bench: Add support for measuring CPU cycles

This adds cycle min/max/avg to the statistics.

Supported on x86 and x86_64 (natively through rdtsc), as well as Linux
(perf syscall).

src/Makefile.bench.include

@@ -22,7 +22,9 @@ bench_bench_bitcoin_SOURCES = \
   bench/mempool_eviction.cpp \
   bench/verify_script.cpp \
   bench/base58.cpp \
-  bench/lockedpool.cpp
+  bench/lockedpool.cpp \
+  bench/perf.cpp \
+  bench/perf.h
 
 nodist_bench_bench_bitcoin_SOURCES = $(GENERATED_TEST_FILES)
 

src/bench/bench.cpp

@@ -3,6 +3,7 @@
 // file COPYING or http://www.opensource.org/licenses/mit-license.php.
 
 #include "bench.h"
+#include "perf.h"
 
 #include <iostream>
 #include <iomanip>
@@ -26,7 +27,9 @@ BenchRunner::BenchRunner(std::string name, BenchFunction func)
 void
 BenchRunner::RunAll(double elapsedTimeForOne)
 {
-    std::cout << "#Benchmark" << "," << "count" << "," << "min" << "," << "max" << "," << "average" << "\n";
+    perf_init();
+    std::cout << "#Benchmark" << "," << "count" << "," << "min" << "," << "max" << "," << "average" << ","
+              << "min_cycles" << "," << "max_cycles" << "," << "average_cycles" << "\n";
 
     for (std::map<std::string,BenchFunction>::iterator it = benchmarks.begin();
          it != benchmarks.end(); ++it) {
@@ -35,6 +38,7 @@ BenchRunner::RunAll(double elapsedTimeForOne)
         BenchFunction& func = it->second;
         func(state);
     }
+    perf_fini();
 }
 
 bool State::KeepRunning()
@@ -44,8 +48,10 @@ bool State::KeepRunning()
       return true;
     }
     double now;
+    uint64_t nowCycles;
     if (count == 0) {
         lastTime = beginTime = now = gettimedouble();
+        lastCycles = beginCycles = nowCycles = perf_cpucycles();
     }
     else {
         now = gettimedouble();
@@ -53,6 +59,13 @@ bool State::KeepRunning()
         double elapsedOne = elapsed * countMaskInv;
         if (elapsedOne < minTime) minTime = elapsedOne;
         if (elapsedOne > maxTime) maxTime = elapsedOne;
+
+        // We only use relative values, so don't have to handle 64-bit wrap-around specially
+        nowCycles = perf_cpucycles();
+        uint64_t elapsedOneCycles = (nowCycles - lastCycles) * countMaskInv;
+        if (elapsedOneCycles < minCycles) minCycles = elapsedOneCycles;
+        if (elapsedOneCycles > maxCycles) maxCycles = elapsedOneCycles;
+
         if (elapsed*128 < maxElapsed) {
           // If the execution was much too fast (1/128th of maxElapsed), increase the count mask by 8x and restart timing.
           // The restart avoids including the overhead of this code in the measurement.
@@ -61,6 +74,8 @@ bool State::KeepRunning()
           count = 0;
           minTime = std::numeric_limits<double>::max();
           maxTime = std::numeric_limits<double>::min();
+          minCycles = std::numeric_limits<uint64_t>::max();
+          maxCycles = std::numeric_limits<uint64_t>::min();
           return true;
         }
         if (elapsed*16 < maxElapsed) {
@@ -72,6 +87,7 @@ bool State::KeepRunning()
         }
     }
     lastTime = now;
+    lastCycles = nowCycles;
     ++count;
 
     if (now - beginTime < maxElapsed) return true; // Keep going
@@ -80,7 +96,9 @@ bool State::KeepRunning()
 
     // Output results
     double average = (now-beginTime)/count;
-    std::cout << std::fixed << std::setprecision(15) << name << "," << count << "," << minTime << "," << maxTime << "," << average << "\n";
+    int64_t averageCycles = (nowCycles-beginCycles)/count;
+    std::cout << std::fixed << std::setprecision(15) << name << "," << count << "," << minTime << "," << maxTime << "," << average << ","
+              << minCycles << "," << maxCycles << "," << averageCycles << "\n";
 
     return false;
 }

src/bench/bench.h

@@ -41,12 +41,18 @@ namespace benchmark {
         double maxElapsed;
         double beginTime;
         double lastTime, minTime, maxTime, countMaskInv;
-        int64_t count;
+        uint64_t count;
-        int64_t countMask;
+        uint64_t countMask;
+        uint64_t beginCycles;
+        uint64_t lastCycles;
+        uint64_t minCycles;
+        uint64_t maxCycles;
     public:
         State(std::string _name, double _maxElapsed) : name(_name), maxElapsed(_maxElapsed), count(0) {
             minTime = std::numeric_limits<double>::max();
             maxTime = std::numeric_limits<double>::min();
+            minCycles = std::numeric_limits<uint64_t>::max();
+            maxCycles = std::numeric_limits<uint64_t>::min();
             countMask = 1;
             countMaskInv = 1./(countMask + 1);
         }

src/bench/perf.cpp

@@ -0,0 +1,53 @@
+// Copyright (c) 2016 The Bitcoin Core developers
+// Distributed under the MIT software license, see the accompanying
+// file COPYING or http://www.opensource.org/licenses/mit-license.php.
+
+#include "perf.h"
+
+#if defined(__i386__) || defined(__x86_64__)
+
+/* These architectures support quering the cycle counter
+ * from user space, no need for any syscall overhead.
+ */
+void perf_init(void) { }
+void perf_fini(void) { }
+
+#elif defined(__linux__)
+
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <linux/perf_event.h>
+
+static int fd = -1;
+static struct perf_event_attr attr;
+
+void perf_init(void)
+{
+    attr.type = PERF_TYPE_HARDWARE;
+    attr.config = PERF_COUNT_HW_CPU_CYCLES;
+    fd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
+}
+
+void perf_fini(void)
+{
+    if (fd != -1) {
+        close(fd);
+    }
+}
+
+uint64_t perf_cpucycles(void)
+{
+    uint64_t result = 0;
+    if (fd == -1 || read(fd, &result, sizeof(result)) < (ssize_t)sizeof(result)) {
+        return 0;
+    }
+    return result;
+}
+
+#else /* Unhandled platform */
+
+void perf_init(void) { }
+void perf_fini(void) { }
+uint64_t perf_cpucycles(void) { return 0; }
+
+#endif

src/bench/perf.h

@@ -0,0 +1,37 @@
+// Copyright (c) 2016 The Bitcoin Core developers
+// Distributed under the MIT software license, see the accompanying
+// file COPYING or http://www.opensource.org/licenses/mit-license.php.
+
+/** Functions for measurement of CPU cycles */
+#ifndef H_PERF
+#define H_PERF
+
+#include <stdint.h>
+
+#if defined(__i386__)
+
+static inline uint64_t perf_cpucycles(void)
+{
+    uint64_t x;
+    __asm__ volatile (".byte 0x0f, 0x31" : "=A" (x));
+    return x;
+}
+
+#elif defined(__x86_64__)
+
+static inline uint64_t perf_cpucycles(void)
+{
+    uint32_t hi, lo;
+    __asm__ __volatile__ ("rdtsc" : "=a"(lo), "=d"(hi));
+    return ((uint64_t)lo)|(((uint64_t)hi)<<32);
+}
+#else
+
+uint64_t perf_cpucycles(void);
+
+#endif
+
+void perf_init(void);
+void perf_fini(void);
+
+#endif // H_PERF