Struct MixedPrecisionPlan 

pub struct MixedPrecisionPlan {
    pub config: MixedPrecisionConfig,
    pub rows: u32,
    pub cols: u32,
    pub nnz: u64,
    pub values_bytes: u64,
    pub values_bytes_full: u64,
}

Pre-computed execution plan for mixed-precision SpMV.

Contains the resolved configuration, estimated bandwidth savings, and memory footprint information needed to launch the kernel.

Fields

config: MixedPrecisionConfig

Resolved configuration (Auto replaced with concrete algorithm).

rows: u32

Number of matrix rows.

cols: u32

Number of matrix columns.

nnz: u64

Number of non-zero entries.

values_bytes: u64

Memory footprint in bytes for the FP16/BF16 values array.

values_bytes_full: u64

Memory footprint in bytes for the same values if stored in compute precision.

Implementations

impl MixedPrecisionPlan

pub fn bandwidth_savings_ratio(&self) -> f64

Returns the bandwidth savings ratio (e.g., ~2.0 for FP16 vs FP32 values).

This only accounts for the values array; row_offsets and col_indices remain unchanged as i32 arrays.

pub fn estimated_gflops(&self, peak_bandwidth_gb_s: f64) -> f64

Estimates peak achievable GFLOPS assuming the kernel is perfectly memory-bandwidth-bound.

Uses a simple roofline model: each non-zero requires one FMA (2 flops), plus the memory traffic to load the value, column index, and vector element.

pub fn avg_nnz_per_row(&self) -> f64

Returns the average non-zeros per row.

Trait Implementations

impl Clone for MixedPrecisionPlan

fn clone(&self) -> MixedPrecisionPlan

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MixedPrecisionPlan

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.