Enzyme.jl

Enzyme v0.10.12

Diff since v0.10.11

Closed issues:

• Segmentation fault when puting a complex number to power 0 (#381)
• Segmentation fault on a linear regression model (arm & x86 chips) (#500)
• Assertion failed: (G[v].count(u) == 0), function minCut (#510)

Merged pull requests:

• Fix large dynamic tape issue (#509) (@wsmoses)
• Add inactive ijl ptls states (#513) (@wsmoses)
• Better type depth reached warning (#515) (@wsmoses)
• Add box float (#516) (@wsmoses)
• Bump JLL (#517) (@vchuravy)

Enzyme v0.10.11

Diff since v0.10.10

Closed issues:

• Box{AbstractFloat}() calling convention (#368)
• Segmentation fault (#488)

Merged pull requests:

• Fix tape size mismatch for Any (#502) (@vchuravy)
• Add EnzymeCore to docs (#503) (@vchuravy)
• Disable needs_dynamic_size_workaround on nightly (#504) (@vchuravy)

Enzyme v0.10.10

Diff since v0.10.9

Merged pull requests:

• Add EnzymeCore (#498) (@vchuravy)

Enzyme v0.10.9

Diff since v0.10.8

Closed issues:

• Captured x breaks autodiff (#205)
• Implement jl_call convention for 1.9 (#492)
• dot product on Vector{Float32} fails (#495)

Merged pull requests:

• Add test for #205 (#419) (@vchuravy)
• adjust to JuliaLang/julia#46965 (#497) (@aviatesk)
• Update JLL and add test for issue (#499) (@vchuravy)

Enzyme v0.10.8

Diff since v0.10.7

Closed issues:

• Gradients for struct fields are zero when a Bool field is present (#431)

Merged pull requests:

• Add julia.call for structv/inactive (#493) (@wsmoses)
• Workaround Julia GC bug (#494) (@wsmoses)

Enzyme v0.10.7

Diff since v0.10.6

Closed issues:

• at_ccallable function can not be called from within Enzyme (#202)
• Alloc -> Malloc downgrade on sret (#337)
• Inconsistent gradients for edge case (Enzyme v0.10.4 & GPUCompiler v0.16.3) (#432)
• Forward-mode autodiff errors with Assertion v != DIFFE_TYPE::OUT_DIFF' failed.` (#451)
• Illegal address space propagation (#457)
• List comprehension of empty arrays no longer works (#460)
• Illegal address space propagation: Minimal Version (#463)
• Readonly error 1.9 (#486)

Merged pull requests:

• Add no free to functions (#453) (@wsmoses)
• Fix recursive GC (#462) (@wsmoses)
• Add gc sret test (#464) (@wsmoses)
• Deduce writeonly from sret (#468) (@wsmoses)
• Nicer error for forward api (#469) (@bitmyte)
• Fix cache of decayed ptr (#470) (@wsmoses)
• Add additional sret test (#471) (@wsmoses)
• Add better firstptr error (#472) (@wsmoses)
• Generic fixes (#473) (@wsmoses)
• Add internal error handling (#474) (@wsmoses)
• Use variadic post cache store (#475) (@wsmoses)
• Fix Address space cast issues (#476) (@wsmoses)
• Handle memcpy in addrcast (#477) (@wsmoses)
• Add functions to nofree (#479) (@swilliamson7)
• Bump Enzyme_jll to 0.0.40 (#481) (@vchuravy)
• Fix post cache store (#482) (@wsmoses)
• Fix nest and add test (#484) (@wsmoses)
• Add error message if using a chunk size of 0 (#485) (@wsmoses)
• Fix test seg warn (#487) (@wsmoses)
• Bump jll (#490) (@wsmoses)
• vc/fix 1.9 (#491) (@vchuravy)

Enzyme v0.10.6

Diff since v0.10.5

Closed issues:

• atan double argument form errors (#459)

Merged pull requests:

• Add atan2 (#461) (@wsmoses)

Enzyme v0.10.5

Diff since v0.10.4

Closed issues:

