On Linux the write watch protects pages, so a write-back from shadow storage into guest memory took a fault
at the first store into every page. plat_mem_prepare_write clears the protection of a range with one
UFFDIO_WRITEPROTECT call; the pages count as written, as they would after the faults. Off by default;
Windows has nothing to clear.
Pull request #327 by theurbanninja1993-sketch (a draft on GitHub). The platform function and its Linux
test are taken as they are. The call moved: the write-back it was written for is now shadow_write_back in
runtime/vk_memory.c (a file other sessions are changing), where it covers the runs of pages a submission
changed, and the setting is registered in cfg.c and described in nx2navi.ini.example, which also says
that the lazy write-back stays off with shadow storage.
The Windows files compile; the Linux code and its test are not compiled or run here.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
With shadow storage (Linux, and Windows builds under Wine) a CPU-visible pool is a host-visible copy with
no memory object of its own, and only pools that owned one went into the device address table: pointers
the game's shaders read from memory reached nothing. Shadow copies are now in the table. A shader that
reads through the table resolves no pool, so nothing queued the pool it reaches for the copy-in: the part
is marked where its pipeline is drawn or dispatched (VkPipeInfo.gpa, vk_mem_gpa_use), and its prologue
brings the shadow of every pool in the table up to date with the CPU's writes and queues the same pages
for those pools' VRAM mirrors (vk_memory.c vk_mem_gpa_sync). --probe-vkmem reads a word through the table
from a pool that is never bound, before and after a plain CPU write.
Pull request #353 by TophC7, ported onto the base copies of the black-world fix:
- taken as it is: the table registration, the mark in vk_exec_draw.c and vk_pipeline.c, the prologue
call and vk_sync_probe_submit in vk_sync.c (two hunks merged by hand), the probe's test;
- written again: vk_mem_gpa_sync, on top of the snapshot-based shadow_sync_pool; a lookup no longer
queues a pool for being in the table, which every shadow pool now is;
- narrowed: the sync runs only with shadow storage. Without it (Windows) a mirrored pool stays unsynced
for the table's readers as before, so the frame there is unchanged, and the probe makes the source's
mirror only with shadow storage and skips when the device has no address table;
- the probe's shader was an array of SPIR-V words in tex_probe.c; it is now the source
runtime/shaders/gpa_read.spvasm, built like the other runtime shaders;
- the exact re-upload count in the texture probe stays checked on Windows.
Files other sessions are changing: runtime/vk_memory.c (vk_pool_init's registration, mem_resolve's touch,
the new section before vk_mem_use).
The Windows files compile and the shader assembles and validates; no probe was run, and nothing was run on
Linux here.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Shadow storage noted a GPU write with the whole range its binding names: the copy-in skipped that range and
the fence copied all of it back, so CPU writes to other fields of the same records never reached the GPU.
Ocarina's lighting parameter is such a field: the lighting came out NaN and the fog pass drew the scene
black. Each page a noted range covers now keeps a base copy, the copy-in moves only the bytes the CPU
changed and the write-back only those the GPU changed (vk_memory.c merge_page). The mirror's merge shader
marks the 4 KiB pages it changed, and one job per submission copies only those back (shadow_wb_cb); the
copy-in reads the CPU-side generation tables through the shared write-watch snapshots; and a caller that
reads guest memory itself resolves with vk_mem_lookup, which does not queue the pool for the copy-in.
Pull request #394 by jacc. Its changes are taken as they are; one hunk met the lazy write-back that came
later (vk_gpu_written_back) and is merged by hand. Added here: the lazy write-back stays off with shadow
storage, because its fetch copies the mirror into the pool's buffer, which is guest memory only for an
imported pool. Files other sessions are changing: runtime/vk_memory.c (the shadow section, the merge
record, vk_gpu_written_back, wb_lazy) and three one-word call changes in runtime/vk_usnap.c.
