search

google / heir

A compiler for homomorphic encryption
https://heir.dev/
Apache License 2.0
306 stars 46 forks source link

bgv: add lowering to poly #109

Open mr0re1 opened 1 year ago

mr0re1 commented 1 year ago

I'm having difficulty with lowering ModulusSwitch, I was trying to lower directly to arith, but faced avalanche of complications (casts, signedness, overflows). Also it feels idiomatically incorrect to lower straight to arith, bypassing poly.

@asraa , @j2kun , Would it be reasonable to have a Poly_ScaleOp that takes poly argument and float (either argument or attribute?) and returns element-wise multiplied poly (rounded down)?

j2kun commented 1 year ago

What do you think about waiting until we have a unified LWE/RLWE dialect? I agree that lowering to arith is weird, but if you want to go through poly now, you could lower to a combination of to_tensor and from_tensor, using tensor/math/arith ops for the rescaling and rounding.

mr0re1 commented 1 year ago

you could lower to a combination of to_tensor and from_tensor, using tensor/math/arith ops for the rescaling and rounding.

That what I'm currently doing, so output is still in poly. My concern was about "going level too deep" during lowering.

What do you think about waiting until we have a unified LWE/RLWE dialect?

I think it will pose the same problem. I will proceed with crude lowering approach and send out PR to gather concrete feedback.