In thread vectorization-for- ... 34043.html the question arose if a vectorization of 3-dimensional vectors makes sense by virtually enhancing the vector size to 4 elements. I wonder if it is possible/already planned to do something similar for all fixed-sized matrices/vectors which are not of size (k*4*double)?

The reason for asking this follows: I made a small benchmark regarding the possible benefits of vectorization using a small energy function.

Given a 9x9 matrix "D" and 9-elemental vector F I computed
vector9 stress=D*F;
scalar energy = 0.5*stress*F;
in one flow.

Since the matrix D is not arbitrary in my case I was able to reduce the problem. By defining a transformation matrix T of size 9x6 I could transform the vector F and temporary products. First I precomputed the reduced 6x6 matrix D_reduced from D.

Then again:

Vector6 F_reduced = linear function of F;
Vector6 stress_reduced = D_reduced*F_reduced;
Vector9 stress = T * stress_reduced;
scalar energy = 0.5*stress_reduced*F_reduced;

As you can see, there are far more operations involved. But still, it is considerable faster (factor >2). This is quite contradictious and gives me a bad feeling in my stomach when I need to chose a complex operation because it is faster compared with the simple one, won't you agree?

There's been a thread on the mailing list recently where this was discussed, and Gael had some ideas,

http://listengine.tuxfamily.org/lists.t ... 00041.html

perhaps in a future release...