This is an interface between Dyalog APL and Python. It allows Python code to be accessed from APL, and vice versa.

• Dyalog APL version 16.0 Unicode
• Python 2.7.9 or higher, or Python 3.4 or higher.

The APL side of the interface is located in Py.dyalog. It can be loaded into the workspace using:

]load Py

Note that it expects the included Python scripts to be in the same directory as the Py.dyalog file.

To start a Python interpreter, make a new instance of the Py.Py class. This will start a Python instance in the background, and connect to it. On Unix, this is done using two pipes; on Windows this is done using a TCP connection.

py ← ⎕NEW Py.Py

The resulting object can be used to interact with the Python interpreter. Once the object is destroyed, the Python interpreter associated with it will also be shut down.

There are several different options that can be given to the Py class, namely:

In particular, the following might be of interest:

py ← ⎕NEW Py.Py('Version' 2) ⍝ use Python 2 instead of 3

⍝ start a Blender instance and control that instead of a normal Python
⍝ (if on Windows, you have to pass in the absolute path to blender.exe instead)
py ← ⎕NEW Py.Py (('PyPath' 'blender') ('ArgFmt' '-P "⍎" -- → ← thread') ('ForceTCP' 1))

The Exec function can be used to run one or more Python statements. It takes one string, which may have newlines in it. The Py.ScriptFollows function can be used to help load scripts.

⍝ run one statement
py.Exec 'import antigravity'

⍝ run a script
py.Exec 'def foobar():',(⎕UCS 10),'  return "abc"'

⍝ or (in a tradfn):
py.Exec #.Py.ScriptFollows
⍝ def foobar():
⍝    return "abc"

An APL object will be available to the Python code, in order for it to call back into the APL code. (See Accessing APL from Python for more information.)

The Eval function can be used to evaluate a Python expression. It takes as its left argument the Python expression to be evaluated, and as its right argument a vector of APL arguments to be substituted into it. Inside the Python expression, the quad (⎕) or the quote-quad (⍞) can be used to refer to these arguments. If the quad is used, the argument will be converted to a (hopefully) suitable Python representation first; if the quote-quad is used, the argument will be exposed on the Python side as an APLArray object. (See Data conversion for more information.)

If the Python expression returns something other than an APLArray, it will be converted back into a suitable APL form before being sent back to APL.

     ⍝ access a variable
     py.Eval '__name__'
pynapl.PyEvaluator

     ⍝ add two numbers
     '⎕+⎕' py.Eval 2 2
4

     ⍝ this is equivalent to ⍴X
     '⍞.rho' py.Eval ⊂5 5⍴⎕A
5 5

     ⍝ round trip
     'APL.eval("2+2")' py.Eval ⍬
4

     ⍝ set a variable on the Python side
     'x' py.Set 42
     py.Eval 'x'
42

     ⍝ alternate syntax when there are no arguments
     py.Eval 'APL.eval("2+2")' 
4

Just as with Exec, an APL object will be made available to the Python expression.

It is also possible to 'import' a Python function to the APL workspace. The PyFn function can be used to create APL functions that call Python functions automatically.

The PyFn function returns a namespace containing two functions, Call and CallVec.

• CallVec takes a vector of arguments as its right argument, and passes those into the Python function. It takes an optional boolean vector as its left argument, which describes whether or not to convert the arguments.

• Call is a "normal" APL function. If used monadically, it calls the Python function with one argument (f(⍵)); if used dyadically it calls the Python function with two arguments (f(⍺,⍵)). The arguments are always converted.

The namespace also includes a reference to the Py object that created it, so it will not be destroyed until such functions themselves are.

Example:

⍝ import a Python module
py.Exec 'import webbrowser'

⍝ define a function from it in APL
⍝ this one handily takes only one argument so can be used monadically
showPage←(py.PyFn 'webbrowser.open').Call

⍝ this will now show a web page
showPage 'https://www.dyalog.com'

By default, Python objects that have APL equivalents are automatically converted. E.g., a Python list becomes an APL vector. (See the "Data Conversion" section.)

