Glassvision/IfcOpenShell
1
0
Fork 0
mirror of https://github.com/IfcOpenShell/IfcOpenShell.git synced 2026-09-27 02:31:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

wgpu backend: 10× perf — megadraw, async HiZ, parallel cull, motion mode

Browse Source 
Closes the perf gap to the GL backend on real BIM benchmarks. On a 10-
sidecar / 380k-instance corpus at a fixed --camera the wgpu binary went
from 110.6 ms to 11.6 ms (vs GL's 23 ms — half the frame time, but
note GL is doing extra work the wgpu backend hasn't ported yet; see
the caveats list at the bottom). Bundled because the pieces interlock
and shipping any of them without the others reintroduces the same wall.

1. Cross-mesh vertex pulling (single mega-draw per model)
   The previous one-drawIndexed-per-(mesh × LOD-bucket) loop was costing
   ~13ms on a 27k-mesh scene. CPU now emits a flat visible_draws[]
   (16 B per visible (mesh,lod,instance)) plus a prefix_sums[] table.
   WGSL binary-searches prefix_sums by @builtin(vertex_index) to find
   the entry, then manually fetches the mesh-local index from a
   storage-bound indices[] and pulls the packed 12 B vertex. No
   setIndexBuffer; the shader reads everything from storage. Bind
   group grew from 4 to 7 entries (vertices, meshes, instances,
   indices, visible_draws, prefix_sums, per-model uniform) — well
   under WebGPU's mandatory 8 storage / 12 uniform floor.

2. Async HiZ readback via ping-pong staging buffers
   Sync wait via wgpuInstanceProcessEvents was costing ~37 ms on a
   real scene (GPU drain). Two staging slots now ping-pong: frame N
   kicks a non-blocking mapAsync on slot K, frame N+1's first action
   is one processEvents drain. Pyramid is 1-2 frames stale — matches
   the "slightly-stale depth, fine" pattern the GL backend already
   documents. encodeHizResolve returns -1 (skip) if both slots are
   in flight; cull keeps using the most recent pyramid.

3. Cull reorder: contribution before HiZ
   HiZ projection is ~10× more expensive than the contribution
   check, yet most contribution-survivors would be HiZ-rejected
   anyway on dense scenes. Computing projected_px first lets
   contribution short-circuit ~80% of HiZ tests with no rejection-
   quality loss. Saved ~34 ms on the dense bench.

4. Motion-mode contribution threshold
   AppSettings::motionMinPixelRadius parity. While the camera is
   changing (orbit/pan/zoom/--benchmark sweep), drop instances
   below 10 px instead of 2 px. Halves visible_objects during
   motion with no perceived quality loss.

5. Parallel cull (std::async across models)
   Per-model cullModelCpu split into Compute (CPU-only, thread-safe)
   + Upload (main-thread wgpu queue writes). std::async fan-outs the
   compute across models; main-thread joins and uploads. Wall-clock
   cull on the 10-model corpus drops from ~17 ms single-threaded to
   ~9 ms across cores.

6. --no-hiz CLI flag + per-phase benchmark timings
   Benchmark now also prints "per-frame avg ms: cull=X
   hiz_readback=Y" so future regressions can be attributed without
   guesswork. --no-hiz toggles the master switch from the CLI.

Honest caveats — wgpu is currently faster mostly because GL is doing
work we haven't ported yet:
  - Edge silhouette pass (stage 9) will add ~3-5 ms back to wgpu.
  - GL's HiZ uses the BVH so it rejects whole subtrees (1.7k vs
    our 358 rejects on the same scene). BVH for HiZ is future work
    (task #13 / a new task) — until then we draw more sub-pixel
    geometry that's behind closer surfaces. Visually correct, perf
    cost paid. Stage 4+5 are unaffected.

Verified pixel-identical on basic.ifc through every change. Real-scene
visual diff against GL pending the --screenshot flag on the GL minimal
(task #10's other half).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Dion Moult
2026-05-27 19:46:13 +10:00
parent 406124ca3d
commit 51dc31a50b
4 changed files with 551 additions and 283 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/ifcviewer-wgpu-minimal/main.cpp
+3
View File
@@ -48,10 +48,13 @@ int main(int argc, char* argv[]) {
parser.addOption({{"c", "camera"}, parser.addOption({{"c", "camera"},
"Set camera as tx,ty,tz,dist,yaw,pitch (same format as IfcViewerMinimal).", "Set camera as tx,ty,tz,dist,yaw,pitch (same format as IfcViewerMinimal).",
"params"}); "params"});
parser.addOption({"no-hiz",
"Disable HiZ occlusion culling for perf diagnostics."});
parser.process(app); parser.process(app);
auto* viewport = new WgpuViewportWindow; auto* viewport = new WgpuViewportWindow;
viewport->resize(1280, 800); viewport->resize(1280, 800);
if (parser.isSet("no-hiz")) viewport->hiz_enabled_ = false;
QWidget* container = QWidget::createWindowContainer(viewport); QWidget* container = QWidget::createWindowContainer(viewport);
container->setMinimumSize(320, 240); container->setMinimumSize(320, 240);
src/ifcviewer-wgpu/WgpuModelGpuData.h
+36 -20
View File
@@ -51,31 +51,47 @@ struct WgpuModelGpuData {
// recompose from placement_transformation when stage matrices change. // recompose from placement_transformation when stage matrices change.
WGPUBuffer instance_storage = nullptr; WGPUBuffer instance_storage = nullptr;
// u32[] of visible instance indices, repacked per frame by the CPU cull // Cross-mesh vertex pulling: one entry per visible (mesh, lod, instance)
// into per-mesh contiguous slices. Sized for the worst case // tuple, written into a single flat buffer per frame. The vertex shader
// (instance_count entries) at applyCachedModel time so we never have to // binary-searches `prefix_sums` to find which entry a given
// recreate it (and the bind group that references it) mid-frame. // gl_VertexID belongs to, then fetches that entry's mesh slice via the
WGPUBuffer visible_buffer = nullptr; // offsets baked here. Pre-sized at applyCachedModel to a generous
size_t visible_buffer_capacity = 0; // entries (u32 count), not bytes // worst case so the bind group stays valid for the model's lifetime.
//
// std430 layout: 16 bytes per entry, naturally aligned.
struct alignas(16) VisibleDrawGpu {
uint32_t mesh_id; // -> meshes[] for quantisation basis
uint32_t instance_idx; // -> instances[] for transform + ids
uint32_t ebo_first_u32; // start of this entry's slice in indices[]
uint32_t base_vertex; // start of this mesh's slice in vertices[]
};
static_assert(sizeof(VisibleDrawGpu) == 16, "VisibleDrawGpu must be 16 bytes");
// Bind group binding the four storage buffers above (group=1 in the WGPUBuffer visible_draws_buffer = nullptr;
// main pipeline). Built in applyCachedModel after the buffers exist. size_t visible_draws_capacity = 0; // entries (not bytes)
// prefix_sums[i] = sum of index_counts for visible_draws[0..i-1].
// Sized to visible_draws_capacity + 1 so prefix_sums[N] = total verts.
WGPUBuffer prefix_sums_buffer = nullptr;
size_t prefix_sums_capacity = 0; // entries (u32 count)
// Per-model uniform: (draw_count, total_vertex_count, _pad, _pad).
// The vertex shader uses draw_count to bound the binary search; CPU
// uses total_vertex_count as the draw() call's vertex count.
WGPUBuffer per_model_uniform = nullptr;
// Bind group binding the storage + uniform buffers above (group=1 in
// the main pipeline). Built in applyCachedModel.
WGPUBindGroup bind_group = nullptr; WGPUBindGroup bind_group = nullptr;
// One per mesh, populated per frame by cullAndCompact. instance_count==0 // Per-frame, populated by cullModelCpu and consumed by render().
// means the mesh contributes nothing this frame and the draw is skipped. uint32_t total_visible_vertices = 0;
struct MeshDraw { uint32_t total_visible_draws = 0;
uint32_t first_instance = 0; // offset into visible_buffer
uint32_t instance_count = 0;
uint32_t first_index = 0; // index buffer offset (in u32 indices)
int32_t base_vertex = 0; // added to every fetched vertex_index
uint32_t index_count = 0;
};
std::vector<MeshDraw> mesh_draws; // size = meshes.size() after first frame
// Scratch reused each frame so per-frame cull doesn't allocate. Sized // Scratch reused each frame so per-frame cull doesn't allocate. Sized
// to instance_count entries on first use; never shrunk. // on first use; never shrunk.
std::vector<uint32_t> visible_flat_scratch; std::vector<VisibleDrawGpu> visible_draws_scratch;
std::vector<uint32_t> prefix_sums_scratch;
// Size mirrors for stats / range checks. // Size mirrors for stats / range checks.
