Worlds, the UGC server and the dashboard open the pair named in
resServer/cdclient-current (CDServer.sqlite without an fdb when there is
none). Worlds switch to a new pair on CDCLIENT_RELOAD between frames.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Appended at the end of MessageType::Master (46); the pin test is updated.
With names it tells a server which fdb copy and CDServer.sqlite to switch
to; without them it asks master to check the client's fdb now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CDClientSnapshot converts a copy of the client's fdb into its own
CDServer-<hash>.sqlite with the cdserver migrations applied, on its own
connection. CDClientDatabase::Reconnect opens the new file before letting go
of the old one; CDClientManager::Reload empties every table (old entries kept
alive so references from spawned objects stay valid), retires the old fdb view
and loads again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
FdbSnapshot names copies cdclient-<hash>.fdb / CDServer-<hash>.sqlite, writes
them under temporary names and renames them in, keeps a pointer file naming
the current pair, removes old copies, and decides when a changed file has
settled enough to hash. Nothing in it logs, so it can run on a worker.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Launching to a property started an extra clone 0 instance of the zone besides the property's own, because
the prep only carried the zone. PrepZone now carries the clone when there is one (written only then, so a
plain prep is unchanged).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixture check now also reads every recorded client game message with the
struct the server reads it with (GameMessageHandler::CreateReceived) and writes
it again: it must read the whole message and give back the same bits. Game
message fields are decoded with the server's structs in the fixture tests.
A synthetic fixture, built in the test from the server's own structs (login,
position update, a client game message), goes through the export steps
(portable, anonymised, saved, read again) and passes the same checks; recorded
fixtures stay in tests/fixtures-local and are never committed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GameDependenciesTest deleted the logger, config and managers but left the
pointers set, so a later test without the fixture (packet capture) logged
through a freed logger and crashed when the whole suite ran.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Staff arm a packet capture on the dashboard; master passes MESSAGE_CAPTURE_CONTROL ARM to
every world, auth and chat and arms its own. Each server's PacketCapture tap (dServer receive,
and a send hook in RakPeer::Send so replica constructions are seen too) records into one
preallocated chunk per server and ships sealed chunks through master on the main loop when
capture_flush_bytes or capture_flush_interval_ms is reached; past capture_buffer_max_mb the
oldest chunks are dropped and the dashboard records a gap. Nothing is armed: one flag check.
- targets: an account (from its login; packets before the login are kept per connection
and added once auth or the world knows whose they are), a character (from when it is
picked), or everything; up to 8 at once (a bit each in the record mask)
- worlds and auth record their clients' packets and the master link messages of a captured
player (session keys by name, zone transfers by request, player added/removed, migration);
chat finds the player in each packet; master records server traffic for everything
- secrets are never recorded: structs that carry them (login request, login response user
key, world validation session key, session key messages between servers) are read,
blanked and written again before recording; auth keeps only the handshake and login
- PacketDecoder: a registry by service and message id names every packet and decodes the
registered structs; CaptureBundle is the file format (DLUBNDL1, metadata, records);
CaptureTools orders records on one timeline, pulls movement out, makes bundles portable
or anonymous and diffs replays
- the dashboard keeps packet captures in message_capture_sessions (capture_kind 1) and
their packets in a file under capture_dir, one write per batch; arming is audited
- MESSAGE_CAPTURE_CONTROL/DATA only gain appended enum values and trailing fields
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An open group (game clients, web clients, a zone's instances) keeps one box
that grows by its rows, at most 8 in view with the rest scrolled. Each row
in view is a link end of its own with only its member's traffic, and the
open box's header has none. Members are filtered by name, account or user,
instance and (with network_ips) address, in a steady order; web clients
show the dashboard user.
