GDVP Oversampling
Oversamplers — GDVP_NODE_UPSAMPLER / GDVP_NODE_DOWNSAMPLER ✅ (auto-injected)
← Node Reference · Manual index
A matched pair of polyphase half-band FIR converters that move signal between the base rate and
2× oversampled rate without aliasing. You never place these manually — the
DAG compiler inserts them automatically around any node that runs
oversampled, so the rate boundary is band-limited rather than naive. Source:
gdvp_dsp_oversampling.c, payloads gdvp_node_upsampler_t / gdvp_node_downsampler_t.
What they do
- Upsampler (1×→2×): interpolates a band-limited sample between each input sample, doubling the rate ahead of an oversampled node. Carries a 5-sample tail across blocks.
- Downsampler (2×→1×): low-pass filters and decimates the 2× signal back to base rate after the oversampled node. Carries a 10-sample tail.
Both use the same 11-tap symmetric half-band FIR with its corner at fs/4 (Nyquist/2 of the
base rate), stored as three Q15 coefficients (the other taps are zero or the unity centre tap).
Filter gain bookkeeping is handled by the shift amounts (>>14 up, >>15 down) so levels are
preserved across the conversion.
Parameters
None that you set. These nodes have no operator parameters and no parameter-updater entry —
they are pure structural plumbing inserted by the compiler. Their only state is the FIR tail and
an ownership link back to the node they wrap (parent_node_id), used so re-compilation stays
idempotent.
Why a node would be oversampled
Oversampling is requested via the node header flag GDVP_NODE_FLAG_OVERSAMPLED (which makes that
node process 128-sample blocks instead of 64). It's used where a nonlinearity would otherwise
alias — bright/aggressive filter settings, saturating VCA drive, FM. When a
node is flagged oversampled but its neighbours are not, the compiler brackets it with an
upsampler/downsampler pair so the rate change itself doesn't introduce aliasing.
This is also where the filter's OSAC correction matters: an oversampled filter uses a different, exact coefficient LUT for the 128-sample rate rather than approximating.