Add VectorizedStatistics dependency, faster sorting in linterp1s! w…

…ith one less buffer required

Project.toml

@@ -1,7 +1,7 @@
 name = "StatGeochemBase"
 uuid = "61e559cd-58b4-4257-8528-26bb26ff2b9a"
 authors = ["C. Brenhin Keller"]
-version = "0.4.2"
+version = "0.4.3"
 
 [deps]
 Colors = "5ae59095-9a9b-59fe-a467-6f913c188581"
@@ -10,6 +10,7 @@ LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 NaNStatistics = "b946abbf-3ea7-4610-9019-9858bfdeaf2d"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
+VectorizedStatistics = "3b853605-1c98-4422-8364-4bd93ee0529e"
 
 [compat]
 Colors = "0.10 - 0.12"
@@ -18,6 +19,7 @@ LoopVectorization = "0.11, 0.12"
 NaNStatistics = "0.3, 0.4, 0.5, 0.6"
 SpecialFunctions = "0.5 - 0.10, 1, 2"
 VectorizationBase = "0.18, 0.19, 0.20, 0.21"
+VectorizedStatistics = "0.4.4"
 julia = "1.6"
 
 [extras]

src/Geochronology.jl

@@ -245,8 +245,8 @@
         # terms that don't depend on a or b
         α = sqrt.(ωx .* ωy)
 
-        x̄ = _vmean(x, :)
-        ȳ = _vmean(y, :)
+        x̄ = vmean(x)
+        ȳ = vmean(y)
         r = sum((x .- x̄).*(y .- ȳ)) ./ (sqrt(sum((x .- x̄).^2)) * sqrt(sum((y .- ȳ).^2)))
 
         ## 3. Perform the York fit (must iterate)

src/Interpolations.jl

@@ -228,18 +228,24 @@
 
     """
     ```julia
-    linterp1s!(yq::DenseArray, x::AbstractArray, y::AbstractArray, xq; extrapolate=:Linear, sort_index=ones(Int, length(x)), knot_index=ones(Int, length(xq)))
+    linterp1s!(yq::AbstractArray, x::AbstractArray, y::AbstractArray, xq; extrapolate=:Linear)
+    linterp1s!(yq::AbstractArray, knot_index::AbstractArray{Int}, x::AbstractArray, y::AbstractArray, xq::AbstractArray; extrapolate=:Linear)
     ```
-    In-place variant of `linterp1s`. Will sort `x` and permute `y` to match.
+    In-place variant of `linterp1s`.
+    Will sort `x` and permute `y` to match, before interpolating at `xq` and storing the result in `yq`.
+
+    An optional temporary working array `knot_index = similar(xq, Int)` may be provided to fully eliminate allocations.
     """
-    function linterp1s!(yq::AbstractArray, x::AbstractArray, y::AbstractArray, xq; extrapolate=:Linear, sort_index::AbstractVector{Int}=ones(Int, length(x)), knot_index::AbstractVector{Int}=ones(Int, length(xq)))
-        @assert eachindex(sort_index) === eachindex(x) === eachindex(y)
-        @assert eachindex(knot_index) === eachindex(xq)
-        if !issorted(x)
-            sortperm!(sort_index, x) # indices to construct sorted array
-            permute!(x, sort_index)
-            permute!(y, sort_index)
-        end
+    function linterp1s!(yq::AbstractArray, x::AbstractArray, y::AbstractArray, xq; extrapolate=:Linear)
+        @assert eachindex(xq) === eachindex(yq)
+        @assert eachindex(x) === eachindex(y)
+        vsort!(y, x) # Sort x and permute y to match
+        return _linterp1!(yq, x, y, xq, extrapolate)
+    end
+    function linterp1s!(yq::AbstractArray, knot_index::AbstractArray{Int}, x::AbstractArray, y::AbstractArray, xq::AbstractArray; extrapolate=:Linear)
+        @assert eachindex(knot_index) === eachindex(xq) === eachindex(yq)
+        @assert eachindex(x) === eachindex(y)
+        vsort!(y, x) # Sort x and permute y to match
         return _linterp1!(yq, knot_index, x, y, xq, extrapolate)
     end
     export linterp1s!

src/Math.jl

@@ -41,16 +41,6 @@
     end
     export fast_inv_sqrt
 
-## --- Fast vectorized mean
-
-    function _vmean(A::DenseArray{T}, ::Colon) where {T<:Number}
-        m = zero(T)
-        @turbo for i ∈ eachindex(A)
-            m += A[i]
-        end
-        return m / length(A)
-    end
-
 
 ## --- Base-10 version of log1p
 #

src/StatGeochemBase.jl

@@ -1,6 +1,7 @@
 module StatGeochemBase
 
     using NaNStatistics
+    using VectorizedStatistics
 
     # AVX vectorziation tools
     using LoopVectorization

test/testInterpolations.jl

@@ -30,7 +30,7 @@
     @test linterp1s!(similar(xq), 1:10, 21:30, xq) == [25.0, 25.5, 26.0]
     @test linterp1s!(similar(xq), collect(10:-1:1), collect(21:30), xq) == [26.0, 25.5, 25.0]
     xq = 5:0.5:7
-    @test linterp1s!(similar(xq), collect(10:-1:1), collect(21:30), xq; knot_index=ones(Int,5), sort_index=ones(Int,10)) == [26.0, 25.5, 25.0, 24.5, 24]
+    @test linterp1s!(similar(xq), rand(Int,length(xq)), collect(10:-1:1), collect(21:30), xq) == [26.0, 25.5, 25.0, 24.5, 24]
     xq = 0:0.01:1
     x = rand(200)
     y = rand(200)