Terrain DEM — HEAVY benchmark results
Bench: benches/heavy_terrain.rs (criterion). Run:
export CARGO_TARGET_DIR=/mnt/work4t/cargo-target/facett-map3d-flight
cargo bench -p facett-map3d-flightMachine (HEAVY — full cores + max memory)
- CPU: AMD Ryzen 9 7900, 12 cores / 24 threads
- Memory: 64 GB (64 955 548 kB total)
- Build target dir:
/mnt/work4t(fast PCIe4 disk, per disk-layout law) - Profile:
cargo benchrelease (opt-level 3) - Note: the CPU terrain pipeline is currently single-threaded; the run had the full 24-thread / 64 GB machine to itself. A rayon fan-out over grid rows / sample batches is an obvious future win (the mesh build is embarrassingly parallel).
Tile size
Default grid 1024 × 1024 vertices = 1 048 576 vertices / 2 093 058 triangles per mesh build. .hgt load bench uses a real 1-arc-second SRTM tile: 3601² = 12 967 201 samples. Sampling benches use 1 000 000 query points. Override the grid with MAP3D_FLIGHT_BENCH_GRID.
Results (median)
| Bench | Time | Throughput | What |
hgt_load/srtm_1arcsec_3601 |
36.8 ms | 352 Melem/s | parse a full 1-arc-sec SRTM .hgt (3601², big-endian i16) |
dem_sample/bilinear_1M |
18.9 ms | 53.0 Melem/s | bilinear elevation_at over 1 M points on a real DEM grid |
procedural_sample/fbm_1M |
64.3 ms | 15.6 Melem/s | synthetic-fallback (4-octave) sampling over 1 M points |
build_mesh_dem/dem_1024x1024 |
45.7 ms | 45.8 Mtri/s | full terrain mesh (verts + normals + indices) from a DEM |
build_mesh_synth/synth_1024x1024 |
67.1 ms | 31.2 Mtri/s | same build over the procedural fallback |
Reading it: a ~2.1 M-triangle Sweden terrain patch builds from a real DEM in ~46 ms single-threaded (≈22 full rebuilds/s; incremental/streamed builds are far cheaper), DEM bilinear sampling runs at 53 M queries/s (the elevation_at the drape pass will hammer per footprint), and a full 1-arc-second SRTM tile parses at 352 M samples/s (~37 ms for a 13 M-sample tile). The procedural fallback is ~4× the sampling cost of a DEM lookup (four trig octaves vs a 4-tap bilinear) but never touches disk — the deterministic CI/no-tile path.