it is (for the most part) possible to find the correct ufunc loop to call, since most ufuncs – such as add – effectively only require np.result_type:

>>> np.add(arr1, arr2).dtype == np.result_type(arr1, arr2)

and ~numpy.result_type is largely identical to ~numpy.common_type.

Fixed, high precision math

Allowing arbitrary precision or higher precision math is important in simulations. For instance mpmath defines a precision:

>>> import mpmath as mp
>>> print(mp.dps)  # the current (default) precision
15

NumPy should be able to construct a native, memory-efficient array from a list of mpmath.mpf floating point objects:

>>> arr_15_dps = np.array(mp.arange(3))  # (mp.arange returns a list)
>>> print(arr_15_dps)  # Must find the correct precision from the objects:
array(['0.0', '1.0', '2.0'], dtype=mpf[dps=15])

We should also be able to specify the desired precision when creating the datatype for an array. Here, we use np.dtype[mp.mpf] to find the DType class (the notation is not part of this NEP), which is then instantiated with the desired parameter. This could also be written as MpfDType class:

>>> arr_100_dps = np.array([1, 2, 3], dtype=np.dtype[mp.mpf](dps=100))
>>> print(arr_15_dps + arr_100_dps)
array(['0.0', '2.0', '4.0'], dtype=mpf[dps=100])

The mpf datatype can decide that the result of the operation should be the higher precision one of the two, so uses a precision of 100. Furthermore, we should be able to define casting, for example as in:

>>> np.can_cast(arr_15_dps.dtype, arr_100_dps.dtype, casting="safe")
True
>>> np.can_cast(arr_100_dps.dtype, arr_15_dps.dtype, casting="safe")
False  # loses precision
>>> np.can_cast(arr_100_dps.dtype, arr_100_dps.dtype, casting="same_kind")
True

Casting from float is a probably always at least a same_kind cast, but in general, it is not safe:

>>> np.can_cast(np.float64, np.dtype[mp.mpf](dps=4), casting="safe")
False

since a float64 has a higer precision than the mpf datatype with dps=4.

Alternatively, we can say that:

>>> np.common_type(np.dtype[mp.mpf](dps=5), np.dtype[mp.mpf](dps=10))
np.dtype[mp.mpf](dps=10)

And possibly even:

>>> np.common_type(np.dtype[mp.mpf](dps=5), np.float64)
np.dtype[mp.mpf](dps=16)  # equivalent precision to float64 (I believe)

since np.float64 can be cast to a np.dtype[mp.mpf](dps=16) safely.

Categoricals

Categoricals are interesting in that they can have fixed, predefined values, or can be dynamic with the ability to modify categories when necessary. The fixed categories (defined ahead of time) is the most straight forward categorical definition. Categoricals are hard, since there are many strategies to implement them, suggesting NumPy should only provide the scaffolding for user-defined categorical types. For instance:

>>> cat = Categorical(["eggs", "spam", "toast"])
>>> breakfast = array(["eggs", "spam", "eggs", "toast"], dtype=cat)

could store the array very efficiently, since it knows that there are only 3 categories. Since a categorical in this sense knows almost nothing about the data stored in it, few operations makes, sense, although equality does:

>>> breakfast2 = array(["eggs", "eggs", "eggs", "eggs"], dtype=cat)
>>> breakfast == breakfast2
array[True, False, True, False])

The categorical datatype could work like a dictionary: no two items names can be equal (checked on dtype creation), so that the equality operation above can be performed very efficiently. If the values define an order, the category labels (internally integers) could be ordered the same way to allow efficient sorting and comparison.

Whether or not casting is defined from one categorical with less to one with strictly more values defined, is something that the Categorical datatype would need to decide. Both options should be available.

