Free, open-source 3D sculpting for Linux, macOS and Windows: dense meshes, remeshing, painting, built-in AI (image/text to 3D), Send to Unreal, game-ready LODs, Gaussian splats and OpenUSD.
C++ 3d3d-sculptingcppdear-imguidigital-sculptingfbxgame-developmentgaussian-splattinggltfimage-to-3dmcpopenglopenusdsculptingtext-to-3dunreal-engine
Perf scoreboard: beta.2 baseline on the E-core floor, laptop and Windows 3060 (15M, 100M, 200M at 25/120/240 px)
islandmed9e2557d3bdb0
diff --git a/docs/design/perf-scoreboard.md b/docs/design/perf-scoreboard.mdindex e3879229..ebe0b5ba 100644--- a/docs/design/perf-scoreboard.md+++ b/docs/design/perf-scoreboard.md@@ -26,7 +26,7 @@ figure is raw minus that floor.2. **E-core floor** (the gate today): `SF_HW_SIMULATE=floor SF_GOVERNOR=0 taskset -c 0,1` on this laptop (M1,Asahi Linux): the floor's budgets, GL 3.3, on the two efficiency cores (about half a fast core each). This is thepessimistic CPU bound; a CPU-bound claim counts only if it holds here (perf-moonshots §0 item 6).-3. **Laptop**: the same machine, no simulation, all cores (Apple M1 Pro GPU through Asahi's GL driver).+3. **Laptop**: the same machine, no simulation, all cores (Apple M1 Pro, 32 GB shared, through Asahi's GL driver).4. **Windows RTX 3060** (Ryzen 5600X, 32 GB): `~/Work/.sf/win.sh`, PC idle (win.sh holds the PC lock). A sanitycheck: a result that holds only here hasn't hit its target.@@ -38,15 +38,53 @@ figure is raw minus that floor.- Committed lag: pointer sample to the swap of the frame whose mesh contains that sample's dab (`lag_ms`).Release: pen up to the last dab in the mesh and shaded (`release_done_ms`). Orbit: frames per second over theorbit phase. Memory: peak resident RAM plus GPU buffers.-- Defaults of the build under test (no toggles), load < 2 on the laptop, two interleaved runs per cell; a cell shows- the range of the runs.+- Defaults of the build under test (no toggles), 6 s strokes, each run started at load < 2 on the laptop, at least+ two interleaved runs per cell; a cell shows the median.- Rows from `python3 tools/perf_scoreboard.py TIER results/*.json`.-## Shipping today: main 0.11.0-beta.2 defaults+## Shipping today: main 0.11.0-beta.2 defaults (11-04)-| Mesh | Brush | Tier | Committed lag p50 / p99 | Release | Orbit fps | Memory (peak RAM + GPU) | Commit |-|---|---|---|---|---|---|---|---|-| (filled below as runs complete) | | | | | | | |+Committed lag p50 / p99, then release (pen up to the stroke in the mesh), in ms; median of the runs. Floor iGPU:+no machine yet. E-core and laptop: 3 runs each (perf/scoreboard e30349e sources; the build reports its parent+8d89d61c); Windows: 2 runs on an idle PC (e30349e). Bold: meets the target (p50 within 16 ms of the spacing floor,+p99 < 33 ms, release < 100 ms).++| Mesh | Brush | Floor iGPU | E-core floor (the gate) | Laptop | Windows RTX 3060 |+|---|---|---|---|---|---|+| 15M | 25 px | - | 19 / 174, 156 | **14** / 109, **51** | **8** / 59, **45** |+| 15M | 120 px | - | 550 / 884, 608 | 177 / 259, **81** | 170 / 255, **73** |+| 15M | 240 px | - | 552 / 2036, 3355 | 405 / 577, 230 | 447 / 737, 249 |+| 100M micro | 25 px | - | 24 / 79, 131 | **14** / **28**, **23** | **6** / **24**, **33** |+| 100M micro | 120 px | - | 240 / 390, 521 | 113 / 186, **68** | 102 / 183, 107 |+| 100M micro | 240 px | - | 829 / 1254, 2747 | 286 / 479, 338 | 298 / 437, 404 |+| 200M micro | 25 px | - | 25 / 90, 109 | **18** / 60, **42** | **6** / 46, **49** |+| 200M micro | 120 px | - | 466 / 582, 960 | 134 / 225, 171 | 122 / 215, 312 |+| 200M micro | 240 px | - | 1498 / 1735, 4675 | 319 / 511, 452 | 440 / 565, 827 |++Orbit (fps, the orbit phase after the strokes) and memory (peak RAM + GPU buffers):++| Mesh | E-core floor | Laptop | Windows RTX 3060 | Memory (laptop / Windows) |+|---|---|---|---|---|+| 15M | **90** | **182** | **319** | 2.3 GB + 233 MB / 2.4 GB + 211 MB |+| 100M micro | 13 | 32 | 47 | **0.8 GB + 330 MB** / 0.9 GB + 330 MB |+| 200M micro | 10 | 27 | 32 | **1.3 GB + 640 MB** / 1.4 GB + 640 MB |++What it says:++- **Small brushes keep up at p50, not at p99.