Enzyme AD

Enzyme Automatic Differentiation Framework

Enzyme Overview

The Enzyme project is a tool for performing reverse-mode automatic differentiation (AD) of statically-analyzable LLVM IR. This allows developers to use Enzyme to automatically create gradients of their source code without much additional work.

double foo(double);

double grad_foo(double x) {
    return __enzyme_autodiff(foo, x);

By differentiating code after optimization, Enzyme is able to create substantially faster derivatives than existing tools that differentiate programs before optimization.


Enzyme is composed of four pieces:

  • An optional preprocessing phase which performs minor transformations that tend to be helpful for AD.
  • A new interprocedural type analysis that deduces the underlying types of memory locations
  • An activity analaysis that determines what instructions or values can impact the derivative computation (common in existing AD systems).
  • An optimization pass which creates any required derivative functions, replacing calls to __enzyme_autodiff with the generated functions.

More resources

For more information on Enzyme, please see:

Citing Enzyme

To cite Enzyme, please cite the following:

title = {Instead of Rewriting Foreign Code for Machine Learning, Automatically Synthesize Fast Gradients},
author = {Moses, William S. and Churavy, Valentin},
booktitle = {Advances in Neural Information Processing Systems 33},
year = {2020},
note = {To appear in},

The original Enzyme is also avaiable as a preprint on arXiv .