Improve performance of weighted sum (#778)

The current code is calling the `AbstractArray` matrix multiplication fallback,
which is slower than BLAS.

Author: nalimilan

src/weights.jl

@@ -382,6 +382,18 @@ Compute the weighted sum of an array `v` with weights `w`, optionally over the d
 """
 wsum(v::AbstractArray, w::AbstractVector, dims::Colon=:) = transpose(w) * vec(v)
 
+# Optimized methods (to ensure we use BLAS when possible)
+for W in (AnalyticWeights, FrequencyWeights, ProbabilityWeights, Weights)
+    @eval begin
+        wsum(v::AbstractArray, w::$W, dims::Colon) = transpose(w.values) * vec(v)
+    end
+end
+
+function wsum(A::AbstractArray, w::UnitWeights, dims::Colon)
+    length(A) != length(w) && throw(DimensionMismatch("Inconsistent array dimension."))
+    return sum(A)
+end
+
 ## wsum along dimension
 #
 #  Brief explanation of the algorithm:
@@ -605,12 +617,6 @@ optionally over the dimension `dims`.
 Base.sum(A::AbstractArray, w::AbstractWeights{<:Real}; dims::Union{Colon,Int}=:) =
     wsum(A, w, dims)
 
-function Base.sum(A::AbstractArray, w::UnitWeights; dims::Union{Colon,Int}=:)
-    a = (dims === :) ? length(A) : size(A, dims)
-    a != length(w) && throw(DimensionMismatch("Inconsistent array dimension."))
-    return sum(A, dims=dims)
-end
-
 ##### Weighted means #####
 
 function wmean(v::AbstractArray{<:Number}, w::AbstractVector)

test/weights.jl

@@ -476,7 +476,7 @@ end
 @testset "Sum, mean, quantiles and variance for unit weights" begin
     wt = uweights(Float64, 3)
 
-    @test sum([1.0, 2.0, 3.0], wt) ≈ 6.0
+    @test sum([1.0, 2.0, 3.0], wt) ≈ wsum([1.0, 2.0, 3.0], wt) ≈ 6.0
     @test mean([1.0, 2.0, 3.0], wt) ≈ 2.0
 
     @test sum(a, wt, dims=1) ≈ sum(a, dims=1)