criterion

Highlights

Criterion is a micro-benchmarking library for modern C++.

• Convenient static registration macros for setting up benchmarks
• Parameterized benchmarks (e.g., vary input size)
• Statistical analysis across multiple runs
• Requires compiler support for C++17 or newer standard
• Header-only library - single header file version available at single_include/
• MIT License

Table of Contents

• Getting Started
  • Simple Benchmark
  • Passing Arguments
  • Passing Arguments (Part 2)
  • CRITERION_BENCHMARK_MAIN and Command-line Options
  • Exporting Results (csv, json etc.)
• Building Library and Samples
• Generating Single Header
• Contributing
• License

Getting Started

Let’s say we have this merge sort implementation that needs to be benchmarked.

template<typename RandomAccessIterator, typename Compare>
void merge_sort(RandomAccessIterator first, RandomAccessIterator last,
                Compare compare, std::size_t size) {
  if (size < 2) return;
  auto middle = first + size / 2;
  merge_sort(first, middle, compare, size / 2);
  merge_sort(middle, last, compare, size - size/2);
  std::inplace_merge(first, middle, last, compare);
}

Simple Benchmark

Include <criterion/criterion.hpp> and you’re good to go.

• Use the BENCHMARK macro to declare a benchmark
• Use SETUP_BENCHMARK and TEARDOWN_BENCHMARK to perform setup and teardown tasks
  • These tasks are not part of the measurement

#include <criterion/criterion.hpp>

BENCHMARK(MergeSort)
{
  SETUP_BENCHMARK(
    const auto size = 100;
    std::vector<int> vec(size, 0); // vector of size 100
  )
 
  // Code to be benchmarked
  merge_sort(vec.begin(), vec.end(), std::less<int>(), size);
  
  TEARDOWN_BENCHMARK(
    vec.clear();
  )
}

CRITERION_BENCHMARK_MAIN()

What if we want to run this benchmark on a variety of sizes?

Passing Arguments

• The BENCHMARK macro can take typed parameters
• Use GET_ARGUMENTS(n) to get the nth argument passed to the benchmark
• For benchmarks that require arguments, use INVOKE_BENCHMARK_FOR_EACH and provide arguments

#include <criterion/criterion.hpp>

BENCHMARK(MergeSort, std::size_t) // <- one parameter to be passed to the benchmark
{
  SETUP_BENCHMARK(
    const auto size = GET_ARGUMENT(0); // <- get the argument passed to the benchmark
    std::vector<int> vec(size, 0);
  )
 
  // Code to be benchmarked
  merge_sort(vec.begin(), vec.end(), std::less<int>(), size);
  
  TEARDOWN_BENCHMARK(
    vec.clear();
  )
}

// Run the above benchmark for a number of inputs:

INVOKE_BENCHMARK_FOR_EACH(MergeSort,
  ("/10", 10),
  ("/100", 100),
  ("/1K", 1000),
  ("/10K", 10000),
  ("/100K", 100000)
)

CRITERION_BENCHMARK_MAIN()

Passing Arguments (Part 2)

Let’s say we have the following struct and we need to create a std::shared_ptr to it.

struct Song {
  std::string artist;
  std::string title;
  Song(const std::string& artist_, const std::string& title_) :
    artist{ artist_ }, title{ title_ } {}
};

Here are two implementations for constructing the std::shared_ptr:

// Functions to be tested
auto Create_With_New() { 
  return std::shared_ptr<Song>(new Song("Black Sabbath", "Paranoid")); 
}

auto Create_With_MakeShared() { 
  return std::make_shared<Song>("Black Sabbath", "Paranoid"); 
}

We can setup a single benchmark that takes a std::function<> and measures performance like below.

