v0.2.1

cuTile v0.2.1

Diff since v0.2.0

Merged pull requests:

• Change function name sanitization to support unicode (#183) (@AntonOresten)
• Add softmax example (#185) (@maleadt)
• Fix reduce identity encoding for floats and integers. (#186) (@maleadt)

v0.2.0

cuTile v0.2.0

Diff since v0.1.2

Breaking changes: see https://juliagpu.org/post/2026-04-08-cutile_0.2/

Merged pull requests:

• Add alias-aware token threading for memory operations. (#89) (@shreyas-omkar)
• Matmul: Switch to trailing batch dims and allow mat-vec, vec-mat (#132) (@AntonOresten)
• Register ghost CGVals for ghost-type kernel arguments (#133) (@maleadt)
• Add mapreduce. (#134) (@maleadt)
• Fix @device_code_tiled crash on kernels with reduce. (#135) (@maleadt)
• Add atomic max, min, or, and, xor operations (#136) (@maleadt)
• Clean-up and optimize mapreduce (#137) (@maleadt)
• Support Constant(nothing) as kernel argument (#138) (@maleadt)
• Align keyword argument with cuTile Python (#139) (@maleadt)
• Switch to EnumX.jl. (#140) (@maleadt)
• Use atomic reductions to eliminate two-pass overhead (#141) (@maleadt)
• Store tile shapes in row-major order (#142) (@maleadt)
• Keep shapes until bytecode emission. (#143) (@maleadt)
• Update examples (#144) (@maleadt)
• Align FFT examples. (#145) (@maleadt)
• Declarative IR rewrite pattern infrastructure (#147) (@maleadt)
• Simplify using new IRStructurizer.jl tools. (#148) (@maleadt)
• Add loop parallel store optimization and DCE pass (#149) (@maleadt)
• Add normalize pass to lower Julia intrinsics to cuTile equivalents. (#150) (@maleadt)
• Update benchmarks scripts and results. (#151) (@maleadt)
• Improve the rewriter (#152) (@maleadt)
• Fix divisibility computation exceeding max_divisor. (#153) (@maleadt)
• Add simple constant folding (#154) (@maleadt)
• Add algebraic simplifications to get rid of 1-based index IR bloat (#156) (@maleadt)
• Extend reflection: support constants, add kwargs. (#157) (@maleadt)
• Canonicalize the IR before optimization (#158) (@maleadt)
• Add reshape to transparent ops in rewrite pattern matching (#159) (@maleadt)
• Constant propagation and identity fold for the rewrite framework (#160) (@maleadt)
• Add cmp strength reduction (#161) (@maleadt)
• Update benchmark times. (#162) (@maleadt)
• Add Mixture of Experts (MoE) example (#163) (@maleadt)
• Fix alias analysis not tracking pointer aliases through getfield/offset (#164) (@maleadt)
• Add a LICM pass (#165) (@maleadt)
• Optimize gather bounds mask calculation (#166) (@maleadt)
• Rewriter: intermediate type inference and transitive worklist propagation (#167) (@maleadt)
• Rewriter: add some algebra rules (#168) (@maleadt)
• Add AttentionFMHA example (#170) (@maleadt)
• Get rid of ScalarShape, represent as RowMajorShape. (#171) (@maleadt)
• Replace per-op floating-point mode args with a context macro. (#172) (@maleadt)
• Add printing functionality (+ rewriter splat functionality) (#173) (@maleadt)
• Use Julia-native for loops (#174) (@maleadt)
• Add support for debug info. (#175) (@maleadt)
• Add pow2 strength reduction (#177) (@maleadt)
• Add isnan overlay and fix ordering (#179) (@maleadt)
• README update and ct.where removal (#180) (@maleadt)
• Add support for Type arguments and fix static parameter codegen (#181) (@AntonOresten)
• Support for 1.13 (#182) (@maleadt)

Closed issues:

• Alias-aware token threading for better parallelism (#1)
• Port additional examples (#14)
• Support for overflow options in Integer cases (#59)
• Matmul broadcasting (#115)
• TagBot: Manual intervention needed for releases (#131)
• Layernorm regression: Token threading requires loop parallel store optimization (#146)
• Broadcast involving scalar fails (#169)

v0.1.2

cuTile v0.1.2

Diff since v0.1.0

Merged pull requests:

• BFloat16 scalar support (#90) (@AntonOresten)
• Insert trailing 1s instead of leading before broadcasting (#91) (@AntonOresten)
• Allow early returns (#92) (@AntonOresten)
• Generic support for ghost types in launch (#93) (@AntonOresten)
• Fix scalar indexing on TileArrays and add codegen test (#94) (@AntonOresten)
• Add tile-indexed atomic operations (#96) (@AntonOresten)
• Support runtime values in ct.full (Intrinsics.constant) (#100) (@0xtaruhi)
• Fix bitwise ops encoding and add execution tests (#106) (@0xtaruhi)
• Add not_int intrinsic for for-loop iteration support (#107) (@0xtaruhi)
• Add CI (#110) (@maleadt)
• Fixes for CUDA 13.2 (#111) (@maleadt)
• Update execution requirements (#113) (@AntonOresten)
• Add Grid Definition to Quick Start (#116) (@dearmachi)
• Support reductions without dims arg. (#118) (@maleadt)
• Update to CompilerCaching v0.2. (#119) (@maleadt)
• Add architecture-specific optimization annotations (#122) (@maleadt)
• Overlay fill, zeros, and ones from Base (#123) (@AntonOresten)
• Allow transpose on 1D tiles, disallow on >2D (#124) (@AntonOresten)
• Fix codegen crash when SSAValue appears as a statement (#125) (@maleadt)
• Upgrade to IRStructurizer v0.2 (#126) (@maleadt)
• Fix @device_code_tiled for kernels with Constant arguments (#127) (@maleadt)
• Support destructuring arbitrary arguments (#128) (@maleadt)
• Provide host-level broadcast (#129) (@maleadt)

