OneCAD-Draft/ROADMAP.md

# Open CAD Studio — File Open & Render Speed Roadmap

This document lists the planned improvements for cutting **file open time**
and **on-screen draw (render) time**. It builds on the already-landed
**Rendering Optimization** work (Phase 1-4); what is left now sits on the
open-time, allocation, and draw-call sides.

Source-scan summary (references):

- File open flow: [`src/io/mod.rs`](src/io/mod.rs)
  (`open_path_with_phase`, `load_file`, `purge_corrupt_entities`).
- Post-open UI work: [`src/app/update.rs:90-192`](src/app/update.rs#L90-L192)
  (`FileOpened` handler — xref resolve, second purge, linetype populate, etc.).
- Derived-cache build: [`src/scene/mod.rs:90-212`](src/scene/mod.rs#L90-L212)
  (`build_derived_caches` — rayon-parallel for hatch / image / mesh).
- Wire tessellation: [`src/scene/mod.rs:1328-1402`](src/scene/mod.rs#L1328-L1402)
  (rayon, zoom-adaptive curve tol).
- Block defn cache: [`src/scene/block_cache.rs:238`](src/scene/block_cache.rs#L238)
  (`build_defn` — single-threaded today; nested expansion is topological).
- Pipeline: [`src/scene/pipeline/mod.rs`](src/scene/pipeline/mod.rs)
  (batched hatch, frustum cull, LOD — Phase 1-4 done).

---

## Phase 1 — File Open Time

**Goal:** measurably halve the wall time between "Open" click and "first
frame" for a 50 MB DWG.

### 1.1 Drop the second `purge_corrupt_entities` ✅ DONE

Today `purge_corrupt_entities` runs once on the
[background thread](src/io/mod.rs#L62-L74) and again in the
[`FileOpened` handler after xref resolve](src/app/update.rs#L138-L145).
XREF content already comes from a separate document — fold the purge
**inline** into xref resolve and delete the outer one. On large files
walking `doc.entities()` again is a measurable cost.

**Work:** make `resolve_xrefs` call purge as it merges each xref; remove
[`update.rs:138`](src/app/update.rs#L138).

### 1.2 Move XREF resolution to the background thread

[`resolve_xrefs`](src/app/update.rs#L132-L166) runs on the UI thread today —
large external references freeze the UI. Move it into the
`open_path_with_phase` worker; have `DerivedCaches` carry the resolved-xref
list back. The UI thread only emits log lines.

**New phase tag:** `PHASE_XREFS` (we already have 3 phases; this is the 4th).

### 1.3 Single-pass entity walk (parse + purge + cache planning)

`load_file` → `purge` → `build_derived_caches` does three separate
`entities()` walks. A single pass can produce:

- corrupt-entity detection,
- hatch / image / mesh handle lists,
- AABB accumulation for `world_offset` (currently a separate pass inside
  `compute_world_offset`).  ✅ the world_offset AABB scan is now folded into
  the cache-handle walk (see 2.4); corrupt-detect + hatch/image/mesh planning
  remain a follow-up.

Target: three `O(N)` passes → one.

### 1.4 Memory-mapped file reads (DWG / DXF)

`DwgReader::from_file` / `DxfReader::from_file` likely load the whole file
into RAM with `std::fs::read`. Switching to `memmap2`:

- eliminates the cold-cache read syscall on large files,
- lets the DWG section index be walked on disk (if the acadrust API
  supports it).

**Dependency:** acadrust upstream may need a `from_reader` / `from_slice`
API; add it in our patched fork (`hakanaktt/acadrust`).

### 1.5 Parallelize the acadrust parser (long-term)

acadrust's DWG parser is single-threaded. Section-based parallelism
(header / classes / objects / blocks / entities — independent offsets) is
the biggest unrealized win. Lives in the upstream fork.

**Order:** profile first — is this really the largest slice? Measure with
`puffin`.

### 1.6 Defer raster image decode

[`build_derived_caches`](src/scene/mod.rs#L177-L186) calls
`ImageModel::from_raster_image` for every `RasterImage` entity — pixel
decode happens up front. Wasted if the entity is off-screen. Defer the
decode until **first render** (per-handle lazy `OnceCell`).

