Alexander Viand

You virtually always want to flatten images before encryption, as FHE ciphertexts are effectively vectors and FHE offers SIMD operations over vectors. For advanced use cases, more complex patterns might...

Thanks for reporting this! I'll have a look and see if this can be fixed by depending on a more revent version of LLVM/MLIR, or if we do indeed need...

> > Safer-MLIR paper ... > > Could you link to this paper? I don't recall it. Maybe [SyFER-MLIR](https://math.mit.edu/research/highschool/primes/materials/2020/Govindarajan-Moses.pdf)? They do discuss a few things around NTT, iirc, but mostly...

I'm trying to get up to speed after RWC/Easter Holidays, and I saw that you mentioned compile times in the order of 50 minutes in #589, which sounds an order...

> > What I think I'd want is just the output MLIR from each of the evaluation artifacts, so that I can compare the rotation counts in them with what...

> * Scrutinize roberts_cross to ensure that HEIR's "better" program is actually correct. 👀 I think HECO generates the same program as porcupine here (which should be optimal, so I'd...

> @AlexanderViand-Intel: do you happen to have any input-output test pairs lying around from HECO? Just to save some time manually generating them. Sorry, everything from HECO would be SEAL...

> So remaining work is: > > * Improve box_blur I think this is the last open item in this issue, right? According to the `box_blur_64x64.mlir` test, HEIR still generates...

> From [#571 (comment)](https://github.com/google/heir/issues/571#issuecomment-2063841078), we agreed that it was 7 because the kernel is 3x3, whereas HECO did 2x2. 🤦 Sorry, I got confused by seeing the intermediate-value rotations, which...

👍 Isn't the classification just "is the element type of the tensor `secret.secret`"? If that's the case, it might not be worth adding a whole pass/analysis and instead bneing able...