Commit 2d088a9f by Dominic Hamon

Merge branch 'ismaelJimenez-added_lambdas'

parents 84cd50b8 e4981431
......@@ -142,6 +142,14 @@ BENCHMARK(BM_StringCompare)
->RangeMultiplier(2)->Range(1<<10, 1<<18)->Complexity();
```
The following code will specify asymptotic complexity with a lambda function,
that might be used to customize high-order term calculation.
```c++
BENCHMARK(BM_StringCompare)->RangeMultiplier(2)
->Range(1<<10, 1<<18)->Complexity([](int n)->double{return n; });
```
### Templated benchmarks
Templated benchmarks work the same way: This example produces and consumes
messages of size `sizeof(v)` `range_x` times. It also outputs throughput in the
......
......@@ -247,9 +247,14 @@ enum BigO {
oNCubed,
oLogN,
oNLogN,
oAuto
oAuto,
oLambda
};
// BigOFunc is passed to a benchmark in order to specify the asymptotic
// computational complexity for the benchmark.
typedef double(BigOFunc)(int);
// State is passed to a running Benchmark and contains state for the
// benchmark to use.
class State {
......@@ -257,7 +262,7 @@ public:
State(size_t max_iters, bool has_x, int x, bool has_y, int y,
int thread_i, int n_threads);
// Returns true iff the benchmark should continue through another iteration.
// Returns true if the benchmark should continue through another iteration.
// NOTE: A benchmark may not return from the test until KeepRunning() has
// returned false.
bool KeepRunning() {
......@@ -358,7 +363,7 @@ public:
// family benchmark, then current benchmark will be part of the computation and complexity_n will
// represent the length of N.
BENCHMARK_ALWAYS_INLINE
void SetComplexityN(size_t complexity_n) {
void SetComplexityN(int complexity_n) {
complexity_n_ = complexity_n;
}
......@@ -439,7 +444,7 @@ private:
size_t bytes_processed_;
size_t items_processed_;
size_t complexity_n_;
int complexity_n_;
public:
// FIXME: Make this private somehow.
......@@ -538,6 +543,10 @@ public:
// the asymptotic computational complexity will be shown on the output.
Benchmark* Complexity(BigO complexity = benchmark::oAuto);
// Set the asymptotic computational complexity for the benchmark. If called
// the asymptotic computational complexity will be shown on the output.
Benchmark* Complexity(BigOFunc* complexity);
// Support for running multiple copies of the same benchmark concurrently
// in multiple threads. This may be useful when measuring the scaling
// of some piece of code.
......
......@@ -86,6 +86,7 @@ class BenchmarkReporter {
// Keep track of arguments to compute asymptotic complexity
BigO complexity;
BigOFunc* complexity_lambda;
int complexity_n;
// Inform print function whether the current run is a complexity report
......@@ -147,7 +148,7 @@ class BenchmarkReporter {
// REQUIRES: 'out' is non-null.
static void PrintBasicContext(std::ostream* out, Context const& context);
private:
private:
std::ostream* output_stream_;
std::ostream* error_stream_;
};
......@@ -159,31 +160,31 @@ class ConsoleReporter : public BenchmarkReporter {
virtual bool ReportContext(const Context& context);
virtual void ReportRuns(const std::vector<Run>& reports);
protected:
protected:
virtual void PrintRunData(const Run& report);
size_t name_field_width_;
};
class JSONReporter : public BenchmarkReporter {
public:
public:
JSONReporter() : first_report_(true) {}
virtual bool ReportContext(const Context& context);
virtual void ReportRuns(const std::vector<Run>& reports);
virtual void Finalize();
private:
private:
void PrintRunData(const Run& report);
bool first_report_;
};
class CSVReporter : public BenchmarkReporter {
public:
public:
virtual bool ReportContext(const Context& context);
virtual void ReportRuns(const std::vector<Run>& reports);
private:
private:
void PrintRunData(const Run& report);
};
......
