Files
solidcraft/vendor/README.md
T
semmelrindeandClaude Opus 5.5 02981d6bcf Lost GPU: a dropped frame no longer crashes the window's close (#7148) (#7400)
* claim: #7148

* Lost GPU: a dropped frame no longer crashes the window's close (#7148)

On a lost device wgpu's present and discard return early and leave the
acquired frame's image in the surface, holding a swapchain acquire
semaphore. Dropping the surface later (the window closing for the
restart after the loss) then panicked in wgpu-hal. The painter now holds
the frame as AcquiredFrame, which destroys a second handle to the
frame's texture after the frame is presented or dropped; wgpu-core's
destroy of a surface texture takes the image back without checking the
device, and does nothing once a present or discard took it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-27 14:48:04 +00:00

338 lines
26 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Vendored third-party crates
Copies of third-party crates that SolidCraft builds with local fixes, wired in with
`[patch.crates-io]` in the workspace `Cargo.toml`. They keep their own licence; changed lines
are marked `SolidCraft:` in the source.
## truck-shapeops 0.4.0 (Apache-2.0, © RICOS Co. Ltd., https://github.com/ricosjp/truck)
Changes:
- `transversal/intersection_curve/mod.rs`: reject empty/single-point polylines and
collapsed or non-finite segments before endpoint indexing and normalization. A hidden
constructor re-export lets the kernel workspace tests exercise the actual vendored code.
- `transversal/integrate/mod.rs`: mesh both shells with `robust_triangulation` (faces made by
earlier booleans have boundaries that lie on their surface only within tolerance).
- `transversal/divide_face/mod.rs`: when projecting a loop onto a face's surface, fall back to
the nearest parameter when the exact search fails (intersection curves are approximations).
- `transversal/integrate/mod.rs`: the result's shells and faces keep the order of the shell
they come from (truck's connected components followed hash order of face addresses, so a
boolean's faces came out in a different order on every run).
- `transversal/loops_store/mod.rs`: a point at an end of a boundary curve (within the point
tolerance) is that end, whatever its parameter: a polyline's parameter counts segments, so
an intersection curve ending 2e-7 mm from another's end could read as inside it, cut it
into a sliver with a vertex of its own, and leave the result's shell open (#391).
- `transversal/loops_store/mod.rs`: an intersection edge added to a face at a vertex that more
than one wire visits (several curves end on it: the touching point of two equal cylinders) goes
into the corner that holds its direction, swept counterclockwise about the face's normal from
the wire's outgoing edge to the way back along its incoming one (`add_edge_by_angle`, told on
the polygon loops and followed by the geometry loops). Taking the first visit of each end
joined wires that are different regions of a face and left a wire visiting a vertex twice.
With no direction to go by, the edge goes where both of its ends are in one wire.
- `transversal/divide_face/mod.rs`: a face piece whose wire is not simple makes the operation
give up (`Face::try_new`) instead of panicking.
- `transversal/divide_face/mod.rs`: a loop whose polygon in the face's parameters has no area
(a small hole's rim divided at the tolerance of a much larger operand: one chord per half
circle) is divided again, finer, before its area's sign tells a hole from a face piece. The
sign of a zero area was random: the hole became a face of its own, inside out, and the face it
was cut in stayed whole.
- `lib.rs`: `tr!` reports where an operation gives up when `SHAPEOPS_TRACE` is set; compiler
warnings no longer fail the build.
- Face and edge history (`transversal/history.rs`, new): `and_with_history` / `or_with_history`
also return, for each face of the result (in `face_iter` order), the operand face it is a
region of and whether it is that face unchanged, and for each edge (distinct edges in order
of first use along the faces) the operand edge it is (a piece of) or the two operand faces
whose intersection made it. The code that creates faces and edges records them as it goes:
`divide_face` (face pieces), `loops_store` (intersection edges, split edges, edges rebuilt on
a moved vertex), across the passes that fold in the operands' extra shells. `and` / `or` are
wrappers that drop the history; the `*_with_history` forms return the same solid with the
history as an `Option` (`None` where a face or edge went unrecorded).
