Struct CpuBatchThresholds 

#[non_exhaustive]
pub struct CpuBatchThresholds {
    pub item_overhead_ns: f64,
    pub muladds_per_ns: f64,
    pub min_work_per_lane_ns: f64,
    pub min_items: usize,
    pub min_items_per_lane: usize,
}

The cost and fan-out model CpuBatchStrategy::Auto applies.

This is tcontract_exec::WorkModel; the old vendor-dimension cutoff is not part of it.

Examples

use tenferro_cpu::CpuBatchThresholds;

assert_eq!(CpuBatchThresholds::default().min_items, 2);

The cost and fan-out model BatchStrategy::Auto applies.

The item rate is the serial GEMM throughput of the host class (host::muladds_per_ns: 27 multiply-adds per nanosecond on aarch64-apple and the generic row, 15.6 on x86_64-avx512; see crates/tcontract-exec/src/host.rs and docs/design/thresholds.md of threadteam-rs); the width a team pays for comes from the token, not from this model.

The three item-cost fields follow tenferro’s restored CpuBatchThresholds: one item costs roughly item_overhead_ns plus one nanosecond per muladds_per_ns multiply-adds, a lane needs min_work_per_lane_ns of estimated work, and a fan-out needs min_items items, at least min_items_per_lane per lane.

Examples

use tcontract_exec::batch::WorkModel;

let model = WorkModel::default().with_min_items(4).with_muladds_per_ns(32.0);
assert_eq!(model.min_items, 4);
assert_eq!(model.muladds_per_ns, 32.0);

Fields (Non-exhaustive)

Non-exhaustive structs could have additional fields added in future. Therefore, non-exhaustive structs cannot be constructed in external crates using the traditional Struct { .. } syntax; cannot be matched against without a wildcard ..; and struct update syntax will not work.

item_overhead_ns: f64

Estimated fixed cost of one item, in nanoseconds.

The default is zero so that Default reproduces the packed ns_per_flop() estimate.

muladds_per_ns: f64

Estimated multiply-adds per nanosecond on one lane.

1.0 / ns_per_flop() / 2.0 (27 on aarch64-apple and the generic row, 15.6 on x86_64-avx512) reproduces 2 * macs * ns_per_flop(): the packed driver counts two flops per multiply-add. A value <= 0 is treated as one multiply-add per nanosecond.

min_work_per_lane_ns: f64

Least estimated work a lane must receive, in nanoseconds; 0 disables the work gate (the packed default).

min_items: usize

Fewest items for which Auto fans out at all.

min_items_per_lane: usize

Fewest items each lane must receive.

Implementations

impl WorkModel

pub fn with_item_overhead_ns(self, nanoseconds: f64) -> WorkModel

Return this model with a different per-item overhead.

pub fn with_muladds_per_ns(self, muladds_per_ns: f64) -> WorkModel

Return this model with a different lane throughput.

Examples

use tcontract_exec::batch::WorkModel;
assert_eq!(WorkModel::default().with_muladds_per_ns(16.0).muladds_per_ns, 16.0);

pub fn with_min_work_per_lane_ns(self, nanoseconds: f64) -> WorkModel

Return this model with a different minimum work per lane.

Examples

use tcontract_exec::batch::WorkModel;
assert_eq!(WorkModel::default().with_min_work_per_lane_ns(8_000.0).min_work_per_lane_ns, 8_000.0);

pub fn with_min_items(self, items: usize) -> WorkModel

Return this model with a different minimum fan-out item count.

pub fn with_min_items_per_lane(self, items: usize) -> WorkModel

Return this model with a different minimum item count per lane.

Trait Implementations

impl Clone for WorkModel

fn clone(&self) -> WorkModel

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for WorkModel

impl Debug for WorkModel

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

Formats the value using the given formatter.

impl Default for WorkModel

fn default() -> WorkModel

Returns the “default value” for a type.

impl PartialEq for WorkModel

fn eq(&self, other: &WorkModel) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for WorkModel