IPPL dot operation seems slower than an explicit dot implementation

[s-mayani]

During the FEM framework implementation, we have noticed some slowdown in our FEMPoissonSolver code in the call to the IPPL dot product expression. More concretely, in this line.

When we replace this line by an explicit implementation of the dot product (so a sum of the components multiplied), the code becomes much faster while retaining correct results (on CPU). However, we have also realised that doing this affects our the code running on GPU (it produces NaNs in the results).

This issue is opened to document this observation. In order to move forward in investigating the performance of the IPPL dot, one would need to write a test case with an isolated call to the IPPL dot product and compare its timings to a manual dot product, without the side effects of the FEM Poisson solver test case.