- `transversal/integrate/mod.rs`: `meshes` and `and_with_history_meshed` take the first
operand's shell meshes made beforehand (copied for each use, since the loops stores cut their
edges): a large body cut down to the regions of many small ones is meshed once, which was most
of each operation's time.
- `transversal/integrate/mod.rs`: in an AND, a face no intersection divides (a shell that meets
nothing of the other) is placed by the crossings of a ray with the other shell counted as each
face is oriented, not by its surface alone, and an inward shell (a void, a cut's inverted tool)
holds what lies outside it. Faces with reversed surfaces (a revolved body's) counted the wrong
way, and a void dropped every face outside it: a cut beside a sealed void came out empty or
with the body gone. An OR keeps truck's count.
## truck-geometry 0.5.0 (Apache-2.0, © RICOS Co. Ltd., https://github.com/ricosjp/truck)
Changes:
- `decorators/intersection_curve.rs`: `IntersectionCurve::parameter_division` divides as
`algo::curve::parameter_division` does (the same split points, in the same order), but each
point's projection onto the two surfaces starts from the parameters of the point that begins
its piece, instead of a search of the whole surfaces (a grid pre-search of a B-spline surface
for every point: a small hole on a large curved face took seconds, #1038). A hinted result
farther from the leader's point than the curve's tolerance (the projection settled on another
part of a surface, such as the far side of a small hole) is dropped for the search from
nothing, as before. `crates/kernel/src/issue_1038_tests.rs` holds it to that search.
- `decorators/revolved_curve.rs`: a point on a revolved surface's seam (angle 0, the same place
as 2π) takes the end of the period nearer the hinted angle. The angle was decided by the sign of
a cross product that is rounding noise there, so a boundary followed point by point could come
back to the seam at the other end of the period and fold its parameter polygon over itself (no
intersection loop then lay inside the face). `crates/kernel/src/seam_faces_tests.rs` holds it
to a cone read from STEP.
- `nurbs/bspsurface.rs`: `uuder` and `vvder` multiply by p·(p − 1), as `BSplineCurve::der2`
does. They multiplied by p only, so the second derivatives of every B-spline and NURBS surface
of degree p ≥ 3 were 1/(p − 1) of their value (a cubic surface's curvature read as half, #763).
`crates/kernel/src/surface_second_derivative_tests.rs` holds them to finite differences.
- `decorators/intersection_curve.rs`: `IntersectionCurve::parameter_division` with a tolerance
below truck's `TOLERANCE` (1e-6), or one that is not a number, divides at `TOLERANCE` instead
of panicking in truck's tolerance assertion; a join failed on that panic (#4623).
`crates/kernel/src/issue_1038_tests.rs` holds it to the division at 1e-6.
- `decorators/intersection_curve.rs`: where the hinted projection behind `parameter_division`
and the search from nothing both fail, the leader's point stands in, the division stops, and
the failure is recorded for `take_projection_failure`. It fell back to `subs`, which unwraps
that same failed search: a revolve cut whose sphere's pole sat on a box's edge panicked
dividing an edge (#5562).
- truck-shapeops `transversal/divide_face/mod.rs`: an edge whose division recorded such a
failure gives up the face division (`None`), so the Boolean attempt fails as quickly as the
panic made it, and the kernel goes on to its other tries. Carrying the approximate points on
made one test's Boolean run for minutes.
## truck-stepio 0.3.0 (Apache-2.0, © RICOS Co. Ltd., https://github.com/ricosjp/truck)
Only the writer (`out`) is used; SolidCraft reads STEP with its own reader.
Changes:
- `out/geometry.rs`: an `INTERSECTION_CURVE`'s second surface was written at the first
surface's entity index, so every boolean-made edge produced duplicate entity ids.
- `out/geometry.rs`: surfaces of revolution carry the right `same_sense` for inverted and
mirrored revolutions (ISO 10303-42 parameterises them by angle then profile, so their normal
is the opposite of truck's; a mirroring transform reverses it again), and the modeling-surface
wrapper passes the revolution's sense through instead of always `true`.
