Make gaus_2d faster (2)

[AlexKurek]

before (but after #273)

    10935    0.200    0.000    0.204    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10935    0.205    0.000    0.208    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10935    0.222    0.000    0.226    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10935    0.196    0.000    0.200    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)

after:

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
    10322    0.173    0.000    0.177    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10322    0.167    0.000    0.171    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10322    0.167    0.000    0.171    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10322    0.167    0.000    0.172    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)
    10322    0.165    0.000    0.169    0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)

~18,4% faster. A lot is done in place now. I think this is now as fast as it gets.
There are very rare differences in the .gaul file at 13th decimal place, but this is normal.

[AlexKurek]

After 291f1e0 it gets faster by another ~24% :
10322 0.128 0.000 0.153 0.000 /home/akurek/miniconda3/envs/pybdsf-313_local_temp/lib/python3.13/site-packages/bdsf/functions.py:202(gaus_2d)

The .gaul file is OK.

[gmloose]

Can you provide me the input file(s) you're using to do the profiling? I'd like to experiment a bit myself, and would like to be able to make a fair comparison. Also, did you just call gaus_2d in isolation, or is this profile part of a different function call? Anyway, please provide all the necessary information to reproduce your experiments.

[AlexKurek]

Since it is an optimisation, I assumed that it would be best to use the latest Python and Numpy, as these are constantly being updated to improve performance. I was not able to build against Python 3.14 nor using pip install . -e. This is the best I can do for now:

Paste all this to the terminal in one go. It will remove conda env and cloned repo from previous experiment and prepare new conda env for installation anew.

cd ~
rm -rf PyBDSF
conda deactivate

conda remove --name pybdsf-313_local_temp --all -y
conda create -n pybdsf-313_local_temp python=3.13 -y
conda activate pybdsf-313_local_temp
conda install -c conda-forge cmake ninja boost libboost-python-devel numpy scipy matplotlib astropy gfortran_linux-64 gcc_linux-64 gxx_linux-64 -y
# Tools for communication Python <-> CMake <-> Fortran
pip install scikit-build setuptools wheel meson setuptools_scm
export CMAKE_ARGS="-DBoost_PYTHON_LIBRARY_RELEASE=$CONDA_PREFIX/lib/libboost_python313.so \
                  -DBoost_PYTHON_INCLUDE_DIR=$CONDA_PREFIX/include \
                  -DBoost_NUMPY_LIBRARY_RELEASE=$CONDA_PREFIX/lib/libboost_numpy313.so \
                  -DBOOST_ROOT=$CONDA_PREFIX"
export FFLAGS="-std=legacy"
export FCFLAGS="-std=legacy"
git clone https://github.com/lofar-astron/PyBDSF.git
cd PyBDSF

Now paste your experimental changes using a text editor to the clonned repo. Then paste all this in one go to the terminal. It will install PyBDSF, move to a different folder to execute a profilled run and display the result for map_2d.

pip install .
cd ..
cd bench_pybdsf
python -m cProfile -o out.prof run_pybdsf.py
echo "stats map_2d" | python -m pstats out.prof

run_pybdsf.py is attached. Here is a link for FITS file (127 MB). Please tell me when I can remove it from the server.
https://www.oa.uj.edu.pl/A.Kurek/NGC6048_9-MFS-image.fits

Please note that the total running time is not stable:

PyBDSF running time: 0:2:38
PyBDSF running time: 0:2:30
PyBDSF running time: 0:2:42
PyBDSF running time: 0:2:35
PyBDSF running time: 0:2:33

As far as I know this is normal for high level languages, as there is a garbage collector going all over the code etc. Thats why I was always doing the profiling few times.

At the first time you can also display full profilling results:

python -m pstats out.prof
sort cumulative
stats 60

to see what should be optimised.

One comment about 326df50. At first I was thinking that it would be optmial to let it compute on positive values as long as possible. But I have profilled your suggestion and it turned out that there is no difference.

run_pybdsf.py