......@@ -317,6 +317,7 @@ struct Benchmark::Instance {
bool use_real_time;
bool use_manual_time;
BigO complexity;
BigOFunc* complexity_lambda;
bool last_benchmark_instance;
int repetitions;
double min_time;
......@@ -362,6 +363,7 @@ public:
void UseRealTime();
void UseManualTime();
void Complexity(BigO complexity);
void ComplexityLambda(BigOFunc* complexity);
void Threads(int t);
void ThreadRange(int min_threads, int max_threads);
void ThreadPerCpu();
......@@ -382,6 +384,7 @@ private:
bool use_real_time_;
bool use_manual_time_;
BigO complexity_;
BigOFunc* complexity_lambda_;
std::vector<int> thread_counts_;
BenchmarkImp& operator=(BenchmarkImp const&);
......@@ -446,6 +449,7 @@ bool BenchmarkFamilies::FindBenchmarks(
instance.use_real_time = family->use_real_time_;
instance.use_manual_time = family->use_manual_time_;
instance.complexity = family->complexity_;
instance.complexity_lambda = family->complexity_lambda_;
instance.threads = num_threads;
instance.multithreaded = !(family->thread_counts_.empty());
......@@ -573,6 +577,10 @@ void BenchmarkImp::Complexity(BigO complexity){
complexity_ = complexity;
}
void BenchmarkImp::ComplexityLambda(BigOFunc* complexity) {
complexity_lambda_ = complexity;
}
void BenchmarkImp::Threads(int t) {
CHECK_GT(t, 0);
thread_counts_.push_back(t);
......@@ -697,6 +705,12 @@ Benchmark* Benchmark::Complexity(BigO complexity) {
return this;
}
Benchmark* Benchmark::Complexity(BigOFunc* complexity) {
imp_->Complexity(oLambda);
imp_->ComplexityLambda(complexity);
return this;
}
Benchmark* Benchmark::Threads(int t) {
imp_->Threads(t);
return this;
......@@ -855,6 +869,7 @@ void RunBenchmark(const benchmark::internal::Benchmark::Instance& b,
report.items_per_second = items_per_second;
report.complexity_n = total.complexity_n;
report.complexity = b.complexity;
report.complexity_lambda = b.complexity_lambda;
if(report.complexity != oNone)
complexity_reports.push_back(report);
}
......
......@@ -17,31 +17,30 @@
#include "benchmark/benchmark_api.h"
#include "complexity.h"
#include <algorithm>
#include <cmath>
#include "check.h"
#include "complexity.h"
#include "stat.h"
#include <cmath>
#include <algorithm>
#include <functional>
namespace benchmark {
// Internal function to calculate the different scalability forms
std::function<double(int)> FittingCurve(BigO complexity) {
BigOFunc* FittingCurve(BigO complexity) {
switch (complexity) {
case oN:
return [](int n) {return n; };
return [](int n) -> double { return n; };
case oNSquared:
return [](int n) {return n*n; };
return [](int n) -> double { return n * n; };
case oNCubed:
return [](int n) {return n*n*n; };
return [](int n) -> double { return n * n * n; };
case oLogN:
return [](int n) {return log2(n); };
return [](int n) { return log2(n); };
case oNLogN:
return [](int n) {return n * log2(n); };
return [](int n) { return n * log2(n); };
case o1:
default:
return [](int) {return 1; };
return [](int) { return 1.0; };
}
}
......@@ -49,19 +48,19 @@ std::function<double(int)> FittingCurve(BigO complexity) {
std::string GetBigOString(BigO complexity) {
switch (complexity) {
case oN:
return "* N";
return "N";
case oNSquared:
return "* N**2";
return "N^2";
case oNCubed:
return "* N**3";
return "N^3";
case oLogN:
return "* lgN";
return "lgN";
case oNLogN:
return "* NlgN";
return "NlgN";
case o1:
return "* 1";
return "(1)";
default:
return "";
return "f(N)";
}
}
......@@ -75,21 +74,9 @@ std::string GetBigOString(BigO complexity) {
// For a deeper explanation on the algorithm logic, look the README file at
// http://github.com/ismaelJimenez/Minimal-Cpp-Least-Squared-Fit
// This interface is currently not used from the oustide, but it has been
// provided for future upgrades. If in the future it is not needed to support
// Cxx03, then all the calculations could be upgraded to use lambdas because
// they are more powerful and provide a cleaner inferface than enumerators,
// but complete implementation with lambdas will not work for Cxx03
// (e.g. lack of std::function).
// In case lambdas are implemented, the interface would be like :
// -> Complexity([](int n) {return n;};)
// and any arbitrary and valid equation would be allowed, but the option to
// calculate the best fit to the most common scalability curves will still
// be kept.
