Bump version, and reuse a couple preallocated vectors.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.6.2"
+version = "0.6.3"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"

src/determinestrategy.jl

@@ -67,8 +67,7 @@ end
 function evaluate_cost_unroll(
     ls::LoopSet, order::Vector{Symbol}, max_cost = typemax(Float64), vectorized::Symbol = first(order)
 )
-    # included_vars = Set{UInt}()
-    included_vars = fill(false, length(operations(ls)))
+    included_vars = fill!(resize!(ls.included_vars, length(operations(ls))), false)
     nested_loop_syms = Symbol[]#Set{Symbol}()
     total_cost = 0.0
     iter = 1.0
@@ -368,9 +367,9 @@ function evaluate_cost_tile(
     unrolled = order[2]
     ops = operations(ls)
     nops = length(ops)
-    included_vars = fill(false, nops)
+    included_vars = fill!(resize!(ls.included_vars, nops), false)
     unrolledtiled = fill(false, 2, nops)
-    descendentsininnerloop = fill(false, nops)
+    descendentsininnerloop = fill!(resize!(ls.place_after_loop, nops), false)
     innerloop = last(order)
     iters = fill(-99.9, nops)
     nested_loop_syms = Symbol[]# Set{Symbol}()

src/graphs.jl

@@ -142,22 +142,22 @@ Base.@propagate_inbounds Base.getindex(lo::LoopOrder, i...) = lo.oporder[LinearI
 
 # Must make it easy to iterate
 # outer_reductions is a vector of indices (within operation vectors) of the reduction operation, eg the vmuladd op in a dot product
+# O(N) search is faster at small sizes
 struct LoopSet
     loopsymbols::Vector{Symbol}
     loops::Vector{Loop}
     opdict::Dict{Symbol,Operation}
     operations::Vector{Operation} # Split them to make it easier to iterate over just a subset
     outer_reductions::Vector{Int} # IDs of reduction operations that need to be reduced at end.
     loop_order::LoopOrder
-    # stridesets::Dict{ShortVector{Symbol},ShortVector{Symbol}}
     preamble::Expr
     preamble_symsym::Vector{Tuple{Int,Symbol}}
     preamble_symint::Vector{Tuple{Int,Int}}
     preamble_symfloat::Vector{Tuple{Int,Float64}}
     preamble_zeros::Vector{Int}
     preamble_ones::Vector{Int}
     includedarrays::Vector{Symbol}
-    syms_aliasing_refs::Vector{Symbol} # O(N) search is faster at small sizes
+    syms_aliasing_refs::Vector{Symbol}
     refs_aliasing_syms::Vector{ArrayReferenceMeta}
     cost_vec::Matrix{Float64}
     reg_pres::Matrix{Int}

src/operation_evaluation_order.jl

@@ -61,10 +61,8 @@ function fillorder!(ls::LoopSet, order::Vector{Symbol}, loopistiled::Bool)
     end
     ops = operations(ls)
     nops = length(ops)
-    included_vars = resize!(ls.included_vars, nops)
-    fill!(included_vars, false)
-    place_after_loop = resize!(ls.place_after_loop, nops)
-    fill!(ls.place_after_loop, true)
+    included_vars = fill!(resize!(ls.included_vars, nops), false)        
+    place_after_loop = fill!(resize!(ls.place_after_loop, nops), true)
     # to go inside out, we just have to include all those not-yet included depending on the current sym
     empty!(lo)
     for _n ∈ 1:nloops