David Schneider

Documenting here for future references: With #1847 and #1849, the two-level initialization works nice and relatively fast for moderate core counts. However, if we go beyond let's say 100 cores,...

Discussing in person. Working on this requires a lot of effort, as [Eigen uses internally signed indices](https://eigen.tuxfamily.org/index.php?title=FAQ#Why_Eigen.27s_API_is_using_signed_integers_for_sizes.2C_indices.2C_etc..3F). If we find out later, that having this is very relevant, we can...

> Open question: should we build these in the CI pipeline as well? Experience states that it will become outdated otherwise.

> values as a matrix of data vectors per vertex msize x ddim "values as a matrix, where each column represents a component for all vertices", would make most sense...

> You are right, I got the dimensions wrong… I don't think so. Values would be stored as a matrix with msize rows and ddata columns, hence `MatrixXd(msize, ddata)` (corresponding...

I think the summation action would be another case, where such a scenario would occur while being valid.

> Adding this functionality requires new API functions for reading data data = read(xB, yB, zB) and writing data write(data, xB, yB, zB). The signature of the `readData` function is...

> A further con is that overloading functions makes bindings more difficult to write and maintain. Thanks for pointing this out, I agree on this point. > Without having the...

> The mapping configuration has become monster. It definitely needs some clean-up at some point. Its massive, yes, but it's still clean, at least when I looked at it the...

> I am struggling with the data flow of JIT. As mentioned above, the data flow is the same as for direct access, i.e., we sample on the remote mesh...