Without shadow storage (Windows outside Wine) the merge pipeline gains an unused fifth binding and two
push constants and nothing else changes. The Windows files compile; nothing was run, on Windows, Wine or
Linux.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A fetch copies the mirror over the whole range, not the merge of the bytes the GPU changed that a write-back does. The
completion thread arms a range after its write-back reached guest memory; a CPU write into those pages in between, not
yet carried up by a sync of the mirror, was then overwritten by the mirror's older bytes at the first fetch. In play
that was single frames with another frame's constants: the night lighting missing from the whole scene in Hyrule
Field every 0.1 to 0.4 s, and a black scene in Kakariko, gone with wb_lazy = 0. Now the pages are made inaccessible
first and the back end then checks the CPU-write page generations against the mirror's last sync and the write-watch
bits nobody has read yet; when either shows a write the mirror lacks, the pages are given back and the range arms at a
later write-back. The lazy line counts the arms refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A review found a gap: a part checked what it had skipped at its end, and a fault between that check and the part's
hand-over to the queue fetched without the part's writes while nothing wrote them back afterwards. The replay now
opens a window from the check to the hand-over, per part, and a fault on an armed page waits for it to close (at most
20 ms; the replay thread's own faults and thaws do not wait and are counted). What a part skipped for a range that is
no longer armed goes back through the usual snapshot and merge, taken in that part's prologue, and not as a plain
copy, which would have put the mirror's words over what the CPU had just written there. A part split off earlier
makes the same check before its prologue. wb_lazy_touch is a test that reads armed ranges on purpose.
Dev build, Hyrule Field: 71.1 fps as before; standing, a fight, pause and unpause, and the touch test made 146 to 338
faults each, none inside a window, no fault of its own.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The 2 GiB of device memory a filled cache cost was not the thousands of compute pipelines but twenty of them: the compute shaders of 0.9 to 2.0 million SPIR-V words, the game's tensor-core upscaler programs, which nothing dispatches. A pipeline of one holds about 100 MiB of device memory from the moment it exists, and the driver keeps that memory when the pipeline is destroyed, so keeping only the recently dispatched pipelines did not help: the boot pass still created the twenty from pipeline-cache.bin at every start. Measured with --probe-prewarm on a copy of an installed copy's caches (RTX 5090, NVIDIA 621.01): the pass over all 4,066 compute shaders took the device memory from 386 to 2,398 MiB and the process's commit up by 2.5 GB; the same pass before the twenty were in the cache left the device memory at 386 MiB, and all 4,046 other pipelines kept at once cost 2 MiB.
A compute shader over NX2NAVI_LIB_PREBUILD_BIG is now built ahead, from the caches or by the background builder, only when the usage record says a run dispatched it. None of the twenty is among the 138 shaders eleven runs dispatched. Should the game dispatch one for the first time, that dispatch compiles it once and later boots build it ahead.
The boot also reads the shader of a pipeline it does not keep as a copy and releases it after the build, instead of leaving 3,900 shaders loaded (240 MB of commit); a first dispatch then reads its shader from the disk, 0.04 to 0.10 ms in the probe.
With both, the same pass on the same caches leaves the device memory at 386 MiB and raises the commit by 47 MB; 256 first dispatches after it took 0.13 ms on average, 1.8 ms at most, all read from the cache file. With compute_keep = all, which builds everything as before, the pass ends at 2,528 MiB. Not yet run in the game.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A draw whose shader or pipeline is still building is left out of its frame, and the frame was shown without it: a black or unlit picture with the interface intact when it was a lighting pass, a light going out, or ground missing for a frame. In a player's session a full-screen draw was skipped after its 50 ms and the frame came out black as a boulder hit.
The frame's present is now left out when draws were skipped (vk_backend.c frame_hold, NX2NAVI_FRAME_HOLD_MS, 120 by default, 0 off). The game runs on, so nothing waits. The hold comes from an allowance that fills again at an eighth of the clock's rate, so a long build or a run of them shows its frames as before and does not freeze the picture; a full-screen draw that would be skipped with the allowance spent waits on instead, 500 ms at most. Not with DLSS Frame Generation.