• Custom pullbacks for PP AD example (#26)
• Crash with Tullio (#99)
• Correctly deal with array alloc of eltype(A) = Any (#118)
• Union of Return (#154)
• TypeAnalysis fails on Julia 1.8 LLVM 13 (#171)
• Crash on multidimensional setindex (#175)
• Understanding a huge crash report (#181)
• Global leads to wrong gradient (#185)
• Kernel crashing issue (#192)
• Union leads to illegal type analysis (#236)
• Segfault when testing package on Julia 1.8 on a mac. (#250)
• Attributor pass broken on 1.8-dev/ LLVM13 (#267)
• foldl with CUDA.jl results in jl_lazy_load_and_lookup (#272)
• Assertion `getMinSignedBits() <= 64 && "Too many bits for int64_t"' failed. (#273)
• Verification or llvm inliner error? (#276)
• Out of place broadcasting creates unexpected results (#278)
• Segfault when used with Optim.jl (#298)
• MethodError on dynamic Val construction (#309)
• Support reshaping via jl_array_reshape (#310)
• Segfault when using with BLAS (#315)
• Nested broadcast not working (#338)
• MethodError when trying to autodiff through a complex matrix (#359)
• Differentiating trigonometric functions with KernelGradients and CUDAKernels (#375)
• Jacobian producing unexpected results (#377)
• Problems for CUDA.jl kernel autodiff (#378)
• Mutation of gradient input in MatMul example for Square Matrices (#382)
• limitations of CUDA kernels differentiation (#383)
• guess_activity(::Type{Union{}}) is ambiguous (#386)
• Correctness bug with 2D array offsets (#388)
• enzyme doesn't work on Apple silicon (#394)
• Segfault with dense matrix multiplication (#391)
• Segfault with sparse matrix multiplication (#393)
• sum() sometimes returns incorrect result (#397)
• exp10 causes illegal type analyses error (#401)
• differentiation works on cpu, fails on gpu, for exponentiation (#412)
• memory error on second invocation of derivative GPU kernel (#413)
• v1.6 only LLVM-level error (#414)
• Concatenating AD for multiple GPU kernels (#415)
• can't push! twice (#417)
• atomic operations error (#421)
• Uncaught Enzyme exception crashes Julia 1.8 (#423)
• Support for tan operation (#424)
• sqrt in CUDA kernel gives error (#426)
• Uncaught LLVM-level error from neural SDE (#429)
• Push Undefined Memory Error (#430)
• Gradients for struct fields are zero when a Bool field is present (#431)
• BatchDuplicated in Forward breaks for sinc (#434)
• BLAS fallback has short-circuiting zero path (#445)
• Gradient fails to deal with GPU movements (#448)
• Gradient through acos errors (#449)
• Pullback for array function gives zero gradients with forward mode (#450)
• orig is not defined (#454)

Merged pull requests:

• Mark aliasing property of error (#104) (@wsmoses)
• Set default activity of any (#187) (@wsmoses)
• Inactive arg copy (#279) (@wsmoses)
• Add fast exp (#339) (@wsmoses)
• Add ijl box variations (#374) (@wsmoses)
• Add preserve nvvm (#379) (@wsmoses)
• Add array reshape rev (#380) (@wsmoses)
• Add runtime activity for fwd (#385) (@wsmoses)
• Box tutorial documentation (#387) (@swilliamson7)
• Fix getwidth bug (#390) (@wsmoses)
• fix cuda tests (#395) (@wsmoses)
• Custom allocator support (#396) (@vchuravy)
• Handle TSCtx on 1.9 (#398) (@vchuravy)
• Use literate for box example (#399) (@vchuravy)
• Use gc_alloc_obj for tapes containing Julia objects (#400) (@vchuravy)
• Handle exp10 (#403) (@wsmoses)
• Better unimlemented vector assertions (#405) (@wsmoses)
• Fix movebefore (#406) (@wsmoses)
• Use Anonymous Struct for sret (#408) (@vchuravy)
• docs for Sparse Array Structure Initializer (#409) (@ryanstoner1)
• Copy over correct IR dumping code (#410) (@pchintalapudi)
• Bump JLL (#411) (@wsmoses)
• support double push (#420) (@vchuravy)
• Rules for sinh and cosh (#425) (@sriharikrishna)
• Detect integer form of eqtable (#427) (@wsmoses)
• Fix nightly (#433) (@wsmoses)
• Fix zero'ing in augmented forward (#436) (@wsmoses)
• Reuse julia type rules (#437) (@wsmoses)
• Dynamic val fix (#438) (@wsmoses)
• Handle not concrete type trees (#439) (@wsmoses)
• Remove preprocessed funcs & fix typeof (#440) (@wsmoses)
• Fix nightly pt2 (#441) (@wsmoses)
• Bump jll (#442) (@wsmoses)
• Add sinc (#444) (@wsmoses)
• Fix typo delf -> delF (#446) (@avik-pal)
• Add stacktraces to errors (#452) (@wsmoses)
• Fix orig name bug (#455) (@wsmoses)
• Add acos (#456) (@wsmoses)
• Cleanup CI (#458) (@vchuravy)

