Open SoniEx2 opened 4 months ago
Isn't more detailed testing and more coverage always better?
it would certainly be able to catch hardware bugs that only manifest with certain lanes.
but it also takes 4x longer?
Is speed a constraint for running tests? For operations that don't map to a single hardware instruction, or for engines that scalarize, testing all lanes could still be helpful in catching implementation bugs.
A lot of the SIMD tests run the same floats through all of the lanes, instead of just assuming lanewise operations to be implemented in a lanewise way. Is there any particular reason to do it this way? You could probably cut down the size of the tests by a factor of 2 or 4 if you just assumed all the lanes behave the same.