A Tensor Compiler with Automatic Data Packing for Simple and Efficient Fully Homomorphic(…)
summary
Fhelipe is a tensor-to-FHE compiler that implements automatic data packing and a layout algebra (bit-permutation + gap bits) to pack tensors into CKKS ciphertexts, performs cheap compaction to merge partially filled ciphertexts, and includes a search that matches manual bootstrapping placement. It compiles large neural-network and linear-algebra workloads in seconds, runs on CPU and a state-of-the-art accelerator, reduces code size 10–48%, and outperforms a prior FHE compiler by gmean 18.5× while matching hand-tuned performance.
AI-written summary · about
chapters (16)
- 0:00Introduction to fully homomorphic encryption
- 1:03FHE adoption challenges
- 2:06Philippe compiler for FHE
- 3:51How FHE works
- 4:54FHE data layout challenges
- 7:45Philippe programming model
- 9:11Philippe compilation process
- 9:52Flexible layout in Philippe
- 11:38Large tensors with compaction
- 13:23Automatic layout in Philippe
- 14:25Philippe performance
- 14:46Philippe benefits summary
- 15:28Q&A memory requirements
- 16:32Q&A noise and bootstrapping
- 17:58Q&A new accelerator backends
- 19:00Q&A layout backtracking
related talks
28mFunctional Sparse Tensor Compilation
24mVerified Tensor-Program Optimization Via High-Level Scheduling Rewrites
11mQiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a...
25mFormat Abstraction for Sparse Tensor Algebra Compilers
20mA Mechanized Algebra of Verified Data Structures for Optimizing Sparse Tensor Programs
38mPrivacy in the Open Source: Homomorphic Encryption
19mCompiler Support for Sparse Tensor Convolutions
20mHEaaN.MLIR: An Optimizing Compiler for Fast Ring-Based Homomorphic Encryption