Skip to content

Clad version 2.1

Latest
Latest

Choose a tag to compare

View all tags
@vgvassilev vgvassilev released this 01 Oct 09:56
· 49 commits to master since this release
v2.1
7596cbc

Introduction

This document contains the release notes for the automatic differentiation plugin for clang Clad, release 2.1. Clad is built on top of
Clang and LLVM compiler infrastructure. Here we describe the status of Clad in some detail, including major improvements from the previous release and new feature work.

Note that if you are reading this file from a git checkout, this document applies to the next release, not the current one.

What's New in Clad 2.1?

Clad 2.1 introduces major advancements in reverse mode differentiation, bringing smarter handling of loops, assignments, and method calls, alongside the new clad::restore_tracker for functions that modify their inputs. Forward mode gains static scheduling for Hessians and higher-order derivatives, while CUDA support expands with custom derivatives for key Thrust algorithms such as reduce, transform, and transform_reduce, plus optimizations that reduce unnecessary GPU atomics. The release also strengthens error estimation, simplifies adjoint initialization, improves tape efficiency, and enhances diagnostics. With a migration to C++17, support extended up to clang-21, and numerous bug fixes, Clad 2.1 delivers faster, safer, and more reliable automatic differentiation across CPU and GPU workflows.

Some of the major new features and improvements to Clad are listed here. Generic improvements to Clad as a whole or to its underlying infrastructure are described first.

External Dependencies

  • Clad now works with clang-11 to clang-21.
  • Switch to C++17 standard.

Forward Mode & Reverse Mode

  • Improve support for differentiation of function/method calls:
    • Move call/argument differentiation into unified helpers.
    • Improve handling of base initializers, delegating constructors, and method bases.
  • Optimizations in Tape-Based Recording (TBR):
    • Enable TBR for pointer arithmetic, nested calls, local variables, loops, and constructors.
    • Improve analysis of pointers, nested derivatives, and division denominators.
    • Remove redundant or unsafe getZeroInit initializations in aggregates and init lists.
  • Statically schedule higher-order derivatives for Hessians.

Reverse Mode

  • Better handling of pullbacks and reverse_forw:
    • Avoid generating empty or redundant pullbacks.
    • Improve type consistency in reverse_forw.
    • Introduce clad::restore_tracker to support functions modifying their arguments.
  • Improve differentiation ordering (differentiate method bases before arguments).
  • Simplify assignments when LHS and RHS are independent.
  • Support for for-loops without increments.
  • Add optimizations to avoid unnecessary storage of unused return values.
  • Consider unions differentiable.
  • Support for not differentiating w.r.t. const references.
  • Numerous fixes in reverse_forw handling of types and aggregates.

CUDA

  • Add custom derivatives for several Thrust algorithms:
    • thrust::reduce, thrust::transform, thrust::transform_reduce,
      thrust::copy, thrust::inner_product.
  • Improve handling of CUDA atomics:
    • Avoid atomics for injective index computations.
    • Added liveness analysis for removing unnecessary atomics.

Error Estimation

  • Reverse mode error estimation now uses pullbacks consistently.
  • Fix _final_error value propagation .

Misc

  • Improve diagnostics for propagator signature mismatches.

  • Full qualification logic for generated code.

  • Better handling of consteval functions, decl refs, and unused parameters.

  • Remove unused STL derivatives and redundant infrastructure.

  • CI improvements:

    • Added valgrind test runs.
    • Updated MacOS runners, godbolt build (clang 21), and codecov handling.

  • Documentation: Update README to clarify plugin usage, C++14 requirement, and
    plugin args.

Fixed Bugs

428 691 752 768 1274 1301 1346 1349 1411 1419 1436 1445 1457 1466 1469 1472 1473 1490 1497 1498 1555 1557 1559 1560 1565 1573 1581 1582

Special Kudos

This release wouldn't have happened without the efforts of our contributors,
listed in the form of Firstname Lastname (#contributions):

FirstName LastName (#commits)

A B (N)

PetroZarytskyi (60; major ad/tbr work)
Vassil Vassilev (30; infrastructure, ci, compiler integration)
Abdelrhman Elrawy (7; cuda/thrust support)
Max Andriychuk (7; va/tbr infrastructure, cfg/loop analysis, cuda atomics)
aditimjoshi (4; tape improvements, benchmarks)
Vipul Cariappa (1; godbolt integration)
Timo Nicolai (1; docs, usage improvements)
Rohan Timmaraju (1; tensor-specific support)
mcbarton (1; ci, runners)
Jonas Rembser (1; build fixes)
Christina Koutsou (1; nondiff structs support)

What's Changed

New Contributors

Full Changelog: v2.0...v2.1

Contributors

vgvassilev, Time0o, and 8 other contributors
Assets 2
Loading