DOCUMENTATION A WORLD IN THE MAKINGFIELD NOTES / MILESTONES
A few documented steps toward a more capable terrain pipeline.
CURRENT DIRECTION
A DAG is a graph executed in dependency order. EONFORM’s resident DAG already composes supported stages on the GPU, but it does not automatically convert C++ nodes into shaders. The next step is reusable typed operators shared by multiple nodes, not a special shader for every graph shape.
The resident graph currently admits Windows D3D SM6; older standalone GPU paths have native Metal test evidence, but those results do not validate the current resident graph on Metal. Unsupported graphs can fall back to CPU. GPU source execution does not mean simulation, mesh construction and commit all run on the GPU.
Full-reference fields can be expensive: one 8129-square float field is about 252 MiB. The default resident payload budget is 256 MiB and the final-field cache is capped at 64 MiB; these are limits, not measured peak memory. Selected regions do not necessarily mean regional-only simulation.
EXPLORE NODE COVERAGE ↗NATIVE MESH TERRAIN / PRODUCT ROADMAP
A separate, planned workstream explores editable caves, overhangs, undercuts, natural arches and sea caves using UE 5.8 Mesh Terrain geometry rather than only heightfield sculpting. These tools could become reusable, artist-directed landform assets, driven by geology and hydrology where appropriate.
Status: feasibility and proof of native 3D geometry are not yet verified. Existing river and lake bed carving lowers terrain height; it does not demonstrate a cave roof or arbitrary volumetric topology. Next gates are native API feasibility, a persisted real 3D cut, editable regeneration, region seams and clean-install testing.
NEXT MILESTONES
Record node path, supported settings, platform, individual test outcomes and remaining gates. A scalar probe or registered test is not complete GPU-node validation.
The November 13 Windows offscreen baseline completed 418 tests: 404 passed, 8 failed, and 6 included skipped checks. Investigate the failures before treating the baseline as green; separate opt-in runs are recorded on their individual test pages.
EXPLORE TESTS ↗One shared shader dispatcher now serves existing typed modifier payloads in procedural, primitive and Combine paths, including resident modifiers. This removes duplicate code; it does not add nodes, fuse resident stages or demonstrate a speedup. A broader typed operator representation is still ahead.
Test branching, requested-output readback, physical X/Y spacing, rectangular worlds, selected windows and seams. Sampling and neighbourhood operations need explicit boundaries and halos; reductions and simulations need their own process contracts.
Test native Metal before admitting the resident graph there. Compare CPU and GPU with matching graphs, including cold and warm time, GPU timestamps, readback, mesh commit and measured peak memory. Keep existing routes until alternatives pass.
RECORDED EVIDENCE
A non-flat Hemisphere and SoftClip graph matched CPU output and exact selected-window seams on a 2400 × 1600 m world at 505 × 329, 1009 × 657 and 2017 × 1313 samples. Common world-coordinate heights matched across those resolutions. This is one fixture, not universal GPU support; no speedup was measured.
RESOLUTION PARITY TEST ↗A typed GPU DAG, full-reference fields and selected-window readback were exercised on Windows D3D12. The recorded nine-stage fixtures do not imply universal node support or production readiness.
GRAPH PARITY TEST ↗The handoff records resident-graph wall times and one final readback. The 4096-square result used two stages, not nine; timings are not GPU timestamps or measured peak VRAM.
BENCHMARK RECORD ↗Pointwise Adjust and standalone Equalize have documented D3D12 fixture results. Global modes, strict admission, other fields and platforms must be assessed separately.
ADJUST NODE PROGRESS ↗Windows D3D12 tests passed 54 AutoClip, 72 DotNoise and 36 Transpose executions or window checks on an RX 7900 XT. These fixtures do not certify every setting, resolution or platform.
AUTOCLIP TEST ↗