Start a newly selected production listener on offline rematch, reuse an existing listener, and cleanly release its transport state when disabled. Balance Winsock ownership and failed starts. Native production runtime and four focused tests pass, including six isolated real listener/bulk-write launch cases. Tracking: beads-eio.3.385.
- Filtered replacement sampling: with any texture filter but nearest, the
replacement is sampled through its mip chain (textureGrad), the footprint
clamped to the primitive's own UV bounds so atlas neighbours cannot bleed
in. Texture windows (repeat) and UVs outside the page keep hd_texel.
- Edge clamp: an out-of-bounds sample clamped to the primitive's UV bounds
takes the edge's sub-texel position too (keeping the outside fraction read
a 1 px sliver of the next glyph at sprite borders).
- GL cache: 4096 slots and a PSX_HD_GL_CACHE_MIB budget (default 2048 MiB)
with a direct-mapped hint table; a 4x pack of one race scene is several
hundred textures, and the old 256 slot / 128 MiB cache re-uploaded and
re-mipped them every frame.
- Invalidation: duck_texture_pack_invalidate returns early when a native
write touches no residency or cache entry (almost every draw under HD
authority lands on the frame buffer), and retire_unused walks the
residency lists once instead of once per source.
- Indexed replacements: a TTTTTTTT-index.png key resolves through the palette
in use at draw time, for packs whose textures are shared by several CLUTs.
- Debug: TCP {"cmd":"hd_uploads","x":..,"y":..} lists the texture uploads
covering a VRAM point (pack authoring).
Measured with a full HD HUD and course pack for R4: steady 60 fps in races
where the old cache dropped to the high 20s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rkyma8rbyPiccs9n2bkyG6
For title-specific lighting presets ([video] post_shader):
- `input <sampler> depth`: view depth sz (GTE screen z) of the presented frame,
R32F at window size, rebuilt from the surfaces' PGXP depth; 0 where nothing
3D drew. Only built while a loaded chain reads it.
- `input <sampler> history`: the pass's own target from the last real frame
(trails, temporal filters); generated frames do not update it.
- `u_time` (seconds since the chain loaded) and `u_proj` (latest GTE
projection: 240/H, screen centre in uv, aspect) in every pass.
- Live uniforms: psx_host_set_post_uniform(name, x, y, z, w) sets a preset
`uniform` from game state (e.g. a light source projected to the screen);
the preset value applies until the first call.
- Debug: TCP {"cmd":"post_shader"} reports the loaded preset, its parse status
and the live values; {"cmd":"post_shader","path":"..."} reloads a preset
(or loads another) without restarting, for shader work.
docs/POST_EFFECTS.md documents the inputs and the uniforms.
Used by a lighting preset for R4 (sun and tunnel light shafts, lamp halos,
speed streaks, paint reflections on the player's car from live car pose),
about 0.7 ms of GPU at 1080p on an RTX 3060.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rkyma8rbyPiccs9n2bkyG6
Reject invalid index products and unsafe paths, cap encoded/decoded residency, recover allocation and partial worker-start failures, release GL movie textures with their context, and serialize diagnostics through the presenter. Real SDL decode/restart/failure test passes and SDL2 C seam compiles warning-free. Guest/present timestamp ownership remains a separate design gate. Tracking: beads-eio.3.388.
An optional pack replaces a title's MDEC movies with high-definition frames
while the game keeps its own audio, timing and skip control:
- fmv_hd/<movie>/index.bin: header, then a luminance signature per source
frame; fmv_hd/<movie>/frames/NNNNNN.jpg: `sub` HD frames per source frame.
- Every presented 24-bit frame is matched by signature (mean absolute
difference), first around the last recognised frame, then over the whole
pack when tracking is lost; a near global match is also tried a few lines
up or down (each movie has its own vertical placement on screen) and the
offset kept. Worker threads decode the upcoming frames ahead into a small
RGBA ring.
- Format FMVHD002 records the native display the movie is shown in and its
position, so the HD frame lands exactly where the original was.
- The last HD frame is held briefly when the next one is not ready or not
recognised (no flicker back to the 4:3 original mid-movie), and is
uploaded to the GPU only when it changes.
- Debug: TCP {"cmd":"fmv_hd"} reports the pack, the matches and lost frames.
No pack: no effect. Pack frames are the user's own upscale of their own disc;
nothing is distributed with the runtime.
Used for R4's intro and ending movies (SeedVR2 upscales, 25 fps sources).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rkyma8rbyPiccs9n2bkyG6
Use a bounded atomic snapshot and presenter-owned reset across all backends. Readers sort their own snapshots; report host API completion cadence. Add deterministic and concurrent reader/reset coverage.
Bound reservations against total retained residency, skip decoding pinned images, and report complete repeated loads truthfully. Cover cap reductions and publication retry in both formats. Tracking: beads-eio.3.386.
Both preload functions now take an aggregate decoded-byte cap and a thread
cap (runtime: 512 MiB, 4 threads). Each image is probed and its RGBA
reserved against the cap before decoding; per-image limits follow the
pack's decode budget, which is no longer grown to usage (pinned images sit
outside it, and outside the Duck 512-entry cap). The pool (texture_pack_
preload.hpp) joins every started thread on unwinding, survives thread
creation failure, and catches worker exceptions. Images not preloaded are
left to the asynchronous budgeted path instead of being marked failed, and
HdTexturePreloadResult reports loaded / over-cap / failed / worker errors
with a complete-or-partial status. The Beetle async worker no longer lets
allocation failures escape its thread entry.
Tests cover both formats: complete preload, over-cap skip with on-demand
fallback, malformed images, injected worker and thread-creation failures,
and budget reservation bounds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replacement images were decoded on demand by a single low-priority
worker, so a texture first seen mid-race appeared late and could be
evicted again. When replacements are on, configuring a pack now decodes
every unambiguous image (DuckStation and Beetle formats) across all
cores and pins it in RAM before the game draws; pinned images are never
evicted by the decode budget. Logs the count, size and time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Some PlayStation 1 titles implement trivial rejection by folding the sign bits of projected vertex screen coordinates across triangle or quad primitives with a chain of AND instructions. Without adjusting these checks during widescreen presentation, geometry in the newly revealed horizontal area is still rejected by the original 4:3 bounds. Rather than embedding title specific instruction patterns directly into the recompiler, this introduces a generalized `[[widescreen.cull.sxy]]` configuration schema within `game.toml`.
Game configurations can now declare multi vertex trivial rejection sequences by their exact AND instruction words, vertex registers and the branch that consumes the result. The loader derives the expected chain from those words and rejects any other shape. Each AND keeps its original value except for the horizontal sign bit, which survives only while that vertex lies beyond the revealed left edge, so vertical and right side rejection are preserved and 4:3 output is bit identical. Shifted coordinate lower edge comparisons reuse the existing `screen_x_sites` handling through a new `slti rt, x, 0` form instead of adding a separate site family.
Full parity is maintained between recompiled code and the dirty RAM interpreter by sharing the same runtime helper and validating every fold and consumer instruction word on both paths, leaving mismatching overlay variants untouched. This eliminates hardcoded game logic from the recompiler while giving titles clean geometry retention in expanded aspect ratios.
Matthew Stanley