Sim;Engine
Sim;Engine is a validated visualization library built for simulation products. It provides a high-performance 2D renderer, scientific visualization paths, and a focused retained 3D stereometry pipeline.
Sim;X is the primary consumer and design driver, but the crate is reusable. Applications provide ready visual state. Sim;Engine owns cameras, drawing, clipping, interpolation, GPU resources, composition, recovery, and rendering diagnostics. Physics, simulation stepping, domain entities, UI navigation, and plugins remain in the host application.
This guide describes Sim;Engine 0.4.1. The crate remains pre-1.0, and its supported release target is Linux with Vulkan. A concrete adapter/driver is supported when the mandatory semantic fixture passes on it; untested drivers are not inferred from Mesa evidence. The minimum supported Rust version is 1.90.
The software-Vulkan CI fixture uses Mesa 25.3+ lavapipe. Mesa 25.2.8 llvmpipe has a reproduced MSAA scissor leak and does not pass the rendering contract; its upstream fix is included in Mesa 25.3. This does not impose an Ubuntu-version requirement on host applications.
The 0.4 API includes Opaque/Mask/Blend 3D materials, sidedness, vertex colors, optional Lambert lighting and distance fog, capacity-reusing mesh updates, mipmapped and repeating textures, atomic region edits, and complete material recovery. CPU-only fonts and reusable pre-shaped text reduce repeated text work; optional GPU timestamps and changing-scene fixtures make costs observable. See the 0.4 integration guide and 0.4.0 changelog for contracts and migration notes from 0.3.0.
Version 0.4.1 focuses on profiling-driven optimization. Texture region edits filter and upload only affected mip rectangles, preserving complete recovery snapshots and old aliases. Its other performance work accelerates Native 3D arithmetic validation, batches compatible repeated surfaces, removes unused edge uploads, and reduces repeated lighting/fog validation. StrictPortable classification reuses bounded per-source calculations and skips polygon clipping when all original triangle proofs already succeed. Frame preparation reuses scratch capacity and uploads only changed camera, instance and edge-uniform ranges, without skipping validation or drawing. Scene resource and visibility totals update at accepted mutations, making statistics queries constant-time without changing ownership or resource budgets. An opt-in Native surface culling path can omit proven offscreen objects after all original validation and budget checks. It is disabled by default; small shared meshes can cost more draw calls when omissions split instanced batches. The 2D/3D renderer, scene styles and regression fixtures are now organized into focused modules, with unchanged public imports and rendering contracts. External tests follow their owning subsystem, use contract-based names, and share raw GPU readback transport without sharing expected-pixel calculations. CI checks the source-layout conventions. Existing 0.4 integrations need no source migration. See the performance guide and 0.4.1 changelog for the complete delta and remaining costs. No universal frame rate or replacement for consumer testing is promised.
Documentation
-
Library documentation - installation, concepts, every rendering path, recovery, performance, architecture, and examples.
-
Changelog - release history and migration notes.
-
Archived 0.3 documentation - the guide for integrations that have not upgraded.
-
Generated API reference:
Capabilities
- validated circles, rectangles, rounded rectangles, lines, and polylines, including sub-ULP radii, rounded corners, and world-width strokes preserved as camera-relative local offsets;
- solid, linear-gradient, and radial-gradient fills;
- logical/world-width strokes with bounded caps, joins, dashes, markers, clipping, and stable layer ordering;
- 2D camera pan, rotation, zoom, fit, and picking;
- fallible tweening and easing;
- streaming, prepared, and dynamic triangle geometry;
- retained sRGB RGBA images, atlas sprite batches, and world-space image quads;
- host-shaped, bounded glyph atlas runs with deterministic logical bounds;
- optional CPU-only
fontsfeature for TTF/OTF loading, shaping and rasterization;textadds GPU atlases, retained runs and preparation from pre-shaped lines; - budgeted instanced particle rendering;
- scalar-field textures, color maps, partial updates, and heatmaps;
- offscreen targets, composition, and bounded trails;
- device recovery with explicit retained-resource restoration;
- typed logical-pixel, 2D/3D world-length, and target-scale boundaries;
- retained 3D meshes, independent transforms, hardware depth, and visible or dashed hidden mathematical edges;
- explicit StrictPortable/Native surface policies and source-triangle diagnostics;
- capacity-reusing, alias-isolated mesh edits and material-only rebinding;
- Opaque/Mask/Blend surfaces, sidedness, vertex colors, optional normals, ambient/directional Lambert lighting and distance fog;
- complete mip chains, isolated repeating tiles and bounded texture-region edits;
- atomic retained-3D scene recovery that preserves stable object IDs and visual state across logical-device replacement;
- optional bounded GPU pass timestamps with correlation IDs and loss counters;
- primitive/source-grouped diagnostics and repeatable release-mode workloads for static UI, prepared/streaming UI, viewports, atlases, glyphs, mixed layers, budget rejection, DPI reconfiguration, real compositor scale transitions, and recovery.
