The author of pull request #353 closed it on 2026-11-11 with the words that they give no permission to use its code. Its port (3134f524) goes: the mark of a pipeline that declares the table (VkPipeInfo.gpa, vk_pipeline.c pipe_add) and of the part that draws with it (vk_mem_gpa_use in vk_exec_draw.c bind_resources and mo_emit), the sync of the table's pools in every such part's prologue (vk_mem_gpa_sync), the submission with a callback in the replay's place as it stood (vk_sync_probe_submit), and --probe-vkmem's read through the table with its shader (test_gpa_read, runtime/shaders/gpa_read.spvasm).
In its place, after pull request #551, which was written without that code: a shadow pool goes into the table and onto a list of the table's shadows (vk_memory.c shadow_table_add); the record of a binding of the table notes the guest submission (vk_exec_draw.c ssbo_info, vk_mem_shadow_table_read); and vk_mem_shadow_sync brings the CPU's writes into every shadow on the list once for that guest submission, before the part is submitted. Once is enough, because the copy-ins read the guest submission's write-watch snapshots, which a second sync would find unchanged. Measured by #551's author on Linux (RTX 4090, Kokiri Forest): 113 and 76 frames over 25 ms in 75 s against 217 and 429 with the port, 43 against 41 fps, the picture complete in both.
What this line had built on the port stays, on the new sync: the test that passes over a pool with nothing written (shadow_pool_clean), the upload of the same pages into a table pool's VRAM mirror, the write-back of stores through the table (NX2NAVI_SHADOW_TABLE_STORES, table_wb_cb, now asked once per guest submission that binds the table) and the counts of NX2NAVI_SHADOW_CHECK. The probes that record their own commands (the lost-device probe, --probe-vkmem's table parts) call vk_sync_submit_own, written anew. vk_mem_shadow_storage stays: one line that returns the mode. Finalized pools in the table were already kept for vk_mem_shutdown here.
--probe-vkmem has a check of its own for the sync (test_table_sync): the shadow of a pool that is never bound keeps its old word through a CPU write and a guest submission that does not bind the table, and holds the new one after a guest submission that does. It passes with NX2NAVI_SHADOW=1 on Windows (RTX 5090), as do the store check with NX2NAVI_SHADOW_TABLE_STORES=1 and the probe without shadow storage; --probe-devlost ends as before. Not run: the game, with or without shadow storage, and anything on Linux.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A pipeline cache over the size limit was dropped at every start and, since the file is no longer written in play, never replaced: every start began empty. The play file is now folded in at the next start whatever its size when the file beside it is over the limit.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A mesh draw becomes its record (the workgroup count, copied on the GPU for an indirect draw), the task dispatch, a
scan that lists the mesh workgroups the tasks started, the mesh dispatch, and an instanced indirect draw with a
generated vertex stage and the program's fragment stage. The compute modules bind the resources of the task and mesh
stages, and their pipelines are made on a thread of their own while the program's draws are skipped.
The indirect mesh ring and the texture pool marks are released at a probe's shutdown.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A save that found no memory for its copy, could not write its file or met a build that had just begun was not tried again until the next new pipeline was built, so a run that built none and then ended without its orderly exit lost what it had compiled since the last file. The autosave thread now repeats the try, and the same bound holds for the save owed after the play file was folded in, which before was tried at every interval for as long as it failed. tests/save_retry_check.py checks the rule.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The library creates its pipelines itself and names no pipeline cache. It is now handed pipeline creates of the runtime's in the place of the device's two, and a create without a cache is asked of the pipeline cache first and compiled into the play cache otherwise. ffx_pipe_cache = 0 turns it off. --probe-ffx-cache makes one upscaling context without a window or the game and says where its pipelines came from, and tests/ffx_cache_check.py runs it.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A draw that waits for a pipeline with no fast link (a mesh pipeline, one without stage libraries, or one whose link failed) needs the whole build, and its job stood where it was queued: the draw queue is first in, first out, so on a burst of new pipelines the build a draw stood waiting for started after the fast links and builds of the draws before it, and the draw was skipped at the end of its wait all the same. graphics now hands the job to job_wanted: while no worker has taken it, it goes to the head of the draw queue and a worker is woken. A wait longer than 500 ms, which is no frame budget (a one-shot draw with oneshot_budget = 0), builds the job on the waiting thread at once when no worker is idle, as a stage library no worker took already is. A wait inside a frame's budget is never turned into a build on the replay thread: the budget is there so that thread is not held for a compile. The worker's build and bookkeeping moved into job_build for both to use; the log's statistics count both cases.
