cheats.c rebuild_if_needed runs in rt_boot after the prewarm thread has started, and ended the process with only
prof_write and rtss_limit_restore. By then vk_pipeline.c may have written pipeline-prewarm-running.txt (pm_set) and
opened the boot's journal of library builds; left in place, the next start took the rebuild's end for a driver crash
in the boot's builds (pm_skipped counts it as a quiet end; three of them skip the prebuild). The rebuild.skip file
keeps it to one quiet end per rebuild.
The rebuild's end now takes the orderly end's guard (rt_exit_begin), puts RTSS's profile back, clears the prewarm
marker, closes the boot's journal through the new vk_pipeline.c vk_pipeline_boot_abandon (journal_close, as
vk_pipeline_save_exit does), writes the summary and ends. No save is mounted and no menu has run yet, so the saves and
nx2navi.ini have nothing to write. A narrow race stays: pm_set writes the marker and then sets pm_on, so a clear in
between leaves the file; the window's close has the same one.
tests/exit_order_check.py checks that order in rebuild_if_needed. Not reproduced before the change;
exit_order_check and the compile checks of every runtime file pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
completion_main calls vk_mem_collect after it publishes each new done, and vk_mem_collect waited for cmd_lock. An
upload flush holds that lock across vk_queue_lock, behind a present call (FIFO, a frame limiter inside the present,
frame generation) or another long queue-lock hold. With wait = 1 it also holds it across its batches' fence waits,
and stg_create waits for batch fences under it. The single completion thread blocked there could not process the next
timeline value, so the next done, and every fence waiter for it, waited for that chain.
vk_mem_collect_try only tries cmd_lock. When the lock is busy it still destroys the finalized pools, which need only
retire_lock, and leaves the deferred objects to a later completion or to the next upload flush's own collect. After
eight busy tries in a row it waits as vk_mem_collect does, which keeps the deferred list bounded, as before. pool_fini
keeps the waiting vk_mem_collect. The trace=perf "mem: sizes" GPU line counts the collects that gave way.
Not reproduced before the change: the completion thread's lock waits are not timed. Checks that passed:
pass_cc_check on every runtime file (vk.h changed), no new -Wall -Wextra warnings in vk_memory.c and vk_sync.c,
sync_hw_check, gpa_check, mem_budget_check, tex_release_check, mem_growth_check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
vk_astc_scratch_trim runs at every present and held the decoder's lock while vk_defer_buffer waited for cmd_lock.
The upload path takes the two locks the other way round: vk_upload_begin takes cmd_lock, then vk_astc_record and
scratch_get take the decoder's lock. A present that trimmed while another thread was between vk_upload_begin and
scratch_get therefore hung both threads. Without the hang, the present still waited for as long as an upload flush
held cmd_lock through its fence waits. vk_memory.c gets vk_defer_buffer_try, which defers only when it gets cmd_lock
at once. When the lock is busy the trim leaves the scratch as it is, and the next present tries again.
tests/astc_trim_check builds the real vk_astc.c with platform/common.c's recursive lock. It checks that the trim gives
way while another thread holds cmd_lock, that it goes through once the lock is free, and that the forced interleaving
ends. Before the change the check reported the deadlock (exit 2). Checks that passed: astc_trim_check, pass_cc_check
on every runtime file (vk.h changed), no new -Wall -Wextra warnings in vk_astc.c and vk_memory.c, and the pass's
test matrix (35 ok; the 6 failures are the baseline's).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In the 2026-11-13 session frames of 60-96 ms came 2 to 6 frames after each area load with the render thread waiting 59-86 ms for the GPU, and no line said what the GPU was doing. With trace = perf such a frame now gets a line at the next present: the timed parts that ended in its interval, their GPU time, the longest gap between two, and the longest parts with their kind (draws, copies into VRAM mirrors, pre-passes, uploads, presents) and their time in the queue. The completion thread keeps the last parts in a ring as it reads their timestamps; nothing waits for the GPU. Measurement only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A scratch grown past 32 MiB for a large texture went back to the usual size only when the next ASTC upload asked for one, so in the 2026-11-13 session the decoder held 224 MiB for the 70 s after the first area load and from a later load to the session's end. Each present now gives it back once no texture needed more for 5 s; nothing is made then, the next request makes the usual scratch, and the buffer is freed once the upload batches recorded so far completed, as before. The present only tries the decoder's lock, which can be held while VMA makes a block. astc_scratch_trim = 0 keeps the scratch until the next upload.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The session of 2026-11-13 ended 8 s after its window lost the focus, after 7 minutes of play, with none of the game's exit lines: no frame-time summary of its last period, no stutter map of its last minute and no growth ranking at all, as the ranking came every 10 minutes. When the window loses the focus in play, the present thread now writes the frame-time summary so far (starting a new period), the stutter map and the growth ranking at its next present, at most once a minute; formatting only, nothing asks the device. The growth ranking comes every 2 minutes of play instead of 10, the stutter map every 5 instead of 10.
