|
| 1 | +#include "performancecounters/benchmarker.h" |
| 2 | +#include "fast_float/fast_float.h" |
| 3 | +#include <charconv> |
| 4 | +#include <cstdint> |
| 5 | +#include <cstdio> |
| 6 | +#include <cstdlib> |
| 7 | +#include <cstring> |
| 8 | +#include <random> |
| 9 | + |
| 10 | +void pretty_print(size_t volume, size_t bytes, std::string name, |
| 11 | + event_aggregate agg) { |
| 12 | + printf("%-40s : ", name.c_str()); |
| 13 | + printf(" %5.2f GB/s ", bytes / agg.fastest_elapsed_ns()); |
| 14 | + printf(" %5.1f Ma/s ", volume * 1000.0 / agg.fastest_elapsed_ns()); |
| 15 | + printf(" %5.2f ns/d ", agg.fastest_elapsed_ns() / volume); |
| 16 | + if (collector.has_events()) { |
| 17 | + printf(" %5.2f GHz ", agg.fastest_cycles() / agg.fastest_elapsed_ns()); |
| 18 | + printf(" %5.2f c/d ", agg.fastest_cycles() / volume); |
| 19 | + printf(" %5.2f i/d ", agg.fastest_instructions() / volume); |
| 20 | + printf(" %5.2f c/b ", agg.fastest_cycles() / bytes); |
| 21 | + printf(" %5.2f i/b ", agg.fastest_instructions() / bytes); |
| 22 | + printf(" %5.2f i/c ", agg.fastest_instructions() / agg.fastest_cycles()); |
| 23 | + } |
| 24 | + printf("\n"); |
| 25 | +} |
| 26 | + |
| 27 | +int parse_u8_fastswar(const char *&p, const char *pend, uint8_t *out) { |
| 28 | + if (p == pend) |
| 29 | + return 0; |
| 30 | + auto r = fast_float::from_chars(p, pend, *out); |
| 31 | + if (r.ec == std::errc()) { |
| 32 | + p = r.ptr; |
| 33 | + return 1; |
| 34 | + } |
| 35 | + return 0; |
| 36 | +} |
| 37 | + |
| 38 | +static inline int parse_u8_fromchars(const char *&p, const char *pend, |
| 39 | + uint8_t *out) { |
| 40 | + if (p == pend) |
| 41 | + return 0; |
| 42 | + auto r = std::from_chars(p, pend, *out); |
| 43 | + if (r.ec == std::errc()) { |
| 44 | + p = r.ptr; |
| 45 | + return 1; |
| 46 | + } |
| 47 | + return 0; |
| 48 | +} |
| 49 | + |
| 50 | +template <typename Parser> |
| 51 | +static inline int parse_ip_line(const char *&p, const char *pend, uint32_t &sum, |
| 52 | + Parser parse_uint8) { |
| 53 | + uint8_t o = 0; |
| 54 | + for (int i = 0; i < 4; ++i) { |
| 55 | + if (!parse_uint8(p, pend, &o)) |
| 56 | + return 0; |
| 57 | + sum += o; |
| 58 | + if (i != 3) { |
| 59 | + if (p == pend || *p != '.') |
| 60 | + return 0; |
| 61 | + ++p; |
| 62 | + } |
| 63 | + } |
| 64 | + // consume optional '\r' |
| 65 | + if (p != pend && *p == '\r') |
| 66 | + ++p; |
| 67 | + // expect '\n' or end |
| 68 | + if (p != pend && *p == '\n') |
| 69 | + ++p; |
| 70 | + return 1; |
| 71 | +} |
| 72 | + |
| 73 | +static std::string make_ip_line(uint8_t a, uint8_t b, uint8_t c, uint8_t d) { |
| 74 | + std::string s; |
| 75 | + s.reserve(16); |
| 76 | + s += std::to_string(a); |
| 77 | + s += '.'; |
| 78 | + s += std::to_string(b); |