The other half of this item was there: the boot's prebuild asks the caches for a large compute shader's pipeline (compute_create, cached only) before it queues the background build (lb_compute), and with the play cache that question now reaches both caches.
Checked: it builds, --probe-cfg is ok, and --probe-prewarm builds the 3,212 recorded pipelines through the moved code as before. The queue jump and the build on the waiting thread need a draw that waits, which no probe makes: not run.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
pipeline-cache.bin was copied out of the driver and written whole at every save after any pipeline was created, also one the cache had held: tonight's logs show three saves of 358.1 MB in a run that added 0.3 MB, and 302 to 306 MB written in each of seven sessions that added about 0.5 MB each. The library warm in play built its 37,572 stage libraries into the same cache, which grew it by 1.6 GB: a save of 2026.7 MB in play, a file over pipeline_cache_max_mb, and the next start dropped it whole and compiled the pipelines it had held.
cache_save now asks a cache for the size of its data first and neither copies nor writes one whose size is what its file holds. What a run compiles goes into a second cache with its own file, pipeline-cache-play.bin (pipeline_cache_play, default 1; needs pipelineCreationCacheControl): pipe_create_in asks the first cache with VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT and builds into the play cache only when that says a compile is needed, and in play only the play file is saved. At a start the play file is read back and merged into the first cache in memory; once it reached pipeline_cache_play_mb (64) it is folded in for good: pipeline-cache.bin is written once, before play, and the play file removed. The preparation modes fold it at any size and build into the one file as before. The library warm's threads create with no cache of ours, so their libraries reach the driver's own cache alone.
Checked with --probe-prewarm on a copy of a dev folder's caches (3,212 recorded pipelines, 333 compute shaders), no game run: with a filled pipeline-cache.bin every pipeline came from it and neither file was written; with none and a cold driver cache the compiles went into pipeline-cache-play.bin (75.9 MB written, pipeline-cache.bin not created); the next start merged it and wrote pipeline-cache.bin once (191.1 MB) and removed it; kept below the fold size, a play file of 68.8 MB gave 1,445 pipelines at 0.3 ms from the first cache where 470 hit without the merge, and was not written again. Not checked in play: the saves during a session, the library warm without a cache, and drivers other than NVIDIA's.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Only the large compute shaders' background builder looked at the system's commit before it started a build, and it
saw nothing of what the builds beside it were about to take. The boot's recorded pipelines (up to 64 builders), the
pool's background builds during play, the boot's compute pipelines and stage libraries and the library warm asked
nobody.
Each of them now asks the gate before a build. A build that starts reserves an estimate of the driver compiler's
working memory until it ends (build_reserve_mb, 64 MB; 4096 MB for a large compute shader), and the gate admits a
build only while the commit headroom holds the part kept free (build_free_mb, 4096 MB), the reservations and its own.
A refused boot builder ends and leaves its pipelines to the pool; the pool leaves a background build queued and asks
again every two seconds; the stage library prebuild and the library warm wait up to 30 seconds and then end, and the
warm goes on at the next start. A pipeline a draw or a dispatch waits for is never refused. The preparation of the
pipeline list runs without the gate: no game runs beside it.
tests/build_gate_check.py checks the gate's arithmetic and, through --probe-buildgate, its answers with the machine's
own commit. What the builders do when refused is not exercised by a test.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The prepared list named its shader variants and nothing about where they came from, so an update that keeps the prepared shaders (shader_carry) kept the old list whatever seed the new kit writes: variants only the new seed names, from a newer pipeline list or a change to which variants are prepared, were translated when play first drew them. The existing identities do not cover this: nx2navi_xlate.id and identities.txt name translators, which decide cache keys and what is carried over, not which variants are listed.
