Funny results on huge data sets

[DBJDBJ]

Code is available online https://godbolt.org/z/cYnaoxM58 . Admittedly with much smaller matrix sizes. My code has design issues and is needlessly complex, but correct and works ok for smaller matrixes.

If one takes that offline and declares large matrix sizes, funny things will happen:

It seems that is an overflow issue in UBENCH. One quick-and-dirty approach might be to report negative timings as zeroes? longjump out of the ubench_main? Will look into it.