On 04/30/2012 06:30 PM, Wes McKinney wrote:
On Mon, Apr 30, 2012 at 11:19 AM, mark florisson
<markflorisso...@gmail.com> wrote:
On 30 April 2012 14:49, Wes McKinney<wesmck...@gmail.com> wrote:
On Sun, Apr 29, 2012 at 5:56 AM, mark florisson
<markflorisso...@gmail.com> wrote:
On 29 April 2012 08:42, Nathaniel Smith<n...@pobox.com> wrote:
On Sat, Apr 28, 2012 at 10:25 PM, mark florisson
<markflorisso...@gmail.com> wrote:
On 28 April 2012 22:04, Nathaniel Smith<n...@pobox.com> wrote:
Was chatting with Wes today about the usual problem many of us have
encountered with needing to use some sort of templating system to
generate code handling multiple types, operations, etc., and a wacky
idea occurred to me. So I thought I'd through it out here.
What if we added a simple macro facility to Cython, that worked at the
AST level? (I.e. I'm talking lisp-style macros, *not* C-style macros.)
Basically some way to write arbitrary Python code into a .pyx file
that gets executed at compile time and can transform the AST, plus
some nice convenience APIs for simple transformations.
E.g., if we steal the illegal token sequence @@ as our marker, we
could have something like:
@@ # alone on a line, starts a block of Python code
from Cython.MacroUtil import replace_ctype
def expand_types(placeholder, typelist):
def my_decorator(function_name, ast):
functions = {}
for typename in typelist:
new_name = "%s_%s" % (function_name, typename)
functions[name] = replace_ctype(ast, placeholder, typename)
return functions
return function_decorator
@@ # this token sequence cannot occur in Python, so it's a safe end-marker
# Compile-time function decorator
# Results in two cdef functions named sum_double and sum_int
@@expand_types("T", ["double", "int"])
cdef T sum(np.ndarray[T] arr):
cdef T start = 0;
for i in range(arr.size):
start += arr[i]
return start
I don't know if this is a good idea, but it seems like it'd be very
easy to do on the Cython side, fairly clean, and be dramatically less
horrible than all the ad-hoc templating stuff people do now.
Presumably there'd be strict limits on how much backwards
compatibility we'd be willing to guarantee for code that went poking
around in the AST by hand, but a small handful of functions like my
notional "replace_ctype" would go a long way, and wouldn't impose much
of a compatibility burden.
-- Nathaniel
_______________________________________________
cython-devel mailing list
cython-devel@python.org
http://mail.python.org/mailman/listinfo/cython-devel
Have you looked at http://wiki.cython.org/enhancements/metaprogramming ?
In general I would like better meta-programming support, maybe even
allow defining new operators (although I'm not sure any of it is very
pythonic), but for templates I think fused types should be used, or
improved when they fall short. Maybe a plugin system could also help
people.
I hadn't seen that, no -- thanks for the link.
I have to say that the examples in that link, though, give me the
impression of a cool solution looking for a problem. I've never wished
I could symbolically differentiate Python expressions at compile time,
or create a mutant Python+SQL hybrid language. Actually I guess I've
only missed define-syntax once in maybe 10 years of hacking in
Python-the-language: it's neat how if you do 'plot(x, log(y))' in R it
will peek at the caller's syntax tree to automagically label the axes
as "x" and "log(y)", and that can't be done in Python. But that's not
exactly a convincing argument for a macro system.
But generating optimized code is Cython's whole selling point, and
people really are doing klugey tricks with string-based preprocessors
just to generate multiple copies of loops in Cython and C.
Also, fused types are great, but: (1) IIUC you can't actually do
ndarray[fused_type] yet, which speaks to the feature's complexity, and
What? Yes you can do that.
I haven't been able to get ndarray[fused_t] to work as we've discussed
off-list. In your own words "Unfortunately, the automatic buffer
dispatch didn't make it into 0.16, so you need to manually
specialize". I'm a bit hamstrung by other users needing to be able to
compile pandas using the latest released Cython.
Well, as I said, it does work, but you need to tell Cython which type
you meant. If you don't want to do that, you have to use this branch:
https://github.com/markflorisson88/cython/tree/_fused_dispatch_rebased
. This never made it in since we had no consensus on whether to allow
the compiler to bootstrap itself and because of possible immaturity of
the branch.