A process that ends through exit() rather than plat_exit now writes the log thread's queue last, after the frame-time summary's exit handler, instead of losing it.
The launcher logged a game's end only for a nonzero code, so the session's end with code 0 left no line anywhere. It now logs every end of a game it started, and when the game's log has no line of the game's own end (the window's close, the game's ExitProcess, a timed run's end, a FATAL or exit line), it notes the end there too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every RX 6000 loss whose pass marks named a pass was in one that starts with a native mesh draw, on every version from 0.6.0 to 0.9.2 and with every other setting tried each way; RDNA 1 cards on the same driver draw through the fallback and RDNA 3 and 4 on AMD's other driver draw natively, and neither has such a report. That the native mesh path is the cause is not shown. The decision moves into vk_mesh_path.h with a test of each device kind and setting; has_mesh_shader keeps saying what the hardware has, so the recorder thread and the low-end card test do not change. On those cards no native mesh pipeline is built any more: the seed list and the boot prewarm skip them. The early tail stays off with the fallback, a combination never measured. The fatal line names the mesh path, and NX2NAVI_VK_NOOPT's default applies with mesh_fallback = 0 alone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A pass mark now counts its mesh draws and those that take their count from the GPU; the gap before a mark notes a group recorded by the workers, whose passes have no marks, where the report said "barriers and state only"; the fatal line names the mesh path (native, the mesh fallback, the emulation). The comment on the RX 6700 XT report in vk_exec.c called the last marked dispatch the replay's own: it is the game's (op_dispatch, cs d7566eb1).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The RX 6000 reports of 0.9.1 and 0.9.2 ran the completion callbacks of submissions the GPU never finished before any call returned VK_ERROR_DEVICE_LOST, although the completion thread polled the device first. Every value a queue hand-over or a host signal gives the submissions' timeline or the executed timeline is now noted before the signal (vk_sync_hw.h); the completion thread and the wait after a submit end the run when a timeline reads above that note. The first occlusion query found unwritten after its fence logs its submission, the submissions' values, both timelines' readings and the GPU breadcrumbs, and the lost device's report starts with the same line, so the next report shows whether the driver gave a value above the note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every report adds a line for video memory outside VMA (the driver's usage and budget per heap against VMA's blocks, and counts of pipelines by kind and by last use, stage libraries, compute pipelines, layouts and descriptor pools, with the swapchain, DLSS and FSR sizes), a line that splits the process's commit by owner (the game's heap and maps, device-local memory, VMA's host blocks, the C runtime heap with the shader caches, pipeline tables, texture cache, log queue and flight recorder in it, the driver's pipeline cache data and the rest), and a line naming what grew by more than 64 MiB since the report before and since play began. Every ten minutes of play and at exit a ranking line orders the categories by growth and marks those that only climbed; tests/mem_growth_check.py checks that bookkeeping.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
F8 now always writes a folder under logs/captures: the area, Link's position and rotation, the camera of the frame on screen and the game's own, the presented frame as a PNG, frame times around the press, the machine, the settings that differ from the defaults, the end of the log, the save, and the game's flags and texts. A note on screen names the capture. The flight recorder's F8 dump goes into the capture's folder.