Unit on the Datatype

There are different ways to define Units, depending on how the internal machinery would be organized, one way is to have a single Unit datatype for every existing numerical type. This will be written as Unit[float64], the unit itself is part of the DType instance Unit[float64]("m") is a float64 with meters attached:

>>> from astropy import units
>>> meters = np.array([1, 2, 3], dtype=np.float64) * units.m  # meters
>>> print(meters)
array([1.0, 2.0, 3.0], dtype=Unit[float64]("m"))

Note that units are a bit tricky. It is debatable, whether:

>>> np.array([1.0, 2.0, 3.0], dtype=Unit[float64]("m"))

should be valid syntax (coercing the float scalars without a unit to meters). Once the array is created, math will work without any issue:

>>> meters / (2 * unit.seconds)
array([0.5, 1.0, 1.5], dtype=Unit[float64]("m/s"))

Casting is not valid from one unit to the other, but can be valid between different scales of the same dimensionality (although this may be “unsafe”):

>>> meters.astype(Unit[float64]("s"))
TypeError: Cannot cast meters to seconds.
>>> meters.astype(Unit[float64]("km"))
>>> # Convert to centimeter-gram-second (cgs) units:
>>> meters.astype(meters.dtype.to_cgs())

The above notation is somewhat clumsy. Functions could be used instead to convert between units. There may be ways to make these more convenient, but those must be left for future discussions:

>>> units.convert(meters, "km")
>>> units.to_cgs(meters)

There are some open questions. For example, whether additional methods on the array object could exist to simplify some of the notions, and how these would percolate from the datatype to the ndarray.

The interaction with other scalars would likely be defined through:

>>> np.common_type(np.float64, Unit)
Unit[np.float64](dimensionless)

Ufunc output datatype determination can be more involved than for simple numerical dtypes since there is no “universal” output type:

>>> np.multiply(meters, seconds).dtype != np.result_type(meters, seconds)

In fact np.result_type(meters, seconds) must error without context of the operation being done. This example highlights how the specific ufunc loop (loop with known, specific DTypes as inputs), has to be able to to make certain decisions before the actual calculation can start.

Implementation

Plan to Approach the Full Refactor

To address these issues in NumPy and enable new datatypes, multiple development stages are required:

  • Phase I: Restructure and extend the datatype infrastructure (This NEP)

    • Organize Datatypes like normal Python classes [PR 15508]_

  • Phase II: Incrementally define or rework API

    • Create a new and easily extensible API for defining new datatypes and related functionality. (NEP 42)

    • Incrementally define all necessary functionality through the new API (NEP 42):

      • Defining operations such as np.common_type.

      • Allowing to define casting between datatypes.

      • Add functionality necessary to create a numpy array from Python scalars (i.e. np.array(...)).

    • Restructure how universal functions work (NEP 43), in order to:

      • make it possible to allow a ~numpy.ufunc such as np.add to be extended by user-defined datatypes such as Units.

      • allow efficient lookup for the correct implementation for user-defined datatypes.

      • enable reuse of existing code. Units should be able to use the normal math loops and add additional logic to determine output type.

  • Phase III: Growth of NumPy and Scientific Python Ecosystem capabilities:

    • Cleanup of legacy behaviour where it is considered buggy or undesirable.

    • Provide a path to define new datatypes from Python.

    • Assist the community in creating types such as Units or Categoricals

    • Allow strings to be used in functions such as np.equal or np.add.

    • Remove legacy code paths within NumPy to improve long term maintainability

This document serves as a basis for phase I and provides the vision and motivation for the full project. Phase I does not introduce any new user-facing features, but is concerned with the necessary conceptual cleanup of the current datatype system. It provides a more “pythonic” datatype Python type object, with a clear class hierarchy.

The second phase is the incremental creation of all APIs necessary to define fully featured datatypes and reorganization of the NumPy datatype system. This phase will thus be primarily concerned with defining an, initially preliminary, stable public API.

Some of the benefits of a large refactor may only become evident after the full deprecation of the current legacy implementation (i.e. larger code removals). However, these steps are necessary for improvements to many parts of the core NumPy API, and are expected to make the implementation generally easier to understand.

