Merge pull request #29 from JuliaSIMD/check-empty

`@turbo check_empty=true` in most all cases

Author: brenhinkeller

Project.toml

@@ -1,7 +1,7 @@
 name = "VectorizedStatistics"
 uuid = "3b853605-1c98-4422-8364-4bd93ee0529e"
 authors = ["C. Brenhin Keller", "Chris Elrod"]
-version = "0.5.4"
+version = "0.5.5"
 
 [deps]
 LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"

src/argsort.jl

@@ -2,7 +2,7 @@
 function sortnans!(I::AbstractArray, A::AbstractArray, iₗ::Int=firstindex(A), iᵤ::Int=lastindex(A))
     # Count up NaNs
     Nₙₐₙ = 0
-    @turbo for i = iₗ:iᵤ
+    @turbo check_empty=true for i = iₗ:iᵤ
         Nₙₐₙ += A[i] != A[i]
     end
     # If none, return early

src/quicksort.jl

@@ -45,7 +45,7 @@ end
 function sortnans!(A::AbstractArray, iₗ::Int=firstindex(A), iᵤ::Int=lastindex(A))
     # Count up NaNs
     Nₙₐₙ = 0
-    @turbo for i = iₗ:iᵤ
+    @turbo check_empty=true for i = iₗ:iᵤ
         Nₙₐₙ += A[i] != A[i]
     end
     # If none, return early
@@ -80,7 +80,7 @@ function quickselect!(A::AbstractArray, iₗ::Int=firstindex(A), iᵤ::Int=lasti
 
     # Count up elements that must be moved to upper partition
     Nᵤ = 0
-    @turbo for i = (iₗ+1):iᵤ
+    @turbo check_empty=true for i = (iₗ+1):iᵤ
         Nᵤ += A[i] >= pivot
     end
     Nₗ = N - Nᵤ
@@ -103,9 +103,15 @@ function quickselect!(A::AbstractArray, iₗ::Int=firstindex(A), iᵤ::Int=lasti
     iₚ = iₗ + Nₗ - 1
     A[iₗ], A[iₚ] = A[iₚ], A[iₗ]
     # Recurse: select from partition containing k
-    (iₗ <= k < iₚ) && quickselect!(A, iₗ, iₚ, k)
-    (iₚ < k <= iᵤ) && quickselect!(A, iₚ+1, iᵤ, k)
-    return A[k]
+    if iₚ==k
+        return A[k]
+    elseif k < iₚ
+        Nₗ == 2 && return A[iₗ]
+        quickselect!(A, iₗ, iₚ, k)
+    else
+        Nᵤ == 2 && return A[iᵤ]
+        quickselect!(A, iₚ+1, iᵤ, k)
+    end
 end
 
 

src/vcov.jl

@@ -1,7 +1,7 @@
 function _vcov(x::AbstractVector, y::AbstractVector, corrected::Bool, μᵪ::Number, μᵧ::Number, multithreaded::False)
     # Calculate covariance
     σᵪᵧ = zero(promote_type(typeof(μᵪ), typeof(μᵧ), Int))
-    @turbo for i ∈ indices((x,y))
+    @turbo check_empty=true for i ∈ indices((x,y))
             δᵪ = x[i] - μᵪ
             δᵧ = y[i] - μᵧ
             σᵪᵧ += δᵪ * δᵧ
@@ -13,7 +13,7 @@ end
 function _vcov(x::AbstractVector, y::AbstractVector, corrected::Bool, μᵪ::Number, μᵧ::Number, multithreaded::True)
     # Calculate covariance
     σᵪᵧ = zero(promote_type(typeof(μᵪ), typeof(μᵧ), Int))
-    @tturbo for i ∈ indices((x,y))
+    @tturbo check_empty=true for i ∈ indices((x,y))
             δᵪ = x[i] - μᵪ
             δᵧ = y[i] - μᵧ
             σᵪᵧ += δᵪ * δᵧ

