Hi,
I found some time to do a pretty complete rewrite of the argument unpacking
code, i.e. the code that unpacks positional args and keyword arguments from
the Python calling convention into the local variables of a function.
Previously, we were using PyArg_ParseTupleAndKeywords with a couple of added
optimisations for common cases. I replaced this with a dedicated
implementation that is generated specifically for the signature of the
function. It uses extremely well optimisable switch statements in the C code
(and no loops unless you use *args/**kwargs or in some of the error cases).
The result is that Cython now generates function init code that is faster than
Python in most cases, even in many cases where you pass keyword arguments
(which wasn't the case before at all). It's actually a lot faster than the
last Cython release, especially for keyword arguments, and quite a bit faster
than Python, which has impressively fast generic argument handling code. I
attached some numbers below.
The only case where we are still a bit slower than Python is when you pass
(almost) all arguments as keyword arguments. I suspect that it would be faster
in this case to iterate over the dictionary and compare the keys with the
argument names than to ask for each key separately. Which strategy to choose
could be decided at runtime if the size of the keyword dictionary is close to
the total number of arguments. However, as the absolute numbers compared to
Python show, this may not be worth the effort.
Have fun,
Stefan
-----------------------------
def generic(*args, **kwargs):
result = (args, kwargs)
def partially_generic(a, b, *args, **kwargs):
result = (a, b, args, kwargs)
def args(*args):
result = (args,)
def positional(a,b,c,d):
result = (a,b,c,d)
def optional(a=1, b=2, c=3, d=4):
result = (a,b,c,d)
-----------------------------
$ TEST=optional; python2.6 -m timeit "test(1,2)"
callbench_python
1000000 loops, best of 3: 1.29 usec per loop
callbench_devel
1000000 loops, best of 3: 0.603 usec per loop
callbench_release
1000000 loops, best of 3: 0.61 usec per loop
$ TEST=args; python2.6 -m timeit "test(1,2)"
callbench_python
1000000 loops, best of 3: 1.09 usec per loop
callbench_devel
1000000 loops, best of 3: 0.547 usec per loop
callbench_release
1000000 loops, best of 3: 0.568 usec per loop
$ TEST=optional; python2.6 -m timeit "test(1,2,3,d=4)"
callbench_python
1000000 loops, best of 3: 1.5 usec per loop
callbench_devel
1000000 loops, best of 3: 1.25 usec per loop
callbench_release
1000000 loops, best of 3: 1.99 usec per loop
$ TEST=positional; python2.6 -m timeit "test(1,2,3,d=4)"
callbench_python
1000000 loops, best of 3: 1.44 usec per loop
callbench_devel
1000000 loops, best of 3: 1.23 usec per loop
callbench_release
100000 loops, best of 3: 2.01 usec per loop
$ TEST=positional; python2.6 -m timeit "test(a=1,b=2,c=3,d=4)"
callbench_python
1000000 loops, best of 3: 1.73 usec per loop
callbench_devel
1000000 loops, best of 3: 1.81 usec per loop
callbench_release
100000 loops, best of 3: 2.41 usec per loop
$ TEST=generic; python2.6 -m timeit "test(a=1,b=2,c=3,d=4)"
callbench_python
100000 loops, best of 3: 2.15 usec per loop
callbench_devel
100000 loops, best of 3: 2.48 usec per loop
callbench_release
100000 loops, best of 3: 2.56 usec per loop
$ TEST=partially_generic; python2.6 -m timeit "test(a=1,b=2,c=3,d=4)"
callbench_python
100000 loops, best of 3: 2.51 usec per loop
callbench_devel
100000 loops, best of 3: 2.96 usec per loop
callbench_release
100000 loops, best of 3: 3.42 usec per loop
$ TEST=partially_generic; python2.6 -m timeit "test(1,b=2,c=3,d=4)"
callbench_python
100000 loops, best of 3: 2.25 usec per loop
callbench_devel
100000 loops, best of 3: 2.52 usec per loop
callbench_release
100000 loops, best of 3: 3.1 usec per loop
$ TEST=partially_generic; python2.6 -m timeit "test(1,2,3,d=4)"
callbench_python
100000 loops, best of 3: 2.01 usec per loop
callbench_devel
100000 loops, best of 3: 2.06 usec per loop
callbench_release
100000 loops, best of 3: 2.46 usec per loop
_______________________________________________
Cython-dev mailing list
[email protected]
http://codespeak.net/mailman/listinfo/cython-dev