The following figure illustrates the proposed design at a high level, and roughly delineates the components of the overall design. Note that this NEP only regards Phase I (shaded area), the rest encompasses Phase II and the design choices are up for discussion, however, it highlights that the DType datatype class is the central, necessary concept:

_images/nep-0041-mindmap.svg

Backward compatibility

While the actual backward compatibility impact of implementing Phase I and II are not yet fully clear, we anticipate, and accept the following changes:

  • Python API:

    • type(np.dtype("f8")) will be a subclass of np.dtype, while right now type(np.dtype("f8")) is np.dtype. Code should use isinstance checks, and in very rare cases may have to be adapted to use it.

  • C-API:

    • In old versions of NumPy PyArray_DescrCheck is a macro which uses type(dtype) is np.dtype. When compiling against an old NumPy version, the macro may have to be replaced with the corresponding PyObject_IsInstance call. (If this is a problem, we could backport fixing the macro)

    • The UFunc machinery changes will break limited parts of the current implementation. Replacing e.g. the default TypeResolver is expected to remain supported for a time, although optimized masked inner loop iteration (which is not even used within NumPy) will no longer be supported.

    • All functions currently defined on the dtypes, such as PyArray_Descr->f->nonzero, will be defined and accessed differently. This means that in the long run lowlevel access code will have to be changed to use the new API. Such changes are expected to be necessary in very few project.

  • dtype implementors (C-API):

    • The array which is currently provided to some functions (such as cast functions), will no longer be provided. For example PyArray_Descr->f->nonzero or PyArray_Descr->f->copyswapn, may instead receive a dummy array object with only some fields (mainly the dtype), being valid. At least in some code paths, a similar mechanism is already used.

    • The scalarkind slot and registration of scalar casting will be

      removed/ignored without replacement. It currently allows partial value-based casting. The PyArray_ScalarKind function will continue to work for builtin types, but will not be used internally and be deprecated.

    • Currently user dtypes are defined as instances of np.dtype. The creation works by the user providing a prototype instance. NumPy will need to modify at least the type during registration. This has no effect for either rational or quaternion and mutation of the structure seems unlikely after registration.

Since there is a fairly large API surface concerning datatypes, further changes or the limitation certain function to currently existing datatypes is likely to occur. For example functions which use the type number as input should be replaced with functions taking DType classes instead. Although public, large parts of this C-API seem to be used rarely, possibly never, by downstream projects.

Detailed Description

This section details the design decisions covered by this NEP. The subsections correspond to the list of design choices presented in the Scope section.

Datatypes as Python Classes (1)

The current NumPy datatypes are not full scale python classes. They are instead (prototype) instances of a single np.dtype class. Changing this means that any special handling, e.g. for datetime can be moved to the Datetime DType class instead, away from monolithic general code (e.g. current PyArray_AdjustFlexibleDType).

The main consequence of this change with respect to the API is that special methods move from the dtype instances to methods on the new DType class. This is the typical design pattern used in Python. Organizing these methods and information in a more Pythonic way provides a solid foundation for refining and extending the API in the future. The current API cannot be extended due to how it is exposed publically. This means for example that the methods currently stored in PyArray_ArrFuncs on each datatype (see NEP 40) will be defined differently in the future and deprecated in the long run.

The most prominent visible side effect of this will be that type(np.dtype(np.float64)) will not be np.dtype anymore. Instead it will be a subclass of np.dtype meaning that isinstance(np.dtype(np.float64), np.dtype) will remain true. This will also add the ability to use isinstance(dtype, np.dtype[float64]) thus removing the need to use dtype.kind, dtype.char, or dtype.type to do this check.

