Skip to content
/ lupa Public
forked from scoder/lupa

Lua in Python + automatic conversion of numpy arrays to a luajit ctype array

License

Notifications You must be signed in to change notification settings

cstraehl/lupa

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Lupa

Lupa integrates the LuaJIT2 runtime into CPython. It is a partial rewrite of LunaticPython in Cython with some additional features such as proper coroutine support.

For questions not answered here, please contact the Lupa mailing list.

  • separate Lua runtime states through a LuaRuntime class
  • Python coroutine wrapper for Lua coroutines
  • iteration support for Python objects in Lua and Lua objects in Python
  • proper encoding and decoding of strings (configurable per runtime, UTF-8 by default)
  • frees the GIL and supports threading in separate runtimes when calling into Lua
  • supports Python 2.x and 3.x, potentially starting with Python 2.3 (currently untested)
  • written for LuaJIT2 (tested with LuaJIT 2.0.0-beta8), but reportedly works with the normal Lua interpreter (5.1+)
  • easy to hack on and extend as it is written in Cython, not C

It complements Python very well. Lua is a language as dynamic as Python, but LuaJIT compiles it to very fast machine code, sometimes faster than many other compiled languages for computational code. The language runtime is extremely small and carefully designed for embedding. The complete binary module of Lupa, including a statically linked LuaJIT2 runtime, is only some 500KB on a 64 bit machine.

However, the Lua ecosystem lacks many of the batteries that Python readily includes, either directly in its standard library or as third party packages. This makes real-world Lua applications harder to write than equivalent Python applications. Lua is therefore not commonly used as primary language for large applications, but it makes for a fast, high-level and resource-friendly backup language inside of Python when raw speed is required and the edit-compile-run cycle of binary extension modules is too heavy and too static for agile development or hot-deployment.

Lupa is a very fast and thin wrapper around LuaJIT. It makes it easy to write dynamic Lua code that accompanies dynamic Python code by switching between the two languages at runtime, based on the tradeoff between simplicity and speed.

>>> import lupa
>>> from lupa import LuaRuntime
>>> lua = LuaRuntime()

>>> lua.eval('1+1')
2

>>> lua_func = lua.eval('function(f, n) return f(n) end')

>>> def py_add1(n): return n+1
>>> lua_func(py_add1, 2)
3

>>> lua.eval('python.eval(" 2 ** 2 ")') == 4
True
>>> lua.eval('python.builtins.str(4)') == '4'
True

Python objects are either converted when passed into Lua (e.g. numbers and strings) or passed as wrapped object references.

>>> lua_type = lua.globals().type   # Lua's type() function
>>> lua_type(1) == 'number'
True
>>> lua_type('abc') == 'string'
True

Wrapped Lua objects get unwrapped when they are passed back into Lua, and arbitrary Python objects get wrapped in different ways:

>>> lua_type(lua_type) == 'function'  # unwrapped Lua function
True
>>> lua_type(eval) == 'userdata'      # wrapped Python function
True
>>> lua_type([]) == 'userdata'        # wrapped Python object
True

Lua supports two main protocols on objects: calling and indexing. It does not distinguish between attribute access and item access like Python does, so the Lua operations obj[x] and obj.x both map to indexing. To decide which Python protocol to use for Lua wrapped objects, Lupa employs a simple heuristic.

Pratically all Python objects allow attribute access, so if the object also has a __getitem__ method, it is preferred when turning it into an indexable Lua object. Otherwise, it becomes a simple object that uses attribute access for indexing from inside Lua.