| 79 | + s += '.'; |
| 80 | + s += std::to_string(c); |
| 81 | + s += '.'; |
| 82 | + s += std::to_string(d); |
| 83 | + s += '\n'; |
| 84 | + return s; |
| 85 | +} |
| 86 | + |
| 87 | +int main() { |
| 88 | + constexpr size_t N = 500000; |
| 89 | + std::mt19937 rng(1234); |
| 90 | + std::uniform_int_distribution<int> dist(0, 255); |
| 91 | + |
| 92 | + std::string buf; |
| 93 | + buf.reserve(N * 16); |
| 94 | + |
| 95 | + for (size_t i = 0; i < N; ++i) { |
| 96 | + uint8_t a = (uint8_t)dist(rng); |
| 97 | + uint8_t b = (uint8_t)dist(rng); |
| 98 | + uint8_t c = (uint8_t)dist(rng); |
| 99 | + uint8_t d = (uint8_t)dist(rng); |
| 100 | + buf += make_ip_line(a, b, c, d); |
| 101 | + } |
| 102 | + |
| 103 | + // sentinel to allow 4-byte loads at end |
| 104 | + buf.append(4, '\0'); |
| 105 | + |
| 106 | + const size_t bytes = buf.size() - 4; // exclude sentinel from throughput |
| 107 | + const size_t volume = N; |
| 108 | + |
| 109 | + // validate correctness |
| 110 | + { |
| 111 | + const char *start = buf.data(); |
| 112 | + const char *end = buf.data() + bytes; |
| 113 | + const char *p = start; |
| 114 | + const char *pend = end; |
| 115 | + uint32_t sum = 0; |
| 116 | + for (size_t i = 0; i < N; ++i) { |
| 117 | + int ok = parse_ip_line(p, pend, sum, parse_u8_fromchars); |
| 118 | + if (!ok) { |
| 119 | + std::fprintf(stderr, "fromchars parse failed at line %zu\n", i); |
| 120 | + std::abort(); |
| 121 | + } |
| 122 | + p = start; |
| 123 | + pend = end; |
| 124 | + ok = parse_ip_line(p, pend, sum, parse_u8_fastswar); |
| 125 | + if (!ok) { |
| 126 | + std::fprintf(stderr, "fastswar parse failed at line %zu\n", i); |
| 127 | + std::abort(); |
| 128 | + } |
| 129 | + } |
| 130 | + } |
| 131 | + |
| 132 | + uint32_t sink = 0; |
| 133 | + |
| 134 | + pretty_print(volume, bytes, "parse_ip_fromchars", bench([&]() { |
| 135 | + const char *p = buf.data(); |
| 136 | + const char *pend = buf.data() + bytes; |
| 137 | + uint32_t sum = 0; |
| 138 | + int ok = 0; |
| 139 | + for (size_t i = 0; i < N; ++i) { |
| 140 | + ok = parse_ip_line(p, pend, sum, parse_u8_fromchars); |
| 141 | + if (!ok) |
| 142 | + std::abort(); |
| 143 | + } |
| 144 | + sink += sum; |
| 145 | + })); |
| 146 | + |
| 147 | + pretty_print(volume, bytes, "parse_ip_fastswar", bench([&]() { |
| 148 | + const char *p = buf.data(); |
| 149 | + const char *pend = buf.data() + bytes; |
| 150 | + uint32_t sum = 0; |
| 151 | + int ok = 0; |
| 152 | + for (size_t i = 0; i < N; ++i) { |
| 153 | + ok = parse_ip_line(p, pend, sum, parse_u8_fastswar); |
| 154 | + if (!ok) |
| 155 | + std::abort(); |
| 156 | + } |
| 157 | + sink += sum; |
| 158 | + })); |
| 159 | + |
| 160 | + std::printf("sink=%u\n", sink); |
| 161 | + return 0; |
| 162 | +} |
0 commit comments