Fix and test leakyrelu (#505)

* ignore oftype

* broadcast fixes

* delete problematic line that accidentally wasn't removed

* no print

* fix order

* ifelse unroll

* Test leakyrelu

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.163"
+version = "0.12.164"
 
 
 [deps]

benchmark/looptests.jl

@@ -76,12 +76,12 @@ function jgemm!(𝐂, 𝐀ᵀ::Adjoint, 𝐁ᵀ::Adjoint)
   end
 end
 gemmavx!(𝐂, 𝐀, 𝐁) = @turbo for m ∈ indices((𝐀, 𝐂), 1), n ∈ indices((𝐁, 𝐂), 2)
-    𝐂ₘₙ = zero(eltype(𝐂))
-    for k ∈ indices((𝐀, 𝐁), (2, 1))
-      𝐂ₘₙ += 𝐀[m, k] * 𝐁[k, n]
-    end
-    𝐂[m, n] = 𝐂ₘₙ
+  𝐂ₘₙ = zero(eltype(𝐂))
+  for k ∈ indices((𝐀, 𝐁), (2, 1))
+    𝐂ₘₙ += 𝐀[m, k] * 𝐁[k, n]
   end
+  𝐂[m, n] = 𝐂ₘₙ
+end
 function gemmavx!(
   Cc::AbstractMatrix{Complex{T}},
   Ac::AbstractMatrix{Complex{T}},
@@ -102,12 +102,12 @@ function gemmavx!(
   end
 end
 gemmavxt!(𝐂, 𝐀, 𝐁) = @tturbo for m ∈ indices((𝐀, 𝐂), 1), n ∈ indices((𝐁, 𝐂), 2)
-    𝐂ₘₙ = zero(eltype(𝐂))
-    for k ∈ indices((𝐀, 𝐁), (2, 1))
-      𝐂ₘₙ += 𝐀[m, k] * 𝐁[k, n]
-    end
-    𝐂[m, n] = 𝐂ₘₙ
+  𝐂ₘₙ = zero(eltype(𝐂))
+  for k ∈ indices((𝐀, 𝐁), (2, 1))
+    𝐂ₘₙ += 𝐀[m, k] * 𝐁[k, n]
   end
+  𝐂[m, n] = 𝐂ₘₙ
+end
 function gemmavxt!(
   Cc::AbstractMatrix{Complex{T}},
   Ac::AbstractMatrix{Complex{T}},
@@ -204,11 +204,11 @@ function jdot3avx(x, A, y)
   s
 end
 jvexp!(b, a) = @inbounds for i ∈ eachindex(a)
-    b[i] = exp(a[i])
-  end
+  b[i] = exp(a[i])
+end
 jvexpavx!(b, a) = @turbo for i ∈ eachindex(a)
-    b[i] = exp(a[i])
-  end
+  b[i] = exp(a[i])
+end
 function jsvexp(a)
   s = zero(eltype(a))
   @inbounds for i ∈ eachindex(a)
@@ -242,12 +242,12 @@ function jgemv!(𝐲, 𝐀ᵀ::Adjoint, 𝐱)
   end
 end
 jgemvavx!(𝐲, 𝐀, 𝐱) = @turbo for i ∈ eachindex(𝐲)
-    𝐲ᵢ = zero(eltype(𝐲))
-    for j ∈ eachindex(𝐱)
-      𝐲ᵢ += 𝐀[i, j] * 𝐱[j]
-    end
-    𝐲[i] = 𝐲ᵢ
+  𝐲ᵢ = zero(eltype(𝐲))
+  for j ∈ eachindex(𝐱)
+    𝐲ᵢ += 𝐀[i, j] * 𝐱[j]
   end
+  𝐲[i] = 𝐲ᵢ
+end
 function jvar!(𝐬², 𝐀, x̄)
   @. s² = zero(eltype(𝐬²))
   @inbounds @fastmath for i ∈ 1:size(𝐀, 2)
@@ -258,14 +258,14 @@ function jvar!(𝐬², 𝐀, x̄)
   end
 end
 jvaravx!(𝐬², 𝐀, x̄) = @turbo for j ∈ eachindex(𝐬²)
-    𝐬²ⱼ = zero(eltype(𝐬²))
-    x̄ⱼ = x̄[j]
-    for i ∈ 1:size(𝐀, 2)
-      δ = 𝐀[j, i] - x̄ⱼ
-      𝐬²ⱼ += δ * δ
-    end
-    𝐬²[j] = 𝐬²ⱼ
+  𝐬²ⱼ = zero(eltype(𝐬²))
+  x̄ⱼ = x̄[j]
+  for i ∈ 1:size(𝐀, 2)
+    δ = 𝐀[j, i] - x̄ⱼ
+    𝐬²ⱼ += δ * δ
   end
+  𝐬²[j] = 𝐬²ⱼ
+end
 japlucBc!(D, a, B, c) = @. D = a + B * c';
 japlucBcavx!(D, a, B, c) = @turbo @. D = a + B * c';
 