### 1.7 File-hash cache (warm re-open)

When re-opening the same file (`(path, mtime, size)` key) keep a disk
snapshot of `CadDocument` + `DerivedCaches` (e.g. `~/.cache/OpenCADStudio/`). Skip
DWG parse entirely. **Win:** most-recently-opened file goes from 1-2 s to
sub-100 ms.

**Risk:** cache invalidation. Stay conservative — load only on exact
`mtime + size` match, otherwise normal parse.

---

## Phase 2 — First-Frame Wire Tessellation

After `FileOpened`, `bump_geometry()` fires; the first frame tessellates
**every** model-space wire. Measurable hitch at ~100 k entities.

### 2.1 Parallelize block-definition build ✅ DONE

[`block_cache::build`](src/scene/block_cache.rs#L127) was single-threaded.
No topological stratification was needed after all: `build_defn` stores
nested INSERTs as by-name references (`LocalSub::Nested`) and never expands
them at build time, so each defn depends only on the read-only `doc` — the
builds are embarrassingly parallel. Now a plain rayon `par_iter().collect()`.
`compute_block_aabbs` is also parallelized: its `defn_aabb_recursive` walk is
read-only over the finished `defns` map and re-walks shared nested defns
(no memo), so the per-name resolves fan out across rayon (read phase), then a
serial phase stores each AABB back.

### 2.2 Incremental wire cache (delta tessellation) ✅ DONE (render path)

A per-entity tessellation memo (`tess_memo`, `Handle → Arc<Vec<WireModel>>`)
sits inside `wires_for_block_culled` on the culled Model render path. It's
keyed by a guard hash of the tessellation parameters (tol / view / anno /
offset / bg / entered viewport); a mismatch (zoom, layout) clears it.
`bump_geometry` clears it (structural change); incremental edits drop just the
changed handle via `mark_entity_dirty` and bump with `bump_geometry_no_blocks`.
So a single-entity edit (line commit, grip commit/cancel, MOVE / ROTATE /
SCALE / MIRROR, COPY, ERASE) re-tessellates only the touched entities and
reuses every other — the per-entity geometry math is skipped for the
unchanged set. `transform_entities` marks the moved handles dirty,
`erase_entities` drops the deleted ones, and `copy_entities`' new handles are
natural memo misses; all use `bump_geometry_no_blocks` so the block cache is
kept too. The hit-test (`view_aabb == None`), paper and
per-viewport paths bypass the memo so their cull params don't thrash it.

Remaining: the assembly still concatenates into one `Vec` and rebuilds the
batched GPU buffer (O(N) clone + upload), and the uncalled hit-test
(`entity_wires_arc`) still rebuilds on edit — driving those to true O(1)
needs per-handle GPU slots, which trade against the single-draw batch (3.3).

`bump_geometry()` invalidates the whole wire cache today
([`scene/mod.rs:650`](src/scene/mod.rs#L650)). Edits usually touch 1-2
entities — re-tessellating the whole doc is waste.

**Fix:** wire cache becomes `HashMap<Handle, (entity_version,
Vec<WireModel>)>`. The editing command bumps the version of the affected
handles; the render path re-tessellates only those, reusing the rest.

Also useful on open: any partial cache (e.g. from block defns) can be
re-used.

**Partial (landed): block cache survives non-block edits.** The
block-definition tessellation cache (`BlockCache::build`, every block defn)
was keyed on `geometry_epoch`, so *any* edit rebuilt all block defns — a
~30 ms baseline wire re-tess turned into a ~400 ms spike on block-heavy
drawings at line-commit / grip-start / grip-release. The cache now keys on a
separate `block_epoch`. `bump_geometry` bumps both (safe default); the
operations that provably can't change a block defn — adding a top-level
entity (`add_entity` for non-Insert/Block) and grip-moving an entity/insert —
call `bump_geometry_no_blocks`, which re-tessellates only the visible wires
(block cache reused). Block-content edits (REFCLOSE, block create, xref,
explode) still go through full `bump_geometry`.