Enzyme v0.10.4

Diff since v0.10.3

Closed issues:

• Incorrect gradients with views (#356)
• Failed when trying to create a new vector whose length depends on input (#366)

Enzyme v0.10.3

Diff since v0.10.2

Merged pull requests:

• Use a custom variant of Ref that is fails instead of handing out null-pointers (#372) (@vchuravy)

Enzyme v0.10.2

Diff since v0.10.1

Closed issues:

• Differentiate dsyevr_64_ (#64)
• Crash with Tullio (#99)
• Enable parallel flag as option within autodiff (#106)
• dgemm_64 support missing (#113)
• dgbtrs support missing (#136)
• sgemv support missing (#162)
• Unsupported Val in CUDA kernel - Enzyme v0.10.0 - GPUCompiler v0.15 (#358)
• Cannot use Enzyme on clean Julia installation (#360)
• Wrong gradients (mutating buffers) (#362)
• Unsupported call through a literal pointer (call to __libc_malloc) (#367)

Merged pull requests:

• Don't restore pointers when doing GPU compilation (#370) (@vchuravy)
• Describe tmp activity (#371) (@wsmoses)

Enzyme v0.10.1

Diff since v0.10.0

Closed issues:

• CUDA support broken since Enzyme v0.8 and GPUCompiler v13 (#230)
• Thread integration broken on Julia 1.8 (#283)
• Obtain gradient function using Enzyme.jl (#357)

Merged pull requests:

• Make same or 1 recursive (#361) (@wsmoses)

Enzyme v0.10.0

Diff since v0.9.6

Closed issues:

• Analysis error ("could not deduce type of integer") on GEP when the stars align correctly (#311)
• Symbols not found: [ cblas_xerbla ] (#312)
• Segmentation fault during NVPTX emission due to debuginfo (#316)
• Forward over Reverse Crash (#320)
• Support for BigFloat (#348)
• AbstractDifferentiation (#349)
• Problem in Windows Subsystem Linux (WSL2) (#353)

Merged pull requests:

• Generalize activity sets (#296) (@wsmoses)
• merge autodiff and fwddiff API (#306) (@vchuravy)
• Handle xerbla (#317) (@wsmoses)
• add tests to Jacobian using LinAlg (#319) (@vchuravy)
• Fix metadata (#323) (@wsmoses)
• Fix debug info2 (#324) (@wsmoses)
• Fix scope bug for arg (#325) (@wsmoses)
• Fix in to haskey (#327) (@wsmoses)
• Add log1p (#328) (@wsmoses)
• Re-enable tests (#329) (@wsmoses)
• Move validation up (#330) (@vchuravy)
• Add type error and array fix (#332) (@wsmoses)
• Rand (#333) (@wsmoses)
• Fix type analysis on 1.8 (#335) (@wsmoses)
• Fix jl_arraycopy on phi (#336) (@wsmoses)
• Ensure that loading CUDA doesn't break things on 1.6 (#341) (@vchuravy)
• Rebase: Move compilation out of threadsfor (#342) (@vchuravy)
• Unbreak CI until JLL bump (#344) (@vchuravy)
• CompatHelper: bump compat for GPUCompiler to 0.16, (keep existing compat) (#345) (@github-actions[bot])
• Jacobian returns one array (#351) (@vchuravy)
• Remove tuple from gradient(::Forward) without chunking (#352) (@vchuravy)
• Bump jll version (#355) (@vchuravy)

