The review found that astc_trim_check, evo_peek_check, heap_lazy_copy_check and gen_wrap_check were in neither the
pass's test matrix nor the docs, so a later change to the code they cover would not know to run them. Each now has a
sentence in the "A change to X passes python tests\Y.py" form: vk_astc.c's scratch trim with vk_defer_buffer_try,
vk_evo_peek.h with vk_exec.c vk_exec_held_ready, memory.c mem_copy_readable with the lazy heap, and vk_memory.c's page
write generations.
Docs only. Checks that passed: check_content --staged; the four checks pass with system Python.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The review of c05e6528 found a window in perf.c ft_frame: it loads walk_t0 and then walk_t1, while perf_census_walk(1)
stores walk_t1 = 0 and then the new start. A walk that starts between the two loads leaves ft_frame with the previous
walk's start and an end of 0, so perf_census.h census_charge took that walk as running since then, the whole frame as
overlapped, and moved all its replay and unattributed time to "census". The window is a few nanoseconds, but the error
is a whole frame. A walk takes 100-260 ms, so census_charge now returns 0 for a walk seen running for longer than
CENSUS_WALK_MAX_NS (2 s).
tests/census_charge_check.c gets a torn read (a start 10 s back with an end of 0), a walk running for 1.9 s, which still
counts, and one running for 2.001 s, which does not; its random frames now reach walks more than 2 s old and check the
moved time exactly (the overlap capped by the frame's room and its replay and unattributed parts) instead of only its
bound. Built against the header before this change, the check fails the torn read, the 2.001 s walk and the first such
random frame; with it, it passes. docs/building.md names the new case.
Diagnostics only: the census part exists only with NX2NAVI_MEM_CENSUS = 1. Not reproduced in a run (the window is a
few nanoseconds). perf.c compiles with the dev flags and adds no -Wall -Wextra warnings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- nx2navi.ini.example said the stutter map is written every 10 minutes and at exit; since 7b0dbeb1 it is every 5
minutes of play (stutter_map.c SM_REPORT_NS), at a focus loss in play and at exit, and since 2f640fb8 a run of
consecutive long frames counts as one hitch with their summed time. The paragraph now says both.
- docs/building.md had no paragraph for tests/freeze_hold_check.py (0e0cfe08), while each other header-alone check has
one: it now says what a change to runtime/freeze_hold.h or vk_memory.c freeze_pass must pass and what it checks.
- vk_texture.c released_cap_bytes named "a thirty-second of the installed memory" as the source of the bound even when
plat_mem_info gave no size and the bound fell back to 1024 MiB; that case now reads "the default without the
installed memory's size".
Text only. check_content passes; freeze_hold_check passes; vk_texture.c compiles with the dev flags and with -Wall
-Wextra with the one warning it had before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
perf.c rt_knob reads "0" as off, "ab" as the A/B blocks and everything else as on, without a word: "cache_die = off",
"crash_filter_guard = false" and "one_per_core = 0 ; x" all read as 1 (the audit's probe, Fv/knob.exe, cut rt_knob out
with cfg.c: three 1s and no log line). A player who wrote such a value meant the opposite and had no way to learn it
from the log.
rt_knob now logs "cfg: NX2NAVI_<name> = '<value>' is not 0, 1 or ab: read as 1" the first time it reads such a value
for a switch. The value still reads as 1: reading words as 0 changes what a written setting does and waits for
approval with F-8. Most of the 109 call sites keep the result in a static but some read it per call, so the warned
names are kept in a table of 32 pointers (the names are string literals) filled by compare-and-swap; 0, 1 and ab pay
only one more string compare.
New tests/rt_knob_check.py cuts rt_knob and its table out of perf.c and builds them with stand-ins for the settings and
the log: 0, 1, ab, empty and unset read as before without a line; off, false and "0 ; a note" read as 1 and log once
each however often they are read. docs/building.md describes it. The fault was shown before the change by the audit's
probe; rt_knob_check passes, and perf.c compiles with the dev flags and with -Wall -Wextra without new warnings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With mem_census = 1 and trace = perf, vk_memory.c census_main walks the address space for 100-240 ms on its own
thread, holding the process's address-space lock. The replaying thread's write-watch queries wait for that lock
("mem: slow write-watch query: 20-34 ms ... 0 pages written"), and ft_frame charged the frame to "replay": in the
0.10.0 fresh session (nx2navi.1.log, 7,974,927 bytes) 63 of the stutter map's 66 replay hitches sat beside such a
walk, 49 of them at frames 600N+1..3. The presenting thread's own 600-frame memory report (10-17 ms, timed in
vk_mem_census since b81a9561) went to "unattributed". And at frame 3600N, census_check walked on the presenting thread
and the same frame's heap_log started census_main's second walk.
- perf.c perf_census_walk brackets census_main's walk (atomics); ft_frame moves the part of a frame over 25 ms (or a
stutter-map candidate) that overlapped the walk out of the replay part first, then out of the unattributed part,
into "census". The rule is runtime/perf_census.h census_charge, capped by the overlap and the frame.
- A new cause "report": vk_mem_census feeds the report's time (perf_report_part), and so do perf.c's memory
snapshots and its 600-frame sizes lines; the next frame takes it, so the overlap rule no longer has to infer it.
The cause list changes together: perf.c FT_C_* and ft_cause_name, stutter_map.c SM_CAUSES and cause_short,
tests/stutter_hitch_check.c names[].
- census_check records the frame it walked (perf_census_walked_frame); heap_log does not start census_main for that
frame, so frame 3600N walks once. A frame stamp, not a time window, since prof_commit_census can fail and leave
perf_note_wc uncalled.
- The "mem: process" line gives the walk's start and end on the run's clock and its length, so a play-test with
mem_census = 0 can tell the walk from the presenting thread's report as the cause of the 600N+1 hitches.
New tests/census_charge_check.py builds runtime/perf_census.h alone: no overlap, partial, contained and still-running
walks, the charge from replay first, capped by the overlap and the parts, and 100,000 random frames in which the census
part never passes the frame; docs/building.md describes it. Diagnostics only: NX2NAVI_MEM_CENSUS defaults to 0, and
nothing in gameplay, pacing or the save changes. Not reproduced before the change (it needs a game run). The pass's
known-issues table, which should say that hitch counts are taken with mem_census = 0 and trace off, is not in the
repository. census_charge_check, stutter_hitch_check and stutter_map_check pass; every runtime file compiles with the
dev flags, and perf.c, stutter_map.c and vk_memory.c add no -Wall -Wextra warnings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The perf lines could not see pacing. In the 0.10.0 fresh session (nx2navi.1.log, 7,974,927 bytes) all 666 windows read
"pacing 0%" and "bound: cpu" and all 70 slow-frame lines "pacing 0.0 drain 0.0", while 610 windows had a 16.x ms
interval at fps_limit 60 under the game's own limiter. The pacer's wait happened where perf does not count it:
vk_backend.c present_image's wait for the present thread to take the previous picture was untimed, perf_end drops
waits on any thread but the presenting and replay ones (a game may acquire its window image on another), and the
acquire's GPU wait was counted twice, as pacing around b->fence_wait and as a fence wait inside vk_sync_fence_wait
(F-7-04).
- host.h: PERF_HANDOFF, the presenting thread's wait to hand the present thread the next picture, timed in
present_image from under pl (where pending is read) and kept in PresentJob.handoff_ns. perf_end_any counts a wait
on any other guest thread into a sum that perf_present adds to the frame before ft_frame reads its parts.
- nvn_api_obj.c nvnWindowAcquireTexture: the pacer's sleep is timed with perf_end_any; the GPU wait counts only as
the fence wait it is; one line names the threads that acquire and present once the first present is known.
- vk_backend.c present_main: the TRACE=present line adds the present call's time (vkQueuePresentKHR, under the queue
lock, where an overlay's frame limiter sleeps) and the job's hand-off wait; with trace=perf, a line per 60 presents
gives the present thread's blit and submit, paced wait, present call (avg, max), time past target and the hand-off.
- perf.c: "handoff" in the window line's shares and the slow-frame line, "hand-off" in the worst-frame lines and as a
cause of its own for frames over 25 ms and the stutter map. The "bound:" rule stays as it is until play-test numbers
exist.
- The cause list changes together: perf.c FT_C_* and ft_cause_name, stutter_map.c SM_CAUSES and cause_short,
tests/stutter_hitch_check.c names[] (now sized from SM_CAUSES), and tools/stutter_map.py, whose cause and hitch
patterns now take a hyphen. flight.c's frames.tsv gets a handoff_ms column (blocked_names checked against PERF_N),
and tools/flight_view.py lists it.
- tools/perf_summary.py reads the hand-off share, the present-thread lines and the TRACE=present lines, and prints
percentiles of the present interval, paced wait, time past target, present call and hand-off.
New tests/perf_summary_check.py makes lines with the runtime's format strings (checked against the C sources) and
checks the percentiles, the window averages and the hand-off share, and that older lines without these fields still
parse; tests/stutter_map_check.py gets a hand-off hitch and area cause, which the old patterns dropped. Diagnostics
only: no wait, pacing or presentation change. Not reproduced before the change (it needs a game run with trace=perf).
perf_summary_check, stutter_map_check, stutter_hitch_check, flight_flash_check and frame_step_check pass; every
runtime file compiles with the dev flags after the host.h change, and the changed files add no -Wall -Wextra warnings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A session of 2026-11-13 grew from 683 to 1095 OS handles in 11 minutes, rising at every area load and never falling,
while the guest's own handles stayed at 116-122 and the threads flat; the VMA block count fell back while the handle
count did not. The log could not show it: the 600-frame "mem: sizes" line had only the guest's handles, and nothing
told which kind of handle grows.
kernel.c kern_sizes now adds "process handles" (plat_handle_count, one system call every 600 frames under trace=perf)
beside the guest's "handles". platform.h gets plat_handle_census: on Windows the system's extended handle table
filtered by our process id, with the type names from NtQueryObject's ObjectTypesInformation; on Linux the
/proc/self/fd links grouped by kind. It never opens, duplicates or names a handle (a name query can block on a pipe).
With NX2NAVI_HANDLE_CENSUS=1 (default 0, read at start) and trace=perf, perf.c starts one long-lived thread that the
600-frame report wakes; when the count grew by 16 or more since its last census it logs "mem: handles frame N: T in
all (+D since the last census): Event a (+x), ...". The census reads a few MiB of the system's table in 2-5 ms, off
the drawing threads; one thread for the run keeps a new thread's own handle out of what it counts. New cfg rows:
HANDLE_CENSUS in known[], sections[] (diagnostics) and metas[] ("0", "0|1", restart), and an nx2navi.ini.example
entry.
New tests/handle_census_check.py builds the platform layer alone: 20 events and 5 sections show as Event +20 and
Section +5, the total stays within 2 of GetProcessHandleCount, the figures return once they are closed, and kinds
past the room given go into the last entry; docs/building.md describes it. Not reproduced before the change (the
growth needs a game run). handle_census_check, sync_wake_check and the cfg row check pass; every runtime file compiles
with the dev flags, and perf.c, kernel.c, win32.c and cfg.c add no -Wall -Wextra warnings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The worker read RTSS's shared memory, decided and wrote the profile in one loop over file statics, so its decisions
could be checked only against a running RTSS. rtss_decide.h now holds the parts that make no system call: rtss_parse
reads the shared memory's header and this process's entry (bounded by the memory's size and 4096 entries, as before),
rtss_decide takes one poll's view (the want, whether it is still current, a closed run, the hooks library, the entry)
and the state between polls (RtssState: the limit applied and the one refused, whether the run changed the profile,
the frame count, the failure) and returns what to do and whether RTSS holds the game, and rtss_poll carries that out
through the profile calls rtss_limit.c hands it (apply and put_back, which keep their I/O and log lines). The worker,
rtss_limit_restore and the menu state read the same state.
No decision changes: the order of the rules, the immediate write before RTSS lists the game, six polls over 1.2 times
the limit, a write per movie and per switch of low latency all stay as they are. The rules that wait for approval are
not in it (the pacer keeping movies and windows with no host present, Vulkan flags, settling and a fresh period
before standing aside, rounding the limit down, the state file).
tests/rtss_worker_check.c gains nine scenarios that pin those decisions against the real worker (a read-only profile,
a refused limit, the frame count rule at 144, 69, 73 fps and unlimited, RTSS quitting, a hooks library that does not
load, low latency switching, a movie); the logs of all 23 scenarios are the same, line for line, before and after this
change. tests/rtss_limit_check.c checks the same decisions on rtss_decide.h alone, with a fake RTSSSharedMemoryV2 for
the reader (a bad header, short entries, an entry past the memory or past the 4096th) and a fake profile behind
RtssOps; mutating the frame count's polls or slack, the refusal, the library failure or the clear of a refused limit
makes it fail. docs/building.md describes the check. Every runtime file compiles; -Wall adds no warning;
frame_step_check passes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This takes back the three commits that added it (cd2f9f1, e60677e, a01910d): its runtime layer and Steam binding, the Co-op tab and what it drew over the game, the launcher's copy of Steam's library file, its settings, tests and docs. The tree is as it was before them, with the bug capture that came after kept: its note on screen no longer makes room for co-op's lines, and dbg_area, which the co-op had brought and the capture uses, stays.
The four settings of that co-op join the retired names, so a settings file that still has them logs no unknown key. A steam_api64.dll that the launcher put beside the game is left where it is: nothing loads it any more.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The character controller's three follow-surface modules kept only a tenth of the floor's velocity whenever the physics' inverse step passed 100, which every update above 100 fps did. The gate now compares with 100 over frames per update from the patch page above 60 fps, and with the game's 100 at 60 and 30 fps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Link is carried by the Deku Tree's floating platform fully at 60 fps, about a tenth at 160 and not at all at 240,
which follows the solver steps an update (4, 2, 1 with physics_steps = 1) rather than those a second. The new set
osolverfull raises the rounded-up count to the 60 Hz frame's 4 in every update; physics_steps = 2 switches it on and
osolver off at start, and the mode is logged. 60 and 30 fps keep the game's 4 and 8. It costs more physics time above
60 fps: 576 solver steps a second at 144 fps and 960 at 240, against 288 and 240 with 1, the default, which stays
until a run confirms the carry. frame_step_check runs the lifted words of all three counts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Up to four players, each in their own game and save, meet in a room. Rooms are Steam lobbies and the games talk through Steam's networking, so nobody forwards a port and nobody runs a server. The game signs in to Steam as its test app (480), so Steam shows the player as playing Spacewar. Off by default; the menu has a Co-op tab with Create a room, Join a room by its ID, Invite a friend, the players, chat and Leave, in English, Russian, Brazilian Portuguese and Simplified Chinese.
runtime/steam.c loads steam_api64.dll from beside the exe at run time and declares the C functions it calls itself; nothing of the Steamworks SDK is in the repository, and a player copies the file from a Steam game they have. runtime/coop.c is the room: the lobby's owner is the host, lets games of the same protocol and game build in and lists them, and every game takes positions, chat and progress only from the players on that list. runtime/coop_game.c reads Link from the Player actor and shares progress, items, songs and quest items through the game's own flag routines, each value only ever raised; small keys, rupees, ammunition and hearts stay each player's own, and the save is copied to coop-backup before anything of a room reaches it. The other players show as a name with a distance where their Link stands, projected with the upscaler's camera.
One Steam account cannot be in a lobby twice, and this PC has not the memory for two games, so the room and the game side were run with one game and tests/coop_peer.py as the second player over runtime/coop_local.c, a stand-in for Steam on 127.0.0.1 that only the environment turns on: the peer was let in, saw the game's Link in Hyrule Field, chat went both ways, the save was copied, one flag the save lacked was set and a flag the game does not share was refused. On Steam itself one game signed in as app 480, made a room, showed its ID and the player's Steam name on the Co-op tab and left. Not run: two Steam accounts in one room, an invite, a join by room ID over Steam, a host that leaves, a removal, and the menu's text probe on this build.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The physics' two water managers and the game's interactive floor simulation (the Deku Tree's ground webs) step their
surfaces in fixed steps of 1/60 s, as many whole steps as the frame's time holds, and drop the fraction, so every
frame shorter than 1/60 s ran none and the surfaces stood still above 60 fps. The new set owater (water_steps, on by
default) calls the page between the division and the game's cut, which keeps the fraction per object; at 60 fps the
count is the game's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hidden beta shared story progress through a relay a player had to run and never showed the other players. A new co-op replaces it. Gone are runtime/coop*.c and their headers, the menu's Co-op tab and its texts in four languages, tools/coop_relay.py, the co-op tests, docs/coop.md, the [coop] section of nx2navi.ini.example, the --probe-coop mode, the coop trace item, and Monocypher, which only this layer used.
The old settings stay known as retired keys, so an nx2navi.ini that still sets them logs no unknown key, and they stay on the list of private settings, so a bug report keeps a room's secret, a relay's address and a name out. The platform layer's UDP sockets and fs_save_snapshot stay for the new co-op; cfg_ini_now and fs_profile_path, which nothing else called, are removed.
Checked: the game builds and starts, the menu shows its four tabs, --probe-cfg and --probe-fs pass, and the launcher's private-settings test passes. tests/save_share_check.py builds without its co-op stub; it cannot run from this worktree because the folder's path is longer than Windows allows, on main too.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The patch set text_speed scales the typewriter speed of a message the game lets the player skip (fast, three times as
fast, its scripted waits as long) or runs the game's own skip on the first frame of each page (instant). Messages the
game marks unskippable, such as those that turn their page on a timer, keep the game's speed. The setting is live and
has a row on the Game tab in all four menu languages; with it at normal the set is off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The cache kept the first eight sizes and formats for the whole run and refused every other one, so after a few
window, resolution or DLSS mode changes the stand-in's blit of a new size was left out, while the old sizes' images
stayed. A new size now takes a free entry or the one used longest ago, whose image goes through the deferred
destruction after the tonemap pass drops its views of it (vk_tonemap_forget), before another image can take its
handle. tests/scratch_lru_check.py checks the choice.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The two sites that count a movie's event frames from the shown picture's timestamp were part of the real-time step's
set, so the only way back to the old count also turned the step off. They are set omovie now, NX2NAVI_MOVIE_PTS
(movie_pts in nx2navi.ini), on by default.
The comment says what the new count changes at the default movie cap: nothing while the decoder keeps up; in a stall
the cues wait with the picture instead of going on with the soundtrack, and step by 2 or more when a picture is
dropped. The counter's reader, main+0x3b3ec10, hands it out as a float frame through a virtual call; its callers were
not traced, so whether a cue on a dropped picture's tick can be missed is said to be unknown, which is what the
switch is for.
The movie cap's comment keeps the measurements it rests on, the narration 2.4 times early and the fade at 65 s of 154
at 144 fps, with one sentence on the new sites. The new test is described in prose with what it does not cover,
instead of a row in the lifter's table.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The window's close ends the process under the detached threads, so a dump file the flight recorder's thread was
writing stayed cut short under its own name. Its three files are now written as .part and renamed when complete
(runtime/part_file.h, with a check that kills a writer in the middle), and the allocator's map goes through
plat_file_write_cached. The shader identities and the other cache files already went through an atomic write.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
With the setting on, textures no shader sampled for ten seconds are copied on the GPU out of the emptiest allocator
blocks into the holes of fuller ones, so the emptied blocks go back to the driver. A pass starts when the free space
inside the device-local blocks reaches NX2NAVI_TEX_COMPACT_MIN_MB (one block under VRAM pressure) and the block
table has blocks that hold nothing but such textures, with room elsewhere for what moves. It takes at most four
textures and NX2NAVI_TEX_COMPACT_FRAME_MB a frame, and none in a frame that loads an area, already made or uploaded
textures, or follows a long one. A released texture the size cap would drop is dropped instead of moved.
Only a texture whose image holds what the game's memory holds moves: uploaded, never written by the GPU, single
sampled, not depth. The entry takes the new image under the texture lock and its address generation moves, so every
slot, pool descriptor and memo resolves again. The old image and its views stay until the texture pool has taken
every set of its ring and the GPU is past the submissions recorded until then, as a retired texture's do.
Each pass logs what it moved and dropped, the blocks and bytes given back, and the time it took. A block emptied of
textures that the allocator keeps is left alone for ten minutes, and a pass that gave nothing back doubles the wait.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A log without the perf trace said nothing about the free space inside the allocator's device-local blocks, which the
driver counts as in use. Every 3600 frames (600 with the perf trace) a "mem: holes" line now gives the blocks and the
allocations in them, and for the memory type the textures use most the blocks that hold a texture, the textures and
VRAM mirrors in them, the blocks at most half full, and those of them that hold nothing but textures that could be
copied elsewhere or dropped: emptying those is the most a compaction could return.
The counting is runtime/vk_tex_compact.h, which also chooses the blocks to empty, with tests/tex_compact_check.py.
A texture entry keeps the block its memory lies in, so the table is built without asking the allocator per entry.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The section described the build of the other game versions with an example for 1.4.3. It now says how make_game.py
takes a kit's profile and when it leaves it out, which LLVM releases write and read the file, which packages the
instrumented build and the merge need, and the steps of a training session for 1.1.1: the build command, where the
game is started, what to play, where the counts land, the merge, and the comparison of both builds with the runtime's
sampling profiler and its frame timings.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
runtime/rumble_hd.c turns one of the game's vibration values (a low and a high band, each with a frequency and an
amplitude) into the four bytes a linear actuator's rumble report takes, clamps the frequencies to each band's ends and
the amplitude to the last safe step, and decides when a report is written: the newest value only, a held one again
every 30 ms, the neutral value three times once the game stops. Nothing calls it yet.
tests/rumble_hd_check.py builds it alone, checks known values, the clamps and a batch reduced to its reports, and
decodes a sweep of values with formulae of its own.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A marker or a journal left by a start that died in the boot's builds was always read as a driver crash. Since the
whole warm those builds run for minutes, so a start ended from outside in that time (the task manager, a program that
stops the game, the power) switched the pipeline prebuild off for that driver for good, or put up to 32 variants on
the skip list that the driver had merely been building. Closing the window was already an orderly end and still is.
The crash handler now leaves a note while the boot's builds run. With the note the next start does what it did: it
skips the prebuild, or the variants the journal names. Without it the end is taken for a stopped start: the pipelines
are built again and the journal's variants are blamed on nothing. Three such ends in a row still skip the prebuild,
for a driver that ends the process without a fault the handler sees. The boot's journal has a file of its own, so the
preparation's keeps its meaning.
The whole warm waits prewarm_full_s seconds at most (600): when they are up the game starts, the builders step down
to the background's thread counts and priority, and the file says "done", so the boot screen has one turn per driver.
tests\boot_warm_check.py covers what a leftover marker means with and without the note.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A start that finds the pipeline cache written by another device or driver has nothing built, and everything was
compiled beside the game: an 8-core integrated GPU drew fewer than 600 frames in its first 11 minutes. Such a start
now waits on the "Preparing shaders" screen, with counts, for every recorded pipeline, for the compute pipelines on
every core at normal priority, and, when a sample says the driver's own cache lacks the prepared stage libraries, for
all of them on every core. The memory gate still holds for the builds, and the library builds write the preparation's
journal while the start waits, so a variant the driver's compiler dies on is skipped from the next start on.
boot-warm-owed.txt beside the caches names the GPU and driver: a start closed during the warm is followed by one that
goes on with it, three at most, and "done" keeps later starts ordinary while the cache file still carries the earlier
driver's header. When the boot stops waiting (prewarm_wait_ms, or a minute without progress) the builders step down
to the background's thread counts and idle priority. prewarm_full = 0 turns the warm off.
With nothing in the driver's cache the probe took 425 s on an RTX 5090 with 16 CPU threads: 87 s for 5,775 recorded
pipelines, 19 s for 4,075 compute pipelines and 317 s for 42,231 prepared variants.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Every guest thread is a host thread, and the game makes and ends them all through play. A thread that slept through plat_wait_ns kept a waitable timer, one that read or wrote a file an event for each, and one that decoded zstd through the host a decoder context of about 100 KiB; nothing closed or freed them when the thread ended, and a detached thread's own record stayed allocated. They are now given back when the thread's function returns: the handles in win32.c thread_main, the decoder context through plat_thread_end_hook, and the record by whichever of the thread and its detach comes last.
tests/thread_end_check.py runs 416 such threads and compares the process's handle count before and after. tests/sync_wake_check.c gets the two stand-ins kernel.c has needed since the service wait log, so that check builds and passes again. The Linux file only gains the hook call and was not compiled here.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A frame that was not rendered when it was due stayed lost: the timer set its
deadline to now once it was more than 100 ms late, and the device-paced path
only filled the ring to its target again, so after every gap the sound was
heard that much later than the picture it belongs to.
Frames that are owed are now rendered and dropped: mixed in full, so the
voices, the effects and the sinks the game reads back advance, and not queued
for the device, whose queue is filled to the same target afterwards. While the
device paces, what is owed is the silence it played for want of frames,
counted in its own samples from an underrun until the prefill is queued again;
the wait for the first prefill and anything owed across a stop do not count.
On the timer a frame is owed when its deadline passed a whole frame ago. A gap
of more than 3 seconds is not caught up: the sound goes on from now.
NX2NAVI_AUDIO_CATCHUP=0 turns it off and keeps the old timing. The audio
statistics line counts the catch-up frames and the gaps that were too long,
and tests/audio_mixer_check.py checks the device's side of the accounting.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The mixer applied each voice's mix volume and each splitter destination's
volume to a whole 5 ms frame as one constant, so the level jumped at the
frame's first sample whenever the game moved a volume, which is heard as a
click. It now keeps what each voice and each splitter destination was mixed
with in the frame before and goes from there to the new value in equal steps
across the frame's samples, as the console's renderer does. An unchanged
volume stays on the constant path.
A voice that has just started, one that was silent in the frame before or was
moved to another mix or splitter, a destination that is new or was moved, and
the first frame after the renderer starts take the new volume at once.
NX2NAVI_AUDIO_RAMP=0 turns the ramps off. The audio statistics line counts the
render passes that ramped a volume. tests/audio_mixer_check.py builds the
mixer with the rest of the runtime stubbed and checks the ramps and the cases
that must not ramp.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
With flight = 1 and its thumbnails kept, the recorder now also writes a dump when a frame's picture differs strongly from the frame before and the frame after while those two agree, which is what a one-frame flicker looks like. Each thumbnail is reduced to 16x9 cells of mean brightness where the present thread already sees it complete, and a frame is a flash when at least three cells are more than 24 of 255 brighter or darker than in both neighbours, the neighbours being alike in those cells, around them and over the picture as a whole. A camera cut, a fade, a pan and shapes crossing the picture are no flash, and a frame beside one without a thumbnail is not judged. The dump is the one F8 writes, named "flash" with the frame's number in its header line, at most one every 30 seconds. flight_flash = 0 turns it off; flight_flash_diff and flight_flash_cells set the two thresholds.
On the thumbnails of earlier dumps the defaults flag the frames whose scene was white once and nothing in the runs without that fault. tests/flight_flash_check.py checks the test itself on drawn pictures.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The water renderer runs trunc(step) times the water's substeps of its ripple simulation a frame, which is none while the step is under one frame: the rings stood still above 60 fps. A new site of the set ostep hands the cut to frame_rate_water_host, which adds the step to a remainder kept for the water object and returns the whole steps in it. The waters have a table of 64 slots of their own, as fewer slots than surfaces would have them take each other's in turn.
frame_step_check.py calls the water and day hooks as their sites do. It also gets the three platform stubs it has lacked since the day clock's hook, without which it did not link.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The log thread wrote as soon as it woke, which outside a burst is once for every line: each line woke it and was a write of its own to the log file and the console. It now waits a quarter of a second from the first line it finds, or until the queue is half full, and writes what was queued by then at once, so the log is up to that much behind the game. What is being gathered stays in the queue, so a fatal error, a crash report and an exit write it out before their own lines as before; a game ended from outside loses those last lines. The queue's test checks the gathering, the half-full write and that a process which ends at once, by an exit or as a fatal error does, leaves every line in its file.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
On Linux every CPU-visible pool had a host-visible shadow copy, because amdgpu treats a host-pointer import as a userptr object that it looks up again at every submission and invalidates whenever the write watch or a freeze re-protects its pages. Now the watched reservations are shared mappings of one memfd, /dev/udmabuf makes dma-bufs of page ranges of it, and the device imports them (VK_KHR_external_memory_fd, VK_EXT_external_memory_dma_buf). A dma-buf pins its pages once and is not tied to the process's page tables, so protection changes leave the GPU's mapping alone, and nothing is copied.
RADV puts every imported memory object into every submission, so pools share their imports: guest addresses are cut into 64 MiB spans (NX2NAVI_DMABUF_MB), and each pool is a buffer bound into its span's memory at its offset. A new span is cut back at the spans already there, so spans overlap only where a pool itself lies across an older span's edge, since amdgpu counts every span's pages in its GTT use. A decommit punches the pages out of the memfd, or zeroes them in place where a dma-buf pins them, and the pages a freed span pinned that nothing committed go back to the system.
It is on by default where /dev/udmabuf opens, udmabuf's size_limit_mb is 1024 or more, the device has both extensions and NVIDIA's kernel module is not loaded. A start-up probe imports one page as the pools will be; any refusal keeps the shadow copies, and the log says why. NX2NAVI_DMABUF=0 keeps the copies. tests/platform_dmabuf_test.c checks the shared mapping without a GPU, and docs/building.md ("Memory") describes the imports and what they need.
RX 6800 XT (RADV, Mesa 26.2.6), one build, interleaved pairs with NX2NAVI_DMABUF=1 and 0: Kokiri Forest 60.7 against 40.3 fps (3 pairs, +51%, standard error 0.4), 1% low 50.9 against 34.0, frames over 25 ms 5 against 768; Hyrule Field 55.7 against 42.4 fps (2 pairs, +31%, standard error 0.4). No pool was left without a span, no span was refused, and --probe-vkmem passes with and without the imports.
The shader preparation modes and the probes return from main, and the C runtime's exit then calls the detach of every library in the process, those other programs put there included. On a player's PC a post-processing injector's Vulkan layer read a null pointer there at the end of every --warm-shaders and --prepare-shaders, four runs over three versions: each printed "variants built into the driver cache", then a FATAL, and ended with code 71, so the launcher logged a finished preparation as one that did not finish. The crash handler now ends such a process with the code its mode returned and one "exit:" line; a fault while the mode still runs, or in the executable's own code on the way out, is reported as before. A crash inside that injector's library also gets the note that the library is not part of the game.
--probe-crash-exit makes the fault from an atexit handler, and tests/exit_fault_check.py checks both cases.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
vk_dlaa_fg_inputs hands out DLSS's new history with the frame's depth and motion, and FSR 4's and NVIDIA's frame
generation pass it on, but vk_fg_frame dropped it: FSR 3.1 takes a cut only in the generation callback, which its
swapchain calls later on its own thread. vk_fg_frame now notes each frame id's cut in a small ring
(runtime/vk_fg_reset.h), and generate sets the callback's reset from it. A frame after one that was not prepared (the
menu open, no inputs) counts as a cut too, and so does an id the ring no longer holds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 8 ms and 16 ms tried in their place left fewer draws out in a scripted walk, but in play a frame that takes 10 to 14 ms longer felt worse to the player than one that is left out, in two sessions on the same install. The wait stays counted once for each pipeline, and a frame that has left a draw out already still waits for nothing more.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A draw that met a new pipeline waited 3 ms for its fast link, in every frame, which is just under what the link takes with the stage libraries in the driver's cache: in a first-visit session on a new install 66 of 87 such draws gave up at the end of the wait and their frame was held back. The wait is now 8 ms for each pipeline, counted from its first miss and shared by all its draws, and 16 ms for a frame, so a pipeline the driver has to compile costs one frame the wait and the frames after it nothing.
The pipeline list tool also takes the compute variants a game folder dispatched, which no pipeline record names and which were translated in the middle of play, and reads a list made before as one of its inputs.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A new install met every pipeline in play: in Hyrule Field one 600-frame stretch of a first visit had 1,769 draws wait 461 ms in all for pipelines, the longest frame 88 ms. The boot now reads seed-pipeline-keys.bin, the release's list, as recorded pipelines after the install's own (keys_open_seed, NX2NAVI_PIPE_SEED, default 1): their pairs go into the vertex output unions and the prewarm builds them as it builds the install's. The preparation builds every listed pipeline (NX2NAVI_PREPARE_PIPELINES now defaults to all), with the unions of the same list, so the first boot of a new install finds them in the caches. The list's mesh pipelines are left out at the boot on a GPU without mesh shaders, as the preparation leaves them out. Nothing of the list goes into pipeline-keys.bin.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A mesh pipeline whose fragment stage reads a location the mesh program never writes now gets a fragment module where
that input is a private variable set to (0, 0, 0, 1) at entry, or to (s, t, 0, 1) of gl_PointCoord when the mesh
program draws points. The fragment shaders recorded with two of the game's point-drawing mesh programs read location
5, which those programs never write; Vulkan leaves such an input undefined.
Checks: tests/fs_fill_inputs_check.py passes; 237 recorded fragment modules rewritten offline pass spirv-val; a
60 s Kakariko run with the change shows the area as before with no Vulkan errors.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A device that reports an axis held from the start (a virtual pad nobody holds, a trigger or pedal mapped onto a stick,
a stick resting far off its middle) turned the camera or moved Link on its own, and since pad_active counted it as
pressed, it could take the gamepad role from the controller in the player's hands whenever that one was idle, and keep
it. input.c now reads every axis through pad_axis: 0 until the axis has been within a quarter of its range of rest
(a trigger's rest is 0), then the device's value. The game's sticks and triggers, the menu's stick readout and the
choice of gamepad all use it. A stick the player holds at the start counts once it has been let go.
Checked: the runtime compiles; new tests/pad_rest_check.py (pad_axis cut from input.c, SDL stood in for) passes: a
held axis reads 0 for 100 polls, a released one reads the device, and axes and gamepads keep separate states;
tests/mouse_stick_check.py still passes. Not run with a real device.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Linux every driver got shadow storage (copies of guest memory beside it), which exists for amdgpu, where importing guest memory costs the write protection of its pages. NVIDIA's own driver imports without that cost, so there the copies are only the slower path: in Kokiri Forest on 0.7.1 (RTX 4090) imports ran at 51.9 frames a second and the copies at 27.4. Until now only the AppImage's play.sh knew, by setting NX2NAVI_SHADOW=0 when NVIDIA's kernel driver was present at build time, so a game started any other way got the copies, and a PC with that kernel driver whose game ran on another card got imports where it should not. The game now chooses by the driver of the device it runs on, play.sh no longer sets the variable, and NX2NAVI_SHADOW=1 or 0 still decides. The driver is asked when the device is picked, before the storage is chosen.
Windows is unchanged: the device probe here reads "mem: mode mirror" with no shadow storage line, and the shadow line with NX2NAVI_SHADOW=1. The Linux path itself was not run: there is no Linux machine here.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The author of the pull request it was ported from closed that request on 2026-11-11 with the words that they give no
permission to use its code. The history (NX2NAVI_VA_REPORT, runtime/vk_vaddr.c, the stack walk it used and its notes
in the docs and the example settings) goes; the lost device's report keeps the breadcrumbs, the pass marks, the video
memory and the driver's fault information.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The author of pull request #353 closed it on 2026-11-11 with the words that they give no permission to use its code. Its port (3134f524) goes: the mark of a pipeline that declares the table (VkPipeInfo.gpa, vk_pipeline.c pipe_add) and of the part that draws with it (vk_mem_gpa_use in vk_exec_draw.c bind_resources and mo_emit), the sync of the table's pools in every such part's prologue (vk_mem_gpa_sync), the submission with a callback in the replay's place as it stood (vk_sync_probe_submit), and --probe-vkmem's read through the table with its shader (test_gpa_read, runtime/shaders/gpa_read.spvasm).
In its place, after pull request #551, which was written without that code: a shadow pool goes into the table and onto a list of the table's shadows (vk_memory.c shadow_table_add); the record of a binding of the table notes the guest submission (vk_exec_draw.c ssbo_info, vk_mem_shadow_table_read); and vk_mem_shadow_sync brings the CPU's writes into every shadow on the list once for that guest submission, before the part is submitted. Once is enough, because the copy-ins read the guest submission's write-watch snapshots, which a second sync would find unchanged. Measured by #551's author on Linux (RTX 4090, Kokiri Forest): 113 and 76 frames over 25 ms in 75 s against 217 and 429 with the port, 43 against 41 fps, the picture complete in both.
What this line had built on the port stays, on the new sync: the test that passes over a pool with nothing written (shadow_pool_clean), the upload of the same pages into a table pool's VRAM mirror, the write-back of stores through the table (NX2NAVI_SHADOW_TABLE_STORES, table_wb_cb, now asked once per guest submission that binds the table) and the counts of NX2NAVI_SHADOW_CHECK. The probes that record their own commands (the lost-device probe, --probe-vkmem's table parts) call vk_sync_submit_own, written anew. vk_mem_shadow_storage stays: one line that returns the mode. Finalized pools in the table were already kept for vk_mem_shutdown here.
--probe-vkmem has a check of its own for the sync (test_table_sync): the shadow of a pool that is never bound keeps its old word through a CPU write and a guest submission that does not bind the table, and holds the new one after a guest submission that does. It passes with NX2NAVI_SHADOW=1 on Windows (RTX 5090), as do the store check with NX2NAVI_SHADOW_TABLE_STORES=1 and the probe without shadow storage; --probe-devlost ends as before. Not run: the game, with or without shadow storage, and anything on Linux.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The probe's programs are written in the shape the translator gives a task or mesh program and need no game files.
With a folder of game translations it converts those instead and builds their compute pipelines.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>