Audio, video, pacing, and hitch behavior

Documentation index

Audio generation

MesenCE emulates the NES pulse, triangle, noise, DMC, and expansion channels at CPU-cycle-accurate event boundaries. Channel amplitude changes feed band-limited resamplers and the NES channels are combined with nonlinear mixing behavior.

The Libretro target outputs 48 kHz audio to RetroArch.

Libretro audio adapter

The adapter converts mono output to interleaved stereo when required and keeps a bounded pending queue. If the frontend accepts only part of a batch, the remaining samples are preserved instead of dropped. Excess fast-forward backlog is trimmed before normal playback resumes.

Current counters track short writes, zero writes, and dropped queued frames.

Video delivery

MesenCE builds a native or HD RenderedFrame and the asynchronous video decoder produces the frontend image. The Libretro layer retains the latest completed decoded frame and supports duplicate-frame submission when the frontend permits it.

Current synchronization policy

RetroArch owns pacing, throttling, fast-forward, slow motion, and host audio synchronization. The core should generate one complete emulated frame per retro_run() and must not run an independent desktop frame limiter.

Known hitch path

A rare expensive HD frame can delay normal-speed return from retro_run(). That can make a visual frame drop and an audio hitch occur together even when the APU itself is correct.

Planned alpha work:

  1. Add measured decoder/renderer/audio queue diagnostics.
  2. Test a smooth delivery mode that duplicates a late video frame instead of blocking the next emulation/audio frame.
  3. Replace abrupt queue trimming with a short splice from the last submitted sample to the new queue head.
  4. Compare the current 96 kHz internal path and final downsampling against a direct 48 kHz band-limited path before changing audio quality.

Audio changes must not introduce pitch drift, duplicate synchronization loops, or save-state differences.