vcount (#471)

* Add vcount

* Test vcount

* Test range of sizes

* Test updates

* Returns not defined in Julia 1.6

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.151"
+version = "0.12.152"
 
 [weakdeps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
@@ -57,5 +57,5 @@ Static = "0.8.4"
 StaticArrayInterface = "1"
 ThreadingUtilities = "0.5"
 UnPack = "1"
-VectorizationBase = "0.21.53"
+VectorizationBase = "0.21.60"
 julia = "1.6"

src/LoopVectorization.jl

@@ -196,7 +196,8 @@ export LowDimArray,
   vfilter,
   vfilter!,
   vmapreduce,
-  vreduce
+  vreduce,
+  vcount
 
 const VECTORWIDTHSYMBOL, ELTYPESYMBOL, MASKSYMBOL =
   Symbol("##Wvecwidth##"), Symbol("##Tloopeltype##"), Symbol("##mask##")
@@ -234,6 +235,7 @@ include("reconstruct_loopset.jl")
 include("constructors.jl")
 include("user_api_conveniences.jl")
 include("simdfunctionals/mapreduce.jl")
+include("simdfunctionals/count.jl")
 include("broadcast.jl")
 
 """

src/simdfunctionals/count.jl

@@ -0,0 +1,41 @@
+_vcount(f) = 0
+function _vcount(f::F, args::Vararg{DenseArray,M}) where {F,M}
+  x = first(args)
+  y = Base.tail(args)
+  foreach(a -> @assert(size(a) == size(x)), y)
+  N = length(x)
+  ptrargs = map(VectorizationBase.zstridedpointer, args)
+  i = 0
+  V = VectorizationBase.pick_vector_width(
+    reduce(promote_type, map(eltype, ptrargs))
+  )
+  W = unwrap(V)
+  UNROLL = 4
+  LOG2UNROLL = 2
+  counts = if VERSION >= v"1.7"
+    VecUnroll(ntuple(Returns(0), Val(UNROLL)))
+  else
+    VecUnroll(ntuple(_ -> (0), Val(UNROLL)))
+  end
+  while i < vsub_nsw(N, ((W << LOG2UNROLL) - 1))
+    index = VectorizationBase.Unroll{1,W,UNROLL,1,W,zero(UInt)}((i,))
+    counts += count_ones(f(VectorizationBase.fmap(vload, ptrargs, index)...))
+    i = vadd_nw(i, StaticInt{UNROLL}() * W)
+  end
+  count = reduce_tup(+, data(counts))
+  while i < vsub_nsw(N, (W - 1)) # stops at 16 when
+    count += count_ones(f(map1(vload, ptrargs, (MM{W}(i),))...))
+    i = vadd_nw(i, W)
+  end
+  if i < N
+    m = mask(StaticInt(W), N & (W - 1))
+    vfinal = f(map1(vload, ptrargs, (MM{W}(i),), m)...)
+    count += count_ones(vfinal & m)
+  end
+  count
+end
+
+@generated function vcount(f::F, args::Vararg{DenseArray,M}) where {F,M}
+  call = Expr(:call, :_vcount, :f)
+  gc_preserve_call_quote(call, M::Int)
+end

src/simdfunctionals/map.jl

@@ -74,8 +74,11 @@ function vmap_singlethread!(
   ::Val{NonTemporal},
   args::Vararg{AbstractArray,A}
 ) where {F,T<:NativeTypes,A,NonTemporal}
-  ptry, ptrargs, N = setup_vmap!(f, y, Val{NonTemporal}(), args...)
-  _vmap_singlethread!(f, ptry, Zero(), N, Val{NonTemporal}(), ptrargs)
+  presy = preserve_buffer(y)
+  GC.@preserve presy begin
+    ptry, ptrargs, N = setup_vmap!(f, y, Val{NonTemporal}(), args...)
+    _vmap_singlethread!(f, ptry, Zero(), N, Val{NonTemporal}(), ptrargs)
+  end
   nothing
 end
 function _vmap_singlethread!(
@@ -263,20 +266,25 @@ function vmap_multithread!(
   end
   nothing
 end
-function gc_preserve_vmap_quote(NonTemporal::Bool, Threaded::Bool, A::Int)
-  m = Threaded ? :vmap_multithread! : :vmap_singlethread!
-  call = Expr(:call, m, :f, :y, Expr(:call, Expr(:curly, :Val, NonTemporal)))
+function gc_preserve_call_quote(call, A::Int)
   q = Expr(:block, Expr(:meta, :inline))
   gcpres = Expr(:gc_preserve, call)
-  for a ∈ 1:Int(A)::Int
+  for a ∈ 1:A
     arg = Symbol(:arg_, a)
     parg = Symbol(:parg_, a)
-    push!(q.args, Expr(:(=), arg, :(@inbounds args[$a])))#Expr(:ref, :args, a)))
+    push!(q.args, Expr(:(=), arg, :($getfield(args, $a))))
     push!(q.args, Expr(:(=), parg, Expr(:call, :preserve_buffer, arg)))
     push!(call.args, arg)
     push!(gcpres.args, parg)
   end
-  push!(q.args, gcpres, :y)
+  push!(q.args, gcpres)
+  q
+end
+function gc_preserve_vmap_quote(NonTemporal::Bool, Threaded::Bool, A::Int)
+  m = Threaded ? :vmap_multithread! : :vmap_singlethread!
+  call = Expr(:call, m, :f, :y, Expr(:call, Expr(:curly, :Val, NonTemporal)))
+  q = gc_preserve_call_quote(call, A)
+  push!(q.args, :y)
   q
 end
 @generated function gc_preserve_vmap!(

test/grouptests.jl

@@ -15,7 +15,7 @@ const START_TIME = time()
   @time if LOOPVECTORIZATION_TEST == "all" || LOOPVECTORIZATION_TEST == "part2"
     if VERSION <= v"1.8" || isempty(VERSION.prerelease)
       using Aqua
-      @time Aqua.test_all(LoopVectorization, ambiguities = false)
+      @time Aqua.test_all(LoopVectorization, ambiguities = false, piracy = false)
     end
     @test isempty(detect_unbound_args(LoopVectorization))
 

test/map.jl

@@ -37,4 +37,10 @@
   @test vmap(abs2, 1:100) == map(abs2, 1:100)
   @test vmapt(abs2, 1:3:10000) == map(abs2, 1:3:10000)
   @test vmapt(abs2, 1.0:3.0:10000.0) ≈ map(abs2, 1.0:3.0:10000.0)
+
+  for n = -64:64
+    let x = rand(UInt8, (1 << 14) + n)
+      @test count(==(UInt8('\n')), x) == vcount(==(UInt8('\n')), x)
+    end
+  end
 end