The list is now version 2 (runtime/prp_file.h): after the head come the sha256 of seed-programs.bin, the sha256 of seed-answers.bin and the translator identity, written by gfx_shader_prepare_seed; the three readers go through gfx_shader_prepared_read and still read a version 1 list. At an update step_shaders writes the seed first (prepare_shaders.py --no-run, about 4 s) and keeps the list only when its digests are that seed's (make_game.py seed_changed); another seed, or a list that does not say, is prepared again. Another translator alone still carries over: input ids do not depend on the translator, and preparing everything again for it is what carrying was made to avoid.
Checked with tests/prp_file_check.py (the C layout against make_game's reader), three launcher tests, and the exe's --prepare-shaders on a seed without records in a folder of its own: the digests it wrote are those of the seed files, one byte more in the seed reads as changed, and after preparing again it does not. A version 1 list of 42,142 variants and the same list as version 2 are both read whole by --probe-prewarm.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
cache_load measured pipeline-cache.bin with ftell into a long, which is 32 bits on Windows: a file past 2 GB measured as -1, so the run started with an empty cache and the log said nothing, not even that the file was over pipeline_cache_max_mb. The read now lives in runtime/vk_cache_file.h: the length comes from the 64-bit tell, the read runs in 64 MB pieces, and a refusal returns its reason and the file's size, which cache_load logs. The Windows atomic write no longer cuts a block of 4 GB or more to a 32-bit count.
tests/pipeline_cache_file_check.py reads a sparse file just over 2 GB through the header. With --probe-prewarm and such a file carrying this GPU's header, the log reads "holds 2048 MB, over NX2NAVI_PIPELINE_CACHE_MAX_MB 512" under the default limit and "2048.0 MB read" with the limit off; the exe before this change logged nothing for the same file.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Verified by runs in the audit: the Market's periodic stall (a pool slot was listed once per refresh pass instead of once; the list fell from 160,385 entries to at most 3,192 and no frame over 20 ms remained while standing), orderly exits delete the pipeline prebuild's running marker (left behind, the next start wrote a skip file and every later start skipped the prebuild), the log and crash report name an area a save asks for that the game does not have, and the menu atlas hash no longer overflows a signed multiply. The other fixes compile and pass the regression route but their faults were never reproduced: see AUDIT.md section 13 for each.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The marker was deleted only when the boot's builds ended. A window closed before that left it, the next start read it as a start that died in the driver's compiler, and every later start on that GPU and driver skipped the prebuild and built each pipeline at its first draw.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A file over pipeline_cache_max_mb starts the run empty, but it was read into memory whole before its size was compared, which is the memory the limit is there to spare on a 16 GB machine. The size is now compared first and a file that is dropped is never read. The limit itself stays at 512 MB: sizing it from the machine's memory, as the contributor proposed, was not run by them or here.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Four "mem: sizes" lines give the entries and bytes of what no other line counted: the guest heap, physical maps, kernel region, lazily written ranges and the fixed kernel, setjmp, save-file, RomFS and zstd tables; pool records, mirrors, deferred destructions, upload and staging buffers, the device address table and the texture, sampler and slot caches; image views, descriptor spares, the completion queue, uniform snapshot arenas, recording copies and pins; and pipelines, layouts, stage libraries, job queues, the shader memory cache and the translator processes. Each subsystem reads its own counters under its lock; settings now count the replaced values they keep. Segment and scene lines say "-" for the world clock before it starts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A slow frame's line and the 60-frame window's line now say how long each kind took on any thread, how often, how many bytes and the longest single one, so a hitch outside the render thread's own waits has a named cause. While NX2NAVI_TRACE=perf is off each site pays one branch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The carry path of an update does not run the shader preparation: it runs --warm-shaders, which builds the listed pipelines from the seed-pipeline-keys.bin the game folder already has. make_game.py now writes that file from the kit's list first, so a newer list reaches an updated install. Its shader variants are not translated on this path; a record whose variant the folder has neither prepared nor met is left out and counted.
