Nils (Jan 05 2024 at 11:28): Nils (Jan 05 2024 at 11:28):Is there a way to get something as nice as
x[x .<= maximum(x)/4] .= 0.0
with the same performance as
cutoff = maximum(x)/4
for i in eachindex(x)
if x[i] <= cutoff
x[i] = 0.0
end
end
?
I see about a 2x performance hit from the first and 3 allocations (which surprised me, I was expecting one from the comparison)
Sukera (Jan 05 2024 at 11:56):I don't think the dots fuse across indexing like that
Sukera (Jan 05 2024 at 11:56):so might be resolved with a @views
Sukera (Jan 05 2024 at 11:56):the result of the comparison will still be allocated though
Mason Protter (Jan 05 2024 at 12:33):I'm actually kinda surprised that julia doesn't support a generator as an index
Mason Protter (Jan 05 2024 at 12:34):then you could write this as something like
x[(i for i in eachindex(x) if x[i] <= maximum(x)/4)] .= 0.0
I think the both cleanest(?) and most performant way to do this would be
x .= ifelse.(x .<= maximum(x)/4, 0.0, x)
julia> f1(x) = x[x .<= maximum(x)/4] .= 0.0;
julia> f2(x) = let cutoff=maximum(x)/4
for i in eachindex(x)
if x[i] <= cutoff
x[i] = 0.0
end
end
x
end;
julia> f3(x) = x .= ifelse.(x .<= maximum(x)/4, 0.0, x);
julia> for N ∈ (100, 1000, 10000)
@show N
x = rand(N)
print("f1 "); @btime f1(x) setup=(x = rand($N)) evals=1
print("f2 "); @btime f2(x) setup=(x = rand($N)) evals=1
print("f3 "); @btime f3(x) setup=(x = rand($N)) evals=1
end
N = 100
f1 290.000 ns (3 allocations: 256 bytes)
f2 130.000 ns (0 allocations: 0 bytes)
f3 110.000 ns (0 allocations: 0 bytes)
N = 1000
f1 1.012 μs (4 allocations: 6.17 KiB)
f2 671.000 ns (0 allocations: 0 bytes)
f3 401.000 ns (0 allocations: 0 bytes)
N = 10000
f1 6.723 μs (5 allocations: 23.53 KiB)
f2 4.799 μs (0 allocations: 0 bytes)
f3 2.174 μs (0 allocations: 0 bytes)
Oh wow, how is that ifelse thing faster?
Nils (Jan 05 2024 at 13:37):Fwiw I don't see a speedup, so might be CPU dependent - in any case it's at least as fast and shorter and cleaner, so wins all round, thank you!
Mason Protter (Jan 05 2024 at 13:46):yeah, the relative performance of ifelse versus actual branching can be very CPU dependant
Mason Protter (Jan 05 2024 at 13:49):ifelse tends to be easier to make work with SIMD, but the branch-predictor can make branching more efficient
Mason Protter (Jan 05 2024 at 13:49):could also be julia-version dependant, since it'll depend a lot on various low-level optimizations LLVM may or may not make
chriselrod (Jan 05 2024 at 15:08):Mason Protter said:
could also be julia-version dependant, since it'll depend a lot on various low-level optimizations LLVM may or may not make
Yeah, for me f2 and f3 perform about the same, both of them SIMDing thanks to LLVM.
AMJ (Jan 05 2024 at 16:14):On my machine, if I am understanding correctly(which is a big if), f3 creates 2 main branches which both are heavily SIMDed, while f2 has one SIMDed branch with fewer instructions. So probably a better vectorization plus a better branch prediction pattern?
Last updated: Dec 28 2024 at 04:38 UTC