Skip to content

Latest commit

 

History

History
73 lines (50 loc) · 4.19 KB

python.md

File metadata and controls

73 lines (50 loc) · 4.19 KB
Raw

Python API

The Python API is found under the python/ directory. It provides convinient wrapper objects for the low-level C API, as well as utilities for building Weld computations and composing Python libraries.

To use the Python API in 'development' mode, run the following from $WELD_HOME/python:

$ cd pyweld;  python setup.py develop; cd ..
$ cd grizzly; python setup.py develop; cd ..

Otherwise, run the following from $WELD_HOME/python:

$ cd pyweld;  python setup.py install; cd ..
$ cd grizzly; python setup.py install; cd ..

Alternatively, you can install our pre-built Python modules by running,

$ pip install pyweld
$ pip install pygrizzly

You should also follow the setup instructions here (in particular, make sure WELD_HOME is set so the libraries Weld uses can be found).

Bindings

The weld.bindings module contains bindings for the C API. Each type in the C API is wrapped as a Python object. Methods on the Python objects call the corresponding C API functions.

As an example, the code below creates a new WeldConf, sets a value on it, and then gets the value back:

>>> import weld.bindings as bnd
>>> conf = bnd.WeldConf() # calls weld_conf_new()
>>> conf.set("myKey", "myValue") # calls weld_conf_set(...)
>>> print conf.get("myKey") # calls weld_conf_get(...)
"myValue"

WeldObject API

The WeldObject API is included in the weld.weldobject module, and an example of how to use it is described here.

This API provides an interface for composing Python programs by building a lazily evaluated Weld computation. The WeldObject tracks values it operates over (called "dependencies") and builds and runs a runnable Weld module using a evaluate method. Dependencies are added using the update function. This function requires a Weld type for the dependency being added; types are available in the weld.types module.

The table below describes the WeldObject API in brief:

Method/Field Description
update(value, ty) Adds value (which has Weld type ty) in Weld as a dependency. Returns a string name which can be used in the object's Weld code to refer to this value.
evaluate(ty) Evaluates the object and returns a value. ty is the expected Weld type of the return value.
weld_code A string field representing the Weld IR for this object. This string is modified to register a computation with this object. See this document for a description of the language.

The general usage pattern for a WeldObject is to initialize it, add some dependencies and Weld code to register a computation, and then evaluate it to get a return value. Here's an example, where we add two numbers:

>>> import weld.weldobject as wo
>>> import weld.encoders as enc
>>> obj = wo.WeldObject(enc.WeldScalarEncoder(), enc.WeldScalarDecoder()) # See more about encoders below
>>> name1 = obj.update(1, WeldI32())
>>> name2 = obj.update(2, WeldI32())
>>> obj.weld_code = name1 + " + " + name2 # Weld IR to add two numbers.

Encoders and Decoders

When data is passed into Weld, it must be marshalled into a binary format which Weld understands (these formats are described in the C API doc. In general, values are formatted using C scalars and structs; Python's ctypes module allows constructing these kinds of representations.

To support custom formats, the WeldObject API takes an encoder, which allows encoding a Python object as a Weld object, and a decoder, which allows decoding a Weld object into a Python object. These encoders and decoders are interfaces which must be implemented by a library writer.

Weld provides some commonly used encoders and decoders in the weld.encoders module. NumPy arrays, for example, are a common way to represent C-style arrays in Python. Weld thus includes a WeldNumPyEncoder and WeldNumPyDecoder class to marshall 1-dimensional NumPy arrays.