NX2NAVI_CAPTURE_REPLAY starts a run on a capture's settings, teleports to its area, walks Link to where he stood and puts the free camera where the camera was; tools/capture_run.py prepares the profile from the capture's save, runs it and compares the two captures. docs/bug-capture.md has the files and the steps.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The line freed the upscaler stand-in's scratch images, FSR 1's and SMAA's images and the DLSS sizes on its own while
this pass did: its versions stay. Of this pass's versions, what they do not cover stays: the stand-in's views of a game
image it read as it is now go with that image (vk_tonemap_forget from vk_exec_image_gone, under a lock), and the
upscaler release frees only what the line's idle release leaves, the reactive mask's snapshots, the FSR upscale
contexts and the DLSS sizes after a failure, at the same 600 presents and with vram_use = high keeping them as well.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each made its images at the picture's or the scene's size the first time it ran and kept them until the process
ended, also after the menu switched it off:
- the HDR encode's picture and menu image (28 and 14 MiB at 2560x1440, 63 and 32 MiB at 2160 lines) now go when the
swapchain is made without HDR, after a wait for the presents that may still read them;
- FSR 1's two images (56 MiB, 127 MiB at 2160 lines) go after 300 presents it did not scale, with the same wait;
- SMAA's three work images (35 MiB, 79 MiB at 2160 lines) go after 600 frames of another anti-aliasing, deferred past
the submission being recorded.
Each makes them anew when it runs again; one log line says when they go.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The scratch images of the fallback blit, one per size and format, were kept for the process, the first eight ever
asked for. Their sizes follow the scene's resolution, which DLSS modes, the dynamic resolution and a window or
picture size change move, at up to 32 MiB each at 2160 lines; a ninth size got no scratch and its blit was left out.
A new size now takes the place of the one used longest ago, which goes once the GPU is past the submission being
recorded, and never one the submission being recorded uses. The cache is under a lock, since replay workers record
blits side by side.
The tonemap pass's views were kept by image handle for the process, a game image's view naming a destroyed image that
a new image could get the handle of, and no view at all past sixteen. A view now goes with its image: at a scratch
image's replacement, and with a game image through the replay's image-gone hook.
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>
Every VMA allocation is counted under the file that made it until it is freed, so the new "mem: by owner" line
splits the device-local memory the driver reports into textures (live, GPU-drawn, idle, released, released too
lately to drop), pools and VRAM mirrors, upscalers, replay buffers, uniform snapshots, the device address table,
the ASTC decoder, present and post-processing, headroom spares, the free space inside VMA's blocks and what VMA
never allocated. "mem: host" gives the host-visible allocations, the process's commit and working set and the
game's heap. "vk: released textures" counts take-backs by how long the texture was released and textures made
again after a drop by a bound or a pressure, with the drop's reason.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The larger pool asks for 655,360 sampled images in one set where 524,288 were asked before. The start now logs the
device's limits for one set and one stage next to that count, warns when the count is over them, and a refused
layout ends the start with the limits named instead of a bare failure.
The log line for a slot registered past the pool names both slots a shader reads instead: the pool's last entry, and
the last of the old length for a translation carried over from the translator before. The comments say what they say
without an issue number.
The translator's identity changes with this pool size: after the update every install translates its shaders again
in the background, the earlier translations serve meanwhile, and a shader gets the ground fix once it is translated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After an update the game keeps the previous translator's translations until a background pass translates them again. Those still declare texture pool arrays of 32768, which spirv_pool_rebind no longer took for pool arrays once VK_POOL_TEX grew, so their textures were never bound and areas loaded with empty images. They now bind to the first 32768 entries of the larger binding.
The game makes a texture pool of 33068 descriptors, and the translator clamped every pool handle to 32767, so the textures registered from 32768 up were drawn with whatever slot 32767 held, often an empty image: ground went see-through after an area load until the area was loaded again. VK_POOL_TEX and the translator's TexturePoolLength go to 40960, and a slot registered past it is logged once.
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>
A mesh pipeline (task and mesh stages: the foliage, the grass, the terrain) had no stage libraries. Each new one was
a whole build, 0.2 to 1.1 s the first time foliage of a kind appeared, and its draws waited or were left out of the
picture until it was done.
Where the graphics pipeline library is on, a mesh pipeline is now linked at its first use from three libraries and
its whole build follows behind, as for a vertex pipeline. Its pre-rasterization library holds the task and mesh
stages and is keyed by both shaders. No vertex input library is linked: a mesh pipeline has no such state. Its layout
is a stage-independent pair, the mesh and task stages' bindings in set 0 and the fragment stage's in set 1, so the
fragment shader library can be the one a vertex pipeline uses. That library's key now also carries what a mesh
pipeline's fragment module gets filled in and whether the shading rate is dynamic state, which the pre-rasterization
and the fragment shader state both hold. The fragment output library is shared as it is.
The fast link on the library workers, the wait of a draw for its libraries and the pipeline hints take mesh pipelines
in. The summary counts them by themselves: fast links and their time, how long the ones a draw missed took to become
usable, and the libraries made for them.
