When a Text/MText entity's anno-scaled glyph height projects to fewer
than 4 screen pixels, drop the per-glyph stroke tessellation and emit
a single filled rectangle in the text's color instead. The user still
sees "text lives here" at the original color and footprint, but the
expensive font-stroke generation is skipped for thousands of unreadable
labels at fit-all / zoom-out views.
Pre-boosts the fill color to compensate for the 0.45 ambient-occlusion
dim the face3d pipeline applies (it's meant for 3D-shaded mesh solids,
not 2D text greeking).
Note: applies only to top-level entities; Text/MText *inside* block
defs is already tessellated into the block cache at build time, so
greeking those would require either a cache rebuild on zoom or a
parallel "low-detail" cache. Left for later if real-world test
suggests it's needed.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Frustum culling only helps at zoom-in (offscreen geometry); zoom-out
still tessellated and uploaded every wire even when an entire entity
projected to a single pixel. Now we also drop wires whose world AABB
projects smaller than 0.5px in the active view — a fit-all view of a
city plan skips sub-pixel symbols, lines, and curves before they ever
hit the GPU.
- `Scene` captures `world_per_pixel` each render from camera ortho_size
/ widget height.
- `expand_insert` / `tessellate_entity` take an optional
`world_per_pixel`; per-LocalWire / per-nested / whole-Insert /
top-level entity AABB → pixel-size test → skip if < MIN_PIXEL_SIZE.
- One-off paths (hit-test, viewport pane) pass `None` so they still
see full geometry regardless of zoom.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Two bugs in the new frustum culler that left the GUI rendering nothing
on file open:
1. `cam.target` is already in **local** f32 space — `fit_to_bounds`
populates it from local wire coords — so subtracting `world_offset`
again pushed the view rect to the negated centroid, with the actual
geometry far outside.
2. The first call to `entity_wires()` after open comes from `fit_all()`
itself, while `camera_generation` is still 0 and the camera is at
its default origin. Culling against that rect discarded every wire
→ fit_all saw an empty list → camera never moved → the user-visible
render stayed empty forever.
`view_world_aabb` now returns `None` until `camera_generation > 0`
(letting fit_all collect a real extent), then derives the view rect
directly from `cam.target` in local space.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Block defns and top-level entities now know their world-space XY
AABB, and `tessellate_entity` / `expand_insert` early-out for anything
that doesn't intersect the camera's view rect. Tested commit-only on
time_open (correct subset emitted with default camera); GUI run
pending user verification.
Cache plumbing:
- `LocalWire` / `BlockDefn` carry `aabb_local` (post-build recursive
union over nested defns).
- `Scene::view_world_aabb()` derives a 25%-margined world rect from
the active camera + last-seen render aspect.
- `wire_cache` / `paper_sheet_cache` keys widen to
`(geometry_epoch, camera_generation)` so pan/zoom invalidates the
cached wire list — and pipeline `cached_epoch` widens to match
so the GPU re-uploads on view change.
- `XrefStatus::Unloaded` ack already shipped — orthogonal to this.
Also adds `examples/xref_states.rs` to dump per-block_record xref
metadata for diagnosing the is_loaded / unloaded distinction.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Tessellates each block_record once into local-coord LocalWires; Insert
references transform-copy + ByBlock-resolve the cache rather than
explode+re-tessellate per reference. Same-style wires inside one Insert
are NaN-batched into a single WireModel to keep wgpu buffer counts
sane — large blocks that previously bailed to a marker now render in
a handful of buffers each.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
A real DWG with an xref containing ~74k sub-entities froze on the first
render: Insert tessellation exploded the block, style-resolved every sub
serially, and one bad sub could hang the loop forever.
- Insert short-circuits to an insertion marker when the referenced block
has more than INSERT_SUB_LIMIT (5_000) sub-entities. Documented as
temporary in RENDERING_OPTIMIZATION.md.
- Re-run purge_corrupt_entities after resolve_xrefs so parser-garbage
inside xref content is caught (initial purge runs before xref load).
- build_derived_caches: single pass over entities collecting handles
per kind; tessellation reads via doc.get_entity instead of cloning
every EntityType. Image decode now parallel too.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Remove MAX_SEGS_PER_WIRE = 100_000 and MAX_TOTAL_SEGS = 3_000_000.
These were stop-gap GPU-memory guards that mutated visible output:
single complex entities were cut mid-geometry, and entire wires past
the 3M total were dropped by arrival order with no spatial awareness.
Either is worse than honestly drawing whatever the file contains —
truncated arcs and missing road segments look like bugs to the user,
not "we ran out of GPU memory."
The matching "Temporary Workarounds" section in
RENDERING_OPTIMIZATION.md goes with them. Per-entity LOD and
viewport culling (Phase 1 + 3.2 in the same doc) remain the right
long-term answer; until they land, prefer correctness over a soft
crash on pathological files.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Civil 3D files can legitimately carry profile geometry at huge Z
(10^6+ units) inside an otherwise small XY footprint — long-section
polylines on a road, for example. fit_all's outlier gate previously
filtered any point whose |x|, |y|, or |z| exceeded 10× the header
extent half-size, which dropped those polylines and made ZOOM
EXTENTS center on the main drawing only, leaving the profile
geometry permanently invisible.
The gate exists to reject origin-stuck entities and Ray/XLine far
points, both of which manifest as large XY excursions after
world_offset subtraction. Z-axis outliers are not a known garbage
pattern, so the gate now only checks XY.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
acadrust's DWG reader occasionally desynchronises on certain files
(observed on render.dwg: 1519 of 1763 LwPolylines arrived with non-unit
normals in the 1e200+ range and 100k phantom vertices). Tessellating
those entities allocates gigabytes and produces NaN-laden wires,
freezing the UI.
Add a pre-tessellation guard that scans freshly loaded documents and
drops any entity failing a cheap sanity check: non-unit normal,
non-finite or absurd coordinate, or vertex count above 100k. Count of
dropped entities is threaded through DerivedCaches and surfaced as a
warning in the command line so the user knows the file was patched up
on the way in.
Same file in DXF form parses cleanly (0 corrupt), so this is strictly
a DWG-path workaround until the upstream parser bug is fixed.
The new examples/inspect_polyline.rs is a one-off diagnostic that dumps
LwPolyline/Polyline2D/Polyline3D fields plus a corrupt-entity count;
useful for narrowing down the acadrust bug.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Entity converters were casting WCS coordinates to f32 before the renderer
subtracted world_offset, which destroyed sub-unit precision on drawings
sitting at large coordinates (typical UTM-style CAD files). The multileader
case made it visible — glyphs warped under rotation — but every entity
that emits TruckObject::Lines/SegmentedLines/key_vertices/fill_tris was
affected.
Type changes (in src/scene/acad_to_truck.rs):
- TruckObject::Lines / SegmentedLines: Vec<[f32; 3]> → Vec<[f64; 3]>
- TruckEntity::key_vertices / fill_tris: f32 → f64
All ~17 producer files updated to keep WCS coordinates in f64 right up to
the TruckEntity boundary. The consumer in scene/tessellate.rs now performs
`(x - ox) as f32` (f64 subtract, then cast) for points, key_vertices, and
fill_tris — preserving precision near the camera origin even when the
drawing sits at million-unit coordinates.
catmull_rom_pts (multileader spline) is now f64; the scene path widens its
offset-relative f32 ctrl points to f64 around the call and narrows back at
the output.
Notes:
- Table layout still uses glam::Vec3 (f32) internally; widened to f64 only
at the TruckEntity boundary for type compatibility. Separate cleanup to
switch its internal math to f64 is tracked.
- snap_pts and tangent_geoms remain f32 — they are UI feedback (snap
candidates, tangent hints) and the precision is sufficient there.
Multileader text was emitted as a single literal stroke pass with no MText
code handling, no rotation, no attachment, no line breaks, no per-piece
color and bad f32 precision under world_offset. Two separate render paths
both had the same bug, and one was masking the fix in the other.
This rewrite, applied to both paths:
- Strips MText format codes (`{\fArial Black|b0|i0|c162|p34;...}`) so the
rendered text is the actual content, not the markup.
- Honors text_rotation (preferring text_direction so transforms survive),
text_attachment_point (Left/Center/Right horizontal anchor), and
text_left_attachment for vertical alignment via a shared
v_offset_for_attachment helper.
- Splits on \P/\n/\N and word-wraps to text_width.
- Applies line_spacing_factor with the same 5/3 × height convention as
MText.
- Resolves the text style via text_style_handle (fallback STANDARD).
- Applies ml.scale_factor everywhere (arrow size, dogleg length, text
height, text width). anno_scale is now only applied when
enable_annotation_scale is set.
- Splits the scene-level output into separate WireModels so leader,
text, frame and background fill each carry their own colour
(ctx.text_color, ml.text_frame, ctx.background_fill_color).
- Mirrors text_direction, base_direction and root.direction in
apply_transform so a mirrored multileader's text follows the geometry
instead of staying at its original rotation.
- Computes the text origin and frame corners in offset-relative f32
space — drawings sit at large absolute coordinates and casting to f32
before subtracting world_offset was wrecking the rotated sub-glyph
precision.
Mirrored INSERTs reverse OCS handedness, but acadrust's explode used to
keep the original normal — making CCW-around-normal arcs and ellipses
sweep the wrong way. The fix lives in acadrust 7c711e9, which flips the
new normal when det < 0 and reprojects center/major_axis/angles into the
new OCS. H7CAD's fix_mirrored_arc workaround is no longer needed.
Use reflection (TAU - sa, TAU - ea) for CW arcs and span.abs() for
segment count, consistent with the CircularArc pattern in both
hatch boundary (mod.rs) and tessellate (tessellate.rs).
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Color was read from the stale HatchModel cache instead of the live
entity, so property-panel color changes appeared on boundaries (wire)
but not on the hatch fill pattern.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Adds two-layer protection against large entities crashing the renderer:
- MAX_SEGS_PER_WIRE (100K): per-entity hard cap, excess segments truncated
- MAX_TOTAL_SEGS (3M): scene-level budget in upload_wires, excess wires skipped
Both limits log a warning to stderr when triggered.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
PolyfaceMesh and PolygonMesh were wireframe-only. They now render with
the same flat-color solid fill as Face3D entities.
- Add fill_tris: Vec<[f32;3]> to TruckEntity and WireModel
- PolyfaceMesh: fan-triangulate each face into fill_tris
- PolygonMesh: triangulate each grid quad into fill_tris
- tessellate.rs Lines branch: apply world_offset to fill_tris
- Face3DGpu::from_wires: collect fill_tris from all wires in addition
to Face3D key_vertices quads
- upload_face3d: pass all_wires so fill_tris are batched into one draw call
Parsing + DerivedCaches construction now runs entirely in a dedicated
OS thread inside open_path(). FileOpened handler only assigns the
pre-built caches, keeping the UI thread free during file load.
- Add build_derived_caches(): parallel hatch + mesh tessellation via rayon
- Extend Message::FileOpened / open_path return type with DerivedCaches
- Drop populate_*_from_document() calls from FileOpened handler
- Hatch pattern_scale multiplied by annotation_scale in model space so
pattern density stays consistent when CANNOSCALE changes
- Annotative check now uses AcAnnoPO xdata (R2007+ marker) instead of
relying solely on group-code-293 flag; falls back to is_annotative
for files with no xdata (old DXF backward compat)
- acadrust patched to local path for is_annotative field
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
- Scale popup dropdown replaces cycle behavior; paper-space viewport scale
editable when a viewport is active/selected
- CANNOSCALE persisted via DXF $USERR1 (read on open, write on save)
- PSLTSCALE: linetype dash lengths scaled by viewport anno_scale so dashes
appear uniform physical size in paper space
- MultiLeader: arrow size, dogleg length, and text height scale with
anno_scale; leader line paths (world coords) stay fixed
- Leader: arrow size and hookline length scale with anno_scale; also
corrects world_offset subtraction that was previously deferred to the
TruckObject::Lines path
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Viewport-projected wires now clip at pixel level to their viewport boundary.
- WireModel.vp_scissor: Option<[f32; 4]> — paper-space bbox set on projection
- WireGpu.vp_scissor carries it through to the draw loop
- Pipeline.compute_wire_scissors() converts paper→NDC→pixel each frame
- render() calls set_scissor_rect() per wire, reset after viewport group
- INTEGRATION_GAPS.md: Viewport ⚠️ → ✅, all 51 gaps resolved
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
LwPolyline PLINEGEN=0 (DXF default): emit SegmentedLines with NaN
separators between vertex-to-vertex segments; wire_gpu resets
arc-length to 0 after each NaN so linetype pattern restarts per vertex.
PLINEGEN=1: existing Contour path unchanged (continuous pattern).
WireModel gains plinegen field propagated through tessellate pipeline.
Solid3D/Region/Body: call SabReader::read() when parse_sat() returns
None due to binary SAB format (DWG files / newer AutoCAD exports).
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
cxf.rs: missing non-ASCII glyphs now emit eprintln! once per font+char
pair via a global Mutex<HashSet> dedup set. ASCII chars silent (expected).
INTEGRATION_GAPS.md: ellipse partial arc snap marked done — endpoints are
correctly emitted via key_vertices → SnapType::Endpoint in snap engine.
Unicode warning gap closed. plinegen behaviour documented.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Dimension measurement origin points (first_point, second_point,
definition_point, angle_vertex, etc.) are now emitted as Node snap
hints so OSNAP Node mode can locate them precisely without relying on
the tessellated geometry wire alone.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Hatch boundaries with arc or ellipse edges were silently dropped from
the wire outline, leaving gaps in the hatch contour. CircularArc and
EllipticArc boundary edges are now tessellated as polyline segments and
their centers are emitted as Center snap hints.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Entities carrying a non-standard extrusion normal (normal ≠ (0,0,1)) were
rendered flat in the XY plane. The DXF arbitrary-axis algorithm is now
implemented in scene::transform::{ocs_axes, ocs_point_to_wcs} and applied
during tessellation of Arc, Circle, and Line so they render correctly on
tilted planes in 3D DXF files.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Block INSERTs are expanded into sub-entity wires for rendering but the
OSNAP Insertion hint at the block reference origin was never generated.
Now a zero-geometry WireModel carrying the Insertion snap hint is appended
so OSNAP Insertion mode can locate the block origin.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Debug output revealed the sheet viewport has id=0 (DWG reader never sets
the field), scale=0.8965 (not ≈1.0, so the 0.02 tolerance missed it),
and center=(0.0, 0.0) — exactly at the paper-space origin.
In DWG format the overall/sheet viewport always has its center at the
origin. Content viewports are placed at their actual paper positions and
always have a non-zero center. Switching from the scale heuristic to the
center-position check makes the detection reliable regardless of what
scale value the sheet viewport happens to carry.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Remove unused `screen_delta_to_world` camera method (superseded by the
corrected pan_active_viewport formula).
Fix `is_content_viewport`: the old geometry fallback required the viewport
center to be near the origin, but DWG files never write the viewport id
field (stays 0) and the sheet viewport center is the paper centre
(e.g. 105, 148.5 mm for A4) — not the origin. This caused the sheet
viewport to be classified as a content viewport in DWG files, making it
both tessellated (border visible) and enterable via double-click.
The new heuristic uses only the scale: a sheet viewport has scale ≈ 1:1
(view_height ≈ paper height); true content viewports have deliberate
drawing scales (1:50, 1:100, …) that are far from 1.0.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
The old formula (distance × 0.001) was a fixed approximation that ignored
the paper camera's current zoom level and the viewport's paper-space size,
causing panning to feel sluggish.
New formula: model_per_pixel = (2 × paper_half_h / canvas_height) × (view_height / vp_height)
This converts screen pixels → paper-space mm → model units correctly,
making the content follow the cursor regardless of paper zoom or scale.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Entity linetype_scale is a per-entity multiplier applied on top of the
document-level $LTSCALE header variable. The previous code used only the
entity scale (defaulting to 1.0), ignoring $LTSCALE entirely. Drawings
with a non-default LTSCALE showed dash patterns at the wrong size.
Effective scale is now: document.header.linetype_scale * entity linetype_scale.
Applied in both render::render_style_for and the complex-linetype path in
tessellate_entity.
Fixes part of #10
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Two layout-related fixes:
1. Sheet viewport (id=1) was rendered as a grey rectangle border in
paper-space layouts. It is the paper boundary, not a user viewport,
and should not appear. Now filtered via is_content_viewport() before
tessellation. The redundant id==1 color branch is removed.
2. Box crossing-selection (right-to-left drag) only detected wires whose
vertex points fell inside the selection box. Large entities like viewport
frames could be visually crossed without any corner inside the box and
would not be selected. Added segment-box edge intersection tests so any
wire segment that crosses the selection rectangle boundary is included.