Closed issues:

• Launching kernels with arrays outside GPU (#98)
• Compiler failures with nested while loops and runtime values in ct.full (#99)
• TagBot trigger issue (#101)
• Nested while loops produce wrong results (#102)
• [Feature Request] Support for loops within the kernel (#103)
• InvalidTerminatorError: dot-broadcast inside while loop causes yield type mismatch (#104)
• Bitwise operations on tiles crash tileiras with ProcessExited(3) (#105)
• Pre-Blackwell support (#109)
• Architecture-specific configuration (#112)
• transpose semantics on Tile (#114)
• Splatted tile sizes to align with Julia functions (#117)
• TagBot doesn't work (#120)
• Downgrading deps makes tests fail (#130)

v0.1.1

What's Changed

• BFloat16 scalar support by @AntonOresten in #90
• Generic support for ghost types in launch by @AntonOresten in #93
• Fix scalar indexing on TileArrays and add codegen test by @AntonOresten in #94
• Support runtime values in ct.full (Intrinsics.constant) by @0xtaruhi in #100
• Allow early returns by @AntonOresten in #92
• Add tile-indexed atomic operations by @AntonOresten in #96
• Insert trailing 1s instead of leading before broadcasting by @AntonOresten in #91

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• Port the batched matmul example. by @maleadt in #8
• IRStructurizer: Switch to code_ircode by @maleadt in #6
• Add and clean-up intrinsics by @maleadt in #12
• Lay out SSAArray as a StructOfArrays by @vchuravy in #5
• Deduce view index type from array fields. by @maleadt in #17
• Permute during reshape to support column-major storage. by @maleadt in #18
• Simplify type extraction for tile-only intrinsics. by @maleadt in #19
• Validate tile sizes to be pow2. by @maleadt in #20
• Fix index calculation in 2D gather/scatter by @AntonOresten in #23
• Add initial CI (only codegen tests) by @maleadt in #28
• Simplify constant emission. by @maleadt in #29
• Fix FFT example and add benchmark. by @maleadt in #30
• Expose entry hints through launch by @AntonOresten in #27
• Remove bad matmul-related overrides. by @maleadt in #33
• Expose load/store optimization hints by @AntonOresten in #32
• Refactor examples by @maleadt in #35
• Add UInt8 support to julia_to_tile_dtype by @arhik in #38
• Allow redefinition of kernel methods by @AntonOresten in #31
• Remove Int64 from encode_signed_varint! signature by @arhik in #42
• replace unsigned only and float only varint calls with encode_varint! by @arhik in #43
• IRStructurizer: handle merge phis for if-then regions by @AntonOresten in #53
• Support BFloat16 by @AntonOresten in #34
• Require terminators + validate them by @maleadt in #54
• Move IRStructurizer and FileCheck subpackages out of the repo. by @maleadt in #55
• Integrate widh CompilerCaching.jl by @maleadt in #46
• Remove unnecessary rtol specifications. by @maleadt in #56
• Add more broadcastable operators. by @maleadt in #57
• Add and shorten some tests by @maleadt in #58
• Support Float8 types by @AntonOresten in #36
• feat: Add integer reduction support for reduce ops by @arhik in #37
• Add support for PermutedDimsArray by @AntonOresten in #48
• Add scan (prefix sum) operations support by @arhik in #39
• Switch to an idiomatic reduce/scan API by @maleadt in #60
• Add some more reduction-like operators by @maleadt in #61
• Support map and generalize broadcast by @maleadt in #62
• Compiler simplifications. by @maleadt in #63
• Use Base.transpose. by @maleadt in #64
• Fix and add reflection macros. by @maleadt in #65
• Replace ct.permute with permutedims by @maleadt in #67
• Encode ArraySpec fields as typevars. by @maleadt in #69
• Fix benchmark scripts. by @maleadt in #70
• Make sure constructed ghosts yield SSA values. by @maleadt in #73
• Add order kwarg to load/store for dimension reordering. by @maleadt in #72
• Fix constants and switch intrinsics to constant value inputs + tfuncs by @maleadt in #79
• Auto-match tile rank in load/store by @AntonOresten in #74
• Emit GlobalRef constants eagerly (#77) by @maleadt in #80
• Add scalar ops on Tile and TileArray. by @maleadt in #76
• Switch to Base.size for TileArray by @AntonOresten in #75
• Replace astype by idiomatic convert/broadcast. by @maleadt in #81
• Add support for assertions. by @maleadt in #83
• Split tests for better parallelism. by @maleadt in #84
• Rework intrinsics by @maleadt in #85
• Sanitize kernel names. by @maleadt in #86
• Support broadcasting unsafe_trunc and trunc by @maleadt in #88
• Pass constants as scalars, infer as constants. by @maleadt in #87

New Contributors

• @maleadt made their first contribution in #8
• @vchuravy made their first contribution in #5
• @AntonOresten made their first contribution in #23
• @arhik made their first contribution in #38

Full Changelog: https://github.com/JuliaGPU/cuTile.jl/commits/v0.1.0