Transonic run python code instead of pythran

Question

After the first good results with pythran, I tried transonic to benefit form the jit and the class support. Unfortunately it does not run as expected.
If I use the @jit decorator the decorated functions are compiled and cached, but during the first run of the code the compiled version is not used, instead the function is processed by python. After the first run the cached version is used.
If I use the @boost decorator and run transonic runmwe.py a compiled version is created in the __pythran__ folder, but running the script with python runmwe.py I receive the following warning and the code is processed by python.

WARNING: Pythran file does not seem to be up-to-date:
<module '__pythran__.runmwe_920d6d0a5cd396436d463468328e997b' from '__pythran__/runmwe_920d6d0a5cd396436d463468328e997b.cpython-38-x86_64-linux-gnu.so'>

Rerunning transonic runmwe.py just produces a warning that the code is already up-to-date.

Do I miss some configuration to use @jit and @boost properly or is this the expected behavior and I use transonic the wrong way?

Used software from conda-forge:
transonic 0.4.5
pythran 0.9.7
python 3.8.6

MWE:

import numpy as np
from transonic import jit,boost

#transonic def looping(float[])

@boost
def looping(np_array):
    shape_x =np_array.shape[0]
    for x in range(shape_x):
                if np_array[x] < 0.5:
                    np_array[x] = 0
                else:
                    np_array[x] = 1
    return np_array

in_arr = np.random.rand(10**7)

looping(in_arr)

Can you please have a look at my new answer and tell me if it works for you? — paugier, Feb 17 '21 at 09:24

paugier · Accepted Answer · 2021-02-17T09:23:15.633

Your problem is due to an old bug in Transonic (now fixed by https://foss.heptapod.net/fluiddyn/transonic/-/merge_requests/88).

You use #transonic def looping(float[]) (without space between # and transonic) and it was working well only with the space after # (from https://www.python.org/dev/peps/pep-0008/#block-comments: "Each line of a block comment starts with a # and a single space").

The easiest way to fix your problem is to add the space, i.e. to write # transonic def looping(float[]). By the way, any code formatter (like https://black.readthedocs.io) would add it for you automatically.

However, Transonic supports type annotations and it would even be nicer to avoid the signatures in comments and to write:

from transonic import boost

@boost
def looping(np_array: "float[]"):
    ...

Thanks for the fix. For the `@boost` decorator the solution works. For the `@jit` decorator the first iteration still has only python speed. I will move the discussion to heptapod. — rqwa, Feb 22 '21 at 11:12

Transonic run python code instead of pythran

1 Answers1