IfcOpenShell/src/ifcviewer/README.md

# IfcViewer

A high-performance native IFC viewer built on IfcOpenShell's C++ geometry
engine with a Qt6 interface and OpenGL 4.5 rendering.

## Architecture

```
+---------------------------------------------------+
|  Qt6 Application (MainWindow)                     |
|  +----------+ +----------------------------------+|
|  | Element  | | 3D Viewport                      ||
|  | Tree     | | (QWindow + OpenGL 4.5 Core)      ||
|  | (per-    | |                                  ||
|  |  model)  | | Per-model: VAO/VBO/EBO           ||
|  +----------+ |            instance SSBO         ||
|  | Property | |            visible SSBO          ||
|  | Table    | |            indirect buffer       ||
|  +----------+ | glMultiDrawElementsIndirect      ||
|  | Status / Progress / Stats                      |
+---------------------------------------------------+
        ^                        ^
        |                        |
  element metadata      MeshChunk / InstanceChunk / Sidecar
        |                        |
+---------------------------------------------------+
|  GeometryStreamer (one per loaded model)          |
|  IfcGeom::Iterator with N threads                 |
|  Dedups representations -> MeshChunk              |
|  Emits one InstanceChunk per placement            |
+---------------------------------------------------+
```

### Key design decisions

- **QWindow viewport** embedded via `QWidget::createWindowContainer()`. Gives
  us a raw native surface for OpenGL, bypassing `QOpenGLWidget`'s compositor
  overhead.
- **GPU instancing as the central pillar.** IFC models are dominated by
  repeated geometry — identical doors, windows, studs, pipes placed at
  different transforms. IfcOpenShell's iterator surfaces representation
  identity, so we upload each unique mesh exactly once and keep per-placement
  data (transform, object id, optional colour override) in a separate SSBO.
  For real projects this collapses tens of millions of triangles of duplicate
  vertex data into a few hundred MB of unique meshes.
- **Per-model GPU buffers**: each loaded model gets its own
  VAO/VBO/EBO/instance-SSBO/visible-SSBO/indirect-buffer. No cross-model
  growth copies. Removing a model frees its GPU memory immediately.
- **Local-coordinate vertex format (28 B):** position (3 floats) + normal
  (3 floats) + packed RGBA8 colour (1 uint). The per-instance transform is
  applied in the vertex shader via an SSBO lookup. No world-baked vertex data.
- **Multi-draw indirect:** every frame the CPU builds a flat list of visible
  instance indices and one `DrawElementsIndirectCommand` per non-empty mesh,
  then issues a single `glMultiDrawElementsIndirect` per model. 50k visible
  instances across 8k unique meshes collapse to one driver-side command
  submission per model.
- **BVH frustum culling over instances**: per-model BVH trees cull whole
  subtrees of placements with one frustum test. Falls back to a linear scan
  during progressive upload and for very small models (< 32 instances).
- **Reflection-aware two-pass draw:** IFC placements can have negative-
  determinant transforms (mirrored families). These flip the screen-space
  winding of their triangles, which would make them vanish under
  `GL_CULL_FACE`. The cull pass buckets visible instances into forward
  (det ≥ 0) and reverse (det < 0) slices and the renderer issues two MDI
  calls per model with `glFrontFace` toggled between them.
- **`reorient-shells` enabled in the iterator:** makes face winding
  consistent within a shell at geometry-gen time — the only place this can
  actually be fixed. Without it, files with inside-out faces produce dark
  patches and swiss-cheese under backface culling. Costs iterator time but
  is cached in the sidecar.
- **Progressive rendering during streaming:** the viewport is drawable
  before `finalizeModel()`. Instances are pushed to the SSBO one at a time
  via `glNamedBufferSubData` as they arrive, and the linear-scan cull path
  handles them until the BVH is built. Orbit and pan remain interactive
  through load.
- **Non-blocking sidecar loading**: sidecars are read on a background
  thread; only the final GPU upload touches the main thread.
- **GPU object picking**: a second render pass writes object IDs into an
  R32UI framebuffer. Click reads back one pixel. No CPU-side raycasting.
- **Multi-model support**: multiple IFCs can be loaded simultaneously.
  Each gets its own `GeometryStreamer` (which owns the `ifcopenshell::file`
  for property lookup). Models load sequentially. Per-model
  hide/show/remove.