Fixes part of #10
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
- Grip points in paper space now use the 2-D linear to_px transform
(tx/ty/ortho_size) instead of the 3-D view_proj matrix, matching the
coordinate system used by PaperCanvas
- world_offset is no longer subtracted from paper-space entity grip
positions — paper-space coordinates are already unoffset (same rule
as wire tessellation in wires_for_block)
- MTEXT v_offset corrected for all attachment points: CXF glyphs sit
on the baseline and extend upward, so Top attachment needs v_offset=-h,
Bottom needs (n_lines-1)*line_h, Middle is the midpoint of the two
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
paper_limits() now reads paper_width/paper_height/plot_rotation from the
Layout object (populated by a patched acadrust from the embedded DWG
PlotSettings). This fixes layouts where min/max_limits do not match the
physical paper size (e.g. A4 portrait content inside A3 landscape limits).
DXF files use codes 44/45/73 from raw_plot_settings_codes; DWG files use
the PlotSettings block embedded in the LAYOUT object. Both fall back to
min/max_limits when paper_width is zero.
Depends on HakanSeven12/acadrust feat/layout-paper-dimensions (PR #21 on
hakanaktt/acadrust); using that fork until the PR is merged.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
- Save camera to named View table entries (H7CAD_Camera_Model / H7CAD_Camera_<layout>)
on every navigation change; restore from these entries on file open and tab switch.
acadrust's DWG writer overrides *Active VPort view_height/view_center with
zoom-extents values, but leaves the View table untouched, so this survives save.
- Fix pitch recovery: clamp vd.z to [-1, 1] instead of [-0.999, 0.999] so that
asin(1.0) = π/2 exactly for plan/top views, preventing the subtle tilt on reload.
- Fix IEEE 754 edge case: atan2(+0, -0) = π for plan view direction (0,0,1);
special-case yaw = 0 when both vd.x and vd.y are near zero.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
view_dir = rotation * Z = (sin(yaw)*cos(pitch), -cos(yaw)*cos(pitch), sin(pitch))
so yaw = atan2(x, -y), not atan2(x, y).
The sign error caused a 180-degree Z rotation on every save/reload cycle.
Applies to apply_active_vport_camera, apply_sheet_viewport_camera, and
camera_for_viewport.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
- On file open and layout tab switch, restore the camera from the
document's saved view (*Active VPort for model space, sheet viewport
for paper space) instead of fitting to wire bounds.
- After any zoom/pan/rotate, write the new camera position back to the
document and mark the file dirty so it is saved with the file.
- Uses last_synced_camera_gen to avoid marking the file dirty on the
initial camera restore.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Model entity wires are tessellated with world_offset subtracted for GPU
precision, but view_target/view_center were raw model coordinates.
Subtracting world_offset from view_target aligns both coordinate systems
so model content projects correctly into viewport rectangles in paper space.
Also fixes scale priority (view_height before custom_scale) and adds
view_center DCS offset to the projection, completing the paper-space
viewport rendering.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
The acadrust DWG reader never assigns a viewport id (always 0).
Some DXF exporters use id=-1 for all viewports. In both cases the
previous id > 1 check excluded every viewport from content rendering.
Replace is_user_viewport_id(id) with is_content_viewport(vp) that
uses a geometry heuristic for unknown ids (0 / negative):
- id=1 → always sheet viewport (excluded)
- id≥2 → always user viewport (included)
- id=0 or id<0 → sheet if center≈(0,0) AND scale≈1.0, else user
This allows DWG files and non-standard DXF files to display model
content inside their paper-space viewports and be entered via
double-click.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
Two root causes prevented paper-space viewports from loading correctly:
1. acadrust's ViewportStatusFlags::from_bits() maps bit 0 → is_on, but
the real DXF/DWG spec uses bit 15 (0x8000) for "viewport on". Post-load
fixup in io/mod.rs now corrects this for both DXF and DWG files: when
bit 0 is clear but bit 15 is set, is_on is forced true and locked is
derived from bit 14 (0x4000).
2. Some DXF exporters write id=-1 for all viewport entities instead of the
standard id=1 (sheet) / id≥2 (user) convention. The is_user_viewport_id()
helper now treats any id ≠ 0 and ≠ 1 as a user viewport, covering both
the standard id≥2 and the non-standard id=-1 cases.
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