LeastSq CalculateLeastSq(const std::vector<int>& n,
LeastSq MinimalLeastSq(const std::vector<int>& n,
const std::vector<double>& time,
std::function<double(int)> fitting_curve) {
BigOFunc* fitting_curve) {
double sigma_gn = 0.0;
double sigma_gn_squared = 0.0;
double sigma_time = 0.0;
......@@ -105,6 +92,7 @@ LeastSq CalculateLeastSq(const std::vector<int>& n,
}
LeastSq result;
result.complexity = oLambda;
// Calculate complexity.
result.coef = sigma_time_gn / sigma_gn_squared;
......@@ -134,29 +122,29 @@ LeastSq MinimalLeastSq(const std::vector<int>& n,
const std::vector<double>& time,
const BigO complexity) {
CHECK_EQ(n.size(), time.size());
CHECK_GE(n.size(), 2); // Do not compute fitting curve is less than two benchmark runs are given
CHECK_GE(n.size(), 2); // Do not compute fitting curve is less than two
// benchmark runs are given
CHECK_NE(complexity, oNone);
LeastSq best_fit;
if(complexity == oAuto) {
std::vector<BigO> fit_curves = {
oLogN, oN, oNLogN, oNSquared, oNCubed };
if (complexity == oAuto) {
std::vector<BigO> fit_curves = {oLogN, oN, oNLogN, oNSquared, oNCubed};
// Take o1 as default best fitting curve
best_fit = CalculateLeastSq(n, time, FittingCurve(o1));
best_fit = MinimalLeastSq(n, time, FittingCurve(o1));
best_fit.complexity = o1;
// Compute all possible fitting curves and stick to the best one
for (const auto& fit : fit_curves) {
LeastSq current_fit = CalculateLeastSq(n, time, FittingCurve(fit));
LeastSq current_fit = MinimalLeastSq(n, time, FittingCurve(fit));
if (current_fit.rms < best_fit.rms) {
best_fit = current_fit;
best_fit.complexity = fit;
}
}
} else {
best_fit = CalculateLeastSq(n, time, FittingCurve(complexity));
best_fit = MinimalLeastSq(n, time, FittingCurve(complexity));
best_fit.complexity = complexity;
}
......@@ -164,14 +152,13 @@ LeastSq MinimalLeastSq(const std::vector<int>& n,
}
std::vector<BenchmarkReporter::Run> ComputeStats(
const std::vector<BenchmarkReporter::Run>& reports)
{
const std::vector<BenchmarkReporter::Run>& reports) {
typedef BenchmarkReporter::Run Run;
std::vector<Run> results;
auto error_count = std::count_if(
reports.begin(), reports.end(),
[](Run const& run) {return run.error_occurred;});
auto error_count =
std::count_if(reports.begin(), reports.end(),
[](Run const& run) { return run.error_occurred; });
if (reports.size() - error_count < 2) {
// We don't report aggregated data if there was a single run.
......@@ -190,12 +177,11 @@ std::vector<BenchmarkReporter::Run> ComputeStats(
for (Run const& run : reports) {
CHECK_EQ(reports[0].benchmark_name, run.benchmark_name);
CHECK_EQ(run_iterations, run.iterations);
if (run.error_occurred)
continue;
if (run.error_occurred) continue;
real_accumulated_time_stat +=
Stat1_d(run.real_accumulated_time/run.iterations, run.iterations);
Stat1_d(run.real_accumulated_time / run.iterations, run.iterations);
cpu_accumulated_time_stat +=
Stat1_d(run.cpu_accumulated_time/run.iterations, run.iterations);
Stat1_d(run.cpu_accumulated_time / run.iterations, run.iterations);
items_per_second_stat += Stat1_d(run.items_per_second, run.iterations);
bytes_per_second_stat += Stat1_d(run.bytes_per_second, run.iterations);
}
......@@ -204,10 +190,10 @@ std::vector<BenchmarkReporter::Run> ComputeStats(
Run mean_data;
mean_data.benchmark_name = reports[0].benchmark_name + "_mean";
mean_data.iterations = run_iterations;
mean_data.real_accumulated_time = real_accumulated_time_stat.Mean() *
run_iterations;
mean_data.cpu_accumulated_time = cpu_accumulated_time_stat.Mean() *
run_iterations;
mean_data.real_accumulated_time =
real_accumulated_time_stat.Mean() * run_iterations;
mean_data.cpu_accumulated_time =
cpu_accumulated_time_stat.Mean() * run_iterations;
mean_data.bytes_per_second = bytes_per_second_stat.Mean();
mean_data.items_per_second = items_per_second_stat.Mean();
......@@ -224,10 +210,8 @@ std::vector<BenchmarkReporter::Run> ComputeStats(
stddev_data.benchmark_name = reports[0].benchmark_name + "_stddev";
stddev_data.report_label = mean_data.report_label;
stddev_data.iterations = 0;
stddev_data.real_accumulated_time =
real_accumulated_time_stat.StdDev();
stddev_data.cpu_accumulated_time =
cpu_accumulated_time_stat.StdDev();
stddev_data.real_accumulated_time = real_accumulated_time_stat.StdDev();
stddev_data.cpu_accumulated_time = cpu_accumulated_time_stat.StdDev();
stddev_data.bytes_per_second = bytes_per_second_stat.StdDev();
stddev_data.items_per_second = items_per_second_stat.StdDev();
......@@ -237,8 +221,7 @@ std::vector<BenchmarkReporter::Run> ComputeStats(
}
std::vector<BenchmarkReporter::Run> ComputeBigO(
const std::vector<BenchmarkReporter::Run>& reports)
{
const std::vector<BenchmarkReporter::Run>& reports) {
typedef BenchmarkReporter::Run Run;
std::vector<Run> results;
......@@ -252,19 +235,22 @@ std::vector<BenchmarkReporter::Run> ComputeBigO(
// Populate the accumulators.
