Loopsimd alignment

This MR adds a additional template parameter to LoopSIMD to specify the alignment.

The default alignment is the alignment of the underlying type.

I'm not sure how to determine the proper alignment for some return types or RebindType. To mitigate that problem I define conversion constructors and assignment operators from LoopSIMD types with different alignment.

This MR depends on !669 (merged).

Feel free to push at the source branch!

added 1 commit

• b25fd482 - set the alignment of the ReboudType

Compare with previous version

added 2 commits

• 7b8601eb - fix template arguments
• b69e9f37 - switch LoopSIMD of Mask to Mask of LoopSIMD

Compare with previous version

resolved all discussions

added 1 commit

• 303b6538 - Apply suggestion to dune/common/simd/loop.hh

Compare with previous version

added 1 commit

• 2f131bbe - add assertion that checks the allignement of a LoopSIMD object

Compare with previous version

added 1 commit

• d7f8f261 - only check whether LoopSIMD is at least as good aligned as it is guaranteed

Compare with previous version

• regarding the RebindType I suggest to take the max alignment of the old and the new type.
• regarding the implementation, please check with Jö, I have the impression that you started with an older master and now there are some inconsistencies regarding the bool return types.

regarding the RebindType I suggest to take the max alignment of the old and the new type.

You are right. I changed this. I recognized that previously even Rebind<double, LoopSIMD<float, 1>> was ill-formed. Since it would be resolved to LoopSIMD<double, 1, 4>, which is ill-formed since the natural alignemnt of doubleis 8.

regarding the implementation, please check with Jö, I have the impression that you started with an older master and now there are some inconsistencies regarding the bool return types.

I don't really know what you mean. I merged the current master. It went through without problems. Just waiting for the CI...

added 48 commits

• d7f8f261...dfa40b10 - 46 commits from branch core:master
• 4356f480 - Merge remote-tracking branch 'origin/master' into loopsimd_alignment
• bbe30918 - use the maximum of A and the alignemnt of the underlying type for a rebind type

Compare with previous version

Is there a reason that this MR is still WIP? Otherwise I'd merge.

assigned to @christi

Probably because Nils is still waiting for input from me after some offline discussion months ago. This input is unlikely to come, and anyway I think it was basically choosing between to incompatible options, and it was unclear which one was better. So go ahead as far as I'm concerned.

unmarked as a Work In Progress

changed the description

merged

mentioned in commit 3c5922d7

Please revert this. MR !669 (merged) already increase the amount of memory needed to build looptest.cc.o with g++-8 -c -O0 to about 9gb. This MR !670 (merged) is even worse: Now I can no longer build this target (and thus build_tests) with 16gb memory.

Also the time used by the CI increased from about 5min to about 1h. The fact that the pipeline failed with a timeout could have been an indicator for this.

Sorry, as the pipeline is failing for various reason recently, I didn't realize this issue.

mentioned in commit ebae92f0

mentioned in merge request !782 (merged)