Kunwar Grover
Kunwar Grover
> I'm quite concerned about the DecomposeExpReduction implementation, as it seems weird to call the greedy Rewriter in the Pass, then receive it as a Builder in AggregatedOpInterface::decomposeOp, only to...
@efric I added a ci-extra trailer to run test_torch can you check?
> Something is off with this runner -- the gpu doesn't show up it seems It's with all OSSCI runners I think.
I think this needs perf numbers. This can lead to weird shuffles if done on a non innermost dimensions and we need to make sure we aren't regressing.
This is less related to flattening, more related to the fact that our broadcast lowering is "bad". https://github.com/iree-org/iree/issues/21978 is bad because of the same reason. The correct way to lower...
@krzysz00 Can you rebase this? I added ci-extra: test_torch to it. If it improves perfs let's land this, this is effectively doing SLP Vectorization which is okay to do here.