Expand description
Bicubic (Keys a=-0.5) NCHW f32 resize via Forge WGSL on the auxiliary GPU circuit.
Matches the CPU Keys cubic used by cv::sr::bicubic_u8 (per-channel NCHW).
Direct WGSL dispatch (not yet a graph OpNode) so product SR can use GPU
without waiting on IR lowerer expansion.
Device routing: vision SR runs on the auxiliary circuit (the iGPU when present) to keep the
discrete GPU free for the LLM. The device comes from
crate::gpu_context::device_registry::try_auxiliary_gpu, whose fallback chain is
auxiliary → primary → CPU: on a single-GPU box it transparently uses the primary device, and on
a GPU-less box it returns None, so this function reports BackendUnavailable and the caller
(super::dispatch) degrades to the CPU oracle. It never panics.
Constants§
Functions§
- bicubic2d_
cpu - CPU bicubic NCHW oracle (same Keys weights as WGSL).
- bicubic2d_
wgsl - WGSL bicubic upsample. params u32:
[c, h_in, w_in, h_out, w_out, 0, 0, 0]. - try_
resize_ bicubic_ shared_ gpu - Attempt bicubic NCHW resize on the auxiliary GPU circuit (iGPU when present; falls back
auxiliary → primary → CPU). Returns
Err(BackendUnavailable)on a GPU-less box so the caller degrades to the CPU oracle. Never panics.