With the design decision of DTypes as full-scale Python classes, the question of subclassing arises. Inheritance, however, appears problematic and a complexity best avoided (at least initially) for container datatypes. Further, subclasses may be more interesting for interoperability for example with GPU backends (CuPy) storing additional methods related to the GPU rather than as a mechanism to define new datatypes. A class hierarchy does provides value, this may be achieved by allowing the creation of abstract datatypes. An example for an abstract datatype would be the datatype equivalent of np.floating, representing any floating point number. These can serve the same purpose as Python’s abstract base classes.

Scalars should not be instances of the datatypes (2)

For simple datatypes such as float64 (see also below), it seems tempting that the instance of a np.dtype("float64") can be the scalar. This idea may be even more appealing due to the fact that scalars, rather than datatypes, currently define a useful type hierarchy.

However, we have specifically decided against this for a number of reasons. First, the new datatypes described herein would be instances of DType classes. Making these instances themselves classes, while possible, adds additional complexity that users need to understand. It would also mean that scalars must have storage information (such as byteorder) which is generally unnecessary and currently is not used. Second, while the simple NumPy scalars such as float64 may be such instances, it should be possible to create datatypes for Python objects without enforcing NumPy as a dependency. However, Python objects that do not depend on NumPy cannot be instances of a NumPy DType. Third, there is a mismatch between the methods and attributes which are useful for scalars and datatypes. For instance to_float() makes sense for a scalar but not for a datatype and newbyteorder is not useful on a scalar (or has a different meaning).

Overall, it seem rather than reducing the complexity, i.e. by merging the two distinct type hierarchies, making scalars instances of DTypes would increase the complexity of both the design and implementation.

A possible future path may be to instead simplify the current NumPy scalars to be much simpler objects which largely derive their behaviour from the datatypes.

C-API for creating new Datatypes (3)

The current C-API with which users can create new datatypes is limited in scope, and requires use of “private” structures. This means the API is not extensible: no new members can be added to the structure without losing binary compatibility. This has already limited the inclusion of new sorting methods into NumPy [new_sort].

The new version shall thus replace the current PyArray_ArrFuncs structure used to define new datatypes. Datatypes that currently exist and are defined using these slots will be supported during a deprecation period.

The most likely solution is to hide the implementation from the user and thus make it extensible in the future is to model the API after Python’s stable API [PEP-384]:

static struct PyArrayMethodDef slots[] = {
    {NPY_dt_method, method_implementation},
    ...,
    {0, NULL}
}

typedef struct{
  PyTypeObject *typeobj;  /* type of python scalar */
  ...;
  PyType_Slot *slots;
} PyArrayDTypeMeta_Spec;

PyObject* PyArray_InitDTypeMetaFromSpec(
        PyArray_DTypeMeta *user_dtype, PyArrayDTypeMeta_Spec *dtype_spec);

The C-side slots should be designed to mirror Python side methods such as dtype.__dtype_method__, although the exposure to Python is a later step in the implementation to reduce the complexity of the initial implementation.

C-API Changes to the UFunc Machinery (4)

Proposed changes to the UFunc machinery will be part of NEP 43. However, the following changes will be necessary (see NEP 40 for a detailed description of the current implementation and its issues):

  • The current UFunc type resolution must be adapted to allow better control for user-defined dtypes as well as resolve current inconsistencies.

  • The inner-loop used in UFuncs must be expanded to include a return value. Further, error reporting must be improved, and passing in dtype-specific information enabled. This requires the modification of the inner-loop function signature and addition of new hooks called before and after the inner-loop is used.

An important goal for any changes to the universal functions will be to allow the reuse of existing loops. It should be easy for a new units datatype to fall back to existing math functions after handling the unit related computations.

Discussion

See NEP 40 for a list of previous meetings and discussions.

Acknowledgments

The effort to create new datatypes for NumPy has been discussed for several years in many different contexts and settings, making it impossible to list everyone involved. We would like to thank especially Stephan Hoyer, Nathaniel Smith, and Eric Wieser for repeated in-depth discussion about datatype design. We are very grateful for the community input in reviewing and revising this NEP and would like to thank especially Ross Barnowski and Ralf Gommers.