LoopSIMD for vector types

This MR adapts LoopSIMD to take another vectorized type as template argument. If Tis a vectorized type, then LoopSIMD<T,k> has lanes<T>()*k lanes and the same scalar as T (Scalar<LoopSIMD<T,k>> == Scalar<T>).

This is useful with dune-vectorclass, to create arbitrary large vector types with explicit SIMD support, for example.

Feel free to push any changes to the source branch!

added 1 commit

• 0f0881c1 - use the mask type explicitly in any* and all* functions of LoopSIMD

Compare with previous version

added 1 commit

• c1ced7de - use the mask type explicitly in any* and all* functions of LoopSIMD

Compare with previous version

added 1 commit

• a47e522f - adapt `cond` method of LoopSIMD

Compare with previous version

resolved all discussions

added 1 commit

• 4863b4e7 - adapt `cond` method of LoopSIMD

Compare with previous version

added 1 commit

• ee4d0000 - fix return type of LoopSIMD::lane for rvalue references

Compare with previous version

mentioned in merge request !670 (merged)

unmarked as a Work In Progress

merged

mentioned in commit 32c2fdd3

mentioned in commit 3c5922d7

Please revert this. While building the tests in dune-common was not a problem before (despite the huge amount of tests just for the SIMD stuff), after merging this, building looptest.cc.o with g++-8 -c -O0 take about 9gb of memory.

@carsten.graeser what is taking that long? The updated SIMD tests, or also the other tests?

The 9gb are just for the looptest.cc.o target, nothing else. I did not check the execution time of the tests.

Thanks for pointing out this issue. @nils.dreier is not in office this week, but we'll try to fix this next weeks.

Thanks!

The reason it's taking so much memory is because Nils added a test for LoopSIMD<LoopSIMD<...> >, but did not add explicit instantiations for that.