Reduce unrolling when unnecessary. Additionally, fix #143.

Author: chriselrod

src/determinestrategy.jl

@@ -402,7 +402,7 @@ function solve_unroll(X, R, u₁L, u₂L, u₁step, u₂step)
     u₂low = min(u₂low, maxunroll)
     u₁high = min(u₁high, maxunroll)
     u₂high = min(u₂high, maxunroll)
-    solve_unroll_iter(X, R, u₁L, u₂L, u₁low:u₁step:u₁high, u₂low:u₂step:u₂high)
+    solve_unroll_iter(X, R, u₁L, u₂L, u₁high:-u₁step:u₁low, u₂high:-u₂step:u₂low)
 end
 
 function solve_unroll_constU(R::AbstractVector, u₁::Int)

src/operations.jl

@@ -324,7 +324,7 @@ end
 function Operation(id::Int, var::Symbol, elementbytes::Int, instr, optype::OperationType, mpref::ArrayReferenceMetaPosition)
     Operation( id, var, elementbytes, instr, optype, mpref.loopdependencies, mpref.reduceddeps, mpref.parents, mpref.mref )
 end
-Base.:(==)(x::ArrayReferenceMetaPosition, y::ArrayReferenceMetaPosition) = x.mref.ref == y.mref.ref
+Base.:(==)(x::ArrayReferenceMetaPosition, y::ArrayReferenceMetaPosition) = x.mref == y.mref
 # Avoid memory allocations by using this for ops that aren't references
 const NOTAREFERENCE = ArrayReferenceMeta(ArrayReference(Symbol(""), Union{Symbol,Int}[]),Bool[],Symbol(""))
 const NOTAREFERENCEMP = ArrayReferenceMetaPosition(NOTAREFERENCE, NOPARENTS, Symbol[], Symbol[],Symbol(""))

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, 8)
+    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (1, 6) : (1, 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, 8)
+    Unum, Tnum = LoopVectorization.REGISTER_COUNT == 16 ? (1, 6) : (1, 8)
     dot3q = :(for m ∈ 1:M, n ∈ 1:N
               s += x[m] * A[m,n] * y[n]
               end);

test/offsetarrays.jl

@@ -87,9 +87,10 @@ using LoopVectorization.VectorizationBase: StaticUnitRange
     end
 
 
-    struct SizedOffsetMatrix{T,LR,UR,LC,RC} <: DenseMatrix{T}
+    struct SizedOffsetMatrix{T,LR,UR,LC,UC} <: DenseMatrix{T}
         data::Matrix{T}
     end
+    Base.size(::SizedOffsetMatrix{<:Any,LR,UR,LC,UC}) where {LR,UR,LC,UC} = (UR-LR+1,UC-LC+1)
     Base.axes(::SizedOffsetMatrix{T,LR,UR,LC,UC}) where {T,LR,UR,LC,UC} = (StaticUnitRange{LR,UR}(),StaticUnitRange{LC,UC}())
     Base.parent(A::SizedOffsetMatrix) = A.data
     @generated function LoopVectorization.stridedpointer(A::SizedOffsetMatrix{T,LR,UR,LC,RC}) where {T,LR,UR,LC,RC}