Struct SqrtEqLayers 

pub struct SqrtEqLayers {
    pub low: EqEvalLayers<EF>,
    pub high: EqEvalLayers<EF>,
}

Square-root decomposition using pre-computed eq evaluation layers.

Instead of folding columns on GPU, we pre-compute all layers of eq evaluations and simply drop the highest layer each round (no GPU work needed).

Fields

low: EqEvalLayers<EF>

Layers for xi[(n + 1) / 2..]. Layers insert xi_i from the front, so low contains the latter half of xi.

high: EqEvalLayers<EF>

Layers for xi[..(n + 1) / 2]. Layers: [ eq(xi[j..(n+1)/2], ..) ] for j = (0..(n+1)/2).rev()

Implementations

impl SqrtEqLayers

pub fn from_xi( xi: &[EF], device_ctx: &GpuDeviceCtx, ) -> Result<Self, KernelError>

Build layers from xi values.

This is meant to match behavior of SqrtHyperBuffer::from_xi but with layers.

Example: This means [a,b,c,d] should be sent to low: [[d], [d, c]], high: [[b], [b, a]].

pub fn max_n(&self) -> usize

pub fn low_n(&self) -> usize

pub fn high_n(&self) -> usize

pub fn drop_layer(&mut self)

Drop the highest layer. Drops from high first, then low. This corresponding to pop_front from xi.