Python objects that do not have such equivalents are sent as references instead, which can be used on the APL side to access their attributes. On the APL side, a stub class will be instantiated which will have attributes corresponding to the Python ones.

      py.Exec'import sys'
      sys←py.Eval'sys'
      sys.version_info
3 5 3  final  0

      5↑sys.⎕NL¯2
 __doc__  __name__  __package__  __stderr__  __stdin__ 

Fields are exposed on the APL side by means of properties, which can be used to set and retrieve the values, and methods are exposed as functions which can be called:

      os←py.Import'os' ⍝ convenience functions
      +os.getpid ⍬
17906

Such functions return a shy result, and take a right argument consisting of a vector of positional arguments, and an optional left argument representing the keyword arguments. This left argument may either be a namespace or a list of key-value pairs.

      json←py.Import'json'
      +(⊂'separators' '--')json.dumps ⊂1 2 3 4
[1-2-3-4]

It is also possible to send these references back to Python and interact with them there:

      '⎕.getpid()' py.Eval os
17906

The resulting classes cannot be instantiated from APL using ⎕NEW, they can only be instantiated by calling the Python constructors. The objects keep a reference to the Py instance that created them, which means the Python interpreter will stay alive as long as any of its objects are still around. On the Python side, the objects are stored by the interface, and released when all APL references to them have been removed.

In some instances, some name mangling is required. Variables are exposed as properties on the APL side, which means that a variable foo will conflict with functions named get_foo and set_foo. In this case, those functions will be renamed ⍙get_foo and ⍙set_foo.

In APL, a Python object reference will have the following members:

• Properties:
  • foo: for each non-callable attribute foo in the object, gets or sets that attribute
  • ∆foo: for each non-callable attribute foo in the object, gets that attribute without object translation (i.e., will always return a Python object, even for lists and numbers).
• Functions:

  • bar: for each callable attribute bar in the object, calls it and returns the result.

  • ∆bar: for each callable attribute bar in the object, calls it and returns the result without object translation.

  • ⍙DF: returns the string representation of the object (using repr), and also sets the display form to it.

  • ⍙Get: retrieves an attribute, given as a character vector as the right argument

  • ⍙Get∆: retrieves an attribute, given as a character vector as the right argument, without object translation on the result.

  • ⍙Set: sets the attribute (left argument) to the value given as the right argument

  • ⍙SetRaw: sets the attribute (left argument) to the value given as the right argument, using ⍞ instead of ⎕.

  • ⍙Call: calls the function (left argument) with the given positional and keyword arguments (right argument, both must be present).

  • ⍙Call∆: like ⍙Call, but without object translation on the result.

If the Python code raises an exception, the interface will signal a DOMAIN ERROR. ⎕DMX.Message will contain the string representation of the Python exception.

The APL.py module contains a function that will start an APL interpreter. Just like the APL side, it expects the Py.dyalog script to be in the same directory.

An APL object can be obtained using the APL.APL function. This will start a Dyalog instance in the background and connect to it.

from pynapl import APL
apl = APL.APL()

An optional dyalog argument may be given to the APL function, to specify the path to the dyalog interpreter. If it is not given, on Unix the dyalog interpreter on the path will be used, on Windows the registry will be consulted. The Dyalog instance will be shut down once the apl object is destroyed.

The fix function takes a string, which will be 2⎕FIX'ed on the APL side. This can be used to load large amounts of APL code into the interpreter.

apl.fix("""
:Namespace Test
foo←42
:EndNamespace
""")

The eval function takes a string, which will be evaluated using ⍎ on the APL side. Any extra arguments passed into eval will be put into a vector and exposed as ∆ on the APL side, and a py object will be available for the APL code to communicate back to the Python interpreter. (See Accessing Python from APL.)

This is a relatively low-level function, and it is probably better to use fn and op.

Conversion of data to APL types is done automatically. (Anything that's not an APLArray is converted.) The result of the evaluation is converted back to the Python format unless raw is set.

