carstenbauer
commented
3 years ago Since I know that you're interested in these things, maybe a bit of background. A C++ friend gave me this code (.txt because GitHub doesn't allow .cpp):
and told me that his Julia code was much slower (>5x) (again .txt because .jl isn't allowed):
I went ahead and tried to reproduce the timings and (using icc -O3 -I ~/.local/lib/eigen/ Ising2D.cpp -o Ising2D) got this entirely different picture:
➜ ~/Desktop/ising julia 2dising.jl
-0.4428173828125
-0.4118935546875
-0.436130859375
-0.460140625
-0.4394482421875
-0.4422109375
-0.4049609375
3.088 s (14 allocations: 822.06 KiB)
➜ ~/Desktop/ising time ./Ising2D
-0.423299
./Ising2D 9.22s user 0.01s system 99% cpu 9.234 totalI asked him what he's doing differently and he told me that he is using icpx instead of icc or clang. For him, this makes the a interestingly HUGE difference:
cecri@cecri-Mint:~/work/Practices/2dIsing$ clang++-9 -O3 -march=native Ising2D.cpp
cecri@cecri-Mint:~/work/Practices/2dIsing$ time ./a.out
-0.434271
real 0m3.337s
user 0m3.330s
sys 0m0.004s
cecri@cecri-Mint:~/work/Practices/2dIsing$ icpx -Ofast -xHost Ising2D.cpp
cecri@cecri-Mint:~/work/Practices/2dIsing$ time ./a.out
-0.406395
real 0m0.667s
user 0m0.663s
sys 0m0.004sSo, while Julia is competitive to clang/icc, perhaps even faster, icpx crushes everyone else! My thought then was, alright I know a package that is known for automagical performance speedups, which is why I tried LoopVectorization.jl after optimising the code a bit (40% speedup on my machine):
using Random
using BenchmarkTools
const rng = MersenneTwister()
function get(spin_conf, x, y)
(Nx, Ny) = size(spin_conf)
@inbounds spin_conf[(x-1+Nx)%Nx + 1, (y-1+Ny)%Ny + 1]
end
function deltaE(spin_conf, i, j)
return 2.0*get(spin_conf,i,j) * (get(spin_conf,i-1,j) + get(spin_conf,i+1,j) +
get(spin_conf,i,j+1) + get(spin_conf,i,j-1))
end
function energy(spin_conf)
(Nx, Ny) = size(spin_conf)
res = 0
for i = 1:Nx
@simd for j = 1:Ny
res += -get(spin_conf,i,j)*(get(spin_conf,i+1,j) + get(spin_conf,i,j+1))
end
end
return res
end
magnetism(spin_conf) = sum(spin_conf)
function sweep!(rec, spin_conf, beta, M)
(Nx, Ny) = size(spin_conf)
@inbounds @simd for i = 1:M
x = rand(rng, 1:Nx)
y = rand(rng, 1:Ny)
r = rand(rng)
dE = deltaE(spin_conf, x, y)
if exp(-beta*dE) > r
spin_conf[x,y] *= -1
end
rec[i] = energy(spin_conf)
end
return rec
end
function main()
beta = 0.2
L = 64
total_iteration = 100_000
spin_conf = rand(rng, (-1,1), L, L)
rec = Vector{Int64}(undef,total_iteration)
sweep!(rec, spin_conf, beta, total_iteration)
s = sum(last(rec, 1000))/1000
s /= L^2
println(s)
return nothing
end
@btime main()This is the background for this issue and any help in getting the magical performance of icpx in Julia would be appreciated :)
(Update: g++ can also give the "magic" when run with the -fwhole-program compiler flag)
chriselrod
DNF2
MWE:
Trying to run
energy(spin_conf)I get"Manually inlining" the
getfunction:I obtain