** 25 px p50 is 6-25 ms on every tier (the target or within a few ms+ of it on the E-core floor); p99 meets 33 ms only on 100M (laptop, 3060) and is 46-174 ms elsewhere.+- **Big brushes fail everywhere, the 3060 included.** 120 px is 100-550 ms behind the pen, 240 px 290-1500 ms.+ On the micro scenes the E-core floor places 9-22 dabs in a 6 s stroke at 120-240 px (the brush can't keep up,+ so dabs merge): the stroke lands, but late and coarse. This is what N2 (micro levels) and Fast Big Brushes+ (15M) are for.+- **Orbit after sculpting on micro levels is 10-47 fps** (it opens and orbits at 60+ fps before any stroke, per+ docs/design/core-200m.md §8): the shading catch-up after the strokes runs during the orbit. 15M orbits at 90+.+- **200M fits a 16 GB integrated-GPU laptop:** 1.3 GB RAM + 640 MB GPU.+- **Noise:** the laptop is shared with other sessions (load 4-13 during runs; the start waits for load < 2). One+ E-core round of three was 2-7x worse on the micro scenes and is outvoted by the median; ranges per cell are in+ the raw rows (`tools/perf_scoreboard.py`). The Windows runs agree within 10%.+- **Windows harness note:** the micro scenes have to start through `cmd /c` (win.sh with any `SFENV`); a direct+ launch from the scheduled task exits silently before writing the scene (not yet understood).## Candidates for beta.3@@ -56,4 +94,4 @@ Each candidate is measured against the row above on the same tier, interleaved,|---|---|---|| perf/n2-wavelet (n2stroke) | N2 wavelet strokes on micro levels, default on; GPU shading (fragment shader) | in progress || perf/beta-boost-2 (betaboost) | Fast Big Brushes with fused batches capped by time; NVIDIA first-dab stall | in progress |-| research/core-200m (core200m) | N1 screen-space strokes under 1 ms a dab | in progress |+| research/core-200m (core200m) | N1 screen-space strokes under 1 ms a dab | 6d410e1f: 0.58-0.73 ms GPU per 240 px dab on the laptop (was 4.2-4.8), live on the floor 9.6 vs the ghost's 8.3 ms GPU; research, off (SF_SCREEN_STROKE=1) |diff --git a/tools/perf_scoreboard.py b/tools/perf_scoreboard.pyindex 79eb0b8d..abc96463 100644--- a/tools/perf_scoreboard.py+++ b/tools/perf_scoreboard.py@@ -7,7 +7,7 @@Each result is a `sculptforge --benchmark <scene> --benchmark-out <file>` JSON. Prints one markdown row per brushphase (committed lag p50 / p99, release, orbit fps, memory); with several results for the same scene and phase-(interleaved repeats) the cells show the range."""+(interleaved repeats) the cells show the median and the range."""import jsonimport sys@@ -15,11 +15,14 @@ PHASES = [("clay small", "25 px"), ("clay large", "120 px"), ("240 px brush", "2def rng(values, fmt="{:.0f}"):- values = [v for v in values if v is not None]+ """The value, or with repeats the median (and the range when they differ)."""+ values = sorted(v for v in values if v is not None)if not values:return "-"- lo, hi = min(values), max(values)- return fmt.format(lo) if fmt.format(lo) == fmt.format(hi) else fmt.format(lo) + "-" + fmt.format(hi)+ lo, hi, mid = values[0], values[-1], values[len(values) // 2]+ if fmt.format(lo) == fmt.format(hi):+ return fmt.format(lo)+ return fmt.format(mid) + " (" + fmt.format(lo) + "-" + fmt.format(hi) + ")"def main():