Le jeudi 12 mars 2015 à 11:01 -0500, Tim Holy a écrit :
> This is something that many people (understandably) have a hard time
> appreciating, so I think this post should be framed and put up on the julia
> wall.
>
> We go to considerable lengths to try to make code work efficiently in the
> general case (check out subarray.jl and subarray2.jl in master some time...),
> but sometimes there's no competing with a hand-rolled version for a
> particular
> case. Folks should not be shy to implement such tricks in their own code.
Though with the new array views in 0.4, the vectorized version should be
more efficient than in 0.3. I've tried it, and indeed it looks like
unrolling is not really needed, though it's still faster and uses less
RAM:
X = rand(100_000, 5)
function f1(X, i, j)
for _ in 1:1000
X[:, i], X[:, j] = X[:, j], X[:, i]
end
end
function f2(X, i, j)
for _ in 1:1000
a = sub(X, :, i)
b = sub(X, :, j)
a[:], b[:] = b, a
end
end
function f3(X, i, j)
for _ in 1:1000
@inbounds for k in 1:size(X, 1)
X[k, i], X[k, j] = X[k, j], X[k, i]
end
end
end
julia> f1(X, 1, 5); f2(X, 1, 5); f3(X, 1, 5);
julia> @time f1(X, 1, 5)
elapsed time: 1.027090951 seconds (1526 MB allocated, 3.63% gc time in
69 pauses with 0 full sweep)
julia> @time f2(X, 1, 5)
elapsed time: 0.172375013 seconds (390 kB allocated)
julia> @time f3(X, 1, 5)
elapsed time: 0.155069259 seconds (80 bytes allocated)
Regards
> --Tim
>
> On Thursday, March 12, 2015 07:49:49 AM Steven G. Johnson wrote:
> > As a general rule, with Julia one needs to unlearn the instinct (from
> > Matlab or Python) that "efficiency == clever use of library functions",
> > which turns all optimization questions into "is there a built-in function
> > for X" (and if the answer is "no" you are out of luck). Loops are fast,
> > and you can easily beat general-purpose library functions with your own
> > special-purpose code.
