ffcc 0.1.3

Compiler for fast floating-point approximations

Fast Math Program Synthesis Tool:

Tool for generating fast approximations of floating point math by (ab)using the floating point representation.

This tool was produced for the ICML paper "Faster Activation Functions at the Edge for Post-Training Speedups" link.

Usage:

Basic usage (ffcc tool):

$ ffcc -e "silu(x) = x / (1 + exp(-x))" --approx=exp --tune=[-6,6] -o torch

This should give you, after some tuning time, the following torch module:

import torch
from torch import nn, Tensor


class FastSilu(nn.Module):
        def forward(self, x: Tensor) -> Tensor:
                v0 = (1064873152.0 + (-12104086.0 * x))
                v1 = v0.type(torch.int32).view(torch.float32)
                return (x / (1.0 + v1))

Flags explained:

• -e $expr provides the input expression to approximate
• -approx=exp approximates exponentiation (log and div can be added as well, though div support is experimental)
• -tune=[-6,6] performs gradient-descent based constant tuning on the domain $[-6,6]$
• -o torch prints the resulting code as a pytorch module

Development Environment:

There's a flake.nix file for all nixos users.

The python dependencies are managed through uv. Setting everything up usually involves running a combination of uv venv; uv sync --all-extras; source .venv/bin/activate.

To run tests, use lit tests/filecheck, there are no pytests yet.

There is an ffcc-opt tool available for testing.

License:

ffcc - fast float compiler - Copyright (C) 2025 Anton Lydike

This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.

You should have received a copy of the GNU Affero General Public License along with this program. If not, see https://www.gnu.org/licenses/.