src/vmean.jl

@@ -55,7 +55,7 @@ function _vmean(A, ::Colon, multithreaded::False)
     # Promote type of accumulator to avoid overflow
     Tₒ = Base.promote_op(/, eltype(A), Int)
     Σ = zero(Tₒ)
-    @turbo for i ∈ eachindex(A)
+    @turbo check_empty=true for i ∈ eachindex(A)
         Σ += A[i]
     end
     return Σ / length(A)
@@ -155,7 +155,7 @@ function _vmean(A, ::Colon, multithreaded::True)
     # Promote type of accumulator to avoid overflow
     Tₒ = Base.promote_op(/, eltype(A), Int)
     Σ = zero(Tₒ)
-    @tturbo for i ∈ eachindex(A)
+    @tturbo check_empty=true for i ∈ eachindex(A)
         Σ += A[i]
     end
     return Σ / length(A)

src/vstd.jl

@@ -39,13 +39,13 @@ sqrt!(x, multithreaded::Symbol) = sqrt!(x, (multithreaded===:auto && length(x) >
 sqrt!(x, multithreaded::Bool) = sqrt!(x, static(multithreaded))
 sqrt!(x::Number, multithreaded::StaticBool) = sqrt(x)
 function sqrt!(A::AbstractArray, multithreaded::False)
-    @turbo for i ∈ eachindex(A)
+    @turbo check_empty=true for i ∈ eachindex(A)
         A[i] = sqrt(A[i])
     end
     return A
 end
 function sqrt!(A::AbstractArray, multithreaded::True)
-    @tturbo for i ∈ eachindex(A)
+    @tturbo check_empty=true for i ∈ eachindex(A)
         A[i] = sqrt(A[i])
     end
     return A

src/vsum.jl

@@ -54,7 +54,7 @@ function _vsum(A, ::Colon,  multithreaded::False)
     # Promote type of accumulator to avoid overflow
     Tₒ = Base.promote_op(+, eltype(A), Int)
     Σ = zero(Tₒ)
-    @turbo for i ∈ eachindex(A)
+    @turbo check_empty=true for i ∈ eachindex(A)
         Σ += A[i]
     end
     return Σ
@@ -149,7 +149,7 @@ function _vsum(A, ::Colon, multithreaded::True)
     # Promote type of accumulator to avoid overflow
     Tₒ = Base.promote_op(+, eltype(A), Int)
     Σ = zero(Tₒ)
-    @tturbo for i ∈ eachindex(A)
+    @tturbo check_empty=true for i ∈ eachindex(A)
         Σ += A[i]
     end
     return Σ

src/vvar.jl

@@ -48,7 +48,7 @@ function _vvar(μ::Number, corrected::Bool, A, ::Colon, multithreaded::False)
     # Reduce all the dims!
     n = length(A)
     σ² = zero(typeof(μ))
-    @turbo for i ∈ eachindex(A)
+    @turbo check_empty=true for i ∈ eachindex(A)
         δ = A[i] - μ
         σ² += δ * δ
     end
@@ -58,7 +58,7 @@ function _vvar(μ::Number, corrected::Bool, A, ::Colon, multithreaded::True)
     # Reduce all the dims!
     n = length(A)
     σ² = zero(typeof(μ))
-    @tturbo for i ∈ eachindex(A)
+    @tturbo check_empty=true for i ∈ eachindex(A)
         δ = A[i] - μ
         σ² += δ * δ
     end
@@ -72,12 +72,12 @@ function _vvar(::Nothing, corrected::Bool, A, ::Colon, multithreaded::False)
     n = length(A)
     Tₒ = Base.promote_op(/, eltype(A), Int)
     Σ = zero(Tₒ)
-    @turbo for i ∈ eachindex(A)
+    @turbo check_empty=true for i ∈ eachindex(A)
             Σ += A[i]
     end
     μ = Σ / n
     σ² = zero(typeof(μ))
-    @turbo for i ∈ eachindex(A)
+    @turbo check_empty=true for i ∈ eachindex(A)
             δ = A[i] - μ
             σ² += δ * δ
     end
@@ -88,12 +88,12 @@ function _vvar(::Nothing, corrected::Bool, A, ::Colon, multithreaded::True)
     n = length(A)
     Tₒ = Base.promote_op(/, eltype(A), Int)
     Σ = zero(Tₒ)
-    @tturbo for i ∈ eachindex(A)
+    @tturbo check_empty=true for i ∈ eachindex(A)
             Σ += A[i]
     end
     μ = Σ / n
     σ² = zero(typeof(μ))
-    @tturbo for i ∈ eachindex(A)
+    @tturbo check_empty=true for i ∈ eachindex(A)
             δ = A[i] - μ
             σ² += δ * δ
     end