Obviously, this heuristic will fail to provide the required behaviour in many cases, e.g. when attribute access is required to an object that happens to support item access. To be explicit about the protocol that should be used, Lupa provides the helper functions as_attrgetter() and as_itemgetter() that restrict the view on an object to a certain protocol, both from Python and from inside Lua:

>>> lua_func = lua.eval('function(obj) return obj["get"] end')
>>> d = {'get' : 'got'}

>>> value = lua_func(d)
>>> value == 'got'
True

>>> dict_get = lua_func( lupa.as_attrgetter(d) )
>>> dict_get('get') == 'got'
True

>>> lua_func = lua.eval(
...     'function(obj) return python.as_attrgetter(obj)["get"] end')
>>> dict_get = lua_func(d)
>>> dict_get('get') == 'got'
True

Note that unlike Lua function objects, callable Python objects are indexable:

>>> def py_func(): pass
>>> py_func.ATTR = 2
>>> lua_func = lua.eval('function(obj) return obj.ATTR end')
>>> lua_func(py_func)
2
>>> lua_func = lua.eval(
...     'function(obj) return python.as_attrgetter(obj).ATTR end')
>>> lua_func(py_func)
2
>>> lua_func = lua.eval(
...     'function(obj) return python.as_attrgetter(obj)["ATTR"] end')
>>> lua_func(py_func)
2

Iteration over Python objects from Lua's for-loop is fully supported. However, Python iterables need to be converted using one of the utility functions which are described here. This is similar to the functions like pairs() in Lua.

To iterate over a plain Python iterable, use the python.iter() function. For example, you can manually copy a Python list into a Lua table like this:

>>> lua_copy = lua.eval('''
...     function(L)
...         local t, i = {}, 1
...         for item in python.iter(L) do
...             t[i] = item
...             i = i + 1
...         end
...         return t
...     end
... ''')

>>> table = lua_copy([1,2,3,4])
>>> len(table)
4
>>> table[1]   # Lua indexing
1

Python's enumerate() function is also supported, so the above could be simplified to:

>>> lua_copy = lua.eval('''
...     function(L)
...         local t = {}
...         for index, item in python.enumerate(L) do
...             t[ index+1 ] = item
...         end
...         return t
...     end
... ''')

>>> table = lua_copy([1,2,3,4])
>>> len(table)
4
>>> table[1]   # Lua indexing
1

For iterators that return tuples, such as dict.iteritems(), it is convenient to use the special python.iterex() function that automatically explodes the tuple items into separate Lua arguments:

>>> lua_copy = lua.eval('''
...     function(d)
...         local t = {}
...         for key, value in python.iterex(d.items()) do
...             t[key] = value
...         end
...         return t
...     end
... ''')

>>> d = dict(a=1, b=2, c=3)
>>> table = lua_copy( lupa.as_attrgetter(d) )
>>> table['b']
2

Note that accessing the d.items method from Lua requires passing the dict as attrgetter. Otherwise, attribute access in Lua would use the getitem protocol of Python dicts.

Lua tables mimic Python's mapping protocol. For the special case of array tables, Lua automatically inserts integer indices as keys into the table. Therefore, indexing starts from 1 as in Lua instead of 0 as in Python. For the same reason, negative indexing does not work. It is best to think of Lua tables as mappings rather than arrays, even for plain array tables.

>>> table = lua.eval('{10,20,30,40}')
>>> table[1]
10
>>> table[4]
40
>>> list(table)
[1, 2, 3, 4]
>>> list(table.values())
[10, 20, 30, 40]
>>> len(table)
4

>>> mapping = lua.eval('{ [1] = -1 }')
>>> list(mapping)
[1]

>>> mapping = lua.eval('{ [20] = -20; [3] = -3 }')
>>> mapping[20]
-20
>>> mapping[3]
-3
>>> sorted(mapping.values())
[-20, -3]
>>> sorted(mapping.items())
[(3, -3), (20, -20)]

>>> mapping[-3] = 3     # -3 used as key, not index!
>>> mapping[-3]
3
>>> sorted(mapping)
[-3, 3, 20]
>>> sorted(mapping.items())
[(-3, 3), (3, -3), (20, -20)]

A lookup of nonexisting keys or indices returns None (actually nil inside of Lua). A lookup is therefore more similar to the .get() method of Python dicts than to a mapping lookup in Python.

