Permute N-D array : why is it faster with real than complex ?

Hi!

I'm encoding a matlab function which calculate the contraction of two tensor. It's quite good but I have a problem with permute : it's faster with real than complex. For example :

>> A=rand(250,10,500,120);

>> B=complex(rand(250,10,500,120),rand(250,10,500,120));

>> sigma=[3 4 1 2];

>> tic; A=permute(A,sigma); toc

Elapsed time is 0.945218 seconds.

>> tic; B=permute(B,sigma); toc

Elapsed time is 3.444950 seconds.

Do you know why is there such a gap ? Can you help me reduce it ?

Thanks!

PS : I'm sorry, my English is not good....

Answer 1

[Matt J]

Do you know why is there such a gap ?

Because there is twice as much data to shuffle around when you have both real and imaginary parts.

Can you help me reduce it ?

Assuming you can tolerate lower precision, you could use type single instead of type double. On my machine,

>> tic; permute(B,sigma); toc
Elapsed time is 0.872804 seconds.

>> B=single(B);

>> tic; permute(B,sigma); toc
Elapsed time is 0.582679 seconds.

Answer 2

[Richard]

Thanks for your answer. It's better but not enough for my problem. I've to try another way, without permutation.