# Workshop performance, 2026-10-30

The flat-six workshop now runs within the real-time audio budget. The measured
simulation cost fell from 104.44 to 8.79 microseconds per sample (11.9 times faster).
At 44,100 Hz the available budget is 22.68 microseconds per sample.

## Cause and changes

The workshop used a different assembly path from the native scenarios. It split
fluid sections at every visual pipe knot, inserted 5 mm inlet stubs, required at
least four fluid substeps per sample, and repeated mechanical/control work in
each substep. Flat6 consequently had a 1.78 mm cell and averaged 12.35 substeps
per sample. Its native scenario had a 34.20 mm minimum cell and about one substep.

- Mesh each pipe by total arclength and its diameter profile. Visual knots remain
  in the editor and renderer; exact frustum averages preserve gas volume.
- Advance shaft, ECU, cams, combustion, carburetors and pump displacement at the
  audio sample rate, matching native engine scheduling. Keep CFL stepping for
  fluid transport and pressure-operated check valves. Compile active part/gate
  lists and cache fixed restriction areas.
- Place the importer's metering assembly farther upstream within its existing
  inlet segment, allowing a sensible outlet stub. The inlet endpoints, route
  length, diameter and total gas volume stay the same. Existing custom pipes are
  never lengthened automatically; actual short pipes still constrain CFL steps.
- Cache local pipe vertices, normals and indices, shared primitive meshes, and
  the part-picker icon atlas. Transforms and pressure updates reuse geometry;
  route/diameter edits rebuild the affected pipe.

Untouched stock imports in `machines/` were compared against the previous
importer's canonical output before replacement. User files, including
`paralell-twin.machine.toml`, were preserved.

## Simulation results

Windows GCC RelWithDebInfo, Rusanov, preset density 20 cells/m, 44,100 Hz,
512 warm-up samples and 32,768 measured samples, motor-driven at 3,000 RPM,
35% pedal. Timings include audio synthesis but exclude editor rendering and
snapshot collection. Paired native/workshop measurements use the same process;
before/after runs have normal timing variation.

| Preset | Workshop before, us/sample | Workshop after | Native after | Workshop cells before / after | Substeps before / after |
| --- | ---: | ---: | ---: | ---: | ---: |
| Single 450 | 12.61 | 3.55 | 2.89 | 56 / 44 | 4.00 / 1.00 |
| Twin | 16.96 | 6.89 | 6.11 | 67 / 58 | 4.00 / 1.72 |
| Flat6 | 104.44 | 8.79 | 8.43 | 161 / 98 | 12.35 / 1.03 |
| V8 H-pipe | 67.51 | 16.01 | 16.56 | 318 / 175 | 4.09 / 1.00 |

At 6,000 RPM and 70% pedal, flat6 measured 9.31 us/sample and V8 16.11.
All benchmark cases reported zero positivity repairs and finite audio.
See [baseline](workshop-before.txt), [after](workshop-after.csv) and
[higher-load results](workshop-high-load.csv).

These are performance comparisons, not claims of identical combustion or sound.
The workshop has explicit metering-body storage, connections and editable
calibration; the native scenarios assemble that hardware differently. Arbitrary
large or finely meshed custom machines can still exceed the audio budget.

## Rendering and playback

Hidden 1440 x 900 OpenGL window, 120 static frames, `glFinish`, no vsync or
simulation: flat6 fell from 1.40 to 0.70 ms/frame; V8 from 1.77 to 0.96 ms/frame.
This measures CPU/GPU render cost, not the displayed frame rate.
See [before](workshop-render-before.csv) and [after](workshop-render-after.csv).

The combined workshop/audio check ran at 60 display frames/s with a held
3,000 RPM shaft, 16 warm-up frames and 120 measured frames per machine:

| Preset | Wall seconds | Simulated seconds | Audio underruns | Buffered samples |
| --- | ---: | ---: | ---: | ---: |
| Flat6 | 2.006 | 2.020 | 0 | 4,096 |
| V8 H-pipe | 2.071 | 2.082 | 0 | 4,096 |

Simulation time includes producer buffering and snapshot timing, so a small
offset from wall time is expected. This is a short playback check, not a long
session guarantee. Raw output: [playback](workshop-playback.csv).

## Validation and reproduction

All 22 CTest cases passed. Machine tests cover import length/volume preservation,
route-knot-independent meshes, a real 5 mm pipe retaining its CFL constraint,
explicit fuel sources, cold starting and ECU cutoff. Pump pressure agreed within
1% between 44.1 and 88.2 kHz. Builder checks cover geometry-cache reuse and
invalidation, pressure coloring, transforms, deletion/ID reuse and the icon atlas.
Editor preview images were visually inspected.

From the repository root after `cmake --build --preset windows`:

```powershell
ctest --preset windows -j 8 --output-on-failure
.\out\build\windows-gcc\fluidsim_machine_benchmark.exe 32768 3000 0.35
.\out\build\windows-gcc\fluidsim_machine_benchmark.exe 32768 6000 0.7
.\out\build\windows-gcc\fluidsim_machine_render_benchmark.exe 120
.\out\build\windows-gcc\fluidsim_machine_render_benchmark.exe 120 --play
.\out\build\windows-gcc\fluidsim_machine_builder_tests.exe --preview out/workshop-performance-preview
```