- `out/geometry.rs`: a surface of revolution under a transform that commutes with turning about
its axis (a scale across the axis that differs from the scale along it, as a shell or offset
makes of a revolved face) is written as the transformed profile revolved about the
transformed axis. Only uniform scales were taken. Other transforms are still refused: the
writer stops with a formatting error, which SolidCraft's `step.rs` now reports as an export
error (`to_string` panicked on it).
- `lib.rs`: compiler warnings no longer fail the build. Example and test targets are dropped
(their sources are not vendored).
## truck-meshalgo 0.4.0 (Apache-2.0, © RICOS Co. Ltd., https://github.com/ricosjp/truck)
Changes:
- `tessellation/triangulation.rs`: a boundary segment that crosses constraints already added
is split where it crosses by `add_split_constraint` (a vertex at each crossing, the pieces
added where they can be), not by spade's `add_constraint_and_split`, which asserts when a
split lands so near another vertex or constraint that the piece up to it is blocked. That
assertion lost whole bodies and operations (#4031, #4598); a blocked piece is now left out, as
truck did before #1653.
- `tessellation/triangulation.rs`, `closed.rs`: retry nearest trim projections outside a
spline's knot domain without the preceding point's hint, preferring a bounded fit within
the caller's chord tolerance, with a bounded damped least-squares fallback where
Newton's method diverges. Affine plane charts remain unbounded; modelling surfaces may
retain genuine continuation when no bounded fit is accurate enough. Geometrically closed,
smooth spline seams get a local periodic chart during meshing, with wrapped evaluation,
so a trim crossing the seam stays a small patch. The stored geometry is unchanged (#1974).
- `tessellation/mod.rs`, `triangulation.rs`: a separate provisional `preview_face`
path accepts at most 4,096 already sampled boundary points and a fixed grid of
at most 16 subdivisions per direction, including periodic seam connections.
It does not run adaptive subdivision or grid balancing. The kernel caller bounds
projection complexity and total allocation before using it; ordinary modelling
tessellation keeps its existing accuracy policy.
- `tessellation/triangulation.rs`: exact, robust and provisional-preview projections prefer an
equivalent in-domain parameter at collapsed surface edges, and snap to exact bounded
endpoints only at collapsed edges when that improves the spatial fit within the existing
search tolerance. Closed spline seams retain their distinct parameter endpoints.
This prevents huge arbitrary pole parameters from corrupting trimmed meshes (#796),
while retaining genuine extrapolation for unbounded planes. Work is bounded to four
endpoint evaluations per projected point; no search or tessellation tolerance is relaxed.
- `tessellation/triangulation.rs`: a boundary loop is projected onto its surface point by point,
each searched from the one before; the first has no hint. When that walk fails, it is tried
once more from the first point an unhinted search finds. An unhinted search can miss a point
that lies on the surface, such as the corner of a fillet strip on a surface extended past it,
and the face was then left unmeshed, which also failed every boolean with the body (#882). A
walk that succeeds is kept as it was: where a loop starts decides how it is laid out across a
periodic surface's seam.
- `tessellation/mod.rs`, `triangulation.rs`: edges are sampled through a new `TrySubs` trait
(`try_subs`, `approx_subs`), a bound of `PolylineableCurve`, instead of `subs`. truck's
`IntersectionCurve::subs` unwraps a projection onto its two surfaces that can fail to
converge, which panicked while meshing (#1252). A failed sample at an interval's end uses the
intersection curve's leader point, so the polyline stays finite; a failed interior probe is
NaN, which already counts as flat. `truck-shapeops`' `ShapeOpsCurve` carries the same bound,
and `truck-modeling` is now a regular dependency (it was a dev-dependency).
- `tessellation/triangulation.rs`: the inside test of a face's boundary skips loops whose
parameter box is clear of the point (it tested every segment of every loop, so a face with
hundreds of holes meshed in quadratic time).
- `tessellation/triangulation.rs`: meshed faces and edges are built unchecked: faces whose loop
runs along a seam edge both ways and closed edges (a full circle on one vertex) are valid,
but the debug-build checks rejected them (debug builds could not mesh them).
