How much of a kernel’s performance is decided by the MLIR pass pipeline before LLVM sees it? One linalg.matmul, 14 lowering pipelines, identical back-end flags, every result verified against a reference.
benchmarking llvm simd compilers code-generation performance-analysis vectorization llvm-clang loop-tiling linalg matmul mlir mlir-lang performance-analysis-and-simulation mlir-llvm linalg-cholesky-from-scratch reproducible-benchmarks transform-dialect affine-dialect
-
Updated
Oct 18, 2026 - Python