On a GPU that has no mesh shaders every mesh record failed with a log line each. They are counted in the step's line instead.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The preparation translates every shader and builds each one's stage library, so a pipeline a draw meets for the first time is linked in 2 to 6 ms. A mesh pipeline has no stage libraries: on an install that never drew it, it compiles whole in play, 0.2 to 1.1 s with its draws left out of the picture (18 of them on a first visit of Kokiri Forest). Whole pipelines were built ahead only from the folder's own pipeline-keys.bin.
tools/pipeline_list.py merge folds the pipeline-keys.bin and shader-answers.bin of game folders that were played into one list file: every key record once, with the answers record of each shader it names. A record names its shaders by input id, which depends on the game's code and the answers alone, and holds the game's pipeline state; what depends on the GPU is derived when the pipeline is built, so a record made on one machine is good on another with the same game version and runtime. Four folders on this PC give 5,432 records, 317 of them mesh pipelines (147 KB of keys).
tools/prepare_shaders.py reads the list (tools/data/pipeline_list.bin, data/pipeline_list.bin in a release folder, or --pipeline-list), adds its shader variants to the seed and writes its records as seed-pipeline-keys.bin. nx2navi.exe --prepare-shaders builds the mesh pipelines among them after the seed is translated and saves the pipeline cache before the stage libraries are built (NX2NAVI_PREPARE_PIPELINES=all builds the vertex pipelines too); --warm-shaders does it again after an update or a new driver, and --prepare-pipelines is that step alone. The pipelines go into the caches and are not kept or added to the folder's own key file, so they cost no start-up time or video memory until a draw meets them. A marker file keeps a driver that crashes on one of them from failing every preparation.
The list is derived from game data and stays out of the repository: check_content.py refuses files recorded from a game folder by their first bytes, and make_kit.py, make_launcher.py and make_release.py carry a list given with --pipeline-list. All three now also carry pipeline_list.py, which prepare_shaders.py imports.
Checked without the game: tests/pipeline_list_test.py, and prepare_shaders.py --no-run on the game files with a list of 317 mesh pipelines (all 266 shader variants found in this game version, 177 new to the seed). The build step in nx2navi.exe compiles and links and has not been run.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The reports of black and white blocks on Navi, on Link next to her and on bright surfaces seen from inside a house come from RX 5000 and 6000 cards, whose Windows driver compiles with "AMD proprietary shader compiler" (vk_caps.amd_sc); an RX 7900 XT (LLPC) draws the same scenes clean. The game's shaders test for NaN with OpIsNan in 48,717 of an install's 54,959 translated modules, 26 tests each on average, and the translator's guard that keeps NaN out of every float colour output, so that one bad texel does not spread through blur and bloom into blocks, is such a test too. SignedZeroInfNanPreserve asks the driver to keep those tests; whether that compiler does is not known.
With that driver, vk_pipeline.c spirv_nan_bits now turns every 32-bit isnan, scalar or vector, into (bits & 0x7fffffff) > 0x7f800000 on the number's bits, and a float select with a constant whose bits are a NaN into a select of the bits (issue #156), when the module is created. NX2NAVI_VK_NAN_BITS=1 does it on every driver, 0 on none. Run over every module of an install's shader cache, the rewrite changed 1,281,675 tests and 57 selects with every output well formed (each id defined once, below the bound, the uint types declared once); --selftest checks it on a small module, and Link's house with it forced on (RX 7900 XT) draws as without it. Not yet tried on an RX 5000 or 6000 card.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Found in the first run of the previous two commits (a first visit of Kokiri Forest from a state that had only seen Hyrule Field):
- The count of frames a target was rendered in started again whenever the guest's texture object changed, and the game initializes its render target objects anew every frame at the same storage: the log showed `seen 2129/2128 in 1 frames`, so no target ever became steady and the rule did nothing. The owner of an image is now the storage address the texture cache keys it by. An image handed on to a texture elsewhere still starts at none.
- A one-shot mesh draw waited its 8 ms in every frame that met the same pipeline while it compiled for 150 to 1100 ms: 17 waits, 170 ms, every draw skipped at the end, up to four in one frame. A mesh pipeline that has been building for longer than that wait is a compile, and the draws after the first do not wait for it (vk_pipeline_building_ms).
- The last wait of a full-screen draw with no picture to hold was 40 ms a draw, and two such draws made 86 ms in one frame. It is 40 ms a frame now.
