Expand description
§ironaccelerator-qnn
Qualcomm AI Engine backend, dispatching through the QNN SDK:
- HTP (Hexagon Tensor Processor) — primary NPU path; INT8 / FP16 with HMX matrix-engine acceleration on Hexagon v73+.
- HVX — fallback Hexagon vector path for ops not yet supported by HMX.
- Adreno GPU — alternative target for FP16 graphs.
Scope is the same as every other backend here: probe the runtime, enumerate targets, report capability bits. QNN graphs are built ahead of time by the SDK, so graph construction and quantisation calibration belong to the consumer, not to this crate.
§Status
v0.1 scaffold. Library probe via libloading; FFI bindings for QnnApi.h,
QnnContext.h, QnnGraph.h, QnnHtp.h are tracked next.
Re-exports§
pub use backend::QnnBackend;pub use backend::QNN_BACKEND;pub use iron_qnn_sys as sys;
Modules§
- backend
- QNN
Backendimpl. Enumerate-by-probing: try HTP / GPU / CPU / DSP and emit aDeviceDescriptorfor each target whose library loads. - drv
- Safe QNN driver layer. Opens a backend library, creates a context, and
exposes graph construction + execute in an API that mirrors the CUDA/ROCm
drvmodules.
Functions§
- register
- Registers the QNN backend with
reg.