Ordinary scenes support explicit construction, tessellation, and upload
budgets, fixed logical-screen geometry, and independent positioned viewport or
offscreen-target rendering. FrameComposer orders streaming and prepared
scenes, dynamic meshes, particles, scalar fields, images, glyph runs, and 2D/3D
color targets under one frame budget and one surface presentation.
DynamicMeshBudget limits caller-provided filled triangles and makes full
updates allocation-fallible and atomic.
GPU transform validation enforces bounded arithmetic for 2D scenes, dynamic
triangles, particles, and retained 3D. Nonzero subnormal
coordinate/transform operands, coordinate magnitudes above 2^120, ambiguous
geometry-branch thresholds, and arithmetic ranges that cannot be proven
portable are rejected with a structured transform error before submission.
Normalized linear color channels are not geometric operands and may include
subnormal values; their contribution is below the precision of current
normalized render targets. This intentionally prefers a clear rejection over
backend-dependent geometry disappearance or topology. Retained 3D defaults to
StrictPortable, which also proves clipping topology and rejects ambiguous
surface classifications. Opt-in Native uses hardware surface clipping while
retaining arithmetic checks; mathematical-edge validation stays strict.
Blend sorts objects, not triangles within a mesh; intersecting transparent surfaces can still overlap incorrectly. General section materials, hatching, projected 3D anchors, 3D picking, point lights, shadows and PBR remain outside 0.4.1. Unlit remains available and is the default.
An additional Intel UHD (CML GT2) / Mesa 26.1.6 Vulkan qualification run did not complete the material readback fixture; the same wait reproduced on the unchanged dev.4 baseline. This configuration is not qualified by the successful NVIDIA or software-Vulkan runs. See known boundaries.
Installation
The default feature set includes the wgpu renderer:
or add it directly to Cargo.toml:
[]
= "0.4.1"
Use core visual-state APIs without GPU dependencies:
[]
= { = "0.4.1", = false }
Run the examples from a repository checkout matching the library version:
The text lab has three pages: 1 compares DejaVu Sans/Serif/Mono, Inter and
DejaVu Math; 2 (default) shows 𝓣𝔂𝓹𝓮 𝓼𝓸𝓶𝓮𝓽𝓱𝓲𝓷𝓰 𝓽𝓸 𝓼𝓽𝓪𝓻𝓽, math
symbols, Japanese, combining accents and one Arabic RTL run; 3 exercises
retained updates, clipping and motion at 16/24/48 logical pixels. Use the wheel
to scroll shorter gallery windows, and --page fonts|unicode|motion to select
the initial page. Licensed font fixtures are embedded, including an 89 KiB
Japanese example subset, not a full CJK font. --font replaces the selected
font comparison and motion samples with your TTF/OTF file. Font support
is opt-in; the default renderer and core-only builds retain their dependency
boundary. UI navigation, audio, and domain logic stay in the host. --acceptance
exercises public update, composition and recovery routes and requires 40 actual
presented frames on each page. Automatic font fallback and mixed-script paragraph layout
are not implied by these independently shaped, explicitly font-selected rows.