>>> table[1000000] is None
True
>>> table['no such key'] is None
True
>>> mapping['no such key'] is None
True

Note that len() does the right thing for array tables but does not work on mappings:

>>> len(table)
4
>>> len(mapping)
0

This is because len() is based on the # (length) operator in Lua and because of the way Lua defines the length of a table. Remember that unset table indices always return nil, including indices outside of the table size. Thus, Lua basically looks for an index that returns nil and returns the index before that. This works well for array tables that do not contain nil values, gives barely predictable results for tables with 'holes' and does not work at all for mapping tables. For tables with both sequential and mapping content, this ignores the mapping part completely.

Note that it is best not to rely on the behaviour of len() for mappings. It might change in a later version of Lupa.

Similar to the table interface provided by Lua, Lupa also supports attribute access to table members:

>>> table = lua.eval('{ a=1, b=2 }')
>>> table.a, table.b
(1, 2)
>>> table.a == table['a']
True

This enables access to Lua 'methods' that are associated with a table, as used by the standard library modules:

>>> string = lua.eval('string')    # get the 'string' library table
>>> print( string.lower('A') )
a

The next is an example of Lua coroutines. A wrapped Lua coroutine behaves exactly like a Python coroutine. It needs to get created at the beginning, either by using the .coroutine() method of a function or by creating it in Lua code. Then, values can be sent into it using the .send() method or it can be iterated over. Note that the .throw() method is not supported, though.

>>> lua_code = '''\
...     function(N)
...         for i=0,N do
...             coroutine.yield( i%2 )
...         end
...     end
... '''
>>> lua = LuaRuntime()
>>> f = lua.eval(lua_code)

>>> gen = f.coroutine(4)
>>> list(enumerate(gen))
[(0, 0), (1, 1), (2, 0), (3, 1), (4, 0)]

An example where values are passed into the coroutine using its .send() method:

>>> lua_code = '''\
...     function()
...         local t,i = {},0
...         local value = coroutine.yield()
...         while value do
...             t[i] = value
...             i = i + 1
...             value = coroutine.yield()
...         end
...         return t
...     end
... '''
>>> f = lua.eval(lua_code)

>>> co = f.coroutine()   # create coroutine
>>> co.send(None)        # start coroutine (stops at first yield)

>>> for i in range(3):
...     co.send(i*2)

>>> mapping = co.send(None)   # loop termination signal
>>> list(mapping.items())
[(0, 0), (1, 2), (2, 4)]

It also works to create coroutines in Lua and to pass them back into Python space:

>>> lua_code = '''\
...   function f(N)
...         for i=0,N do
...             coroutine.yield( i%2 )
...         end
...   end ;
...   co1 = coroutine.create(f) ;
...   co2 = coroutine.create(f) ;
...
...   status, first_result = coroutine.resume(co2, 2) ;   -- starting!
...
...   return f, co1, co2, status, first_result
... '''

>>> lua = LuaRuntime()
>>> f, co, lua_gen, status, first_result = lua.execute(lua_code)

>>> # a running coroutine:

>>> status
True
>>> first_result
0
>>> list(lua_gen)
[1, 0]
>>> list(lua_gen)
[]

>>> # an uninitialised coroutine:

>>> gen = co(4)
>>> list(enumerate(gen))
[(0, 0), (1, 1), (2, 0), (3, 1), (4, 0)]

>>> gen = co(2)
>>> list(enumerate(gen))
[(0, 0), (1, 1), (2, 0)]

>>> # a plain function:

>>> gen = f.coroutine(4)
>>> list(enumerate(gen))
[(0, 0), (1, 1), (2, 0), (3, 1), (4, 0)]

The following example calculates a mandelbrot image in parallel threads and displays the result in PIL. It is based on a benchmark implementation for the Computer Language Benchmarks Game.

