v0.3.0

cuTile v0.3.0

Diff since v0.2.2

Breaking changes: Replace ct.launch with @cuda backend=cuTile

Merged pull requests:

• Include (c - a*b) pattern in FMA rewrite pipeline. (#190) (@maleadt)
• Support for array slicing (#191) (@maleadt)
• Add generic dataflow framework; port constant & alias analyses (#192) (@maleadt)
• Add kernel and host RNG (#193) (@maleadt)
• Stop special-casing TileArray in codegen; add permutedims/transpose/reshape (#194) (@maleadt)
• Document intrinsics. (#198) (@maleadt)
• Add transform-side control-flow helpers. (#199) (@maleadt)
• Add KernelState plumbing for per-launch ambient state. (#200) (@maleadt)
• Various small fixes (#201) (@maleadt)
• Test BFloat16 broadcast subtraction. (#202) (@maleadt)
• Minor fixes given the Tile IR spec (#203) (@maleadt)
• Improve Random.jl coverage: randn and randexp (#204) (@maleadt)
• Extend assumption analysis: divisibility, bounds, no-wrap (#205) (@maleadt)
• Add lightweight CSE on StructuredIRCode. (#207) (@maleadt)
• Update benchmarks (#208) (@maleadt)
• Benchmark harness improvements (#210) (@maleadt)
• Recompute IR flags from efunc when the rewriter changes opcodes. (#211) (@maleadt)
• Fold contiguous-axis stride in gather/scatter offset chains (#212) (@maleadt)
• Add rewrite rule to drop contiguous-axis stride in scatter/gather offsets + unified AssumeOp injection (#213) (@maleadt)
• Integrate with CUDA.jl + reduce launch overhead (#214) (@maleadt)
• Fix UndefVarError in README Quick Start (#215) (@AntonOresten)
• TTFX improvements (#216) (@maleadt)
• Backport fixes from cuTile Python (#218) (@maleadt)
• Suppress divby assume on tile-of-pointers from offset. (#219) (@maleadt)
• Run normalization rewrites to fixpoint before FMA fusion. (#220) (@maleadt)
• Update benchmark timings. (#221) (@maleadt)

Closed issues:

• Random Number Generation (#189)
• Consider CSE/LICM for TileView construction (#195)
• Make assumption emission into a pass (#196)
• Emit DivBy assumptions (#197)