Unexpectedly slow class method calls

[denik]

Calling a class method is unexpectedly much slower (10-20x) than calling a function. This is not observed with comparable Cython code, so not an issue with Python class in C extensions.

pydust: calling function is fast (compiled with "zig build -Doptimize=ReleaseFast")

$ python -m timeit -s 'from fibonacci._lib import get1' 'get1()'
5000000 loops, best of 5: 37.7 nsec per loop

$ python -m timeit -s 'from fibonacci._lib import get1' 'get1()'
5000000 loops, best of 5: 46.1 nsec per loop

Calling class method is slow:

$ python -m timeit -s 'from fibonacci._lib import Class; s = Class(10)' 's.get()'
200000 loops, best of 5: 981 nsec per loop

$ python -m timeit -s 'from fibonacci._lib import Class; s = Class(10)' 's.get()'
200000 loops, best of 5: 922 nsec per loop

Code: diff against fresh checkout of ziggy-pydust-template:

$ git diff
diff --git a/src/fib.zig b/src/fib.zig
index 3fde4f7..c32e91c 100644
--- a/src/fib.zig
+++ b/src/fib.zig
@@ -80,6 +80,28 @@ pub const FibonacciIterator = py.class(struct {
     }
 });
 
+pub const Class = py.class(struct {
+    const Self = @This();
+
+    i: u64,
+
+    pub fn __new__(args: struct { i: u64 }) !Self {
+        return .{ .i = args.i };
+    }
+
+    pub fn get(self: *Self) u64 {
+        return self.i;
+    }
+});
+
+pub fn get1() u64 {
+    return 10;
+}

Same thing done with Cython - same performance for function and method call:

$ python -m timeit -s 'from testclass import get1' 'get1()'
5000000 loops, best of 5: 39.3 nsec per loop

$ python -m timeit -s 'from testclass import get1' 'get1()'
5000000 loops, best of 5: 44.1 nsec per loop

$ python -m timeit -s 'from testclass import Class; s = Class(10)' 's.get()'
5000000 loops, best of 5: 44.5 nsec per loop

$ python -m timeit -s 'from testclass import Class; s = Class(10)' 's.get()'
5000000 loops, best of 5: 47 nsec per loop

$ cat testclass.pyx 
cdef class Class:
    cdef int i

    def __init__(self, i):
        self.i = i

    cpdef get(self):
        return self.i

cpdef get1():
    return 10

$ cat setup.py 
from setuptools import setup
from Cython.Build import cythonize

setup(
    ext_modules = cythonize("testclass.pyx")
)

$ python setup.py build_ext --inplace

[gatesn]

Thank you for reporting this, I'll take a look. My initial suspicion is that we have some naive code that asserts the type of the self argument - which is probably unnecessary given we know it's a class/instance method.

[gatesn]

Yup, so a trace confirmed all the time is spent doing an import then isinstance check on the self parameter.

I've put up a PR #228 that seems to fix it. With optimize=Debug:

(ziggy-pydust-py3.11) ➜  ziggy-pydust git:(ngates/class-method-perf) ✗ python -m timeit -s 'from example.fib import get1' 'get1()'
5000000 loops, best of 5: 85.2 nsec per loop
(ziggy-pydust-py3.11) ➜  ziggy-pydust git:(ngates/class-method-perf) ✗ python -m timeit -s 'from example.fib import Class; s = Class(10)' 's.get()'
5000000 loops, best of 5: 95.5 nsec per loop

I'll try to put up a proper fix to this (which I believe to be #227) tomorrow sometime.