faster faster!

speed.jl

using SIMD
using BenchmarkTools

foo(x, s, t) = x[ x .< s] .= t
foo2(x, s, t) = for i in eachindex(x) x[i] = x[i] < s ? t : x[i] end
foo3(x, s, t) = @inbounds for i in eachindex(x) x[i] = x[i] < s ? t : x[i] end
function foo3b(x, s, t) 
   @inbounds for i in eachindex(x)
      (x[i] < s) && (x[i] = t) 
   end
end
function foo4(x::Array{T}, s::T, t::T, ::Type{Vec{N,T}}) where {N, T}
    #@assert length(x) % N == 0
    sv = Vec{N,T}(s)
    tv = Vec{N,T}(t)
    @inbounds for i in 1:N:length(x)
        xv = vload(Vec{N,T}, x, i)
        mask = xv < sv
        xv = vifelse(mask, xv, tv)
        vstore(xv, x, i)
    end
end

function foo_t(x, s, t)
   nt = Threads.nthreads()
   N = length(x)
   #@assert N % nt == 0
   chunk = div(N, nt)
   Threads.@threads for t=1:nt
      chunk_range = ((t-1)*chunk + 1):(t*chunk)
      @inbounds for i=chunk_range
         x[i] = x[i] < s ? t : x[i]
      end
   end
end

T = Float64
x = rand(T, 200*200)
s = T(0.5)
t = T(0.11)
#@btime foo(x, s, t);
@btime foo2(x, s, t);
@btime foo3(x, s, t);
@btime foo3b(x, s, t);
@btime foo4(x, s, t, Vec{8,T});

@btime foo_t(x, s, t);