**Partial (landed): grip drag.** Dragging an entity's grip called
`scene.apply_grip` every move, which `bump_geometry`'d → a full model
re-tessellation per move (plus an O(N) clone of all wires for snapping). Now
the first move hides the edited entity from the base (one re-tess) and shows
it as a one-entity overlay preview; subsequent moves only re-tessellate that
one entity (cheap) — the base stays a cache hit. Snapping runs against the
set directly (the edited entity is already hidden, so no clone and no
self-snap). The drag commits on release / Esc: un-hide + one final re-tess.

**Partial (landed): preview / interim overlays.** `set_preview_wires`,
`clear_preview_wire` and `set_interim_wire` used to `bump_geometry`, so every
rubber-band frame of a drawing command re-tessellated the whole model (the
same 30 ms → 400 ms spike, but while drawing). Preview / interim wires are an
overlay appended to the cached base wire set in `build_primitive`, not part of
the tessellation cache — so they no longer bump geometry. The base stays a
cache hit (no re-tess), and the overlay forces only a GPU wire re-upload via
the `has_overlay` content-id path. Snap/hit-test (`entity_wires_arc`) also
stops re-tessellating per preview frame.

### 2.3 Progressive first render

On the first frame emit a **coarse**-tol wire pass (e.g. 4× the normal
tol); refine to full tol on the second frame. The user sees *something*
within 16 ms; detail snaps in smoothly afterwards.

### 2.4 Merge the world-offset scan into the single-pass walk ✅ DONE

[`compute_world_offset`](src/scene/mod.rs#L128) walks the whole MSPACE
AABB when the header is unreliable. That scan should join the single-pass
walk from 1.3 (we are already iterating `entities()`).

---

## Phase 3 — Per-Frame Render Cost

After Phase 1-4 culling/LOD, what's left is **upload bytes** and **draw
call count**.

### 3.1 Camera-only invalidation: don't re-tessellate

The wire cache key today is `(geometry_epoch, camera_generation)`
([`scene/mod.rs:414`](src/scene/mod.rs#L414)). A camera change should not
force re-tessellation — only zoom-adaptive curve-tol changes need
resampling, and only for curve entities (Arc / Spline / Ellipse). Straight
geometry is camera-invariant.

**Practical:** split the wire cache in two:

- `tess_cache[handle] → WireModel` (rebuild only if tol-invariant content
  changed),
- `frame_visible[handle] → bool` (recomputed per `camera_generation`).

**Partial (landed): pan reuse.** The Model-tile wire cache no longer keys on
the exact camera hash. It now keys on `(geometry_epoch, pan_invariant_hash,
tessellated_region)` where `pan_invariant_hash` covers rotation + tol (`wpp`)
but NOT the pan target. A pure pan keeps the epoch + signature and only shifts
the view, so as long as the new visible rect still fits inside the
1.25×-margin region the wires were culled to, the tessellation is reused
outright — no re-tessellation, `tess ms` drops to ~0 on the PERF HUD. Zoom
(changes `wpp`), orbit (changes rotation) and edits (change epoch) rebuild
exactly as before; the cull margin is unchanged, so miss cost is identical
(no zoom regression).

**Partial (landed): selection decoupled.** Picking an entity used to call
`bump_geometry`, invalidating the wire cache and re-tessellating the WHOLE
model just to repaint one entity (30 ms → 400 ms on a large drawing). The
highlight is no longer baked into tessellation: wires are always base-coloured
(`sel` is empty in `wires_for_block_culled`), and the selection highlight is
applied in the GPU xray overlay from the live `selected ∪ hover` set,
recoloured to `WireModel::SELECTED`. Selection / hover now bump a cheap
`selection_generation` instead of `geometry_epoch`, so the overlay refreshes
without any re-tessellation or main-buffer re-upload. `tess ms` stays flat
when selecting.

