render() was re-running the full cull every 16 ms timer tick even when
nothing had changed — the camera matrices, scene state, and therefore
visible set were all identical to the previous frame's. The GPU was
still happy to redraw from the cached indirect buffer, but the CPU was
burning 21 ms/frame rebuilding the same visible list.
Detect the no-op case by comparing view/proj against last_cull_view_ /
last_cull_proj_ and checking a scene-dirty flag (have_cached_cull_)
that every mutator on models_gpu_ invalidates — finalizeModel,
applyCachedModel, applyLodExtension, hide/show/remove/reset, and
uploadInstanceChunk. When the check passes we skip both
cullAndUploadVisible and buildHizPyramid (the depth buffer is
bit-identical, so re-reading it produces the same pyramid).
Per-model visible_objects / visible_triangles stats now live on
ModelGpuData so the stats line reports correct numbers on skipped
frames instead of reading from a stale indirect_scratch_.
Measured on a 569k-object overview: still frames go 22 fps → 62 fps;
orbiting goes 23 fps → ~30-50 fps depending on how hard you move the
mouse (the cull only pays its full cost on the ~25 % of frames where
the camera actually moved). The stats line gains a "skipped N/M"
field so you can see the ratio live.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
cullAndUploadVisible was reading each instance's AABB through
m.instances[idx] — a 104-byte InstanceCpu struct — for the frustum /
contribution / HiZ tests. Only 24 of those bytes (the two float[3]
AABBs) are actually used by the tests; the rest (4×4 transform +
header) is pure cache-line waste, and with 569k instances the array
is 59 MB, well past any cache.
bvh_items[idx] already stores a 1:1 compact 28-byte record with the
same AABB, built unconditionally in buildBvhForModel(). Switch the
hot test path to read from it, and only touch InstanceCpu once an
instance has passed all three tests (for mesh_id). Modest ~20 %
drop in cull-traverse time on a 569k-object overview (26 ms → 21 ms).
Also add four cull-phase timers (clr / trv / emt / upl) to the
per-second stats line so future optimisation work has concrete
numbers to chase. Confirmed via these timers that bucket clears,
emit and GPU upload are all <1 ms combined; traversal is where the
remaining CPU cost lives.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
After the main draw, blit the MSAA default-framebuffer depth to a
single-sample 256×128 depth texture, read it back, and build a CPU
max-reduced mip pyramid. Next frame's cullAndUploadVisible projects
each BVH node / instance AABB through the previous frame's VP and
compares the AABB's nearest depth against the pyramid's deepest value
at the matching mip level; strictly-beyond AABBs are rejected.
Conservative direction (aabb_near > hiz_max) — never wrongly rejects a
visible instance, so no flicker. BVH subtree-level test lets a single
8-corner projection reject up to a leaf's worth of instances.
Tuning knobs: IFC_NO_HIZ=1 disables; IFC_HIZ_SIZE overrides base width.
New stats counter hiz_rej shows rejects/frame.
Measured: big win on interior views (GPU-bound), roughly zero net
effect on exterior overviews (CPU-bound on cull traversal, so the
saved GPU work is masked). Tried a 3-deep PBO ring for async readback
and reverted — the extra frame of staleness produced visible flicker
on fast orbit, and the synchronous readback wasn't actually a measured
bottleneck at 256×128.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Decimate each unique mesh once at sidecar-build time and swap to the
reduced index slice per-instance per-frame when projected sphere radius
drops below IFC_LOD1_PX (default 30). Same VBO, same SSBO, just a
different firstIndex/count in the indirect command.
