Le lundi 26 janvier 2015 à 03:07 -0800, Yuuki Soho a écrit :
> Coming back to my original problem, I did a simplified version of it,
>
> which is about 10x slower than a vectorized matlab version. Have I
> missed anything here ?
Your code looks fine to me, at least it doesn't seem to suffer from the
classic performance traps. Could you show the MATLAB code you're using?
It may well be calling a highly-tuned BLAS behind the scenes.
BTW, zero(Float64) is stricly equivalent to 0. or 0.0. zero() is only
useful to write generic code when you don't know in advance the type of
the variables your function will operate on.
Regards
> A = ones(50,40,40);
>
> B = ones(50,100)/2;
>
> C = ones(40,100)/3;
>
> D = ones(40,100)/4;
>
> E = ones(100,100)/5;
>
> idx = int([100:-1:1]);
>
>
>
> function
> testSum(A::Array{Float64,3},B::Array{Float64,2},C::Array{Float64,2},
>
>
> D::Array{Float64,2},E::Array{Float64,2},idx::Array{Int64,1})
>
>
>
> alpha = zeros(100)
>
> tmp_1 = zero(Float64)
>
> tmp_2 = zero(Float64)
>
> tmp_3 = zero(Float64)
>
>
>
> @inbounds for t = 1:100
>
> for thp = 1:50
>
>
>
> tmp_3 = zero(tmp_3)
>
> for x_3 = 1:40
>
>
>
> tmp_2 = zero(tmp_2)
>
> for x_2 = 1:40
>
>
>
> tmp_1 = zero(tmp_1)
>
> @simd for x_1 = 1:50
>
> tmp_1 += A[x_1,x_2,x_3] * B[x_1,t]
>
>
> end
>
> tmp_2 += tmp_1 * C[x_2,t]
>
> end
>
> tmp_3 += tmp_2 * D[x_3,t]
>
> end
>
>
>
> alpha[t] = E[t,idx[t]] * tmp_3
>
> end
>
> end
>
>
>
> return alpha
>
>
>
> end
>
>
>
>
