Review of #561. Guest time now has two entry points (psx_cycles.h): the
CPU charge psx_cpu_charge() (instruction base, load fudge/region/completion,
i-cache refills, BIOS HLE costs), which applies [timing] guest_cycle_scale,
and the unscaled device advance psx_advance_cycles() (DMA, idle skips,
mul/div and GTE deadline waits). The emitters' block-mode charges emit
psx_cpu_charge; overlay DLLs accumulate it raw and publish through the
cpu_charge callback (was advance_cycles, same slot) into the same host
function, so nothing is scaled at the DLL boundary. The batched
psx_cyc_charge path, the dirty-RAM interpreter, psx_cyc_readmem, i-cache
refills and bios_hle.c all charge through it. At scale 1 it is exactly
psx_advance_cycles.
Gate in psxrecomp: [timing] guest_cycle_scale_gate, RAM predicates judged
at every VBlank edge (deterministic, works online, no mod needed), plus
psx_mod_set_guest_cycle_scale_gate() in mod_plugins.h for custom logic or
to tie the scale to a mod (guest_cycle_scale_gated = true; every mod
session reset shuts it). The carried fraction and both gate states are a
boot_state section (BS_SEC_GCS, written only when the scale is not 1), so
savestates and the netplay rollback snapshot re-simulate the same cycles.
psx_guest_cycle_scale_gate_test pins tier identity, the gates and rollback
re-simulation against the real psx_cycles.c.
OpenBIOS regenerated (emitter stamp); retail SCPH1001 is not.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
No CPU is emulated: recompiled code charges the guest clock a fixed cost
per instruction. guest_cycle_scale = N charges 1/N of it (instruction,
load, mul/div, GTE latencies, overlay DLL batches) while VBlank, timers,
CD, SPU and DMA keep hardware time, so the game never runs out of frame
time. A title constant in game.toml (1 = faithful, 1..64), never a player
setting; guest_cycle_scale_gated applies it only while the title's mod
opens the gate (psx_guest_cycle_scale_gate_open). PSX_GUEST_CYCLE_SCALE
overrides for testing. A scale other than 1 is folded into the netplay
content fingerprint as a safety net.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The UI prepass re-read aperture node payloads through the RAM-folded key, so
THPS2's relocated prim arena discarded the HUD squash every frame.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit fd6267e952)
A full-screen rect backdrop that stops short of the reveal vetoes polygon overhang (Tomba village, beads-eio.4.28).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The remaining tunnel cracks came from vertex sources, not from missing
depth: 46 % of R4 lookups were marked ambiguous because the same 3D point
is projected several times per frame (corners shared by separate meshes,
other functions), so those vertices fell back to a triangle-mean guess.
- The same point projected again is one vertex (the smaller identity, so
the choice does not follow draw order).
- Up to two different points on one pixel are kept; gpu.c picks the one
projected by the same function as the triangle's other vertices, else
the one nearest their depth.
- Per-vertex motion is no longer rounded to 1/8 px: at high internal
scales the rounding opened hairlines along long T-junction seams.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Corner and tunnel runs showed light cracks along seams and near the view
edge in generated frames:
- weld: every placed vertex at one screen position of a view moves alike
(the same corner arrives from several projections);
- a vertex without a projection on an edge between placed vertices moves
as that point of the edge (T-junctions from the game's polygon splits);
- a vertex the in-between camera passes is projected at a clamped depth,
and such a game frame is not generated (no near-plane clipping: tunnel
ceilings, overhead bridges);
- debug: "clamped", "guessed" (CPU-built vertices placed by a triangle
mean over a >1 px motion spread).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pairing triangles of consecutive draw lists by key and draw order paired
the wrong ones whenever the game reordered, culled or switched LODs (R4:
a grandstand of one repeated texture turned into a staircase; on a
synthetic version 225 of 226 pairs were wrong). No primitive is paired by
draw order any more.
gte.cpp notes each RTPS/RTPT output (only while generation is on) with an
identity (issuing ra + model-space vertex), its camera-space position and
H, indexed by screen word; gpu.c records each polygon vertex's source
ahead of the triangle (RTH_FG_SRC, never drawn). Per view the camera
motion between the two frames is fitted from the paired camera-space
vertices (RANSAC + least squares); one real frame is redrawn from the
in-between camera: world vertices re-projected, vertices moving against
the world (cars) lerped in camera space, CPU-built vertices moved with
their triangle, HUD/2D unchanged. Moving forward the older frame is
redrawn; uncovered areas show the newer real frame. A verdict rejects
the generated frame (the real one shows) without projections, with few
pairs/inliers, or with a motion no camera makes. Real frames are
unchanged (gl_frame_gen_test).
