Added vfilter to resolve #42.

Author: chriselrod

Manifest.toml

@@ -49,9 +49,9 @@ uuid = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 
 [[SIMDPirates]]
 deps = ["VectorizationBase"]
-git-tree-sha1 = "d652a1a47fd0806ca3f9109a0d729c3f7701e897"
+git-tree-sha1 = "a27b812034efdb062ec3d9e787bc299510057b3d"
 uuid = "21efa798-c60a-11e8-04d3-e1a92915a26a"
-version = "0.3.4"
+version = "0.3.5"
 
 [[SLEEFPirates]]
 deps = ["Libdl", "SIMDPirates", "VectorizationBase"]

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.6.6"
+version = "0.6.7"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
@@ -12,7 +12,7 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
 Parameters = "0"
-SIMDPirates = "0.3.4, 0.4, 0.5"
+SIMDPirates = "0.3.5, 0.4, 0.5"
 SLEEFPirates = "0.3.2, 0.4, 0.5"
 VectorizationBase = "0.2.5, 0.3, 0.4"
 julia = "1.1"

src/LoopVectorization.jl

@@ -11,13 +11,16 @@ using SIMDPirates: VECTOR_SYMBOLS, evadd, evmul, vrange, reduced_add, reduced_pr
 using Base.Broadcast: Broadcasted, DefaultArrayStyle
 using LinearAlgebra: Adjoint, Transpose
 
+const SUPPORTED_TYPES = Union{Float16,Float32,Float64,Integer}
 
 export LowDimArray, stridedpointer, vectorizable,
     @avx, @_avx, *ˡ, _avx_!,
-    vmap, vmap!
+    vmap, vmap!,
+    vfilter, vfilter!
 
 
 include("map.jl")
+include("filter.jl")
 include("costs.jl")
 include("operations.jl")
 include("graphs.jl")

src/broadcast.jl

@@ -188,7 +188,7 @@ end
 # function vmaterialize!(
 @generated function vmaterialize!(
     dest::AbstractArray{T,N}, bc::BC, ::Val{Mod}
-) where {T <: Union{Float32,Float64}, N, BC <: Broadcasted, Mod}
+) where {T <: SUPPORTED_TYPES, N, BC <: Broadcasted, Mod}
     # we have an N dimensional loop.
     # need to construct the LoopSet
     loopsyms = [gensym(:n) for n ∈ 1:N]
@@ -212,7 +212,7 @@ end
 end
 @generated function vmaterialize!(
     dest′::Union{Adjoint{T,A},Transpose{T,A}}, bc::BC, ::Val{Mod}
-) where {T <: Union{Float32,Float64}, N, A <: AbstractArray{T,N}, BC <: Broadcasted, Mod}
+) where {T <: SUPPORTED_TYPES, N, A <: AbstractArray{T,N}, BC <: Broadcasted, Mod}
     # we have an N dimensional loop.
     # need to construct the LoopSet
     loopsyms = [gensym(:n) for n ∈ 1:N]

src/filter.jl

@@ -0,0 +1,29 @@
+
+function vfilter!(f::F, x::Vector{T}, y::AbstractArray{T}) where {F,T <: SUPPORTED_TYPES}
+    W, Wshift = VectorizationBase.pick_vector_width_shift(T)
+    N = length(y)
+    Nrep = N >>> Wshift
+    Nrem = N & (W - 1)
+    i = 0
+    j = 0
+    GC.@preserve x y begin
+        ptr_x = pointer(x)
+        ptr_y = pointer(y)
+        for _ ∈ 1:Nrep
+            vy = vload(Vec{W,T}, gep(ptr_y, i))
+            mask = f(SVec(vy))
+            SIMDPirates.compressstore!(gep(ptr_x, j), vy, mask)
+            i += W
+            j += count_ones(mask)
+        end
+        rem_mask = VectorizationBase.mask(T, Nrem)
+        vy = vload(Vec{W,T}, gep(ptr_y, i), rem_mask)
+        mask = rem_mask & f(SVec(vy))
+        SIMDPirates.compressstore!(gep(ptr_x, j), vy, mask)
+        j += count_ones(mask)
+        resize!(x, j)
+    end
+    x
+end
+vfilter(f, y::AbstractArray{T}) where {T<:SUPPORTED_TYPES} = vfilter!(f, Vector{T}(undef, length(y)), y)
+