- A one-shot vertex draw skipped at the end of its wait gets its log line however short the wait was (seven were skipped with no line, so their targets could not be named).
That run against the same cell of the earlier comparison (same exe path, cache state, route and settings; one run each): the load's worst wait of the draws in one frame 99 ms against 735 ms, no mesh wait over 13 ms against 524 and 206 ms. Two long frames remain in both runs (628 and 862 ms here, 660 and 854 ms there): two compute pipelines built at their first dispatch, about 395 and 397 ms.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A session's log gave the counts of slow pipelines and slow frames but not what they were. Now:
- Each pipeline a draw missed that took 50 ms or longer to become usable gets a line (the first 48 of a run): mesh or vertex, fast-linked or built whole, whether pipeline-keys.bin named it, its shaders' input ids and where their translations came from (the shader cache, a previous translator's, or translated in this session). The summary line counts the same: how many were mesh pipelines, how many the key file named and the prebuild had not reached, how many had a shader translated in this session.
- Each shader library that took 25 ms or longer gets a line (the first 32): its shader, the translation's origin, and whether outputs were added for its fragment shader, which is a module the preparation never built.
- The count of draw lookups that "found shader libraries missing" says what it counts: a library not made in this process yet, not one the driver's cache lacks.
- After the boot prewarm one line says how many of the recorded pipelines' shaders are a previous translator's translations, beside the pipeline cache hits that show whether any pipeline was compiled again.
- The frame-time summary's "over 25 ms by cause" took the replay's whole time as one part, so every frame that waited for a pipeline, a translation or a texture upload inside the replay counted as "replay". The replay's part is now its time without those waits.
gfx_shader_origin gives a translation's origin.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Eight commits: a draw that misses its pipeline has the rest of its submission read ahead, the library workers link
the pipelines of skipped draws, a pipeline whose fragment shader reads a target of its own pass is fast-linked like
any other, and diagnostics. The merge was clean. Their before and after matrix had not finished when this was merged.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The next draw with a new pipeline was only found when the replay reached it, so the draws of an area's first frame,
or of a turn that brings several new materials into view, paid for their pipelines one after another while the
workers that could have built them together sat idle. The submission is recorded before it is replayed: the first
draw of a submission that finds a shader or pipeline missing now has the ops after it read (vk_exec.c vk_exec_ahead)
with the binds that decide a draw's shaders and shader libraries applied to a copy of the replay state. Each draw
there looks its shaders up as the real draw will, which also queues a shader that is not in memory, and names its
state to vk_pipeline_hint, which queues the pre-rasterization and fragment shader libraries that are missing. No
pipeline is entered and no key recorded, so a wrong guess costs two libraries at most. The draw's own lookup then
links on its thread when both libraries are there.
On a first visit to Kokiri Forest a scan read 6,900 ops in 0.78 ms on average (the longest 5.5 ms; a scan ends
after 3 ms), and the scans of the load and the walk named 377 states without a pipeline and queued 104 libraries
for them.
pipe_ahead = 0 (NX2NAVI_PIPE_AHEAD) turns it off.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A draw that is skipped while its pipeline builds leaves the pipeline to the pool: a job is fast-linked first and goes
back in the queue for its optimised build. The pool keeps one worker out of optimised builds; with the other three
inside one (100 to 350 ms each) that worker linked every new pipeline of a burst one after another. The library
workers, which so far served only draws that wait, now take the fast link of a draw-queue job too (vk_pipeline.c
fast_take, lib_worker), and there are as many of them as pool workers (four on sixteen logical processors, two
before).
Measured on a first visit to Kokiri Forest (RTX 5090, libraries in the driver's cache): of the 444 pipelines draws
missed, 180 took over 50 ms to become usable, the longest 1.3 s, and their draws were missing from every frame until
then. With this change and the two beside it (feedback loops, the look-ahead) 23 did, 19 of them mesh pipelines,
which have no fast link.
pipe_fast_workers = 0 (NX2NAVI_PIPE_FAST_WORKERS) gives the old behaviour.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Such a pipeline (soft particles and water reading the depth buffer, full-screen passes that read the target they
write) had no fast link: gpl_link refused it, so its draws were skipped during, or waited for, the whole optimised
build. The flags belong to the fragment output state alone, so the fragment output library carries them and its key
names them (FoKey.fbl); the shader libraries are the ones the same shaders use without a loop.
