SIMD.jl

SIMDjl_performance.jl

julia> function vector_dot(B, C)
           a = zero(eltype(B))
           for i in eachindex(B,C)
               @inbounds a += B[i] * C[i]
           end
           return a
       end
vector_dot (generic function with 1 method)

julia> function vector_dot_macro(B, C)
           a = zero(eltype(B))
           @simd for i in eachindex(B,C)
               @inbounds a += B[i] * C[i]
           end
           return a
       end
vector_dot_macro (generic function with 1 method)

julia> using SIMD

julia> function vector_dot_SIMD(B::Vector{T}, C::Vector{T}) where T
           N = 8 # SIMD width
           @assert length(B) == length(C)
           @assert length(B) % N == 0
           a = zero(T)
           @inbounds for i in 1:N:length(B)
               Bvec = vload(Vec{N,T}, B, i)
               Cvec = vload(Vec{N,T}, C, i)
               a += sum(Bvec + Cvec)
           end
           return a
       end
vector_dot_SIMD (generic function with 1 method)

julia> using BenchmarkTools

julia> B = rand(51200); C = rand(51200);

julia> @btime vector_dot($B,$C);
  43.630 μs (0 allocations: 0 bytes)

julia> @btime vector_dot_macro($B,$C);
  8.930 μs (0 allocations: 0 bytes)

julia> @btime vector_dot_SIMD($B,$C);
  8.907 μs (0 allocations: 0 bytes)