Fix unnecessary allocations in maximization due to insufficient broadcasting (#1189)

* Fix unnecessary allocations in maximization due to insufficient broadcasting

* Fix a few more cases

---------

Co-authored-by: Patrick Kofod Mogensen <[email protected]>

Author: devmotion

src/maximize.jl

@@ -9,12 +9,16 @@ res(r::MaximizationWrapper) = r.res
 # Univariate warppers
 # ==============================================================================
 function maximize(f, lb::Real, ub::Real, method::AbstractOptimizer; kwargs...)
-    fmax = x -> -f(x)
+    fmax = let f=f
+        x -> -f(x)
+    end
     MaximizationWrapper(optimize(fmax, lb, ub, method; kwargs...))
 end
 
 function maximize(f, lb::Real, ub::Real; kwargs...)
-    fmax = x -> -f(x)
+    fmax = let f=f
+        x -> -f(x)
+    end
     MaximizationWrapper(optimize(fmax, lb, ub; kwargs...))
 end
 
@@ -23,7 +27,9 @@ end
 # Multivariate warppers
 # ==============================================================================
 function maximize(f, x0::AbstractArray, options::Options = Options())
-    fmax = x -> -f(x)
+    fmax = let f=f
+        x -> -f(x)
+    end
     MaximizationWrapper(optimize(fmax, x0, options))
 end
 function maximize(
@@ -32,7 +38,9 @@ function maximize(
     method::AbstractOptimizer,
     options = Optim.Options()
 )
-    fmax = x -> -f(x)
+    fmax = let f=f
+        x -> -f(x)
+    end
     MaximizationWrapper(optimize(fmax, x0, method, options))
 end
 function maximize(
@@ -42,8 +50,12 @@ function maximize(
     method::AbstractOptimizer,
     options = Optim.Options()
 )
-    fmax = x -> -f(x)
-    gmax = (G, x) -> (g(G, x); G .= -G)
+    fmax = let f=f
+        x -> -f(x)
+    end
+    gmax = let g=g
+        (G, x) -> (g(G, x); G .= .-G)
+    end
     MaximizationWrapper(optimize(fmax, gmax, x0, method, options))
 end
 
@@ -55,9 +67,15 @@ function maximize(
     method::AbstractOptimizer,
     options = Optim.Options()
 )
-    fmax = x -> -f(x)
-    gmax = (G, x) -> (g(G, x); G .= -G)
-    hmax = (H, x) -> (h(H, x); H .= -H)
+    fmax = let f=f
+        x -> -f(x)
+    end
+    gmax = let g=g
+        (G, x) -> (g(G, x); G .= .-G)
+    end
+    hmax = let h=h
+        (H, x) -> (h(H, x); H .= .-H)
+    end
     MaximizationWrapper(optimize(fmax, gmax, hmax, x0, method, options))
 end
 

src/multivariate/solvers/first_order/cg.jl

@@ -110,7 +110,7 @@ function reset!(cg::ConjugateGradient, cgs::ConjugateGradientState, obj, x)
     if cg.P !== nothing
         project_tangent!(cg.manifold, cgs.pg, x)
     end
-    cgs.s .= -cgs.pg
+    cgs.s .= .-cgs.pg
     cgs.f_x_previous = typeof(cgs.f_x_previous)(NaN)
 end
 function initial_state(method::ConjugateGradient, options, d, initial_x)

src/multivariate/solvers/second_order/krylov_trust_region.jl

@@ -113,7 +113,7 @@ function cg_steihaug!(
 
     fill!(z, 0.0)  # the search direction is initialized to the 0 vector,
     r .= g  # so at first the whole gradient is the residual.
-    d .= -r # the first direction is the direction of steepest descent.
+    d .= .-r # the first direction is the direction of steepest descent.
     rho0 = 1e100  # just a big number
 
     state.cg_iters = 0

test/general/maximize.jl

@@ -37,8 +37,8 @@
     g! = gradient(prob)
     h! = hessian(prob)
     fmax(x) = -f(x)
-    gmax = (G, x) -> (g!(G, x); G .= -G)
-    hmax = (H, x) -> (h!(H, x); H .= -H)
+    gmax = (G, x) -> (g!(G, x); G .= .-G)
+    hmax = (H, x) -> (h!(H, x); H .= .-H)
 
     function test_same_content(f1, f2)
         for prop in (:minimizer, :minimum)