this PR #10993 fixed a real bug but performance dropped by ~10x (i found out after trying to apply some tunings on code that was searched before the bug).

Fix

regain performance but retain the bug fix

@griwes @fbusato we approved the external PR despite the SASS bot ringing the bell. Let's ask authors to post benchmarks to justify their changes when they change our kernels?

(i got a fix and pushing it rn)