9 releases (4 breaking)
new 0.5.0 | Oct 22, 2024 |
---|---|
0.4.1 | Oct 10, 2024 |
0.4.0 | Sep 30, 2024 |
0.4.0-alpha.0 | Jul 17, 2024 |
0.1.3 | Jan 30, 2024 |
#1 in #fully
407 downloads per month
Used in tfhe
1MB
15K
SLoC
TFHE Cuda backend
Introduction
The tfhe-cuda-backend
holds the code for GPU acceleration of Zama's variant of TFHE.
It implements CUDA/C++ functions to perform homomorphic operations on LWE ciphertexts.
It provides functions to allocate memory on the GPU, to copy data back and forth between the CPU and the GPU, to create and destroy Cuda streams, etc.:
cuda_create_stream
,cuda_destroy_stream
cuda_malloc
,cuda_check_valid_malloc
cuda_memcpy_async_to_cpu
,cuda_memcpy_async_to_gpu
cuda_get_number_of_gpus
cuda_synchronize_device
The cryptographic operations it provides are:- an amortized implementation of the TFHE programmable bootstrap:
cuda_bootstrap_amortized_lwe_ciphertext_vector_32
andcuda_bootstrap_amortized_lwe_ciphertext_vector_64
- a low latency implementation of the TFHE programmable bootstrap:
cuda_bootstrap_low latency_lwe_ciphertext_vector_32
andcuda_bootstrap_low_latency_lwe_ciphertext_vector_64
- the keyswitch:
cuda_keyswitch_lwe_ciphertext_vector_32
andcuda_keyswitch_lwe_ciphertext_vector_64
- the larger precision programmable bootstrap (wop PBS, which supports up to 16 bits of message while the classical PBS only supports up to 8 bits of message) and its sub-components:
cuda_wop_pbs_64
,cuda_extract_bits_64
,cuda_circuit_bootstrap_64
,cuda_cmux_tree_64
,cuda_blind_rotation_sample_extraction_64
- acceleration for leveled operations:
cuda_negate_lwe_ciphertext_vector_64
,cuda_add_lwe_ciphertext_vector_64
,cuda_add_lwe_ciphertext_vector_plaintext_vector_64
,cuda_mult_lwe_ciphertext_vector_cleartext_vector
.
Dependencies
Disclaimer: Compilation on Windows/Mac is not supported yet. Only Nvidia GPUs are supported.
- nvidia driver - for example, if you're running Ubuntu 20.04 check this page for installation
- nvcc >= 10.0
- gcc >= 8.0 - check this page for more details about nvcc/gcc compatible versions
- cmake >= 3.24
Build
The Cuda project held in tfhe-cuda-backend
can be compiled independently from TFHE-rs in the following way:
git clone git@github.com:zama-ai/tfhe-rs
cd backends/tfhe-cuda-backend/cuda
mkdir build
cd build
cmake ..
make
The compute capability is detected automatically (with the first GPU information) and set accordingly. If your machine does not have an available Nvidia GPU, the compilation will work if you have the nvcc compiler installed. The generated executable will target a 7.0 compute capability (sm_70).
Links
License
This software is distributed under the BSD-3-Clause-Clear license. If you have any questions,
please contact us at hello@zama.ai
.