R009 F0/N second pair: Resize in the shortcut

Device: SM8550 / Hexagon V73. QAIRT 2.46.0.260424. Graph built through the QNN C API from PowerShell on device; nothing compiled.

Result

ValidateRc=0 FinalizeRc=0 ExecuteRc=0, no non-zero add codes, context 3,338,240 bytes. Oracle full scale 29.8425, so the residual is 0.09% of scale; max relative error on values above 1.0 is 0.002442. 45 of 66,560 values exceed the stage's absolute 0.02 tolerance, split 22/23 even/odd in time and with nothing at t=0 or t=129 — no parity bias, no edge concentration.

Cause

Upstream kokoro (dfb907a, kokoro/istftnet.py) builds AdainResBlk1d with two upsample paths:

• residual: self.pool = weight_norm(nn.ConvTranspose1d(dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1)) — depthwise ConvTranspose
• shortcut: self.upsample = UpSample1d(upsample) — F.interpolate, lowering to Resize

The earlier polyphase work replaced the depthwise ConvTranspose and left Resize in the shortcut. Both are miscomputed on HTP V73, so out = residual + shortcut was half wrong. Op census of the two prior modules:

• Kokoro.F0NSecondPairR009.psm1: TransposeConv2d, Resize
• Kokoro.F0NSecondPairPolyphaseR009.psm1: Resize

Change

Nearest-neighbour x2 along the time axis, built from ops that are exact:

[void](& $reshape "$branch.shortcut.lift"      "$branch.ShortcutFMajor65"   "$branch.ShortcutLifted"     @(1,65,1,512))
[void](& $concat  "$branch.shortcut.duplicate" @("$branch.ShortcutLifted","$branch.ShortcutLifted") "$branch.ShortcutDuplicated" @(1,65,2,512) 2)
[void](& $reshape "$branch.shortcut.expand"    "$branch.ShortcutDuplicated" "$branch.ShortcutFourD"      @(1,1,130,512))

Reshape to [1,65,2,512] then flatten row-major gives out[2t] = out[2t+1] = x[t], which is exactly nearest-neighbour x2.

Open

The stage still reports Passed=False because its gate is an absolute 0.02 on a tensor spanning +/-29.8. That gate was calibrated for a 58-op stage; it should be relative and sized to the fp16 floor (see fp16-null-result-20260922.md). The gate has not been changed.