Replaying the measured Cybermorph R3x trace, 25% held vsync on through an
intermittent deficit (11% fewer new images than v3.9.0); 15% matches v3.9.0's
image count with 13 instead of 153 toggles per 60 s. A 250 ms stall and a
3-frame 40 ms hitch on a core with time to spare change nothing.
Refs #952
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A transient (Cybermorph's 14 ms phase, a state load, a loaded machine) that
releases the swap interval must not cost the vsync for the session: a core
back at 11-12.5 ms holds a vblank-locked loop. Adds the heavy-phase and
hitch cases.
Refs #952
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Fake-clock cases for #952: 8 ms core on 60/59.94/120/119.88 Hz with jitter
(zero toggles, no batching habit), vsync-off throughput and sleeping,
sustained deficit with vsync on, edge and alternating-cost traces measured
against the v3.9.0 loop. These fail on the #956 pacer (14+ failures).
Refs #952
Co-Authored-By: Claude Opus 5.5 <[email protected]>
When the core cannot hold full speed (RISC overclock, a heavy title) the loop
zeroed the backlog after each catch-up batch and then slept most of a period
(vsync off) or blocked in SDL_GL_SwapWindow until the next vblank (vsync on),
on the thread that runs the emulation. Doom at RISC 300% ran 54.6 fps headless
but 45 fps in the window.
Pacing moves into pacing.h (Pacer): the time spent running frames is charged to
the backlog so 'behind' is exact, the sleep is skipped while behind, and the
swap interval is released while behind and restored after ~0.5 s of keeping up.
tests/pacing_test.cpp drives it with a fake clock.
Refs #952
Co-Authored-By: Claude Opus 5.5 <[email protected]>