The filter and rows are HTML over the SVG, made once and then moved and
updated in place, so live updates keep the typed text, the focus and the
scroll; links follow the list's scroll. The box grows and shrinks with its
list, and the List view nests the members under their group with the same
filter.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Web clients are one entry per address and signed-in account, with user and
account_id; the user is shown without network_ips too, the address then
masked as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each address's requests are counted per signed-in account (the dashboard's
session or API key, as the auth middleware found it), so several people
behind one address stay apart, with the account's user name; a WebSocket
upgrade counts under its account.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dashboard's own loop is framed too, with a scope per module update. Frame time and stacked phase charts per server, the servers' loop summary, longest frames, packet handling times, the last 50 slow frames with a nested timeline, and profiling sessions (profiling_run, GM 8) drawn as a flame graph with folded stacks to download. PerfHistory keeps it in memory and is unit tested; the layouts are tested with node. Task 96.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An optional section after marker 3 carries the report's frame timing; older readers stop before it and reports without it still read. Phase times go with their count so readers with fewer or more phases read them. PROFILE_REQUEST and PROFILE_RESULT are appended to the master messages (44, 45). Task 96.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Profiler.h: frames, named scopes and phases, recorded on the main thread only (other threads' scopes do nothing). Per second: frames, total and longest frame time, a mergeable frame time histogram (TrafficStats::Histogram) and time per phase. Per report: the packet types that took longest, the worst frames and the frames over a threshold with their heaviest scopes. Profiling sessions merge every frame's scope tree for a while into one (folded stacks for flame graphs). Optional Tracy client (DLU_TRACY, off by default) gets the same frames and scopes. Task 96.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GameTextJs scans templates, scripts and the routes' strings for the
game's zone, item and character names and currency labels, with an
allowlist for legitimate uses and a self check of the scanner.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One helper for game text on the dashboard: zone names, locale phrases,
localized table columns, %[key] expansion and the game's currency words,
looked up in the viewer's language (a cookie pick, else Accept-Language),
falling back to en_US and then the key or id. Unit tested.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Skipped unless DLU_FDB_BENCH_XML names a character save; prints startup
time, inventory load time and resident memory (private and file-backed,
on Linux) for DLU_FDB_BENCH_MODE=fdb or sqlite.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The world and master servers pass the client's res/cdclient.fdb to
CDClientManager. When it opens, these three tables, all looked up by
their first column, find rows through the fdb's buckets instead of each
process caching the whole table: ComponentsRegistry keeps nothing,
ItemComponent and Objects keep only the entries asked for (their API
returns references). Ids whose rows CDServer.sqlite changes are loaded
from SQLite at startup and win. Without an fdb (or with one whose
columns don't match) the tables load from CDServer.sqlite as before.
ItemComponent and Objects now fill entries from one template for both
sources instead of copies of the same field list.
Tests cover the SQLite changes on top of the fdb, the no-fdb and
unmapped paths, and, when DLU_CLIENT_RES points at a client, every id of
the three tables through the fdb against CDServer.sqlite.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CDFdb opens cdclient.fdb once per process and hands out a table when its
columns match the CDServer.sqlite table of the same name. CDServer.sqlite
stays the source of truth: the CDServer migrations change a few rows, so
FindChangedKeys compares both files row by row (a hash per row, summed
per key) and returns the keys that differ, for the tables to read from
SQLite. RowFields reads an fdb row with CppSQLite3Query's accessors and
defaults, so a table fills its entries from either file with one piece
of code.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
FdbMappedFile maps a file read-only (CreateFileMapping/MapViewOfFile on
Windows, mmap elsewhere) and falls back to reading it into memory when
mapping fails. FdbReader reads the table and column headers from it and
looks rows up by their first column through the fdb's own buckets,
decoding every integer as little-endian with bounds checks, so the rows
never get copied out of the file.
Tests write small fdb files (collisions, text, int64, nulls) and read
them both mapped and from memory.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The client's TacArcBehavior::Cast (0x00fb2d10) sorts the targets in the arc nearest first, or by weight when
distance_weight or angle_weight is set (sortWithWeights, 0x00f58cd0: distance_weight * (max range - distance) /
max range + angle_weight * (180 - angle) / 180, heaviest first). With use_attack_priority, SortByAttackPriority
(0x00f72900) then buckets them by GetAttackPriority, lowest first, keeping that order inside each bucket; only the
DestroyableComponent answers it, and an object without one counts as 1. DoHit keeps the first max targets.
Nothing else ranks targets: enemies are taken over nearer smashables only because their attack_priority (1) is
lower than most smashables' (10). use_attack_priority is off when a behavior does not set it (TacArcBehavior::
Initialize, 0x00f9b980), as before.
The server sorted by distance only and ignored the flag, so a one-target swing hit the nearest crate instead of the
enemy behind it. OrderTargets now does the client's ordering, with equal targets in ascending id order (the order
the client's id set hands them over in); the chosen targets are still written in ascending id order (issue 1045).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Account pages get a Client system info (as reported) card (client_sysinfo):
every description the account's client sent at login, newest first, with the
raw values, the memory text split into numbers and each field's caveat. The
address is only shown with logs_audit. The Client System Info page (Logs &
Health) shows the spread across players from each account's newest report:
Windows version, video card, memory buckets, processor count and client build,
marked as client-reported and approximate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
client_sysinfo (migrations mysql 102, sqlite 85) keeps the system description
each account's client sent at login, exactly as sent, plus the physical memory
read from it. While nothing but the memory in use changes, the account's newest
row gets the new time and one more login; otherwise a new row starts. Log
pruning deletes rows not seen for a while (eLog::CLIENT_SYSINFO) and deleting
an account deletes its rows.
Tests on SQLite alone (dDatabaseSqliteTests) and in the MySQL parity tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The login request's memoryStats is a run of short parts with no separators
(process working set and private bytes, memory load, physical memory, commit
limit, the 32-bit client's own address space, peaks). ParseMemoryStats splits
it into numbers; a text cut short keeps the parts before the cut.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A player on the servers' machine (or behind the same address as other players) was merged with the
servers' own links and everyone else there.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The layout the viewer makes is kept in the browser, with a Reset layout button. A link's colour now
compares it with the busiest link right now instead of its own recent peak, which made steady links
look saturated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Next to Diagnostics (health_view): game and web clients, auth, chat,
worlds per zone (expandable), other reporting servers, master, the
dashboard and the UGC server, with a lane each way per link sized by
bytes/s, animated by packets/s and coloured by load against its own
5 minute peak. Estimated links (servers without a split) are dashed.
Clicking a box shows its links, busiest message types and a sparkline,
with a link to Diagnostics (which now takes ?server=). A connections
table lists each remote address; single clients can be drawn too.
Narrow screens get a list. Drawing stops while the tab is hidden.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The traffic topic now carries each server's rates with its split by
peer (null for servers that report none), link statistics and gauges.
New routes: /api/diagnostics/network, /network/server (message types
and a 10 minute series) and /network/connections (remote ends grouped
by address). Addresses need the new network_ips permission; without it
each is a salted token. They stay in memory from the last report only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two optional sections at the end of the report, each after a marker
byte: every second's packets by peer with its HTTP requests from and to
other servers, and the busiest remote ends with the rest summed.
Reports without them still read (older servers), and older readers stop
before them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every server now splits its packet counts by peer: its own connections
(players on auth and worlds), its master link, or other servers (the
worlds on chat, every server on master, a world's chat link, which is
now counted too). HTTP requests carrying X-Darkflame-Server count as
another server's, the dashboard counts its own requests to the UGC
server, and each HTTP client address is counted. The report also gets
each RakNet connection's statistics (worlds name the player on it),
trimmed to the 32 busiest with the rest summed. Counting stays on the
main loop, except the dashboard's UGC fetches, which only touch the
locked recorder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Guild chat per guild (/chat_log/guild/<id>, from the guild's card), team chat by
team (/chat_log/teams) and a character's whispers by conversation
(/characters/<id>/whispers, chat_dms; opening one is audited), paged on the
server. The Chat Log filters by account and time range, shows the conversation
around a message, links to its history and marks flagged messages; messages can
be picked (shift-click for a range) to flag. The character page links to all of
them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
chat_log gets guild_id, team_id and filtered (migrations mysql 100, sqlite 83),
and the chat log queries take a time range, a guild, a team, a conversation
between two characters and paging back by id. Whispers are their own visibility
switch (includeWhispers), apart from team and guild chat (includePrivate).
Whisper partners and teams that talked are listed with counts.
chat_flags, chat_flag_messages and chat_flag_events (mysql 101, sqlite 84) hold
flagged chat with a copy of the messages around it, its status, note, linked
player report and history. Open flags are in the dashboard snapshot.
Tests on SQLite alone (dDatabaseSqliteTests) and in the MySQL parity tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The client's render component sets the object's own position and
rotation on the root node it loads, over the stored ones, so a root's
turn never shows in game (its scale stays). NifFile now reads roots the
same way. LU Toolbox's .nif (root turned 90 degrees about X, the
NiLODNodes turned back) now stands up in icons and the dashboard's
views; every game .nif has an unturned root and reads as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MovementAIComponent::SetDestination cuts an enemy's path (chasing, tethering,
wandering) where it first walks into a wall its collision group can't cross,
so enemies no longer walk through the Sentinel camp walls, the Crux Prime and
property navmesh carvers, or the enemy blocking volumes. With nothing left to
walk the enemy stays put. Patrols along a level path, and movers without combat
AI, are left alone.
Adds a scenario test: an enemy chasing across an enemy only volume and across a
carver stops in front of it; the clear threat wall is a trigger, not a wall.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Enemies are moved along navmesh paths without collision, so a wall only stops
them if their path stops at it. dpWorld now keeps a list of such walls, each
with the collision filter it blocks with, and cuts a path where it first walks
into one (a little short of the wall; a mover inside one can walk out).
Which objects block comes from data (dpMovementBlockers::BlockingFilter):
- navmesh carvers (navmesh_carver in the level config, which the client reads
with add_to_navmesh and carver_only) block every mover;
- solid objects whose collision group touches enemies and not players (group 18,
e.g. "FV - Enemy Blocking Volume", "PR - Pet Blocker") block what that group
touches.
dpShapeBox gains SegmentEntry, a segment test against the rotated box.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
UgcToolbox runs LU Toolbox itself in an external Blender that stays up
and makes one model after another (dlu_toolbox_worker.py runs
LU-Toolbox-Standalone's steps: import, Process Model, Bake Lighting,
niftools export). JSON lines over its stdin and original stdout; it is
restarted after a crash, a timeout, 100 models or new settings, runs at
worker_nice with toolbox_threads, and its CPU time is charged to the
worker, which pauses it (SIGSTOP) while the CPU budget is overdrawn.
ProcessModelToolbox reads LU Toolbox's .nif back, counts its triangles
and draws the icon from it. Tests: option parsing, the fallback, the
protocol framing, the worker with a stand-in Blender (crash, hang,
failure, give-up), and LU Toolbox itself when DLU_TEST_TOOLBOX_* are set.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After serviceType the reply now sends major, minor, patch, a flags byte
(build kind in bits 0-1, dirty in bit 2), the first 32 bits of the commit
hash and a u16 length-prefixed build string, instead of the stale fixed
ASCII "0.1.3". unknown stays "DLU3". The 1.10.64 client reads only
netVersion and serviceType and never checks the length, so the extra
bytes are ignored. The build string is optional on read.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Live GF Large Crates (LOT 1859, path CrateMast) and FV small white shrines
(LOT 3141, path gate_statue_quickbuild) dropped 1-point powerups outside
their template matrices, from the path's smashable_loot_matrix=1:29 with
smashable_loot_matrix_set=7:1. DLU already carries a spawner path's
waypoint config into the spawned entity's settings; this pins it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A quickbuild's HasItem preconditions took their items only when the build
completed and never gave them back. Live took them when the build started
and gave them back when it was cancelled: the FV Stone Warrior pedestal
(LOT 8551, precondition 99: 5 of LOT 6194) took the items at Building 5
times and added them back with the loot source Quickbuild on each of the 3
cancels in the live captures.
Preconditions now report their item costs instead of removing items while
checking. A quickbuild takes them at the start, gives them back on cancel
or a reset during the build, keeps them on completion, and gives them back
when the builder leaves the world mid-build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ActivityRewards coins were always read from CurrencyTable npcminlevel 1.
Live used the reward row's ChallengeRating as the npcminlevel, and level 1
when the currency index has no row for that level:
- FV foot races (ChallengeRating 4, index 1) gave 36 and 48 coins, which
only level 4 (30-50) fits; level 1 is 3-5.
- Frakjaw's chest (activity 58, ChallengeRating 6) gave 250 each to a team
of 2: its currency indices 123-126 only have a level 6 row (500), so DLU
gave nothing.
- Quickbuilds, wishing wells and chests have ChallengeRating 1; survival
and the shooting galleries have ratings with no row and keep level 1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The third ray backend, so every machine has a library for it: Embree on x86
CPUs (embree), HIPRT on AMD and NVIDIA GPUs (hiprt), Embree through SYCL on
Intel Arc and Xe GPUs (embree-gpu).
- CMake option DLU_EMBREE_SYCL (off). dUgcServer/EmbreeSycl is a project of
its own built by a SYCL compiler (DLU_SYCL_CXX; icpx through ONEAPI_ROOT or
the path, or the open source DPC++'s clang++ through DPCPP_ROOT) as an
external project: Embree 4.4 with EMBREE_SYCL_SUPPORT, linked statically and
bound inside (-Bsymbolic, only its C functions exported, so it never meets
the servers' own Embree), and the GPU kernels (nearest hit skipping a ray's
triangle, any hit), into libdlu_embree_sycl next to the servers
- UgcRaysEmbreeGpu loads it the first time embree-gpu is asked for; one GPU for
the process (embree_gpu_device picks it), the workers take turns, the
occlusion rays in batches as for hiprt
- without the build, the library or a supported Intel GPU it falls back to
embree and says why (the UGC server's log at start, --make-model on stderr)
- the option names, the settings page, the dashboard's picker, /reprocessproperty
Verified here: the default build and ctest; the SYCL build with the open source
DPC++ 7.1.0 (compiles, links against oneAPI's libsycl.so.9, exports only its C
functions); on this machine (no Intel GPU) it loads, finds no GPU and falls
back to embree. Not verified: tracing on an Intel GPU (none here).
Check: on a machine with an Intel Arc or Xe GPU and oneAPI, configure with
-DDLU_EMBREE_SYCL=ON and run UgcServer --make-model x.lxfml out embree-gpu;
the UGC tests then compare it with Embree.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Embree 4 (always built) replaces the UGC server's own bounding volume
hierarchies (the nearest-hit one and the occlusion rays' any-hit one), with no
fallback to them. ray_backend is embree (default) or hiprt (optional build;
Embree when it can't be used).
Settings, stored options and stats that say builtin still read: it is embree
(UgcRays::Parse, UgcProcessOptions::Parse). The dashboard's picker,
/reprocessproperty and --make-model offer embree and hiprt.
Tests: the backends are compared with Embree (hiprt when built), Embree against
rays whose hits are known, and the clutter's occlusion against what the old
hierarchy worked out (296 vertices summing to 114.5, 26 open, 183 dark); the
pinned model hashes are unchanged with Embree.
Check: ray_backend=builtin in an ini still starts and uses embree; the
settings page offers Embree and HIPRT.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The path tracer that imitated LU Toolbox's Remove Hidden Faces took 17 to 100
times as long as the renders from 42 directions around the model, too heavy to
use. It goes with its settings (hsr_method, hsr_samples, hsr_bounces,
hsr_sample_spacing, hsr_min_points), its side by side tracing for GPUs and its
tests. The renders (UgcRender::VisibleFromAround) remove hidden faces as before
that method existed; their size is hsr_resolution (1024), and the memory
estimate counts their buffers again.
The hidden-face method is no longer a processing option: the dashboard's picker,
/reprocessproperty and --make-model take only the ray backend (the occlusion's)
and denoising. Options stored before (made_options, process_options,
ugc_process_runs, --make-model arguments) that name toolbox or fast still
parse; the word is skipped.
Check: the settings page has no hsr_method or path settings and has
hsr_resolution; /reprocessproperty embree toolbox still works (toolbox
ignored); models made with the defaults keep their hashes (UGC tests).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HIPRT (MIT) behind the CMake option DLU_HIPRT (off): its headers come from its
SDK (HIPRT_ROOT, else ROCm's /opt/rocm) and are copied next to the servers; its
library is loaded when first used (hiprtew), as HIP or CUDA are by Orochi
(MIT, fetched pinned by hash; CUDA when its toolkit is found). The trace
kernels (nearest hit skipping the triangle a ray leaves, any hit) are compiled
the first time and kept in cache/hiprt. One GPU context for the process
(hiprt_device picks the GPU); the workers take turns on it. When HIPRT, the
GPU or a scene's upload fails, Embree is used instead, and the UGC server logs
why at start.
For a GPU the rays go in batches (UgcRays::Scene gets batch queries; the CPU
backends answer them a ray at a time):
- hidden faces: with a batch backend the paths are traced side by side, a
bounce at a time (the path code split into Start, Scatter and Bounce, the one
by one tracing unchanged); the same paths with the same random numbers, so
the same triangles are decided (tested with builtin side by side)
- the occlusion bake and the denoised icons' traced occlusion always ask in
batches (the same rays, the same results)
Its symbols are hidden: the servers export theirs (-rdynamic), and HIPRT's
library, which has an Orochi of its own, would otherwise call ours.
Check: configure with -DDLU_HIPRT=ON on a machine with ROCm (or HIPRT's SDK)
and an AMD RDNA or NVIDIA GPU; UgcServer --make-model x.lxfml out hiprt; the
UGC tests (hits, hidden faces and occlusion against builtin); a build without
it leaves everything as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Intel Open Image Denoise 2 (Apache-2.0) behind the CMake option DLU_OIDN (off):
an installed OIDN is used when found, else Intel's release package (pinned by
hash) is downloaded and its libraries copied next to the servers.
A denoiser only removes noise that differs from pixel to pixel; the icons'
occlusion comes from the bake, per vertex, which it leaves as it is (checked:
white noise 0.17 -> 0.006 relative spread, per-vertex blocks unchanged). So
with denoise=oidn a model's icon is drawn from model.noao.nif (its colors
before the bake) with its occlusion traced per pixel of the supersampled image
(denoise_samples rays, default 4, with the bake's distance and strength and the
ray backend), box filtered and denoised at the icon's size, guided by the colors
and normals. The model keeps its baked occlusion. Icons drawn again from stored
files use the stored model.noao.nif the same way.
OIDN works on a thread of its own; its time is charged to the job's thread
(UgcThrottle::Charge), so the CPU budget and the recorded CPU time include it.
Check: configure with -DDLU_OIDN=ON; UgcServer --make-model x.lxfml out oidn
and compare its icon.png with one made with off; an OFF build leaves icons as
they were.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