PSX_PRESENT_SHOT_BURST=N captures the next N composed frames, real and
generated, for checking in-between frames against their neighbours.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A render thread a frame behind lowers the plan's ceiling by one (recovering
one per 2 s) instead of tripping the breaker; the guest waiting on the queue
trips it only when the wait was a stall (a quarter of the game frame). The
real-frame cost plans with CPU time; the GPU span is reported only. The
generated-frame blit copies colour only and marks the stencil stale.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Armed at init (session start, render-thread start, a new resolution) and on
a savestate load, game entry or window resize. While armed, an overrun jumps
straight to the highest level whose pure-area prediction is at or below
0.78, then is judged like any step; a fit after a step, an up step, a failed
judgement or 6 s disarm it and single steps fine-tune. A fixed cost only
makes the target dearer than predicted, so a jump never lands below the
level that fits.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Opt-in under the render thread. Composed frames go to a ring of offscreen
slots at the window's size (binds of framebuffer 0 on the composing context
land on the current slot); a present thread with a second, shared context on
the same window waits on the slot's fence, copies it 1:1 and swaps. The
compositor's wait moves off rendering; the composing side waits only while
every slot is queued or on screen. Falls back to direct swaps when the shared
context or sync objects are unavailable.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
[video] frame_generation (opt-in, with render_thread): the render thread keeps
each game frame's replayed draw list, matches the triangles of the last two
(frame_gen.c), and presents in-between frames at the display refresh by
redrawing the newer list into separate surfaces with matched vertices
interpolated. It never runs guest code. Real frames are composed at their
flip and presented after the generated ones, with the same pixels and
order. Generation uses only measured surplus: a plan from render-thread CPU,
generated-frame cost and swap time, a breaker after backpressure or late
guest frames, and a pause while dynamic resolution is over budget. The
render-thread core gains a timed tick hook. Frame-cost queries pause around
generation, so dynamic resolution measures real frames only.
Tests: frame_gen_test, render_thread_test (tick), gl_frame_gen_test (real
frames identical off/on; phase 0/1 equal the endpoints). Docs:
docs/FRAME_GENERATION.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hr (VRAM) surface stayed allocated at the ceiling. On GL-on-Metal every
render pass loads and stores the whole attachment, so its cost followed the
ceiling rather than the level: R4 2P VS at 5x ran 40-45 Hz for its first
seconds with a 10x ceiling and 60 Hz with surfaces allocated at 5x. Allocate
a new hr surface at each step and reseed it from the 1x image and the saved
rects; a failed allocation falls back to the in-place step down, or refuses
a step up.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A native-wide surface was allocated at the dynamic-resolution ceiling and
rendered into its lower-left corner. On GL-on-Metal every render pass loads
and stores the whole attachment, so the wide surfaces cost the ceiling's
bandwidth at every level. Allocate them at the level they render at and
reallocate at each step, rescaling the presented rows and margins from the
old surface. The limit check stays at the ceiling.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
--headless-opengl captured the presented image into the ring every frame
through gl_renderer_capture_display_hires, a sync point, so a headless
run (fp_identity) held the context for most of every frame. The capture
is now recorded at the same point of the frame (gl_renderer_ring_capture)
and done by the render thread; ring readers already sync first. A rect the
GL surface cannot serve is skipped there rather than resolved through a
CPU readback, and ensure_cpu refuses to run on the render thread at all:
its readback would land in the private upload source and mark the CPU
side current without gpu.c's array seeing the pixels.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One frame behind is the normal pipelined state when the render thread is
the slower side; skipping on it dropped nearly every present at 4K
widescreen. rt_frames_ahead() counts the complete frames queued after
the one being replayed; the backend skips only at two, where the
emulation thread is at the in-flight bound waiting on the render thread.
