X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=folly%2FBenchmark.cpp;h=de379ec598367516f421c3c46aeb6af2c43b7307;hb=1509cebb0407aee1ca0a381af6f737d53aa2325b;hp=545cefccb574a08cc065b132a97980eaf7876647;hpb=efdc68945456c0c71a05d4a58cf8491454af3ce3;p=folly.git

diff --git a/folly/Benchmark.cpp b/folly/Benchmark.cpp
index 545cefcc..de379ec5 100644
--- a/folly/Benchmark.cpp
+++ b/folly/Benchmark.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright 2012 Facebook, Inc.
+ * Copyright 2014 Facebook, Inc.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,10 +16,10 @@
 
 // @author Andrei Alexandrescu (andrei.alexandrescu@fb.com)
 
-#include "Benchmark.h"
-#include "Foreach.h"
-#include "json.h"
-#include "String.h"
+#include <folly/Benchmark.h>
+#include <folly/Foreach.h>
+#include <folly/json.h>
+#include <folly/String.h>
 
 #include <algorithm>
 #include <boost/regex.hpp>
@@ -28,6 +28,7 @@
 #include <limits>
 #include <utility>
 #include <vector>
+#include <cstring>
 
 using namespace std;
 
@@ -40,6 +41,9 @@ DEFINE_string(bm_regex, "",
 DEFINE_int64(bm_min_usec, 100,
              "Minimum # of microseconds we'll accept for each benchmark.");
 
+DEFINE_int64(bm_min_iters, 1,
+             "Minimum # of iterations we'll try for each benchmark.");
+
 DEFINE_int32(bm_max_secs, 1,
              "Maximum # of seconds we'll spend on each benchmark.");
 
@@ -48,17 +52,45 @@ namespace folly {
 
 BenchmarkSuspender::NanosecondsSpent BenchmarkSuspender::nsSpent;
 
-typedef function<uint64_t(unsigned int)> BenchmarkFun;
-static vector<tuple<const char*, const char*, BenchmarkFun>> benchmarks;
+typedef function<detail::TimeIterPair(unsigned int)> BenchmarkFun;
+
+
+vector<tuple<const char*, const char*, BenchmarkFun>>& benchmarks() {
+  static vector<tuple<const char*, const char*, BenchmarkFun>> _benchmarks;
+  return _benchmarks;
+}
+
+#define FB_FOLLY_GLOBAL_BENCHMARK_BASELINE fbFollyGlobalBenchmarkBaseline
+#define FB_STRINGIZE_X2(x) FB_STRINGIZE(x)
 
 // Add the global baseline
-BENCHMARK(globalBenchmarkBaseline) {
+BENCHMARK(FB_FOLLY_GLOBAL_BENCHMARK_BASELINE) {
+#ifdef _MSC_VER
+  _ReadWriteBarrier();
+#else
   asm volatile("");
+#endif
+}
+
+int getGlobalBenchmarkBaselineIndex() {
+  const char *global = FB_STRINGIZE_X2(FB_FOLLY_GLOBAL_BENCHMARK_BASELINE);
+  auto it = std::find_if(
+    benchmarks().begin(),
+    benchmarks().end(),
+    [global](const tuple<const char*, const char*, BenchmarkFun> &v) {
+      return std::strcmp(get<1>(v), global) == 0;
+    }
+  );
+  CHECK(it != benchmarks().end());
+  return it - benchmarks().begin();
 }
 
+#undef FB_STRINGIZE_X2
+#undef FB_FOLLY_GLOBAL_BENCHMARK_BASELINE
+
 void detail::addBenchmarkImpl(const char* file, const char* name,
                               BenchmarkFun fun) {
-  benchmarks.emplace_back(file, name, std::move(fun));
+  benchmarks().emplace_back(file, name, std::move(fun));
 }
 
 /**
@@ -207,14 +239,15 @@ static double runBenchmarkGetNSPerIteration(const BenchmarkFun& fun,
   // the clock resolution is worse than that, it will be larger. In
   // essence we're aiming at making the quantization noise 0.01%.
   static const auto minNanoseconds =
-    max(FLAGS_bm_min_usec * 1000UL, min(resolutionInNs * 100000, 1000000000UL));
+    max<uint64_t>(FLAGS_bm_min_usec * 1000UL,
+        min<uint64_t>(resolutionInNs * 100000, 1000000000ULL));
 
   // We do measurements in several epochs and take the minimum, to
   // account for jitter.
   static const unsigned int epochs = 1000;
   // We establish a total time budget as we don't want a measurement
   // to take too long. This will curtail the number of actual epochs.
-  const uint64_t timeBudgetInNs = FLAGS_bm_max_secs * 1000000000;
+  const uint64_t timeBudgetInNs = FLAGS_bm_max_secs * 1000000000ULL;
   timespec global;
   CHECK_EQ(0, clock_gettime(CLOCK_REALTIME, &global));
 
