Edward Chen

Here's a brief [walkthrough of viaduct-he from a code perspective](https://www.notion.so/Code-A-Compiler-from-Array-Programs-to-Vectorized-Homomorphic-Encryption-3830208c5bbc46a19c3b84040352aa40?pvs=4).

> The basic approach of "pick a batching approach, then generate the circuit to get a cost estimate" does seem somewhat limiting if one were to do this at a...

Here are some notes on the different packing abstractions from Fhelipe and Viaduct. At a high level, both abstractions are interchangeable and can support all three use-cases listed (packing multiple...

> This would also allow us to use the full hypercube view of packing/rotations (which I'm not sure any of the "custom" approaches currently seem to support?) and would of...

> Though I'm not sure how we could represent the "replication" style things that, e.g., TileTensor notation has. Replication within a ciphertext can be represented by using _empty dimensions_ in...

> In the meantime, I wanted to ask if you could elaborate on the (dis)advantages of the TileTensor notation, too? I added the TileTensors notation for Fhelipe's LogReg program to...

> The diagonal packing can be directly expressed as an affine_map today. Can this be expressed in Fhelipe? Out of the box, we can't express diagonalization in Fhelipe. However, if...

Does HEIR support lowering affine_map expressions, e.g. diagonalization, into and FHE circuit? If so, how would it reason about the required data movement and rotations?

One thing to consider when using the affine_map style representation is how to derive the output packing from a given (edit:tensor) operation. Here was one example where I got stuck...

Hi everyone! I've been working on a representation and layout assignment protocol that utilizes diagonal packing. Specifically, my representation builds on top of the layout abstraction defined in [Viaduct](https://arxiv.org/pdf/2311.06142), and...