Module warp_eval 

Per-vertex warp evaluation.

Each frame the warp mesh’s cols × rows vertices need their UVs warped according to the preset’s per-vertex (Milkdrop “per_pixel”) equations and the MilkDrop 2 zoom/rot/stretch/translate/warp formula.

Hot path strategy

WarpExecutor precompiles each per-vertex equation into an evalexpr::Node once at preset load and then reuses one scratch MilkContext across all vertices of a frame. This replaces the previous “clone evaluator state per vertex” approach, which allocated a fresh HashMapContext (≈60 vars + math-function map) for every one of the cols × rows vertices — 768 allocations per frame at the default 32×24 mesh, up to 18 432 at 192×96.

Per-vertex independence (each vertex sees the same per-frame state on entry, regardless of what previous vertices wrote) is preserved by snapshotting the motion outputs (zoom, zoomexp, rot, cx, cy, dx, dy, sx, sy, warp) and the q1..q32 channels at the start of the frame and restoring them before each vertex eval.

Structs

PerVertexMotion 🔒

Per-vertex motion outputs read from the evaluator after running per-vertex equations.

WarpExecutor

Pre-compiled per-vertex equation executor.

Constants

PARALLEL_EQ_THRESHOLD 🔒

Minimum compiled-equation count before we hand the per-vertex loop to rayon. Empirically: presets with ~3 cheap equations stay faster sequential (warp-exec time is dominated by per-thread MilkContext clone + work-stealing overhead at ~1 ms/frame baseline cost), while presets with 10+ equations see a 2-3× wall-clock win on a 12-thread box. Picked 8 as the breakpoint after measuring three reference presets (3 / 26 / 35 equations) on the bench_render tool.

Q_CHANNEL_COUNT 🔒

Number of q* channels we snapshot/restore between vertices. MilkDrop 2 exposes q1..q32; the underlying MilkContext reserves 64 slots but only the first 32 are user-facing.

Functions

read_motion 🔒

read_q_snapshot 🔒

warp_uv_md2 🔒

MilkDrop 2 warp UV formula.

write_motion 🔒

write_q_snapshot 🔒