Extends MeshInfo (48→56 B) with lod1_ebo_byte_offset + lod1_index_count
and bumps the sidecar to v5. buildLods() runs inside
onStreamingFinished, appends decimated indices to sd.indices,
applyLodExtension pushes the EBO suffix to the live GPU state, and the
sidecar is written with LOD1 baked in.
simplifySloppy (voxel clustering) is used instead of the default
edge-collapse meshopt_simplify because BIM brep output is per-triangle-
unwelded and non-manifold after welding — simplify returned the input
unchanged for every mesh tested. Sloppy ignores topology. Knobs
(IFC_LOD_SLOPPY, IFC_LOD_ERROR, IFC_LOD_RATIO, IFC_LOD_MIN_SAVINGS,
IFC_LOD_LOCK_BORDER, IFC_LOD_DEBUG) are available for A/B tuning.
Result on the 128M-tri 10-model test scene (GTX 1650, 2px contribution
cull): 20.2 → 43.2 fps, 40M → 14M visible triangles, no change in
object count. LOD build adds 100–600 ms per model on first open,
cached thereafter.
README Phase 3B section is now a full writeup of pipeline, selection,
decimator-choice rationale, env vars, and measured numbers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Reject frustum-visible objects whose bounding sphere projects below a
pixel-radius threshold. Applied at both BVH-node level (whole subtrees
pruned) and per-instance level; short-circuits when the camera is
inside the AABB so nothing-you're-standing-next-to is ever lost.
Pick pass passes threshold 0 so sub-pixel objects stay clickable.
Threshold defaults to 2 px (radius), overridable via IFC_MIN_PX env
var. Measured on the 128 M-tri test scene (GTX 1650):
0 px (off): 6.7 fps, 128 M tris
2 px: 20.2 fps, 40 M tris (31%)
4 px: 30.3 fps, 15 M tris (12%)
The metric is sphere-based (cheap: one sqrt per test) rather than
AABB-corner projection; loses a little precision on very elongated
bounds but costs ~5x less per test and the BVH-node pre-cull means
the long-tail-of-small-things case is already handled by subtree
pruning before we touch individual instances.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Earlier probes pointed at per-frame glNamedBufferSubData uploads as the
bottleneck (60 fps when those two calls were commented out). That was a
false reading — zeroing the uploads also emptied the indirect buffer, so
MDI drew nothing. "No upload" and "no draw" were indistinguishable.
Two new diagnostic env vars in render() isolate the real costs:
IFC_SKIP_MDI=1 keep cull + upload + binds, skip only the MDI
draws. Gives 62 fps with everything else running,
confirming the non-draw path fits in ~16 ms.
IFC_MAX_SUBDRAWS=N cap each MDI's drawcount. 67k -> 30k sub-draws
saves 0 ms, confirming sub-draw count itself is
not the bottleneck; the long tail of sub-draws
carries ~no triangles.
On a GTX 1650 with 128 M triangles in view, nvidia-smi sits at 95 %
GPU util and FPS scales with triangle work, not sub-draw count. The
card is simply rasterising at ~850 M tri/s. No CPU-side or upload
trick recovers it.
Revised Phase 3 is therefore shedding triangles, not bytes:
3A screen-space contribution culling (next)
3B LOD
3C HiZ occlusion
3D GPU-side compute culling
README Phase 3 section rewritten around the diagnosis, including the
false lead, so future work doesn't re-tread the upload path. The
aborted staging+resident ring-buffer implementation was reverted (the
uncommitted working tree is gone — pure glNamedBufferSubData retained
for the visible + indirect buffers, which we now know is fine).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Enables GL_CULL_FACE by default (user-toggleable in Settings) so
closed solids skip shading their back halves. The catch is that
IFC placements can contain reflections (mat4 with det<0 — mirrored
families, symmetric instances). Naively culling would make every
mirrored instance vanish because the rasterizer sees its screen-space
winding as backwards.
Fix: detect reflections at upload time via determinant sign, bucket
visible instances into forward (det>=0) and reverse (det<0) per mesh
during culling, and issue two glMultiDrawElementsIndirect calls per
model with glFrontFace toggled CCW/CW between them. The indirect
buffer is still one buffer — just split into a forward slice followed
by a reverse slice, with m.indirect_forward_count recording the split.
