Expand description
STELLAR §A AH-track H0 — host topology + capability sensor (decision D23/D24).
The engine must sense then route instead of grabbing one adapter. Today gpu_context:: shared_gpu() requests a single PowerPreference::HighPerformance device — on a discrete +
integrated box that silently takes only the discrete GPU and ignores the integrated GPU and all
of system RAM. This module enumerates every adapter, classifies the memory topology
(discrete vs unified), reads host RAM, and computes the bounded OS floor (D24). The residency
planner (H2) and heterogeneous/cluster dispatch (H3/H5) consume this; it makes no routing
decision itself and changes no existing behaviour.
Native only — enumerate_adapters is not available on the wasm/WebGPU path.
Structs§
- Adapter
Desc - One enumerated adapter.
- Host
Topology - The sensed host: adapters, topology, host memory, and the bounded OS floor (D24).
Enums§
- Adapter
Class - Coarse adapter class (maps
wgpu::DeviceType). - Host
Memory Topology - Whole-host memory topology: is there a dedicated VRAM pool, or one shared pool?
Functions§
- probe_
host_ topology - Probe the host: enumerate all adapters (deduped across backends), classify, size the floor.