size_t vertex_bytes = 0; size_t vertex_bytes = 0;
src/ifcviewer-wgpu/WgpuViewportWindow.cpp
+443 -243
View File
@@ -33,8 +33,10 @@
#include <webgpu/wgpu.h> // wgpu-native extensions (logging, MULTI_DRAW_INDIRECT, …) #include <webgpu/wgpu.h> // wgpu-native extensions (logging, MULTI_DRAW_INDIRECT, …)
#include <algorithm> #include <algorithm>
#include <atomic>
#include <cmath> #include <cmath>
#include <cstring> #include <cstring>
#include <future>
#include <limits> #include <limits>
#include <utility> #include <utility>
@@ -130,32 +132,42 @@ static WGPUBuffer createBufferWithData(WGPUDevice device, WGPUQueue queue,
} }
void releaseWgpuModelGpuData(WgpuModelGpuData& m) { void releaseWgpuModelGpuData(WgpuModelGpuData& m) {
if (m.bind_group) { wgpuBindGroupRelease(m.bind_group); m.bind_group = nullptr; } if (m.bind_group) { wgpuBindGroupRelease(m.bind_group); m.bind_group = nullptr; }
if (m.vertex_storage) { wgpuBufferRelease(m.vertex_storage); m.vertex_storage = nullptr; } if (m.vertex_storage) { wgpuBufferRelease(m.vertex_storage); m.vertex_storage = nullptr; }
if (m.index_buffer) { wgpuBufferRelease(m.index_buffer); m.index_buffer = nullptr; } if (m.index_buffer) { wgpuBufferRelease(m.index_buffer); m.index_buffer = nullptr; }
if (m.mesh_storage) { wgpuBufferRelease(m.mesh_storage); m.mesh_storage = nullptr; } if (m.mesh_storage) { wgpuBufferRelease(m.mesh_storage); m.mesh_storage = nullptr; }
if (m.instance_storage) { wgpuBufferRelease(m.instance_storage); m.instance_storage = nullptr; } if (m.instance_storage) { wgpuBufferRelease(m.instance_storage); m.instance_storage = nullptr; }
if (m.visible_buffer) { wgpuBufferRelease(m.visible_buffer); m.visible_buffer = nullptr; } if (m.visible_draws_buffer) { wgpuBufferRelease(m.visible_draws_buffer); m.visible_draws_buffer = nullptr; }
if (m.prefix_sums_buffer) { wgpuBufferRelease(m.prefix_sums_buffer); m.prefix_sums_buffer = nullptr; }
if (m.per_model_uniform) { wgpuBufferRelease(m.per_model_uniform); m.per_model_uniform = nullptr; }
m.vertex_bytes = 0; m.vertex_bytes = 0;
m.index_count = 0; m.index_count = 0;
m.mesh_count = 0; m.mesh_count = 0;
m.instance_count = 0; m.instance_count = 0;
m.visible_buffer_capacity = 0; m.visible_draws_capacity = 0;
m.prefix_sums_capacity = 0;
m.total_visible_vertices = 0;
m.total_visible_draws = 0;
m.meshes.clear(); m.meshes.clear();
m.instances.clear(); m.instances.clear();
m.mesh_draws.clear(); m.visible_draws_scratch.clear();
m.visible_flat_scratch.clear(); m.prefix_sums_scratch.clear();
} }
// ----------------------------------------------------------------------------- // -----------------------------------------------------------------------------
// WGSL main pipeline — vertex pulling. // WGSL main pipeline — cross-mesh vertex pulling.
// //
// Vertex bytes are read manually from a u32 storage buffer (3 u32 per vertex, // We issue ONE draw() call per model per frame. The vertex shader binary-
// matching INSTANCED_VERTEX_STRIDE_BYTES=12). gl_BaseVertex is folded into // searches the prefix-sum table to find which visible-draw entry the current
// vertex_index by the WebGPU IA, so per-mesh draws set baseVertex = // @builtin(vertex_index) belongs to, then manually fetches the index and the
// mesh.vbo_byte_offset / 12 and the shader indexes the global storage array // 12-byte packed vertex from storage buffers. This avoids the N-drawcalls-per-
// directly. firstInstance carries the global instance slot, fetched as // frame CPU overhead of per-mesh draws (which dominated on scenes with many
// @builtin(instance_index). // unique meshes — wgpu-native overhead is ~5 µs/draw, so 27k draws = 135ms).
//
// Binary search cost is O(log N) per vertex, with N up to a few hundred
// thousand on dense scenes. Adjacent vertices in the same draw entry share
// the search result inside a warp, so memory-coherence keeps this cheap on
// GPU.
// ----------------------------------------------------------------------------- // -----------------------------------------------------------------------------
static const char* MAIN_WGSL = R"( static const char* MAIN_WGSL = R"(
@@ -180,15 +192,29 @@ struct FrameUniforms {
ground_color: vec4<f32>, ground_color: vec4<f32>,
}; };
struct VisibleDraw {
mesh_id: u32,
instance_idx: u32,
ebo_first_u32: u32,
base_vertex: u32,
};
struct PerModel {
draw_count: u32,
total_vertex_count: u32,
_pad0: u32,
_pad1: u32,
};
@group(0) @binding(0) var<uniform> u_frame: FrameUniforms; @group(0) @binding(0) var<uniform> u_frame: FrameUniforms;
@group(1) @binding(0) var<storage, read> vertices: array<u32>; @group(1) @binding(0) var<storage, read> vertices: array<u32>;
@group(1) @binding(1) var<storage, read> meshes: array<MeshQuant>; @group(1) @binding(1) var<storage, read> meshes: array<MeshQuant>;
@group(1) @binding(2) var<storage, read> instances: array<InstanceRecord>; @group(1) @binding(2) var<storage, read> instances: array<InstanceRecord>;
// Per-frame compacted visible list: visible[firstInstance + i] picks the @group(1) @binding(3) var<storage, read> indices: array<u32>;
// real instance for this draw slot. firstInstance is set per drawIndexed @group(1) @binding(4) var<storage, read> visible_draws: array<VisibleDraw>;
// call so each mesh reads its own contiguous slice of `visible`. @group(1) @binding(5) var<storage, read> prefix_sums: array<u32>;
@group(1) @binding(3) var<storage, read> visible: array<u32>; @group(1) @binding(6) var<uniform> u_model: PerModel;
struct VsOut { struct VsOut {
@builtin(position) clip_pos: vec4<f32>, @builtin(position) clip_pos: vec4<f32>,
@@ -214,16 +240,46 @@ fn octDecode(e: vec2<f32>) -> vec3<f32> {
return normalize(n); return normalize(n);
} }
@vertex // Binary search for the largest i in [0, draw_count) with prefix_sums[i] <= vid.
fn vs_main(@builtin(vertex_index) vid: u32, // prefix_sums is monotonic non-decreasing and contains draw_count+1 entries
@builtin(instance_index) iid: u32) -> VsOut { // (prefix_sums[draw_count] == total_vertex_count).
let inst_idx = visible[iid]; fn find_draw(vid: u32) -> u32 {
let inst = instances[inst_idx]; var lo: u32 = 0u;
let mq = meshes[inst.mesh_id]; var hi: u32 = u_model.draw_count;
while (lo + 1u < hi) {
let mid = (lo + hi) >> 1u;
if (prefix_sums[mid] <= vid) {
lo = mid;
} else {
hi = mid;
}
}
return lo;
}
let w0 = vertices[vid * 3u + 0u]; @vertex
let w1 = vertices[vid * 3u + 1u]; fn vs_main(@builtin(vertex_index) vid: u32) -> VsOut {
let w2 = vertices[vid * 3u + 2u]; // Saturate past the end (shouldn't happen given draw() count, but safe).
if (vid >= u_model.total_vertex_count) {
var degen: VsOut;
degen.clip_pos = vec4<f32>(0.0, 0.0, 0.0, 0.0);
return degen;
}
let draw_idx = find_draw(vid);
let local_v = vid - prefix_sums[draw_idx];
let item = visible_draws[draw_idx];
// Fetch the mesh-local index then the global vertex index.
let mesh_local_index = indices[item.ebo_first_u32 + local_v];
let v_global = item.base_vertex + mesh_local_index;
let inst = instances[item.instance_idx];
let mq = meshes[item.mesh_id];
let w0 = vertices[v_global * 3u + 0u];
let w1 = vertices[v_global * 3u + 1u];
let w2 = vertices[v_global * 3u + 2u];
let px = f32(w0 & 0xFFFFu) / 65535.0; let px = f32(w0 & 0xFFFFu) / 65535.0;
let py = f32((w0 >> 16u) & 0xFFFFu) / 65535.0; let py = f32((w0 >> 16u) & 0xFFFFu) / 65535.0;
@@ -428,11 +484,13 @@ void WgpuViewportWindow::applyCachedModel(uint32_t model_id, SidecarData data) {
WGPUBufferUsage_Storage, WGPUBufferUsage_Storage,
"model.vertex_storage"); "model.vertex_storage");
// Index buffer — mesh-local u32 indices; baseVertex applied per-draw. // Index buffer — mesh-local u32 indices. Bound as a read-only storage
// buffer in the vertex shader (cross-mesh vertex pulling manually fetches
// indices); Index usage is kept for forward compatibility but never used.
m.index_buffer = createBufferWithData( m.index_buffer = createBufferWithData(
device_, queue_, device_, queue_,
data.indices.data(), data.indices.size() * sizeof(uint32_t), data.indices.data(), data.indices.size() * sizeof(uint32_t),
WGPUBufferUsage_Index, WGPUBufferUsage_Storage | WGPUBufferUsage_Index,
"model.index_buffer"); "model.index_buffer");
// Derive MeshGpu[] (vec4 aabb_min + vec4 aabb_max) from MeshInfo's // Derive MeshGpu[] (vec4 aabb_min + vec4 aabb_max) from MeshInfo's
@@ -476,24 +534,43 @@ void WgpuViewportWindow::applyCachedModel(uint32_t model_id, SidecarData data) {
WGPUBufferUsage_Storage, WGPUBufferUsage_Storage,
"model.instance_storage"); "model.instance_storage");
// Pre-size the visible buffer for the worst case (every instance visible). // Per-frame buffers for cross-mesh vertex pulling. Sized for the worst
// Per-frame cull writes a u32 list into it via wgpuQueueWriteBuffer; we // case so cullModelCpu only needs wgpuQueueWriteBuffer (never recreate)
// never grow it so the bind group reference stays valid for the model's // — keeps the bind group reference valid for the model's lifetime.
