CPUBench

A Windows 8+ Microbenchmarking library that does not require test signed drivers. Somewhat restricted by the Windows Kernel Logger, but sufficient for most use cases.

Requirements

Windows 10 SDK

Compilation flags

Make sure to enable full compiler optimizations, such as /O2 on MSVC or -O3 on gcc

Example Usage

void benchmark_mispredicts(CPUBench::BenchmarkContext& context) {
	//Intentionally similar to Google Benchmarking, set up necessary state and pass in anything via user_context
	
	volatile const char c{ 0 };
	std::minstd_rand simple_random;
	simple_random.seed(0xF00D);

	auto val = std::any_cast<int>(context.m_user_context);

	//range style iteration to repeat the code to be microbenchmarked
	for (auto _ : context) {
		if (simple_random() % 100 > val) {
			(void)c;
		}
	}
}

CPUBench::BenchmarkProperties props{};
props.profiling_flags = ETWUtils::ProfilerFlags::BranchMispredictions
		| ETWUtils::ProfilerFlags::BranchInstructions
		| ETWUtils::ProfilerFlags::UnhaltedCoreCycles;
props.repetitions = 10'000'000;
props.name = "MispredictionBenchmark";
props.warmup_cache = true;
props.user_context = 30;

CPUBench::Benchmark benchmark(benchmark_mispredicts, props);

CPUBench::run_benchmarks();

//load the result, set the logger to a custom logger that has the default behavior but also times the function and outputs the misprediction rate (%)
auto& result = CPUBench::get_result("MispredictionBenchmark");
result.set_logger([](CPUBench::BenchmarkResult& res) {
	CPUBench::default_logger(res); //start with default logger behavior 
 
	//add extra output based on counters we want to get
	unsigned int branches = 0;
	unsigned int mispredicts = 0;
	unsigned int core_cycles = 0; //will read unhalted core cycles
 
	for (auto& counter : res.m_counters) {
		if (counter.name == L"BranchInstructions")
			branches = counter.value;
		else if (counter.name == L"BranchMispredictions")
			mispredicts = counter.value;
		else if (counter.name == L"UnhaltedCoreCycles")
			core_cycles = counter.value;
	}
  
	//float precis is fine, we are close to 1.0f
	std::cout << "Branch miss rate: " << mispredicts * 100.f / branches << "%\n";
	std::cout << "Unhalted core time: " << static_cast<double>(core_cycles) / (CPUBench::get_baseclock_mhz() * 1'000'000) << "s\n";
});
result.log();