Creating A New Code Frontend

yt is designed to support analysis and visualization of data from multiple different simulation codes. For a list of codes and the level of support they enjoy, see Code Support.

We’d like to support a broad range of codes, both Adaptive Mesh Refinement (AMR)-based and otherwise. To add support for a new code, a few things need to be put into place. These necessary structures can be classified into a couple categories:

  • Data meaning: This is the set of parameters that convert the data into physically relevant units; things like spatial and mass conversions, time units, and so on.

  • Data localization: These are structures that help make a “first pass” at data loading. Essentially, we need to be able to make a first pass at guessing where data in a given physical region would be located on disk. With AMR data, this is typically quite easy: the grid patches are the “first pass” at localization.

  • Data reading: This is the set of routines that actually perform a read of either all data in a region or a subset of that data.

Note that a frontend can be built as an external package. This is useful to develop and maintain a maturing frontend at your own pace. For technical details, see Frontends as extensions.

If you are interested in adding a new code, be sure to drop us a line on yt-dev!

Bootstrapping a new frontend

To get started

  • make a new directory in yt/frontends with the name of your code and add the name into yt/frontends/api.py:_frontends (in alphabetical order).

  • copy the contents of the yt/frontends/_skeleton directory, and replace every occurrence of Skeleton with your frontend’s name (preserving case). This adds a lot of boilerplate for the required classes and methods that are needed.

Data Meaning Structures

You will need to create a subclass of Dataset in the data_structures.py file. This subclass will need to handle conversion between the different physical units and the code units (typically in the _set_code_unit_attributes() method), read in metadata describing the overall data on disk (via the _parse_parameter_file() method), and provide a classmethod called _is_valid() that lets the yt.load method help identify an input file as belonging to this particular Dataset subclass (see How to make yt.load magically detect your data format ?). For the most part, the examples of yt.frontends.boxlib.data_structures.OrionDataset and yt.frontends.enzo.data_structures.EnzoDataset should be followed, but yt.frontends.chombo.data_structures.ChomboDataset, as a slightly newer addition, can also be used as an instructive example.

A new set of fields must be added in the file fields.py in your new directory. For the most part this means subclassing FieldInfoContainer and adding the necessary fields specific to your code. Here is a snippet from the base BoxLib field container:

from yt.fields.field_info_container import FieldInfoContainer


class BoxlibFieldInfo(FieldInfoContainer):
    known_other_fields = (
        ("density", (rho_units, ["density"], None)),
        ("eden", (eden_units, ["energy_density"], None)),
        ("xmom", (mom_units, ["momentum_x"], None)),
        ("ymom", (mom_units, ["momentum_y"], None)),
        ("zmom", (mom_units, ["momentum_z"], None)),
        ("temperature", ("K", ["temperature"], None)),
        ("Temp", ("K", ["temperature"], None)),
        ("x_velocity", ("cm/s", ["velocity_x"], None)),
        ("y_velocity", ("cm/s", ["velocity_y"], None)),
        ("z_velocity", ("cm/s", ["velocity_z"], None)),
        ("xvel", ("cm/s", ["velocity_x"], None)),
        ("yvel", ("cm/s", ["velocity_y"], None)),
        ("zvel", ("cm/s", ["velocity_z"], None)),
    )

    known_particle_fields = (
        ("particle_mass", ("code_mass", [], None)),
        ("particle_position_x", ("code_length", [], None)),
        ("particle_position_y", ("code_length", [], None)),
        ("particle_position_z", ("code_length", [], None)),
        ("particle_momentum_x", (mom_units, [], None)),
        ("particle_momentum_y", (mom_units, [], None)),
        ("particle_momentum_z", (mom_units, [], None)),
        ("particle_angmomen_x", ("code_length**2/code_time", [], None)),
        ("particle_angmomen_y", ("code_length**2/code_time", [], None)),
        ("particle_angmomen_z", ("code_length**2/code_time", [], None)),
        ("particle_id", ("", ["particle_index"], None)),
        ("particle_mdot", ("code_mass/code_time", [], None)),
    )

