chomsky-cost

Multi-dimensional cost modeling for the Chomsky optimization framework.

Overview

In the vast space of equivalent program forms managed by the E-Graph, chomsky-cost provides the "fitness function" to evaluate which form is best for a given target. It models multiple metrics like latency, throughput, code size, and energy consumption.

Features

• Multi-dimensional Cost: The Cost struct tracks latency, throughput, size, and energy.
• Cost Models: The CostModel trait allows defining how each IKun operator impacts the total cost on different hardware.
• Customizable Weighting: Support for scalarizing multi-dimensional costs into a single score using weighted sums, allowing users to prioritize different optimization targets (e.g., optimize for size vs. optimize for speed).
• Default Models: Includes a DefaultCostModel and target-specific examples like JsCostModel.

Cost Metrics

• Latency: Estimated execution time of a single operation.
• Throughput: Number of operations that can be processed in parallel or pipelined.
• Size: Estimated binary or source code size of the operation.
• Energy: Estimated energy consumption.

Usage

use chomsky_cost::{Cost, CostModel, DefaultCostModel};
use chomsky_uir::IKun;

let model = DefaultCostModel;
let node = IKun::Constant(42);
let cost = model.cost(&node);

println!("Score: {}", cost.score());