**Partial (landed): per-frame split.** `build_primitive` ran
`split_face3d_wires` every frame — an O(N) per-wire handle lookup + clone to
separate Face3D wires — even on a pan that reused the tessellation. It's now
memoized by the tile's wire content id, and the non-overlay frame reuses the
`other` Arc with no clone at all. So a pan reuses tessellation, the split, and
the GPU upload; only the uniform + scissors update per frame.

**Partial (landed): hover.** `set_hover_highlight` no longer bumps the geometry
epoch (a full re-tessellation) when the hovered entity is already selected —
the effective highlight set `selected ∪ {hover}` is then unchanged, so the
tessellation output is identical. Hovering over / between selected entities
is now free. The full camera/selection-from-tessellation split is still open
and needs running-app verification (highlight colour is baked into
`WireModel.color` across several tessellation sites).

### 3.2 Persistent GPU buffer pool — diff upload ✅ DONE (wire pan path)

Wire vertex buffers are world-space, so a camera move alone never changes
them — only the `view_proj` uniform (already uploaded per frame). The wire
upload was gated on `(geometry_epoch, camera_generation)`, re-sending every
pan. Now each Model-tile tessellation is stamped with a monotonic content id
(`WIRE_CONTENT_GEN`), reused when a pan reuses the tessellation; the pipeline
holds the resident buffer's id and `upload_wires` is skipped when it matches.
Gate is independent of the camera tick so a preview/interim change still
uploads. Non-tile paths and overlay frames force a fresh id (unchanged
behaviour). Monotonic id avoids the ABA hazard of a raw `Arc` pointer.

Still open: a true `HashMap<Handle, GpuSlot>` per-entity pool for partial
edits (re-upload only the changed slots); this covers the whole-buffer
pan/idle case, the dominant one.

Today every wire GPU buffer is re-uploaded when
[`cached_epoch`](src/scene/pipeline/mod.rs#L101) changes. A persistent
pool — `HashMap<Handle, GpuSlot>` — uploads only the slots that actually
changed. Big win in CAD-edit scenarios.

### 3.3 Single-draw batched wire pipeline (Phase 4-B-style) ✅ DONE

`upload_wires` made one GPU buffer + one draw call per `WireModel` (tens of
thousands on a large drawing). Now it merges maximal runs of *consecutive*
wires sharing scissor + mesh-edge state into one concatenated instance buffer
each (`WireGpu::from_run`), so the existing draw loop issues one draw per run
— a 2D model collapses to a single buffer + single draw. Runs stay
*consecutive* (not globally regrouped) so the sorted draw order is preserved
bit-for-bit; depth bias and alpha blending are unchanged. The `WireInstance`
layout, shader, scissor logic and draw loop are untouched — only the buffer
packing changed. (No iced widget-pipeline limits were hit: the change lives
entirely inside the existing custom wgpu pipeline.)

Every `WireModel` today costs one draw call plus a bind-group swap. Port
the batched hatch pipeline (`hatch_batched_gpu.rs`) to wires:

- pack all wire vertices into one storage buffer,
- per-instance `(color, pattern_id, lw_px, visibility)` in a side buffer,
- vertex shader pulls instance data via `instance_index`,
- a single `pass.draw(0..V, 0..N)` covers everything.

At 100 k wires that collapses thousands of draw calls into one. If iced
0.14's widget-pipeline limits allow, immediate win.

**Follow-up (landed): separate overlay buffer.** Command-preview / interim /
grip-drag wires were appended into the main wire buffer, so every drag frame
re-uploaded the entire (batched) base set. They now ride in their own small
per-frame `gpu_preview_wires` buffer drawn on top in the wire pass, so the
resident base buffer stays untouched during a drag — only the tiny overlay
re-uploads. Combined with the no-re-tess work, a grip/command drag now costs
one small overlay upload per move, nothing else.

### 3.4 Hardware instancing for repeated block inserts

When the same block defn is `INSERT`-ed N times (every door / window in
an architectural drawing) each instance currently renders as its own wire
set. Hardware instancing:

- upload the block defn vertex buffer once,
- one 4×4 transform row per Insert in an instance buffer,
- `pass.draw_indexed(0..V, 0..N_instances)`.

Typical architectural DWGs: 10-100× faster.