// lifetime. Minimum size 4 bytes — wgpu rejects zero-sized buffers. //
const size_t visible_bytes = std::max<size_t>( // VisibleDraw capacity: up to one entry per (instance × 2 LODs). LOD1 is
size_t(data.instances.size()) * sizeof(uint32_t), 4u); // only used when bake produced one, but allocating headroom is cheap
WGPUBufferDescriptor vb_desc = {}; // (~3 MB worst-case for a 100k-instance model).
vb_desc.size = visible_bytes; const size_t draw_cap = std::max<size_t>(size_t(data.instances.size()) * 2u, 1u);
vb_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst; const size_t draws_bytes = draw_cap * sizeof(WgpuModelGpuData::VisibleDrawGpu);
vb_desc.label = svFromCStr("model.visible_buffer"); WGPUBufferDescriptor vd_desc = {};
m.visible_buffer = wgpuDeviceCreateBuffer(device_, &vb_desc); vd_desc.size = std::max<uint64_t>(draws_bytes, 16);
m.visible_buffer_capacity = std::max<size_t>(data.instances.size(), 1u); vd_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
vd_desc.label = svFromCStr("model.visible_draws");
m.visible_draws_buffer = wgpuDeviceCreateBuffer(device_, &vd_desc);
m.visible_draws_capacity = draw_cap;
// prefix_sums: draw_cap + 1 entries (final entry = total vertex count).
const size_t ps_cap = draw_cap + 1;
WGPUBufferDescriptor ps_desc = {};
ps_desc.size = std::max<uint64_t>(ps_cap * sizeof(uint32_t), 16);
ps_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
ps_desc.label = svFromCStr("model.prefix_sums");
m.prefix_sums_buffer = wgpuDeviceCreateBuffer(device_, &ps_desc);
m.prefix_sums_capacity = ps_cap;
// 16-byte per-model uniform: (draw_count, total_vertex_count, _pad, _pad).
WGPUBufferDescriptor mu_desc = {};
mu_desc.size = 16;
mu_desc.usage = WGPUBufferUsage_Uniform | WGPUBufferUsage_CopyDst;
mu_desc.label = svFromCStr("model.per_model_uniform");
m.per_model_uniform = wgpuDeviceCreateBuffer(device_, &mu_desc);
// Hand off CPU mirrors (cull / picking will need them later). // Hand off CPU mirrors (cull / picking will need them later).
m.meshes = std::move(data.meshes); m.meshes = std::move(data.meshes);
m.instances = std::move(data.instances); m.instances = std::move(data.instances);
m.mesh_draws.assign(m.meshes.size(), WgpuModelGpuData::MeshDraw{}); m.visible_draws_scratch.reserve(draw_cap);
m.visible_flat_scratch.reserve(m.instances.size()); m.prefix_sums_scratch.reserve(ps_cap);
auto [inserted, _] = models_gpu_.emplace(model_id, std::move(m)); auto [inserted, _] = models_gpu_.emplace(model_id, std::move(m));
WgpuModelGpuData& mref = inserted->second; WgpuModelGpuData& mref = inserted->second;
@@ -1005,7 +1082,19 @@ void WgpuViewportWindow::ensureHizTextures(int viewport_w, int viewport_h) {
if (hiz_resolve_view_) { wgpuTextureViewRelease(hiz_resolve_view_); hiz_resolve_view_ = nullptr; } if (hiz_resolve_view_) { wgpuTextureViewRelease(hiz_resolve_view_); hiz_resolve_view_ = nullptr; }
if (hiz_resolve_texture_) { wgpuTextureRelease(hiz_resolve_texture_); hiz_resolve_texture_ = nullptr; } if (hiz_resolve_texture_) { wgpuTextureRelease(hiz_resolve_texture_); hiz_resolve_texture_ = nullptr; }
if (hiz_staging_buffer_) { wgpuBufferRelease(hiz_staging_buffer_); hiz_staging_buffer_ = nullptr; } for (int s = 0; s < HIZ_SLOTS; ++s) {
if (hiz_staging_buffers_[s]) {
// Force any pending map to finish before release (defensive: shouldn't happen on resize).
if (hiz_slot_state_[s] == HizSlotState::Mapped) {
wgpuBufferUnmap(hiz_staging_buffers_[s]);
}
wgpuBufferRelease(hiz_staging_buffers_[s]);
hiz_staging_buffers_[s] = nullptr;
}
hiz_slot_state_[s] = HizSlotState::Idle;
}
hiz_write_idx_ = 0;
hiz_valid_ = false;
if (hiz_bind_group_) { wgpuBindGroupRelease(hiz_bind_group_); hiz_bind_group_ = nullptr; } if (hiz_bind_group_) { wgpuBindGroupRelease(hiz_bind_group_); hiz_bind_group_ = nullptr; }
WGPUTextureDescriptor desc = {}; WGPUTextureDescriptor desc = {};
@@ -1028,15 +1117,19 @@ void WgpuViewportWindow::ensureHizTextures(int viewport_w, int viewport_h) {
vdesc.aspect = WGPUTextureAspect_DepthOnly; vdesc.aspect = WGPUTextureAspect_DepthOnly;
hiz_resolve_view_ = wgpuTextureCreateView(hiz_resolve_texture_, &vdesc); hiz_resolve_view_ = wgpuTextureCreateView(hiz_resolve_texture_, &vdesc);
// Staging buffer: pad each row to 256-byte alignment. // Staging buffers: pad each row to 256-byte alignment. Two slots
// ping-pong so GPU fill of slot N overlaps CPU read of slot N-1.
hiz_padded_bpr_ = uint32_t( hiz_padded_bpr_ = uint32_t(
(dst_w * sizeof(float) + WGPU_BYTES_PER_ROW_ALIGN - 1) (dst_w * sizeof(float) + WGPU_BYTES_PER_ROW_ALIGN - 1)
/ WGPU_BYTES_PER_ROW_ALIGN * WGPU_BYTES_PER_ROW_ALIGN); / WGPU_BYTES_PER_ROW_ALIGN * WGPU_BYTES_PER_ROW_ALIGN);
WGPUBufferDescriptor bdesc = {}; for (int s = 0; s < HIZ_SLOTS; ++s) {
bdesc.size = uint64_t(hiz_padded_bpr_) * uint64_t(dst_h); WGPUBufferDescriptor bdesc = {};
bdesc.usage = WGPUBufferUsage_CopyDst | WGPUBufferUsage_MapRead; bdesc.size = uint64_t(hiz_padded_bpr_) * uint64_t(dst_h);
bdesc.label = svFromCStr("ifcviewer-wgpu.hiz_staging"); bdesc.usage = WGPUBufferUsage_CopyDst | WGPUBufferUsage_MapRead;
hiz_staging_buffer_ = wgpuDeviceCreateBuffer(device_, &bdesc); bdesc.label = svFromCStr(s == 0 ? "ifcviewer-wgpu.hiz_staging[0]"
: "ifcviewer-wgpu.hiz_staging[1]");
hiz_staging_buffers_[s] = wgpuDeviceCreateBuffer(device_, &bdesc);
}
hiz_resolve_w_ = dst_w; hiz_resolve_w_ = dst_w;
hiz_resolve_h_ = dst_h; hiz_resolve_h_ = dst_h;
@@ -1048,7 +1141,17 @@ void WgpuViewportWindow::releaseHizResources() {
if (hiz_uniform_buffer_) { wgpuBufferRelease(hiz_uniform_buffer_); hiz_uniform_buffer_ = nullptr; } if (hiz_uniform_buffer_) { wgpuBufferRelease(hiz_uniform_buffer_); hiz_uniform_buffer_ = nullptr; }
if (hiz_resolve_view_) { wgpuTextureViewRelease(hiz_resolve_view_); hiz_resolve_view_ = nullptr; } if (hiz_resolve_view_) { wgpuTextureViewRelease(hiz_resolve_view_); hiz_resolve_view_ = nullptr; }
if (hiz_resolve_texture_) { wgpuTextureRelease(hiz_resolve_texture_); hiz_resolve_texture_ = nullptr; } if (hiz_resolve_texture_) { wgpuTextureRelease(hiz_resolve_texture_); hiz_resolve_texture_ = nullptr; }
if (hiz_staging_buffer_) { wgpuBufferRelease(hiz_staging_buffer_); hiz_staging_buffer_ = nullptr; } for (int s = 0; s < HIZ_SLOTS; ++s) {
if (hiz_staging_buffers_[s]) {
if (hiz_slot_state_[s] == HizSlotState::Mapped) {
wgpuBufferUnmap(hiz_staging_buffers_[s]);
}
wgpuBufferRelease(hiz_staging_buffers_[s]);
hiz_staging_buffers_[s] = nullptr;
}
hiz_slot_state_[s] = HizSlotState::Idle;
}
hiz_write_idx_ = 0;
if (hiz_pipeline_) { wgpuRenderPipelineRelease(hiz_pipeline_); hiz_pipeline_ = nullptr; } if (hiz_pipeline_) { wgpuRenderPipelineRelease(hiz_pipeline_); hiz_pipeline_ = nullptr; }
if (hiz_shader_module_) { wgpuShaderModuleRelease(hiz_shader_module_); hiz_shader_module_ = nullptr; } if (hiz_shader_module_) { wgpuShaderModuleRelease(hiz_shader_module_); hiz_shader_module_ = nullptr; }
if (hiz_pipeline_layout_) { wgpuPipelineLayoutRelease(hiz_pipeline_layout_); hiz_pipeline_layout_ = nullptr; } if (hiz_pipeline_layout_) { wgpuPipelineLayoutRelease(hiz_pipeline_layout_); hiz_pipeline_layout_ = nullptr; }
@@ -1061,8 +1164,19 @@ void WgpuViewportWindow::releaseHizResources() {
hiz_mip_h_.clear(); hiz_mip_h_.clear();
} }
void WgpuViewportWindow::encodeHizResolve(WGPUCommandEncoder enc) { int WgpuViewportWindow::encodeHizResolve(WGPUCommandEncoder enc) {
if (!hiz_enabled_ || !hiz_pipeline_ || !hiz_resolve_view_ || !depth_view_) return; if (!hiz_enabled_ || !hiz_pipeline_ || !hiz_resolve_view_ || !depth_view_) return -1;
// Pick an idle ping-pong slot. If both slots are in flight, skip the
// resolve for this frame — the cull keeps using whatever pyramid we
// already built (slightly more stale than usual, but never blocks).