The unit test checks the count never exceeds the bound and that both
cases occur under a slow consumer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the render thread running, the backend's facade table records every
state change, primitive, fill, copy, CPU->VRAM upload (payload copied)
and native-wide call, and the present, into the command ring; the render
thread replays them through the unchanged glb_* functions. With it off
gl_backend_get returns the existing table and nothing new runs.
Sync points: every other GL entry point reached on the emulation thread
(readbacks, savestate restage, render passes, OpenXR, settings,
diagnostics) drains the ring and takes the context (GL_RT_SYNC); the
emulation thread then runs the rest of that frame directly, exactly as
without the thread, and hands the context back at the frame boundary when
the frame may run asynchronously again (no depth24 display, dual-raster,
frame interpolation, render pass, netplay or OpenXR).
Guest-visible VRAM stays gpu.c's array: while the render thread owns the
context, uploads stage from a private copy that receives each payload in
command order. Guest state the backend reads while drawing (per-primitive
widescreen tags, the native-wide fast-path latch) is captured at record
time; host overlays travel with the present; a present that a newer
recorded frame already supersedes is skipped. The emulation thread answers
projective/wide-present/draw-area queries from mirrors of the recorded
calls. The per-frame display-ring readback is queued (fbo_peek_deferred)
instead of being a sync point.
The GL fixtures link render_thread.c and stub the two facade hooks; the
scale fixture also gains the gpu_ws_background_requires_full_composite
stub it was already missing (it failed to link on master).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Backend-neutral core for the opt-in [video] render_thread: the emulation
thread appends records to a single-producer/single-consumer ring and the
render thread replays them in order with the graphics context current.
rt_frame_end bounds closed frames in flight (backpressure), the render
thread can tell a frame is stale, and rt_acquire is the sync point:
drain, then move the context to the emulation thread until rt_release.
No lock is held while a record executes.
render_thread_test drives it with a fake executor and context token:
ordering and payload integrity across ring wraps, oversize refusal, the
in-flight bound under a slow consumer, stale detection, randomized
acquire/release interleaving and stop while held (TSan clean).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A step controller (dynamic_resolution.c) lowers or raises the GL internal scale from frame-time telemetry between a floor and the configured ceiling; off unless a title or the player enables it. Tested by dynamic_resolution_test (synthetic load) and the real-GL gl_scale_invariance_test step modes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gl_renderer_set_dynamic_resolution(1) keeps the hr surface and the
native-wide surfaces allocated at the scale context init gave them (the
ceiling) and lets the internal scale step to any integer level from 1 to
it while the game runs. A lower level renders into the surfaces'
lower-left VRAM*S corner; every viewport, scissor, pack, copy and capture
already takes the scale at use, so only allocation and the present's
normalised UVs read the allocated size. Off (the default) nothing is
allocated differently and no step can run.
A step runs after a present, never inside a render pass:
- it takes the top-left sample of every native block of the old surface
(the sample the pack reads) into a 1x image and reseeds the new corner
from it, so the guest-visible VRAM is unchanged byte for byte;
- it rescales the displayed rect and the draw area from their old pixels,
so neither the frame on screen nor the one being drawn shows 1x blocks;
- it rescales the native-wide margins in place, in row bands ordered so
no source row is overwritten before it is read;
- it marks the mask stencil stale (rebuilt at once while mask checking is
on) and drops the render-pass backup and images.
Reallocating per step instead measured 30-106 ms at 7x-10x on an Apple
M4, mostly the first touch of the new surface. Steps are not offered in
the windowed high-resolution mode or with dual raster.
gl_scale_invariance_test gains "steps" runs: rendering after a step into
a ceiling-sized surface equals the fixed-scale run at that level; a chain
of steps (9x down to 1x and back, 5x) leaves the native VRAM unchanged at
every step, the two rects and the wide margins exact nearest rescales of
their old pixels, everything else the old top-left samples, and mask
checking right after a step correct; the window mode refuses steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Preserve negotiated content and dual-raster pass gates. Retain both fixture assertion sets and HD-off seams. Regenerate both BIOS with actual native emitter. Fix type-only rollback corrections and exclude extra-pad changes from digital-only promote gates.
Matthew Stanley