@@ -222,14 +255,15 @@ static double runBenchmarkGetNSPerIteration(const BenchmarkFun& fun,
   size_t actualEpochs = 0;
 
   for (; actualEpochs < epochs; ++actualEpochs) {
-    for (unsigned int n = 1; n < (1UL << 30); n *= 2) {
-      auto const nsecs = fun(n);
-      if (nsecs < minNanoseconds) {
+    for (unsigned int n = FLAGS_bm_min_iters; n < (1UL << 30); n *= 2) {
+      auto const nsecsAndIter = fun(n);
+      if (nsecsAndIter.first < minNanoseconds) {
         continue;
       }
       // We got an accurate enough timing, done. But only save if
       // smaller than the current result.
-      epochResults[actualEpochs] = max(0.0, double(nsecs) / n - globalBaseline);
+      epochResults[actualEpochs] = max(0.0, double(nsecsAndIter.first) /
+                                       nsecsAndIter.second - globalBaseline);
       // Done with the current epoch, we got a meaningful timing.
       break;
     }
@@ -263,7 +297,7 @@ static const ScaleInfo kTimeSuffixes[] {
   { 1E-9, "ns" },
   { 1E-12, "ps" },
   { 1E-15, "fs" },
-  { 0, NULL },
+  { 0, nullptr },
 };
 
 static const ScaleInfo kMetricSuffixes[] {
@@ -285,7 +319,7 @@ static const ScaleInfo kMetricSuffixes[] {
   { 1E-18, "a" }, // atto
   { 1E-21, "z" }, // zepto
   { 1E-24, "y" }, // yocto
-  { 0, NULL },
+  { 0, nullptr },
 };
 
 static string humanReadable(double n, unsigned int decimals,
@@ -296,7 +330,7 @@ static string humanReadable(double n, unsigned int decimals,
 
   const double absValue = fabs(n);
   const ScaleInfo* scale = scales;
-  while (absValue < scale[0].boundary && scale[1].suffix != NULL) {
+  while (absValue < scale[0].boundary && scale[1].suffix != nullptr) {
     ++scale;
   }
 
@@ -315,12 +349,12 @@ static string metricReadable(double n, unsigned int decimals) {
 static void printBenchmarkResultsAsTable(
   const vector<tuple<const char*, const char*, double> >& data) {
   // Width available
-  static const uint columns = 76;
+  static const unsigned int columns = 76;
 
   // Compute the longest benchmark name
   size_t longestName = 0;
-  FOR_EACH_RANGE (i, 1, benchmarks.size()) {
-    longestName = max(longestName, strlen(get<1>(benchmarks[i])));
+  FOR_EACH_RANGE (i, 1, benchmarks().size()) {
+    longestName = max(longestName, strlen(get<1>(benchmarks()[i])));
   }
 
   // Print a horizontal rule
@@ -404,10 +438,10 @@ static void printBenchmarkResults(
 }
 
 void runBenchmarks() {
-  CHECK(!benchmarks.empty());
+  CHECK(!benchmarks().empty());
 
   vector<tuple<const char*, const char*, double>> results;
-  results.reserve(benchmarks.size() - 1);
+  results.reserve(benchmarks().size() - 1);
 
   std::unique_ptr<boost::regex> bmRegex;
   if (!FLAGS_bm_regex.empty()) {
@@ -416,19 +450,24 @@ void runBenchmarks() {
 
   // PLEASE KEEP QUIET. MEASUREMENTS IN PROGRESS.
 
-  auto const globalBaseline = runBenchmarkGetNSPerIteration(
-    get<2>(benchmarks.front()), 0);
-  FOR_EACH_RANGE (i, 1, benchmarks.size()) {
+  unsigned int baselineIndex = getGlobalBenchmarkBaselineIndex();
+
+  auto const globalBaseline =
+      runBenchmarkGetNSPerIteration(get<2>(benchmarks()[baselineIndex]), 0);
+  FOR_EACH_RANGE (i, 0, benchmarks().size()) {
+    if (i == baselineIndex) {
+      continue;
+    }
     double elapsed = 0.0;
-    if (!strcmp(get<1>(benchmarks[i]), "-") == 0) { // skip separators
-      if (bmRegex && !boost::regex_search(get<1>(benchmarks[i]), *bmRegex)) {
+    if (strcmp(get<1>(benchmarks()[i]), "-") != 0) { // skip separators
+      if (bmRegex && !boost::regex_search(get<1>(benchmarks()[i]), *bmRegex)) {
         continue;
       }
-      elapsed = runBenchmarkGetNSPerIteration(get<2>(benchmarks[i]),
+      elapsed = runBenchmarkGetNSPerIteration(get<2>(benchmarks()[i]),
                                               globalBaseline);
     }
-    results.emplace_back(get<0>(benchmarks[i]),
-                         get<1>(benchmarks[i]), elapsed);
+    results.emplace_back(get<0>(benchmarks()[i]),
+                         get<1>(benchmarks()[i]), elapsed);
   }
 
   // PLEASE MAKE NOISE. MEASUREMENTS DONE.