for (const Run& run : reports) {
n.push_back(run.complexity_n);
real_time.push_back(run.real_accumulated_time/run.iterations);
cpu_time.push_back(run.cpu_accumulated_time/run.iterations);
real_time.push_back(run.real_accumulated_time / run.iterations);
cpu_time.push_back(run.cpu_accumulated_time / run.iterations);
}
LeastSq result_cpu = MinimalLeastSq(n, cpu_time, reports[0].complexity);
// result_cpu.complexity is passed as parameter to result_real because in case
// reports[0].complexity is oAuto, the noise on the measured data could make
// the best fit function of Cpu and Real differ. In order to solve this, we
// take the best fitting function for the Cpu, and apply it to Real data.
LeastSq result_real = MinimalLeastSq(n, real_time, result_cpu.complexity);
LeastSq result_cpu;
LeastSq result_real;
std::string benchmark_name = reports[0].benchmark_name.substr(0, reports[0].benchmark_name.find('/'));
if (reports[0].complexity == oLambda) {
result_cpu = MinimalLeastSq(n, cpu_time, reports[0].complexity_lambda);
result_real = MinimalLeastSq(n, real_time, reports[0].complexity_lambda);
} else {
result_cpu = MinimalLeastSq(n, cpu_time, reports[0].complexity);
result_real = MinimalLeastSq(n, real_time, result_cpu.complexity);
}
std::string benchmark_name =
reports[0].benchmark_name.substr(0, reports[0].benchmark_name.find('/'));
// Get the data from the accumulator to BenchmarkReporter::Run's.
Run big_o;
......
......@@ -60,11 +60,5 @@ struct LeastSq {
// Function to return an string for the calculated complexity
std::string GetBigOString(BigO complexity);
// Find the coefficient for the high-order term in the running time, by
// minimizing the sum of squares of relative error.
LeastSq MinimalLeastSq(const std::vector<int>& n,
const std::vector<double>& time,
const BigO complexity = oAuto);
} // end namespace benchmark
#endif // COMPLEXITY_H_
......@@ -15,9 +15,9 @@
#include "benchmark/reporter.h"
#include "complexity.h"
#include <algorithm>
#include <cstdint>
#include <cstdio>
#include <algorithm>
#include <iostream>
#include <string>
#include <tuple>
......@@ -62,8 +62,8 @@ void ConsoleReporter::ReportRuns(const std::vector<Run>& reports) {
void ConsoleReporter::PrintRunData(const Run& result) {
auto& Out = GetOutputStream();
auto name_color = (result.report_big_o || result.report_rms)
? COLOR_BLUE : COLOR_GREEN;
auto name_color =
(result.report_big_o || result.report_rms) ? COLOR_BLUE : COLOR_GREEN;
ColorPrintf(Out, name_color, "%-*s ", name_field_width_,
result.benchmark_name.c_str());
......@@ -89,20 +89,20 @@ void ConsoleReporter::PrintRunData(const Run& result) {
const double real_time = result.GetAdjustedRealTime();
const double cpu_time = result.GetAdjustedCPUTime();
if(result.report_big_o) {
std::string big_o = result.report_big_o ? GetBigOString(result.complexity) : "";
ColorPrintf(Out, COLOR_YELLOW, "%10.4f %s %10.4f %s ",
real_time, big_o.c_str(), cpu_time, big_o.c_str());
} else if(result.report_rms) {
ColorPrintf(Out, COLOR_YELLOW, "%10.0f %% %10.0f %% ",
real_time * 100, cpu_time * 100);
if (result.report_big_o) {
std::string big_o = GetBigOString(result.complexity);
ColorPrintf(Out, COLOR_YELLOW, "%10.2f %s %10.2f %s ", real_time,
big_o.c_str(), cpu_time, big_o.c_str());
} else if (result.report_rms) {
ColorPrintf(Out, COLOR_YELLOW, "%10.0f %% %10.0f %% ", real_time * 100,
cpu_time * 100);
} else {
const char* timeLabel = GetTimeUnitString(result.time_unit);
ColorPrintf(Out, COLOR_YELLOW, "%10.0f %s %10.0f %s ",
real_time, timeLabel, cpu_time, timeLabel);
ColorPrintf(Out, COLOR_YELLOW, "%10.0f %s %10.0f %s ", real_time, timeLabel,
cpu_time, timeLabel);
}
if(!result.report_big_o && !result.report_rms) {
if (!result.report_big_o && !result.report_rms) {
ColorPrintf(Out, COLOR_CYAN, "%10lld", result.iterations);
}
......
