[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Python-Dev] Micro-benchmarks for PEP 580

On Tue, Jul 10, 2018 at 8:55 PM Jeroen Demeyer <J.Demeyer at ugent.be> wrote:
> OK, I tried with --duplicate 200 and you can see the results at
> https://gist.github.com/jdemeyer/f0d63be8f30dc34cc989cd11d43df248
> In short, the timings with and without PEP 580 are roughly the same
> (which is to be expected). Interestingly, a small but significant
> improvement can be seen when calling *unbound* methods.
> The real improvement comes from supporting a new calling protocol:
> formerly custom classes could only implement tp_call, but now they can
> use FASTCALL just like built-in functions/methods. For this, there is an
> improvement of roughly a factor 1.2 for calls without arguments, 1.6 for
> calls with positional arguments and 2.8 for calls with keywords.

We know it when we introduced FASTCALL.

What I want know is "how often" tp_call in custom type is called
in real application.  Does it boost real application performance
significantly?  5%? 10%?

If it's not significant enough, I want to wait make FASTCALL public until
more evolutionary optimization happened.  There are some remaining
possible optimizations.

For example, let's assume cfunction like this:

static PyObject*
myfunc_impl(PyObject *self, Py_ssize_t i)

static PyObject*
myfunc(PyObject *self, PyObject *arg)
    Py_ssize_t i;
    if (!PyArg_Parse(arg, "n;myfunc", &i)) {
        return NULL;
    return myfunc_impl(self, i);

Then, the function is called from another C extension like this:

PyObject_CallFunction(func, "n", 42);

Currently, we create temporary long object for passing argument.
If there is protocol for exposeing format used by PyArg_Parse*, we can
bypass temporal Python object and call myfunc_impl directly.

I think optimization like this idea can boost application performance
using Cython heavily.
But in Python and stdlib, there are no enough "call C function from C function"
scenarios, compared with Cython based applications.  We really
need help from Cython world for this area.

INADA Naoki  <songofacandy at gmail.com>