- `tessellation/triangulation.rs`: on a periodic surface, a boundary point's projected parameter
is moved by whole periods to within half a period of the point before it, however many periods
away the search put it (it was moved by two periods at most). A hinted search on a revolved
surface can return an angle millions of turns from its hint; the loop then spanned those turns,
the polyline closing it across the seam had millions of points, and meshing never finished,
which hung the STEP export that reads its own file back (#1985).
- `tessellation/triangulation.rs`: parameter searches that land outside a surface's bounded,
non-periodic domain are rejected (a hinted B-spline search can run off and extrapolate to
thousands of parameter units, and meshing that grid never finishes); the last-resort nearest
search is clamped into the domain.
- `tessellation/triangulation.rs`, bounded work (hostile files must not hang): edge polylines
and surface grids come from our own bounded divisions (at most 20 000 points per edge, a
grid budget per face shared out of 10 million points per shell, 24 refinement rounds; a
point that does not evaluate counts as flat instead of recursing). Surface grids refine only
the direction that bends, and a ruling left undivided gets cells comparable to the other
direction (long triangles from a trim loop to the grid's far ends cut through cylinders).
- `tessellation/triangulation.rs`: an edge's division evaluates each parameter once. A piece
carries the points of its ends and its midpoint is an end of both halves: 2 evaluations per
piece instead of 4, and the same polyline (a point of an intersection curve costs two grid
searches of its surfaces). `bounded_polyline` and `Deflection::turn2`/`turns` are public but
hidden, for the kernel's regression against the old division.
- `tessellation/mod.rs`, `tessellation/triangulation.rs`, for display: a `Deflection` is the
chord tolerance plus an optional bound on the angle one facet spans, and
`RoundMeshableShape::robust_triangulation_within` divides by it: an edge or a grid direction
within the chord tolerance is still halved while a piece turns more than the angle, until its
sag is under a floor (a radius that is small beside the tolerance was one flat facet).
Bounded work: a face that would be bounded by more than 4096 segments has its curved edges
divided at twice the angle, and again, down to the chord tolerance alone, and the angle
stops refining a face's grid at 16 384 points. Each edge is still divided once for both its
faces. `triangulation` and `robust_triangulation` divide by the chord tolerance alone, as
before.
- `analyzers/collision.rs`: the collision search builds each triangle's box once, filters the
sweep's open list by box overlap as it goes (it cloned the list for every triangle and built
both boxes for every pair), and runs the exact triangle tests on the rayon pool, in order
(sequential on wasm). Same collisions in the same order. The old search is kept as
`analyzers::collision_test_hooks`, behind a `test-hooks` feature that only
`solidcraft-kernel`'s tests turn on, and `crates/kernel/tests/mesh_collision_equivalence.rs`
compares the two.
- `tessellation/triangulation.rs`: the inside test of a face's boundary also skips runs of 32
segments whose box, grown by the tolerance, the test's ray cannot meet. Same answers. The
old test is kept as `tessellation::include_test_hooks`, behind a `test-hooks` feature that
only `solidcraft-kernel`'s tests turn on, and `crates/kernel/tests/tess_include_equivalence.rs`
compares the two.
- `tessellation/triangulation.rs`: the constrained Delaunay triangulation of a trimmed face
works in (u, v) scaled by the surface's mean speed along each (a power of two, so exact both
ways). In raw parameters, a loft side ruled across (about 30 mm) and long along the loft
(about 600 mm) is a strip stretched twentyfold, and a trimming node was fanned to boundary
nodes hundreds of millimetres away: a flat sheet through a joined tube.
- `tessellation/triangulation.rs`: every segment of a face's boundary becomes a constraint; one
that crosses or overlaps a constraint already added (an intersection edge's polyline can hook
back a few micrometres past its vertex) is split where they meet. truck dropped such a
segment and joined its start to the first later node it could reach, which on a joined tube
left a trimming curve's 130 nodes behind one constraint across the face: slivers from it
sagged millimetres off the surface, and a face pinched to a point (#1648) got no triangles.
All meshes; faces whose segments all go in mesh as before.
- `tessellation/mod.rs`: `Deflection::rulings` is now `Deflection::display`: a mesh to display
or export (the kernel's `tessellate` and display meshes), which gets the ruling grid lines,
and the speed-scaled triangulation above. The meshes a Boolean operation intersects
(`Deflection::chord`) keep truck's (u, v) and insertion order.
