Comparing Python, Julia, and C++

tkoolen · October 31, 2018, 6:15pm

Not sure about the original question, but for the Julia code you’re not actually initializing the input arrays:

ziotom78/python-julia-c-/blob/d6a5a1faa3350498b321c9719293d25752112a1d/julia-speed.jl#L13


      
          using Statistics: minimum
          using Printf
          
          
f(x1, x2, x3, x4, x5, x6) = @. x1 + x2
          g(x1, x2, x3, x4, x5, x6) = @. x1 + x2 - x3
          h(x1, x2, x3, x4, x5, x6) = @. x1 + x2 - x3 + x4
          i(x1, x2, x3, x4, x5, x6) = @. x1 + x2 - x3 + x4 - x5
          j(x1, x2, x3, x4, x5, x6) = @. x1 + x2 - x3 + x4 - x5 + x6
          
          
N = 1000000
          x = [Array{Float64}(undef, N) for i in 1:6]
          
          
function print_result(num, b)
              best = minimum(b)
              # The member "time" is in ns
              @printf("%d\t%8.2f\t%.2f\n", num, best.time / 1e6, best.memory / 1024^2)
          end
          
          
println("Terms\tSpeed [ms]\tMemory [MB]")
          print_result(2, @benchmark f(x...))
          print_result(3, @benchmark g(x...))

This may result in subnormal numbers in your test inputs, which can adversely affect performance. See e.g. 50x speed difference in gemv for different values in vector - #3 by StefanKarpinski.