This repository provides an improved implementation of the number theoretic transform (NTT) that leverages Tensor Cores on NVIDIA GPUs.
NTT is a generalization of the fast Fourier transform (FFT) over modular integers, and it is used to perform convolutions and polynomial/multiple-precision arithmetic exactly, in contrast to FFT, which uses floating-point arithmetic and thus suffers from rounding errors.
Through careful analysis, we successfully removed the redundant operations in the previous Tensor Core-based NTT implementations. Along with efficient modular reduction algorithm, we achieved higher performance than the previous work in most settings.
See the paper by Y. Sugizaki and D. Takahashi titled "Improved Implementation of Number Theoretic Transform on NVIDIA GPU with Tensor Cores" (doi:10.1145/3773656.3773673), which was accepted to SupercomputingAsia 2026 / International Conference on High Performance Computing in Asia-Pacific Region 2026 (SCA/HPCAsia 2026), for more details.
For license and copyright notices, see the SPDX file tags in each file. Unless otherwise noted, files in this project are licensed under the Apache License, Version 2.0 (SPDX short-form identifier: Apache-2.0) and copyrighted by the contributors.
Everyone is encouraged to contribute to this project. See the CONTRIBUTING.md file for instructions.