Measured on a first visit to Kokiri Forest (RTX 5090, stage libraries in the driver's cache, DLSS Frame Generation
on, where no frame is held back and a full-screen draw waits on): full-screen draws of such pipelines waited 267, 42
and 61 ms in the area's first frame and 410 ms in the middle of the walk; fast-linked they wait 2 to 3 ms.
gpl_feedback = 0 in nx2navi.ini (NX2NAVI_GPL_FEEDBACK) gives the old behaviour.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- Every pipeline a draw missed is timed from the miss to its first usable pipeline, the fast link or, without one,
the build itself (vk_pipeline.c ready_note): the gpl lines of a save give the counts under 4, 8, 17, 50 and 150 ms
and over, the average, the longest, and how many had no fast link. That time is how long the pipeline's draws were
missing from the picture or waited.
- The stage library lines add the libraries of 25 ms or more and their sum, and the pre-rasterization libraries that
had outputs added for their fragment shaders: those are another module than the one the preparation built, so the
driver's cache has none of them.
- "a draw waited" lines name a pipeline with a feedback loop.
- The "draws skipped for a pipeline" counter (overlay, perf lines, flight records) was raised by every lookup that
found the pipeline building, also when the draw then waited for its link and was drawn. It is raised where the draw
is skipped now (vk_exec_draw.c draw_skipped).
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The driver keeps the prepared shaders' libraries in its shader cache, and under Windows that cache has a size limit
set in NVIDIA's settings. A limit smaller than the set drops the oldest libraries as new ones come in (a 2048 MB cache
could not hold them on the reporter's RTX 5070 Ti), so every start builds them again in the background and play
stutters until that is done. The cold lines of the library warm now say so on NVIDIA under Windows, most plainly when
the cache is cold right after a warm that finished on the same driver; the build kit's notes and the example file say
it too. tests/warm_hint_check.py checks which hint each case gets.
From pull request #403 by tarballz; two conflicts resolved (an include and the build kit's notes).
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>
Only a shader a run dispatched reaches that builder now, so its pipeline stays, entered before its cache is merged, as before the usage record existed. The branch that destroyed an unkept pipeline before the merge could no longer be reached; had it been, a dispatch in that window would have found the pipeline in no table and in no cache and compiled it again.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Found by a review of the compute prebuild change:
- cf_log formatted and wrote its log line while holding the pipeline lock shared, on the autosave thread, which runs below normal priority; every dispatch needs that lock. It now copies the numbers and writes after the unlock.
- The record could be written by the autosave thread and by a save at exit at the same time, so the older copy could land on the newer file; and a failed write cleared the changed mark, so nothing tried again. One writer at a time now, the copy is allocated outside the lock a dispatch takes, and a failed write leaves the record marked.
- compute_keep read anything that was neither "all" nor a number as 0, which keeps nothing; it now reads as the default.
- The statistics called a pipeline "read from the pipeline cache file" when the driver said the pipeline cache held it, which is also true of what the same run built into the cache earlier. The lines now say the pipeline cache held it. The probe's skip line names the settings that also lead to it.
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>
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)
After an update or a driver update every recorded compute pipeline is compiled again in the background: about 140 ms each on an RTX 5090's driver, so minutes of two threads for a long play history. Those compiles ran at normal priority, because a dispatch of a shader being compiled waits for that build and an idle-priority compile under load had held such a frame for seconds. On a CPU with few cores that put two more threads in equal competition with the game's own.
