Speed up interpolation during delay calculation

[Askaholic]

I profiled the zenith delay calculation at query points using the following command:

raiderDelay.py --date 20200103 --time 23:00:00 --latlon lat.dat lon.dat --model HRRR --zref 20000 -v --out test_output/

Where my lat/lon file consisted of 100x100 points spaced equally over the bounding box (SNWE) 37 40 -120 -116.

I found that the time was pretty evenly split between interpolation and reprojection:

         963425 function calls (913599 primitive calls) in 20.541 seconds

   Ordered by: internal time
   List reduced from 2679 to 10 due to restriction <10>

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        2    8.496    4.248    8.496    4.248 {method '_transform' of 'pyproj._transformer._Transformer' objects}
        2    5.090    2.545    6.403    3.202 interpolate.py:2499(_evaluate_linear)
        2    1.971    0.985    2.926    1.463 interpolate.py:2519(_find_indices)
60259/20139    1.278    0.000    2.736    0.000 {built-in method numpy.core._multiarray_umath.implement_array_function}
        6    0.954    0.159    0.954    0.159 {method 'searchsorted' of 'numpy.ndarray' objects}
        2    0.432    0.216    0.432    0.216 {built-in method scipy.interpolate.interpnd._ndim_coords_from_arrays}
       40    0.164    0.004    0.164    0.004 {method 'astype' of 'numpy.ndarray' objects}
        2    0.148    0.074    0.148    0.074 interpolate.py:2505(<listcomp>)
        1    0.145    0.145    0.145    0.145 {RAiDER.makePoints.makePoints1D}
        2    0.140    0.070    0.140    0.070 {pyproj._transformer.from_crs}

Where the cumulative time spent in interpolation made up about 50% of the total computation time:

  ncalls  tottime  percall  cumtime  percall filename:lineno(function)
     2    0.049    0.025    9.811    4.906 interpolate.py:2450(__call__)

After replacing this with the C++ interpolator, the interpolation which took 10 seconds before, took less than 1 second:

         963731 function calls (913898 primitive calls) in 11.178 seconds

   Random listing order was used
   List reduced from 2655 to 13 due to restriction <'interpolate'>

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.001    0.001   10.652   10.652 delay.py:28(interpolateDelay)
        2    0.052    0.026    0.769    0.385 delayFcns.py:303(interpolate2)
        2    0.246    0.123    0.246    0.123 {built-in method RAiDER.interpolate.interpolate}

I also verified that the results are the same:

>>> old = gdal_open("test_output/old/HRRR_hydro_23_00_00_std.envi")
>>> new = gdal_open("test_output/HRRR_hydro_23_00_00_std.envi")
>>> np.nanmean((old-new)**2)
0.0

[dbekaert]

@Askaholic this looks good to me. @jlmaurer any issues to merge?

[jlmaurer]

Looks good to me!