Vectorization of Fortran loops and other optimizations #227
Description
Just a question / suggestion: this are the results of profiling:
ncalls tottime percall cumtime percall filename:lineno(function)
571 7110.513 12.453 7110.513 12.453 {built-in method posix.waitpid}
4 1365.157 341.289 2005.113 501.278 wavelet_atrous.py:667(check_islands_for_overlap)
1 830.799 830.799 830.799 830.799 psf_vary.py:685(tess_simple)
3407 381.475 0.112 381.497 0.112 necompiler.py:978(re_evaluate)
1544 312.586 0.202 312.586 0.202 {method 'partition' of 'numpy.ndarray' objects}
20 305.964 15.298 305.964 15.298 {built-in method scipy.ndimage._nd_image.geometric_transform}
44 288.652 6.560 288.652 6.560 {built-in method gc.collect}
5 270.768 54.154 1012.391 202.478 functions.py:2223(bstat)
214 168.266 0.786 190.097 0.888 wavelet_atrous.py:649(renumber_islands)
314391 137.325 0.000 225.320 0.001 _methods.py:196(_var)
5110219 134.055 0.000 134.055 0.000 {method 'reduce' of 'numpy.ufunc' objects}
6 117.511 19.585 251.401 41.900 make_residimage.py:25(__call__)
27523 106.675 0.004 106.675 0.004 {method 'flatten' of 'numpy.ndarray' objects}
10056453/7133976 97.635 0.000 817.598 0.000 {built-in method numpy.core._multiarray_umath.implement_array_function}
31858 55.493 0.002 55.493 0.002 {built-in method posix.read}
Have you tried to vectorize loops there:
https://github.com/lofar-astron/PyBDSF/blob/master/src/fortran/pytess_simple.f#L16C1-L29C15
?
ChatGPT seems to be doing this:
program vectorized_code
implicit none
integer, parameter :: m = 100, n = 100, ngens = 10
integer :: i, j, k
real(8), dimension(n, m) :: volrank
real(8), dimension(ngens) :: xgens, ygens, wts
real(8), dimension(n, m, ngens) :: dist
real(8) :: dumr
integer, dimension(n, m) :: minind
! Initialize generators and weights (example values)
do k = 1, ngens
xgens(k) = k
ygens(k) = k
wts(k) = 1.0d0
end do
! Vectorized computation
volrank = 0.d0
do k = 1, ngens
! Calculate distance from generator k to all points (i, j)
dist(:, :, k) = sqrt((reshape(([(i, i = 1, n)], [n, m]) - xgens(k)), [n, m]) ** 2 + &
(reshape(([(j, j = 1, m)], [m, n]) - ygens(k)), [m, n]) ** 2) / wts(k)
end do
! Find the minimum distance and corresponding generator index
do j = 1, m
do i = 1, n
dumr = minval(dist(i, j, :))
minind(i, j) = minloc(dist(i, j, :), dim = 1)
end do
end do
end program vectorized_code
but Im bad at Fortran so I cannot verify if this is OK.