### Files

| File | Purpose |
|------|---------|
| `main.cpp` | Application entry, GL 4.5 surface format, CLI argument parsing |
| `MainWindow.h/cpp` | Qt main window: multi-model project, element tree, properties, status |
| `ViewportWindow.h/cpp` | OpenGL 4.5 Core renderer: shaders, buffers, camera, culling, MDI draw, picking |
| `GeometryStreamer.h/cpp` | Background iterator runner; emits `MeshChunk` + `InstanceChunk` |
| `InstancedGeometry.h` | Shared structs: `MeshInfo`, `InstanceCpu`, `InstanceGpu`, chunk records |
| `BvhAccel.h/cpp` | Median-split BVH builder; operates on instance world-AABBs |
| `SidecarCache.h/cpp` | Raw binary `.ifcview` (v4) sidecar read/write |
| `AppSettings.h/cpp` | Persisted preferences (geometry library, stats overlay, backface culling) |
| `SettingsWindow.h/cpp` | Settings dialog |
| `CMakeLists.txt` | Build configuration |

## Dependencies

- **Qt6** (Core, Gui, Widgets, OpenGL)
- **OpenGL 4.5** with `GL_ARB_direct_state_access` and
  `GL_ARB_shader_draw_parameters` — available on Windows and Linux. macOS
  will need a Vulkan/MoltenVK backend (not yet implemented; macOS caps out
  at GL 4.1).
- **IfcOpenShell C++ libraries** (IfcParse, IfcGeom, and their
  dependencies: Open CASCADE, Boost, Eigen3, optionally CGAL).

## Building

IfcViewer is part of the IfcOpenShell CMake project. From the repo root:

```sh
mkdir build && cd build

cmake ../cmake \
    -DCMAKE_BUILD_TYPE=Release \
    -DBUILD_IFCVIEWER=ON \
    -DBUILD_CONVERT=OFF \
    -DBUILD_IFCPYTHON=OFF \
    -DBUILD_GEOMSERVER=OFF \
    -DBUILD_DOCUMENTATION=OFF \
    -DBUILD_EXAMPLES=OFF \
    -DCOLLADA_SUPPORT=OFF \
    -DGLTF_SUPPORT=OFF \
    -DHDF5_SUPPORT=OFF

make -j$(nproc) IfcViewer
```

If Qt6 is not in a standard location, pass `-DQT_DIR=/path/to/qt6`.

## Usage

```sh
./IfcViewer arch.ifc struct.ifc mep.ifc
./IfcViewer                   # then File -> Add Files
```

### Controls

| Input | Action |
|-------|--------|
| Middle mouse drag | Orbit camera |
| Shift + middle mouse drag | Pan camera |
| Scroll wheel | Zoom |
| Left click | Select object |

### Keyboard

| Key | Action |
|-----|--------|
| Ctrl+O | Add files |
| Ctrl+Q | Quit |

### Settings

- **Geometry Library** — kernel string passed to IfcOpenShell (default
  `hybrid-cgal-simple-opencascade`).
- **Show Performance Stats** — overlay FPS / object / triangle / draw
  counts in the status bar.
- **Backface Culling** — `GL_CULL_FACE` on closed solids. Default on.
  Disable if a model uses open shells and you see missing faces.

## Performance Strategy

The viewer targets smooth orbiting at 60 fps on real-world multi-discipline
BIM projects (a "real job" being ~50 models, several million placements,
hundreds of millions of rasterised triangles when everything is in view).

Rendering performance has evolved in phases. Each builds on the previous,
and smaller models never pay for optimisations they don't need.

### Phase 1 — Per-object Frustum Culling

**Status:** implemented (and still the fallback for small models / during
streaming).

Six view-frustum planes are extracted from the view-projection matrix each
frame. Each instance's world AABB is tested with the p-vertex / n-vertex
method (one dot product + one compare per plane, 6 planes).

Surviving instance indices are written into a per-mesh bucket, then
flattened into a single `uint[]` (the "visible SSBO", binding = 1) and
accompanied by one `DrawElementsIndirectCommand` per non-empty mesh.
One `glMultiDrawElementsIndirect` call per model draws everything.

Cost: ~6 dot products per instance per frame. Fine up to ~100 k instances
per frame; above that the linear scan shows up in profiles, motivating
Phase 2.

### Phase 2 — BVH Acceleration + Sidecar Cache

**Status:** implemented.

For models exceeding ~32 instances, a bounding volume hierarchy groups
nearby placements into a binary tree and culls entire subtrees with a
single frustum test. This reduces per-frame work from O(N) to O(log N) in
the best case (camera zoomed to a corner) and remains well under 1 ms for
100 k instances in the worst case (everything on screen).

A BVH was chosen over an octree because BIM data is spatially non-uniform
— dense MEP risers in one zone, sparse open atria in another. An octree
subdivides space uniformly, wasting nodes on empty regions and creating
deep chains in dense ones. A BVH adapts its splits to the actual
placement distribution.

#### Activation

The BVH is optional and non-disruptive. Until it is built, the Phase 1
linear scan handles culling. The renderer checks for a BVH per model and
falls back to the scan for any model that doesn't have one.

It activates in one of two ways:

1. **Sidecar hit** — the `.ifcview` file next to the `.ifc` is found and
   valid; its instance data is uploaded and the BVH rebuilt on the fly
   from the restored AABBs (cheap — `< 100 ms` for 100 k placements).
2. **After streaming** — `finalizeModel()` builds the BVH synchronously
   once all chunks are in (instances already live on the GPU, so there's
   no EBO re-sort to do). The sidecar is written afterwards.

Models under 32 instances skip the BVH.

#### BVH node layout (32 B, two per cache line)

```cpp
struct BvhNode {
    float    aabb_min[3];      // 12 B
    float    aabb_max[3];      // 12 B
    uint32_t right_or_first;   // interior: right child index; leaf: first item index
    uint16_t count;            // 0 = interior, >0 = leaf
    uint16_t axis;             // 0/1/2 for interior; unused for leaf
};
```

Left child is always the next node (pre-order DFS). Leaf items are
indices into the per-model `instances` array; the parallel `bvh_items[]`
array carries the world AABBs.

#### Build: object-median split

1. Compute centroid of each item's AABB.
2. Pick the longest axis of the node's AABB.
3. `std::nth_element` partitions at the median on that axis — O(n).
4. Recurse until a leaf holds ≤ 8 items.

O(n log n) total. No SAH — for frustum culling (6-plane tests, early
subtree reject) the quality difference vs median is negligible.

#### Traversal: stack-based, no recursion

```
stack[64] = { 0 }                     // root
while stack not empty:
    node = nodes[stack.pop()]
    if node.aabb outside frustum: continue
    if leaf:
        for each item in node:
            if item.aabb in frustum: emit to visible list
    else:
        push right child, push left child   // left processed first (DFS)
```

Depth 64 is enough for billions of items on any balanced tree. The stack
is on the C++ stack, zero per-frame allocation.

#### Sidecar format (`.ifcview`, v4)

Raw memory dump, Blender-`.blend`-style — no serialisation, no parsing.
Stores everything needed to skip the `IfcGeom::Iterator` pass:

```
SidecarHeader            (magic "IFVW", version, endian, ...)
uint64_t                 source_file_size
uint32_t + float[]       vertex data    (7 floats × N_verts, local coords)
uint32_t + uint32_t[]    index data     (mesh-local)
uint32_t + MeshInfo[]    per-unique-mesh metadata (48 B each)
uint32_t + InstanceCpu[] per-placement records (transform + AABB + ids)
uint32_t + PackedElementInfo[]   element tree records
uint32_t + char[]        string table
```

Staleness check: `source_file_size` vs actual file size. Mismatched →
reject and rebuild. Endianness marker rejects cross-arch caches.

### GPU Instancing pipeline (the central pillar)

Everything above plugs into a single data-flow, worth documenting on its
own because it's what makes the whole thing fast.

Per-model state on the GPU:

| Buffer | Contents | Lifetime |
|--------|----------|----------|
| `VBO` | Interleaved local-coord vertex data (28 B/vert). One range per unique representation. | Grow-on-demand during streaming; static after finalize. |
| `EBO` | Mesh-local uint32 indices. One range per unique representation. | Same. |
| `SSBO` (binding 0) | `InstanceGpu[]` (80 B each: mat4 transform, object_id, color_override, pad). | Appended during streaming, static after finalize. |
| `visible SSBO` (binding 1) | `uint32[]` — flat list of visible instance indices, ordered by mesh, uploaded each frame. | Rewritten every frame. |
| Draw-indirect buffer | `DrawElementsIndirectCommand[]` — one per non-empty mesh, uploaded each frame. | Rewritten every frame. |

