QuantLibAAD: QuantLib with XAD Automatic Differentiation in C++

Part of the XAD ecosystem - the fastest automatic differentiation library for C++. If this is useful to you, please star the core XAD repository.

QuantLibAAD builds QuantLib with XAD, so a pricing can compute all its risks (sensitivities) in one adjoint pass. QuantLib's library code is unchanged; your code registers the inputs, records the pricing and runs the adjoint, as the samples show. That is 4× to 29× faster than bumping on the QuantLib samples below, and up to 270× faster on Monte Carlo with XAD's commercial code generation add-on. It shows XAD in production-quality use on a code base of several hundred thousand lines.

This repository contains integration headers, examples, and tests required for this integration. It is not usable stand-alone.

Performance

Risk with XAD costs little more than pricing alone, and far less than bumping.

Cost of AD types

About 10% overhead on QuantLib's examples and test suite: QuantLib built with XAD's active type, running without computing derivatives, against the same build with double:

	AD / double
QuantLib test suite (1,714 cases)	1.13×
QuantLib benchmark	1.21×
QuantLib examples (20)	1.10×

Sensitivities: AD vs bump-and-revalue

All sensitivities in one adjoint pass, against one-sided bumping, one revaluation per input:

Sample	Inputs	XAD	Bumping
AdjointEuropeanEquityOption	98	29× faster · 0.24 s	6.90 s
AdjointMulticurveBootstrapping	69	27× faster · 1.85 s	50.5 s
AdjointSwap	55	26× faster · 0.45 s	11.5 s
AdjointReplication	13	4.4× faster · 1.33 s	5.81 s

A full set of sensitivities costs 2–3.5 valuations.

Monte Carlo: tape, split mode and JIT

benchmark_aad, 100,000 paths:

	Swaption (45 inputs)	CVA (90 inputs)
Bumping	24.5 s	48.9 s
XAD tape	6.8× faster · 3.62 s	13× faster · 3.63 s
XAD split mode	11× faster · 2.18 s	22× faster · 2.21 s
JIT with xad-codegen ¹	158× faster · 0.15 s	270× faster · 0.18 s

¹ xad-codegen is XAD's native code generation backend, available under a commercial license.

Setup

AMD EPYC (Zen 5), one core, Linux, GCC 15, XAD 2.2.0, release build; median of 3 runs.

Using QuantLibAAD in production under commercial terms, with support, the XAD validation pack or code generation: see XAD subscriptions or contact us.

Getting Started

For detailed build instructions with XAD and QuantLib, please refer to the XAD documentation site.

Build options

The defaults are tuned for speed: fast AD code, and fast builds of QuantLib and its tests. They need QuantLibAAD listed before XAD in QL_EXTERNAL_SUBDIRECTORIES, e.g. -DQL_EXTERNAL_SUBDIRECTORIES="<path>/QuantLibAAD;<path>/xad"; with XAD first, the configure stops with an error: fix the order and reconfigure in a fresh build dir, or pass -DQLAAD_ALLOW_XAD_FIRST=ON to keep XAD's own settings (CMake then warns which defaults it misses).

Any of them can be switched off on the QuantLib command line: XAD's tape options with -DXAD_<option>=OFF (XAD_USE_STRONG_INLINE, XAD_LINEAR_SLOTS, XAD_TLS_INITIAL_EXEC, XAD_HUGE_PAGES), a shared QuantLib with -DBUILD_SHARED_LIBS=ON, and -DQLAAD_<option>=OFF for QLAAD_LTO, QLAAD_FP_CONTRACT_OFF, QLAAD_FAST_BUILD (or only QLAAD_PRECOMPILE_HEADERS, with QLAAD_UNITY_BATCH_SIZE=0 for no unity build), QLAAD_USE_CCACHE and QLAAD_CTEST_PER_SUITE. Options that don't apply to a platform, compiler or Debug build (with single-config generators) are left off there; link-time optimisation is on by default with GCC, except on Windows.

• Set XAD's options with -D on the command line: other cache entries for them (from an old-order configure, ccmake, cmake-gui or -C) are replaced by the defaults (for XAD_NO_THREADLOCAL, XAD's own default, OFF), with a warning naming each entry that changes.
• Linear slots take more tape memory, and stay off with XAD_COPY_ELISION, XAD_TAPE_REUSE_SLOTS or XAD_TAPE_TRACK_LIVE_SLOTS.
• Bindings and other modules loaded with dlopen need -DXAD_TLS_INITIAL_EXEC=OFF (it is off by default for a shared QuantLib).
• When QuantLib is built inside another project, that project's settings (BUILD_SHARED_LIBS, XAD's options) decide; with an installed XAD package, XAD's own settings apply.
• A static QuantLib needs -lxad too for users that link through pkg-config or quantlib-config (or build with -DBUILD_SHARED_LIBS=ON); remove an older shared QuantLib from the install prefix.
• One CTest test per suite needs CMake 3.28 or newer; older versions keep QuantLib's single test.
• An installed LTO build links only with the same toolchain: installing for other toolchains needs -DQLAAD_LTO_FAT_OBJECTS=ON (GCC; it covers XAD's and QuantLib's libraries, not xad-codegen's), or -DQLAAD_LTO=OFF.

Code Generation

XAD is optimized for computing sensitivities efficiently in a single evaluation pass using adjoint mode. For workflows that require repeated evaluation across many scenarios — such as Monte Carlo simulations, XVA calculations, regulatory stress testing, or scenario-based risk analysis — XAD also supports recording computations into a JITGraph that can be compiled and re-evaluated efficiently using a JIT backend.

xad-codegen compiles a recorded graph to native code and evaluates it vectorised across paths. It is a commercial add-on, available with XAD Enterprise subscriptions: see XAD editions.

Getting Help

If you have found an issue, want to report a bug, or have a feature request, please raise a GitHub issue.

For general questions about QuantLibAAD, sharing ideas, engaging with community members, etc, please use GitHub Discussions.

Contributing

Please read CONTRIBUTING for the process of contributing to this project. Please also obey our Code of Conduct in all communication.

Help wanted: more QuantLib tests ported to AAD, and more examples.

Related Projects

• XAD Comprehensive automatic differentiation in Python and C++
• QuantLibAAD: Fast risk evaluations in Python and C++

License

Due to the nature of this repository, two different licenses have to be used for different part of the code-base. The tests and examples folders are containing code taken and modified from QuantLib where the QuantLib license applies. The ql folder contains adaptor modules for XAD, where the GNU AGPL applies. This is clearly indicated by having separate license files in each folder.

Authors

• Various contributors from Xcelerit
• See also the list of contributors who participated in the project.