| name | gpu |
| description | Plan abi GPU/backend work — Metal on macOS, CPU SIMD fallback, and the disclosed honest-stub backends (accelerator, shaders, mlir, mobile). Use when asked about GPU/backends, why accelerated=false, or when planning backend work. Routes to backend-diagnostics and abi-superpower-gpu. Never claims native CUDA/ANE/Metal-kernel execution — those are disclosed non-goals. |
gpu
Entry point for abi's GPU/backend surface (src/features/gpu/ + the four
honest-stub feature modules). Routes to specialists:
| You want to… | Use |
|---|
| Report GPU/accelerator/shader/MLIR status + compute matrix | backend-diagnostics |
| Deep-dive the GPU/Metal superpower | abi-superpower-gpu |
Honest status (trust the source flags over any prose)
- Metal on macOS is real: linked at build time;
accelerated=false is the
normal state until g_metal_context.init() succeeds at runtime; mid-run
failure degrades to CPU. No -Dgpu-backend option exists.
- Honest stubs (
available=false / native_dispatch=false in each
src/features/*/mod.zig): accelerator (selection report + CPU SIMD
fallback only), shaders (validate + checksum, no compiler), mlir
(textual lower only, no LLVM toolchain), mobile (profile report only, no
runtime).
- ANE execution is a disclosed non-goal (100% Zig constraint; requires
CoreML/ObjC). Native CUDA/Vulkan/Metal-kernel execution is not linked.
Hard rule
Do not claim real shader compilation, MLIR/LLVM lowering, or native
accelerator dispatch — per docs/contracts/external-claims-audit.mdx.