On a first walk through a village with no pipeline recorded, 21 of the 23 frames with skipped draws were held back; in the Lost Woods, 7 of 8.
The performance overlay: the VRAM line is the card's dedicated memory in use, which counts every program on it: a player read 14.5 GB as the game's while the game held 7.3 GB. The line's third cell now gives what this process holds on the device-local heaps (the driver's budget figure), in place of the GPU clock, which stays where the driver gives no budget.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Standing in Hyrule Field the GPU spent 13 ms a frame, of which 2.4 ms were the game's passes and 7 to 9 ms the copy
of 66 MiB of shader-written buffers back into guest memory at the end of every submission, in case the CPU reads
them. The write-back probe found no CPU access to them in 40 s. Now a range goes back once and its pages are made
inaccessible; while they stay so, later writes stay in the mirror. The first CPU access faults, the bytes are fetched
with a GPU copy of their own after everything submitted so far, the pages are given back, and the range is written
back every submission again for a while (1 s, doubling to 32 s while it keeps being touched). Thaws before kernel
reads and writes and before mapping changes release armed ranges the same way, and so does a mirror that goes away.
Dev build, same spot, 80 s each: 51.5 to 71.0 fps median, 1% low 43.5 to 56.3, GPU time 13.1 to 7.5 ms a frame,
fence waits 9.8 to 4.4 ms. In that run the game faulted on armed pages 226 times and 124 fetches took 51 ms in all;
a walk into a fight made 283 faults and 186 fetches, and a test that touches eight armed ranges every 60 frames
(wb_lazy_touch) 480 faults and 364 fetches, none with a fault of its own. wb_lazy = 0 is the old behaviour.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
An installed copy's warm list has 42,344 variants, 42,142 of them from shader-prepared.bin, and its 4,066 compute shaders are what the boot kept pipelines for. The comments and the settings text said the list grew with the play history; a copy with prepared shaders held the 2.3 GiB from its first boot. Comments, docs and the ini text only.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
(cherry picked from commit ea984fe05b995c5dad1e0f49c5f44de9ee8f19b2)
The boot prebuild built the pipeline of every compute shader play ever recorded and kept all of them for the process. A pipeline that exists holds driver memory whether or not anything dispatches it: with a history of 4,066 compute shaders they held 2.3 GiB of device memory, and its share of the process's commit, from the title screen on (RTX 5090, NVIDIA 621.01), while a boot into the first village dispatched 95 of them.
The boot now keeps the pipelines of the 512 compute shaders dispatched most recently (compute-used.bin next to pipeline-keys.bin records which run last dispatched each; the setting is compute_keep, with "all" for the earlier behaviour). Every other one is still built at boot, so that the pipeline cache holds it after an update or a driver update, but destroyed at once, and created from the cache at its first dispatch. A dispatch that comes while such a build runs gets that build's pipeline.
Each boot stage now logs the device memory and the process commit it leaves ("mem: boot stage"), and the log says what the first dispatches cost. nx2navi.exe --probe-prewarm runs these stages without the game from the caches next to the executable and times first dispatches of pipelines the boot did not keep: 256 of them took 0.19 ms on average and 3 ms at most, all read from the pipeline cache file, and 372 pipelines built and destroyed at boot left the device memory where it was.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
(cherry picked from commit 16a796d623686bca5304eaf1b29981247453a413)
The pre-snap of a submission took the pools the submission before it had synced. In Ocarina those are another submission's pools: the frame's first and fourth submissions use no pool and queried the pools of the ones before them, and the submissions after them then queried their own pools one at a time at first use. The pools are now remembered by the submission's place in its frame.
NX2NAVI_PRESNAP_PLACE=0 gives the old choice back and =ab switches it in the B blocks.
Co-Authored-By: Claude Fable 5.2 <[EMAIL]>
vk_mem_collect destroyed dead images, views and buffers while it held cmd_lock, and the replay's uploads wait for that lock. Destroying an image and its memory takes up to milliseconds. The collect now moves the objects out under the lock and destroys them after releasing it.
