pyTFHE GPU Acceleration Library

Kernels and NTT are based on the cuFHE library (https://github.com/vernamlab/cuFHE).

Requirements

• CUDA Toolkit (Tested on 13.0)
• pybind11
• python3-dev

Installation

First, clone and cd into this repo.

Then, run the following commands:

pip3 install .

Enjoy.

Usage

import pyTFHEGPU
# Initialize cuda device. Must be done before creating any TFHE objects.
pyTFHEGPU.init_cuda()

# These are the gate_types
NAND = 0
OR = 1
AND = 2
NOR = 3
XOR = 4
XNOR = 5
ANDNY = 6  # (NOT input1) AND input2
ANDYN = 7  # input1 AND (NOT input2)


t = pyTFHEGPU.BTGM()
# Create a new empty batch
b1 = pyTFHEGPU.Batch()

# ========= add each ciphertext =================
# the 2nd argument is the ctxt_unique_id
#     ctxt_unique_id can be non-continuous but must be unique
# the 3rd argument is the message (0 or 1)
t.add_ctxt(b1, 0, 0)
t.add_ctxt(b1, 1, 1)
t.add_ctxt(b1, 2, 0)

# ========= add each gate =================
# - the first argument is the gate_unique_id
#     gate_unique_id can be non-continuous but must be unique
# - the second argument is the gate_type
# - the third argument is the input1_unique_id
# - the fourth argument is the input2_unique_id
# - the fifth argument is the output_unique_id
#
# For example, this means Ctxt[2] = Ctxt[0] NAND Ctxt[1]
t.add_gate(b1, 0, NAND, 0, 1, 2)
# Feel free to add more gates

# Don't forget to build dependency graph
# - verbose = False
t.build_dependency_graph(b1, False)

# Evaluate the graph
t.eval(b1, False)

# Now let's see the results
# The first argument is the ctxt_unique_id
print(t.get_ctxt_value(b1, 2))

Check the test directory for more examples.

Benchmark

RTX A5000 reports 4122.82 gates/sec when there's no dependency between the gates.