......@@ -13,9 +13,10 @@
// limitations under the License.
#include "benchmark/reporter.h"
#include "complexity.h"
#include <cstdint>
#include <algorithm>
#include <cstdint>
#include <iostream>
#include <string>
#include <tuple>
......@@ -79,7 +80,7 @@ void CSVReporter::PrintRunData(const Run & run) {
}
// Do not print iteration on bigO and RMS report
if(!run.report_big_o && !run.report_rms) {
if (!run.report_big_o && !run.report_rms) {
Out << run.iterations;
}
Out << ",";
......@@ -87,8 +88,10 @@ void CSVReporter::PrintRunData(const Run & run) {
Out << run.GetAdjustedRealTime() << ",";
Out << run.GetAdjustedCPUTime() << ",";
// Do not print timeLabel on RMS report
if(!run.report_rms) {
// Do not print timeLabel on bigO and RMS report
if (run.report_big_o) {
Out << GetBigOString(run.complexity);
} else if (!run.report_rms) {
Out << GetTimeUnitString(run.time_unit);
}
Out << ",";
......
......@@ -13,9 +13,10 @@
// limitations under the License.
#include "benchmark/reporter.h"
#include "complexity.h"
#include <cstdint>
#include <algorithm>
#include <cstdint>
#include <iostream>
#include <string>
#include <tuple>
......@@ -128,30 +129,47 @@ void JSONReporter::PrintRunData(Run const& run) {
<< FormatKV("error_message", run.error_message)
<< ",\n";
}
if(!run.report_big_o && !run.report_rms) {
if (!run.report_big_o && !run.report_rms) {
out << indent
<< FormatKV("iterations", run.iterations)
<< ",\n";
}
out << indent
<< FormatKV("real_time", RoundDouble(run.GetAdjustedRealTime()))
<< ",\n";
out << indent
<< FormatKV("cpu_time", RoundDouble(run.GetAdjustedCPUTime()));
if(!run.report_rms) {
out << ",\n" << indent
<< FormatKV("time_unit", GetTimeUnitString(run.time_unit));
} else if (run.report_big_o) {
out << indent
<< FormatKV("cpu_coefficient", RoundDouble(run.GetAdjustedCPUTime()))
<< ",\n";
out << indent
<< FormatKV("real_coefficient", RoundDouble(run.GetAdjustedRealTime()))
<< ",\n";
out << indent
<< FormatKV("big_o", GetBigOString(run.complexity))
<< ",\n";
out << indent
<< FormatKV("time_unit", GetTimeUnitString(run.time_unit));
} else if(run.report_rms) {
out << indent
<< FormatKV("rms", RoundDouble(run.GetAdjustedCPUTime()*100))
<< '%';
}
if (run.bytes_per_second > 0.0) {
out << ",\n" << indent
out << ",\n"
<< indent
<< FormatKV("bytes_per_second", RoundDouble(run.bytes_per_second));
}
if (run.items_per_second > 0.0) {
out << ",\n" << indent
out << ",\n"
<< indent
<< FormatKV("items_per_second", RoundDouble(run.items_per_second));
}
if (!run.report_label.empty()) {
out << ",\n" << indent
out << ",\n"
<< indent
<< FormatKV("label", run.report_label);
}
out << '\n';
......
......@@ -189,7 +189,7 @@ void BM_Complexity_O1(benchmark::State& state) {
}
BENCHMARK(BM_Complexity_O1)->Range(1, 1<<18)->Complexity(benchmark::o1);
std::string bigOStr = "[0-9]+\\.[0-9]+ \\* [0-9]+";
std::string bigOStr = "[0-9]+\\.[0-9]+ \\([0-9]+\\)";
ADD_CASES(&ConsoleOutputTests, {
{join("^BM_Complexity_O1_BigO", bigOStr, bigOStr) + "[ ]*$"},
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment