Adopt the reviewer's suggestion on #9396: instead of counting how many
skips each duplicate identity has earned, record which duplicate
identities have already built once. The first occurrence builds, the
rest are skipped. Distinct-identity loops that only become duplicates
after point mapping behave as before. Set semantics are also robust if
the helper is ever driven over the same shell twice, where a consumed
counter would under-build.
Port of #8772 to v0.9.0 (open_cascade_kernel::faceset_helper is the
renamed class, same logic). When an IfcConnectedFaceSet repeats the
same IfcFace, wires() still dropped every occurrence via duplicates_,
leaving a hole in the shell; this counts the redundant occurrences
and skips only those, keeping one copy of each repeated face.
ifcopenshell.geom.create_shape constructs a geometry kernel, converter and
mapping on every call and discards them afterwards. For hybrid kernels the
construction alone rescans the plugin directory per component, which #9417
addresses on the C++ side by caching the resolved component ids. This
prototypes the alternative aothms suggested in the #9417 review: reuse the
constructed kernel itself from Python.
The wrapper gains a geometry_kernel class holding the converter (and through
it the kernel, mapping and conversion caches) built once per
(geometry_library, file, settings) triple, with create_shape(instance[,
representation]) delegating to the same helper the free create_shape uses,
now split so both paths share one body. Python gains
ifcopenshell.geom.kernel(settings, file, geometry_library) mirroring the
iterator constructor signature. The binding is fixed at construction:
settings are copied the way converter already copies them, and a guard
rejects instances from a different file because the mapping is file-bound.
On PGSuper_Import_Model.ifc (322 products, single-threaded, arm64 Linux),
hybrid-cgal-simple-opencascade drops from 71.97 ms/call with per-call
create_shape to 23.35 ms/call through the reused kernel, matching the
iterator (22.74 ms/call) without needing the #9417 cache; plain opencascade
is unaffected (23.30 vs 22.28 ms/call).
Constructing a hybrid kernel (e.g. hybrid-cgal-simple-opencascade) ran
find_kernel_match once per component on every construction. Each call
creates a plugin manager, walks the plugin search paths and loads every
geometry_kernel_* module from disk before matching the component name.
Named kernels only pay this once because they are registered in the
static registry after the first load, but a hybrid name is never
registered, so every iterator or create_shape construction repeated the
full scan (measured 28-100 ms per component on the reference model,
~70 ms extra per create_shape call; far worse on platforms where
library loading is expensive).
Cache the resolved component backend ids per hybrid name. The first
construction still performs the full discovery and registers the
matched modules; subsequent constructions build the components straight
from the registry without touching the disk.
Follow-up to cb08191. The comma-separated pset value was resolved by
tool.Ifc.resolve_uri() as a single path before add_stylesheet() split it,
so normpath collapsed the whole string down to the last entry.
Split the value into paths first, then resolve each one, and concatenate
the stylesheets in listed order into a single <style> element.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
set_attribute_value_py turned every Python None into blank{}, so
createIfcSIUnit(None, ...) stored a null in the derived Dimensions slot.
The STEP output still shows "*" and the Python accessor still resolves
the value, but validate() reads the raw slot and reports "Attribute is
derived in subtype" for every such instance. 0.8.5 accepted the same
call and kept the derived marker. Mirror populate_derived_() and store
derived{} when the target slot is derived.
get_special_type_for_prop() reached the declaration through
entity.wrapped_data, which v0.9.0 removed when entity instances started
inheriting from the wrapper type directly. Ten ci-bonsai-daily tests
fail on it (quantification, pset editing, templated quantities). Use the
declaration property the way entity_instance.is_entity() itself does.
v0.9.0 already carries the ifcopenshell::file rename this harness was
originally adapted for, but moved Logger into the ifcopenshell namespace
too (logger::root() -> ifcopenshell::logger::root()) and never had a
BUILD_ONLY_COMMON_SCHEMAS cmake option - schema selection there has always
been via the SCHEMA_VERSIONS list. Verified with -fsyntax-only against the
system-installed v0.9.0 headers.
Logger::SetOutput was never called by the harness, so every parse
warning/error (e.g. "Overwriting instance with name #N") was silently
discarded. This directly cost time root-causing a leak: grepping stdout
for an expected warning found nothing, looking like it ruled out a
hypothesis that was actually correct, because the message was just never
printed anywhere.
Only enable it when the binary is given an explicit file argument
(single-input repro, e.g. `-runs=1 <file>`), not during a real campaign
against a corpus directory, where logging on every execution would
dominate the runtime. Verified with -fsyntax-only against this branch's
headers (a real CMake build of ifcviewer-wgpu wasn't attempted, same as
the harness's prior namespace-rename commit).
Generated with the assistance of an AI coding tool.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
ifcviewer-wgpu renamed IfcParse::IfcFile to ifcopenshell::file (header
moved to ifcparse/file.h) and Base::toString to Base::to_string. Update
the harness and README to match; verified with a syntax-only compile
against this branch's headers.
Construction of IfcFile already tokenizes, type-checks, and resolves
every attribute of every instance, so toString() isn't what makes
tokenizer/argument bugs reachable.
A coverage-guided libFuzzer harness (src/ifcfuzz/ifcparse_fuzzer.cpp) that
constructs IfcFile directly from in-memory input and calls toString() on
every parsed instance to force full lazy attribute evaluation, rather than
only observing IfcConvert's exit code from a fuzzed subprocess.
Gated behind a new BUILD_FUZZERS option (OFF by default) so it has no
effect on existing builds; enabling it requires a Clang toolchain built
with -fsanitize=fuzzer. -fsanitize=fuzzer itself stays scoped to the one
new target rather than going into the global compiler flags, since it
supplies its own main() and would otherwise break every other target
including CMake's own compiler checks.
Already found and fixed three real bugs this way: two null-pointer
dereferences (in header parsing and reference resolution) and a leak of
IfcSpfLexer on early return/exception during file scanning.
See src/ifcfuzz/README.md for build and usage instructions.
Bonsai's Pset/Qto editor could display a property or quantity's own Unit
override, but had no UI to author one -- only the project-level Project
Units panel existed, which sets defaults, not per-instance overrides.
Builds on the edit_pset/edit_qto Unit-wrapping support and the
get_unit_scale/get_candidate_units helpers added in the previous commit.
- bim/prop.py: Attribute gains unit_id (the STEP id of the property's own
override, 0 = project default) and unit_id_enum (the dropdown-driving
dynamic enum, "Default (<symbol>)" plus every candidate unit for the
attribute's measure type). update_attribute_unit_id converts the stored
value live when a different unit is picked, so the physical quantity is
preserved rather than the number being silently relabeled.
- tool/pset.py: is_measurable_special_type/get_candidate_units_for_special_type/
resolve_effective_unit/convert_attribute_unit support the picker and the
live conversion. get_special_type_for_prop classifies a property by its
value's own declared measure type, falling back to an explicitly-attached
Unit for generic numeric types (e.g. IfcReal) whose spec carries no unit
semantics of its own but which may still legitimately carry one. Seeding
in import_pset_from_existing ignores a stray Unit attached to a property
whose value has no numeric/measure semantics at all (e.g. text), which
used to crash trying to select an identifier the picker's enum items
never include.
- bim/module/pset/ui.py: the picker widget itself, next to the value field
in edit mode, gated on the attribute being measurable.
- bim/module/pset/operator.py: EditPset wraps measurable values with their
chosen Unit on save, for both properties and quantities. The qto
rounding-loop fix reaches into the wrapped dict instead of assuming a
bare float/int, which would otherwise zero out every unit-overridden
quantity.
Adds regression tests across all of the above, including conversion
correctness, explicit-clear/default round-trips, an unrelated sibling
property's override surviving untouched, and the stray-Unit crash guard.
edit_pset()'s unpack_unit_value() couldn't distinguish "no Unit dict was
passed" from "{"Unit": None, ...} passed to explicitly clear an existing
override" -- both collapsed to a bare None, and every consuming call site
checked truthiness, so there was no way to actually clear a previously-set
property Unit override once one existed. Fixed with a private _NO_UNIT
sentinel; bare (unwrapped) values still leave Unit untouched exactly as
before.
edit_qto() had no Unit-handling capability at all: neither
update_existing_property() nor add_new_properties() ever read or wrote a
quantity's Unit attribute. Added the same {"Unit": ..., "NominalValue": ...}
wrapped-dict convention edit_pset() already supports, disambiguated from
the pre-existing IfcPhysicalComplexQuantity dict convention
({"Discrimination": ..., "HasQuantities": ...}) by checking for a "Unit"
key -- a complex-quantity spec never contains one.
ifcopenshell.util.unit gains two small helpers:
- get_unit_scale(unit): dispatches to get_derived_unit_scale/
get_named_unit_scale depending on unit type, also used to de-duplicate
calculate_unit_scale()'s own inline dispatch of the same logic.
- get_candidate_units(ifc_file, unit_type): all units in a file matching a
given unit type, unlike get_project_unit()'s single-default lookup.
Adds regression tests for all of the above, including explicit-clear,
bare-value-preserves-override, and complex-quantity-routing-unaffected
cases.
Previously, unit symbols only appeared while a Pset/Qto was in edit mode
(pencil icon) -- the read-only summary view read raw {name: value} dicts
straight from ifcopenshell.util.element.get_psets(), a completely separate
path from the Attribute/unit_symbol machinery, so it never showed a label
even after the earlier fixes. This matters for the "someone in the field
just looking at values" use case, not just editing.
- bim/module/pset/data.py: switch to get_psets(verbose=True) to get each
property's own entity id, then resolve its unit symbol the same
override-aware way the edit-mode path does (tool.Pset.get_unit_symbol_for_prop).
Falls back gracefully (empty symbol) for IfcPreDefinedPropertySet
attributes, which aren't IfcProperty entities and can't carry a Unit
override.
- bim/module/pset/ui.py: read-only value button now shows "250 mm" instead
of just "250".
Also adds the regression tests planned but not yet committed:
- test/tool/test_pset.py: edit a property with its own Unit override and
write it back, confirming no rescale and the override survives.
- test/bim/test_prop.py (new): get_display_name() falls back to the plain
name (no crash) when no unit is resolvable or the project has no units
assigned at all.
NominalValue is optional -- IfcPropertySingleValue permits a null value --
but get_property_unit() unconditionally accessed prop.NominalValue.is_a(),
crashing on any single-value property that's legitimately blank.
Also adds a regression test confirming IfcContextDependentUnit symbols
("each", "boxes", etc.) aren't shadowed by the IfcDerivedUnit branch added
in the previous commit.
IfcSIUnit.Dimensions is a schema-derived attribute that isn't computed for
Bonsai's SQLite-linked "large model" file representation, returning None
there instead of an IfcDimensionalExponents entity. #9278 added an
unconditional unit.Dimensions.LengthExponent access to every IfcSIUnit
processed by calculate_unit_scale(), so it crashed project loading for
any linked file, even ones with no unit prefixes at all -- not just the
prefixed-area/volume case the fix targeted.
Fixed by reading dimensions from the existing si_dimensions table (keyed
by the unit's stored Name, not the unresolvable derived attribute) instead
of unit.Dimensions.
See the PR discussion for a standalone reproduction script.
* Modifications to build C++ with Visual Studio 2026 and the v145 toolset.
* Fixes linker settings for rocksdb for Debug and Release builds
* module is a C++ 20 keyword. Explicitly stating namespace allows cpp20 projects to build against the library
* Fixes crash when initializing an object with the initialize function when some of the attributes are empty, {}, or omitted, std::nullopt
* cleanup for vs2026 v145 toolset per @aothms review
* Fixes bug, IfcCurveSegment.setStartLength was setSegmentLength in alignment_helper.cpp
* Bumps boost to 1.92
The imperial list ran the architectural scales from 1'=1'-0" down to
1/128"=1'-0", then restarted at 1"=10' for the engineering scales. Merge
both groups into a single sequence ordered by ratio, largest scale first.
The metric list was already ordered by ratio and is unchanged.
Also fix the enum cache invalidation, which compared the cached list's
length against hardcoded 13/31 while the imperial list has 32 entries, so
switching a scene from imperial back to metric kept showing imperial
scales. Track the unit system the cache was built for instead.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
First v0.9.0alpha0 binary set, so the version prefix moves with it.
The bump trackers had drifted (bonsai's OLD pointed at 3e7b739 while
ifcopenshell-python pinned e333c1c), so this was done by hand;
'make bump' works again from here.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Upstream binary builds no longer produce macos64 zips (build_osx builds
arm64 only since wgpu Qt), and Blender dropped Intel Mac support in 5.0,
so there is nothing left to package for that platform.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The coarse motion threshold (15 px vs the 3 px still floor) followed the
per-frame "did the camera move" test directly. During a slow drag on a
janky main thread — the 66-model web session at 20 fps, mouse events
coalesced — some frames see no camera change, so the cull alternated
between thresholds every few frames: 84% of the visible set vanishing
and reappearing (139k <-> 22k objects in the log), with a full
visible-set re-upload at each flip feeding the very jank that caused the
gaps. On screen it read as the model sporadically jumping and returning
while orbiting slowly, easing as streaming and caching settled — which
is exactly how it was reported.
The motion state now latches: any camera movement arms it, and it only
drops after 250 ms of stillness, with the render loop kept alive over
the hold so the fine-threshold re-cull actually runs in an on-demand
loop. A drag degrades once at its start and restores once shortly after
it ends. Measured with a deliberately gappy scripted drag: two
transitions for the whole drag where each 120 ms pause previously
flipped it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On the single-threaded web build the CPU cull WAS the frame: 52-60 ms
of a 60 ms frame at 640k instances (desktop hides the same cost across
cores via std::async, which web cannot use without COOP/COEP+pthreads).
Two changes, both also helping desktop:
- ModelGpuData::CullInstance packs the six AABB floats and three ids the
cull reads into 40 contiguous bytes. InstanceInfo is 232 bytes with
the AABB 200 bytes away from the ids, so the walk paid two or three
cache lines per instance. Rebuilt by rebuildCullInstances at model
apply and inside uploadInstanceRecords, which every recompose,
transform and colour-override change already funnels through.
Measured on web: 94 ns/instance -> 36 ns/instance during a continuous
orbit (~2.6x).
- render() re-culls only when a cull input changed: the camera, a
cull-relevant setting (contribution px, LOD px, x-ray, HiZ on/off), a
fresh HiZ pyramid, or scene_epoch_ — bumped by chunk residency,
visibility, colours, transforms, model add/remove/hide/unload. A
frame requested for an overlay redraw, pick feedback, or a streaming
tick where nothing landed draws from the buffers the last cull
uploaded and skips the walk entirely. Benchmarks are exempt so bench
numbers keep measuring the real cull.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Filling only from the viewer's reads meant the cache converged on the
bytes the camera had needed — a user had to orbit every model into view
(unloading others to get there) before an entry could finish. Now a
filling entry fetches its uncovered spans in order, 8 MB at a time,
whenever the viewer has been quiet for 1.5 s, yielding the moment real
reads resume so interactive streaming always wins. A 42 MB model that
levelled off at 85% viewed now completes seconds after load with no
interaction.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Streaming a federation over the network re-downloads everything on every
visit: browsers do not populate their HTTP cache from ranged fetches
(measured at 0 of 78 range requests served from cache even with a strong
ETag). Host pages have started hand-rolling OPFS caches against the
library's own source seam — this is the second app to port the same ~350
lines — so the capability moves into the library.
The design keeps what those pages got right: the cache fills FROM THE
VIEWER'S OWN RANGED READS (no second download, and only bytes the camera
actually needed), entries are keyed by a hash of the URL and validated
by ETag (falling back to Last-Modified + size), and a byte-span ledger
guarantees a partial copy is never mistaken for a whole one. What it
fixes: writes go through a FileSystemSyncAccessHandle in an inline
worker — positional writes with no copy-on-open, where the pages'
createWritable({keepExistingData}) paid a whole-file copy per flush
(quadratic as the cache fills) and buffered up to 48 MB per model in JS
to compensate — the handle's exclusive lock makes a second tab fall back
to plain network instead of corrupting the entry; a complete copy now
opens when the server is unreachable (offline was dead before despite
the bytes being local); and entry names are hashes, where prefix-matched
sanitised names could delete a sibling model's cache.
viewer.cacheInfo() reports entries and the storage estimate;
viewer.clearCache(url?) drops one or all. Browsers without OPFS or sync
handles, servers without validators, and second tabs all degrade to
exactly today's network streaming.
Verified: the sample round-trips to zero range requests on reload, and a
42 MB model goes from 108 range requests to 5 on the second visit — the
85% the camera had viewed comes off disk, coverage honestly reports
incomplete for the bytes streaming never needed. The test server now
sends a content-hash ETag so the specs exercise real validation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On web there is no device-memory query, so the budget sat at the wasm
heap cap while the pool grew until Chrome's GPU process refused
(observed at 1920 MB on a 66-model session). Nothing acted on that
refusal: the cache kept the last byte, and the next attachment
reallocation (orbit resize, 76 MB) had to fail first — a few frames of
invalid-TextureView errors — before pressure feedback carved out room.
The refusal IS the query-less platform's device report. render() now
answers the first one by lowering the budget by the required-tier
margin and shrinking the pool to it, so attachments and model buffers
find headroom without ever failing. Desktop gets the same fallback for
drivers GpuMemory cannot answer for.
Reproduced under Playwright with a native process squeezing the GPU:
Chrome refuses at 512 MB, the margin (256 MB) is released on the next
frame, and the session continues with zero uncaptured WebGPU errors —
previously the same squeeze produced invalid-view frames before
recovery.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
getObjects() is on every real host page's path — a click hands back an
object id, and resolving it to a GlobalId/name/type needs the element
tables (the JS layer also builds its GUID index from this call). The
implementation materialised a vector of ElementRef (three fresh
std::strings per element), serialised the entire scene into one JSON
string grown by +=, and UTF8ToString'd the whole thing — several
hundred MB simultaneously alive at ~600k elements. The wasm heap never
returns pages, so that transient became the session's permanent floor.
ViewportCore::visitModelElements hands out one model's elements as
slices into its string table (no per-element copies), and the export
serialises straight from those, one model per batch, reusing one string
whose capacity grows only to the largest model. The JS side accumulates
batches and resolves the same array as before — the page API is
unchanged. Peak is now one model's JSON instead of the scene's.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Petru Conduraru