Vertex shader flips the normal when the transform has negative
determinant, keeping lighting correct on mirrored instances. The
fragment shader keeps the gl_FrontFacing fallback as a safety net
when culling is disabled (e.g. for files with open shells).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two bugs conflated as "weird colors":
1. Two-sided lighting. IFC placements often embed reflection
matrices (mirrored families). Transforming a_normal by
mat3(inst.transform) produces a normal pointing the wrong way
on those instances, and max(n·L, 0) then clamps the surface to
pure ambient — reads as dark / washed out. Use gl_FrontFacing
to flip n in the fragment shader so both winding orientations
shade correctly. The proper fix (ship an inverse-transpose
normal matrix or a det-sign bit per instance) is still owed;
that would unlock re-enabling GL_CULL_FACE for a big fragment-
work win on closed solids.
2. Stats label "inst_draws" was counting indirect sub-draws, not
actual GL draw calls — misleading since MDI collapses N sub-
draws into one glMultiDrawElementsIndirect. Split into
gl_draw_calls (real GL calls, = drawn-model count) and
indirect_sub_draws (packed sub-commands). For a BIM model
with 47k unique meshes at full view this now correctly reads
"1 gl_draws (47092 sub)" rather than suggesting 47k driver
dispatches.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Each visible model now issues a single glMultiDrawElementsIndirect
call instead of one glDrawElementsInstancedBaseVertex per mesh. The
CPU BVH cull populates an array of DrawElementsIndirectCommand
records plus the flat visible-instance list, uploads both, and draws
the whole model in one GL call.
Vertex shaders switch from a uniform u_instance_offset to
gl_BaseInstanceARB (ARB_shader_draw_parameters), so per-draw offset
comes from the indirect command's baseInstance field.
Draw-call counts for BIM scenes with hundreds of unique meshes drop
from hundreds-per-frame to one-per-model, cutting driver overhead.
This also sets up the plumbing for the follow-up compute-shader cull
that will populate the indirect buffer entirely on-GPU.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pre-allocate the instance SSBO on model creation (4 MB, grow-on-demand)
and append each arriving InstanceChunk directly to the GPU-side
InstanceGpu array in uploadInstanceChunk. This makes a model drawable
as soon as its first mesh + first instance chunk land, rather than
waiting for finalizeModel.
The visible-list architecture already decouples SSBO order from the
draw path, so appending in insertion order is correct — no sorting
required. finalizeModel collapses to:
- compute per-mesh instance counts (for stats + sidecar round-trip)
- build the per-model BVH over instance world AABBs
Render / pick loops now gate on ssbo_instance_count > 0 rather than
the finalized flag. Stats include in-progress models in totals
(excluding only hidden).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Re-wires the BVH acceleration structure on top of the new instanced
renderer. Per model, build a BVH over per-instance world AABBs at
finalize (and on sidecar apply). Each frame, traverse the BVH against
the camera frustum to produce a visible-instance index list, bucket by
mesh_id, and upload to a per-model SSBO at binding=1. The main and
pick vertex shaders do a double-indirection
`instances[visible[u_offset + gl_InstanceID]]` so draws only touch
instances that passed the frustum test.
Models with fewer than BVH_MIN_OBJECTS instances skip the BVH build
and fall back to a linear per-instance frustum test.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Commit B of the instancing migration. The sidecar on-disk format is
reintroduced at version 4 with MeshInfo + InstanceCpu sections in place
of v3's flat per-object draw-info array.
After streaming finishes, MainWindow asks the viewport for a post-
finalise snapshot (VBO + EBO are read back from the GPU, meshes and
instances come from the CPU-side arrays) and writes it alongside
PackedElementInfo + the string table. On a subsequent load,
readSidecar rehydrates the whole struct and ViewportWindow::
applyCachedModel uploads VBO/EBO/SSBO in a single step, bypassing the
iterator entirely.