benchmark/plotbenchmarks.jl

@@ -29,7 +29,8 @@ else
   # const COLOR_MAP = Dict{String,RGB{Float64}}()
   # const COLOR_MAP = Dict{String,RGB{Colors.N0f8}}()
   const COLOR_MAP64 = Dict{String,RGB{Float64}}()
-  getcolor(s::String) = get!(COLOR_MAP64, s) do
+  getcolor(s::String) =
+    get!(COLOR_MAP64, s) do
       COLORS[length(COLOR_MAP64)+1]
     end
   replace_and(str) = replace(str, '&' => "with")

ext/ForwardDiffExt.jl

@@ -157,8 +157,8 @@ end
   end
 end
 
-@generated function ifelse(
-  m::AbstractMask,
+@generated function _ifelse(
+  m::Union{AbstractMask,VecUnroll{<:Any,<:Any,Bit,<:AbstractMask}},
   x::ForwardDiff.Dual{TAG,V,P},
   y::ForwardDiff.Dual{TAG,V,P}
 ) where {TAG,V,P}
@@ -171,8 +171,8 @@ end
     ForwardDiff.Dual{$TAG}(z, ForwardDiff.Partials(p))
   end
 end
-@generated function ifelse(
-  m::AbstractMask,
+@generated function _ifelse(
+  m::Union{AbstractMask,VecUnroll{<:Any,<:Any,Bit,<:AbstractMask}},
   x::Number,
   y::ForwardDiff.Dual{TAG,V,P}
 ) where {TAG,V,P}
@@ -184,8 +184,8 @@ end
     ForwardDiff.Dual{$TAG}(z, ForwardDiff.Partials(p))
   end
 end
-@generated function ifelse(
-  m::AbstractMask,
+@generated function _ifelse(
+  m::Union{AbstractMask,VecUnroll{<:Any,<:Any,Bit,<:AbstractMask}},
   x::ForwardDiff.Dual{TAG,V,P},
   y::Number
 ) where {TAG,V,P}
@@ -197,6 +197,29 @@ end
     ForwardDiff.Dual{$TAG}(z, ForwardDiff.Partials(p))
   end
 end
+@inline ifelse(m::AbstractMask, x::ForwardDiff.Dual, y::Number) =
+  _ifelse(m, x, y)
+@inline ifelse(m::AbstractMask, x::ForwardDiff.Dual, y::ForwardDiff.Dual) =
+  _ifelse(m, x, y)
+@inline ifelse(m::AbstractMask, y::Number, x::ForwardDiff.Dual) =
+  _ifelse(m, y, x)
+
+@inline ifelse(
+  m::VecUnroll{<:Any,<:Any,Bit,<:AbstractMask},
+  x::ForwardDiff.Dual,
+  y::Number
+) = _ifelse(m, x, y)
+@inline ifelse(
+  m::VecUnroll{<:Any,<:Any,Bit,<:AbstractMask},
+  x::ForwardDiff.Dual,
+  y::ForwardDiff.Dual
+) = _ifelse(m, x, y)
+@inline ifelse(
+  m::VecUnroll{<:Any,<:Any,Bit,<:AbstractMask},
+  y::Number,
+  x::ForwardDiff.Dual
+) = _ifelse(m, y, x)
+
 @inline function SLEEFPirates.softplus(x::ForwardDiff.Dual{TAG}) where {TAG}
   val = ForwardDiff.value(x)
   expx = exp(val)

src/LoopVectorization.jl

@@ -108,7 +108,8 @@ using VectorizationBase:
   contract_or,
   collapse_or,
   max_mask,
-  maybestaticsize,zero_mask
+  maybestaticsize,
+  zero_mask
 
