AscendC backend

The AscendC backend compiles typed SSA with the local CANN CCEC compiler and launches the resulting library on the current torch_npu stream. Install a matching CANN and torch_npu environment and source the toolkit’s set_env.sh before using the backend. Compiler discovery uses NINETOOTHED_ASCENDC_COMPILER, PATH and the installed toolkit paths.

Select backend="ascendc" in ninetoothed.make. For the Ascend910B2C, platform="ascend-910b2c" records the concrete device contract. Execution uses NPU devices even with the legacy generic platform description. The only backend option is arch: native by default, or dav-c220 and supported Ascend 910B family names. NINETOOTHED_ASCENDC_ARCH can supply a concrete architecture for native selection. The resolved architecture and local compiler identity participate in the compilation cache key.

The baseline uses the shared SSA walker for indexed and masked memory access, broadcasting, loops, reductions and decomposed linear algebra. A single scalar executor visits the logical task domain, including tails and in-place writes. GM access uses byte-counted DMA through a reserved UB scratch slot, avoiding scalar DCache writeback conflicts between disjoint views on different streams. Transfers complete before the slot is reused; eligible tile arenas cannot overlap this scratch region. Eligible loop and reduction programs reuse the shared bounded tile planner. Its local storage is allocated from one UB arena, with typed lifetimes and simultaneous region yields. Pure elementwise values may be recomputed at their consumers to fit the resource budget. A nonoverlap guard protects the tile path; programs outside its legality conditions retain scalar access.

FP16 and BF16 operations preserve their declared result boundaries. Scalar reductions accumulate these types in FP32 and convert the result back to its declared type. Integer data and indices use native integer arithmetic. FP16 host scalar arguments use an explicit 16-bit storage ABI. Built artifacts support AOT construction and ninetoothed.compiler.load_built_artifact reload, including validation of this scalar ABI.

Atomic updates are rejected: serial execution within a kernel does not provide atomicity with other streams. Float64 and float8 device arithmetic are also unsupported. This baseline has no vector instruction selection, overlapped DMA schedule, Cube matrix primitives or autotuning. Large fused kernels can therefore be slow; target performance optimizations belong to subsequent changes.