Only mask updating reductions.

Author: chriselrod

Project.toml

@@ -17,7 +17,7 @@ Parameters = "0.12.0"
 SIMDPirates = "0.1.0"
 SLEEFPirates = "0.1.1"
 VectorizationBase = "0.1.3"
-julia = "1.0.0"
+julia = "1.3.0"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

src/graphs.jl

@@ -1,9 +1,9 @@
 # lv(x) = Expr(:(.), :LoopVectorization, QuoteNode(x))
-@static if VERSION  < v"1.3.0"
-    lv(x) = Expr(:(.), :LoopVectorization, QuoteNode(x))
-else
+# @static if VERSION  < v"1.3.0"
+    # lv(x) = Expr(:(.), :LoopVectorization, QuoteNode(x))
+# else
     lv(x) = GlobalRef(LoopVectorization, x)
-end
+# end
 
 isdense(::Type{<:DenseArray}) = true
 

src/lowering.jl

@@ -281,7 +281,7 @@ function lower_compute!(
     # @show instr parentsunrolled
     # parentsyms = [opp.variable for opp ∈ parents(op)]
     Uiter = opunrolled ? U - 1 : 0
-    maskreduct = mask !== nothing && isreduction(op)#any(opp -> opp.variable === var, parents_op)
+    maskreduct = mask !== nothing && isreduction(op) && any(opp -> opp.variable === var, parents_op)
     # if a parent is not unrolled, the compiler should handle broadcasting CSE.
     # because unrolled/tiled parents result in an unrolled/tiled dependendency,
     # we handle both the tiled and untiled case here.