RustNotMC/benchmarks/entity-model-document-2026-10-14

Entity model document loading

The three startup batches (boats/chest boats, minecarts, armor stands/player armor) now share one parsed document. For this fixture, that replaces 43 full document reads/parses (boat discovery plus 42 selected layers) with one. Selected models still produce independent Part trees; the document is dropped after the batches are built. Other entity loaders retain their existing behavior.

Reproduce

From the repository root:

python3 benchmarks/entity-model-document-2026-10-14/generate_fixture.py /tmp/horizon-entity-model-document-fixture.json
ENTITY_MODEL_BENCH_DOCUMENT=/tmp/horizon-entity-model-document-fixture.json \
  CARGO_TARGET_DIR=/home/bob/Documents/horizon/target \
  cargo test -p horizon-client --profile fast --bin horizon-client \
  entity_model_document_benchmark -- --ignored --nocapture

CARGO_TARGET_DIR is optional and can be changed for another machine. Supply a prepared entity-models.json through ENTITY_MODEL_BENCH_DOCUMENT to benchmark that document instead; the harness copies only that file to a temporary root. ENTITY_MODEL_BENCH_REPETITIONS controls repetitions (default 15).

The opt-in Rust test calls the actual three optimized production loaders and compares every selected part name, pose, polygon normal/vertex and child against the previous algorithm using the compatible per-model load_model API. Timings include reading, parsing, constructing, normalizing into a common result map, and dropping the result. Each variant gets three warmups and 15 measured calls, with order alternating. black_box consumes inputs/results. No other builds or optimization benchmarks ran during the measured calls.

Input and limitations

Prepared resources were unavailable in the repository and local Horizon asset directories. This is a synthetic 2,473,695-byte document with 458 model layers, each containing 12 posed parts with six exported quads and normals. It selects 22 boat/chest-boat layers, six minecart layers and 14 armor/player layers (42 total). One wood name is synthetic (test_wood); boat discovery accepts arbitrary wood names. The extra layers model the cost of parsing a full production-size document for each selected model.

File data is warm in the OS page cache. Equipment files are absent, so equipment texture loading is excluded. This measures the entity-model batch only and does not establish overall startup, frame time or FPS improvement. Shapes/layer counts in real prepared documents can produce different absolute timings.

Machine: AMD Ryzen 5 5600X (6 cores / 12 threads), Linux x86_64, Rust 1.99.0 (b940084d7, LLVM 23.1.1), Cargo fast profile (opt-level 3, 16 codegen units, no LTO). Baseline algorithm corresponds to main 73a019967e9665c6b96332988c13cb66679d8477.

Results

Variant	Median batch time	Full document reads/parses
Previous per-model algorithm	1,434.974 ms	43
Shared document	33.420 ms	1

The measured batch is 42.94× faster on this synthetic input. All 15 samples for both variants, warmup count, document size and selected model count are in results.json.

Validation

cargo fmt --all -- --check
CARGO_TARGET_DIR=/home/bob/Documents/horizon/target \
  cargo test -p horizon-client --profile fast --bin horizon-client \
  entity_model_loading_tests -- --nocapture
CARGO_TARGET_DIR=/home/bob/Documents/horizon/target \
  cargo clippy -p horizon-client --profile fast --lib --bin horizon-client --tests -- -D warnings

The regular tests cover full/sparse documents, unrelated layers, a missing file, invalid JSON, non-object JSON, missing model keys, independent model trees, and equipment loading when the document is unavailable.