Hello everyone,

I am getting bad performance when computing

Code: Select all

B = A.transpose() * A

. When A is a 903x13 Matrix it takes around 13 ms.
After checking in the forum I found this thread, where this issue is discussed. There someone proposed a better way of doing this operation by using

Code: Select all

B.selfadjointView<Upper>().rankUpdate(A);

Is this correct? Because, this code does not produce the same matrix as A.transpose() * A. I am a bit lost right now...
If this is not the correct way, then what is the best way for computing A^T * A?

Thank you and best regards,
G. Ros.

Ok, I have realized that the correct expression should be:

Code: Select all

B.selfadjointView<Upper>().rankUpdate(A.transpose());

However, it seems to be that the transpose of A is making the operation very slow. When A is used instead of A.transpose() the time goes down to 1.5 ms.

Is there a way of doing this in an efficient fashion?

Best regards,
G. Ros

My fault. Problem solved.

What was the error?

Also, regarding your first post, 13ms for this computation means 23GFLOPS for a full product. Of course, using the selfadjointView allows to reduce the number of computations, but for such small outputs I don't expect much gain from this trick.