For changing-text diagnostics, use frame_cache_benchmark --case recolored --labels 1000 --cache on; compare with --upload-staging 0 to isolate bounded
uniform-upload packing without disabling the other caches. Transfer memory and
GPU copy work are reported separately from host-upload bytes.
Loading a font
Enable the optional text feature:
= { = "0.4.1", = ["text"] }
#
#
#
#
Then create a TextAtlas2d, prepare a TextRun2d, and draw it through the
frame composer. The font guide
covers the complete render/update/DPI/recovery workflow. Supply trusted,
appropriately licensed font assets. Automatic font fallback, paragraph bidi,
wrapping, color emoji, and font-file I/O are not built into the engine.
For CPU-only font work, use default-features = false, features = ["fonts"].
TextShapingSession reuses shaping state; TextAtlas2d::prepare_from_shaped
and update_from_shaped consume its results without shaping again.
Quick Start
use ;
Window creation stays in the host. Renderer initialization is asynchronous; steady-state updates and rendering are synchronous submissions. Standalone retained-resource mutations submit their transfer immediately (without waiting for GPU completion), while a skipped surface frame enqueues no uploads:
let options = new?;
let mut renderer = new_with_options.await?;
let notify_window = window.clone;
renderer.set_pre_present_notify;
let mut frame = renderer.begin_frame?;
frame.draw_scene?;
let report = frame.present?;
Choose A Rendering Path
| Workload | Preferred API |
|---|---|
| Small changing scene | Scene |
| Static geometry with a moving camera | PreparedScene |
| Frequently changing triangles | DynamicMesh2d |
| Images and atlas sprites | Image2d plus ImageBatch2d |
TTF/OTF labels and counters (text feature) |
FontFace, TextAtlas2d and TextRun2d |
| Host-shaped scientific text | GlyphAtlas2d plus GlyphRun2d |
| Large point or circle population | ParticleField2d |
| Dense scalar grid | ScalarFieldTexture |
| Mixed UI, viewports, fields, and targets | FrameComposer |
| Bounded gas plus particle overlay | render_layered_visualization |
| Stereometry solids and hidden edges | Mesh3d plus Scene3d |
| Changing 3D geometry with immutable aliases | update_scene3d_mesh with DynamicMesh3dBudget |
| Material or background changes without geometry upload | Scene3d::set_texture_material / set_background |
| Partial 3D texture edits | update_scene3d_texture_region |
Examples
From a Sim;Engine repository checkout, run the complete named
performance/contract matrix on a Vulkan-capable Linux machine with
./scripts/rendering_benchmark_matrix.sh. Release scripts are repository
tooling and are intentionally not shipped in the crates.io source archive.
Gated surface workloads
require a 1280x720 physical surface at scale 1.0; both values are printed
in every result. Absolute timings are comparable only when surface extent,
scale, adapter, driver, backend, present mode, and workload are recorded
together; deterministic command/vertex/upload counters are portable.
The matrix first probes the real production surface. It pins the selected
high-performance Vulkan adapter by PCI bus address as well as backend, name,
vendor, and model, then requires the offscreen semantic oracle and every
surface/HiDPI process to use that same physical GPU. The oracle uses the
production surface format and its selected MSAA count. A nested compositor may
advertise a different surface format; its HiDPI evidence remains pinned by the
physical PCI identity rather than pretending two surfaces have identical
capabilities. Every warmup and
measured report must be Drawn; a timeout, occlusion, or outdated surface
fails the fixture instead of inflating throughput with skipped attempts.
Gated workloads use three independent 120-frame trials. Every trial must meet
the wall-throughput floor; the reported median is diagnostic only.
Renderer-work percentiles cover all 360 drawn frames, so one failed trial
cannot be hidden by two faster trials.
