One recorded graph node, ready for WgpuComputeContext::submit_graph to play into a
shared command encoder: the compiled pipeline, its bind group, the workgroup count, and
the optional GPU→GPU hand-off copy (src in the read_write slab → dst in the read
slab) emitted after the node’s compute pass. The pipeline and bind group are owned so the
executor can build them all up front and submit the whole graph in one go.
A topology-aware ring buffer designed to eliminate heap allocations
during hot-path kernel dispatch. Maintains strict bounds invariant
so the write_head never laps the read_head.
How a BufferView is bound in a dispatch. Determines which physical slab
backs it: wgpu forbids a single buffer being bound as both read-only and
read-write storage in one dispatch, so read-write outputs live in a separate
buffer from read-only/uniform inputs.
The hardware surface the differential oracle drives. Implemented by both the
wgpu (super::WgpuComputeContext) and the CUDA ([super::CudaComputeContext])
backends so each cross-backend kernel evaluator is written once, generically.
A strictly stateless execution bridge designed to execute WGSL Forge IR operations.
It operates exclusively on pre-allocated, heap-free BufferView slices
provisioned by a higher-level QualiaSlabAllocator.