Documentation: https://bdgiffin.github.io/remat/

REMAT is a proof-of-concept reversible physics library predicated on the use of fixed-precision numerics and integer arithmetic operations to ensure that all computations are exactly bit-reversible. Consequently, prior solution states may be precisely rematerialized through direct reversal of the forward-in-time operations. This is useful for the purpose of computing adjoint sensitivities of a time-dependent simulation without the need for solution checkpointing.

A central concept underpinning the developed bit-reversible computations pertains to the use of "dual" numbers which preserve information (bits) that would otherwise be lost due to finite precision rounding errors. In particular, consider the following example involving integer addition:

$$x=7 \qquad y=x+5=12$$

This operation may be exactly reversed through integer subtraction to recover the addend:

$$x=y-5=7$$

However, the operation of integer division cannot be exactly reversed through multiplication, i.e.

$$x=7 \qquad y=x\div5=1$$

and

$$x\neq y\times5=5$$

To ensure that the integer division operation can be exactly reversed, we must introduce an auxiliary variable to store the remainder:

$$x=7 \qquad y=x\div5=1 \qquad r=x \, \text{mod} \, 5=2$$

and thus the division may be reversed through the following correction involving the remainder:

$$x=y\times5+r=7$$

Ordinarily, the remainders of all such multiplication operations would need to be stored until the operation is later reversed, but this may entail intractable memory requirements if many such operations are carried out. To reduce memory overheads, a similar approach to that of Maclaurin (2015) is adopted, whereby the ancillary variables $x^{*}$ and $y^{*}$ are introduced such that:

$$x=7 \qquad x^{*}=n$$

$$y=x\div5=1 \qquad y^{*}=x^{*}\times5+x \, \text{mod} \, 5=n\times5+2$$

and

$$x=y\times5+y^{*} \, \text{mod} \, 5=7 \qquad x^{*}=y^{*}\div5=n$$

This concept is applied such that any persistent stable variables $x$ are endowed with corresponding ancillary state variables $x^{*}$ subject to an extended set of rules for all basic arithmetic operations which ensure that the pairing of $x$ and $x^{*}$ efficiently preserves all bits that would otherwise be erased due to round-off.

Dependencies

The core functionality of REMAT is primarily written in C++. A Python API wrapper module (implemented using ctypes) is provided to facilitate the use of REMAT in Python projects. Additional pre- and post-processing utilities are provided in Python, and demonstrated through several accompanying examples. While the underlying C++ framework does not have any dependencies apart from the standard library, the following Python packages are used by the API wrapper module and pre/post-processing utilities: - numpy - scipy - pygmsh - ezdxf - pyexodus

REMAT utilizes a CMake-based build framework in combination with the funtionality provided by BLT. CMake version 3.14+ is required to build/install REMAT from source, whereas BLT is included as a git submodule.

The accompanying examples additionally make use of the following Python packages to enable interactive visualization: - pygame - pygame_widgets

If you intend to export recorded animations using the functionality provided in the Animation.py module, you will need to install the command line utility ffmpeg.

Many of the examples can be built for execution in a web browser using Emscripten to compile the underlying C++ library into WebAssembly, and pygbag to package the pygame-based visualizations.

Getting started

To build REMAT, you will first need to clone this repository and build the underlying C++ shared object library files from source. If you are cloning a new repository, you will need to obtain all required submodules (BLT) via: git submodule update --init --recursive

To configure, build, and locally install REMAT: mkdir build cd build cmake .. make install This will create a new local install subdirectory within the root REMAT project directory, containing all packaged Python modules and compiled C++ shared object libraries required to import and run the Python-based examples.

If compiling the REMAT library to WebAssembly, be sure to configure CMake to use the Emscripten toolchain using emcmake to wrap the call to cmake: emcmake cmake .. Otherwise, the build/installation process described above remains the same.

To import the locally installed REMAT package within your Python project: import sys sys.path.append("PATH/TO/REMAT/install/package/") import REMAT The examples subdirectory provides further illustrative cases of invocations of the library within the context of a Python workflow.

When packaging a pygame project that uses REMAT using pygbag for execution in a web browser, make sure that you have compiled the REMAT library using Emscripten (see above note on configuring CMake using emcmake), and include the install/package files directly within the local directory for your project. The provided examples offer a demonstration of how this can be accomplished, with the appropriate invocations of pygbag included in the examples/Makefile.