 using HostCPUFeatures:
   pick_vector_width,

src/codegen/split_loops.jl

@@ -76,7 +76,8 @@ function add_operation!(
   opnew
 end
 
-append_if_included!(vnew, vold, included) = for (i, v) ∈ vold
+append_if_included!(vnew, vold, included) =
+  for (i, v) ∈ vold
     id = included[i]
     iszero(id) || push!(vnew, (id, v))
   end

src/modeling/costs.jl

@@ -13,7 +13,8 @@ struct Instruction
 end
 # lower(instr::Instruction) = Expr(:(.), instr.mod, QuoteNode(instr.instr))
 # Base.convert(::Type{Expr}, instr::Instruction) = Expr(:(.), instr.mod, QuoteNode(instr.instr))
-callexpr(instr::Instruction) = if instr.mod === :LoopVectorization
+callexpr(instr::Instruction) =
+  if instr.mod === :LoopVectorization
     Expr(:call, lv(instr.instr))
   else#if instr.mod === :Main
     Expr(:call, instr.instr)
@@ -563,7 +564,8 @@ function reduction_to_single_vector(x::Float64)
     throw("Reduction not found.")
   end
 end
-reduce_to_onevecunroll(x::Float64) = if x == ADDITIVE_IN_REDUCTIONS
+reduce_to_onevecunroll(x::Float64) =
+  if x == ADDITIVE_IN_REDUCTIONS
     :+
   elseif x == MULTIPLICATIVE_IN_REDUCTIONS
     :*
@@ -578,7 +580,8 @@ reduce_to_onevecunroll(x::Float64) = if x == ADDITIVE_IN_REDUCTIONS
   else
     throw("Reduction not found.")
   end
-reduce_number_of_vectors(x::Float64) = if x == ADDITIVE_IN_REDUCTIONS
+reduce_number_of_vectors(x::Float64) =
+  if x == ADDITIVE_IN_REDUCTIONS
     :contract_add
   elseif x == MULTIPLICATIVE_IN_REDUCTIONS
     :contract_mul
@@ -593,7 +596,8 @@ reduce_number_of_vectors(x::Float64) = if x == ADDITIVE_IN_REDUCTIONS
   else
     throw("Reduction not found.")
   end
-reduction_to_scalar(x::Float64) = if x == ADDITIVE_IN_REDUCTIONS
+reduction_to_scalar(x::Float64) =
+  if x == ADDITIVE_IN_REDUCTIONS
     :vsum
   elseif x == MULTIPLICATIVE_IN_REDUCTIONS
     :vprod

src/predicates.jl

@@ -11,7 +11,7 @@ isscopedname(:(Base.Checked.checked_add), (:Base, :Checked), :checked_add)
 function isscopedname(ex, modpath, name::Symbol)
   isexpr(ex, :(.), 2) &&
     (a = ex.args[2]; isa(a, QuoteNode) && a.value === name) &&
-            hasscope(ex.args[1], modpath)
+    hasscope(ex.args[1], modpath)
 end
 hasscope(modex, mod::Symbol) = modex === mod
 hasscope(modex, mod::Tuple{Symbol}) = hasscope(modex, mod[1])

src/reconstruct_loopset.jl

@@ -27,7 +27,7 @@ Base.promote_rule(
   ::Type{UpperBoundedInteger{N,T}},
   ::Type{T}
 ) where {N,T<:Base.BitInteger} = T
-Base.convert(::Type{T}, i::UpperBoundedInteger) where {T<:Number} =
+Base.convert(::Type{T}, i::UpperBoundedInteger) where {T<:Integer} =
   convert(T, i.i)
 Base.convert(
   ::Type{UpperBoundedInteger{N,T}},

src/simdfunctionals/mapreduce.jl

@@ -115,7 +115,8 @@ end
 Vectorized version of `sum`. Providing a function as the first argument
 will apply the function to each element of `A` before summing.
 """
-@inline vsum(f::F, A::AbstractArray{T}) where {F,T<:NativeTypes} = vmapreduce(f, +, A)
+@inline vsum(f::F, A::AbstractArray{T}) where {F,T<:NativeTypes} =
+  vmapreduce(f, +, A)
 @inline vsum(A::AbstractArray{T}) where {T<:NativeTypes} = vsum(identity, A)
 
 length_one_axis(::Base.OneTo) = Base.OneTo(1)