int slot = -1;
for (int s = 0; s < HIZ_SLOTS; ++s) {
const int idx = (hiz_write_idx_ + s) % HIZ_SLOTS;
if (hiz_slot_state_[idx] == HizSlotState::Idle) { slot = idx; break; }
}
if (slot < 0) return -1;
hiz_write_idx_ = (slot + 1) % HIZ_SLOTS;
// (Re)build the bind group every frame is wasteful; only rebuild when the // (Re)build the bind group every frame is wasteful; only rebuild when the
// depth view itself was replaced (driven by surface resize). For now we // depth view itself was replaced (driven by surface resize). For now we
@@ -1110,13 +1224,13 @@ void WgpuViewportWindow::encodeHizResolve(WGPUCommandEncoder enc) {
wgpuRenderPassEncoderEnd(pass); wgpuRenderPassEncoderEnd(pass);
wgpuRenderPassEncoderRelease(pass); wgpuRenderPassEncoderRelease(pass);
// Now copy the small resolved depth texture into the staging buffer. // Copy the small resolved depth texture into the chosen staging slot.
WGPUTexelCopyTextureInfo src = {}; WGPUTexelCopyTextureInfo src = {};
src.texture = hiz_resolve_texture_; src.texture = hiz_resolve_texture_;
src.aspect = WGPUTextureAspect_DepthOnly; src.aspect = WGPUTextureAspect_DepthOnly;
WGPUTexelCopyBufferInfo dst = {}; WGPUTexelCopyBufferInfo dst = {};
dst.buffer = hiz_staging_buffer_; dst.buffer = hiz_staging_buffers_[slot];
dst.layout.bytesPerRow = hiz_padded_bpr_; dst.layout.bytesPerRow = hiz_padded_bpr_;
dst.layout.rowsPerImage = hiz_resolve_h_; dst.layout.rowsPerImage = hiz_resolve_h_;
@@ -1126,89 +1240,112 @@ void WgpuViewportWindow::encodeHizResolve(WGPUCommandEncoder enc) {
extent.depthOrArrayLayers = 1; extent.depthOrArrayLayers = 1;
wgpuCommandEncoderCopyTextureToBuffer(enc, &src, &dst, &extent); wgpuCommandEncoderCopyTextureToBuffer(enc, &src, &dst, &extent);
return slot;
} }
void WgpuViewportWindow::readbackAndBuildHizPyramid(const QMatrix4x4& vp_used) { void WgpuViewportWindow::startHizMap(int slot, const QMatrix4x4& vp_used) {
if (!hiz_enabled_ || !hiz_staging_buffer_ || hiz_resolve_w_ == 0) return; if (slot < 0 || slot >= HIZ_SLOTS) return;
if (!hiz_staging_buffers_[slot] || hiz_resolve_w_ == 0) return;
struct MapReq { bool done = false; bool ok = false; }; hiz_slot_vp_[slot] = vp_used;
MapReq req; hiz_slot_state_[slot] = HizSlotState::Mapping;
struct MapCtx { WgpuViewportWindow* self; int slot; };
auto* ctx = new MapCtx{ this, slot };
WGPUBufferMapCallbackInfo mcb = {}; WGPUBufferMapCallbackInfo mcb = {};
mcb.mode = WGPUCallbackMode_AllowProcessEvents; mcb.mode = WGPUCallbackMode_AllowProcessEvents;
mcb.callback = [](WGPUMapAsyncStatus status, WGPUStringView /*msg*/, mcb.callback = [](WGPUMapAsyncStatus status, WGPUStringView /*msg*/,
void* ud1, void* /*ud2*/) { void* ud1, void* /*ud2*/) {
auto* r = static_cast<MapReq*>(ud1); auto* c = static_cast<MapCtx*>(ud1);
r->done = true; if (status == WGPUMapAsyncStatus_Success) {
r->ok = (status == WGPUMapAsyncStatus_Success); c->self->hiz_slot_state_[c->slot] = HizSlotState::Mapped;
} else {
c->self->hiz_slot_state_[c->slot] = HizSlotState::Idle;
}
delete c;
}; };
mcb.userdata1 = &req; mcb.userdata1 = ctx;
const size_t map_size = size_t(hiz_padded_bpr_) * size_t(hiz_resolve_h_); const size_t map_size = size_t(hiz_padded_bpr_) * size_t(hiz_resolve_h_);
wgpuBufferMapAsync(hiz_staging_buffer_, WGPUMapMode_Read, 0, map_size, mcb); wgpuBufferMapAsync(hiz_staging_buffers_[slot], WGPUMapMode_Read,
while (!req.done) wgpuInstanceProcessEvents(instance_); 0, map_size, mcb);
if (!req.ok) return; }
const uint8_t* mapped = static_cast<const uint8_t*>( void WgpuViewportWindow::drainHizReadbacks() {
wgpuBufferGetConstMappedRange(hiz_staging_buffer_, 0, map_size)); if (!hiz_enabled_ || hiz_resolve_w_ == 0) return;
// Process any callbacks that have fired since last frame. Does NOT block:
// wgpuInstanceProcessEvents returns immediately after running ready
// callbacks. The mapAsync mode is AllowProcessEvents, so this is the
// correct drainage point.
wgpuInstanceProcessEvents(instance_);
// Mip 0: tight copy, stripping per-row padding. for (int slot = 0; slot < HIZ_SLOTS; ++slot) {
const uint32_t W0 = hiz_resolve_w_; if (hiz_slot_state_[slot] != HizSlotState::Mapped) continue;
const uint32_t H0 = hiz_resolve_h_;
// Pre-size storage: count levels until we hit 1×1. const size_t map_size = size_t(hiz_padded_bpr_) * size_t(hiz_resolve_h_);
hiz_mip_offset_.clear(); const uint8_t* mapped = static_cast<const uint8_t*>(
hiz_mip_w_.clear(); wgpuBufferGetConstMappedRange(hiz_staging_buffers_[slot], 0, map_size));
hiz_mip_h_.clear();
uint32_t total_texels = 0; const uint32_t W0 = hiz_resolve_w_;
{ const uint32_t H0 = hiz_resolve_h_;
uint32_t w = W0, h = H0;
while (true) { // (Re)build mip pyramid metadata if dimensions changed.
hiz_mip_offset_.push_back(total_texels); if (hiz_mip_offset_.empty()
hiz_mip_w_.push_back(w); || hiz_mip_w_.empty() || hiz_mip_w_[0] != W0
hiz_mip_h_.push_back(h); || hiz_mip_h_.empty() || hiz_mip_h_[0] != H0) {
total_texels += w * h; hiz_mip_offset_.clear();
if (w == 1 && h == 1) break; hiz_mip_w_.clear();
w = std::max(1u, w / 2u); hiz_mip_h_.clear();
h = std::max(1u, h / 2u); uint32_t total = 0;
uint32_t w = W0, h = H0;
while (true) {
hiz_mip_offset_.push_back(total);
hiz_mip_w_.push_back(w);
hiz_mip_h_.push_back(h);
total += w * h;
if (w == 1 && h == 1) break;
w = std::max(1u, w / 2u);
h = std::max(1u, h / 2u);
}
hiz_pyramid_.assign(total, 0.0f);
} }
}
hiz_pyramid_.assign(total_texels, 0.0f);
// Mip 0: strip row padding. // Mip 0: strip per-row padding.
for (uint32_t y = 0; y < H0; ++y) { for (uint32_t y = 0; y < H0; ++y) {
std::memcpy(&hiz_pyramid_[y * W0], std::memcpy(&hiz_pyramid_[y * W0],
mapped + size_t(y) * hiz_padded_bpr_, mapped + size_t(y) * hiz_padded_bpr_,
W0 * sizeof(float)); W0 * sizeof(float));
} }
wgpuBufferUnmap(hiz_staging_buffer_); wgpuBufferUnmap(hiz_staging_buffers_[slot]);
hiz_slot_state_[slot] = HizSlotState::Idle;
// Higher mips: max-reduce 2×2 children. Edge handling: if parent // Higher mips: max-reduce 2×2 children.
