v0.2.10

Reactant v0.2.10

Diff since v0.2.9

Merged pull requests:

• fix: correct usage of Ops.select for Base.ifelse (#332) (@avik-pal)
• feat: add dynamic_update_slice_const_prop pass + tril + triu (#334) (@avik-pal)
• Fix for interpreter (#341) (@wsmoses)

v0.2.9

Reactant v0.2.9

Diff since v0.2.8

Merged pull requests:

• No rules v2 (#326) (@wsmoses)
• perf: move all benchmarks to buildkite (#327) (@avik-pal)
• add select op shortcut (#328) (@mofeing)
• add shortcut to dot_general op (#330) (@mofeing)

Closed issues:

• Move benchmarks running in GitHub Runner to Buildkite CI (#207)
• conv gradient is not implemented in EnzymeJAX (#214)

v0.2.8

Reactant v0.2.8

Diff since v0.2.7

Merged pull requests:

• Add IFRT C-API bindings (#178) (@mofeing)
• Reactant interpreter overload (#259) (@wsmoses)
• First batch of shortcuts to HLO ops (#267) (@mofeing)
• fix attribute for arrray of bools (#279) (@Pangoraw)
• enable new optimizations (#280) (@Pangoraw)
• gensym in @code_hlo (#281) (@jumerckx)
• test for traced_type where array isn't captured by type param (#282) (@jumerckx)
• fix: don't print "Module" (#283) (@avik-pal)
• docs: nicer documentation setup (#284) (@avik-pal)
• add check for null in show(::Operation). (#285) (@jumerckx)
• build(deps): bump codecov/codecov-action from 4 to 5 (#286) (@dependabot[bot])
• Format code of branch "main" (#289) (@github-actions[bot])
• CompatHelper: add new compat entry for EnzymeCore at version 0.8, (keep existing compat) (#290) (@github-actions[bot])
• chore: specify sources for local dev (#292) (@avik-pal)
• feat: define aos_to_soa (#294) (@avik-pal)
• add fill to TracedRNumber (#295) (@glou-nes)
• Format code of branch "main" (#296) (@github-actions[bot])
• fix: vararg compilation (#299) (@avik-pal)
• ci: add downgrade testing (#300) (@avik-pal)
• Fix tracedarray (#302) (@wsmoses)
• fix create_result(::Array) for non-vectors (#304) (@mofeing)
• implement collect for TracedRArray (#305) (@mofeing)
• fix YaoBlocks.mat for rotation gates with TracedRNumber parameters (#306) (@mofeing)
• use row major when building attributes (#307) (@Pangoraw)
• Format code of branch "main" (#309) (@github-actions[bot])
• Speed up invoke usage (#313) (@wsmoses)
• specify aarch64-linux toolchain for Yggdrasil build (#316) (@mofeing)
• add optional debug info to Ops calls (#317) (@mofeing)
• Implement NNlib.∇conv_data! and NNlib.∇conv_filter! (#318) (@Pangoraw)
• Ops remaining basic numerical test (#320) (@glou-nes)
• fix: preserve IOContext for printing (#322) (@avik-pal)
• fix: #196 (#323) (@glou-nes)
• [JLL] Add conversion from llvm to mlir (#324) (@wsmoses)

Closed issues:

• IFRT tracking issue (#176)
• Abstract interpretation over ReactantInterpreter sometimes skips overriden methods (#189)
• crash with parameter less function (#196)
• Mismatch computing Enzyme.gradient between regular array and ConcreteRArray (#236)
• Add downgrade testing CI (#271)
• Inconsistent result for iszero for booleans (#277)
• Implement fill for TracedRNumber (#288)
• compilation failure of getindex with a type-assert (#293)

v0.2.7

Reactant v0.2.7

Diff since v0.2.6

Merged pull requests:

• fix: before_enzyme optimizations (#275) (@avik-pal)
• docs: add note on installing cuda version of Reactant on login nodes (#276) (@avik-pal)

Closed issues:

• What is a device to the RArray? (#16)

v0.2.6

Reactant v0.2.6

Diff since v0.2.5

Merged pull requests:

• feat: functionalities for supporting NeuralOperators.jl (#217) (@avik-pal)
• CompatHelper: add new compat entry for AbstractFFTs in [weakdeps] at version 1, (keep existing compat) (#229) (@github-actions[bot])
• feat: overload mul! (#230) (@avik-pal)
• feat: extend mul! work for subarrays (#231) (@avik-pal)
• fix abs of complex number (#239) (@Pangoraw)
• feat: support setindex with views (#240) (@avik-pal)
• fix: bypass segfault with fill complex (#245) (@avik-pal)
• feat: add support for clamp and clamp! (#247) (@avik-pal)
• fix: reduction of complex arrays (#248) (@avik-pal)
• feat: add support for NNlib make causal mask (#249) (@avik-pal)
• feat: partial NNlib.gather support + better indexing support (#252) (@avik-pal)
• Add 32bit CI (#254) (@wsmoses)
• implement @trace for (#255) (@Pangoraw)
• Format code of branch "main" (#257) (@github-actions[bot])
• feat: more coverage for NNlib functions (#258) (@avik-pal)
• feat: inherit scalar indexing functionality from GPUArraysCore (#268) (@avik-pal)
• Disable verifier in main pass manager pipeline (#269) (@Pangoraw)
• chore: bump ReactantCore version (#270) (@avik-pal)
• fix: add GPU backend as default if available (#272) (@avik-pal)
• Use gensym-ed variable in @compile and @jit (#274) (@jumerckx)

Closed issues:

• YaoBlocks.cunmat has problems during tracing (#174)
• Lux example MLP with TrainState API throws JIT session errors (#220)
• Use GPUArraysCore for scalar indexing flags (#232)
• Reactant fails to compile a function accepting complex parameters (#235)
• Name clash when calling Reactant.@compile over a function that is named f (#237)
• getindex assumes static indexing (#243)
• mlirDenseElementsAttrSplatGet with ComplexNumbers lead to a SegFault (#244)
• conv! (#251)
• 32-Bit CI (#253)

v0.2.5

Reactant v0.2.5

Diff since v0.2.4

Merged pull requests:

• Replace MLIR_CAPI_EXPORTED for extern "C" (#228) (@mofeing)

v0.2.4

Reactant v0.2.4

Diff since v0.2.3

Merged pull requests:

• Move Julia bindings to MLIR dialects out of JLL (#166) (@mofeing)
• Implement conj, conj! for TracedRArray (#169) (@mofeing)
• feat: allow passing a optimize option to MLIR (#171) (@avik-pal)
• Format code of branch "main" (#175) (@github-actions[bot])
• Regenerate MLIR Dialects (#177) (@github-actions[bot])
• Update Project.toml (#179) (@wsmoses)
• Fix similar on TracedRArray (#184) (@mofeing)
• feat: support conditionals (#185) (@avik-pal)
• Improve GitHub Actions settings (#186) (@giordano)
• build(deps): bump codecov/codecov-action from 1 to 4 (#187) (@dependabot[bot])
• build(deps): bump peter-evans/create-pull-request from 6 to 7 (#188) (@dependabot[bot])
• Add needs primal for forward mode (#192) (@wsmoses)
• fix primitive_type for complex (#193) (@Pangoraw)
• Add jit macro for simplified testing (#194) (@wsmoses)
• complex DenseElementsAttribute (#195) (@Pangoraw)
• Make libMLIR_h automatically regenerate (#197) (@Pangoraw)
• prevent docs warning (#199) (@Pangoraw)
• update libMLIR_h bindings and fix bindings update step n°2 (#202) (@Pangoraw)
• Refactor regenerate-mlir-bindings.yml workflow (#204) (@mofeing)
• feat: support tracing scalars (#205) (@avik-pal)
• Add C-API for constructing Complex Attributes (#208) (@mofeing)
• feat: allow type-casting numbers to tracednumbers (#209) (@avik-pal)
• feat: specialize dispatches for faster concrete array generation (#213) (@avik-pal)
• feat: add sync to synchonize all result buffers (#216) (@avik-pal)
• fix complex attr c-api (#218) (@mofeing)
• fix resource exhaustion bug #190 (#219) (@mofeing)
• Refactor mlirComplexAttrDoubleGet, mlirComplexAttrDoubleGetChecked (#221) (@mofeing)
• feat: allow conversion to numbers (#222) (@avik-pal)
• Regenerate MLIR Bindings (#223) (@github-actions[bot])
• refactor: use @jit in the tests (#224) (@avik-pal)
• feat: make compile and code_hlo macros work with broadcasting (#225) (@avik-pal)

Closed issues:

• Shortcuts for to_rarray (#138)
• Infinte recursion when compiling a function (#168)
• ConcreteRArray on Complex array zeroes the buffer (#170)
• XlaRuntimeError when passing a Complex buffer to a compiled function (#172)
• Forward-mode differentiation is broken (#183)
• Resource exhaustion when compiling many functions (#190)
• Error on regenerating "libMLIR_h.jl" on GitHub runner (#201)

v0.2.3

Reactant v0.2.3

Diff since v0.2.2

Merged pull requests:

• Export GetDefaultTargetTriple symbol (#160) (@mofeing)
• feat: implement a separate TracedRNumber (#161) (@avik-pal)
• Generalize Base._cat to non-Val, typed Base._cat_t and implement typed_hcat, typed_vcat, typed_hvcat, typed_hvncat (#163) (@mofeing)
• Bypass identical construction on TracedRArray (#165) (@mofeing)
• test: separate out test groups (#167) (@avik-pal)

Closed issues:

• Compiling leads to KeyError (#107)
• Error with vcat (#139)
• Inconsistent semantics of 0-dim RArrays compared to Julia semantics (#141)
• Missing hcat/vcat/hvcat dispatches (#148)

v0.2.2

Reactant v0.2.2

Diff since v0.2.1

Merged pull requests:

• feat: compiling vision transformers (#130) (@avik-pal)
• feat: robust handling of wrapped arrays of reactant arrays (#133) (@avik-pal)
• Replace getfield for traced_getfield in set! (#134) (@mofeing)
• perf: benchmarking CI (#136) (@avik-pal)
• Add fallback gpu setindex (#137) (@wsmoses)
• Bump dependencies (#143) (@wsmoses)
• Build (#145) (@wsmoses)
• feat: generalize indexing to all wrappers (#146) (@avik-pal)
• Arm64 (#147) (@wsmoses)
• Format code of branch "main" (#153) (@github-actions[bot])
• fix: broadcasted type casting (#156) (@avik-pal)
• Fix vararg (#158) (@wsmoses)

Closed issues:

• Handling exponents (#27)
• Incosistent semantics of element-wise application of array functions compared to Julia (#51)
• NNlib activations cannot be compiled without manual dispatches (#54)
• Linear algebra on TracedRArray (#131)
• scalar broadcasting not functional (#135)
• libdevice not found (#149)
• Vararg compilation broken (#151)
• Can I compile in-place functions? (#152)
• Optimisers.jl fails to compile (#154)
• Type-casting doesn't compile (#155)
• Unbreak Reactant benchmark CI (#157)

v0.2.1

Reactant v0.2.1

Diff since v0.2.0

Merged pull requests:

• Format code of branch "main" (#101) (@github-actions[bot])
• Compile NNlib.maxpool and NNlib.meanpool (#102) (@Pangoraw)
• feat: handle cat/hcat/vcat (#103) (@avik-pal)
• feat: implementing view/getindex/setindex (#104) (@avik-pal)
• perf: test compile times for neural networks -- VGG16 (#105) (@avik-pal)
• fix: local JLL build script (#108) (@avik-pal)
• NNlib: forward feature group counts to convolution (#109) (@Pangoraw)
• tests: more comprehensive NN testing (#110) (@avik-pal)
• ci(github-actions): update the formatter action (#111) (@avik-pal)
• Format code of branch "main" (#112) (@github-actions[bot])
• Format code of branch "main" (#113) (@github-actions[bot])
• feat: allow building for CPU target on CUDA-enabled machines (#114) (@avik-pal)
• NNlib: handle kernel flip with reverse instead of window_reversal (#115) (@Pangoraw)
• Format code of branch "main" (#116) (@github-actions[bot])
• Configure CompatHelper workflow (#118) (@mofeing)
• chore: testing Enzyme 0.13 (#119) (@avik-pal)
• Fix non-used argument (#123) (@wsmoses)
• Fix typo in mlirIsNull call (#124) (@mofeing)
• Format code of branch "main" (#125) (@github-actions[bot])
• Refactor compile and test buffer donation (#126) (@mofeing)
• Replace IdDict for OrderedIdDict (#128) (@mofeing)
• CompatHelper: add new compat entry for OrderedCollections at version 1, (keep existing compat) (#129) (@github-actions[bot])

Closed issues:

• Set up GPU CI (#41)
• Print "donated" attribute on @code_hlo (#97)
• Grouped Convolution generates incorrect code (#106)
• Compiling leads to KeyError (#107)
• Update to Enzyme v0.13 (#117)
• gelu needs to be fixed (#121)
• Max Pooling doesn't give correct answers for certain padding (#122)