So what doesn't work is automatic dispatch for Python functions (def
functions and the object version of a cpdef function). They don't
automatically select the right specialization for buffer arguments.
Anything else should work, otherwise it's a bug.
Note also that figuring out which specialization to call dynamically
(i.e. not from Cython space at compile time, but from Python space at
runtime) has non-trivial overhead on top of just argument unpacking.
But you can't say "doesn't work" without giving a concrete example of
what doesn't work besides automatic dispatch, and how it fails.
Sorry, I meant automatic dispatch re "doesn't work" and want to
reiterate how much I appreciate the work you're doing. To give some
context, my code is riddled with stuff like this:
lib.inner_join_indexer_float64
lib.inner_join_indexer_int32
lib.inner_join_indexer_int64
lib.inner_join_indexer_object
where the only difference between these functions is the type of the
buffer in the two arrays passed in. I have a template string for these
functions that looks like this:
inner_join_template = """@cython.wraparound(False)
@cython.boundscheck(False)
def inner_join_indexer_%(name)s(ndarray[%(c_type)s] left,
ndarray[%(c_type)s] right):
'''
...
I would _love_ to replace this with fused types.
In any case, lately I've been sort of yearning for the kinds of things
you can do with an APL-variant like J. Like here's a groupby in J:
labels
1 1 2 2 2 3 1
data
3 4 5.5 6 7.5 _2 8.3
labels</. data
┌───────┬─────────┬──┐
│3 4 8.3│5.5 6 7.5│_2│
└───────┴─────────┴──┘
Here< is box and /. is categorize.
Replacing the box< operator with +/ (sum), I get the group sums:
labels +/ /. data
15.3 19 _2
Have 2-dimensional data?
data
0 1 2 3 4 5 6
7 8 9 10 11 12 13
14 15 16 17 18 19 20
21 22 23 24 25 26 27
28 29 30 31 32 33 34
35 36 37 38 39 40 41
42 43 44 45 46 47 48
labels</."1 data
┌────────┬────────┬──┐
│0 1 6 │2 3 4 │5 │
├────────┼────────┼──┤
│7 8 13 │9 10 11 │12│
├────────┼────────┼──┤
│14 15 20│16 17 18│19│
├────────┼────────┼──┤
│21 22 27│23 24 25│26│
├────────┼────────┼──┤
│28 29 34│30 31 32│33│
├────────┼────────┼──┤
│35 36 41│37 38 39│40│
├────────┼────────┼──┤
│42 43 48│44 45 46│47│
└────────┴────────┴──┘
labels +//."1 data
7 9 5
28 30 12
49 51 19
70 72 26
91 93 33
112 114 40
133 135 47
However, J and other APLs are interpreted. If you generate C or
JIT-compile I think you can do really well performance wise and have
very expressive code for writing data algorithms without all this
boilerplate.
I know how you feel. On one hand I really like metaprogramming; on the
other hand I think it is very difficult to get right when done
compile-time (just look at C++ -- I've heard that D is a bit better
though I didn't really try it).
JIT is really the way to go. It is one thing that a JIT could optimize
the case where you pass a callback to a function and inline it run-time.
But even if it doesn't get that fancy, it'd be great to just be able to
write something like "cython.eval(s)" and have that be compiled (I guess
you could do that now, but the sheer overhead of the C compiler and all
the .so files involved means nobody would sanely use that as the main
way of stringing together something like pandas).
I think that's really the way to get "Pythonic" metaprogramming, where
you mix and match runtime and compile-time, and can hook into arbitrary
Python code in the meta-programming step.
Guido may even accept some syntax hooks to more easily express macros
without resorting to strings, if I got the right impression at PyData on
his take on DSLs.
Without a JIT, all we seem to come up with will be kludges. So I'm
sceptical about more metaprogramming features in Cython now since it
will be outdated in five years -- by then, somebody will either have
gotten our stuff together and JIT-ed both Python and Cython, or we'll
all be using something else (like R or Julia).
Dag
_______________________________________________
cython-devel mailing list
cython-devel@python.org
http://mail.python.org/mailman/listinfo/cython-devel
