Book contents +

XI · Hexagon NPU (welvet.qnn) / CHAPTER 74

QNN: opt-in Hexagon HTP (welvet.qnn)

Samuel Watson · Welvet feature bookOriginal source ↗

github.com/openfluke/welvet/nativeqnn✅ windows/arm64 + cgo


Why it exists

Qualcomm Hexagon HTP is a sibling engine (welvet.qnn 1.0.0), not a silent fallback inside Go. Opt-in BackendQNN + explicit SyncToQNN keep CPU/SIMD/WebGPU as defaults. Caps reject unsupported layer/dtype mounts.

What it is

nativeqnn.Available(), Version(), SupportedLayers(), Dense SyncToQNN → PrepareQNN → Forward with BackendQNN. FormatNone host weights; HTP may run f16/i8/… inside the DLL.

Needs Windows ARM64, cgo, and shipped libwelvet_qnn.dll (plus Qualcomm QAIRT runtime on PATH). Other OS/arch → Available()==false and examples print SKIP.

Layer tour: §80. Large-shape matrix + % wins: §81. NativeCPU twin: §76.

Go host (FormatNone f32)     SyncToQNN / mirror      welvet.qnn DLL
  Dense / Softmax / …  ──►  caps-check + Capture  ──►  HTP graphExecute
  BackendQNN Forward   ◄──  unpack y to Tensor   ◄──  dtype inside NPU
Opt-in path: never Prefer as default.

Go example

examples/79-qnn/main.go

Run:cd welvet/examples/79-qnn && source ../env.sh && go run .
package main

import (
	"fmt"

	"github.com/openfluke/welvet/core"
	"github.com/openfluke/welvet/layers/dense"
	"github.com/openfluke/welvet/nativeqnn"
	"github.com/openfluke/welvet/quant"
)

func main() {
	fmt.Println("version", nativeqnn.Version())
	fmt.Println("available", nativeqnn.Available())
	fmt.Println("default", core.DefaultBackend())
	if !nativeqnn.Available() {
		fmt.Println("SKIP: need windows/arm64 + libwelvet_qnn (+ QAIRT)")
		return
	}
	fmt.Println("layers", nativeqnn.SupportedLayers())

	init := make([]float32, 4*4)
	for i := range init {
		init[i] = float32(i+1) * 0.05
	}
	l, err := dense.NewConfigured(4, 4, core.ActivationLinear, core.DTypeFloat32, quant.FormatNone, init)
	must(err)
	must(l.SyncToQNN())
	must(l.PrepareQNN(2))
	l.Exec.Backend = core.BackendQNN

	x := core.NewTensor[float32](2, 4)
	for i := range x.Data {
		x.Data[i] = float32(i+1) * 0.1
	}
	_, post, err := dense.Forward(l, x)
	must(err)
	fmt.Println("fwd", len(post.Data), "backend", l.Exec.Backend)
	fmt.Println("qnn done")
}

func must(err error) {
	if err != nil {
		panic(err)
	}
}

Output

exit 0 · last run via go run .

version 
available false
default cpu-tiled
SKIP: need windows/arm64 + libwelvet_qnn (+ QAIRT)

Validate (harness)

cd welvet/examples/79-qnn && source ../env.sh && go run .