I have two lists of fixed size vectorizable matrices, that I want to frequently copy back and forth between each other (so I can do calculations on a trial modification, and then either accept [copy back] or reject [recopy original] the trial).

I'm looking at how best to do the copying quickly. My first thought was:

Code: Select all

Matrix<double,16,16> mainSet[40], trialSet[40]
...
memcpy(&trialSet, &mainSet, 40*sizeof(Matrix<double,16,16>))


So my questions are:

1. Can I do this? (Or is there additional data in Matrix<double,16,16> that will not tolerate this kind of bitwise copy?)

2. Should I do this? (Or is there no guarantee that it will continue to work with future versions of Eigen?)

3. Is there a faster way? In particular can I force the compiler to use vectorization since I know the data are all aligned?

With respect to question 3, I'm assuming that while it would make use of vectorization the very naive:

Code: Select all

for (int i = 0; i < 40; ++i)
    trialSet[i] = mainSet[i];


would still be much slower than a memcpy? But if people think maybe not I can try and profile it...

yes for 1 and 2. Regarding 3, usually memcpy makes use of vectorization, so I also doubt the for loop will be faster. Both solutions should be equivalent, but the only way to be sure is to benchmark.