Make sure relu, sigmoid_fast, and tanh_fast can_Avx

Author: chriselrod

src/costs.jl

@@ -244,7 +244,8 @@ const COST = Dict{Symbol,InstructionCost}(
     :cospi_fast => InstructionCost(18,15.0,68.0,26),
     :sincospi_fast => InstructionCost(25,22.0,70.0,26),
     :tanh => InstructionCost(40,40.0,40.0,26), # FIXME
-    # :tanh_fast => InstructionCost(25,22.0,70.0,26), # FIXME
+    :tanh_fast => InstructionCost(25,22.0,70.0,26), # FIXME
+    :sigmoid_fast => InstructionCost(25,22.0,70.0,26), # FIXME
     :identity => InstructionCost(0,0.0,0.0,0),
     :adjoint => InstructionCost(0,0.0,0.0,0),
     :conj => InstructionCost(0,0.0,0.0,0),
@@ -475,7 +476,8 @@ const FUNCTIONSYMBOLS = IdDict{Type{<:Function},Instruction}(
     typeof(Base.FastMath.sincos_fast) => :sincos,
     typeof(SLEEFPirates.sincos) => :sincos,
     typeof(Base.tanh) => :tanh,
-    # typeof(SLEEFPirates.tanh_fast) => :tanh_fast,
+    typeof(tanh_fast) => :tanh_fast,
+    typeof(sigmoid_fast) => :sigmoid_fast,
     typeof(max) => :max,
     typeof(min) => :min,
     typeof(max_fast) => :max_fast,

test/can_avx.jl

@@ -11,7 +11,11 @@
     @test LoopVectorization.ArrayInterface.can_avx(Base.FastMath.add_fast)
     @test LoopVectorization.ArrayInterface.can_avx(/)
     @test LoopVectorization.ArrayInterface.can_avx(sqrt)
+    @test LoopVectorization.ArrayInterface.can_avx(tanh_fast)
+    @test LoopVectorization.ArrayInterface.can_avx(sigmoid_fast)
+    @test LoopVectorization.ArrayInterface.can_avx(VectorizationBase.relu)
     @test !LoopVectorization.ArrayInterface.can_avx(clenshaw)
+    @test !LoopVectorization.ArrayInterface.can_avx(println)
 
 end