CHANGELOG.md

SparseConnectivityTracer.jl

Version v1.2.2

• Add support for cholesky (#322)

Version v1.2.1

• Fix NaNMath.pow method ambiguity (#292)

Version v1.2.0

• Add support for fma (#289)
• Fix bugs in Hessian code generation (#290)
• Rephrase "Don't overload manually" (#285)
• Remove unused custom set types (#287)

Version v1.1.3

• Remove superflous overloads on identity and invalidations caused by them (#280)

Version v1.1.2

• Return Dual on float, widen, big (#279)

Version v1.1.1

• Fix type conversions of primal value in Dual (#271)

Version v1.1.0

• Add ChainRulesCore extension, support ignore_derivatives (#267)
• Fix det on matrix of Duals (#268)
• Minor updates to documentation (0ce0650, #266)

Version v1.0.2

• Avoid invalidations on >, >= by falling back to <, <= (#261) (#259)
• Test invalidations in CI (#264)
• Run CI tests on Julia 1.12, update JET (#246)

Version v1.0.1

• Fix invalidations on ismissing and isnothing (#259)

Version v1.0.0

This release is only breaking for users who touched unexported internals.

• Refactor internal types (#254)
• Select internal data structures for sparsity pattern representations via the TracerSparsityDetector and TracerLocalSparsityDetector keyword arguments gradient_pattern_type and hessian_pattern_type. Refer to the new doc page on "Performance" for more information. (#254, #256)
• Move DataInterpolations.jl extension out of SCT (#255)

Version v0.6.21

• Document limitations on stateful code (#249)

Version v0.6.20

• Revert PR #243 due to concerns of non-spare patterns (#248)

Version v0.6.19

• Stop tracing through multiplication by zero (#243)
• Update code style and formatter to Runic.jl

Version v0.6.18

• DataInterpolations v8 compatiblity (#234)

Version v0.6.17

• Performance optimization in pattern creation (#239)

Version v0.6.16

• Add more matrix division methods (#236)

Version v0.6.15

• Add stable API for tracer type via jacobian_eltype and hessian_eltype (#233)

Version v0.6.14

• Add stable API for allocation of buffers via jacobian_buffer and hessian_buffer (#232)

Version v0.6.13

• Return Dual on zero and friends (#231)

Version v0.6.12

• Fix for method ambiguities resulting from the 3-argument dot methods introduced in v0.6.11 (#228)

Version v0.6.11

• SCT has a new preprint! 🎉 Check it out on the arXiv: Sparser, Better, Faster, Stronger: Efficient Automatic Differentiation for Sparse Jacobians and Hessians (#225)
• Add overloads for 3-argument dot (#226)

Version v0.6.10

• Fix jacobian_sparsity output initialization for inplace functions (#223)

Version v0.6.9

• Relax type annotations in Jacobian output parsing (#217)
• Simplify DataInterpolations.jl extension (#210)

Version v0.6.8

• Support clamp and clamp! (#208)
• Remove internal set type DuplicateVector (#209)

Version v0.6.7

• Drop compatibility with Julia <1.10 to improve tracer performance (#204, #205)

Version v0.6.6

• Fix detector display by replacing println with print (#201)
• Improve code generation for 2-to-1 overloads on arbitrary types (#197, #202)
• Update package tests and CI workflow (#198, #199)

Version v0.6.5

• Fix LogExpFunctions.jl compat entry (#195)
• Fix "How it works" documentation (#193)

Version v0.6.4

• Shorter printing of default detectors (#190)
• Consistently refer to TracerSparsityDetector as detector (#191)
• Make imports explicit, test with ExplicitImports.jl (#188)

Version v0.6.3

• Add DataInterpolations.jl package extension (#178)
• Add LogExpFunctions.jl package extension (#184)
• Add NaNMath.jl package extension ([#187])
• Support two-argument atan and log (#185)
• Document limitations of operator overloading utils (#180)
• Reexport ADTypes interface (#182)
• Update developer documentation URLs (#186)
• Reorganize code and update code generation utilities (#179, #183)

Version v0.6.2

• Return only primal value when applying non-differentiable methods to Dual numbers (#169)
• Increase sparsity for Diagonal inputs (#165)
• Add more methods on round, rand (#162)
• This release brings large updates to the documentation:
  • Document limitations (#175)
  • Document global vs. local patterns (#176)
  • Add "How it works" developer documentation (#174)
  • Add developer documentation on custom overloads (#177)
• Refactor type conversions (#173, #168, #166)
• Make comparisons regular operators (#169)
• Fix Hessian on NNlib activation functions celu, elu, selu, hardswish (#162)
• Fix isless (#161)

Version v0.6.1

• Improve the performance of Hessian pattern tracing by an order of magnitude:
  • Internally represent Hessian patterns with dictionaries (#155, #158)
  • Increase performance via symmetry of Hessian (#151)
• Support ComponentArrays (#146)
• Support boolean not (!) (#150)
• Support isless (#149)

Version v0.6.0

• Remove ConnectivityTracer (#140)
• Remove legacy interface (#140)
  • instead of jacobian_pattern(f, x), use jacobian_sparsity(f, x, TracerSparsityDetector())
  • instead of hessian_pattern(f, x), use hessian_sparsity(f, x, TracerSparsityDetector())
  • instead of local_jacobian_pattern(f, x), use jacobian_sparsity(f, x, TracerLocalSparsityDetector())
  • instead of local_hessian_pattern(f, x), use hessian_sparsity(f, x, TracerLocalSparsityDetector())
• Remove overloads on similar to reduce amount of invalidations (#132)
• Fix sparse array construction (#142)
• Add array overloads (#131)
• Generalize sparsity pattern representations (#139, #119)
• Reduce allocations of new tracers (#128)
• Reduce compile times (#119)