Alex Jurkiewicz 365402bbc7 perf(map-generator): major CPU and memory optimisations (#3860) ...

## Description:

Five performance improvements to the map generator, measured on three
maps of increasing size. End-to-end time on `world` improved ~15×, heap
allocations ~19×.

| Map | Before | After | Speedup |
|-----|--------|-------|---------|
| bosphorusstraits (~612K tiles) | 578ms / 594MB | 45ms / 134MB | 13× /
4.4× |
| world (~2M tiles) | 2333ms / 2128MB | 150ms / 553MB | 15× / 3.8× |
| giantworldmap (~8M tiles) | 10701ms / 9300MB | 635ms / 2509MB | 17× /
3.7× |

Changes (one commit each):
- **`--workers` flag**: bounds concurrent map processing to limit peak
memory
- **Flat `[]bool` visited sets**: replaced `map[string]bool` keyed by
`fmt.Sprintf` with flat `[]bool` indexed `x*height+y` — the dominant
cost
- **`neighborCoords` with stack buffer**: eliminates per-call slice
allocation for neighbour lookups
- **`Terrain` struct 24→16 bytes**: field reorder + `uint8` type for
`TerrainType`
- **Nil buffers early**: releases image/terrain arrays as soon as
they're no longer needed
- **BFS mark-visited on push**: each tile enters the queue once instead
of up to 4×, halving queue memory


also fixes a bug (according to Claude):

Here's the bug: createMiniMap downscales by averaging/sampling 2x2
blocks, copying field values across — including Ocean=true from the
parent scale. When a single connected ocean at 1x splits into multiple
disconnected bodies at 4x (because narrow water channels disappear when
you halve resolution), those smaller fragments still carry Ocean=true
from the carryover. The 4x processWater call picks the new largest
fragment and sets it to Ocean=true, but never clears the others — so
multiple disconnected bodies end up flagged as Ocean.

This PR's fix: before the new BFS pass, zero out every Ocean flag, so
only the truly-largest body at the current scale ends up marked.

It's incidental to the perf work but it's a real semantic change — the
on-disk .bin files will differ from main on any map where ocean splits
across downscaling. The PR doesn't mention it, which is why I flagged
it.

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

2026-05-06 14:39:59 -06:00

4.8 MiB

Executable File

Raw History

View Raw