half a LinearInterpolation POC

SciML · Jul 7, 2024 · 9737c50 · 9737c50
1 parent 152bcbf
commit 9737c50
Show file tree

Hide file tree

Showing 2 changed files with 44 additions and 15 deletions.
diff --git a/ext/DataInterpolationsChainRulesCoreExt.jl b/ext/DataInterpolationsChainRulesCoreExt.jl
@@ -3,26 +3,50 @@ module DataInterpolationsChainRulesCoreExt
 if isdefined(Base, :get_extension)
     using DataInterpolations: _interpolate, derivative, AbstractInterpolation,
                               LagrangeInterpolation, AkimaInterpolation,
-                              BSplineInterpolation, BSplineApprox
+                              BSplineInterpolation, BSplineApprox, LinearInterpolation,
+                              linear_interpolation_parameters
     using ChainRulesCore
 else
     using ..DataInterpolations: _interpolate, derivative, AbstractInterpolation,
                                 LagrangeInterpolation, AkimaInterpolation,
-                                BSplineInterpolation, BSplineApprox
+                                BSplineInterpolation, BSplineApprox, LinearInterpolation,
+                                linear_interpolation_parameters
     using ..ChainRulesCore
 end
 
-function ChainRulesCore.rrule(::typeof(_interpolate),
-        A::Union{
-            LagrangeInterpolation,
-            AkimaInterpolation,
-            BSplineInterpolation,
-            BSplineApprox
-        },
-        t::Number)
-    deriv = derivative(A, t)
-    interpolate_pullback(Δ) = (NoTangent(), NoTangent(), deriv * Δ)
-    return _interpolate(A, t), interpolate_pullback
+function ChainRulesCore.rrule(::typeof(linear_interpolation_parameters), u, t, idx)
+    slope = linear_interpolation_parameters(u, t, idx)
+
+    function linear_interpolation_parameters_pullback(Δslope)
+        du = @thunk(Δslope*...) # TODO: How to handle sparsity?
+        dt = @thunk(Δslope*...) # TODO: How to handle sparsity?
+        df = NoTangent()
+        didx = NoTangent()
+
+        return (df, du, dt, didx)
+    end
+
+    return slope, linear_interpolation_parameters_pullback
+end
+
+function _tangent_u(A::LinearInterpolation, t)
+    ... # TODO: How to handle sparsity?
+end
+
+function _tangent_t(A::LinearInterpolation, t)
+    ... # TODO: How to handle sparsity?
+end
+
+function ChainRulesCore.rrule(::typeof(_interpolate), A::AType, t, iguess) where {AType}
+    u = _interpolate(A, t, iguess)[1]
+    function _interpolate_pullback(Δ)
+        df = NoTangent()
+        dA = Tangent{AType}(; u = _tangent_u(A, t), t = _tangent_t(A, t))
+        dt = @thunk(derivative(A, t)*Δ)
+        diguess = NoTangent()
+        return df, dA, dt, diguess
+    end
+    u, _interpolate_pullback
 end
 
 function ChainRulesCore.frule((_, _, Δt), ::typeof(_interpolate), A::AbstractInterpolation,

diff --git a/src/parameter_caches.jl b/src/parameter_caches.jl
@@ -3,11 +3,16 @@ struct LinearParameterCache{pType}
 end
 
 function LinearParameterCache(u, t)
-    slope = linear_interpolation_parameters.(Ref(u), Ref(t), 1:(length(t) - 1))
+    slope_prototype = linear_interpolation_parameters(u, t, 1)
+    idxs = 1:(length(t) - 1)
+    slope = [zero(slope_prototype) for i in idxs]
+    for idx in idxs
+        slope[idx] = linear_interpolation_parameters(u, t, idx)
+    end
     return LinearParameterCache(slope)
 end
 
-function linear_interpolation_parameters(u, t, idx)
+function linear_interpolation_parameters(u, t, idx::Integer)
     Δu = u isa AbstractMatrix ? u[:, idx + 1] - u[:, idx] : u[idx + 1] - u[idx]
     Δt = t[idx + 1] - t[idx]
     slope = Δu / Δt