The batched scaled charge fell through to the scale-1 batch path, which
also defers while g_psx_cyc_bb_defer is nonzero and routes through the VLC
local sink. Both are host call-stack state (a longjmp escape leaks or
resets the depth), so which charge published the batch -- and so which
clock each device saw -- differed between hosts. R4 2P rollback Mac<->PC
with modern controls split by a few cycles in live play (clk/tim fork at
~3296 with cpu/ram equal), then baseline mismatches and a realign storm
dropped the match; with the batch off the same pair matched.
The scaled path now batches on the guest clock only (device deadline or
the 64-cycle soft cap). New check: the service log is identical with
bb_defer 0 or 338 and with a local sink installed (fails before). Re-stamp
the OpenBIOS emitter sha (psx_cyc.h is in the codegen hash).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With [timing] guest_cycle_scale active, every one-cycle psx_cyc_charge went
straight to psx_cpu_charge and psx_advance_cycles, bypassing the batch the
scale-1 path uses. [timing] guest_cycle_scale_batch (PSX_GCS_BATCH=0/1)
scales the charge in place, with the same fractional carry and order as
psx_cpu_charge, then batches it like the scale-1 path. It flushes at the same
observation points. Default off.
R4 exactness, frame_fingerprint off vs on, 9000 frames each: headless boot
through intro, title and attract; a 1P Grand Prix race; and a 2P VS race.
cyc, mmio, mc and qc are identical on every frame. sp, wr, pc and ws are
identical at the end. The only differences are single-frame VBlank straddles
of wc/sc (never two frames in a row). 0 dispatch misses.
Speed, release builds, races at 60 Hz:
- M4 (macOS sample): psx_cpu_charge 4.3 % -> 0 % of emulation-thread samples;
busy 46.0 -> 44.4 % (2 pairs).
- 9800X3D: no measurable change (30.0 vs 30.5 %; one core 44.6 vs 44.2 %).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Smooth motion lost most of its generated frames with bloom on (R4 race,
120 Hz: ~30% of bloom-off). Measured and fixed in order of cost:
- Depth-stencil blits (CPU milliseconds each on macOS) are gone:
generation surfaces no longer copy the marks and are not marked while
drawn (generated frames reuse the newer real frame's mask); the
native-wide centre copy is colour only again and the mask takes the
centre's marks from the hr surface; the mask itself is resolved in the
presented surface's own pixels through an R8 target that shares the
surface's depth-stencil renderbuffer, then mapped to the window.
- Marking needs one pass per batch, not two: 3D batches only set, 2D
opaque ones only clear, semi-transparent ones only set; only batches of
unknown make (queued replays, lines) still do both.
- No full-size copy of the image: `source_down` (a blit straight to the
pass's size) feeds the bright pass, and the composite screen-blends onto
the window (`blend screen`; unlit pixels are discarded, so they keep their
exact value). The debug split/bloom-only views go with the copy.
M4, R4 race, render thread + reprojection, PSX_FRAME_GEN_FORCE, a host
loaded by other work (absolute rates low): bloom pass on real / generated
frames 1.2 / 0.7 ms at 1080p, 1.3 / 0.9 ms at 1440p, 2.6 / 1.6 ms at 4K
(glFinish-bracketed); generated frames with bloom 79-80% of bloom-off at
all three (was ~30%). A 1/8-resolution blur chain was tried and kept no
more frames, so the 1/4 chain (the look) stays.
Tests: parser (blend, source_down, bloom preset reads no full copy), GL
fixture (source_down + blend screen chain exact; HUD and mask checks
unchanged), FXAA unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review of #584: the default must not make guest-visible VRAM depend on the
internal scale. Mode 1 ("on", default; "scaled" kept as an alias) now
puts the PS1 pattern on the native VRAM grid; mode 2 ("fine") is the
elective per-internal-pixel pattern. gl_dithering_test checks the default
mode's native VRAM is identical at 4x and 1x (equal-colour Gouraud now
averages 15.50 in the default at every scale). Docs keep the
semi-transparent approximation explicit. OpenBIOS stamp regenerated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The OpenGL renderer has always drawn shaded and texture-modulated pixels
at 8 bits per channel with no dither pattern (true colour). That stays the
default. 'on' applies the PS1 4x4 offset matrix and 5-bit truncation per
internal-resolution pixel; 'scaled' puts the pattern on the native grid,
so it reads like the console at any scale and the native VRAM is the same
at every internal resolution.
GP0(E1h) bit 9 reaches the backend through a new optional facade hook
(gr_set_dither) recorded for the render thread and Smooth motion redraws.
With the mode off nothing flushes and no uniform changes. Flat, raw and
other 5-bit-exact pixels are never touched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- config: [video] post_shader is a std::filesystem::path (like memcard_dir);
the frontend turns it into UTF-8 bytes for the GL API. Assigning the path
to a std::string did not compile on MinGW.
- .pfx `generated reuse`: a Smooth motion generated frame keeps that pass's
target from the last real frame. The bloom preset marks its bright and
blur passes so; generated frames run only the masked composite (over their
own picture and mask). bloom debug reports real/generated pass times and
reused passes.
- Mask marking no longer queries GL state per batch: each call site puts
back the blend and mask-bit stencil it set (no driver round trips on the
render thread).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Post effects now run as chains of full-screen fragment passes on one
executor (gl_postfx.inc), as asked in review. A chain is a small .pfx preset
(postfx_chain.c, GL-free): passes with a shader (a file next to the preset or
builtin:<name>), inputs (source, the optional primitive-derived mask,
previous or a named earlier pass), a scale, format, filter and uniforms.
The executor owns target allocation and lifetime, program build/cache,
uniform binding (u_rcp, u_size, u_org) and GL state save/restore.
Stages, in order, each from a fresh copy of the image: the built-in "bloom"
chain ([video] bloom), a user chain ([video] post_shader / PSX_POST_SHADER,
elective, off by default), and the built-in "fxaa"/"fxaa_hq" chain
([video] antialiasing_mode). FXAA folds in as a one-pass chain and its
output is byte-identical to the former fixed pass; it still runs on real
frames only. Generated (Smooth motion) frames run bloom and the user chain.
The "may glow" mask is an optional pipeline input: draws keep stencil bit 1
only while an active chain reads `mask`, and the window-sized mask texture is
built once per present for the chains that use it.
Tests: postfx_chain_test (format, built-ins, errors, file import, no GL);
gl_bloom_test gains an imported two-pass chain (byte-exact invert), the mask
input (HUD unlit, 3D lit), a bad preset, a pass that does not compile, and GL
state restored with bloom + a mask chain + FXAA on, all on the caller thread
and the render thread. docs/POST_EFFECTS.md documents the format.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
[video] bloom = 0.0 (off, default) .. 2.0 adds a glow around lights on the
OpenGL backend. PSX_BLOOM overrides it for one run; {"cmd":"bloom"} reads or
changes strength, threshold and a debug view live. It is a plain [video] key,
so [quality.*] presets can switch it off on low-end machines.
The HUD never glows: with bloom on every textured and untextured draw is
followed by two stencil-only passes of the same vertices that keep a
"may glow" mark in stencil bit 1 (bit 0 stays the mask-bit mirror). Opaque
primitives clear it under their coverage, then PGXP-proven 3D primitives and
additive semi-transparent texels/primitives (modes 1 and 3) set it, painter
ordered like the colour; 3D and non-3D primitives never share a batch while
bloom is on. Semi-transparent primitives leave the mark alone and the PGXP
depth-clear pass (colour writes off) marks nothing. Without PGXP geometry only
additive primitives glow.
The pass runs on the composed game image before AA and the OSD: the presented
surface's marks are blitted into a quarter-resolution stencil, a soft-knee
bright pass runs under it, three separable Gaussian iterations spread it and a
screen blend adds it back. The marks follow the picture: native-wide centre
copies and Smooth motion surfaces (redraw and reprojection) copy stencil with
colour, and generated frames get the same pass. Off: nothing runs, no state
changes, and the composed frame is byte-identical to master.
test_gl_bloom (gl_bloom_test): off is byte-identical before/after bloom and
with --dump against master; on, a bright 2D panel and a panel over a lit 3D
polygon are byte-identical to off, an additive and a 3D polygon glow; guest
VRAM is unchanged; on this thread and on the render thread.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The seed path collapsed ls.texture_filter to a boolean, so a launcher's
xBR (3) was saved and applied as bilinear, and the rematch paths passed
the raw value through unchecked. psx_texture_filter_from_ui is now the
one translation (in-range values pass, others are nearest). Test:
video_texture_filter_test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixture now completes a pad exchange (sio_tick) so SIO holds a pending RX
byte with RX_RDY set and the ACK visible-read counter > 0, and asserts both
fields directly after the host-output callback (sio_peek_ack_visible_reads,
side-effect free, added for this). Without the fix the callback's reads
clear RX_RDY and age the ACK counter, and the test fails on those checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A raw dump with trailing bytes can be a multiple of 2048 and mounted as a cooked ISO
(Legend of Dragoon disc 1: PVD read as zeros, BIOS SystemError B/906).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A plugin that only reads guest state to drive this player's host output
(rumble, haptics) can now declare netplay = "host_output". Online it stays
on per player: its function-entry hooks run in the shared simulation while
the rest of the own-view plan is parked, a completion request is ignored and
every psx_mod_write_* it attempts is refused, so it can never diverge peers.
Offline it is an ordinary plugin. netplay_view_mods_test covers both.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The lobby server starts a 2-player room the host relay cannot carry on ICE
(recomp-net-server start_transport, launch transport "ice"). This client
knew only "host" and the legacy server relay. It inferred a relay from
equal host/guest endpoints -- which is exactly what two peers behind one NAT
advertise -- and each peer dialled its own public address: the match stayed
at "linking" forever.
- psx_lobby_launch_transport.h: one pure decision (host / ice / relay /
direct) with a unit test covering ICE, same-NAT endpoints, host relay and
the legacy relay cases.
- PsxLobbyJoinInfo.transport_ice; the launch handler records it and never
turns it into a relay; launch fill and resolve_use_ice start ICE p2p
(signalled over the lobby) for it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep both netplay_peer_frontier.h and netplay_fmv_confirm.h includes; advance lib/recomp-net to a48d355 (main, includes #38).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
PSX_RB_DET_SELFTEST on R4 4P showed a replayed save of tick T differing
from the live save of T. Each cause, and the fix:
- Replay's finish_frame poll passed no PC hint (Live used the IRQ resume
chain); it now uses psx_netplay_snap_resume_hint.
- Devices are serviced lazily up to the next event, so root counters and
cycles_since_vblank lagged the clock by a host-path amount at save; a
load re-anchors them at the clock and drops the lag. The save now
publishes pending cycle charges and services every device to the clock
first, and renders the SPU up to it.
- GTE / mul-div completion stamps and the load-delay give-back were
re-anchored at the restored clock on every load, so a stall in flight at
the snap point cost the live run cycles but the replay none. New
BS_SEC_CPUTIM carries them relative to the clock.
- The SPU CD-input FIFO was outside the snap and wiped on rollback loads;
new BS_SEC_SPUCD carries the queued frames and the load keeps them.
- LZCR is serialized as derived from LZCS (the raw backing stays 0 until
the first LZCS write; guests only ever read the derived value).
Older snaps without the new sections load as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replay's finish_frame saved its tick snap with no PC hint while Live used
the IRQ resume / last check / cpu->pc chain; the same tick's snap could
carry a different resume PC (psx_netplay_snap_resume_hint now serves both).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Live realign restores the snap at T, sets the session to T+1 and
longjmps back into the guest from the admit spin. The resume ran before
T+1 was admitted, so that frame latched no inputs and did not advance;
the next frame was then counted as T+1 and the realigned seat ran one
frame ahead of its tick label from there on (ring saves after a NACK
realign carried the next tick's clock, 2x the cycles in the first
tick). Only the realigned seat shifted, so it forked from every other
seat; with four seats this produced the "resim core diverge" splits
and the later hc-fork / hang cascade.
flush_resume now admits the tick first on the Live realign path and
keeps spinning until it can.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Verifies the four PS1 semi-transparency modes, set-mask, mask check and
texel STP at 4x internal resolution with PGXP depth on and off, directly
and on the render thread, against the PS1 formulas and the software
rasterizer (new gl_blend_accuracy_test).
Finding: with the historical surface bases (fills and texels stored as
k*255/31, everything else as k*8) single-layer blends are off by one 5-bit
step in places (1312 of the test's checked pixels, e.g. B+F/4 over a fill).
[video] accurate_blending = true stores every VRAM write as k*8 so GL's
8-bit blend truncated by the readback is exactly the PS1 arithmetic; whites
of fills/texels then present as 248 like uploads already do. Default off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The lobby server starts a 2-player room the host relay cannot carry on ICE
(recomp-net-server start_transport, launch transport "ice"). This client
knew only "host" and the legacy server relay. It inferred a relay from
equal host/guest endpoints -- which is exactly what two peers behind one NAT
advertise -- and each peer dialled its own public address: the match stayed
at "linking" forever.
- psx_lobby_launch_transport.h: one pure decision (host / ice / relay /
direct) with a unit test covering ICE, same-NAT endpoints, host relay and
the legacy relay cases.
- PsxLobbyJoinInfo.transport_ice; the launch handler records it and never
turns it into a relay; launch fill and resolve_use_ice start ICE p2p
(signalled over the lobby) for it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With three or more seats two guests can mispredict the same tick and both
BEGIN. The tie-break (lower initiator slot wins) ran only on initiators,
so a follower followed whichever BEGIN arrived first and stashed the rest.
On a 4-seat WAN race the host followed slot 3's epoch while slot 1's won;
the seats split across epochs, each waited for baselines that never came,
and the match ended after 20 s at rb_baseline.
The epoch id names its initiator, so a follower now applies the same rule
to the episode it follows (netplay_rb_concurrent_begin): yield during the
handshake and follow the winner, or drop the loser's BEGIN. A follower
seeing a new epoch from the initiator it already follows (abort and
reopen, the only case two seats can produce) is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two rollback bookkeeping bugs that only show with more than two seats
over a real link:
- g_peer_resolved_through took the highest RESOLVED from any peer. With
three seats the fastest follower stood in for the others, so the host
opened loads past the slower followers' frontiers (3P WAN: load 2400 vs
frontiers 2344 and 2319); every follower refused, episodes aborted and
the session diverged at 2496 before the race started. Adverts are now
kept per seat (netplay_peer_frontier.h) and the lowest across the
peer roster is the frontier.
- The §55 bisect cap (next load < L after a baseline mismatch at L) was
only cleared by an episode commit. An abort followed by a Live heal
left it armed for the rest of the session; once its loads left the
snapshot ring every episode was refused with "no confirmed snap"
(4P WAN: cap 1928, agreed 3136, ring 2656..3166) and a later fork at
3150 was never repaired. A live hash-confirmed agreed ADVANCE at or past
the cap now retires it.
Bumps lib/recomp-net for RB_RESOLVED sender attribution and rb_* queue
retransmit coalescing (RetroPortingToolKit/recomp-net fix/rb-ctrl-queue-dedupe).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Enhancement arenas (GPU DMA aperture, Expansion-1 mod memory) were untracked, so Crash widescreen drew every vertex native with no depth. Adds depth stats/clear causes to pgxp, depth to pgxp_tri_ring, input_route_start at_frame, integer number keys in game.toml.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Games that read ENVX right after KEYON saw 0 until the per-frame pump and keyed the voice off (Crash spin sfx at 21:9).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Matthew Stanley