Better performance of the beam interpolation with faster domain checking and spline reuse

CHANGELOG.md

@@ -3,6 +3,16 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+### Changed
+- Improved `UVBeam` azimuth/zenith spline interpolation performance by sharing the
+tensor-product basis across beam data slices and using vectorized domain checks.
+`scipy.interpolate.NdBSpline` is now used whenever `spline_opts` requests only the
+spline orders `kx`/`ky` (the default), independent of `reuse_spline`;
+`scipy.interpolate.RectBivariateSpline` remains in use for smoothing (`s != 0`) and
+for any other RectBivariateSpline-only option. Both engines solve the same
+interpolating spline for a given `kx`/`ky` with `s = 0`.
+- Updated minimum dependency versions: scipy>=1.15
+
 ### Fixed
 - A bug where the longitudinal coordinate (e.g., RA) of "ephem" phase centers was not
 unwrapped, leading to potential discontinuities when performing interpolation.

README.md

@@ -160,7 +160,7 @@ Required:
 * pyerfa >= 2.0.1.1
 * python >= 3.11
 * pyyaml >= 5.4.1
-* scipy >= 1.9
+* scipy >= 1.15
 * setuptools_scm >= 8.1
 
 Optional:

ci/pyuvdata_min_deps_tests.yml

@@ -18,6 +18,6 @@ dependencies:
   - pytest-cov
   - pytest-xdist
   - pyyaml>=5.4.1
-  - scipy>=1.9
+  - scipy>=1.15
   - setuptools_scm>=8.1
   - sybil>=8.0

ci/pyuvdata_min_versions_tests.yml

@@ -22,7 +22,7 @@ dependencies:
   - pytest-xdist
   - python-casacore==3.5.2
   - pyyaml==5.4.1
-  - scipy==1.9.*
+  - scipy==1.15.*
   - setuptools==65.*
   - setuptools_scm==8.1
   - sybil==8.0

ci/pyuvdata_tests.yml

@@ -22,7 +22,7 @@ dependencies:
   - pytest-xdist
   - python-casacore>=3.5.2
   - pyyaml>=5.4.1
-  - scipy>=1.9
+  - scipy>=1.15
   - setuptools_scm>=8.1
   - sybil>=8.0
   - pip:

ci/pyuvdata_tests_mac_arm.yml

@@ -22,7 +22,7 @@ dependencies:
   - pytest-xdist
   - python-casacore>=3.5.2
   - pyyaml>=5.4.1
-  - scipy>=1.9
+  - scipy>=1.15
   - setuptools_scm>=8.1
   - sybil>=8.0
   - pip:

ci/pyuvdata_tests_windows.yml

@@ -21,7 +21,7 @@ dependencies:
   - pytest-cov
   - pytest-xdist
   - pyyaml>=5.4.1
-  - scipy>=1.9
+  - scipy>=1.15
   - cython>=3.0
   - setuptools_scm>=8.1
   - sybil>=8.0

environment.yaml

@@ -24,7 +24,7 @@ dependencies:
   - pytest-xdist
   - python-casacore>=3.5.2
   - pyyaml>=5.4.1
-  - scipy>=1.9
+  - scipy>=1.15
   - setuptools_scm>=8.1
   - sphinx
   - sybil>=8.0

pyproject.toml

@@ -44,7 +44,7 @@ dependencies = [
     "numpy>=1.23",
     "pyerfa>=2.0.1.1",
     "pyyaml>=5.4.1",
-    "scipy>=1.9",
+    "scipy>=1.15",
     "setuptools_scm>=8.1",
 ]
 requires-python = ">=3.11"

src/pyuvdata/beam_interface.py

@@ -279,13 +279,33 @@ def compute_response(
             the specified tolerance then return the beam at each nearest neighbor,
             otherwise interpolate the beam. Only applies if beam is a UVBeam.
         reuse_spline : bool
-            Save the interpolation functions for reuse.  Only applies if beam is
-            a UVBeam and interpolation_function is "az_za_simple".
+            Save the fitted splines on the beam for reuse. Only applies if beam is
+            a UVBeam and interpolation_function is "az_za_simple", where it applies
+            to both of the engines described under `spline_opts`. The cache holds
+            spline coefficients, so it costs roughly as much memory as the beam
+            data it was fit to.
         spline_opts : dict
-            Provide options to numpy.RectBivariateSpline. This includes spline
-            order parameters `kx` and `ky`, and smoothing parameter `s`. Only
-            applies if beam is a UVBeam and interpolation_function is "az_za_simple"
-            or "az_za_map_coordinates".
+            Options for the spline fit: the spline orders `kx` (zenith angle axis)
+            and `ky` (azimuth axis), and the smoothing parameter `s`. Only applies
+            if beam is a UVBeam. For "az_za_simple" this dict is also what selects
+            the interpolation engine:
+
+            - **Default** (`spline_opts` is None, or contains only `kx`, `ky`
+              and/or `s=0`): :class:`scipy.interpolate.NdBSpline`. The
+              tensor-product basis is built once and every vector/feed/frequency
+              slice is evaluated against it together, which is substantially
+              faster.
+            - **Fallback** (`s` is nonzero, or any other
+              :class:`scipy.interpolate.RectBivariateSpline` keyword such as
+              `bbox` is given): RectBivariateSpline, fit and evaluated one slice
+              at a time. Smoothing is only available on this path.
+
+            For the same `kx` and `ky` with `s=0` the two engines solve the same
+            interpolating spline and agree to floating point rounding (of order
+            1e-15 relative), so this choice affects speed rather than results.
+
+            For "az_za_map_coordinates" this dict is instead passed to
+            :func:`scipy.ndimage.map_coordinates` (e.g. `order`, `mode`, `cval`).
         check_azza_domain : bool
             Whether to check the domain of az/za to ensure that they are covered by the
             intrinsic data array. Checking them can be quite computationally expensive.