Variadic MHLO reduce support about mlir-emitc HOT 6 CLOSED

I assume we are talking about multiple returns? This is indeed not possible, see https://github.com/iml130/mlir-emitc/blob/cbbeed65364e9e1b62737fec18e08587e3945bee/lib/Dialect/EmitC/Conversion/MHLORegionOpsToEmitC.cpp#L153%C3%9FL157 and reported in our MHLO HLO Op Coverage docs. Is this a requirement for the work on your side?

from mlir-emitc.

Indeed else for example argmax lowering to MHLO doesn't work which rules out a large class of models.

from mlir-emitc.

On the one hand, we can of course work on supporting multiple returns to mhlo.reduce (a reproducer that doesn't requires to be touched before running it would be super helpful!).
On the other hand, with lowering argmax, you're referring to lowering https://www.tensorflow.org/api_docs/python/tf/math/argmax to MHLO? In that case #278 would be an alternative to improve the mhlo.reduce support or not? Since our resources are limited, do you have a preference?

from mlir-emitc.

I've briefly looked into this. It should be fairly easy to support the two results case via an overload, but I don't see an easy way to generalize the reduction function to arbitrary arity.

from mlir-emitc.

Two result case would go a long way (I think it is probably second most common post plain single result). That would be very helpful.

For the arbitrary arity case, could it be expanded to a for loop constantly reapplying function? (for (...) { x = f(val[it], x) } kind). So you'd need a function to be able to get the slice of the tensor at iteration I kind.

from mlir-emitc.

The two result case is now supported with #285 landed.

from mlir-emitc.