Draw command:

```c
struct DrawElementsIndirectCommand {
    uint32_t count;         // mesh.index_count
    uint32_t instanceCount; // visible-list length for this mesh
    uint32_t firstIndex;    // mesh.ebo_byte_offset / 4
    uint32_t baseVertex;    // mesh.vbo_byte_offset / 28
    uint32_t baseInstance;  // offset into the flat visible-index array
};
```

The vertex shader reads `visible[gl_BaseInstanceARB + gl_InstanceID]` to
get the real instance id, then indexes into the instance SSBO:

```glsl
uint slot = uint(gl_BaseInstanceARB) + uint(gl_InstanceID);
uint iid  = visible[slot];
InstanceRecord inst = instances[iid];
gl_Position = u_view_projection * inst.transform * vec4(a_position, 1.0);
```

`gl_BaseInstanceARB` requires `GL_ARB_shader_draw_parameters`, which is
available on all GL-4.6-capable drivers.

Reflection handling: at upload time we store a parallel
`instance_reflected[]` byte array (1 if the transform's upper-3×3 has
det < 0). The cull pass produces two flat visible-list slices — fwd
(non-reflected) first, rev (reflected) after — concatenated into one
buffer. The renderer issues MDI twice: fwd with `glFrontFace(GL_CCW)`,
rev with `glFrontFace(GL_CW)`. `GL_CULL_FACE` stays on and does the
right thing in both passes.

### Current bottleneck — draw-bound, not upload-bound

The original README's Phase 3 ("GPU-driven indirect draw") described
moving draw submission to the GPU via compute. In the meantime, GPU
instancing and MDI made the CPU-side draw cost essentially free (10
`glMultiDrawElementsIndirect` calls per frame for 10 models). **That
goal is met.** The real ceiling lies elsewhere, and it took a couple of
bad hypotheses to pin down.

#### Profiled scene

10 models / 379 k instances / 128 M triangles, everything in view, no
camera motion, GTX 1650 (PCIe dGPU, 4 GB VRAM):

| Metric | Value |
|--------|-------|
| FPS | 6.7 |
| Frame time | 149 ms |
| gl_draws | 10 |
| Sub-draws packed in indirect buffers | 67 037 |

`nvidia-smi` reports 95 % GPU utilisation during render — the GPU is
the thing that's pinned.

#### False lead: "the per-frame uploads are the bottleneck"

The first round of probes pointed at the two `glNamedBufferSubData`
calls per model per frame (visible list ~1.5 MB + indirect buffer
~1.3 MB):

| Probe | Result | Initial interpretation |
|-------|--------|------------------------|
| Camera off-screen (nothing visible) | 60 fps | GPU idle → CPU path cheap |
| Comment out the two `glNamedBufferSubData` | 60 fps, blank screen | Uploads are the bottleneck |

This led to an aborted Phase 3A implementation of persistent-mapped
triple-buffered rings (and then staging + VRAM-resident with
`glCopyNamedBufferSubData`). Neither moved the FPS needle — both still
sat at 6.7 fps.

The probe was wrong: **commenting out the uploads emptied the indirect
buffer, so MDI drew zero triangles. "No upload" and "no draw" were
indistinguishable in the test.**

#### What actually isolates the draw cost

Two diagnostic env vars now live in `render()`:

- `IFC_SKIP_MDI=1` — keep everything (cull, upload, binds) but skip the
  actual `glMultiDrawElementsIndirect` calls.
- `IFC_MAX_SUBDRAWS=N` — truncate each MDI's drawcount to N while still
  running the rest of the frame.

Results on the profiled scene:

| Probe | FPS | Frame time |
|-------|-----|-----------|
| baseline | 6.7 | 149 ms |
| `IFC_SKIP_MDI=1` | 62.5 | 16 ms |
| `IFC_MAX_SUBDRAWS=30000` | 6.7 | 149 ms |
| `IFC_MAX_SUBDRAWS=10000` | 7.5 | 133 ms |
| `IFC_MAX_SUBDRAWS=1000` | 20.2 | 49 ms |

Readings:

1. `SKIP_MDI` gives 62 fps with all upload/bind machinery still running
   — the non-draw path fits in ~16 ms easily. **Not upload-bound.**
2. Halving the sub-draw count (67 k → 30 k) saves 0 ms. If per-sub-draw
   command-processor overhead were material, dropping 37 k sub-draws
   would save measurable time no matter which sub-draws were dropped.
   It doesn't. **67 k sub-draws is not the bottleneck** — the long tail
   carries almost no triangles, and the heavyweights dominate.
3. Time only starts coming down once the cap is low enough to shed bulk
   triangle work (1000 sub-draws → 49 ms). The curve is consistent with
   a long-tailed distribution: a handful of very big meshes × instance
   counts do most of the rasterisation.

**Conclusion: the GTX 1650 is rasterising 128 M triangles at ~850 M
tri/s, and that eats ~133 ms of the 149 ms frame.** No CPU-side or
upload-side work will recover it. The only way forward is to draw
fewer triangles.

### Phase 3 (revised) — Shed triangles, not bytes

In order of effort/payoff for BIM workloads:

#### 3A. Screen-space contribution culling (near-term)

Project each visible-instance AABB to screen space during BVH
traversal. Reject instances whose projected size is below a threshold
(~4 px). In BIM this is the single biggest win: at viewer zoom levels
that encompass a whole building, most MEP fittings, fixings, furniture
legs, door hardware etc. occupy < 1 px and contribute nothing.

Scope: a projection + pixel-area test inside
`ViewportWindow::cullAndUploadVisible`. Zero new GPU state. Expect
10–30× reduction in drawn triangles on plant/MEP-dense scenes; full
buildings viewed in overview should approach 60 fps.

#### 3B. Distance / contribution LOD (medium-term)

Pre-simplify unique representations at ingress time (store LOD 0 / 1 /
2 meshes in the VBO/EBO with offsets), select LOD per instance per
frame by the same projected-size metric as 3A. The visible-SSBO
plumbing and MDI structure don't change — only `firstIndex`/`count` in
the indirect command does. Ingress side needs a decimation pass
(`meshoptimizer` or similar); GPU side is nearly free.

#### 3C. Hierarchical-Z occlusion culling (longer-term)

Render large occluders first, build a depth pyramid, test instance
AABBs against it. In dense BIM most geometry is behind other geometry
from any given interior viewpoint; historically a 3–10× reduction in
drawn instances. Most valuable *after* 3A+3B, which together handle
the far-away and small-detail cases. Pairs naturally with GPU-side
culling (a compute shader doing the HiZ test and writing the visible
list + indirect buffer in place).

#### 3D. GPU-side culling via compute (longer-term)

Push the cull loop to a compute shader reading the per-instance SSBO +
frustum planes + HiZ pyramid, emitting the visible list and indirect
commands with atomic counters. Eliminates all CPU→GPU per-frame bytes
and lets 3C scale to millions of instances. Worth doing once 3A–3C
have stabilised the CPU-side algorithm we'd be porting.

### Planned follow-ups (post-Phase-3)

- **Mesh shaders / meshlets.** Ceiling-raising, but overkill until the
  above are exhausted and we've hit silicon limits on vertex/raster
  throughput.

## Summary table

```
Scene size                      Bottleneck              Fix
-----------                     ----------              ---
< 100k instances                CPU cull scan           Phase 1 only
100k–500k                       CPU cull scan           BVH (Phase 2) — done
500k+ tris / overview shot      GPU vertex + raster     Phase 3A contribution cull
                                                        (+ 3B LOD for close-ups)
multi-million + occluders       redundant rasterisation Phase 3C HiZ occlusion
```

## Roadmap

- [x] Material colour support (per-vertex RGBA8)
- [x] Per-model GPU buffers (VAO/VBO/EBO per model, no cross-model copies)
- [x] Per-object frustum culling (Phase 1)
- [x] BVH acceleration with per-model trees (Phase 2)
- [x] Raw binary `.ifcview` sidecar cache
- [x] Non-blocking sidecar loading (background thread I/O)
- [x] Progressive GPU upload (VBO/EBO growth + streaming-time instance appends)
- [x] GPU instancing (unique meshes + per-placement SSBO)
- [x] `glMultiDrawElementsIndirect` draw path
- [x] Reflection-aware two-pass draw for mirrored placements
- [x] Backface culling (user-toggleable, default on)
- [x] `reorient-shells` enabled in iterator
- [x] Perf diagnostic env vars (`IFC_SKIP_MDI`, `IFC_MAX_SUBDRAWS`)
- [ ] **Phase 3A — screen-space contribution culling** (next)
- [ ] Phase 3B — distance / contribution LOD
- [ ] Phase 3C — Hierarchical-Z occlusion culling
- [ ] Phase 3D — GPU-side compute-shader culling
- [ ] Vulkan/MoltenVK backend for macOS
- [ ] Embedded Python scripting console