Co-Authored-By: Claude Fable 5.2 <[EMAIL]>
Every frame the render thread waited about 5 ms for the GPU after the frame's first submission, with the game
CPU-bound in every area. The wait is there so that the GPU has run that submission's passes before the game
rewrites what they read. But the submission also ends with the write-backs of what its shaders wrote, 60 MiB merged
into guest memory in 3 to 4 ms of GPU time, and the wait covered those too.
The write-backs now go into a part of their own. The part before them signals a second timeline, the executed
timeline, and the first-submission wait waits on that; the write-back part signals the timeline as before, so every
fence the game asks for still covers them, in the same queue order. wb_split = 0 keeps one part and the old wait.
Not measured in the game yet.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- The mesh stage's storage table starts 0x90 bytes before the end of the 2 KiB constant bank, and all sixteen slots
were written: 0x6C bytes past a stack buffer on every mesh draw. It holds the nine slots that fit.
- The prebuilt runtime is compiled without AVX, so its preemption stub saved only XMM0-15 while the lifted code of
an AVX2 processor keeps data in the upper YMM halves, which the stub's callees may clear. Both bodies are built
now and the one for the processor is chosen at the start.
- After an FSR 3.1 frame generation error the window's swapchain was still AMD's object but went to the plain Vulkan
calls. Swapchain calls follow whose swapchain it is, and a failure takes it down as a switch in the menu does.
- A movie is paced at its own 60 fps under a lower limit too: at 40 fps it ended 25 s behind its sound.
- A lost device left the wait for an upload spinning for good; it ends the run with the device's report.
- The compute pipeline key had four bytes of padding that were hashed and compared, so a prebuilt pipeline could be
missed and built again.
- A thread's TLS pointer was published before the TLS was allocated, so the slot was always 0.
- A fatal exit puts RivaTuner Statistics Server's profile back, and an ini saved as UTF-16 is read.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Per-pass GPU timestamps put the write-back of GPU-written storage buffers at 3.6 to 5.3 ms of GPU time a frame in play, against 1.8 to 2.5 ms for all of the game's own passes and dispatches: about 67 MiB a frame goes from the mirrors back into guest memory, from the same eleven pools all session, in case the CPU or a host path reads it. Nothing said whether anything does.
With wb_probe = 1 the 2 MiB regions that lie wholly inside a range just written back are made inaccessible. The first read or write of such a region faults; the handler gives the region back the protection it had (read-only when frozen, else read-write), notes the address, the instruction and the thread, and the access runs again. The bytes are in guest memory all along, so nothing changes for the game. A log block every perf window gives the regions armed, the probes a read or a write ended, and the first accesses by instruction. Every thaw ends the probe of its region first, which covers file reads into guest memory and mapping changes; a file write from guest memory thaws its buffer while the probe is on; a probed region is not frozen anew.
The platform layer gets a hook for every read or write access fault beside the write-fault hook; it is installed only with the setting. --probe-gfx-abi checks that nothing is armed without it and, run with NX2NAVI_WB_PROBE=1, that reads and writes end a probe and run again, that a frozen region thaws after it and that mem_thaw ends and holds.
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>
VMA's blocks on large heaps are 64 MiB instead of 256 (vma_block_mb): one session held 2.1 GiB free inside the
device blocks beside 8.2 GiB of allocations after a few area changes, and the driver counts a block whole until it
empties. The headroom thread keeps its spare block only for the device-local types textures and mirrors grow in; the
spare blocks it made for the host types and the PCIe BAR type at boot stayed empty for the whole run. The ASTC
decoder's scratch is 32 MiB, so it stays inside a block, and an area load that makes one per upload batch holds half
as much at once. A DLSS feature no frame used for 3600 evaluations is freed with its work images, and a model change
now frees all five sizes.