BENCHMARK(ConstructSharedPtr, std::function<std::shared_ptr<Song>()>) 
{
  SETUP_BENCHMARK(
    auto test_function = GET_ARGUMENT(0);
  )

  // Code to be benchmarked
  auto song_ptr = test_function();
}

INVOKE_BENCHMARK_FOR_EACH(ConstructSharedPtr, 
  ("/new", Create_With_New),
  ("/make_shared", Create_With_MakeShared)
)

CRITERION_BENCHMARK_MAIN()

CRITERION_BENCHMARK_MAIN and Command-line Options

CRITERION_BENCHMARK_MAIN() provides a main function that:

1. Handles command-line arguments,
2. Runs the registered benchmarks
3. Exports results to file if requested by user.

Here’s the help/man generated by the main function:

foo@bar:~$ ./benchmarks -h

NAME
     ./benchmarks -- Run Criterion benchmarks

SYNOPSIS
     ./benchmarks
           [-w,--warmup <number>]
           [-l,--list] [--list_filtered <regex>] [-r,--run_filtered <regex>]
           [-e,--export_results {csv,json,md,asciidoc} <filename>]
           [-q,--quiet] [-h,--help]
DESCRIPTION
     This microbenchmarking utility repeatedly executes a list of benchmarks,
     statistically analyzing and reporting on the temporal behavior of the executed code.

     The options are as follows:

     -w,--warmup number
          Number of warmup runs (at least 1) to execute before the benchmark (default=3)

     -l,--list
          Print the list of available benchmarks

     --list_filtered regex
          Print a filtered list of available benchmarks (based on user-provided regex)

     -r,--run_filtered regex
          Run a filtered list of available benchmarks (based on user-provided regex)

     -e,--export_results format filename
          Export benchmark results to file. The following are the supported formats.

          csv       Comma separated values (CSV) delimited text file
          json      JavaScript Object Notation (JSON) text file
          md        Markdown (md) text file
          asciidoc  AsciiDoc (asciidoc) text file

     -q,--quiet
          Run benchmarks quietly, suppressing activity indicators

     -h,--help
          Print this help message

Exporting Results (csv, json, etc.)

Benchmarks can be exported to one of a number of formats: .csv, .json, .md, and .asciidoc.

Use --export_results (or -e) to export results to one of the supported formats.

foo@bar:~$ ./vector_sort -e json results.json -q # run quietly and export to JSON

foo@bar:~$ cat results.json
{
  "benchmarks": [
    {
      "name": "VectorSort/100",
      "warmup_runs": 2,
      "iterations": 2857140,
      "mean_execution_time": 168.70,
      "fastest_execution_time": 73.00,
      "slowest_execution_time": 88809.00,
      "lowest_rsd_execution_time": 84.05,
      "lowest_rsd_percentage": 3.29,
      "lowest_rsd_index": 57278,
      "average_iteration_performance": 5927600.84,
      "fastest_iteration_performance": 13698630.14,
      "slowest_iteration_performance": 11260.12
    },
    {
      "name": "VectorSort/1000",
      "warmup_runs": 2,
      "iterations": 2254280,
      "mean_execution_time": 1007.70,
      "fastest_execution_time": 640.00,
      "slowest_execution_time": 102530.00,
      "lowest_rsd_execution_time": 647.45,
      "lowest_rsd_percentage": 0.83,
      "lowest_rsd_index": 14098,
      "average_iteration_performance": 992355.48,
      "fastest_iteration_performance": 1562500.00,
      "slowest_iteration_performance": 9753.24
    },
    {
      "name": "VectorSort/10000",
      "warmup_runs": 2,
      "iterations": 259320,
      "mean_execution_time": 8833.26,
      "fastest_execution_time": 6276.00,
      "slowest_execution_time": 114548.00,
      "lowest_rsd_execution_time": 8374.15,
      "lowest_rsd_percentage": 0.11,
      "lowest_rsd_index": 7905,
      "average_iteration_performance": 113208.45,
      "fastest_iteration_performance": 159337.16,
      "slowest_iteration_performance": 8729.96
    }
  ]
}

Building Library and Samples

cmake -Hall -Bbuild
cmake --build build

# run `merge_sort` sample
./build/samples/merge_sort/merge_sort

Generating Single Header

python3 utils/amalgamate/amalgamate.py -c single_include.json -s .

Contributing

Contributions are welcome, have a look at the CONTRIBUTING.md document for more information.

License

The project is available under the MIT license.