Modify stride penalty, and fix removal of check for indices like A[3 - i] in search for constant offsets (we need a subtraction for a reversal, so constant offset [3 in this example] may as well be rolled in there).

Author: chriselrod

src/determinestrategy.jl

@@ -519,8 +519,11 @@ function stride_penalty(ls::LoopSet, op::Operation, order::Vector{Symbol}, loopf
         opstrides[1] = 1.0
     end
     # loops = map(s -> getloop(ls, s), loopdeps)
+    l = length(getloop(ls, first(loopdeps)))
     for i ∈ 2:length(loopdeps)
-        opstrides[i] = opstrides[i-1] * length(getloop(ls, loopdeps[i-1]))
+        looplength = length(getloop(ls, loopdeps[i-1]))
+        opstrides[i] = opstrides[i-1] * looplength
+        l *= looplength
         # opstrides[i] = opstrides[i-1] * length(loops[i-1])
     end
     penalty = 0.0
@@ -530,7 +533,7 @@ function stride_penalty(ls::LoopSet, op::Operation, order::Vector{Symbol}, loopf
             penalty += loopfreqs[i] * opstrides[id]
         end
     end
-    penalty
+    penalty * l
 end
 function stride_penalty(ls::LoopSet, order::Vector{Symbol})
     stridepenaltydict = Dict{Symbol,Vector{Float64}}()
@@ -545,7 +548,7 @@ function stride_penalty(ls::LoopSet, order::Vector{Symbol})
             push!(v, stride_penalty(ls, op, order, loopfreqs))
         end
     end
-    sum(maximum, values(stridepenaltydict)) #* prod(length, ls.loops) / 1024^length(order)
+    sum(maximum, values(stridepenaltydict)) * 10 / 1024^length(order) #* prod(length, ls.loops)
 end
 function isoptranslation(ls::LoopSet, op::Operation, unrollsyms::UnrollSymbols)
     @unpack u₁loopsym, u₂loopsym, vectorized = unrollsyms

src/memory_ops_common.jl

@@ -112,7 +112,7 @@ function checkforoffset!(
     factor = f === :+ ? 1 : -1
     arg1 = ind.args[2]
     arg2 = ind.args[3]
-    if arg1 isa Integer# && isone(factor)
+    if arg1 isa Integer && isone(factor) # we want to return false when we're subtracting the index, e.g. A[3 - i] 
         if arg2 isa Symbol && arg2 ∈ ls.loopsymbols
             addoffset!(ls, indices, offsets, loopedindex, loopdependencies, arg2, arg1 * factor)
         elseif arg2 isa Expr

test/gemv.jl

@@ -3,7 +3,7 @@ using Test
 # T = Float32
 @testset "GEMV" begin
     # Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (3, 4) : (4, 6)
-    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (2, 6) : (2, 10)
+    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (2, 6) : (2, 8)
     gemvq = :(for i ∈ eachindex(y)
               yᵢ = 0.0
               for j ∈ eachindex(x)

test/miscellaneous.jl

@@ -4,7 +4,7 @@ using Test
 
 @testset "Miscellaneous" begin
 
-    Unum, Tnum = LoopVectorization.REGISTER_COUNT == 16 ? (2, 6) : (2, 10)
+    Unum, Tnum = LoopVectorization.REGISTER_COUNT == 16 ? (2, 6) : (2, 8)
     dot3q = :(for m ∈ 1:M, n ∈ 1:N
               s += x[m] * A[m,n] * y[n]
               end);