The perf log gets a "vram by use" line (textures, released textures, mirrors, physical pools, ASTC scratch, headroom
spare, free space inside the blocks, memory outside VMA and DLSS's own share), the free ranges per memory type, the
addresses whose duplicate texture entries hold the most, and a line when memory outside VMA moves by 256 MiB.
astc_bc_probe in nx2navi.ini turns on the ASTC opacity probe.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 99cfc2fa14ae208b7b23e057ff9b87986cd6a5e0)
A shader's store or atomic through the table goes to a pool's VRAM mirror and again to its guest copy, which on a
discrete GPU crosses PCIe. On PCIe 3.0 (RTX 4090, Kokiri Forest, DLSS Quality) turning the twins off took the GPU from
16 to 13.1 ms a frame; the writes that went wrong without them were foliage counts and records in pools of 20 and
44 KiB. gpa_twin_max_kb = n gives a twin only to mirrors of pools up to n KiB. A larger mirror's record names the mirror
twice, so its second write goes to the table's scratch page, its buffer fills stay in VRAM (vk_mem_guest_twin), and it
is never evicted, since guest memory lacks what the shaders wrote into it. The mirror report counts the mirrors made
without a twin. Default 0: every mirror has one, as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A device lost seen on RTX 4090, 5090 and 5070 Ti cards in 0.5.0 and 0.6.0 (#177, #61, and a 0.6.0 test run): a
"read invalid" at the start of a 128x128 scene capture pass, re-rendered at area loads. destroy_tex purged the views of
the image and deferred the image, then freed its record at once, on whichever thread released the texture. The replay
keeps &e->pub as a render target for the rest of a command buffer (vk_exec.c resolve_target) and DLSS for the frame:
a pass that began again (each indirect mesh draw ends the pass) read the freed record and cached a view of the dead
image, which nothing removed. Vulkan gives a destroyed image's handle to a later image, which then got that view, and
the pass loaded its attachments from freed memory.
- destroy_tex marks the record gone and frees it through the deferred destruction (vk_defer_free, a host kind stamped
with the submission being recorded, as the image is).
- A gone texture's attachment view is made per command buffer (vk_exec.c view_of) or deferred with it
(vk_exec_target_view, for DLSS and the contact shadows), never cached.
- A new image drops any view still cached under its handle, and the log counts it ("cached views of a destroyed image").
No rendering changes: the views and images live exactly as long as before, the records slightly longer.
Co-Authored-By: Claude Opus 5.5 <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>
With trace = perf, every 600 frames heap_log started a thread that walks the whole address space with VirtualQuery
(plat_mem_census) for the "mem: process" line. The walk takes 100-240 ms, and though it runs on its own thread it
holds the process's address-space lock, which the game's threads take at every page fault and VirtualAlloc: the
present after each walk arrived 65-120 ms late, a hitch every ten seconds in every run with the trace on, and the
runs people attach to issues carried it. The walk now runs only with mem_census = 1 (NX2NAVI_MEM_CENSUS); the rest of
the memory lines stay.
Kokiri Forest, trace = perf: before, "slow frame" lines at 601, 1201, 1801, 2401 and 3002 of 65-120 ms with nothing
of it in the pipeline, texture or fence columns; after, none at those frames.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
vk_gpu_written's stamps, which the texture cache needs, made mirror_sync copy the same pages back up into the mirror
the write-back had read them from: outdoors about 57 of the 60 MiB copied up a frame, about 5 ms of GPU time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HtayeqEG2x5woYzWuEQBGV
The 16 MiB merges keep up to 256 MiB of idle snapshot chunks for the submissions in flight. Those chunks prefer device
memory and nothing evicts them, so on a 6 or 8 GiB card they took room the mirror and texture eviction defends. Large
cards keep the full 256 MiB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ranges just above the old 4 MiB limit were copied back over PCIe at the end of every submission. A larger snapshot pool
keeps the extra merges from allocating on the replay thread. NX2NAVI_WB_MERGE_MAX sets the limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HtayeqEG2x5woYzWuEQBGV
texup_log_ms logs a "texup:" line for each frame whose texture work on the render or replay thread
reaches that many milliseconds. The line splits the time into texture creation, staging, storage
copies, ASTC recording, scratch buffers and new device memory, to show where an area load's hitch goes.
