BangC backend

The BangC backend compiles typed SSA to Cambricon BANG C and executes it through the local CNCC/CNRT SDK. Install the matching torch_mlu package and make cncc available through PATH, NEUWARE_HOME or NINETOOTHED_BANGC_COMPILER.

Select backend="bangc" in ninetoothed.make. The optional platform="cambricon-mlu590" selects the concrete MLU590 platform contract. Use this explicit platform when exporting precise device metadata; the legacy generic profile description is retained, while BangC execution uses MLU. The arch backend option defaults to native; explicit values are compute_50 or compute_30. The compiler and resolved architecture are recorded in the compilation cache identity.

This baseline follows the shared SSA access maps, masks, loops, reductions and decomposed linear algebra. Bounded temporary tensor values in eligible SSA loops and programs with multiple reductions are stored locally with typed storage and region-yield lifetimes; this avoids unbounded producer recomputation in attention and entropy programs. The storage plan uses an explicit resource budget and a nonoverlap guard, retaining the original access path when it cannot prove legality. One task otherwise walks a scalar output chunk. Large logical task domains are split into bounded CNRT launches on the current MLU stream. Atomic updates use a single task over the complete domain, preserving updates without relying on unsupported hardware scalar atomics.

Integer data use paired 32-bit words for exact signed and unsigned 64-bit arithmetic; address and shape calculations retain native index arithmetic. Low-precision arithmetic preserves explicit FP16/BF16 rounding. Float8 is not supported by this toolchain mapping.

Native built artifacts support AOT construction and ninetoothed.compiler.load_built_artifact reload. The shared native runtime performs device and argument validation, binds the current MLU queue and pins the compiled library identity.

The baseline intentionally has no vectorized NRAM dataflow, DMA schedules, batched row execution or native matrix primitives. Those optimizations can be added independently; large fused kernels may be slow on the scalar path.