Staleness check is still by source file size.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Commit A of the instancing migration (Phase 3a). The streamer now runs
the iterator with use-world-coords=false and dedupes by the geometry's
representation id, emitting a MeshChunk once per unique geometry and an
InstanceChunk per placement. The viewport keeps geometry in local
coordinates (28 B/vertex, down from 32) and applies the per-instance
transform in the vertex shader via an std430 SSBO indexed by
gl_InstanceID + a per-draw uniform offset. After streaming finishes
finalizeModel() stable-sorts instances by mesh_id, assigns each mesh a
contiguous range, and uploads the SSBO; render then issues one
glDrawElementsInstancedBaseVertex per mesh.
BvhAccel is reshaped to operate on a generic BvhItem (world AABB +
model_id) so it can drive instance-level culling, but the path is not
wired in yet -- every instance is drawn every frame in this commit.
Progressive-during-streaming rendering is likewise disabled: a model
appears when its SSBO is uploaded, not incrementally. Sidecar cache
is stubbed (reads miss, writes are no-ops); the v4 on-disk format with
MeshInfo + InstanceGpu sections lands in Commit B.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a BVH leaf passes the frustum test, emit a single glMultiDrawElements
record covering the leaf's entire index range instead of one per object.
Leaves are contiguous in the EBO after reorderEbo, so the range is just
[first_object.index_offset, sum(index_count)]. Cuts draw calls by ~8x
(BVH_MAX_LEAF_SIZE) and shifts the bottleneck from CPU/driver per-draw
overhead toward GPU vertex throughput.
Per-object features (selection highlight, per-vertex color, object_id
picking) are unchanged — they operate on vertex attributes, not draw
state. Future per-object hide/override will use SSBO lookups sampled
by object_id in the fragment shader.
Slight overdraw from skipping per-object frustum tests within a leaf is
negligible given median-split BVH tightness and spare tri throughput.
Also adds visible_objects_ counter so stats still report true object
counts (not leaf counts), plus leaf_draws/model_draws breakdown in the
per-second frame log.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Per-second frame log reports fps/ms, visible/total object & triangle
ratios, VRAM breakdown (VBO+EBO), model count, and pending uploads.
Upload-complete log includes per-model VBO/EBO MB and scene total VRAM.
Streamer runs an instancing analysis keyed on geom.id(): total shapes,
unique representations, dedup ratio, theoretical VBO/EBO/SSBO sizes if
instanced, potential savings, and top-5 most-duplicated representations.
Used to validate whether GPU instancing is worth the architectural
rewrite for a given dataset.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Phase 2 performance: BVH acceleration with median-split build, per-model
trees, and EBO re-sorting for GPU cache coherence. Raw binary .ifcview
sidecar stores full geometry + BVH for instant subsequent loads (skip
tessellation entirely).
Per-model GPU buffers (VAO/VBO/EBO per model) eliminate cross-model buffer
copies on growth. Sidecar reads happen on a background thread. Bulk GPU
uploads are progressive (48 MB/frame chunks) so the viewport stays
interactive while multi-GB models stream in.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduce ModelHandle and per-model GeometryStreamers so multiple IFC
files can be loaded simultaneously. Object IDs are globally unique
(monotonically increasing across models). File picker is now multiselect.
Each model gets a top-level tree node. Property lookup uses the correct
model's ifcopenshell::file. ViewportWindow supports hide/show/remove
per model via model_id filtering in the frustum cull pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Show FPS, frame time, visible/total objects, and visible/total
triangles in the status bar. Toggled via Settings > Show Performance
Stats, persisted in app settings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Track per-object AABB and index range during upload. Each frame,
extract frustum planes from the view-projection matrix and cull
objects whose AABB is entirely outside any plane. Draw only visible
objects via glMultiDrawElements. Document the three-phase rendering
performance strategy in README.md.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Dion Moult