// dimension shrunk to 1, just read the single child. for (size_t L = 1; L < hiz_mip_offset_.size(); ++L) {
for (size_t L = 1; L < hiz_mip_offset_.size(); ++L) { const uint32_t prev_w = hiz_mip_w_[L - 1];
const uint32_t prev_w = hiz_mip_w_[L - 1]; const uint32_t prev_h = hiz_mip_h_[L - 1];
const uint32_t prev_h = hiz_mip_h_[L - 1]; const uint32_t this_w = hiz_mip_w_[L];
const uint32_t this_w = hiz_mip_w_[L]; const uint32_t this_h = hiz_mip_h_[L];
const uint32_t this_h = hiz_mip_h_[L]; const float* src = &hiz_pyramid_[hiz_mip_offset_[L - 1]];
const float* src = &hiz_pyramid_[hiz_mip_offset_[L - 1]]; float* dst = &hiz_pyramid_[hiz_mip_offset_[L]];
float* dst = &hiz_pyramid_[hiz_mip_offset_[L]]; for (uint32_t y = 0; y < this_h; ++y) {
for (uint32_t y = 0; y < this_h; ++y) { for (uint32_t x = 0; x < this_w; ++x) {
for (uint32_t x = 0; x < this_w; ++x) { const uint32_t x0 = std::min(prev_w - 1, x * 2u);
const uint32_t x0 = std::min(prev_w - 1, x * 2u); const uint32_t y0 = std::min(prev_h - 1, y * 2u);
const uint32_t y0 = std::min(prev_h - 1, y * 2u); const uint32_t x1 = std::min(prev_w - 1, x0 + 1u);
const uint32_t x1 = std::min(prev_w - 1, x0 + 1u); const uint32_t y1 = std::min(prev_h - 1, y0 + 1u);
const uint32_t y1 = std::min(prev_h - 1, y0 + 1u); const float a = src[y0 * prev_w + x0];
const float a = src[y0 * prev_w + x0]; const float b = src[y0 * prev_w + x1];
const float b = src[y0 * prev_w + x1]; const float c = src[y1 * prev_w + x0];
const float c = src[y1 * prev_w + x0]; const float d = src[y1 * prev_w + x1];
const float d = src[y1 * prev_w + x1]; dst[y * this_w + x] = std::max(std::max(a, b), std::max(c, d));
dst[y * this_w + x] = std::max(std::max(a, b), std::max(c, d)); }
} }
} }
}
hiz_vp_ = vp_used; hiz_vp_ = hiz_slot_vp_[slot];
hiz_valid_ = true; hiz_valid_ = true;
}
} }
bool WgpuViewportWindow::aabbOccludedByHiz(const float mn[3], const float mx[3]) const { bool WgpuViewportWindow::aabbOccludedByHiz(const float mn[3], const float mx[3]) const {
@@ -1302,47 +1439,41 @@ void WgpuViewportWindow::setBenchmarkFrames(int frames) {
if (isExposed() && bench_total_ > 0) requestUpdate(); if (isExposed() && bench_total_ > 0) requestUpdate();
} }
void WgpuViewportWindow::cullModelCpu(WgpuModelGpuData& m, uint32_t WgpuViewportWindow::cullModelCpuCompute(WgpuModelGpuData& m,
const float planes[6][4], const float planes[6][4],
const float eye[3], const float forward[3], const float eye[3], const float forward[3],
float focal_px, float focal_px,
float min_radius_px, float min_radius_px,
float lod1_threshold_px, float lod1_threshold_px,
bool hiz_enabled) { bool hiz_enabled) const {
if (m.instances.empty() || m.meshes.empty() || !m.visible_buffer) { m.total_visible_vertices = 0;
for (auto& d : m.mesh_draws) d.instance_count = 0; m.total_visible_draws = 0;
return; uint32_t hiz_rejects = 0;
}
// Per-mesh visible-instance buckets split by LOD. Per-frame scratch held if (m.instances.empty() || m.meshes.empty()
// thread-locally so the cull path doesn't allocate. || !m.visible_draws_buffer || !m.prefix_sums_buffer || !m.per_model_uniform) {
static thread_local std::vector<std::vector<uint32_t>> per_mesh_lod0; return 0;
static thread_local std::vector<std::vector<uint32_t>> per_mesh_lod1;
if (per_mesh_lod0.size() < m.meshes.size()) per_mesh_lod0.resize(m.meshes.size());
if (per_mesh_lod1.size() < m.meshes.size()) per_mesh_lod1.resize(m.meshes.size());
for (size_t mi = 0; mi < m.meshes.size(); ++mi) {
per_mesh_lod0[mi].clear();
per_mesh_lod1[mi].clear();
} }
const bool contrib_enabled = (min_radius_px > 0.0f); const bool contrib_enabled = (min_radius_px > 0.0f);
const bool lod_enabled = (lod1_threshold_px > 0.0f); const bool lod_enabled = (lod1_threshold_px > 0.0f);
m.visible_draws_scratch.clear();
m.prefix_sums_scratch.clear();
m.prefix_sums_scratch.push_back(0); // prefix_sums[0] = 0
uint32_t running_vertex_count = 0;
for (uint32_t i = 0; i < uint32_t(m.instances.size()); ++i) { for (uint32_t i = 0; i < uint32_t(m.instances.size()); ++i) {
const auto& inst = m.instances[i]; const auto& inst = m.instances[i];
if (inst.mesh_id >= m.meshes.size()) continue; if (inst.mesh_id >= m.meshes.size()) continue;
if (!aabbInFrustum(inst.world_aabb_min, inst.world_aabb_max, planes)) continue; if (!aabbInFrustum(inst.world_aabb_min, inst.world_aabb_max, planes)) continue;
if (hiz_enabled
&& aabbOccludedByHiz(inst.world_aabb_min, inst.world_aabb_max)) {
++hiz_reject_count_;
continue;
}
const MeshInfo& mesh = m.meshes[inst.mesh_id]; const MeshInfo& mesh = m.meshes[inst.mesh_id];
// Projected bounding-sphere radius in pixels, shared between the // Projected bounding-sphere radius in pixels — shared between the
// contribution-cull and LOD-pick decisions. Computed once per // contribution-cull and LOD-pick decisions. Computed before HiZ so
// instance; guarded against behind-near-plane and division by zero. // contribution can short-circuit the per-instance HiZ projection
// (which is the bulk of cull cost on dense scenes).
float projected_px = std::numeric_limits<float>::infinity(); float projected_px = std::numeric_limits<float>::infinity();
if (contrib_enabled || (lod_enabled && mesh.lod1_index_count > 0)) { if (contrib_enabled || (lod_enabled && mesh.lod1_index_count > 0)) {
const float cx = 0.5f * (inst.world_aabb_min[0] + inst.world_aabb_max[0]); const float cx = 0.5f * (inst.world_aabb_min[0] + inst.world_aabb_max[0]);
@@ -1360,65 +1491,69 @@ void WgpuViewportWindow::cullModelCpu(WgpuModelGpuData& m,
} }
} }
// Contribution cull: drop instances projected below threshold. Most // Contribution cull before HiZ: HiZ is by far the most expensive
// of the per-frame cost on dense scenes — typically rejects 80-90% // per-instance test (8-corner projection + mip pyramid sample), so
// of frustum-visible instances on real BIM models. // letting cheap contribution drops happen first cuts the HiZ-tested
// population by ~5× on real scenes.
if (contrib_enabled && projected_px < min_radius_px) continue; if (contrib_enabled && projected_px < min_radius_px) continue;
// LOD pick: switch to LOD1 if the mesh has one baked and we're below if (hiz_enabled
// the LOD threshold. && aabbOccludedByHiz(inst.world_aabb_min, inst.world_aabb_max)) {
++hiz_rejects;
continue;
}
const bool use_lod1 = lod_enabled const bool use_lod1 = lod_enabled
&& mesh.lod1_index_count > 0 && mesh.lod1_index_count > 0
&& projected_px < lod1_threshold_px; && projected_px < lod1_threshold_px;
if (use_lod1) per_mesh_lod1[inst.mesh_id].push_back(i); // Emit one VisibleDraw entry per visible instance. Cross-mesh vertex
else per_mesh_lod0[inst.mesh_id].push_back(i); // pulling means the shader binary-searches prefix_sums to locate
// this entry from a global vertex_index.