>>> apl.eval("2+2")
4
>>> apl.eval("⎕A")
u'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
>>> apl.eval("⎕A", raw=True)
<Array.APLArray object at 0x7fb704e36310>
>>> apl.eval("'2+2' py.Eval ⍬") # round trip
4

The fn function can be used to import an APL function to Python. The function may be niladic (if called with no arguments), monadic (if called with one argument), or dyadic (if called with two).

As with eval, a named argument raw can be set to prevent automatic data conversion.

>>> aplsum = apl.fn("+/")
>>> aplsum([1,2,3,4,5])
15
>>> aplsum(3, [1,2,3,4,5])
[6, 9, 12]

The function may be an anonymous dfn and may contain newlines. It may not be a definition of a tradfn (those can be defined using fix or tradfn, then referred to by name using fn).

>>> factorial = apl.fn("""
{ ⍵≤0:1
  ⍵×∇⍵-1
}
""")
>>> factorial(5)
120

Apart from using fix, a tradfn can also be defined using the tradfn function:

>>> foo = apl.tradfn("""
r←foo x
r←x+x
""")
>>> foo(5)
10
>>> apl.fn("foo")(5)
10

Python does not make the difference between functions and operators that APL makes. Therefore, APL operators can be exposed as Python functions using the op function.

>>> scan = apl.op("\\") # note the extra backslash for escaping

The operator is then exposed as a Python function, which takes one or two arguments, depending on whether the operator is monadic or dyadic.

The arguments may be either values or Python functions. If you want to pass an APL function to an APL operator via Python, you must first import the APL function using fn.

>>> apl_add = apl.fn("+")
>>> py_add = lambda x, y: x+y
>>> apl_sumscan = scan(apl_add) # equivalent to "+\"
>>> py_sumscan = scan(py_add)   # uses the Python "+"
>>> apl_sumscan([1,2,3])
[1, 3, 6]
>>> py_sumscan([1,2,3])
[1, 3, 6]

If an APL operator is applied to an APL function via Python, as in the apl_sumscan example, this is detected, and the application is done in APL without calling back into Python.

Just as APL may make use of Python objects, Python may make use of APL objects.

If a call to eval or to an APL function returns an instance of an APL object, it is represented on the Python side by an instance of the APLObject class, in which public functions and variables will appear as attributes.

>>> apl.fix("""
... :Class foo
...    :Field Public n←0
...    ∇x←getN
...       :Access Public
...       x←n
...    ∇
...    ∇setN x
...       :Access Public
...       n←x
...    ∇
... :EndClass
... """)
['foo']
>>> a = apl.eval("+a ← ⎕NEW foo")
>>> a
<pynapl.Array.APLObject object at 0x7f1d3a9e0ac8>
>>> a.n
0
>>> a.getN()
0
>>> a.setN(20)
[]
>>> a.n
20
>>> a.n = 30
>>> a.getN()
30

The APLObject class contains a reference to the object on the APL side, so changes to its state are reflected on the APL side, and vice versa:

>>> apl.eval("a.n")
30
>>> apl.eval("a.n←40")
40
>>> a.n
40

Members whose names are not valid names in Python can be accessed via getattr and setattr. Python 2 requires attribute names to be ASCII only, so members whose names contain non-ASCII characters cannot be accessed. Python 3 does not have this limitation.

If a signal is raised by the APL code, an APLError will be raised on the Python side. The exception object will contain a dmx field, which is a dictionary that contains the fields from ⎕DMX.

When an interrupt is raised, the message will be "Interrupt" and dmx will be None.

• Numbers (any kind) or boolean: number
• One-character strings: characters
• Other string: character vector
• List or tuple: vector
• NoneType: empty numeric vector
• Dictionary: namespace

In addition, any kind of iterable object (objects that are instances of collections.Iterable or that implement __iter__, and objects that implement both __len__ and __getitem__) will be iterated over, and the results sent as a vector to APL. This allows for most kinds of custom container objects to be used.

NOTE: if the object is an infinite generator, it will cause a hang.