### 3.5 Glyph-stroke batching

`tessellate.rs` produces one `WireModel` per glyph stroke today — one text
entity = dozens of models. Cache stroke geometry per font once
(`HashMap<(font, glyph), Vec<Point2>>`), then per-text only a transform
matters.

---

## Phase 4 — Allocation & Memory

### 4.1 Swap `HashMap` for `rustc-hash::FxHashMap` ✅ DONE

`Handle` is an integer wrapper; the default `SipHash` is overkill.
`FxHashMap` gives 20-40 % in hash-heavy sites (block_cache, hatches /
images / meshes, viewport_wire_cache).

### 4.2 Arena (`bumpalo`) for transient wire vertices

Tessellation allocates millions of small `Vec<Vec3>`s. A bump arena —
single allocation, frame-end reset — kills the per-vertex malloc cost.
`bumpalo` plays well with rayon (per-thread arenas).

### 4.3 `SmallVec` for small collections

`Polyline.vertices`, `Hatch.boundary_paths`, glyph-stroke lists are
typically < 8-16 entries. `SmallVec<[T; 8]>` skips the heap on the common
case.

### 4.4 Compact entity-ID representation

`Handle` is 8 bytes. 100 k entities → 800 KB just in keys. Hot handle
HashSet / HashMap usage can be flattened to `Vec<u32>` indices plus a
single `FxHashMap<Handle, u32>` translation table — cache-friendlier.

---

## Phase 5 — Profiling Infrastructure (prerequisite)

Don't start any of the above **without measuring first**.

### 5.1 Add `puffin` or `tracy` spans

- `io::open_path_with_phase` → `parse`, `purge`, `caches` spans.
- `Scene::wires_for_block` → `block_cache`, `tess`, `sort` spans.
- `Pipeline::prepare` → `upload`, `cull`, `draw` spans.

Gate behind `debug_assertions` or a `--features profile` flag.

### 5.2 Open-time breakdown log ✅ DONE

When open completes, push to the command line:

```
Opened "x.dwg" — 84321 entities — parse 1.2s, purge 80ms, caches 340ms, xref 60ms, first frame 210ms
```

Regressions are visible immediately.

### 5.3 Frame-budget HUD ✅ DONE (CPU tess slice)

A CLI `PERF` toggle overlays the cost of the most recent wire
re-tessellation (ms + wire count + geometry epoch) on the active viewport,
anchored top-left. Reads ~0 ms while the wire cache is warm (idle
pan/zoom), so it isolates exactly the work a cache miss costs — the slice
every render-path change (2.2 / 3.1 / 3.3) moves. Timed at the miss paths
in `model_tile_wires_arc` / `paper_sheet_wires_arc` and stored on `Scene`.

Still open: GPU-side `upload` / `draw` / `GPU-wait` spans need wgpu
timestamp queries; the CPU tessellation slice covers the current hot path.

---

## Priority Order

**Phase 5 first** (profiling) — avoids speculation.

Then, measurement-guided:

1. **Phase 1.1 + 1.2** (cheap, low-risk, certain win).
2. **Phase 1.3 + 1.6** (single-pass + lazy image).
3. **Phase 2.2** (incremental wire cache — wins on both edit and open).
4. **Phase 3.1** (camera-only invalidation — users pan/zoom constantly).
5. **Phase 3.3** (batched wire pipeline) and **Phase 3.4** (instancing) —
   biggest render win, highest complexity.
6. **Phase 1.7** (warm cache) — dramatic UX, but invalidation must be
   correct or it creates nasty bugs.
7. **Phase 1.5** (acadrust parallel parse) — hardest, longest-term; only
   worth it if profiling confirms it is the dominant slice.

## Deliberate non-goals (for now)

- **GPU compute culling:** for orthographic 2D CAD the CPU quadtree is
  enough. Already covered by Phase 1-4.
- **Out-of-core entity streaming:** meaningful for 100 MB+ single files;
  typical Open CAD Studio files are not there yet.
- **Multi-frame async tessellation pipeline:** if 2.3 progressive render
  works cleanly, this isn't needed.