Enzyme v0.9.6

Diff since v0.9.5

Merged pull requests:

• remove Manifest now that 0.15.1 is out (#300) (@vchuravy)
• Upgrade to Enzyme 0.0.31 (#302) (@vchuravy)
• Apply safe_name to new functions (#305) (@vchuravy)
• fix doctests (#307) (@vchuravy)
• Convert sentinal for sub_transfer (#308) (@vchuravy)
• Backports for 0.9.6 (#314) (@vchuravy)

Enzyme v0.9.5

Diff since v0.9.4

Merged pull requests:

• Backports 0.9.5 (#297) (@vchuravy)

Enzyme v0.9.4

Diff since v0.9.3

Closed issues:

• [Vector][Forward Mode] Illegal replacement (#281)
• Augmented forward pass missing for grow_end (#287)

Merged pull requests:

• Update Enzyme_jll (#286) (@vchuravy)
• fix misnomer and add ijl variant for grow_end (#288) (@vchuravy)
• Better error on unknown concrete type (#290) (@vchuravy)
• Backports for 0.9.4 (#291) (@vchuravy)

Enzyme v0.9.3

Diff since v0.9.2

Merged pull requests:

• Initial array push handling (#105) (@wsmoses)
• Cleanup generic_call a bit (#268) (@vchuravy)
• add an assertion build to CI (#269) (@vchuravy)
• Make sure that the GC passes run over our functions (#271) (@vchuravy)

Enzyme v0.9.2

Diff since v0.9.1

Closed issues:

• Illegal updateAnalysis prev:{[-1]:Integer} new: {[-1]:Pointer} (#212)

Merged pull requests:

• Add dictionary test (#264) (@vchuravy)

Enzyme v0.9.1

Diff since v0.9.0

Merged pull requests:

• Nix the Runtime, we shouldn't need it (#229) (@vchuravy)
• Test that exceptions are thrown (#243) (@vchuravy)
• Better errors for all (#263) (@vchuravy)

Enzyme v0.9.0

Diff since v0.8.5

Closed issues:

• const global ref seems to be internalized during compilation (#203)
• Enzyme.jl hits an unknown intrisic error when diffing through sincos. (#219)
• Zygote.@adjoint from Enzyme.autodiff (#233)
• Segfault / Bus Error (#238)
• MPI Type error (#244)
• Exception (about missing gemm!?) not throwing at the Julia level and instead by LLVM (#249)

Merged pull requests:

• Handle sincos (#224) (@wsmoses)
• add buildkite (#231) (@vchuravy)
• Add Buildkite to stable (#232) (@vchuravy)
• Adjust to threads ABI in 1.8 (#234) (@vchuravy)
• Add BLAS detection (#237) (@wsmoses)
• MISC BUDE.jl and LULESH.jl fixes (#239) (@wsmoses)
• Add ABI change (#242) (@wsmoses)
• Additional BLAS fixes (#246) (@wsmoses)
• Debug Macos (#247) (@wsmoses)
• Cleanup FFI support (#248) (@vchuravy)
• Fix alloc handler (#251) (@wsmoses)
• Add stored inline inactive (#252) (@wsmoses)
• Bump GPUCompiler compat (#253) (@wsmoses)
• Fix blas handle (#254) (@wsmoses)
• Parallel without GPUCompiler caching (#255) (@wsmoses)
• require GPUCompiler 0.14 (#256) (@vchuravy)
• Bump JLL (#257) (@wsmoses)
• Fix BLAS on Windows (#260) (@vchuravy)