If the numpy library is available, numpy matrices will be automatically converted to APL matrices.

If the object is none of these, an object reference will be sent to APL, where it can be used to access its attributes. Python code can also send an object reference explicitly by using the apl.obj function.

     py.Eval 'sys' ⍝ ask for module object
#.Py.⍙PythonObject.[module]
     py.Eval '[1,2,3,4]' ⍝ send a list
1 2 3 4
     py.Eval 'apl.obj([1,2,3,4])' ⍝ send a list as an object
#.Py.⍙PythonObject.[list]

• Numbers: int, long, or float, depending on which fits best
• Simple (non-nested) character vector: Unicode string
• Numeric vector / nested vector: List
• Higher-rank array: nested list (the equivalent of {↓⍣((⊃⍴⍴⍵)-1)⊢⍵} is done).
• Namespace containing values: dictionary

This is a class on the Python side that can be used to communicate with APL without going through the conversion. It is a multidimensional array, which may contain nested APLArray objects.

An APLArray object can be indexed using a list or a tuple. The index should be given as if ⎕IO=0, e.g.:

>>> foo = apl.eval("5 5⍴⎕A", raw=True)
>>> foo.rho
[5, 5]
>>> foo[2,3]  # ⎕IO←0 ⋄ (5 5⍴⎕A)[2;3]
u'N'

Assignment to individual items is possible in the same manner. Conversion will be done automatically if it is necessary. An IndexError will be raised if the coordinates are out of range.

On Unix, there are two ways in which the connection between APL and Python can be made.

1. The default way is by using two named pipes, which the initiating side will create (using mkfifo) and pass to the client program.

2. It can also be set to use a TCP connection. The initiating side will start up a TCP server (using Conga on the APL side) on an unused port, and listen for a connection from the client side.

On Windows, only TCP mode is supported. On Unix, TCP mode may be necessary to use non-standard interpreters (Blender in particular does not like pipes much). TCP mode has about twice the latency as pipe mode.

The both programs communicate by sending each other messages, as described below.

The underlying format used for messages consists of a 5-byte header and then a body.

The first byte of the header denotes the message type, the next four give the length of the body in bytes (high-endian).

The contents of the body are UTF-8 encoded text, usually JSON.

The EVAL message is a JSON list containing two elements. The first element should be a string containing the expression to evaluate, the second element should be a (possibly empty) list of arguments.

The ERR message is a JSON dictionary containing at least a Message field, which contains the error message. Errors coming from APL may also contain a DMX field, which contains the JSON representation of Dyalog APL's ⎕DMX object.

The message handling code on both sides supports handling messages while waiting for the result of another. E.g., if an EVAL is sent, it may cause another EVAL to be sent back, which will then be handled before the corresponding EVALRET is received. This way, the evaluation of an expression may switch back and forth between the two sides as needed.

Example:

Python side:
>>> x = apl.eval("2+'2+2' py.Eval ⍬")
 # Python sends to APL: EVAL '2+2' py.Eval ⍬
 
APL side:
     2+'2+2' py.Eval ⍬
 ⍝ APL sends back to Python: EVAL 2+2

Python side:
>>> 2+2
 # this evaluates to 4
 # Python sends to APL: EVALRET 4

APL side:
 ⍝ receives EVALRET 4
     2 + 4
 ⍝ this evaluates to 6
 ⍝ APL sends to Python: EVALRET 6

Python side:
 # receives EVALRET 6
 # the final answer is 6 
>>> x
6

To run unit tests from Python:

$ python2 -m unittest pynapl.test.test_APL pynapl.test.test_Array
$ python3 -m unittest pynapl.test.test_APL pynapl.test.test_Array

To run unit tests from APL:

      ]load pynapl/Py       
      ]load pynapl/test/Test
      Test.RunTests 2 ⍝ test with Python 2
      Test.RunTests 3 ⍝ test with Python 3

There is also an interactive GUI test that uses TkInter:

      ]load pynapl/Py
      ]load pynapl/PyTest
      PyTest.PyTest