Skip to content

Releases

AetherCompiler v1.0.0 — MLIR/Triton SSA Graph Compiler, Polyhedral H100 Autotuner & C++20 Arena Planner

Choose a tag to compare

@nagarjunak-pixel nagarjunak-pixel released this 16 Nov 22:57

Initial v1.0.0 release of AetherCompiler (aether-compiler):

  • Typed SSA Tensor Graph IR & Multi-Pass Optimizer: AlgebraicSimplify+CSE -> VerticalEpilogueAndResidualFusion -> DeadCodeElimination (17 -> 9 IR nodes, 2.18x HBM traffic reduction, |Δ| = 2.38e-5, strict SSA dominance & shape verified).
  • Native C++20 Linear-Scan Static Arena & Tiled Fused GEMM: clang++ -O3 best-fit 256B-aligned liveness allocator compressing 2-layer Transformer activations from 432 MiB -> 120 MiB (3.60x, 0 live-buffer overlaps, 100% C++20-vs-Python offset parity, fused GEMM+Bias+SiLU |Δ| = 8.3e-7).
  • Polyhedral Hopper H100 SXM5 Tile & CuTe Bank-Conflict Autotuner: Eliminates 32-way -> 1-way shared-memory bank conflicts via bijective CuTe XOR swizzling and selects CTA_256x256x32_S4_XOR (411.65 TFLOP/s, 31.02x speedup over naive 32x32).
  • Live qwen3.5:4b Hardware Cost-Model CEGIS Gate: Lifts raw Pass@1 from 4/6 (66.7%) to 6/6 (100.0%, +33.3% lift, 0 OOM/misalign leaks).