Remove use of implicit NLSolversBase.value and NLSolversBase.gradient

Author: devmotion

src/LineSearches.jl

@@ -29,10 +29,10 @@ function make_ϕdϕ(df, x_new, x, s)
         x_new .= x .+ α.*s
 
         # Evaluate ∇f(x+α*s)
-        NLSolversBase.value_gradient!(df, x_new)
+        f_x_new, g_x_new = NLSolversBase.value_gradient!(df, x_new)
 
         # Calculate ϕ(a_i), ϕ'(a_i)
-        NLSolversBase.value(df), real(dot(NLSolversBase.gradient(df), s))
+        return f_x_new, real(dot(g_x_new, s))
     end
     ϕdϕ
 end
@@ -42,48 +42,18 @@ function make_ϕ_dϕ(df, x_new, x, s)
         x_new .= x .+ α.*s
 
         # Evaluate ∇f(x+α*s)
-        NLSolversBase.gradient!(df, x_new)
+        g_x_new = NLSolversBase.gradient!(df, x_new)
 
         # Calculate ϕ'(a_i)
-        real(dot(NLSolversBase.gradient(df), s))
+        return real(dot(g_x_new, s))
     end
     make_ϕ(df, x_new, x, s), dϕ
 end
 function make_ϕ_dϕ_ϕdϕ(df, x_new, x, s)
-    function dϕ(α)
-        # Move a distance of alpha in the direction of s
-        x_new .= x .+ α.*s
-
-        # Evaluate f(x+α*s) and ∇f(x+α*s)
-        NLSolversBase.gradient!(df, x_new)
-
-        # Calculate ϕ'(a_i)
-        real(dot(NLSolversBase.gradient(df), s))
-    end
-    function ϕdϕ(α)
-        # Move a distance of alpha in the direction of s
-        x_new .= x .+ α.*s
-
-        # Evaluate ∇f(x+α*s)
-        NLSolversBase.value_gradient!(df, x_new)
-
-        # Calculate ϕ'(a_i)
-        NLSolversBase.value(df), real(dot(NLSolversBase.gradient(df), s))
-    end
-    make_ϕ(df, x_new, x, s), dϕ, ϕdϕ
+    make_ϕ_dϕ(df, x_new, x, s)..., make_ϕdϕ(df, x_new, x, s)
 end
 function make_ϕ_ϕdϕ(df, x_new, x, s)
-    function ϕdϕ(α)
-        # Move a distance of alpha in the direction of s
-        x_new .= x .+ α.*s
-
-        # Evaluate ∇f(x+α*s)
-        NLSolversBase.value_gradient!(df, x_new)
-
-        # Calculate ϕ'(a_i)
-        NLSolversBase.value(df), real(dot(NLSolversBase.gradient(df), s))
-    end
-    make_ϕ(df, x_new, x, s), ϕdϕ
+    make_ϕ(df, x_new, x, s), make_ϕdϕ(df, x_new, x, s)
 end
 
 include("types.jl")

src/initialguess.jl

@@ -74,7 +74,8 @@ function (is::InitialQuadratic{T})(ls, state, phi_0, dphi_0, df) where T
         # If we're at the first iteration
         αguess = is.α0
     else
-        αguess = 2 * (NLSolversBase.value(df) - state.f_x_previous) / dphi_0
+        f_x_ls = NLSolversBase.value!(df, state.x_ls)
+        αguess = 2 * (f_x_ls - state.f_x_previous) / dphi_0
         αguess = NaNMath.max(is.αmin, state.alpha*is.ρ, αguess)
         αguess = NaNMath.min(is.αmax, αguess)
         # if αguess ≈ 1, then make it 1 (Newton-type behaviour)
@@ -180,8 +181,8 @@ function (is::InitialHagerZhang)(ls::Tls, state, phi_0, dphi_0, df) where Tls
         # and the user has not provided an initial step size (is.α0 is NaN),
         # then we
         # pick the initial step size according to HZ #I0
-        state.alpha = _hzI0(state.x, NLSolversBase.gradient(df),
-                            NLSolversBase.value(df),
+        f_x, g_x = NLSolversBase.value_gradient!(df, state.x)
+        state.alpha = _hzI0(state.x, g_x, f_x,
                             is.αmax,
                             convert(eltype(state.x), is.ψ0)) # Hack to deal with type instability between is{T} and state.x
         if Tls <: HagerZhang

test/arbitrary_precision.jl

@@ -98,6 +98,7 @@ doublefloatstypes = [Double64, Double32, Double16]
     ls = HagerZhang{T}()
     state = getstate()
     state.f_x_previous = 2*phi_0
+    state.x_ls = zero(x)
     is = InitialQuadratic{T}(snap2one=(convert(T, 0.9),convert(T, Inf)))
     is(ls, state, phi_0, dphi_0, df)
     @test !isnan(state.alpha)
@@ -107,6 +108,7 @@ doublefloatstypes = [Double64, Double32, Double16]
     ls = HagerZhang{T}()
     state = getstate()
     state.f_x_previous = 2*phi_0
+    state.x_ls = zero(x)
     is = InitialQuadratic{T}(snap2one=(convert(T, 0.75),convert(T, Inf)))
     is(ls, state, phi_0, dphi_0, df)
     @test !isnan(state.alpha)

test/initial.jl

@@ -88,6 +88,7 @@
     ls = HagerZhang()
     state = getstate()
     state.f_x_previous = 2*phi_0
+    state.x_ls = zero(x)
     is = InitialQuadratic(snap2one=(0.9,Inf))
     is(ls, state, phi_0, dphi_0, df)
     @test state.alpha == 1.0
@@ -97,6 +98,7 @@
     ls = HagerZhang()
     state = getstate()
     state.f_x_previous = 2*phi_0
+    state.x_ls = zero(x)
     is = InitialQuadratic(snap2one=(0.75,Inf))
     is(ls, state, phi_0, dphi_0, df)
     @test state.alpha == 1.0