Struct PostStackHead 

pub struct PostStackHead {
    pub conv: Conv1dDyn,
    pub activation: ActivationType,
    pub state: WaveNetLayerState,
    /* private fields */
}

Post-stack head sub-object for WaveNet / ConvNet architectures.

Contains a causal Conv1D + activation that processes the signal after the stack of layer arrays, before the final head_scale gain. Mirrors the _Head structure in NAMCore’s convnet.h:108-118.

Fields

conv: Conv1dDyn

Causal 1D convolution (dynamic runtime dimensions).

activation: ActivationType

Activation function applied after convolution.

state: WaveNetLayerState

Ring buffer state for causal convolution lookback.

Implementations

impl PostStackHead

pub fn from_config(config: &HeadConfig, in_channels: usize) -> Result<Self>

Creates a new PostStackHead from the parsed HeadConfig and the input channel count from the last layer array.

Missing fields in HeadConfig fall back to sensible defaults:

• channels → in_channels (same as the last array’s head projection)
• out_channels → 1 (mono output)
• kernel_size → 3
• bias → false
• activation → “Tanh”

Weight and bias arrays are zero-initialized and must be populated by the dispatcher via set_weights and set_bias.

pub fn receptive_field(&self) -> usize

Returns the receptive field contribution of this head (kernel size). Must be added to the global model receptive field for prewarm.

pub fn out_channels(&self) -> usize

Number of output channels produced by this head.

pub fn in_channels(&self) -> usize

Number of input channels expected by this head.

pub fn set_weights(&mut self, weights: &[f32])

Loads convolution weights from a flat f32 slice.

pub fn set_bias(&mut self, bias: &[f32])

Loads convolution bias from a flat f32 slice, if present.

pub unsafe fn process_block( &mut self, input: &[f32], output: &mut [f32], num_frames: usize, )

Public dispatch wrapper that selects the optimal SIMD path.

Safety

Input and output slices must have sizes compatible with the head dimensions: input.len() == num_frames * in_ch, output.len() == num_frames * out_ch. The ring buffer state must have been properly initialized (via prewarm or sufficient prior processing) to cover the causal receptive field.

pub unsafe fn process_block_internal<M: SimdMath>( &mut self, input: &[f32], output: &mut [f32], num_frames: usize, )

SIMD-dispatched processing kernel.

Writes num_frames of input into the ring buffer, runs the causal Conv1D, applies activation, and writes results to output.

Input layout: frame-interleaved [f0_c0, f0_c1, ..., f1_c0, ...]. Output layout: frame-interleaved [f0_c0, f0_c1, ..., f1_c0, ...].

Safety

input and output must have sizes num_frames * in_ch and num_frames * out_ch respectively. The ring buffer must have been properly initialized.

pub fn prewarm(&mut self)

Public prewarm wrapper with SIMD dispatch.

pub unsafe fn prewarm_internal<M: SimdMath>(&mut self)

Fills the conv state buffer with a single frame of silence replicated backward to cover the entire receptive field.

Safety

Must be called via dispatch_simd! macro. The state buffer must be properly allocated and the ring buffer start pointer must be valid.

Trait Implementations

impl Clone for PostStackHead

fn clone(&self) -> PostStackHead

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.