Files
prometeu-runtime/crates/tools/pbxgen-stress/examples/MEASUREMENTS.md
T
bquarkz c2bf8d055c feat(stress): add a deterministic particle demo with sequenced music
Move Storm Eagle playback and source assets to Stress Console and keep Dummy Boy silent. Exercise VM math, seeded Random and persistent typed particles with a projected tunnel, bounded 512-sprite heavy workload, reproducible replays and measured costs.
2026-11-20 22:15:47 +00:00

4.5 KiB
Raw Blame History

Stress Console local measurement

Executed on 2026-11-20: Apple M2 Pro, macOS/Darwin 27.2.0, aarch64, rustc 1.98.1 (48a229cea 2026-09-01), Cargo release profile, VM resource-diagnostics enabled. The host was not isolated and its power/thermal state was not controlled. These are elapsed-time observations from one run.

Reproduce from the runtime root:

cargo run -p pbxgen-stress --release --example demo_measure -- 120 /tmp/stress-measure

Seed: 1347571533. Particle position-travel multiplier: 1.5. Each preset loads independently, executes its first frame, then measures 120 steady logical frames. All measured frames compose the native 640x360 image and process 800 stereo audio frames (48 kHz). Initialization is excluded from steady distributions. Percentiles are nearest-rank over the 120 samples; they do not estimate a guaranteed upper bound.

Preset VM only mean ms Headless mean ms p95 ms Max ms Mean virtual cycles Host ticks/logical frame
light 1.982 2.683 2.853 2.996 84701.2 1
normal 4.154 4.907 5.056 5.184 178972.8 1
heavy 19.183 19.685 20.553 21.700 774621.4 4

VM-only uses the same PBX on the real VM and SpriteController; other native calls are counted. Headless uses actual firmware, assets, composition and audio. Virtual cycles are machine accounting, not CPU cycles or milliseconds. The heavy preset spans four 250,000-cycle host slices per logical frame; light and normal use one. advance_frame completes all required slices, so these timings are not physical presentation cadence or a promise of 60 FPS in another host.

Preset Host boot ms First frame ms GuestBytes used Live handles/objects AuxiliaryBytes used TemporaryBytes peak Sprites accepted
light 1.730 3.037 22640 66 8680 3840 192
normal 3.421 5.213 44656 130 16360 7680 256
heavy 3.278 20.683 132720 386 31720 15360 512

Each preset recorded zero sprite overflows and zero guest composer errors. Music play was submitted once. The two-second observation reached eight active voices; the separate 52.4-second audio regression reached ten, with no clipped samples. Processing 800 audio frames averaged 0.123–0.128 ms in this run.

The guest retains exactly N Particle structs plus one array and one RNG object. Rebirth reuses those objects. Quota accounting excludes process RSS, framebuffers, resident banks, renderer caches and allocator overhead. The musical bank has 60,358 PCM bytes and canonical decoded size 112,224 bytes; those are independent of the workload preset. Boot/first-frame times include their respective host work and are not isolated codec-decoding measurements.

The current default is heavy: 384 particles, four real simulation passes and 128 tunnel markers. It deliberately selects the largest existing numerical workload. Normal (128 particles/two passes) and light remain explicit options for comparison. The current run includes the 1.5 position-travel multiplier and the larger heavy population. Relative to the previous 256-particle heavy workload, sprite admission rises from 384 to 512, mean virtual cycles from about 531,444 to 774,621 and host slices from three to four. Its observed mean headless cost increased from 13.649 to 19.685 ms. These observations do not provide a handheld frame-rate guarantee.

The example saves report.json, generated cartridges and PPM captures. Diagnostic state hashes after the initial frame plus 120 measured frames were:

Preset Particle-state hash Native frame hash
light b53cd663d1741a24 87f646410e9c3ade
normal 16b062d046d8efc9 c33ceb1e514ff1a8
heavy f934f3e61ae1be77 6c3302f4a16903bc

The tests additionally pin six-frame numeric and visual baselines and execute them in debug and release. A separate probe checks actual guest displacement against the updated velocity, establishing the 1.5 multiplier independently of the replay hashes. The native desktop integration test exercises actual cartridge music through AudioCallback without a physical output device.

No physical callback deadlines, underruns, GPU presentation, VSync, battery consumption, thermal throttling or affordable handheld hardware were measured. The headless harness drives audio explicitly; it cannot observe a device underrun. Use dedicated subsystem tests for numeric edge cases, GC churn, storage/input, crossfade, SFX and full 32-voice load.