Skip to content

README.md

Latest commit

 

History

History
94 lines (68 loc) · 3.73 KB

README.md

File metadata and controls

94 lines (68 loc) · 3.73 KB

Zigen Performance Benchmarks

Performance comparison testing system for Zigen vs Eigen 5.0.0.

Quick Start

cd bench

# Default: dim 64, f64
./run_benchmark.sh

# Custom dimension
./run_benchmark.sh --dim 128

# Full sweep: dims 64,256,1024 × f32,f64
./run_benchmark.sh --all

# Single precision
./run_benchmark.sh --precision f32

# Fixed random seed
./run_benchmark.sh --seed 42

Output

Results are generated in bench/tmp/output_<platform>/:

tmp/output_mac26_aarch_m3pro/
├── results.md              # Markdown report with tables
├── dim64_f64/
│   ├── zigen_bench.txt     # Raw Zigen timing output
│   ├── eigen_bench.txt     # Raw Eigen timing output
│   ├── zigen_results.txt   # Zigen numeric results
│   └── eigen_results.txt   # Eigen numeric results
└── test_data_dim64/        # Generated test matrices

Test Suite

28 tests covering all major operations:

No. Category Tests
1-6 Matrix Ops Add, Sub, Mul, Transpose, Trace, Determinant
7-10 Vector Ops Dot, Cross, Normalize, Norm
11-13 Decompositions LU, QR, Cholesky
14-17 Quaternions Mul, Normalize, toRotationMatrix, SLERP
18-19 Extended SVD, EigenSolver
20-21 Solve Inverse, Linear Solve Ax=b
22-24, 27-28 Dynamic Mul, LU, QR, Cholesky, Transpose
25-26 Sparse SpMV, SparseLU Solve

Zero-Allocation Design

All benchmark loops use zero-allocation APIs to ensure fair comparison with Eigen:

  • Workspace reuse: init() + computeFrom() for decompositions
  • In-place operations: addInto(), mulInto(), transposeInto()
  • Buffer writes: solveInto(), mulVecInto(), inverseDynamicInto()

This eliminates allocator overhead from timing measurements, matching Eigen's stack-based approach.

Files

File Description
run_benchmark.sh Main benchmark script (compile, run, verify, report)
benchmark_zigen.zig Zigen benchmark entry point
benchmark_eigen.cpp Eigen benchmark entry point
generate_test_data.py Test data generator (NumPy)
generate_report.py Markdown report generator
zigen/*.zig Zigen benchmark modules (basic_matrix, dynamic, extended, sparse, etc.)
eigen/*.hpp Eigen benchmark modules

Requirements

  • Zig 0.16.0+
  • C++17 compiler (clang++/g++)
  • Eigen 5.0.0 (at ../eigen-5.0.0 or set EIGEN_PATH)
  • Python 3 (with NumPy, for data generation)

Correctness Verification

Before reporting performance, the script verifies that Zigen and Eigen produce matching results:

Precision Absolute Tolerance
f32 0.0001
f64 0.000001

Note: For very large values (e.g., determinants at high dimensions), the absolute difference may exceed the tolerance even though the relative error is at machine epsilon. These are flagged as warnings but do not prevent report generation.