astc_scratch_keep keeps the ASTC decoder's scratch buffer across upload batches instead of making a
new 64 MiB buffer for each batch. astc_arena stages GPU-decoded ASTC blocks in the staging arena
instead of a buffer per texture. Both aim at the allocation churn seen during area loads and stay off
until a play test measures them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the replay on the submitting thread (the default), vk_texture_release and entry_of destroy textures on the game's
loader threads, where vk_exec_recording_fence is 0: the deferral took the last submitted value, so the image could go
while another thread's submission, or the recorder behind it, still named it. Such deferrals now wait for issued + 1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Performance: draws reuse the bindings that did not change and record only the dynamic states that differ from what
the command buffer holds; small uploads come from a staging arena; finished submissions are held briefly and handed to
the queue in batches; the recorder thread can keep a CPU core of its own; ASTC block-linear textures are deswizzled on
the GPU; dynamic resolution steps the upscaler mode to hold a frame-time target. Each has a setting in nx2navi.ini
(bind_reuse, dyn_delta, stg_arena, submit_batch, recorder_core, recorder_prio, astc_gpu_deswizzle, dyn_res) and the
staging cache now frees what it holds at shutdown.
Capture: NX2NAVI_DUMP=key saves the frame after each press of Insert, with mesh draws listed by task, mesh and fragment
shader, depth test and face culling, and mesh_prims.txt giving each mesh draw's primitive, clipping and fragment counts.
All of it is off unless NX2NAVI_DUMP is set.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In a test run with the budget taken as 16 GiB, 113 pools refused for the budget asked again 1,795 times in 6,000 frames, every 16 submissions, though nothing could make room: the world's textures were all in use. Now the idle mirrors go only when together they free what the new mirror needs, and a refused pool asks the textures for room and comes back soon once; after that it asks again every 600 submissions until it gets its mirror.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A test run with the budget taken as 16 GiB showed two faults. The evictions measured the driver's heap use, which keeps counting a VMA block whole while one allocation lives in it, so freeing textures and mirrors in the middle of blocks did not move it: they went on until 5700 of 7232 textures and 395 of 410 mirrors were gone (9.2 GiB free inside blocks). And mirrors evicted under pressure during a menu were refused again when the world came back, so the draws read the pools over PCIe at 4 fps. The evictions now count the free space inside VMA's blocks as room (vk_mem_vram_live), mirrors are no longer evicted to make room for textures, and a mirror that does not fit after the idle mirrors made room asks idle textures for the rest (vk_texture_want_room) and is asked again 16 submissions later.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A pool whose mirror did not fit the VRAM budget stayed imported for good, however hot, while mirrors nothing had used for minutes kept their memory. Now a new mirror that does not fit takes the room of mirrors no submission synced for 120 submissions, least recently synced first; a pool refused for the budget asks again after 600 submissions; and above the texture pressure threshold (90% of the budget) mirrors idle for 600 submissions go every 8 submissions so the textures get the room. A victim has nothing in the submission being recorded and its GPU writes are already back in guest memory; its device address table pages move back to the imported buffer (vk_gpa_remap_plain) and the mirror is destroyed once the submission is done. It is mirrored again from a whole copy at its next buffer use. NX2NAVI_MIRROR_EVICT=0 turns it off.