lua_code = '''\
    function(N, i, total)
        local char, unpack = string.char, unpack
        local result = ""
        local M, ba, bb, buf = 2/N, 2^(N%8+1)-1, 2^(8-N%8), {}
        local start_line, end_line = N/total * (i-1), N/total * i - 1
        for y=start_line,end_line do
            local Ci, b, p = y*M-1, 1, 0
            for x=0,N-1 do
                local Cr = x*M-1.5
                local Zr, Zi, Zrq, Ziq = Cr, Ci, Cr*Cr, Ci*Ci
                b = b + b
                for i=1,49 do
                    Zi = Zr*Zi*2 + Ci
                    Zr = Zrq-Ziq + Cr
                    Ziq = Zi*Zi
                    Zrq = Zr*Zr
                    if Zrq+Ziq > 4.0 then b = b + 1; break; end
                end
                if b >= 256 then p = p + 1; buf[p] = 511 - b; b = 1; end
            end
            if b ~= 1 then p = p + 1; buf[p] = (ba-b)*bb; end
            result = result .. char(unpack(buf, 1, p))
        end
        return result
    end
'''

image_size = 1280   # == 1280 x 1280
thread_count = 8

from lupa import LuaRuntime
lua_funcs = [ LuaRuntime(encoding=None).eval(lua_code)
              for _ in range(thread_count) ]

results = [None] * thread_count
def mandelbrot(i, lua_func):
    results[i] = lua_func(image_size, i+1, thread_count)

import threading
threads = [ threading.Thread(target=mandelbrot, args=(i,lua_func))
            for i, lua_func in enumerate(lua_funcs) ]
for thread in threads:
    thread.start()
for thread in threads:
    thread.join()

result_buffer = b''.join(results)

# use PIL to display the image
import Image
image = Image.fromstring('1', (image_size, image_size), result_buffer)
image.show()

Note how the example creates a separate LuaRuntime for each thread to enable parallel execution. Each LuaRuntime is protected by a global lock that prevents concurrent access to it. The low memory footprint of Lua makes it reasonable to use multiple runtimes, but this setup also means that values cannot easily be exchanged between threads inside of Lua. They must either get copied through Python space (passing table references will not work, either) or use some Lua mechanism for explicit communication, such as a pipe or some kind of shared memory setup.

Lupa provides a simple mechanism to control access to Python objects. Each attribute access can be passed through a filter function as follows:

>>> def filter_attribute_access(obj, attr_name, is_setting):
...     if isinstance(attr_name, unicode):
...         if not attr_name.startswith('_'):
...             return attr_name
...     raise AttributeError('access denied')

>>> lua = lupa.LuaRuntime(
...           register_eval=False,
...           attribute_filter=filter_attribute_access)
>>> func = lua.eval('function(x) return x.__class__ end')
>>> func(lua)
Traceback (most recent call last):
 ...
AttributeError: access denied

The is_setting flag indicates whether the attribute is being read or set.

Note that the attributes of Python functions provide access to the current globals() and therefore to the builtins etc. If you want to safely restrict access to a known set of Python objects, it is best to work with a whitelist of safe attribute names. One way to do that could be to use a well selected list of dedicated API objects that you provide to Lua code, and to only allow Python attribute access to the set of public attribute/method names of these objects.

This will usually work as is, but here are the details, in case anything goes wrong for you.

To use binary modules in Lua, you need to compile them against the header files of the LuaJIT sources that you used to build Lupa, but do not link them against the LuaJIT library.

Furthermore, CPython needs to enable global symbol visibility for shared libraries before loading the Lupa module. This can be done by calling sys.setdlopenflags(flag_values). Importing the lupa module will automatically try to set up the correct dlopen flags if it can find the platform specific DLFCN Python module that defines the necessary flag constants. In that case, using binary modules in Lua should work out of the box.

If this setup fails, however, you have to set the flags manually. When using the above configuration call, the argument flag_values must represent the sum of your system's values for RTLD_NEW and RTLD_GLOBAL. If RTLD_NEW is 2 and RTLD_GLOBAL is 256, you need to call sys.setdlopenflags(258).

Assuming that the Lua luaposix (posix) module is available, the following should work on a Linux system:

>>> import sys
>>> orig_dlflags = sys.getdlopenflags()
>>> sys.setdlopenflags(258)
>>> import lupa
>>> sys.setdlopenflags(orig_dlflags)

>>> lua = lupa.LuaRuntime()
>>> posix_module = lua.require('posix')     # doctest: +SKIP

About

Lua in Python + automatic conversion of numpy arrays to a luajit ctype array

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 100.0%