Jim Cownie | 4cc4bb4 | 2014-10-07 16:25:50 +0000 | [diff] [blame] | 1 | /** @file kmp_stats_timing.cpp |
| 2 | * Timing functions |
| 3 | */ |
| 4 | |
| 5 | |
| 6 | //===----------------------------------------------------------------------===// |
| 7 | // |
| 8 | // The LLVM Compiler Infrastructure |
| 9 | // |
| 10 | // This file is dual licensed under the MIT and the University of Illinois Open |
| 11 | // Source Licenses. See LICENSE.txt for details. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | |
| 16 | #include <stdlib.h> |
| 17 | #include <unistd.h> |
| 18 | |
| 19 | #include <iostream> |
| 20 | #include <iomanip> |
| 21 | #include <sstream> |
| 22 | |
| 23 | #include "kmp_stats_timing.h" |
| 24 | |
| 25 | using namespace std; |
| 26 | |
| 27 | #if KMP_OS_LINUX |
| 28 | # if KMP_MIC |
| 29 | double tsc_tick_count::tick_time() |
| 30 | { |
| 31 | // pretty bad assumption of 1GHz clock for MIC |
| 32 | return 1/((double)1000*1.e6); |
| 33 | } |
| 34 | # else |
| 35 | # include <string.h> |
| 36 | // Extract the value from the CPUID information |
| 37 | double tsc_tick_count::tick_time() |
| 38 | { |
| 39 | static double result = 0.0; |
| 40 | |
| 41 | if (result == 0.0) |
| 42 | { |
| 43 | int cpuinfo[4]; |
| 44 | char brand[256]; |
| 45 | |
| 46 | __cpuid(cpuinfo, 0x80000000); |
| 47 | memset(brand, 0, sizeof(brand)); |
| 48 | int ids = cpuinfo[0]; |
| 49 | |
| 50 | for (unsigned int i=2; i<(ids^0x80000000)+2; i++) |
| 51 | __cpuid(brand+(i-2)*sizeof(cpuinfo), i | 0x80000000); |
| 52 | |
| 53 | char * start = &brand[0]; |
| 54 | for (;*start == ' '; start++) |
| 55 | ; |
| 56 | |
Andrey Churbanov | 74bf17b | 2015-04-02 13:27:08 +0000 | [diff] [blame] | 57 | char * end = brand + KMP_STRLEN(brand) - 3; |
Jim Cownie | 4cc4bb4 | 2014-10-07 16:25:50 +0000 | [diff] [blame] | 58 | uint64_t multiplier; |
| 59 | |
| 60 | if (*end == 'M') multiplier = 1000LL*1000LL; |
| 61 | else if (*end == 'G') multiplier = 1000LL*1000LL*1000LL; |
| 62 | else if (*end == 'T') multiplier = 1000LL*1000LL*1000LL*1000LL; |
| 63 | else |
| 64 | { |
| 65 | cout << "Error determining multiplier '" << *end << "'\n"; |
| 66 | exit (-1); |
| 67 | } |
| 68 | *end = 0; |
| 69 | while (*end != ' ') end--; |
| 70 | end++; |
| 71 | |
| 72 | double freq = strtod(end, &start); |
| 73 | if (freq == 0.0) |
| 74 | { |
| 75 | cout << "Error calculating frequency " << end << "\n"; |
| 76 | exit (-1); |
| 77 | } |
| 78 | |
| 79 | result = ((double)1.0)/(freq * multiplier); |
| 80 | } |
| 81 | return result; |
| 82 | } |
| 83 | # endif |
| 84 | #endif |
| 85 | |
| 86 | static bool useSI = true; |
| 87 | |
| 88 | // Return a formatted string after normalising the value into |
| 89 | // engineering style and using a suitable unit prefix (e.g. ms, us, ns). |
| 90 | std::string formatSI(double interval, int width, char unit) |
| 91 | { |
| 92 | std::stringstream os; |
| 93 | |
| 94 | if (useSI) |
| 95 | { |
| 96 | // Preserve accuracy for small numbers, since we only multiply and the positive powers |
| 97 | // of ten are precisely representable. |
| 98 | static struct { double scale; char prefix; } ranges[] = { |
| 99 | {1.e12,'f'}, |
| 100 | {1.e9, 'p'}, |
| 101 | {1.e6, 'n'}, |
| 102 | {1.e3, 'u'}, |
| 103 | {1.0, 'm'}, |
| 104 | {1.e-3,' '}, |
| 105 | {1.e-6,'k'}, |
| 106 | {1.e-9,'M'}, |
| 107 | {1.e-12,'G'}, |
| 108 | {1.e-15,'T'}, |
| 109 | {1.e-18,'P'}, |
| 110 | {1.e-21,'E'}, |
| 111 | {1.e-24,'Z'}, |
| 112 | {1.e-27,'Y'} |
| 113 | }; |
| 114 | |
| 115 | if (interval == 0.0) |
| 116 | { |
| 117 | os << std::setw(width-3) << std::right << "0.00" << std::setw(3) << unit; |
| 118 | return os.str(); |
| 119 | } |
| 120 | |
| 121 | bool negative = false; |
| 122 | if (interval < 0.0) |
| 123 | { |
| 124 | negative = true; |
| 125 | interval = -interval; |
| 126 | } |
| 127 | |
| 128 | for (int i=0; i<(int)(sizeof(ranges)/sizeof(ranges[0])); i++) |
| 129 | { |
| 130 | if (interval*ranges[i].scale < 1.e0) |
| 131 | { |
| 132 | interval = interval * 1000.e0 * ranges[i].scale; |
| 133 | os << std::fixed << std::setprecision(2) << std::setw(width-3) << std::right << |
| 134 | (negative ? -interval : interval) << std::setw(2) << ranges[i].prefix << std::setw(1) << unit; |
| 135 | |
| 136 | return os.str(); |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 | os << std::setprecision(2) << std::fixed << std::right << std::setw(width-3) << interval << std::setw(3) << unit; |
| 141 | |
| 142 | return os.str(); |
| 143 | } |
| 144 | |
| 145 | tsc_tick_count::tsc_interval_t computeLastInLastOutInterval(timePair * times, int nTimes) |
| 146 | { |
| 147 | timePair lastTimes = times[0]; |
| 148 | tsc_tick_count * startp = lastTimes.get_startp(); |
| 149 | tsc_tick_count * endp = lastTimes.get_endp(); |
| 150 | |
| 151 | for (int i=1; i<nTimes; i++) |
| 152 | { |
| 153 | (*startp) = startp->later(times[i].get_start()); |
| 154 | (*endp) = endp->later (times[i].get_end()); |
| 155 | } |
| 156 | |
| 157 | return lastTimes.duration(); |
| 158 | } |
| 159 | |
| 160 | std::string timePair::format() const |
| 161 | { |
| 162 | std::ostringstream oss; |
| 163 | |
| 164 | oss << start.getValue() << ":" << end.getValue() << " = " << (end-start).getValue(); |
| 165 | |
| 166 | return oss.str(); |
| 167 | } |