Checkpointing.jl Checkpointing Library for Timestepping Codes

@Michel_Schanen and I have created a package that provides checkpointing schemes for adjoint computation using automatic differentiation/autodiff (AD) of time stepping loops. The schemes are agnostic to the ADTool being used and can be easily interfaced with any Julia AD tool.

We welcome contributions and suggestions for improvement. The repository is at:

Currently the package provides the following checkpointing schemes:

  1. Revolve/Binomial checkpointing [1]
  2. Periodic checkpointing


add Checkpointing

Interface with an AD Tool

Currently, Checkpointing.jl interfaces with an AD tool through the computation of a Jacobian by implementing a jacobian method. The following describes the interface for ReverseDiff.jl

using Checkpointing

struct ReverseDiffADTool <: AbstractADTool end

function Checkpointing.jacobian(tobedifferentiated, F_H, ::ReverseDiffADTool)
    return ReverseDiff.jacobian(tobedifferentiated, F_H)

The interfaces for Diffractor.jl, Enzyme.jl, ForwardDiff.jl, ReverseDiff.jl, and Zygote.jl are implemented in examples/adtools.jl.


The macro @checkpointing covers the transformation of for loops with 1:steps ranges where steps is the number of timesteps:

@checkpoint scheme adtool for i in 1:steps
    F_H = [F[1], F[2]]
    F = advance(F_H,t,h)
    t += h

where adtool is one of the interface AD tools (e.g. ReverseDiffADTool()) and scheme is a adjoint checkpointing scheme like for example revolve.

function store(x::Vector, c::Vector, t::Int64, s::Int64)
    c[1,s] = x[1]
    c[2,s] = x[2]
    c[3,s] = t

function restore(c, i)
    x = [c[1,i], c[2,i]]
    t = c[3,i]
    return x, t
Revolve(steps, checkpoints, store, store; verbose=verbose)

Revolve needs the functions store and restore for storing and restoring the i-th checkpoint c[i] with variables x and t. steps is the total number of timesteps while checkpoints is the number of available checkpoints.


The following features are planned for development:

  • Integration with ChainRules.jl to generate rules for timestepping loops
  • Online checkpointing schemes for adaptive timestepping
  • Composition of checkpointing schemes
  • Multi-level checkpointing schemes

[1] Andreas Griewank and Andrea Walther. 2000. Algorithm 799: revolve: an implementation of checkpointing for the reverse or adjoint mode of computational differentiation. ACM Trans. Math. Softw. 26, 1 (March 2000), 19–45. DOI: