A lost device used to end the process from whichever submit or wait noticed it first, while the present only logged it
and went on and the other threads kept calling the driver. The completion thread could take the submissions' timeline
as reached and run the callbacks of submissions that never finished.
vk_device_lost_check now takes the result of every submit, present, acquire and wait. The first thread to see the loss
sets a flag, writes the report and ends the process as a fatal error does (the log written, a save commit under way
finished), with an exit code of its own; later threads stop in that call, and a watchdog ends the process after ten
seconds should the first thread hang. The replay, its submit, the uploads, the present and the lazy fetch read the flag
and return without touching the device. The completion thread asks the device with a call of its own before a new
value counts and runs no callback once the flag is set.
The executable's --probe-devlost and tests/device_lost_check.py check this without a lost device, and
device_lost_at in nx2navi.ini ends a real run the same way.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Both are about the time between two frames. Ocarina waits for the previous frame's fence before it lets the next frame's GPU work start: its main thread signals an event from the CPU, and the main queue's commands begin with a wait for it.
The completion thread ran a submission's callbacks before it let the submission's waiters go, and the occlusion counts were among them: one vkGetQueryPoolResults and one vkResetQueryPool per reported query, about a hundred queries a frame in Hyrule Field. On an RX 7900 XT (26.10.1) each call goes into the kernel, 3.9 ms of every frame, so the fence was reached 4 to 6 ms after the GPU had finished with every game thread waiting. Slots come from the ring in order, so one call now reads up to 256 slots from the one asked for, the available results are kept for the segments that follow, and the slots taken are reset in runs (vk_exec_xfer.c oq_take, NX2NAVI_OQ_BATCH). Standing in Hyrule Field at 3440x1440 with FSR: 52.7 fps against 44.6, the longest frame of the measured minute 46 ms against 55; the completion thread spends 0.3 ms a frame in the callback.
When the CPU's signal came a millisecond or two after the main queue's submit call, the submission took its copies, stopped at the wait and was held; the render thread polled the event, presented, and the present replayed the frame under a value the game's fence does not stand for. The next frame then skipped its wait for the GPU and was replayed inside its submit call, and frames of 23 and 15 ms took turns. A submission that begins with a wait for an event the CPU has signalled before now waits for the signal in the submit call, before its copies are taken, 10 ms at most (vk_exec.c vk_exec_gate_wait, NX2NAVI_GATE_WAIT_MS): 54.5 fps and a longest frame of 33 ms in the same place, 58.6 fps with the recorder thread on. With a camera sweep the pictures went out 21.0 ms apart on average with a 99th percentile of 55 ms, against 25.4 and 64. A village went from 48.6 to 53.7 fps and a graveyard from 56.5 to 64.2. 900 consecutive frames of a still scene show no dark or changed frame, with the recorder thread and without.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Each frame the game sets one of two events to 0 from the CPU, submits its worker queues and the main queue, whose first command waits for that event to be 1, and signals 1 with nvnEventSignal once it has written what the frame's commands read. The recording of a submission takes the VRAM mirror sync and the uniform snapshot at its start and tested the wait a moment later, so a signal that arrived in between passed the wait and the frame was drawn from copies taken before the signal: single frames with the whole scene's shadowing missing (6 to 7 of 255 brighter, the sky and the interface unchanged) or a black scene, more of them the busier the CPU.
Every nvnEventSignal is now numbered and keeps the value it replaced, the recording notes the number before it takes its copies, and a signal with a later number does not count for that submission. The submission stops at the wait like any wait reached before its signal and goes on at once in a submission of its own, whose copies are taken after the signal. NX2NAVI_EVENT_AT_SYNC=0 gives the earlier test. NX2NAVI_EVO_TRACE=1 logs the order in which every frame's submissions, waits and signals were replayed, which is how the frames were found.
Measured with the flight recorder's thumbnails, 1,800 presented frames a run, standing in Hyrule Field by day with the lazy write-back on, DLSS Performance, frame generation 4x and HDR: the build before this change showed 34 single-frame spikes with eight busy threads beside the game, 20 with four and 8 with six, most of them in frames whose signal came during the replay; with this change two runs with eight busy threads showed 0 and 0, with 667 and 724 such frames. A session of the user's with the lazy write-back off ran 25,200 frames without a spike in its three automatic dumps while the change stopped 4,744 wait tests.
Not measured: runs of this same build with the switch off (the two tried failed for other reasons, a device loss during the boot beside another game and a load with cold caches), any run with the lazy write-back off under load, the frame rate and GPU time against the switch off, and the differential tests, which this change does not touch. The flicker with the lazy write-back off that was reported has not been reproduced in a capture; its mechanism is the same signal and occurs there too.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A session with per-pass GPU timestamps showed the game's own render passes and compute dispatches at 1.6 to 2.3 ms of GPU time a frame while the perf line's GPU time was 7.7 to 13.5 ms: most of the frame's GPU time is in commands the replay records between the marks, and nothing said which.
The timed frames now also get a timestamp at the start and the end of each part of a submission, and every stretch between two marks (or a mark and the part's edge) is a gap. The replay names the kind of commands it records in a gap: buffer and texture copies, fills and clears, the temporal upscaler, write-back merges and copies, depth twin copies, the texture pool marks, the mesh draw counts and block-compressed target copies, each with a size where it has one. The report adds the gaps' time per frame summed by the kinds noted in them (a gap nobody noted holds barriers and state only), the parts' own time for comparison with the perf line, and each kind's count and size per frame. Nothing runs outside the timed frames.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
From the local fixes package (fix 11), measured in Kokiri Forest at 4K on an RTX 4090 over PCIe 3.0: 40.5 to 62.6 fps
native and 45.6 to 74.9 fps with DLSS Quality, the mirror copies from 6.5 to 0.3 ms of GPU time a frame.
- Pages a VRAM mirror wrote back into guest memory are not copied up into it again: the mirror sync reads page stamps
of their own that a mirror's write-back leaves alone (vk_gpu_written_back, decided at record time by
vk_mem_mirrored). This replaces the lookup table NX2NAVI_WB_NO_REUP had, keeps its name, and is now on by default.
- The game's flushes of mirrored memory are not copied where the write watch covers the region, as every page the game
wrote there is stamped already (47-57 MiB a frame); NX2NAVI_FLUSH_UPLOAD=1 copies them as before.
- Written storage ranges up to NX2NAVI_WB_MERGE_MAX (now 32 MiB) are merged in 4 MiB pieces, so a range is no longer
bound to what one dispatch covers; the snapshot pool keeps its budget-scaled size with 128 slots.
- A submission's first part goes to the GPU at the first pass boundary while the GPU is idle (split_idle), long passes
are split too (split_in_pass), and the part's time limit reads the clock less often.
- A render target that needs an image of its own takes the image of a released one with the same parameters, released
two or more frames before (tex_reuse); memory requirements are kept per parameter set; a released entry another
descriptor's entry replaced at its address is kept and taken back.
- The memory check reads a page's write-watch bit and stamp under the lock that keeps them together.
- The launcher reads the gamepad only while its own window is in front, and does not start a second copy of the game.
- make_game.py says why a Python package did not load, not only that it is missing.
The fix's per-pass GPU timing is left out: NX2NAVI_GPU_PASSES does the same in main.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lock_time (NX2NAVI_LOCK_TIME) makes the replay threads time each lock they have to wait for, kept per asking site (platform.h plat_lock_timing, plat_mutex_lock_at and plat_rmutex_lock_at carry the site), and the frame-time summary lists the sites with the most wait, the recorder's full-ring and drain waits, and the texture lock's holds over 1 ms with the code that took it. replay_time now also keeps each frame's replay ops: a frame over 25 ms charged to the replay says how many draws, binds and transfers it replayed, their time, its slowest op, and how much of the replay's wall time the thread was on the CPU. NX2NAVI_PROFILE_REPLAY=1 samples only the replay thread without call chains, and the profiler now follows the thread that replays instead of the one that started the back end.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Large compute shaders that no cache holds are built on one background thread after the boot prebuild, smallest first, so their first dispatch no longer compiles them on the replay thread. NX2NAVI_BIG_BUILD_MAX_MB stops that builder once the process has grown by that much: on a 31 GB machine one of these shaders took the process up by 12.8 GB and the game stuttered while Windows paged. The builder takes the queue under its lock and copies each entry out before compiling it, so a shader queued meanwhile cannot free the arrays under it. The same change reports what pipeline-cache.bin gives back, and starts over when the file grows past NX2NAVI_PIPELINE_CACHE_MAX_MB.
A tensor-core program is now tested before its shader is asked for, in both dispatch paths. It runs the stand-in while NX2NAVI_NATIVE_TENSOR is off, and translating it anyway recorded four AI-upscaler shaders of 0.87 to 4.4 million SPIR-V words that this game never dispatches natively. Process memory went from 29.3 GB to 24 GB and the multi-gigabyte builds during play are gone.
The frame-time summary says which part a slow frame went to, and with NX2NAVI_REPLAY_TIME how the replay splits between draws, binds and transfers; NX2NAVI_SVC_SLOW_MS logs the kernel services that block, with the thread they blocked on. prof_report.py reads the map this build writes and adds the image base to a runtime Rip, so a profile resolves to function names instead of question marks.
build_xlate.py applies the patches with GIT_CEILING_DIRECTORIES set to the build folder's parent. Inside a checkout git read 0021-prmt-byte-permute.patch's paths from the repository's top, skipped every file of it and still exited 0, so PRMT was missing from locally built translators and every byte permute was a no-op. A Skipped patch line now fails the build. The note in AmpereLowering.cs that said the translator emits nothing for PRMT is corrected.
Ocarina runs compute work on a second queue that trades results with the main one through events: the worker's command buffers wait (WaitEvent) for signals the main queue gives in its submission a few submits later, and the main queue waits for the worker's signals in turn. Replayed one submission after the other, the worker's steps ran before the work they wait for and read the previous frame's results, and the characters came out much too dark at any time of day (Kokiri Forest: about 13 waits a frame, 7 of them reached before their signal).
A top-level wait whose condition does not hold in replay order now stops its submission: the rest is held, with private copies of its recordings and the queue's state, and the queue's later submissions line up behind it. After each replayed signal the held work whose wait now holds goes on inside the command buffer being recorded, between full barriers, with its own state swapped in. Work still held after 36 submissions goes on regardless, as before (counted in the NX2NAVI_TRACE=perf line). The queue reaches the back end as gfx_submit_queue. Only with the replay on the submitting thread; NX2NAVI_EVENT_ORDER=0 turns it off.
Tested by play in adult Hyrule Field with the tree fix of the previous commit: Link lit normally at dusk and at noon, trees steady, lighting otherwise unchanged. The white specks in the grass remain (they predate both changes).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The culling passes count each frame's visible foliage with atomics through the device address table into small records (an indirect draw's count, then 1, 1). Stores through the table already reached both the VRAM mirror and its guest copy, but atomics reached the mirror only, so guest memory held the CPU's template of zeros. The CPU writes other words of the same pages every frame, and the sync of such a page copied those zeros over the counts: in Hyrule Field the trees flickered, went missing or stretched into columns reaching the sky. NX2NAVI_MEMORY=verify found 5 to 10 such pages a minute, all in ten pools of 20 and 44 KiB.
Translator patch 0036 repeats every table atomic on the twin (a compare-exchange with the same operands), leaving its result unused so the driver need not wait on guest memory. A buffer fill into a mirror now fills its guest copy at once (vk_mem_guest_twin), so a counter the GPU resets that way counts on from the same value in both; that mismatch is why table atomics stayed out of the twin before. The memory check's mismatch lines now also give the pool's last written binding, its last use, the CPU's write stamp for the page and the first bytes of both copies.
With the change the trees stayed steady at the spot where they flickered, all 37,848 shaders translated without a failure, and the GPU time per frame stayed at about 9 ms (75 fps uncapped).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Small ranges the game invalidates and reads back (the exposure meter) are copied from guest memory into the mirror at every submission that uses their pool, so a CPU write made between recording and the GPU's run is not lost; NX2NAVI_READBACK_SYNC=0 turns it off.
While the boot builds the recorded pipelines a marker file exists next to the pipeline cache. A start that finds it left over records the GPU and driver in pipeline-prewarm-skip.txt, and starts on them build those pipelines at their first draw instead.
CommandBufferWaitEvent is now recorded as OP_WAIT_EVENT. The replay waits for nothing but counts the waits it reaches before their signal, and how many submissions later that signal comes, in the NX2NAVI_TRACE=gfx frame line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each file that contains copied code names its copyright holder and license and points at THIRD_PARTY_NOTICES.md; zlib code keeps its full notice and the mark that it was altered. Comments and probe text only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comments say what the code does and why, docs/maxwell.md keeps the Maxwell facts, and fetch_refs.sh clones only the checkout the build reads. Code is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A texture slot whose content the CPU writes now follows another image at its address only when that image differs in
swizzle alone or holds GPU-written content; Ocarina reuses freed glyph atlas memory for captions of another width, and
following the older image drew the Basic Controls screen's "Target" with the wrong row length. Mesh-shader pipelines
are built on the workers and recorded in pipeline-keys.bin (version 4 with the task stage; version 3 files are
converted) instead of stopping the frame for 0.26-0.51 s. The ASTC decoder reuses one scratch buffer per upload command
buffer, and CPU-provided textures up to 32 MiB are sub-allocated unless the driver requires dedicated memory, which
takes the scene-load frame from about 1.25 s to 0.7 s. The feedback scan looks only at bound texture slots instead of
the bindless pool arrays, per-draw knob lookups are cached and the texture address buckets grow to 65536, so Kokiri
Forest's replay falls from about 16 to 7-9 ms a frame and the forest holds 60 fps. A float clear of zero on an integer
image now clears instead of being skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The texture pool descriptors are now a ring of sets, one per submission, brought up to date from a change log, and the first submission of each frame waits for its GPU work before the game goes on: the GPU ran late and part of a frame read data of the next one, and distant characters drew NaN that bloom spread over the frame. The window acquire also waits for the image's previous frame. Function locals start at zero unless DLAA is on, which removes the per-warp light tiles on the Great Fairy Fountain floor and Link's dark, speckled look. Texture slots follow only live entries of their own target, so the volumetric fog gets its 3D volume. The pool refresh walks only the slots that changed.
The translator implements PRMT (patch 0021), which it only logged before; the button prompts render again. HFMA2.RELU no longer depends on it, HMMA.F16 rounds once and HMMA.F32 is lowered, checked against GPU vectors with a new Maxwell warp emulator.
The frame dump gains a re-arming white-frame trigger, a trigger on a depth target's first pass, NaN pixels in magenta, image handles and view targets, and depth copies of sampled depth textures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The NVN2 front end and the SM86 to Maxwell shader lowering (tools/xlate/AmpereLowering.cs, mesh and task stages,
bindless texture pools, BRX jump tables, the NVN2 compute c0 layout) with four translator patches (0017-0020) bring the
opening of the game to 58-60 fps on the Vulkan renderer.
This round fixes the haze: SM86 IPA is lowered as a plain Maxwell IPA.PASS (texture coordinates were corrected for
perspective twice); LDSM, TXD, both F2IP byte forms, HMMA, MOVM and HFMA2.RELU are lowered and spill mode has eight
uniform register groups; conditional rendering keeps the newest counter report per address and treats a report as
pending until its value is in guest memory, so occlusion tests and the depth clears around the depth copy behave; and
the upscaler stand-in decodes the post pass's sRGB picture before writing the linear target. Tensor-core programs keep
the stand-in unless NX2NAVI_NATIVE_TENSOR=1, because their lowering does not match the hardware yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>