The compile now runs below normal priority, and the dispatch that comes to wait for it raises the thread to normal, as it already does for the background builder of large compute shaders. Not measured in play.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
(cherry picked from commit e2b38ffb22690b1656d22fbc504d000554877e0d)
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 preparation builds every stage library once so the driver's disk cache holds it and a library a draw needs later takes under a millisecond. Turning dynamic vertex input off for AMD's RDNA 3 and newer driver changed the dynamic states every library lists, which gave each one another entry in the driver's cache, and nothing built them again: an update carried the prepared shaders over without the driver step. On an RX 7900 XT (26.10.1) all 42,142 prepared libraries were cold, 93 ms a vertex and 242 ms a fragment library on average. With no pipeline recorded, a walk through a village skipped 11,180 draws while their pipelines built and made 94 draws wait 5.5 s in all, four of them 1.0 to 2.0 s; a player's session log shows up to 114,834 draws skipped in 600 frames and 181 waits of 251 ms on average. Skipped ground and walls appear and disappear, and a skipped lighting pass is a black frame with the interface intact.
- The game checks the driver's cache at every start (vk_pipeline.c lw_start): a sample of the prepared fragment libraries is built, and when their median time says the cache is cold the whole prepared list is built into it again on idle-priority threads while the game runs. shader-warm.txt records a warm in progress, so the next start goes on with one a run ended, and stops the tries on a driver that does not keep the libraries. This also covers a driver update and a cache the driver emptied.
- An update that carries the prepared shaders over runs the game's --warm-shaders (tools/make_game.py warm_driver_cache): 11 s with the libraries there, about 8 minutes with all of them cold on that card.
- A draw whose pipeline is missing waits for its fast link from a budget of 6 ms a presented frame, 3 ms at most each (NX2NAVI_PIPE_LINK_MS), before it is skipped: with warm libraries the link is done within that.
- Full-screen draws wait from their 50 ms budget again instead of being skipped (NX2NAVI_PIPE_SKIP = any is the default; fullscreen stays selectable). With warm libraries the wait is a millisecond.
The same walk afterwards: libraries 0.3 and 0.7 ms on average, 1,605 draws skipped, 1,473 of them in the scene's first 600 frames and 8 to 80 in each 600 after, and 19 waits of 60 ms in all. First visits to four other saves (a mountain pass, the Lost Woods, Kokiri Forest, a river) waited 12 to 84 ms in all except the Lost Woods, where mesh shader pipelines, which have no stage libraries, and vertex libraries that need added outputs still compiled once: 23 waits, 2.1 s in all, the longest 776 ms.
The texture cache counted its waits in presents chosen at 60 a second: at 165 presents a second the 1800 that stood for thirty seconds were 11 s, so a released render target went sooner than meant and a texture out of view for 11 s was retired and showed its stand-in for some frames when it came back. The counts are now taken times the present rate over 60 (presents_for), and a texture is retired after two minutes without a sample instead of thirty seconds.
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>
Main brought the fixes of eleven pull requests (the texture pool set sealed when an early replay part is submitted, every location reflected in interface arrays, RGB9E5 texels recognised for integer storage views, the upscaler a frame started with kept, GPU pass timestamps without vendor breadcrumbs, the Linux build, window and NGX fixes), render targets released when drawn over at another address, the vertex layout kept in the pipeline on newer AMD drivers, and the upscaler choice read once per evaluation.
Four hunks conflicted and take both sides: pool_sync returns for a sealed set before the per-frame creation budget's clock (vk_exec_draw.c); make_module keeps the test line's early_ok argument and main's module hash (vk_pipeline.c); a texture taken back counts as sampled again and records how long it was away, and the texture line logs the disuse counters and the overwritten render targets (vk_texture.c).
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
0.6.1.14 faulted the GPU (a read of an invalid address in a pass) about 30 seconds after its last disuse evictions, and
no path by which an evicted image is still read was found in the code. An evicted texture is now first retired: the
pool refresh gives its slots the stand-in while the image stays whole. It is destroyed only once the texture pool has
taken every set of its ring since and the GPU has finished the submissions recorded until then, so no pool set can
still be bound with its view. A mark on one of its slots or any lookup outside the pool refresh takes it back as it is,
and the perf line counts those, which shows whether the marks miss a real use. tex_disuse = 2 retires and never frees,
so the next fault can be put on the freeing or on the marks and stand-ins.
