Dipolarkernal speedup (#473)

* Seperate Kinterpolator into its own function

Interpolation in Scipy is very slow, currently the same interpolation is being rerun for every calculation of the dipolarkernal. This is now cached to speed it up.

* Only run orientation selection on grid and integral based kernals

This hunk of code is not needed when fresnel integrals are used so does not need to be evaluated

* Update changelog

deerlab/dipolarkernel.py

@@ -367,7 +367,7 @@ def dipolarkernel(t, r, *, pathways=None, mod=None, bg=None, method='fresnel', e
 
     if tinterp is not None:
         # Construct interpolator, this way elementarykernel_twospin is executed only once independently of how many pathways there are
-        Kinterpolator = [interp1d(tinterp,elementarykernel_twospin(tinterp,r_q,method,excbandwidth,gridsize,g,orisel,complex),axis=0,kind='cubic') for r_q in r]
+        Kinterpolator = _elementarykernel_twospin_interp(tinterp,r,method,excbandwidth,gridsize,g,orisel,complex)
         withinInterpolation = lambda tdip: np.all((np.max(tinterp) >= np.max(tdip)) & (np.min(tinterp) <= np.min(tdip)))
 
     # Define kernel matrix auxiliary functions
@@ -435,8 +435,14 @@ def K0_3spin(tdip):
 
     return K
 #==============================================================================
-
-
+@cached(max_size=100)
+def _elementarykernel_twospin_interp(tinterp,r,method,excbandwidth,gridsize,g,orisel,complex):
+    """
+    Construct interpolator, this way elementarykernel_twospin is executed only once independently of how many pathways there are
+    Cached for performance reasons, interpolation is slow.
+    """
+    Kinterpolator = [interp1d(tinterp,elementarykernel_twospin(tinterp,r_q,method,excbandwidth,gridsize,g,orisel,complex),axis=0,kind='cubic') for r_q in r]
+    return Kinterpolator
 
 #==============================================================================
 #   TWO-SPIN ELEMENTARY DIPOLAR KERNEL
@@ -468,14 +474,15 @@ def elementarykernel_twospin(tdip,r,method,ωex,gridsize,g,Pθ,complex):
     ωr = (μ0/2)*μB**2*g[0]*g[1]/h*1e21/(r**3)  # rad μs^-1
 
     # Orientation selection
-    orientationselection = Pθ is not None 
-    if orientationselection:
-        # Ensure zero-derivatives at [0,π/2]
-        θ = np.linspace(0,π/2,50) # rad
-        Pθ_ = make_interp_spline(θ, Pθ(θ),bc_type="clamped")
-        # Ensure normalization of probability density function (∫P(cosθ)dcosθ=1)
-        Pθnorm,_ = quad(lambda cosθ: Pθ_(np.arccos(cosθ)),0,1,limit=1000)
-        Pθ = lambda θ: Pθ_(θ)/Pθnorm
+    if method != 'fresnel':
+        orientationselection = Pθ is not None 
+        if orientationselection:
+            # Ensure zero-derivatives at [0,π/2]
+            θ = np.linspace(0,π/2,50) # rad
+            Pθ_ = make_interp_spline(θ, Pθ(θ),bc_type="clamped")
+            # Ensure normalization of probability density function (∫P(cosθ)dcosθ=1)
+            Pθnorm,_ = quad(lambda cosθ: Pθ_(np.arccos(cosθ)),0,1,limit=1000)
+            Pθ = lambda θ: Pθ_(θ)/Pθnorm
 
     def elementarykernel_fresnel(tdip):
     #------------------------------------------------------------------------

docsrc/source/changelog.rst

@@ -27,6 +27,7 @@ Release Notes
 Release ``v1.1.3`` - Ongoing
 ------------------------------------------
 - |fix| : Removes unnecessary files from the docs
+- |efficiency| : Improves the performance of the ``dipolarkernel`` function by 10-30% (:pr:`473`), by caching the interpolation of he effective dipolar evolution time vector.
 
 Release ``v1.1.2`` - November 2023
 ------------------------------------------