QNN investigation (historical)

Kokoro-Hexagon 0a03be39Updated 2026-09-27

This entire document is a historical QNN/oracle investigation, not the product architecture or an implementation recipe. Do not copy its graph rewrites, prepared inputs, runtime, or APK requirements into the PowerShell product. The current release boundary and remaining work are in ROADMAP.md. Its numerical and AdaIN findings remain useful comparison evidence; its QNN passes, libraries, and contexts are not production dependencies.

All findings below were verified on the S23 (SM8550, Hexagon V73) with QAIRT 2.46.0.260424, by per-layer probes compared against PyTorch on the device.

Static capacity with masked normalization

Kokoro's AdaIN1d uses InstanceNorm1d, which normalizes over the whole time axis. Sliding windows therefore change the statistics: 64-frame windows with a 12-frame overlap (the Piper/Melo pattern) reached only 5 dB SNR. Instead, one fixed-capacity window per phrase is used, with statistics computed over valid frames only (a length mask). Error against the full-length model: 24.2 dB SNR, 0.62 dB log-mel. Phrases are split at natural boundaries by the front end.

Harmonic source on the host (for now)

SineGen uses rand/randn and a whole-sequence phase cumsum. It is computed on the host and passed to the generator as an input, so the HTP graph has no random ops.

QNN passes

Pass	Reason
`Pow(x,2)` → `Mul(x,x)`	`Pow(x,2)` is wrong on HTP V73 (variance −10 dB; `Mul` 61 dB)
Depthwise `ConvTranspose1d` → polyphase; `Resize` → expand/reshape	the upsampling `decode` block broke on HTP (3.7 dB → 58.5 dB after rewrite)
fp16-safe statistics (scale by per-channel max before squaring)	fp16 overflow of the variance sum on HTP
Time lengths padded to a multiple of 8 (masked)	length 19,201 fails finalize; 19,208 compiles
Stage-resolution masks as graph inputs	removes `Resize` fallbacks from traced masks
Per-voice gamma/beta table (`gb`) instead of `fc(style)`	removes 48 `Gemm` per call; exact
ORT basic optimization	folds shape plumbing; 3,098 → 1,674 nodes, bit-exact

Runtime

Contexts are loaded from a memory-mapped file with contextCreateFromBinary and graphRetrieve (src/runspace/Qnn.Context.psm1). Tensor IDs and shapes come from the binary itself (tools/Read-QnnContextInfo.ps1).
The HTP performance vote (QnnHtpPerfInfrastructure, DCVS v3 burst) cut the generator from 4.69 s to 1.90 s. 8 MB VTCM cut it further to 1.57 s.
Host overhead per call is about 1.5 ms; HTP reports no wait time, so the remaining time is inside the graph.

Android host used in this historical experiment

The diagnostic APK declared <uses-native-library android:name="libcdsprpc.so">.
QNN libraries were preloaded by full path from app storage, with ADSP_LIBRARY_PATH set to the library directory plus the platform defaults (/vendor/lib/rfsa/adsp;/vendor/dsp/cdsp;/dsp).

Open

w8a16 (int8 weights, int16 activations) does not finalize yet, including a conv-only variant under test.
16 generator convs are dilated (3 and 5). Rewriting them as dilation-1 convs over interleaved phases is exact and keeps per-channel int8 legal on HTP.
Per-op profiling is refused (QNN_GRAPH_ERROR_SET_PROFILE) for these contexts; sub-graph timing is used instead.
Residual dropout is not a speed knob for this model: skipping one of 18 generator resblock branches costs about 10 dB log-mel.