Error while implementing Pythran.

[khushi-411]

I'm trying to accelerate my NumPy's implementation using the Pythran library.

Code: here

I'm getting the following outputs:

• 1st: For Command:

  taskset -c 6,7,14,15 python -m pythran.run -v python/acc_pythran2.py data/input16.txt 10.0

  Error:

  INFO:      sys file exists: /usr/lib/python3.9/site-packages/pythran/pythran.cfg
  INFO: platform file exists: /usr/lib/python3.9/site-packages/pythran/pythran-linux.cfg
  WARNING:     user file does not exist: /home/khushi/.pythranrc
  INFO: pythranrc section [pythran] is valid and options are correct
  INFO: pythranrc section [typing] is valid and options are correct
  INFO: pythranrc section [compiler] is valid and options are correct
  CRITICAL: I am in trouble. Your input file does not seem to match Pythran's constraints...
  python/acc_pythran2.py:None:None error: Module 'sys' not found.

• 2nd: For Command:

  pythran python/acc_pythran2.py data/input16.txt 10.0

  Error:

  CRITICAL: I am in trouble. Your input file does not seem to match Pythran's constraints...
  python/acc_pythran2.py:None:None error: Module 'sys' not found.

• 3rd: For Command:

  taskset -c 6,7,14,15 python python/acc_pythran2.py data/input16.txt 10.0

  Result:

  time taken:  4.634182117999444

My Viewpoints

• For 1st & 2nd:
  • I had imported sys module, yet it shows "Module 'sys' not found.". I am concluding that the error is not due to the sys module.
  • Still, I wanted to configure. Hence, I hard-coded datasets and again tried to run the codes. I got: "Module 'timeit' not found.".
  • I am concluding that the error is while computing accelerations in compute_accelerations functions.
• For 3rd: Here, the time taken by Pythran's implementation is almost similar to that of NumPy. This should not be the case since Pythran is an accelerator. I'm expecting the performance of Pythran > Performance of NumPy.

Solution

• We can observe that the error is while calculating accelerations of the masses. I made some changes in the function, here. It works well with jitted compilation using Pythran: Transonic.

Drawbacks with the above solution

• The solution is able to produce better results. But steps must be the same in all kinds of implementation for benchmarking. Does that matter in a bigger view? Yeah, I feel so.
• One solution to the above problem I can think of, is to change the functions in all the other implementations. But that will degrade NumPy's performance. Hence, I'm not up for this solution.

I'm looking forward to your inputs. Thanks! cc: @mattip

I'm referring to the following reference:

• https://pythran.readthedocs.io/en/latest/MANUAL.html
• https://serge-sans-paille.github.io/pythran-stories/pythran-as-a-bridge-between-fast-prototyping-and-code-deployment.html
• https://pythran.readthedocs.io/en/latest/examples/Third%20Party%20Libraries.html#Pythran-Code
• https://github.com/serge-sans-paille/pythran-stories/blob/gh-pages/notebooks/Pythran%20Tutorial.ipynb