Expand description
Evidence-based wgpu tensor compute site discovery.
This crate discovers physical adapters deterministically, requests a bounded device profile, records granted limits and features, and exports SIM tensor sites only for adapters that pass transfer, mapping, and allocation probes.
Structs§
- Allocation
Attempt - One bounded allocation attempt.
- Compute
Wgpu Lib - Loadable library that registers only successful wgpu adapter sites.
- Probe
Evidence - Probe evidence required before a site is exported.
- Probe
Policy - Bounded probe policy.
- Requested
Wgpu Profile - Requested limits and optional features passed to
wgpu. - Transfer
Evidence - Transfer and mapping probe evidence.
- Wgpu
Adapter Evidence - Adapter identity and requested/granted capability evidence.
- Wgpu
Adapter Probe - One successful adapter probe.
- Wgpu
Allocation Id - Opaque resident allocation id inside a planned wgpu arena.
- Wgpu
Arena Allocation - Allocation record returned by the arena.
- Wgpu
Arena Snapshot - Snapshot of resident arena pressure.
- Wgpu
Capability Evidence - Feature evidence recorded from a granted device.
- Wgpu
Discovery - Complete discovery result.
- Wgpu
Discovery Error - Discovery failure for infrastructure-level probe setup.
- Wgpu
Limit Evidence - Limits recorded from either the requested or granted device contract.
- Wgpu
Materialization Cache - Caches exactly one synchronized host materialization result.
- Wgpu
Pipeline Cache - Small FIFO cache for validated pipelines.
- Wgpu
Pipeline Cache Snapshot - Snapshot of cache pressure and reuse.
- Wgpu
Pipeline Key - Cache key for a validated portable pipeline.
- Wgpu
Pipeline Record - Validated pipeline evidence retained in the cache.
- Wgpu
Queue Limits - Queue bounds for resident tensor submissions.
- Wgpu
Queue Snapshot - Snapshot of queued wgpu tensor submissions.
- Wgpu
Resident Arena - Bounded resident arena with oldest-first eviction.
- Wgpu
Resident Segment - One contiguous resident buffer segment.
- Wgpu
Resident Storage - Resident storage that records a validated wgpu pipeline result.
- Wgpu
Segment Plan - Checked segment plan for a tensor payload.
- Wgpu
Submission Queue - Deterministic bounded queue used before command submission.
- Wgpu
Tensor Executor - Tensor executor descriptor backed by a successful wgpu probe.
- Wgpu
Tile Profile - Portable tile profile selected from granted adapter limits.
- Wgpu
Transfer Plan - Planned transfer spans for a segment plan.
- Wgpu
Transfer Span - One host-to-device or device-to-host copy span.
Enums§
- Wgpu
KernelD Type - Dtype strategy selected for a portable kernel.
- Wgpu
Kernel Op - Portable operation implemented by a wgpu tensor kernel.
Statics§
- RECIPES
- Cookbook recipes for this lib, embedded at build time.
Functions§
- compute_
wgpu_ capability - Capability required before a wgpu hardware tensor site can be realized.
- compute_
wgpu_ lib_ symbol - Stable symbol for the wgpu runtime library.
- compute_
wgpu_ site_ symbol - Site symbol exported for a successful wgpu adapter.
- discover_
wgpu_ adapters - Enumerates adapters and returns only probe-backed site candidates.
- wgpu_
executor_ symbol - Stable symbol for a wgpu tensor executor.