pub struct PtxBackend { /* private fields */ }Expand description
PTX backend — Backend impl wrapping a MultiEmitterBackend so
the v0.1.0 scaffold drives through the same routing the future
rustc_codegen_nvvm + aprender-gpu quorum will use.
Implementations§
Source§impl PtxBackend
impl PtxBackend
pub fn new() -> Self
Sourcepub fn new_with_matmul_specialist() -> Self
pub fn new_with_matmul_specialist() -> Self
PMAT-280 — End-to-end validation constructor for Section 29’s multi-emitter routing.
Builds a PtxBackend whose MultiEmitterBackend carries the
ScaffoldPtxEmitter in the general slot AND a
[MatmulSpecialistEmitter] in the specialist slot under
QuorumPolicy::PreferSpecialist. The specialist matches only
modules whose name starts with matmul_ — the shape filter
real specialists like aprender-gpu would use to claim
GEMM-shaped kernels.
This isn’t registered in default_session() — production at
v0.1.0+ still uses PtxBackend::new. The constructor exists
so tests + future integrations can exercise the
MultiEmitterBackend::new_with_specialist path against real
production code (not just mock tests). It’s the smallest
concrete proof that the §29 routing layer is end-to-end
usable, ahead of the heavy rustc_codegen_nvvm / aprender-gpu
integrations that will eventually replace these placeholders.