WgpuModelGpuData::VisibleDrawGpu d;
d.mesh_id = inst.mesh_id;
d.instance_idx = i;
d.ebo_first_u32 = (use_lod1 ? mesh.lod1_ebo_byte_offset : mesh.ebo_byte_offset)
/ uint32_t(sizeof(uint32_t));
d.base_vertex = mesh.vbo_byte_offset / INSTANCED_VERTEX_STRIDE_BYTES;
m.visible_draws_scratch.push_back(d);
const uint32_t entry_vert_count = use_lod1 ? mesh.lod1_index_count
: mesh.index_count;
running_vertex_count += entry_vert_count;
m.prefix_sums_scratch.push_back(running_vertex_count);
} }
// Flatten into m.visible_flat_scratch as [mesh0 lod0 | mesh0 lod1 | mesh1 m.total_visible_draws = uint32_t(m.visible_draws_scratch.size());
// lod0 | mesh1 lod1 | …] and emit one MeshDraw per non-empty bucket. m.total_visible_vertices = running_vertex_count;
m.visible_flat_scratch.clear(); return hiz_rejects;
m.mesh_draws.clear(); }
m.mesh_draws.reserve(m.meshes.size() * 2);
for (uint32_t mi = 0; mi < m.meshes.size(); ++mi) {
const MeshInfo& mesh = m.meshes[mi];
if (!per_mesh_lod0[mi].empty()) { void WgpuViewportWindow::cullModelCpuUpload(WgpuModelGpuData& m) {
WgpuModelGpuData::MeshDraw d; if (!m.visible_draws_buffer || !m.prefix_sums_buffer || !m.per_model_uniform) return;
d.first_instance = uint32_t(m.visible_flat_scratch.size());
d.instance_count = uint32_t(per_mesh_lod0[mi].size()); if (m.total_visible_draws == 0) {
d.first_index = mesh.ebo_byte_offset / uint32_t(sizeof(uint32_t)); // No visible geometry this frame; render() will skip the draw.
d.base_vertex = int32_t(mesh.vbo_byte_offset / INSTANCED_VERTEX_STRIDE_BYTES); // Still update the uniform so the shader sees zero (defensive).
d.index_count = mesh.index_count; const uint32_t um[4] = { 0, 0, 0, 0 };
for (uint32_t inst_idx : per_mesh_lod0[mi]) { wgpuQueueWriteBuffer(queue_, m.per_model_uniform, 0, um, sizeof(um));
m.visible_flat_scratch.push_back(inst_idx); return;
}
m.mesh_draws.push_back(d);
}
if (!per_mesh_lod1[mi].empty()) {
WgpuModelGpuData::MeshDraw d;
d.first_instance = uint32_t(m.visible_flat_scratch.size());
d.instance_count = uint32_t(per_mesh_lod1[mi].size());
d.first_index = mesh.lod1_ebo_byte_offset / uint32_t(sizeof(uint32_t));
d.base_vertex = int32_t(mesh.vbo_byte_offset / INSTANCED_VERTEX_STRIDE_BYTES);
d.index_count = mesh.lod1_index_count;
for (uint32_t inst_idx : per_mesh_lod1[mi]) {
m.visible_flat_scratch.push_back(inst_idx);
}
m.mesh_draws.push_back(d);
}
} }
// Upload the visible list. Round up to 4-byte multiple (always true for wgpuQueueWriteBuffer(queue_, m.visible_draws_buffer, 0,
// u32 arrays). Empty visible list still uploads one zero so the bind m.visible_draws_scratch.data(),
// group has something well-defined; the per-mesh loop will skip the draw. m.visible_draws_scratch.size()
const size_t bytes_to_upload = std::max<size_t>( * sizeof(WgpuModelGpuData::VisibleDrawGpu));
m.visible_flat_scratch.size() * sizeof(uint32_t), sizeof(uint32_t)); wgpuQueueWriteBuffer(queue_, m.prefix_sums_buffer, 0,
static const uint32_t zero = 0; m.prefix_sums_scratch.data(),
const void* src = m.visible_flat_scratch.empty() m.prefix_sums_scratch.size() * sizeof(uint32_t));
? static_cast<const void*>(&zero)
: static_cast<const void*>(m.visible_flat_scratch.data()); const uint32_t um[4] = {
wgpuQueueWriteBuffer(queue_, m.visible_buffer, 0, src, bytes_to_upload); m.total_visible_draws,
m.total_visible_vertices,
0, 0,
};
wgpuQueueWriteBuffer(queue_, m.per_model_uniform, 0, um, sizeof(um));
} }
void WgpuViewportWindow::render() { void WgpuViewportWindow::render() {
@@ -1427,6 +1562,10 @@ void WgpuViewportWindow::render() {
QElapsedTimer frame_timer; QElapsedTimer frame_timer;
if (bench_total_ > 0) frame_timer.start(); if (bench_total_ > 0) frame_timer.start();
// Drain any HiZ async readbacks that completed since last frame so the
// pyramid is as fresh as it can be before cull runs.
if (hiz_enabled_) drainHizReadbacks();
WGPUSurfaceTexture surf_tex = {}; WGPUSurfaceTexture surf_tex = {};
wgpuSurfaceGetCurrentTexture(surface_, &surf_tex); wgpuSurfaceGetCurrentTexture(surface_, &surf_tex);
@@ -1462,6 +1601,8 @@ void WgpuViewportWindow::render() {
last_visible_triangles_ = 0; last_visible_triangles_ = 0;
last_sub_draws_ = 0; last_sub_draws_ = 0;
hiz_reject_count_ = 0; hiz_reject_count_ = 0;
QElapsedTimer cull_timer;
if (bench_total_ > 0) cull_timer.start();
QMatrix4x4 vp_this_frame; QMatrix4x4 vp_this_frame;
{ {
const QVector3D target(camera_target_[0], camera_target_[1], camera_target_[2]); const QVector3D target(camera_target_[0], camera_target_[1], camera_target_[2]);
@@ -1488,18 +1629,62 @@ void WgpuViewportWindow::render() {
/ std::tan(qDegreesToRadians(camera_fov_y_deg_) * 0.5f)) / std::tan(qDegreesToRadians(camera_fov_y_deg_) * 0.5f))
: 0.0f; : 0.0f;
// Motion detection: any change in camera state since last frame
// bumps the contribution threshold to motion_min_pixel_radius_
// (mirrors GL's NavPreset behaviour, drops more sub-pixel work
// during orbit/pan/zoom).
const bool camera_moved = has_prev_camera_
&& (camera_target_[0] != prev_camera_target_[0]
|| camera_target_[1] != prev_camera_target_[1]
|| camera_target_[2] != prev_camera_target_[2]
|| camera_distance_ != prev_camera_distance_
|| camera_yaw_deg_ != prev_camera_yaw_deg_
|| camera_pitch_deg_ != prev_camera_pitch_deg_);
const float effective_min_px =
(camera_moved && motion_min_pixel_radius_ > min_pixel_radius_)
? motion_min_pixel_radius_
: min_pixel_radius_;
// Cull each model on its own worker thread. wgpu queue writes are
// serialised on the main thread after the parallel compute joins —
// wgpu-native doesn't guarantee thread-safety on queue ops.
std::vector<std::pair<uint32_t, std::future<uint32_t>>> futures;
futures.reserve(models_gpu_.size());
for (auto& [mid, m] : models_gpu_) { for (auto& [mid, m] : models_gpu_) {
if (m.hidden) continue; if (m.hidden) continue;
cullModelCpu(m, planes, eye_a, fwd_a, focal_px, auto& m_ref = m;
min_pixel_radius_, lod1_pixel_threshold_, futures.emplace_back(mid, std::async(std::launch::async,
hiz_enabled_); [this, &m_ref, &planes, &eye_a, &fwd_a,
for (const auto& d : m.mesh_draws) { focal_px, effective_min_px]() {
if (d.instance_count == 0 || d.index_count == 0) continue; return cullModelCpuCompute(
last_visible_objects_ += d.instance_count; m_ref, planes, eye_a, fwd_a, focal_px,
last_visible_triangles_ += (d.index_count / 3u) * d.instance_count; effective_min_px, lod1_pixel_threshold_,
last_sub_draws_ += 1; hiz_enabled_);
} }));
} }
for (auto& [mid, fut] : futures) {
hiz_reject_count_ += fut.get();
}
for (auto& [mid, m] : models_gpu_) {
if (m.hidden) continue;
cullModelCpuUpload(m);
last_visible_objects_ += m.total_visible_draws;
last_visible_triangles_ += m.total_visible_vertices / 3u;
// One CPU drawcall per model now (cross-mesh vertex pulling).
if (m.total_visible_draws > 0) last_sub_draws_ += 1;
}
}
// Snapshot camera state for next frame's motion detection.
