gradients with min are the problem

Author: Michael Abbott

src/macro.jl

@@ -1071,7 +1071,7 @@ function make_many_actors(act!, args, ex1, outer::Vector, ex3, inner::Vector, ex
         isempty(store.unsafeleft)
     else # working on ∇act!
         isempty(store.unsaferight) &&
-            store.redfun == :+  && store.grad != :Dual # Disable @avx except for simplest grad, #53
+            store.redfun == :+  # Disable @avx for min/max grad, #53
     end
 
     if safe && store.avx != false && isdefined(store.mod, :LoopVectorization)

test/gradients.jl

@@ -259,24 +259,24 @@ if Tullio._GRAD[] != :Dual
         dv = ForwardDiff.gradient(v -> sum(f6(m2,v)), v2)
         @test dv ≈ _gradient(sum∘f6, m2, v2)[2]
 
-        f7(x,y) = @tullio (max) z[i] := x[i,j]^2 / sqrt(y[i]) + exp(y[j])
+        f7(x,y) = @tullio (max) z[i] := x[i,j]^2 / sqrt(y[i]) + exp(y[j]) avx=false
 
         dm = ForwardDiff.gradient(m -> sum(f7(m,v2)), m2)
-        @test dm ≈_gradient(sum∘f7, m2, v2)[1]
+        @test dm ≈_gradient(sum∘f7, m2, v2)[1]  # gives wrong answers with avx, 1.4 in tests
         dv = ForwardDiff.gradient(v -> sum(f7(m2,v)), v2)
         @test dv ≈ _gradient(sum∘f7, m2, v2)[2]
 
-        f8(x,y) = @tullio (max) z[i,l] := log(x[i,j,k,l]) / y[j]^1/3 avx=false # gives wrong answers
+        f8(x,y) = @tullio (max) z[i,l] := log(x[i,j,k,l]) / y[j]^1/3 avx=false
         f9(x,y) = @tullio (min) z[i,j] := log(x[i,j,k,l]) / y[j]^1/3 avx=false
 
         dm = ForwardDiff.gradient(m -> sum(f8(m,v2)), m4)
-        @test dm ≈_gradient(sum∘f8, m4, v2)[1]
+        @test dm ≈ _gradient(sum∘f8, m4, v2)[1]  # gives wrong answers with avx, 1.5 in tests
         dv = ForwardDiff.gradient(v -> sum(f8(m4,v)), v2)
         @test dv ≈ _gradient(sum∘f8, m4, v2)[2]
         dm = ForwardDiff.gradient(m -> sum(f9(m,v2)), m4)
-        @test dm ≈_gradient(sum∘f9, m4, v2)[1]
+        @test dm ≈_gradient(sum∘f9, m4, v2)[1]  # gives wrong answers with avx, repl
         dv = ForwardDiff.gradient(v -> sum(f9(m4,v)), v2)
-        @test dv ≈ _gradient(sum∘f9, m4, v2)[2]
+        @test dv ≈ _gradient(sum∘f9, m4, v2)[2]  # gives wrong answers with avx
 
     end