Small MLIR compiler: a dsp tensor dialect lowered to linalg, then tiled and vectorized by Transform-dialect schedules derived from a target model (NEON, AVX2). Includes int8 qmatmul, SIMD Mojo kernels and a C++ oracle, all tested bit-exact.
mojo compiler cpp dsl dsp neon llvm tiling image-processing simd compilers avx2 tensor vectorization quantization loop-optimization linalg mlir hardware-optimization transform-dialect
-
Updated
Oct 8, 2026 - C++