The registration shortcut also read a slot's cached entry without checking that the resolution was current, so it could
read the record of a destroyed texture; it now checks the slot's generation first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 4c5b2ae2e030ab00a910d8f071c030e610fd78e1)
The translator's mesh shaders are SPIR-V 1.4, and the marks took only 1.3, so the first mesh shader that indexed the
pool turned eviction by disuse off for the whole run. From 1.4 on an entry point lists every global variable it uses,
so each entry point now gets the marks variable too. Over the 44,255 shaders of a test install's cache, 4,728 are
marked, none is refused, and every marked one passes spirv-val for Vulkan 1.3; 11 of them are 1.4 mesh shaders. The
log names up to eight refused modules by stage, SPIR-V version, size, hash and reason.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit af50808a1cb56f450f99938c9d420e1f5deb333b)
AMD's driver for RDNA 3 and newer compiles a pipeline that has dynamic vertex input again at every boot: neither pipeline-cache.bin nor the driver's own disk cache gives it back. On an RX 7900 XT (26.10.1) a player's key file of 3,979 records used the whole 20 s boot budget at every boot. Dynamic vertex input is now off by default on that driver (vk_caps.amd_llpc); with the vertex layout in the pipeline the third boot of that file built 4,273 pipelines in 0.9 s, all from the file. NX2NAVI_VK_DYN_VERTEX=1 turns it back on. The RDNA 1 and 2 driver branch is not measured and keeps it.
The background builder of large compute shaders now compiles into a pipeline cache of its own and merges it into the main one when the build is done. A create into the main cache holds the cache's lock for its whole length and the cache is only saved while nobody holds it, so a build of minutes kept pipeline-cache.bin from being written at all: on the same card one compute shader of 4,874,502 SPIR-V words was still compiling 590 s after the boot, and nothing built in that session reached the file.
NX2NAVI_PIPE_FEEDBACK=2 also logs every graphics pipeline build with hashes of its modules and its fixed state, which is how the recompiles were found.
Co-Authored-By: Claude <noreply@anthropic.com>
The texture pool refresh looks up every registered slot, so the textures of every area the game keeps loaded never
counted as idle and their memory only grew. Marked shaders now store into a buffer with one word per pool slot. Every
four submissions that recorded work, the replay copies it out and clears it, and the texture side notes when each
texture was last sampled. Under VRAM pressure or low commit, once idle textures are gone, live textures that come back
intact and were not sampled for 1800 presents go too. Their keys are held, so the pool refresh keeps the stand-in until
a shader samples that slot again, and the next refresh makes the texture again.
Fragment shaders that neither discard nor write depth, sample mask, stencil reference or memory keep early depth tests
(EarlyFragmentTests), so the marks do not take away early depth rejection. Alpha to coverage leaves the dynamic state
while marks are on, so each pipeline knows whether early tests are safe. Disuse eviction waits until every module that
indexes the pool was marked and two reads arrived. NX2NAVI_TEX_DISUSE=0 (tex_disuse = 0) turns it off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 1a4c896150a748f9275c3024e96af3055db94af4)
vk_pipeline.c spirv_pool_mark gives each module that indexes the texture pool a storage buffer at set 2, binding
VK_POOL_MARK_BINDING, and a store of 1 into the slot after every access chain into a pool texture array; modules that
cannot be marked are counted in vk_pool_marks_missed. NX2NAVI_SPV_MARK_TEST=<list> runs the rewrite over files for
spirv-val under --selftest. Still to do: the marks buffer and its read-back (vk_exec_draw.c), the pool lookup context,
the holds and the eviction by disuse (vk_texture.c).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 7ae86070f166d935a9949e95dae7a95e2cf3272b)
The fog's first compute pass skips froxels behind a coarse depth table from the previous frame and writes them as empty, so the columns a near object had just left came out with no fog, and the fog's history smeared that into a trail, worst at the screen edges. The comparison is now never true, so those froxels are computed. The shader is found by its shape, so this holds across translator updates. NX2NAVI_FOG_CULL=1 keeps the game's skip.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fog's first compute pass skips froxels behind a coarse depth table from the previous frame and writes them as empty, so the columns a near object had just left came out with no fog, and the fog's history smeared that into a trail, worst at the screen edges. The comparison is now never true, so those froxels are computed. The shader is found by its shape, so this holds across translator updates. NX2NAVI_FOG_CULL=1 keeps the game's skip.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>