Speed up QuadSurface::gen() warps by hoisting the separable x axis

[robert-kofler42]

What

QuadSurface::gen()'s four warp helpers (warpBilinearReplicateVec3f, warpNearestConstU8, warpNearestConstVec3f, warpBilinearConstVec3f) implement an axis-aligned scale+translate resample: src_x = ox + dx*sx, src_y = oy + dy*sy.

The mapping is separable: the source-x sample (x0/x1/wx for bilinear, the rounded source index for nearest) depends only on dx, not dy. But the inner loop recomputed it for every one of the dh rows — including an std::lround per pixel in the nearest variants. This hoists the x-axis computation into a per-column array computed once per call and reused across all rows (and hoists iwy = 1 - wy to per-row).

Correctness — bit-identical

Bit-identical for all finite warp parameters (the only kind gen() produces): the same float expressions are evaluated in the same order, just memoized.

Honest edge note: on non-finite source coordinates (not reachable from gen(), whose warp params are always finite), the bilinear-const variant now writes the border value instead of the original's out-of-bounds read (int(NaN) → row[INT_MIN]) — i.e. strictly safer, not a regression.

Verified with:

• an FNV-1a checksum over all output float bits across a sweep of scales/offsets — identical before/after (18b53e9196982b08);
• the existing render/quadsurface unit tests (16 pass).

Performance

Profiled with perf on a synthetic 4000×4000 surface: the warp helpers were ~53% of gen() self-time, __lroundf alone 8.9%.

Benchmark (gen() with normals, median of 5, 32-core machine, RelWithDebInfo):

scenario	before	after	speedup
interactive, 512×512 tile (one gen() call, internally parallel)	2.54 ms	1.43 ms	1.78×
interactive, 1024×1024 tile	10.13 ms	6.09 ms	1.66×
tiled batch render (outer OMP over tiles)	0.90 ms	0.90 ms	~0 (memory-bandwidth bound)

The win lands on the compute-bound interactive render path (the VC3D viewer issues one gen() per view, which parallelizes internally) — i.e. pan/zoom latency. The tiled batch renderer (vc_render_tifxyz) is memory-bandwidth bound and unchanged, but does the same total work with fewer instructions.

Notes

• Independent of Make QuadSurface::gen() safe for concurrent rendering #1075 (touches different functions in the same file); applies cleanly on main.
• The per-column arrays are small function-local std::vectors filled single-threaded before the row loop and read-only inside it (safe under the warps' internal OpenMP).

[vercel]

Someone is attempting to deploy a commit to the scroll Team on Vercel.

A member of the Team first needs to authorize it.