Generic Matrix kernel specializations

[AlexRTer]

This PR adds implementations using the generic Matrix type for most kernels. They only use methods that are sure to be supported by Matrix.h based classes, i.e. get, set, prepare-/finish-/append.

Hence, new classes based on Matrix.h do not necessarily have to implement specializations for all kernels in order to make use of them. The implementations do not assume any additional constraints (e.g. memory layout) which limits possibilities for optimization in some cases. Therefore, they do not replace specialized kernels entirely but save a great amount of time by providing basic utility "out of the box". Note that this PR only adds the necessary specialized kernels and does not make them accessible to the runtime yet.

Changes:

• Matrix.h specializations have been added to almost all local kernels with exception of e.g. Neural Network related kernels.

• Existing tests have been extended to cover these as well. Currently, the kernels always default to a DenseMatrix if res was given as a nullptr, so tests also use a DenseMatrix through GenGivenVals.

  • Since empty matrices and views cannot be obtained this way, some tests use DTEmpty or DTView to generate these before they are cast to the appropriate type.
  • e.g. using DTEmpty = DenseMatrix<typename DTArg::VT>;
static_cast<DTArg *>(DataObjectFactory::create<DTEmpty>(0, 0, false));
  • Some kernels like Filter* also expect sel to be a DenseMatrix, in which cases DTSel is sometimes used.
  • Some tests have a different DTRes type than DTArg, so DT might be set using std::conditional

• DenseMatrix.h has also been changed to handle empty matrices if prepare-/finish- Append is called on them. I suspect this might also be an issue with e.g. CSRMatrix but I have not yet confirmed this.

To do: The cast (l. 327) in Order.h should be fine, though restrictive, as VTRes = VTArg is guaranteed in that case but please confirm this. The problem lies in extractRow expecting res and arg to have the same value type, which is true if returnIdx = false, but the compiler seems to also check if a cast is legal for the case of returnIdx = true in which value types are not guaranteed to be the same - so a static cast does not work.

[pdamme]

Thanks also for cleaning up the existing kernels and their test cases a bit (e.g., consistent file structure, corrections of misformattings and misnamings, compaction of multiple DataObjectFactory::destroy() into a single one, adding missing destroys() etc.). Very much appreciated!

[AlexRTer]

Thank you for looking at the cast in the Order.h kernel. You can go ahead and merge the PR.