v2.2.0

What's New

Faster recording and adjoint sweeps

• Recording passes the multiplier through sums, differences and negations, so higher-order types build fewer nested products
• std::complex<AReal> arithmetic records fewer tape statements
• The adjoint sweep visits each statement's operands last to first, and chunk crossings stay out of the hot path in recording and in the sweep
• Faster forward mode: packed masking of zero tangents, one reciprocal per vector division, larger FReal operands held by reference
• Higher-order tapes over FReal write their partials straight into the tape

New and improved tape options

• XAD_COPY_ELISION (new, off by default): copies share the slot of their source and record nothing; see the Build Options page for the supported use
• XAD_REDUCED_MEMORY, XAD_SKIP_ZERO_PARTIALS, XAD_TAPE_TRACK_LIVE_SLOTS, XAD_OUTLINE_RECORDING, XAD_TAPE_REUSE_SLOTS, XAD_LINEAR_SLOTS, XAD_HUGE_PAGES, XAD_TLS_INITIAL_EXEC and XAD_SHARED_LIBS are new or improved since 2.0.0
• New Build Options page: what each performance-related option does, when it helps, and starting sets per kind of workload

Direct mode, samples and fixes

• Direct mode types (ARealDirect, FRealDirect) are drop-in replacements for AReal/FReal
• New samples: SIR calibration, Nelson-Siegel curve fit, projectile aiming, Eigen decompositions
• Many fixes in tape chunk handling, checkpointing, resetTo, JIT recording, Eigen and libc++ compatibility

Upgrade notes

• Expression types defined outside XAD (deriving from xad::Expression) need calc_derivatives(info, tape) and calc_derivatives(info, tape, detail::minus_unit) overloads
• Forward-mode expressions over larger FReal types hold their operands by reference: don't keep one past its full expression
• std::numeric_limits<T>::is_iec559 is false for active types

See the CHANGELOG for the full list.

If XAD helps you, a GitHub star helps others find it