Struct WarpExecutor 

pub struct WarpExecutor {
    sources: Vec<String>,
    compiled: CompiledBlock,
}

Pre-compiled per-vertex equation executor.

One instance lives for the lifetime of the engine; set_equations is called whenever a new preset is loaded.

Fields

sources: Vec<String>

Source equations as last set, retained so set_equations can no-op when called with an unchanged set (common case during steady-state rendering).

compiled: CompiledBlock

Compiled equations packaged in a CompiledBlock — auto-lowers to bytecode when every node is supported (~80 % of corpus per_pixel blocks), falls back to evalexpr Node walk otherwise. Per-vertex eval is the second-densest CPU phase after wave per_point: 768 vertices × N equations every frame, often in rayon. Skipping evalexpr’s recursive operator dispatch on each op trims a measurable slice of warp_compute. May contain fewer nodes than sources if some equations failed to compile — failures are logged and silently dropped, matching MD2’s “never crash on a bad preset” stance.

Implementations

impl WarpExecutor

pub fn new() -> Self

pub fn set_equations(&mut self, evaluator: &mut MilkEvaluator, eqs: &[String])

Replace the compiled equation set. No-op when eqs matches the previously compiled sources.

evaluator is the per-frame evaluator; it is borrowed mutably so the preprocess pass can register any newly seen variable names on its context (preserving the same auto-init semantics as MilkEvaluator::eval).

pub fn compute( &mut self, mesh: &WarpMesh, evaluator: &MilkEvaluator, time: f32, ) -> Vec<WarpVertex>

Compute one WarpVertex per mesh vertex, ready to upload to the GPU.

Pre-condition: evaluator has just finished its per-frame phase so motion outputs, audio levels, and q* channels are up to date.

Parallelism

When the preset has any per-vertex equations, the inner loop fans out across rayon’s global thread pool: each worker gets one cloned MilkContext (via map_init) and chews through its slice of the 768-vertex mesh. Per-MD2-spec, per-vertex equations have no cross-vertex carryover, so the parallel order is invisible from the outside. The init clone is ~1 µs (HashMap with per-frame state plus math-function pointers), so amortised across hundreds of vertices per worker it’s free.

Empty-equation presets stay on the simple sequential path: nothing to evaluate, so threading would only add overhead.

pub fn compiled_count(&self) -> usize

Number of equations currently compiled. Test/debug helper.

Trait Implementations

impl Default for WarpExecutor

fn default() -> Self

Returns the “default value” for a type.