- `lib.rs`: compiler warnings no longer fail the build. Example, test and bench targets are
dropped (their sources are not vendored).
- `tessellation/triangulation.rs`: adaptive division also splits a cell that bows off its
corners' plane by more than the chord tolerance (a quarter of its twist `p00 - p01 - p10 +
p11`, across that plane, at the middle). A bilinear patch equals the blend of its corners the
division compared against, so a twisted ruled face (a loft between polylines) went out as two
triangles per face (#1666); a flat cell such as a cone's trapezoid between rulings is not split.
- `tessellation/triangulation.rs`: a shell's face meshes share their edges' points exactly. A
face's mesh puts each point at its parameters on the face's own surface, so a point of an edge
that lies off a surface within tolerance (an intersection-curve sample whose projection did not
converge, an imported edge) came out in a different place on each of the edge's faces, and the
exported mesh had a crack there. In display and export meshes (`Deflection::display`), a
vertex at a point of a face's boundary loops that is farther than `TOLERANCE` from the nearest
point of the face's edges now goes to that point, when it is within the chord tolerance
(`onto_edges`). Points within `TOLERANCE` stay on the surface, so exact geometry meshes as
before. The meshes Boolean operations intersect keep truck's points.
- `tessellation/triangulation.rs`: in display and export meshes (`Deflection::display`), a
direction left undivided along the rulings of a face ruled one way also gets a grid line
through each point of the face's boundary that lies inside the face along it
(`with_boundary_lines`), when there are at most 16 such points: corners standing into the
face, not a boundary curve running through it, which the triangulation already follows. A hole in a cone or cylinder smaller than one cell (a notch into a
chamfer) or a trim crossing the rulings of a narrow band (an oblique plane through a bore's
chamfer) had no grid point beside it, and the triangulation fanned from it to the far ends of
the face, many degrees round: those triangles cut through the surface and folded over each
other (a cone face with a small hole meshed several mm² too large, and the boolean that made
it was refused for its volume).
- `tessellation/triangulation.rs`: a face that comes out of the triangulation with no triangle
at all, though the points of its one boundary loop enclose an area in space, is given the
triangles of that loop (ears clipped on the plane of the loop, no point added, so the faces on
either side of its edges still meet it). Its two edges can be different curves that, sampled
as they are, run along the same line in the face's parameters (a lens thinner than the
tolerance), and the loop then has no area there. The shell's mesh had a hole. Faces that
mesh are untouched, and so are preview meshes (they have their own step grid). Only a face
with one loop of at most 256 points is filled, and not one whose loop crosses itself, whose
ears would overlap (the pinched face of #1648 is one): it stays as it was.
`crates/kernel/src/thin_face_tests.rs` holds it, `crates/kernel/tests/tess_loop_fill.rs` the
ear clipping (through the `test-hooks` feature).
## matext4cgmath 0.1.0 (Apache-2.0, © RICOS Co. Ltd., https://github.com/ricosjp/matext4cgmath)
truck-geometry uses it for a matrix's operator norm (an intersection curve scales its tolerance
by it whenever it is moved).
Changes:
- `solver.rs`: the Newton refinement of cubic and quartic roots stops after 100 steps. Near a
double root it could cycle forever: moving a body whose edges a boolean made (the retries
move both bodies and the result back) hung the app.
- `lib.rs`: compiler warnings no longer fail the build. Test targets are dropped (their sources
are not vendored).
## egui-wgpu 0.36.2 (MIT OR Apache-2.0, Emil Ernerfeldt and contributors)
Original package sources from https://github.com/emilk/egui/tree/0.36.2/crates/egui-wgpu,
with both original licence texts retained. No images or bundled documentation are copied.
SolidCraft changes:
- `src/setup.rs`, `src/lib.rs`: let a named adapter selector defer to the normal wgpu
request when the requested card is unavailable or incompatible with the window surface.
- `src/lib.rs`: use 1x MSAA on the selected OpenGL adapter before creating pipelines,
exposing the effective sample count so host paint callbacks can match it.
- `src/winit.rs`: use that same effective count for depth and resolve attachments.
