v0.2.21

Reactant v0.2.21

Diff since v0.2.20

Merged pull requests:

• PTX fma and other flags (#585) (@wsmoses)

Closed issues:

• CUDA support broken in 0.2.19? (#583)

v0.2.20

Reactant v0.2.20

Diff since v0.2.19

Merged pull requests:

• [ReactantExtra] Use XLA commit for building with CUDA 12.1 (#579) (@giordano)
• Regenerate MLIR Bindings (#580) (@github-actions[bot])
• Profiler annotations & tutorial (#582) (@Pangoraw)
• Fix for unknown cuda drivers (#586) (@wsmoses)

Closed issues:

• Profiling Tutorial (#581)

v0.2.19

Reactant v0.2.19

Diff since v0.2.18

Merged pull requests:

• respect scopping rules in for (#310) (@Pangoraw)
• Despecialize make_tracer (#540) (@wsmoses)
• XLA profiler (#541) (@Pangoraw)
• feat: add isnan and isfinite dispatches (#544) (@avik-pal)
• unionnone (#545) (@wsmoses)
• print (#547) (@wsmoses)
• add int override (#549) (@wsmoses)
• docs: missing doc links in sidebar and navbar (#551) (@avik-pal)
• Fix downgrader CI job (#553) (@giordano)
• Simplify process to builds docs (#554) (@giordano)
• fix: define getindexing into sub reshaped array (#556) (@avik-pal)
• fix: inconsistent return dims (#558) (@avik-pal)
• [CI] Format generated files twice to work around JuliaFormatter bug (#560) (@giordano)
• Regenerate MLIR Bindings (#561) (@github-actions[bot])
• Format code of branch "main" (#562) (@github-actions[bot])
• [GHA] Add paths settings for workflow triggers (#563) (@giordano)
• CUDA: fix nv intrinsic errs (#564) (@wsmoses)
• feat: support arbitrary structures in control flow (#565) (@avik-pal)
• [GHA] Fix syntax of regenerate MLIR bindings workflow (#566) (@giordano)
• More jll/cuda stuff (#567) (@wsmoses)
• Format code of branch "main" (#568) (@github-actions[bot])
• fix: reduction of integers (#573) (@avik-pal)
• profiler: Add option to generate perfetto url (#575) (@Pangoraw)
• [CI] Remove useless call to Pkg.instantiate (#576) (@giordano)
• fix: specialize / on integer types (#577) (@avik-pal)

Closed issues:

• setindex! doesn't work with @trace (#210)
• Incorrect code-generation for @trace for ... (#301)
• Precompiling Reactant errors in GPU-related symbol (#526)
• How to check NaN? (#542)
• MethodError for setindex! in sum! (#548)
• Error in fill! (#550)
• [Docs] Stableurl doesn't go anywhere (#552)
• Subarray indexing error (#555)
• Mark all kernel arguments with the llvm.noalias attribute (#571)
• [Profiling] Add option to autogenerate perfetto and/or tensorboad url (#572)
• Incorrect division of Integers (/ operator not div) (#574)

v0.2.18

Reactant v0.2.18

Diff since v0.2.17

Merged pull requests:

• linearize kernel args (#497) (@mofeing)
• Ka2 (#498) (@wsmoses)
• Regenerate MLIR Bindings (#501) (@github-actions[bot])
• linearize aliased kernel args (#504) (@jumerckx)
• Split should_rewrite_ft for call and invoke expressions, and overlay Base._unique_dims (#505) (@mofeing)
• feat: add rsqrt simplification (#506) (@avik-pal)
• Regenerate MLIR Bindings (#507) (@github-actions[bot])
• Format code of branch "main" (#509) (@github-actions[bot])
• feat: optimization passes (#510) (@avik-pal)
• Regenerate MLIR Bindings (#513) (@github-actions[bot])
• Make v and hcat with numbers work. (#514) (@jaeminoh)
• XLA Allocator stats (#517) (@Pangoraw)
• fix: generalize broadcast_in_dims for setindex (#518) (@avik-pal)
• Format code of branch "main" (#520) (@github-actions[bot])
• WIP: adapt to sroa jll (#521) (@wsmoses)
• Kernel: support constant input arg (#522) (@wsmoses)
• Implement isnan and isfinite for TracedRNumber (#525) (@Pangoraw)
• Format code of branch "main" (#528) (@github-actions[bot])
• feat: sorting and related functions (#529) (@avik-pal)
• Regenerate MLIR Bindings (#531) (@github-actions[bot])
• Format code of branch "main" (#533) (@github-actions[bot])
• Generalize precompilation support (#534) (@wsmoses)
• More constprop (#536) (@wsmoses)
• Fix tolerance on loggamma integration test (#537) (@wsmoses)
• Fix missing dialects in docs (#538) (@wsmoses)

Closed issues:

• minimize XLA error in gemm_autotuner for CUDA (#444)
• Not support partialsortperm ? (#485)
• KernelAbstractions + Reactant: UndefVarError: pop not defined (#488)
• Infinite recursion on unique(::Vector{Symbol}) within Reactant (#493)
• Concatenation of scalar and ConcreteRArray gives a Vector (#511)
• Incorrect broadcast_to_size implementation (#512)
• How to set NaN values in an RArray to a certain number? (#524)
• Precompilation of Reactant 0.2.1x fails (#527)

v0.2.17

Reactant v0.2.17

Diff since v0.2.16

Merged pull requests:

• gpu, llvm and nvvm dialect bindings (#315) (@jumerckx)
• fix: recursion in AbsInt working (#483) (@avik-pal)
• add test for #226 (#486) (@jumerckx)
• fix: missing scalar indexing check for setindex (#491) (@avik-pal)
• fix: incorrect IR for traced RNGs (#494) (@avik-pal)

Closed issues:

• Add custom kernel infrastructure for CUDA and AMD calls (#30)
• passing same argument twice leads to a bounds error while compiling (#226)
• AbstractInterpreter and Recursive Calls into Overlay-ed Functions (#393)
• make_mlir_fn broken on Julia 1.10 (#482)
• Incorrect IR generated (#492)

v0.2.16

Reactant v0.2.16

Diff since v0.2.15

Merged pull requests:

• CuArray tracing (#475) (@wsmoses)
• Improve error message on broadcast (#476) (@wsmoses)
• Fix traced rnumber promotion (#477) (@wsmoses)
• Fix type unstable stack (#478) (@wsmoses)
• Rewrite apply iterate (#479) (@wsmoses)
• Improve cuda error (#480) (@wsmoses)
• Format code of branch "main" (#481) (@github-actions[bot])

v0.2.15

Reactant v0.2.15

Diff since v0.2.14

Merged pull requests:

• precompile first interpretation (#353) (@glou-nes)
• [GHA] Set author of PRs created by scheduled jobs to Enzyme bot (#455) (@giordano)
• feat: partial support for boolean indexing (#457) (@avik-pal)
• feat: no_nan mode and more passes (#458) (@avik-pal)
• Fix cuda isdefined error (#459) (@wsmoses)
• feat: more trigonometric functions (#462) (@avik-pal)
• fix: check for name and module in function (#463) (@avik-pal)
• Fix offsetarrays support (#464) (@wsmoses)
• Fix opaqueclosure Worldage issue (#465) (@wsmoses)
• Format code of branch "main" (#466) (@github-actions[bot])
• CompatHelper: add new compat entry for OffsetArrays in [weakdeps] at version 1 for package ., (keep existing compat) (#467) (@github-actions[bot])
• CompatHelper: add new compat entry for OffsetArrays at version 1 for package test, (keep existing compat) (#468) (@github-actions[bot])
• Update index.md (#469) (@x66ccff)
• Format code of branch "main" (#471) (@github-actions[bot])
• feat: missing mean(f, ...) dispatches (#472) (@avik-pal)
• Update index.md, delete the
  tag (#473) (@x66ccff)
• Provide better error message if calling thunk with wrong types (#474) (@wsmoses)

Closed issues:

• Precompile first interpretation (#348)
• How to empty cache? (#460)

v0.2.14

Reactant v0.2.14

Diff since v0.2.13

Merged pull requests:

• More mul overloads (#446) (@wsmoses)
• fix: use pip for jax (#450) (@avik-pal)
• Pipeline for nested enzyme differentiation (#452) (@wsmoses)
• Revert modularize bazel build to unbreak yggy (#453) (@wsmoses)
• CompatHelper: add new compat entry for SpecialFunctions in [weakdeps] at version 2 for package ., (keep existing compat) (#454) (@github-actions[bot])

Closed issues:

• Bad interaction between JAX and Reactant on aarch64-linux (#438)
• [2nd Order AD] Regularization term in loss function (#449)

v0.2.13

Reactant v0.2.13

Diff since v0.2.12

Merged pull requests:

• SpecialFunctions simple functions (#384) (@glou-nes)
• Modularize Bazel build (#421) (@mofeing)
• feat: support Base.stack (#433) (@avik-pal)
• Fix mul overload (#440) (@wsmoses)
• fix: handle aos for mul (#441) (@avik-pal)
• feat: missing BF16 dispatches (#443) (@avik-pal)
• refactor: rework TracedRNG to be similar to other types (#448) (@avik-pal)

Closed issues:

• 2nd order AD fails (#298)

v0.2.12

Reactant v0.2.12

Diff since v0.2.11

Merged pull requests:

• feat: indexing using traced values (#434) (@avik-pal)
• Regenerate MLIR Bindings (#435) (@github-actions[bot])
• fix: try building with cudnn 9.4 (#436) (@avik-pal)
• chore: bump jll (#437) (@avik-pal)

Closed issues:

• Scalar indexing triggered when requesting a view of a traced array (#408)