Expand description
GpuBackend — wgpu compute-shader implementation of ArithmeticBackend.
wgpu is always compiled in (not feature-gated). At startup GpuBackend::try_init
probes for a compatible adapter; if none is found it returns Err and the
crate::backend::Executor transparently falls back to the CPU backend.
Each shader thread handles one (batch_item × channel) pair. The buffer
layout is identical to crate::batch::RnsBatch, so there is no reformatting
on upload beyond the u64 -> u32 narrowing (safe because all moduli < 2^16).
Structs§
- GpuBackend
- GPU backend holding a device, queue, and the pre-built compute pipelines.
Enums§
- GpuError
- Errors that can occur while bringing up the GPU backend.