The matrix always enforces fixture-specific p95 ceilings for renderer work
excluding surface acquisition. Immediate additionally requires 60 presented
FPS. Mailbox/FIFO wall throughput must reach 95% of the confirmed current
monitor refresh, bounded to a 30-60 Hz release floor. The window is mapped by
an unmeasured drawn frame before any output metadata is used. Immediate does
not require refresh metadata; Mailbox/FIFO require a positive refresh rate
from the window's confirmed current monitor. Zero/unknown refresh and
primary/first-monitor fallbacks are never accepted for synchronized evidence.
Warmup and measured frames advance one RedrawRequested at a time, allowing
compositor events between every sample. A resize, scale transition, or changed
current-output identity invalidates the confirmation and discards partial
timing samples; a surface outside the fixed release extent/scale fails rather
than being measured as a cheaper workload. The benchmark registers
Window::pre_present_notify with the renderer. The callback runs after queue
submission and immediately before present only for the concrete FIFO, FIFO-
relaxed, or Mailbox modes; Immediate remains unpaced. The final measured
present then yields through a compositor-aware redraw boundary and repeats the
generation/output check before evidence is published.
Acquire percentiles remain reported but do not independently flip the verdict
from one scheduler-sensitive sample.
The mandatory workloads include the fused bounded particle/scalar path and a
retained-3D target-composition path with 4,096 budgeted dynamic triangles, so
per-particle portability/culling/upload, dynamic-triangle proofs, and per-object
surface/hidden-edge validation cannot regress outside the release performance
evidence.
The matrix also drives a nested KWin compositor through a real scale
1.00-to-1.25 transition and records the paired event and successful redraw for
the exact revision on that same physical adapter.
The automatic transition requires dbus-run-session, kwin_wayland, and
kscreen-doctor; a missing executable fails the matrix instead of silently
skipping HiDPI evidence.
The wrapper, standalone matrix, and standalone HiDPI gate reject a dirty
worktree, then build and execute from a read-only detached worktree of the
captured commit with a separate Cargo target directory. They recheck the
calling checkout before accepting results. Replacement refs and legacy grafts
are rejected, and all Git operations run with replacement objects disabled.
Evidence therefore remains bound to Git object data for the exact recorded SHA
even if the mutable calling worktree changes while a long gate is running.
Child gates write only to hidden staging. The launcher atomically publishes a
completed bundle after the entire requested gate succeeds; failed, killed, or
crashed runs cannot leave a bundle that claims success. A nonblocking process
lock rejects concurrent release gates before either invocation can invalidate
or publish evidence.
stroke_gallery is the visual oracle for the stroke contract. Pages 1-4
show every cap/join, half-alpha overlap probes, bounded animated dashes,
outward endpoint arrow markers, miter fallback, camera rotation, and the
accepted 0.005-world-unit line at zoom 10,000. Press Space, arrows, +/-, or R
to pause, scrub, zoom, and reset. Page 5 adds the exact-tip scientific arrows
introduced in 0.3.
ui_demo includes a retained host-rasterized scientific glyph probe above all
four independently changing scene workloads. It demonstrates one-time atlas
and run creation without requiring the optional font loader or shaping path.
Verification
From a clean Sim;Engine repository checkout, the Linux release gate checks the declared Rust 1.90 MSRV, all targets with and without the renderer and CPU fonts, strict clippy, rustdoc, Vulkan semantic GPU readback with a backend assertion, the Vulkan performance matrix on a real surface, real text/standalone-mesh recovery, a transactional nested-compositor HiDPI transition, and the publishable package boundary:
After all 11 steps pass, the launcher atomically publishes
target/linux-release-evidence/. Its completion.txt binds the successful
gate and exact SHA; the same directory contains linux-vulkan-surface.txt,
linux-vulkan-adapter.txt, linux-vulkan-performance.txt, and
linux-hidpi-transition.txt. The performance manifest preserves the complete
fixture counters, timings, thresholds, and passed verdicts for all nine
surface runs instead of leaving them only in terminal output.
Maintainers should follow the complete official release procedure. Publishing is complete only when crates.io exposes the immutable package, the annotated Git tag points at the evidenced commit, and the matching GitHub Release is public.
License
Licensed under either MIT or Apache-2.0, at your option.