NX2NAVI_GPL_MESH=0 (gpl_mesh in nx2navi.ini) keeps the whole build. It is also kept without the pipeline library, on
a device that draws no mesh pipeline, and where mesh pipelines alone are built without the driver's optimiser; the
start logs which is in use and why. The mesh emulation of the low-end GPU mode is not touched.
A mesh pipeline with libraries is another pipeline to the driver than the one built before, so an install's recorded
mesh pipelines are compiled once more at the first start with this change.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
With the setting off the device is made exactly as before the mode existed, so the mode cannot change anything for a player who never turns it on. Turning it on in the menu then takes a restart, and the row says so; in a run that started with it on, changes still apply at once.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
NX2NAVI_LOW_LATENCY (0 off, 1 on, 2 on with boost; live) is for low latency without frame generation: NVIDIA's mode
through VK_NV_low_latency2 and AMD's through VK_AMD_anti_lag, each enabled on the device only where the driver lists
it, NVIDIA's with VK_KHR_present_id, which it needs, and AMD's with its feature. A device made by the DLSS Frame
Generation proxies gets neither: that path keeps its own low latency. vk_lowlat.c holds the calls, the frame numbering
and the statistics line; nothing calls into it yet beyond the start and the end of the device.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The mark counter and the ring were written as if only the replay thread made draw marks, but a pass's chunks are
recorded by the replay's workers side by side, and each of them marks its draws. The number is taken atomically now,
and each number has its own slot of the ring. The report says what the numbers mean with workers on: the two marks
the GPU names are exact, the numbers between them are in recording order and not all later in the queue; the mark
the GPU last reached is always printed, also when the list does not come to it.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The table still named a new mirror at record time, as the pages' twin: parts already handed to the queue wrote
through the table into the new block before its first copy, the same early access the delayed read was made to stop.
The reads then moved to the mirror some submissions later by a second change that took a second record, cleared its
note before it knew the record was there, and until it ran left the table's readers on guest memory, without what
other shaders had written into the mirror through a buffer binding.
Nothing is changed when the mirror is made now. The prologue that holds the mirror's first copy also holds a store
into the pool's record, after the copy: what was queued before it reads and writes guest memory through the table as
if no mirror existed, what comes after reads the mirror and writes both. No second record, no second change, no
window in which the table is behind. gpa_late = 0 still changes the table at once, as up to 0.9.2.
--probe-vkmem makes a mirror in a recorded submission and reads the table's own memory back: the page names the
mirror with guest memory as the twin, and the mirror holds the pool's bytes. tests\gpa_check.py runs the store as the
GPU would, before and after. A table with no record left now says so once in the log.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A scene draw whose vertex shader reads a matrix palette and whose vertex layout feeds its joint inputs unsigned integers is drawn a second time into the motion vectors with last frame's blocks and camera, as the player's meshes are, with the same depth test and the same guard for bones that turn sharply. Which draw of last frame it continues is decided when the vectors are made, so the replay's threads cannot change it; a draw in doubt keeps the camera's motion. The player's own draws stay with their ownership records.
skin_motion (NX2NAVI_SKIN_MOTION) is 1 by default and 0 turns it off. The motion status line now counts the skinned draws seen, those drawn with their own motion and the others by reason. It has compiled and passed its checks; it has not been played.
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>
A ClearBuffer or ReportValue is a fill the GPU makes where it stands in the stream, and guest memory has its bytes once the GPU got there. A counter report reaches guest memory as a CPU write after its submission's fence, and a render condition reads the record when it is replayed. So a report followed by a clear of the same record ended as the report (in one submission always, across two when the completion thread was behind the GPU), and a condition replayed after the clear read what memory held before the fill, or the 0 of the report the clear had overwritten.
A clear of at most 256 bytes into CPU-visible memory is now also applied at its place in the replay to the records it covers that a counter report went into: the covered bytes get the value in guest memory at once, reports replayed before the clear leave those bytes alone when their write comes, a condition reads them from memory, and a report replayed after the clear writes the whole record as before. The GPU's fill is recorded as before. Other small clears stay the GPU's fill alone: the GPU reads guest memory too, at once in an imported pool and through the next upload of the written page in a mirrored one, so a write ahead of it into a draw count or an atomic counter would reach passes that still want the old value. NX2NAVI_CLEAR_SMALL_CPU=0 turns it off.