prev_camera_target_[0] = camera_target_[0];
prev_camera_target_[1] = camera_target_[1];
prev_camera_target_[2] = camera_target_[2];
prev_camera_distance_ = camera_distance_;
prev_camera_yaw_deg_ = camera_yaw_deg_;
prev_camera_pitch_deg_ = camera_pitch_deg_;
has_prev_camera_ = true;
if (bench_total_ > 0 && bench_count_ >= bench_warmup_) {
bench_cull_ms_total_ += double(cull_timer.nsecsElapsed()) / 1e6;
} }
WGPUCommandEncoder enc = wgpuDeviceCreateCommandEncoder(device_, nullptr); WGPUCommandEncoder enc = wgpuDeviceCreateCommandEncoder(device_, nullptr);
@@ -1539,37 +1724,22 @@ void WgpuViewportWindow::render() {
wgpuRenderPassEncoderSetBindGroup(pass, 0, frame_bind_group_, 0, nullptr); wgpuRenderPassEncoderSetBindGroup(pass, 0, frame_bind_group_, 0, nullptr);
for (const auto& [mid, m] : models_gpu_) { for (const auto& [mid, m] : models_gpu_) {
if (m.hidden || !m.bind_group || !m.index_buffer || m.index_count == 0) continue; if (m.hidden || !m.bind_group || m.total_visible_vertices == 0) continue;
wgpuRenderPassEncoderSetBindGroup(pass, 1, m.bind_group, 0, nullptr); wgpuRenderPassEncoderSetBindGroup(pass, 1, m.bind_group, 0, nullptr);
wgpuRenderPassEncoderSetIndexBuffer(pass, m.index_buffer, // Single CPU drawcall per model. The vertex shader binary-
WGPUIndexFormat_Uint32, // searches prefix_sums by vertex_index to find (mesh, instance)
0, WGPU_WHOLE_SIZE); // and pulls index + vertex data manually from storage buffers.
wgpuRenderPassEncoderDraw(pass, m.total_visible_vertices, 1, 0, 0);
// One drawIndexed per non-empty mesh. instanceCount is the number
// of frustum-surviving instances of this mesh; firstInstance is
// their offset into m.visible_buffer (consumed by the WGSL
// `visible[iid]` indirection). All-instances-culled meshes are
// skipped — no draw call issued at all.
for (const auto& d : m.mesh_draws) {
if (d.instance_count == 0 || d.index_count == 0) continue;
wgpuRenderPassEncoderDrawIndexed(
pass,
d.index_count,
d.instance_count,
d.first_index,
d.base_vertex,
d.first_instance);
}
} }
} }
wgpuRenderPassEncoderEnd(pass); wgpuRenderPassEncoderEnd(pass);
wgpuRenderPassEncoderRelease(pass); wgpuRenderPassEncoderRelease(pass);
// ---- HiZ: resolve MSAA depth → small single-sample → staging buffer // ---- HiZ: resolve MSAA depth → small single-sample → ping-pong slot
int hiz_submitted_slot = -1;
if (hiz_enabled_) { if (hiz_enabled_) {
encodeHizResolve(enc); hiz_submitted_slot = encodeHizResolve(enc);
} }
// ---- Optional capture: encode copy on the same command buffer ------- // ---- Optional capture: encode copy on the same command buffer -------
@@ -1683,11 +1853,17 @@ void WgpuViewportWindow::render() {
wgpuSurfacePresent(surface_); wgpuSurfacePresent(surface_);
wgpuTextureRelease(surf_tex.texture); wgpuTextureRelease(surf_tex.texture);
// ---- HiZ readback + mip pyramid for the *next* frame's cull --------- // ---- HiZ async readback handoff -------------------------------------
// Sync wait via processEvents — the staging buffer is small (≈ 256×160×4) // Don't block — just kick off the mapAsync for the slot we filled this
// so the stall is well under a millisecond on every backend. // frame. Drainage happens at the top of the *next* frame via
if (hiz_enabled_) { // drainHizReadbacks(), giving the GPU at least one frame of headroom.
readbackAndBuildHizPyramid(vp_this_frame); if (hiz_enabled_ && hiz_submitted_slot >= 0) {
QElapsedTimer hiz_timer;
if (bench_total_ > 0) hiz_timer.start();
startHizMap(hiz_submitted_slot, vp_this_frame);
if (bench_total_ > 0 && bench_count_ >= bench_warmup_) {
bench_hiz_readback_ms_total_ += double(hiz_timer.nsecsElapsed()) / 1e6;
}
} }
// ---- Benchmark integration + auto-quit ------------------------------- // ---- Benchmark integration + auto-quit -------------------------------
@@ -1736,6 +1912,11 @@ void WgpuViewportWindow::render() {
<< " tri " << last_visible_triangles_ << " tri " << last_visible_triangles_
<< " sub_draws " << last_sub_draws_ << " sub_draws " << last_sub_draws_
<< " hiz_rej " << hiz_reject_count_; << " hiz_rej " << hiz_reject_count_;
const double n = double(std::max(1, bench_total_));
qInfo().noquote().nospace()
<< " per-frame avg ms: cull=" << bench_cull_ms_total_ / n
<< " hiz_readback=" << bench_hiz_readback_ms_total_ / n
<< " hiz=" << (hiz_enabled_ ? "on" : "off");
qInfo().noquote() << "=== END BENCHMARK ===\n"; qInfo().noquote() << "=== END BENCHMARK ===\n";
bench_total_ = 0; bench_total_ = 0;
@@ -1764,14 +1945,22 @@ bool WgpuViewportWindow::buildPipelines() {
frame_bgl_desc.label = svFromCStr("ifcviewer-wgpu.frame_bgl"); frame_bgl_desc.label = svFromCStr("ifcviewer-wgpu.frame_bgl");
frame_bgl_ = wgpuDeviceCreateBindGroupLayout(device_, &frame_bgl_desc); frame_bgl_ = wgpuDeviceCreateBindGroupLayout(device_, &frame_bgl_desc);
WGPUBindGroupLayoutEntry model_entries[4] = {}; // 6 read-only storage buffers (vertices, meshes, instances, indices,
for (int i = 0; i < 4; ++i) { // visible_draws, prefix_sums) + 1 uniform (per-model count). All read
// in the vertex shader. WebGPU's mandatory min is 8 storage / 12 uniform
// per stage, so we're comfortably under the cap.
WGPUBindGroupLayoutEntry model_entries[7] = {};
for (int i = 0; i < 6; ++i) {
model_entries[i].binding = uint32_t(i); model_entries[i].binding = uint32_t(i);
model_entries[i].visibility = WGPUShaderStage_Vertex; model_entries[i].visibility = WGPUShaderStage_Vertex;
model_entries[i].buffer.type = WGPUBufferBindingType_ReadOnlyStorage; model_entries[i].buffer.type = WGPUBufferBindingType_ReadOnlyStorage;
} }
model_entries[6].binding = 6;
model_entries[6].visibility = WGPUShaderStage_Vertex;
model_entries[6].buffer.type = WGPUBufferBindingType_Uniform;
model_entries[6].buffer.minBindingSize = 16;
WGPUBindGroupLayoutDescriptor model_bgl_desc = {}; WGPUBindGroupLayoutDescriptor model_bgl_desc = {};
model_bgl_desc.entryCount = 4; model_bgl_desc.entryCount = 7;
model_bgl_desc.entries = model_entries; model_bgl_desc.entries = model_entries;
model_bgl_desc.label = svFromCStr("ifcviewer-wgpu.model_bgl"); model_bgl_desc.label = svFromCStr("ifcviewer-wgpu.model_bgl");
model_bgl_ = wgpuDeviceCreateBindGroupLayout(device_, &model_bgl_desc); model_bgl_ = wgpuDeviceCreateBindGroupLayout(device_, &model_bgl_desc);
@@ -1859,12 +2048,14 @@ void WgpuViewportWindow::buildModelBindGroup(WgpuModelGpuData& m) {
wgpuBindGroupRelease(m.bind_group); wgpuBindGroupRelease(m.bind_group);
m.bind_group = nullptr; m.bind_group = nullptr;
} }
if (!m.vertex_storage || !m.mesh_storage || !m.instance_storage || !m.visible_buffer) { if (!m.vertex_storage || !m.mesh_storage || !m.instance_storage
|| !m.index_buffer || !m.visible_draws_buffer || !m.prefix_sums_buffer
|| !m.per_model_uniform) {
// Empty model — no bind group needed; the draw loop will skip it. // Empty model — no bind group needed; the draw loop will skip it.
return; return;
} }
WGPUBindGroupEntry entries[4] = {}; WGPUBindGroupEntry entries[7] = {};
entries[0].binding = 0; entries[0].binding = 0;
entries[0].buffer = m.vertex_storage; entries[0].buffer = m.vertex_storage;
entries[0].size = WGPU_WHOLE_SIZE; entries[0].size = WGPU_WHOLE_SIZE;
@@ -1875,12 +2066,21 @@ void WgpuViewportWindow::buildModelBindGroup(WgpuModelGpuData& m) {
entries[2].buffer = m.instance_storage; entries[2].buffer = m.instance_storage;
entries[2].size = WGPU_WHOLE_SIZE; entries[2].size = WGPU_WHOLE_SIZE;
entries[3].binding = 3; entries[3].binding = 3;
entries[3].buffer = m.visible_buffer; entries[3].buffer = m.index_buffer;
entries[3].size = WGPU_WHOLE_SIZE; entries[3].size = WGPU_WHOLE_SIZE;
entries[4].binding = 4;
entries[4].buffer = m.visible_draws_buffer;
entries[4].size = WGPU_WHOLE_SIZE;
entries[5].binding = 5;
entries[5].buffer = m.prefix_sums_buffer;
entries[5].size = WGPU_WHOLE_SIZE;
entries[6].binding = 6;
entries[6].buffer = m.per_model_uniform;
entries[6].size = 16;
WGPUBindGroupDescriptor desc = {}; WGPUBindGroupDescriptor desc = {};
desc.layout = model_bgl_; desc.layout = model_bgl_;
desc.entryCount = 4; desc.entryCount = 7;
desc.entries = entries; desc.entries = entries;
desc.label = svFromCStr("ifcviewer-wgpu.model_bind_group"); desc.label = svFromCStr("ifcviewer-wgpu.model_bind_group");
m.bind_group = wgpuDeviceCreateBindGroup(device_, &desc); m.bind_group = wgpuDeviceCreateBindGroup(device_, &desc);
src/ifcviewer-wgpu/WgpuViewportWindow.h
+69 -20
View File
@@ -124,13 +124,19 @@ private:
void ensureHizTextures(int viewport_w, int viewport_h); void ensureHizTextures(int viewport_w, int viewport_h);
void releaseHizResources(); void releaseHizResources();
// Resolves the just-rendered MSAA depth into the small single-sample // Resolves the just-rendered MSAA depth into the small single-sample
// HiZ texture and copies it to the staging buffer. Encoded onto `enc` // HiZ texture and copies it to whichever staging slot is currently
// so it ships in the same command buffer as the main draw. // idle. Returns the slot index used, or -1 if both slots are still
void encodeHizResolve(WGPUCommandEncoder enc); // in flight (resolve is skipped this frame — fine, we already have
// Maps the staging buffer (waits via processEvents), max-reduces a CPU // a recent pyramid). Encoded onto `enc` so it ships in the same
// mip pyramid, stores the VP used. Run after submitting the encoder so // command buffer as the main draw.