The tuples, known_other_fields and known_particle_fields contain entries, which are tuples of the form ("name", ("units", ["fields", "to", "alias"], "display_name")). "name" is the name of a field stored on-disk in the dataset. "units" corresponds to the units of that field. The list ["fields", "to", "alias"] allows you to specify additional aliases to this particular field; for example, if your on-disk field for the x-direction velocity were "x-direction-velocity", maybe you’d prefer to alias to the more terse name of "xvel". By convention in yt we use a set of “universal” fields. Currently these fields are enumerated in the stream frontend. If you take a look at yt/frontends/stream/fields.py, you will see a listing of fields following the format described above with field names that will be recognized by the rest of the built-in yt field system. In the example from the boxlib frontend above many of the fields in the known_other_fields tuple follow this convention. If you would like your frontend to mesh nicely with the rest of yt’s built-in fields, it is probably a good idea to alias your frontend’s field names to the yt “universal” field names. Finally, “display_name”`` is an optional parameter that can be used to specify how you want the field to be displayed on a plot; this can be LaTeX code, for example the density field could have a display name of r"\rho". Omitting the "display_name" will result in using a capitalized version of the "name".

How to make yt.load magically detect your data format ?

yt.load takes in a file or directory name, as well as any number of positional and keyword arguments. On call, yt.load attempts to determine what Dataset subclasses are compatible with the set of arguments it received. It does so by passing its arguments to every Dataset subclasses’ _is_valid method. These methods are intended to be heuristics that quickly determine whether the arguments (in particular the file/directory) can be loaded with their respective classes. In some cases, more than one class might be detected as valid. If all candidate classes are siblings, yt.load will select the most specialized one.

When writing a new frontend, it is important to write _is_valid methods to be as specific as possible, otherwise one might constrain the design space for future frontends or in some cases deny their ability to leverage yt.load’s magic.

Performance is also critical since the new method is going to get called every single time along with yt.load, even for unrelated data formats.

Note that yt.load knows about every Dataset subclass because they are automatically registered on creation.

Creating Aliases for Magnetic Fields

Setting up access to the magnetic fields in your dataset requires special handling, because in different unit systems magnetic fields have different dimensions (see Magnetic Fields for an explanation). If your dataset includes magnetic fields, you should include them in known_other_fields, but do not set up aliases for them–instead use the special handling function setup_magnetic_field_aliases(). It takes as arguments the FieldInfoContainer instance, the field type of the frontend, and the list of magnetic fields from the frontend. Here is an example of how this is implemented in the FLASH frontend:

class FLASHFieldInfo(FieldInfoContainer):
    known_other_fields = (
        ("magx", (b_units, [], "B_x")),  # Note there is no alias here
        ("magy", (b_units, [], "B_y")),
        ("magz", (b_units, [], "B_z")),
        ...,
    )

    def setup_fluid_fields(self):
        from yt.fields.magnetic_field import setup_magnetic_field_aliases

        ...
        setup_magnetic_field_aliases(self, "flash", ["mag%s" % ax for ax in "xyz"])

This function should always be imported and called from within the setup_fluid_fields method of the FieldInfoContainer. If this function is used, converting between magnetic fields in different unit systems will be handled automatically.

Data Localization Structures

These functions and classes let yt know about how the arrangement of data on disk corresponds to the physical arrangement of data within the simulation. yt has grid datastructures for handling both patch-based and octree-based AMR codes. The terms ‘patch-based’ and ‘octree-based’ are used somewhat loosely here. For example, traditionally, the FLASH code used the paramesh AMR library, which is based on a tree structure, but the FLASH frontend in yt utilizes yt’s patch-based datastructures. It is up to the frontend developer to determine which yt datastructures best match the datastructures of their simulation code.

Both approaches – patch-based and octree-based – have a concept of a Hierarchy or Index (used somewhat interchangeably in the code) of datastructures and something that describes the elements that make up the Hierarchy or Index. For patch-based codes, the Index is a collection of AMRGridPatch objects that describe a block of zones. For octree-based codes, the Index contains datastructures that hold information about the individual octs, namely an OctreeContainer.

Hierarchy or Index

To set up data localization, a GridIndex subclass for patch-based codes or an OctreeIndex subclass for octree-based codes must be added in the file data_structures.py. Examples of these different types of Index can be found in, for example, the yt.frontends.chombo.data_structures.ChomboHierarchy for patch-based codes and yt.frontends.ramses.data_structures.RAMSESIndex for octree-based codes.

