Expand description
Single-op CPU reference evaluation via rlx-cpu’s thunk executor (the same
kernels the CPU backend uses, so results are bit-for-bit the reference).
Two callers:
- On a host with no Level Zero device (the macOS dev box / CI),
backend.rswalks the whole legalized graph through this evaluator — every op is served by the CPU reference, so the backend is fully correct without Intel HW. - On Intel hardware, ops with no native SPIR-V kernel yet route here (read from the USM arena, eval, write back) — exactly rlx-vulkan’s host-fallback.
Enums§
- HostBuf
- One host-eval input: f32 activations, or raw bytes for a packed quant weight
(U8/I8 operands such as the GGUF weight of
Op::DequantMatMul).
Functions§
- eval
- Run a single op on the CPU reference and return its f32 output.
inputs[i]is(declared_shape, buffer).