// the GPU has begun the copy. Updates hiz_valid_ to true on success. int encodeHizResolve(WGPUCommandEncoder enc);
void readbackAndBuildHizPyramid(const QMatrix4x4& vp_used); // Issues a non-blocking mapAsync on `slot` after submit, so the
// callback can fire whenever the GPU has actually finished writing.
void startHizMap(int slot, const QMatrix4x4& vp_used);
// Drains pending mapAsync callbacks (via processEvents — does NOT
// block on GPU work). For any slot that just signalled Mapped, reads
// it, unmaps it, max-reduces the mip pyramid, and updates hiz_vp_.
void drainHizReadbacks();
// Project AABB through hiz_vp_ and test against the pyramid. False // Project AABB through hiz_vp_ and test against the pyramid. False
// (keep) if HiZ isn't valid yet, AABB straddles the near plane, or // (keep) if HiZ isn't valid yet, AABB straddles the near plane, or
// any projection is unreliable. True (cull) when AABB is provably // any projection is unreliable. True (cull) when AABB is provably
@@ -156,13 +162,22 @@ private:
// `lod1_threshold_px` — survivors projected below this get the mesh's // `lod1_threshold_px` — survivors projected below this get the mesh's
// LOD1 index slice when one was baked. // LOD1 index slice when one was baked.
// min_radius_px == 0 disables contribution culling. // min_radius_px == 0 disables contribution culling.
void cullModelCpu(WgpuModelGpuData& m, // CPU-only phase of cull: produces m.visible_draws_scratch /
const float planes[6][4], // prefix_sums_scratch and sets total_visible_draws / total_visible_
const float eye[3], const float forward[3], // vertices. Touches no wgpu state, so this can run on a worker thread
float focal_px, // (multiple models culled in parallel). Returns the number of HiZ
float min_radius_px, // rejections accumulated (caller adds to the per-frame stat).
float lod1_threshold_px, uint32_t cullModelCpuCompute(WgpuModelGpuData& m,
bool hiz_enabled); const float planes[6][4],
const float eye[3], const float forward[3],
float focal_px,
float min_radius_px,
float lod1_threshold_px,
bool hiz_enabled) const;
// Upload phase: wgpuQueueWriteBuffer for visible_draws / prefix_sums /
// per-model uniform. Main-thread only (wgpu queue ops are not all
// thread-safe).
void cullModelCpuUpload(WgpuModelGpuData& m);
bool wgpu_initialized_ = false; bool wgpu_initialized_ = false;
bool surface_configured_ = false; bool surface_configured_ = false;
@@ -221,11 +236,25 @@ private:
WGPUTexture hiz_resolve_texture_ = nullptr; WGPUTexture hiz_resolve_texture_ = nullptr;
WGPUTextureView hiz_resolve_view_ = nullptr; WGPUTextureView hiz_resolve_view_ = nullptr;
WGPUBuffer hiz_staging_buffer_ = nullptr;
uint32_t hiz_resolve_w_ = 0; uint32_t hiz_resolve_w_ = 0;
uint32_t hiz_resolve_h_ = 0; uint32_t hiz_resolve_h_ = 0;
uint32_t hiz_padded_bpr_ = 0; // bytes per row in the staging buffer uint32_t hiz_padded_bpr_ = 0; // bytes per row in the staging buffer
// Ping-pong async readback. Frame N submits a copy into slot
// hiz_write_idx_ and calls mapAsync (non-blocking) on that slot. Frame
// N+K (K ≥ 1) calls processEvents to drain callbacks; whichever slot
// signalled completion is mapped, read into hiz_pyramid_, and unmapped
// — making the pyramid 1+ frames stale, which is fine ("slightly-stale
// depth" pattern the GL backend already documents). Two slots overlap
// GPU write with CPU read; we never block on the readback.
enum class HizSlotState : uint8_t { Idle, Mapping, Mapped };
static constexpr int HIZ_SLOTS = 2;
WGPUBuffer hiz_staging_buffers_[HIZ_SLOTS] = { nullptr, nullptr };
QMatrix4x4 hiz_slot_vp_ [HIZ_SLOTS];
HizSlotState hiz_slot_state_ [HIZ_SLOTS] = { HizSlotState::Idle,
HizSlotState::Idle };
int hiz_write_idx_ = 0;
// CPU mip pyramid (max-reduce). hiz_pyramid_[hiz_mip_offset_[L] + y*W + x]. // CPU mip pyramid (max-reduce). hiz_pyramid_[hiz_mip_offset_[L] + y*W + x].
std::vector<float> hiz_pyramid_; std::vector<float> hiz_pyramid_;
std::vector<uint32_t> hiz_mip_offset_; std::vector<uint32_t> hiz_mip_offset_;
@@ -247,16 +276,20 @@ private:
float camera_near_ = 0.1f; float camera_near_ = 0.1f;
float camera_far_ = 10000.0f; float camera_far_ = 10000.0f;
// Drop instances whose projected bounding-sphere radius is below this // Contribution-cull thresholds. Still-frame uses min_pixel_radius_;
// many pixels. Mirrors AppSettings::minPixelRadius() (GL default 2.0; // when the camera changed since last frame, the bigger motion threshold
// motion mode uses 10.0 but we don't differentiate yet — that arrives // kicks in to drop more sub-pixel detail (and slash per-frame cull cost).
// with mouse-driven motion-state tracking later). // Matches AppSettings::minPixelRadius / motionMinPixelRadius in GL.
float min_pixel_radius_ = 2.0f; float min_pixel_radius_ = 2.0f;
float motion_min_pixel_radius_ = 10.0f;
public:
// Master switch for HiZ occlusion. Set false to skip the depth resolve // Master switch for HiZ occlusion. Set false to skip the depth resolve
// + readback + cull test entirely (matches IFC_NO_HIZ in the GL backend). // + readback + cull test entirely (matches IFC_NO_HIZ in the GL backend).
bool hiz_enabled_ = true; bool hiz_enabled_ = true;
private:
// Switch to LOD1 when an instance's projected bounding-sphere radius // Switch to LOD1 when an instance's projected bounding-sphere radius
// drops below this many pixels. 0 disables (always LOD0). Defaults // drops below this many pixels. 0 disables (always LOD0). Defaults
// mirror AppSettings::lod1PixelThreshold() in the GL backend. // mirror AppSettings::lod1PixelThreshold() in the GL backend.
@@ -274,6 +307,15 @@ private:
// user pointed it. // user pointed it.
bool initial_view_applied_ = false; bool initial_view_applied_ = false;
// Camera state at the previous render() for motion detection. Any
// change means we apply the motion contribution threshold this frame
// (drops more sub-pixel work mid-orbit; matches GL behaviour).
float prev_camera_target_[3] = { 0, 0, 0 };
float prev_camera_distance_ = 0.0f;
float prev_camera_yaw_deg_ = 0.0f;
float prev_camera_pitch_deg_ = 0.0f;
bool has_prev_camera_ = false;
// Pending one-shot screenshot, captured at the end of the next render(). // Pending one-shot screenshot, captured at the end of the next render().
QString pending_screenshot_path_; QString pending_screenshot_path_;
bool pending_screenshot_quit_ = false; bool pending_screenshot_quit_ = false;
@@ -301,6 +343,13 @@ private:
uint32_t last_visible_objects_ = 0; uint32_t last_visible_objects_ = 0;
uint32_t last_visible_triangles_ = 0; uint32_t last_visible_triangles_ = 0;
uint32_t last_sub_draws_ = 0; uint32_t last_sub_draws_ = 0;
// Phase-time accumulators for benchmark mode. Each window measures a
// distinct slice of render() so we can attribute frame cost. Totals
// across the timed window are divided by bench_total_ on print.
double bench_cull_ms_total_ = 0.0;
double bench_hiz_readback_ms_total_ = 0.0;
double bench_submit_ms_total_ = 0.0;
}; };
#endif // WGPUVIEWPORTWINDOW_H #endif // WGPUVIEWPORTWINDOW_H