|
| 1 | +// Benchmark on libm functions |
| 2 | + |
| 3 | +#include <math.h> |
| 4 | +#include <sys/time.h> |
| 5 | +#include <stdio.h> |
| 6 | +#include <stdlib.h> |
| 7 | + |
| 8 | + |
| 9 | +// Timing facilities |
| 10 | + |
| 11 | +#ifdef __MACH__ |
| 12 | + |
| 13 | +#include <mach/mach_time.h> |
| 14 | + |
| 15 | +class stimer |
| 16 | +{ |
| 17 | +public: |
| 18 | + typedef uint64_t time_type; |
| 19 | + |
| 20 | + stimer() |
| 21 | + { |
| 22 | + ::mach_timebase_info(&m_baseinfo); |
| 23 | + } |
| 24 | + |
| 25 | + time_type current() const |
| 26 | + { |
| 27 | + return ::mach_absolute_time(); |
| 28 | + } |
| 29 | + |
| 30 | + double span(const time_type& t0, const time_type& t1) const |
| 31 | + { |
| 32 | + uint64_t d = (m_baseinfo.numer * (t1 - t0)) / m_baseinfo.denom; |
| 33 | + return static_cast<double>(d) / 1.0e9; |
| 34 | + } |
| 35 | + |
| 36 | +private: |
| 37 | + mach_timebase_info_data_t m_baseinfo; |
| 38 | +}; |
| 39 | + |
| 40 | +#else |
| 41 | + |
| 42 | +class stimer |
| 43 | +{ |
| 44 | +public: |
| 45 | + typedef timespec time_type; |
| 46 | + |
| 47 | + time_type current() const |
| 48 | + { |
| 49 | + time_type t; |
| 50 | + ::clock_gettime(CLOCK_REALTIME, &t); |
| 51 | + return t; |
| 52 | + } |
| 53 | + |
| 54 | + double span(const time_type& t0, const time_type& t1) const |
| 55 | + { |
| 56 | + return double(t1.tv_sec - t0.tv_sec) + |
| 57 | + double(t1.tv_nsec - t0.tv_nsec) * 1.0e-9; |
| 58 | + } |
| 59 | +}; |
| 60 | + |
| 61 | +#endif |
| 62 | + |
| 63 | + |
| 64 | +inline double sec2mps(double s, long n) |
| 65 | +{ |
| 66 | + return n / (s * 1e6); |
| 67 | +} |
| 68 | + |
| 69 | + |
| 70 | +const long ARR_LEN = 1024; |
| 71 | + |
| 72 | +double a[ARR_LEN]; |
| 73 | +double b[ARR_LEN]; |
| 74 | +double r[ARR_LEN]; |
| 75 | + |
| 76 | +#define TFUN1(FNAME) \ |
| 77 | + void test_##FNAME(long n) { \ |
| 78 | + for (int j = 0; j < ARR_LEN; ++j) r[j] = FNAME(a[j]); \ |
| 79 | + stimer tm; \ |
| 80 | + stimer::time_type t0 = tm.current(); \ |
| 81 | + for(int i = 0; i < n; ++i) { \ |
| 82 | + for (int j = 0; j < ARR_LEN; ++j) r[j] = FNAME(a[j]); \ |
| 83 | + } \ |
| 84 | + double s = tm.span(t0, tm.current()); \ |
| 85 | + double mps = sec2mps(s, n * ARR_LEN); \ |
| 86 | + printf(" %-8s: %7.4f MPS\n", #FNAME, mps); } |
| 87 | + |
| 88 | +#define TFUN2(FNAME) \ |
| 89 | + void test_##FNAME(long n) { \ |
| 90 | + for (int j = 0; j < ARR_LEN; ++j) r[j] = FNAME(a[j], b[j]); \ |
| 91 | + stimer tm; \ |
| 92 | + stimer::time_type t0 = tm.current(); \ |
| 93 | + for(int i = 0; i < n; ++i) { \ |
| 94 | + for (int j = 0; j < ARR_LEN; ++j) r[j] = FNAME(a[j], b[j]); \ |
| 95 | + } \ |
| 96 | + double s = tm.span(t0, tm.current()); \ |
| 97 | + double mps = sec2mps(s, n * ARR_LEN); \ |
| 98 | + printf(" %-8s: %7.4f MPS\n", #FNAME, mps); } |
| 99 | + |
| 100 | + |
| 101 | +#define TCALL(FNAME) test_##FNAME(20000) |
| 102 | + |
| 103 | +// define benchmark functions |
| 104 | + |
| 105 | +TFUN2(pow) |
| 106 | +TFUN2(hypot) |
| 107 | + |
| 108 | +TFUN1(exp) |
| 109 | +TFUN1(log) |
| 110 | +TFUN1(log10) |
| 111 | +TFUN1(sin) |
| 112 | +TFUN1(cos) |
| 113 | +TFUN1(tan) |
| 114 | +TFUN1(asin) |
| 115 | +TFUN1(acos) |
| 116 | +TFUN1(atan) |
| 117 | +TFUN2(atan2) |
| 118 | + |
| 119 | +int main(int argc, char *argv[]) |
| 120 | +{ |
| 121 | + // initialize array contents |
| 122 | + for (int i = 0; i < ARR_LEN; ++i) |
| 123 | + { |
| 124 | + a[i] = rand() / (double) RAND_MAX; |
| 125 | + b[i] = rand() / (double) RAND_MAX; |
| 126 | + } |
| 127 | + |
| 128 | + TCALL(pow); |
| 129 | + TCALL(hypot); |
| 130 | + TCALL(exp); |
| 131 | + TCALL(log); |
| 132 | + TCALL(log10); |
| 133 | + TCALL(sin); |
| 134 | + TCALL(cos); |
| 135 | + TCALL(tan); |
| 136 | + TCALL(asin); |
| 137 | + TCALL(acos); |
| 138 | + TCALL(atan); |
| 139 | + TCALL(atan2); |
| 140 | + |
| 141 | + return 0; |
| 142 | +} |
0 commit comments