Whether the game clears counter records is not known. The pass timing captures of 75 scripted runs show one buffer fill a frame and nothing else about it, and the frame dump does not list buffer fills. The periodic vk_usnap line now counts the small clears, the last one's size, value and address, how many covered a record with a report, the reports into bytes cleared before, and the report writes that left cleared bytes alone.
Checked with a build of the dev exe, the launcher tests, the device address table check, and a scratch link of vk_exec_xfer.c against stubs that replays reports and clears in both orders. Not run in the game.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The game waits each frame until the GPU has finished the frame before. In Hyrule Field with FSR 4 the last of that
work is one part of 4.7 ms, the temporal upscaler and the interface drawn after it, then 1.4 ms of write-backs, and
what the game reads of the frame is done before the upscaler runs. With NX2NAVI_TAIL_EARLY the replay records that
tail into a command buffer of its own, and when it hands the CPU nothing but timestamps, writes no pool memory but
inline uniform updates of mirrored blocks and reads no guest memory late, the write-backs go to the queue before it
with the executed timeline's signal. The game's waits end there, and the GPU runs the tail while the next frame is
prepared. The GPU's commands are the same; a tail that fails a test goes as before. Off by default.
RX 7900 XT, Ryzen 7 5800X3D, 3440x1440, FSR 4 Native AA, one batch, replay workers off: 50.7 against 42.5 fps
standing and 56.0 against 45.9 along the field route; with FSR 3.1, 51.8 against 47.8 and 54.9 against 52.3. With
three replay workers, FSR 4, standing: 63.5 against 49.4. The screenshots matched those without it standing, along
the route and through warps to five areas, with no black frame outside the loading screens of both.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The stand-in for the game's upscaler enlarged the scene with a linear filter, which is softer than the console's picture. It now enlarges with the edge-adaptive filter the present already scales to the window with: the same shader, built a second time with a constant set so it squares what it filtered, as the linear pass does. The stored square-root values are filtered as before and the decode is the same. standin_filter = bilinear brings the linear pass back, and the setting is live. The linear pass also stays where nothing is enlarged and while DLSS or FSR 3.1 runs, since that upscaler scales the scene itself and its picture goes over the stand-in's. The movies do not go through this pass and are not changed.
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.
A pass mark names the pass or dispatch the GPU had started and the last it finished. When the two are the same mark the GPU stopped in between, and the report named nothing: in a player's log of 0.9.2 (RX 6700 XT) both were one compute dispatch, with 51 later marks made and none started. The report now adds the mark that never started and the kinds of commands the replay recorded between the two (copies, clears, the upscaler, write-backs), which the gap notes already name for the GPU pass timing and now keep for every mark.
draw_crumbs = 1 (off by default, AMD buffer marks only) also writes a mark wherever the shaders or the pipeline state change inside a pass. The device losses reported on RX 6600 and RX 6800 class cards end inside one scene pass of 30 to 100 draws and the draw that hangs is not known; with the setting on, the report lists the shaders of the draws the GPU had not finished.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A registration the game made while a submission was being replayed changed the slot table under that submission,
whose texture pool set is written as the replay goes and read when it runs. Such a registration now waits and takes
effect once the submission has gone to the queue; with the replay on its own thread it goes into the replay queue
ahead of the next item. NX2NAVI_REG_ORDER=0 restores the old behaviour. In a town session the white kerb, stair and
wall-top edges of single frames were gone with it; the log counts the registrations that waited.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Conflicts: the settings key lists take both sides; frame generation keeps the camera-cut handling and logs its counters every 600 frames with the cuts among them; the pass-time mark moves into draw_viewport; mo_desc keeps the pose depth and notes the reader.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
At every submission that binds the texture pools the replay checks each listed texture for CPU writes, and takes the write-watch snapshot of the 2 MiB regions that have none yet as the list comes to them: one query a region, since the list is in no address order. Standing still that is 5 to 10 queries and the check takes 0.11 ms. While the player runs through Kakariko or Castle Town the game writes about 32 textures into new storage every four or five frames, a region the CPU wrote is not frozen again for 1,800 submissions, and the check asked for 100 to 500 regions one by one, 2.7 times a frame: 0.5 to 1.5 ms a check, about 3 us a query, in five sessions of one player. The regions are now collected for the whole list first and asked in one query per run of adjacent regions. On this machine 400 queries of 2 MiB take 727 us and one query over the same 800 MiB takes 360 us. The same regions are queried and the same pages stamped, all of them before the list is looked at, as each entry's were before its own look; a run across the edge of a watched reservation is left to the queries by range. pool_presnap_runs = 0 asks region by region as before.