Measured with the budget taken as 16 GiB, same route: off, 691 mirrors (910 MiB) and 133 pools refused for the budget; on, 405 mirrors (799 MiB): 240 idle ones (98 MiB, title-screen and loading pools) evicted, 114 pools refused. That exe still evicted idle mirrors under pressure for the textures, which a later commit removes, and asked refused pools again every 16 submissions (1,795 times), which another commit spaces out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With it set, the mirror budget check and the texture pressure checks compare with at most that many MiB, so a large card behaves as a small one does when video memory runs low. The heap lines still log what the driver reports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every 600 frames with NX2NAVI_TRACE=perf: the live textures' own memory by kind (decoded ASTC as BC3 or RGBA8, other CPU-written, GPU-written) with the part no lookup reached for 240 submissions and 120 presents, entries sharing a storage address with another, the VkFormats holding the most and the largest entries; then the imported pools the GPU resolved and those it never did, and the VRAM mirrors idle for 120 submissions. Textures now carry the submission and present of their last lookup for it.
Measured with it (user's save, RTX 5090, 2560x1440 with DLSS performance, astc_bc = 1): in the world the live textures held 5,166 MiB (3456 decoded ASTC as BC3, 3438 MiB; 3412 other CPU-written, 965 MiB; 216 GPU-written, 763 MiB; 46 sharing a storage address, 115 MiB) and the mirrors 927 MiB, and of the 5,007 MiB of imported pools the GPU never resolved 4,077 MiB (3787 pools). With fine ASTC footprints kept as RGBA8 (the main checkout's uncommitted rule for astc_bc set by hand), decoded ASTC took 10,382 MiB instead of 3,438.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NX2NAVI_TRACE=audio logs voices starting, stopping and sharing a wave buffer; NX2NAVI_DUMP_MOVIE dumps movie frames; NX2NAVI_LOG_CS_BINDS says why the stand-in for an untranslated compute program skipped a dispatch; and NX2NAVI_TRACE=perf adds heap and commit census lines on their own thread, with VMA maps when write-combined memory grows by 2 GiB. From a tester's fixes.
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>
Copying the invalidated ranges from guest memory into the mirror at every submission put the guest's copy of those pages over GPU results that live only in the mirror. In adult Hyrule Field that wiped tree records, and the trees stretched into columns reaching the sky again. With NX2NAVI_READBACK_SYNC=0 the columns did not appear at the same spot. The memory check (NX2NAVI_MEMORY=verify) still finds mirror pages newer than guest memory there with or without the refresh, about 3 per 60 frames, which is left for later.
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>
Shaders: a switch whose index is clamped against a register that a MOV of an immediate set before it now finds its
jump table (AmpereLowering.cs BrxTargets), so all 36,031 of the game's shader variants translate and no draw is skipped
for a failed one. The compare-and-swap loop behind storage-buffer atomics (patch 0033) takes the value a failed swap
returns as its next expected value instead of loading the word again, which on some drivers kept returning a stale
copy.
Memory: an unused physical pool outlives the submission being recorded, as virtual pages may still reach it through
device addresses; an image whose creation failed is not tried again for 2 s.
The installer and Linux texts name this game.
The README has the project icon, the current status, features, requirements and known issues.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Kokiri Forest drew black, in tiles and in shadowed and indoor areas, where shaders read light, probe and instance data through pointers they load from other buffers at an index (a phi of a pushed pointer and a loaded one, a pointer spilled to local memory): those reads returned 0. Translator patch 0030 keeps such accesses global, with a new first source naming the table, and the SPIR-V back end looks the address's 4 KiB page up in it and loads through the buffer device address it gives (PhysicalStorageBuffer64). runtime/vk_gpa.c keeps the table: two levels over the 40-bit GPU address space, page entries naming immutable records, records reused only after the GPU finished with them. Imported pools, their VRAM mirrors and the pages of virtual pools go in it; pools in it are destroyed after the submissions that may still read them. The last failing compute program needed MATCH.ALL (subgroupAllEqual) and the aggregated shared increment ATOMS.POPC.INC, now lowered in AmpereLowering.cs; Kokiri Forest translates without a failure.
NX2NAVI_GPA=0 turns the table off and NX2NAVI_GPA_EMPTY=1 keeps it empty. A first try synced every VRAM mirror at each table reader's draw, which overwrote what the GPU had written into them and made the black worse; mirrors sync only when bound.
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>