Play with nvidia dev suggestion for unroll+jam

[gbaraldi]

This seems like an interesting suggestion llvm/llvm-project#97029

[gbaraldi]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[oscardssmith]

looks mostly bad.

[Zentrik]

Yh and some of the big gains, seem to have been benchmarks that ran unusually slow on master so they might not even be real.

E.g. the below two benchmarks were quite slow on the master commit used as comparison (note the last entry in the plots is very slow)

Also, [["sparse", "transpose", ("transpose", "(20000, 20000)")]] is very noisy. Plot of benchmarks from master below

Plots from https://tealquaternion.camdvr.org/?end=57492&nonRelevant=true