Module analysis 

Modules

baseline

cgp baseline — Save and load performance baselines. Baselines stored in .cgp-baselines/ directory.

bench

cgp bench — Enhanced criterion benchmarking with hardware counters. Spec section 2.3: run cargo bench, capture criterion output, optionally overlay perf stat counters, check regression.

compare

cgp profile compare — Cross-backend comparison. Spec section 2.2: run the same workload across multiple backends and produce a comparison table with TFLOP/s, bandwidth, and speedup ratios.

compete

cgp compete — Head-to-head competitor comparison. Spec section 2.5: run two or more commands, measure wall time, compute TFLOP/s, and produce a comparison table.

contracts

Performance contract verification (CI/CD gate). Extends provable-contracts framework to performance bounds. See spec section 3.4 and 7.1.

diff

Profile diff: compare two CGP profiles and detect regressions. Spec section 2.6: cgp diff. Must complete in <100ms for two saved JSONs (FALSIFY-CGP-062).

explain

cgp explain — Static code analysis for PTX, SIMD assembly, and WGSL shaders. Spec section 2.7: wraps trueno-explain or performs inline analysis. Detects register pressure, instruction mix, and common performance pitfalls.

muda

Muda (Waste) Detection Engine — Seven categories of GPU compute waste. Mapped from Toyota Production System (Ohno, 1988) [7].

regression

Regression detection using bootstrap confidence intervals. Methodology from Hoefler & Belli (2015) [8]: “Scientific Benchmarking of Parallel Computing Systems.” Also supports PELT changepoint detection [43].

roofline

Roofline model implementation per Williams, Waterman & Patterson (2009) [4]. Supports hierarchical GPU roofline per Yang et al. (2020) [13]. Uses Empirical Roofline Toolkit (ERT) methodology [6].