- `src/renderer.rs`: when the GPU refuses a staging buffer for a frame's meshes (device lost
or out of memory), log a warning and draw nothing that frame instead of panicking (#1469).
- Other backends retain requested MSAA, without another adapter or device probe.
- `src/lib.rs`: clamp MSAA to the largest sample count the device supports for the surface and
depth formats (`supported_msaa_samples`, `RenderState::msaa_supported`): a saved 8x on a
device without it failed pipeline validation, a panic at start.
- `src/setup.rs`: the default device requests `TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES` when
the adapter has it, so 2x and 8x work where the adapter supports them (WebGPU only
guarantees 1x and 4x).
- `src/setup.rs`: the default device asks for the adapter's largest 2D texture (at least 8192)
instead of exactly 8192.
- `src/lib.rs`, `src/winit.rs`: `set_startup_hook` is told before each step of getting the
device and the window's surface up (instance, surface, adapter request, the chosen adapter,
device, renderer, surface configure, first present), so the app's crash journal can say which
step a crash inside the graphics driver happened in.
- `src/lib.rs`, `src/winit.rs`: a window larger than the device's largest texture gets a surface
clamped to it (`clamp_surface_extent`, aspect ratio kept) and egui is drawn at the matching
smaller scale, instead of a `Surface::configure` validation panic (#869). The resize path and
an unsupported surface no longer unwrap.
- `src/winit.rs`: acquire a drawable before preparing paint callbacks, so a skipped
surface frame cannot discard initialization commands for cached viewport meshes.
- `src/lib.rs`, `src/winit.rs`: surfaces are configured under `capture_errors` (error scopes):
a configuration the device refuses (D3D12 "Failed to wait for GPU to come idle before
reconfiguring the Surface", #1935) is logged, the frame is skipped and the configure is tried
again, instead of the device's uncaptured-error handler panicking. A streak of refusals warns
once (debug after), waits 16 ms doubling up to 1 s between tries and recreates the surface
every `SURFACE_RECREATE_AFTER` refusals (`ConfigureRetry`). Each surface keeps the deadline of
its next try: until it is due, no configure or recreate happens, whether the repaint comes
from input, an animation, a resize (its size is kept for the retry) or an outdated surface,
and the frame is skipped with a repaint asked for when it is due. The streak and the deadline
survive a recreate; a configure the device takes ends the streak, a recreate it asked for
included. A recreate for a lost surface goes ahead as before.
- `src/lib.rs`, `src/winit.rs`: the device's lost callback marks `RenderState::device_lost`
(`DeviceLost`), and the painter draws nothing on a lost device. Before a frame's uploads it
submits nothing (`DeviceLost::probe`), so a device the driver lost is found by that
submission rather than by the first queue write, which panicked inside wgpu-core
(`PendingWrites::activate`, #4172).
- `src/lib.rs`, `src/winit.rs`: a surface configure that meets a lost device is not taken as
configured (`configure_captured`): wgpu reports device-loss errors to no error scope, and on
D3D12 a driver reset found by the configure has already dropped the surface's configuration.
The next acquire failed with "Surface is not configured for presentation" and the device's
default handler panicked (#7007). An unsupported surface is a refused configure too. The
acquire runs under `capture_errors`, and the `Validation` status it then returns reconfigures
the surface instead of skipping every later frame.
- `src/lib.rs`, `src/winit.rs`: the painter holds an acquired frame as `AcquiredFrame`, which
destroys a second handle to the frame's texture once the frame is presented or dropped. On a
lost device wgpu's present and discard keep the frame's image in the surface, and dropping
the surface (the window closing for the restart) then panicked in wgpu-hal ("Trying to
destroy a SwapchainAcquireSemaphore that is still in use by a SurfaceTexture", #7148); the
destroy takes the image back whatever the device's state.
- `src/lib.rs`: creation of the egui renderer uses validation, internal and out-of-memory
error scopes, awaiting every result before returning it. A refused UI pipeline (D3D12
`egui_pipeline` shader compilation, #6648) returns `WgpuError::CreateRendererError` through
eframe so startup can retry another backend instead of invoking wgpu's panic handler.