A released entry stays in the list until it is destroyed, the check does nothing for it, and the stamp it compares stays as it was at the release, so one whose region took a write later missed the region check at every submission from then on and was fetched only to be passed by: 1,000 to 6,300 entries a check, with up to 8,192 released entries kept. The list now carries the released state, set at the release and cleared at the take-back, and the check passes such an entry by without touching it or asking for its regions.
The pool log's freshness line counts the released entries passed by, and its snapshot queries count the queries either way. --probe-vkmem has a case for the queries by run.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
gpl noted them in ssbo_info, which the bind skip and the kept descriptors leave out once they have warmed up: the
submissions were taken as reading no guest memory again, and the grass and moss flickered again after a minute of play
(every submit wait of a 2100-frame run ended without the GPU). gpa_reader looks once per pipeline for a table binding
and notes the submission at each bind_resources and mo_desc.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The heap limits were fixed at device creation, and only texture images looked at the budget. FidelityFX allocates
outside VMA (FSR 3.1 contexts, frame generation): 882 MiB of it with 4644 MiB of VMA's filled the 5526 MiB budget of
heap 0 while heap 2 had 1757 MiB free, and an upload submit failed with VK_ERROR_OUT_OF_DEVICE_MEMORY. vmaCreateImage,
vmaCreateBuffer and vmaAllocateMemory now go through vk_vma_budget.c, which leaves out a device-local heap whose
driver-reported use is within NX2NAVI_VRAM_RESERVE_MB (default 512) of its budget; a Lama patches row sets it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each draw's indices are now a range of its own and its task records too, reserved by an allocation shader (one
invocation, mesh_emu_alloc_spirv) for the run of draw records queued: before the task stages the task records and the
index ranges of draws without a task stage, after them the index ranges of the others (mesh workgroups with output,
which the task stage now counts in the record, times OutputPrimitives x vertices per primitive). The mesh stage takes
its indices from the draw's range with an atomic in its record; vertex slots stay the arena's, as the indices name
them. So no draw's stages need another's, and the queue (MEQ) records the fills as the draws come, then at the
pre-pass command buffer's end (vk_sync.c pre_end -> vk_exec_mesh_flush): allocation, every task stage, allocation,
every mesh stage, a barrier between each. One by one each draw took three barriers and copied the arena's counters
after the draw before it: an RX 580 spent 5.5 ms of a frame's GPU time on 69 draws' stages in Kokiri Forest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On 0.9.1 this is the release's own mirror capture (NX2NAVI_MIRROR_CAPTURE_KB) turned on: with
NX2NAVI_MIRROR_STAGE=1 (the default) the pages the CPU wrote are copied out of guest memory, up to 64 MiB a
part, when the part is handed to the queue. A submission whose prologue copied everything that way reads no
guest memory late, so the game's wait after a frame's first submission ends once the replay has submitted it
(vk_sync_recorded_wait). One with a range left to the GPU-time copy (a pool with a write-back in flight or a
twin) or a draw that reads a pool without a mirror is noted (vk_mem_note_live), and its wait is the GPU's.
Also 8 texture pool sets.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The perf line's "idle with it queued" was the gap before each replay part handed over before the previous one ended:
13-17 ms a frame on an RX 580 in Kokiri Forest on 0.7.1, unexplained. Uploads and presents were not timed, a held
submission (submit batching) counted as handed over while it still waited in vk_sync, and with cuts and pre-pass
command buffers a submission closed more than the four timestamp pairs ts_unsent had room for: the rest went to the
completion queue without a submit time and out of the GPU's order. Now every timed piece of GPU work (replay parts,
mirror prologues, pre-pass command buffers, upload batches, presents) is noted with when it was handed to vk_sync, when
it really went to the queue and its GPU start and end, and every 60 frames two lines split the idle time: starved
(nothing in the queue, and how much of that a held submission waited in the batching) or with work in the queue, by
the kind of piece that started after the gap, with the average delay from hand-over to start on an idle GPU per kind.
ts_unsent has room for every command buffer of a submission, and the pairs are read in queue order.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>