For the most part, the GridIndex subclass must override (at a minimum) the following methods:

  • _detect_output_fields(): self.field_list must be populated as a list of strings corresponding to “native” fields in the data files.

  • _count_grids(): this must set self.num_grids to be the total number of grids (equivalently AMRGridPatch’es) in the simulation.

  • _parse_index(): this must fill in grid_left_edge, grid_right_edge, grid_particle_count, grid_dimensions and grid_levels with the appropriate information. Each of these variables is an array, with an entry for each of the self.num_grids grids. Additionally, grids must be an array of AMRGridPatch objects that already know their IDs.

  • _populate_grid_objects(): this initializes the grids by calling _prepare_grid() and _setup_dx() on all of them. Additionally, it should set up Children and Parent lists on each grid object.

The OctreeIndex has somewhat analogous methods, but often with different names; both OctreeIndex and GridIndex are subclasses of the Index class. In particular, for the OctreeIndex, the method _initialize_oct_handler() setups up much of the oct metadata that is analogous to the grid metadata created in the GridIndex methods _count_grids(), _parse_index(), and _populate_grid_objects().

Grids

Note

This section only applies to the approach using yt’s patch-based datastructures. For the octree-based approach, one does not create a grid object, but rather an OctreeSubset, which has methods for filling out portions of the octree structure. Again, see the code in yt.frontends.ramses.data_structures for an example of the octree approach.

A new grid object, subclassing AMRGridPatch, will also have to be added in data_structures.py. For the most part, this may be all that is needed:

class ChomboGrid(AMRGridPatch):
    _id_offset = 0
    __slots__ = ["_level_id"]

    def __init__(self, id, index, level=-1):
        AMRGridPatch.__init__(self, id, filename=index.index_filename, index=index)
        self.Parent = None
        self.Children = []
        self.Level = level

Even one of the more complex grid objects, yt.frontends.boxlib.BoxlibGrid, is still relatively simple.

Data Reading Functions

In io.py, there are a number of IO handlers that handle the mechanisms by which data is read off disk. To implement a new data reader, you must subclass BaseIOHandler. The various frontend IO handlers are stored in an IO registry - essentially a dictionary that uses the name of the frontend as a key, and the specific IO handler as a value. It is important, therefore, to set the dataset_type attribute of your subclass, which is what is used as the key in the IO registry. For example:

class IOHandlerBoxlib(BaseIOHandler):
    _dataset_type = "boxlib_native"
    ...

At a minimum, one should also override the following methods

  • _read_fluid_selection(): this receives a collection of data “chunks”, a selector describing which “chunks” you are concerned with, a list of fields, and the size of the data to read. It should create and return a dictionary whose keys are the fields, and whose values are numpy arrays containing the data. The data should actually be read via the _read_chunk_data() method.

  • _read_chunk_data(): this method receives a “chunk” of data along with a list of fields we want to read. It loops over all the grid objects within the “chunk” of data and reads from disk the specific fields, returning a dictionary whose keys are the fields and whose values are numpy arrays of the data.

If your dataset has particle information, you’ll want to override the _read_particle_coords() and read_particle_fields() methods as well. Each code is going to read data from disk in a different fashion, but the yt.frontends.boxlib.io.IOHandlerBoxlib is a decent place to start.

And that just about covers it. Please feel free to email yt-users or yt-dev with any questions, or to let us know you’re thinking about adding a new code to yt.

How to add extra dependencies ?

Note

This section covers the technical details of how optional runtime dependencies are implemented and used in yt. If your frontend has specific or complicated dependencies other than yt’s, we advise writing your frontend as an extension package Frontends as extensions

It is required that a specific target be added to pyproject.toml to define a list of additional requirements (even if empty), see Additional requirements (pip only).

At runtime, extra third party dependencies should be loaded lazily, meaning their import needs to be delayed until actually needed. This is achieved by importing a wrapper from yt.utitilies.on_demand_imports.py, instead of the actual package like so

from yt.utilities.on_demand_imports import _mypackage as mypackage

Such import statements can live at the top of a module without generating overhead or errors in case the actual package isn’t installed.

If the extra third party dependency is new, a new import wrapper must also be added. To do so, follow the example of the existing wrappers in yt.utilities.on_demand_imports.py.