Parallelize + hoist invariants in QuadSurface resample/rotate warps (bit-identical)

[robert-kofler42]

What

The two linear point-warp helpers in QuadSurface.cpp were fully serial and recomputed loop-invariant work per pixel. Both are now parallelized over output rows with their invariants hoisted, bit-identically:

• resamplePointsLinearPreservingInvalids — the source-x sample (x0/x1/fx) depends only on dstX, so precompute it once per column instead of per pixel (same separable-axis technique as the gen() warp helpers in Speed up QuadSurface::gen() warps by hoisting the separable x axis #1076), then run the row loop under OpenMP.
• warpAffinePointsLinearPreservingInvalids (rotate path) — the affine is not x-separable (srcX depends on both dstX and dstY), so only the six dstToSrc.at<double>() coefficients are lifted into locals (the compiler reloaded them every pixel because it can't prove the cv::Mat doesn't alias the store) and the row loop is parallelized. The explicit m*dstX + m*dstY + m form is kept to preserve float rounding.

Correctness — bit-identical

Same float ops in the same order, only memoized / CSE'd; rows are independent so OpenMP doesn't reorder any accumulation. Verified with an FNV-1a checksum over all output floats across a sweep of resample factors {0.5, 0.37, 2.0} and rotation angles {30, 90, 12.5, −47}° — identical before/after (9b62b8c86523bfb0). Existing quadsurface/render tests (10) pass.

Performance

End-to-end timing of resample() / rotate() on a 1500×1500 surface (8 threads, RelWithDebInfo). These include the fixed surface-rebuild/clone overhead, so the warp-only speedup is larger:

op	before	after	speedup
resample()	38.8 ms	31.6 ms	−19%
rotate()	95.6 ms	63.8 ms	−33%

These are user-triggered (resample/rotate/ABF-regrid), off the per-frame gen() hot path, so the absolute wins are modest — but free given they're bit-identical. Independent of #1075/#1076; applies cleanly on main.

[vercel]

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

A member of the Team first needs to authorize it.