Crate rightkit_gpu 

Shared wgpu device context for Right Suite apps (#1159): per-OS backend policy, a crash sentinel around device creation, adapter pick and limits, an exit gate for in-flight GPU work, and a device-memory budget that adapts to out-of-memory. Native only; the apps’ kernels stay in the apps.

• backend: which backends wgpu may load (Windows DX12 alone, Metal on Apple, Vulkan elsewhere, overridable by the app’s own environment variable) and the crash sentinel (backend::set_init_marker, backend::take_init_marker) that lets the next launch know a driver killed the last one during device creation.
• GpuContext::create: adapter by name or default, software adapters refused, limits, error and device-loss callbacks.
• exit: every entry into the device goes through exit::enter; the host ends the process with exit::begin_shutdown and exit::wait_idle.
• MemoryBudget: renders ask for room, wait fairly, halve the budget on out-of-memory and grow it back with each success.

wgpu is pinned at 30 here (the version Ember builds on, and rightkit-compositor’s). CutRight moves from 29.0.1 to 30 to use this crate; the crate does not carry both.

Adapted from Ember crates/gpu (backend.rs, exit.rs, budget.rs, the generic part of ctx.rs); see NOTICE.

Re-exports

pub use wgpu;

Modules

backend

Which wgpu backends an app lets wgpu use, for its compute device and for its window, and the crash sentinel around device creation (issue #136). Adapted from Ember crates/gpu.

exit

Leaving: no GPU work may be in flight when the process ends (issue #620). Adapted from Ember crates/gpu.

Structs

ContextOptions

How the app wants the device made.

GpuContext

A wgpu device and queue on a chosen adapter.

LimitsCap

An upper bound on the device’s limits.

MemoryBudget

Permit

Room on the device for one render, released when dropped.

Functions

adapter_names

The names of the usable adapters for backends. Loads the drivers; the device is not created.

footprint

Device memory a render is expected to hold: the source, two working copies and the output, as float RGBA, at the source’s size.

initial

Where the budget starts: 1.5 × the largest buffer the device allows (about the device’s memory on small cards), at least 1 GiB.

parse_limits_cap

Parses a limits cap the way Ember’s LIGHTCRAFT_GPU_LIMITS does: webgpu / downlevel / default = the WebGPU defaults (128 MiB storage bindings, 256 MiB buffers), <n> (or <n>MiB, <n>MB) = n MiB storage bindings and 2n MiB buffers. Empty or unparsable: None.

pick_adapter

The index of the adapter called wanted in names (the first with that name), or None when there is none, so the default choice is used.