|
| 1 | +// Copyright (c) 2024, DeepLink. |
| 2 | +#include "AscendDeviceActivity.h" |
| 3 | + |
| 4 | +#include <ctime> |
| 5 | +#include <dirent.h> |
| 6 | +#include <output_base.h> |
| 7 | +#include <sys/stat.h> |
| 8 | +#include <unistd.h> |
| 9 | + |
| 10 | +#include "csrc_dipu/base/environ.hpp" |
| 11 | +#include "csrc_dipu/utils/Log.h" |
| 12 | + |
| 13 | +namespace dipu { |
| 14 | + |
| 15 | +static const uint64_t kNpuEvents = 431; |
| 16 | +static const uint64_t kAicoreMetrics = 1; |
| 17 | + |
| 18 | +using libkineto::GenericTraceActivity; |
| 19 | + |
| 20 | +AscendDeviceActivity::AscendDeviceActivity() {} |
| 21 | + |
| 22 | +AscendDeviceActivity& AscendDeviceActivity::instance() { |
| 23 | + static AscendDeviceActivity instance; |
| 24 | + return instance; |
| 25 | +} |
| 26 | + |
| 27 | +// override pure virtual function, do nothing |
| 28 | +void AscendDeviceActivity::pushCorrelationID( |
| 29 | + uint64_t id, DeviceActivityInterface::CorrelationFlowType type) {} |
| 30 | + |
| 31 | +// override pure virtual function, do nothing |
| 32 | +void AscendDeviceActivity::popCorrelationID( |
| 33 | + DeviceActivityInterface::CorrelationFlowType type) {} |
| 34 | + |
| 35 | +// override pure virtual function, do nothing |
| 36 | +void AscendDeviceActivity::enableActivities( |
| 37 | + const std::set<libkineto::ActivityType>& selected_activities) {} |
| 38 | + |
| 39 | +// override pure virtual function, do nothing |
| 40 | +void AscendDeviceActivity::disableActivities( |
| 41 | + const std::set<libkineto::ActivityType>& selected_activities) {} |
| 42 | + |
| 43 | +// override pure virtual function, do nothing |
| 44 | +void AscendDeviceActivity::clearActivities() {} |
| 45 | + |
| 46 | +bool AscendDeviceActivity::remove_temp_dump_path_(const std::string& path) { |
| 47 | + std::unique_ptr<DIR, decltype(&closedir)> dir(opendir(path.c_str()), |
| 48 | + &closedir); |
| 49 | + if (!dir) { |
| 50 | + return false; |
| 51 | + } |
| 52 | + |
| 53 | + dirent* entry; |
| 54 | + while ((entry = readdir(dir.get())) != nullptr) { |
| 55 | + std::string entry_name = entry->d_name; |
| 56 | + if (entry_name == "." || entry_name == "..") { |
| 57 | + continue; |
| 58 | + } |
| 59 | + |
| 60 | + std::string entry_path = path + "/" + entry_name; |
| 61 | + struct stat entry_stat; |
| 62 | + if (stat(entry_path.c_str(), &entry_stat) == -1) { |
| 63 | + return false; |
| 64 | + } |
| 65 | + |
| 66 | + if (S_ISDIR(entry_stat.st_mode)) { |
| 67 | + if (!remove_temp_dump_path_(entry_path)) { |
| 68 | + return false; |
| 69 | + } |
| 70 | + } else { |
| 71 | + if (remove(entry_path.c_str()) != 0) { |
| 72 | + return false; |
| 73 | + } |
| 74 | + } |
| 75 | + } |
| 76 | + return rmdir(path.c_str()) == 0; |
| 77 | +} |
| 78 | + |
| 79 | +int32_t AscendDeviceActivity::processActivities( |
| 80 | + libkineto::ActivityLogger& logger, |
| 81 | + std::function<const libkineto::ITraceActivity*(int32_t)> linked_activity, |
| 82 | + int64_t start_time, int64_t end_time) { |
| 83 | + struct timespec tx; |
| 84 | + struct timespec ts; |
| 85 | + struct timespec ty; |
| 86 | + |
| 87 | + DIPU_CALLACLRT(aclprofFinalize()); |
| 88 | + |
| 89 | + // clang-format off |
| 90 | + // the difference between torch realtime and monotonic_raw, |
| 91 | + // as well as the difference between aclprof realtime and monotonic_raw, |
| 92 | + // are statistically calculated to align the timestamps of aclprof and torch |
| 93 | + // for example: torch's time difference is 10, and aclprof's time difference |
| 94 | + // is 3, we assume that monotonic_raw time is the correct baseline, then, |
| 95 | + // the time diff between torch and aclprof is 10 - 3 = 7 |
| 96 | + // aclprof's time difference is in generated file by msprof tool: PROF_XXXX/host/end_info |
| 97 | + // clang-format on |
| 98 | + |
| 99 | + clock_gettime(CLOCK_REALTIME, &tx); |
| 100 | + clock_gettime(CLOCK_MONOTONIC_RAW, &ts); |
| 101 | + clock_gettime(CLOCK_REALTIME, &ty); |
| 102 | + |
| 103 | + // clang-format off |
| 104 | + // time interval between these two calculation(aclprof and torch) is the smaller the better |
| 105 | + // and aclprof will generate its time difference in start_info and end_info |
| 106 | + // the reason why I use aclprof end_info and put the calculation right after aclprofFinalize is that, |
| 107 | + // after test, this is more precise and less stable than calculate at startTrace() and compare with start_info |
| 108 | + // clang-format on |
| 109 | + |
| 110 | + const int S_TO_NS_FACTOR = 1000000000; |
| 111 | + auto time_x = tx.tv_sec * S_TO_NS_FACTOR + tx.tv_nsec; |
| 112 | + auto monotonic_raw = ts.tv_sec * S_TO_NS_FACTOR + ts.tv_nsec; |
| 113 | + auto time_y = ty.tv_sec * S_TO_NS_FACTOR + ty.tv_nsec; |
| 114 | + |
| 115 | + int64_t diff = (time_x >> 1) + (time_y >> 1) - monotonic_raw; |
| 116 | + |
| 117 | + GenericTraceActivity time_diff; |
| 118 | + time_diff.activityName = "torch_time_diff:" + std::to_string(diff); |
| 119 | + time_diff.activityType = libkineto::ActivityType::USER_ANNOTATION; |
| 120 | + time_diff.startTime = start_time; |
| 121 | + time_diff.endTime = end_time; |
| 122 | + |
| 123 | + GenericTraceActivity tmp_path; |
| 124 | + tmp_path.activityName = "random_temp_dir:" + current_dump_path_; |
| 125 | + tmp_path.activityType = libkineto::ActivityType::USER_ANNOTATION; |
| 126 | + tmp_path.startTime = start_time; |
| 127 | + tmp_path.endTime = end_time; |
| 128 | + logger.handleGenericActivity(tmp_path); |
| 129 | + logger.handleGenericActivity(time_diff); |
| 130 | + |
| 131 | + std::string temp_path_prefix = "./tmp/aclprof"; |
| 132 | + if (last_dump_path_.compare(0, temp_path_prefix.size(), temp_path_prefix) == |
| 133 | + 0) { |
| 134 | + if (remove_temp_dump_path_(last_dump_path_) == false) { |
| 135 | + DIPU_LOGW( |
| 136 | + "remove ascend profiler temp file failed, may need to remove " |
| 137 | + "manually"); |
| 138 | + } |
| 139 | + } |
| 140 | + |
| 141 | + return 0; |
| 142 | +} |
| 143 | + |
| 144 | +void AscendDeviceActivity::startTrace( |
| 145 | + const std::set<libkineto::ActivityType>& selected_activities) { |
| 146 | + if (enable_) { |
| 147 | + DIPU_LOGW("ascend profiler has already enabled"); |
| 148 | + return; |
| 149 | + } |
| 150 | + enable_ = true; |
| 151 | + |
| 152 | + last_dump_path_ = current_dump_path_; |
| 153 | + |
| 154 | + struct stat st; |
| 155 | + if (stat("./tmp", &st) == -1) { |
| 156 | + mkdir("./tmp", 0644); |
| 157 | + } |
| 158 | + if (stat("./tmp/aclprof", &st) == -1) { |
| 159 | + mkdir("./tmp/aclprof", 0644); |
| 160 | + } |
| 161 | + |
| 162 | + char dump_path_template[] = "./tmp/aclprof/aclprofXXXXXX"; |
| 163 | + char* dump_path_cstring = mkdtemp(dump_path_template); |
| 164 | + |
| 165 | + if (dump_path_cstring != nullptr) { |
| 166 | + current_dump_path_ = dump_path_cstring; |
| 167 | + } else { |
| 168 | + DIPU_LOGE( |
| 169 | + "aclprof random dump path generate failed, the export results may be " |
| 170 | + "incorrect"); |
| 171 | + current_dump_path_ = "./tmp/aclprof/aclprof_error"; |
| 172 | + } |
| 173 | + |
| 174 | + DIPU_CALLACLRT( |
| 175 | + aclprofInit(current_dump_path_.c_str(), current_dump_path_.size())); |
| 176 | + DIPU_CALLACLRT(aclrtSynchronizeDevice()); |
| 177 | + |
| 178 | + int32_t device_index = 0; |
| 179 | + DIPU_CALLACLRT(aclrtGetDevice(&device_index)); |
| 180 | + |
| 181 | + std::array<uint32_t, 1> device_ids = {static_cast<uint32_t>(device_index)}; |
| 182 | + aclprofAicoreEvents* events = nullptr; |
| 183 | + config_ = aclprofCreateConfig( |
| 184 | + device_ids.data(), device_ids.size(), |
| 185 | + static_cast<aclprofAicoreMetrics>( kAicoreMetrics), events, kNpuEvents); |
| 186 | + TORCH_CHECK(config_ != nullptr, |
| 187 | + "aclprofCreateConfig fail, device_index = ", device_index, |
| 188 | + "npu_event = ", kNpuEvents, "aicore_metrics = ", kAicoreMetrics); |
| 189 | + |
| 190 | + DIPU_CALLACLRT(aclprofStart(config_)); |
| 191 | +} |
| 192 | + |
| 193 | +void AscendDeviceActivity::stopTrace( |
| 194 | + const std::set<libkineto::ActivityType>& selected_activities) { |
| 195 | + if (!enable_) { |
| 196 | + DIPU_LOGW("ascend profiler has already disabled"); |
| 197 | + return; |
| 198 | + } |
| 199 | + |
| 200 | + DIPU_CALLACLRT(aclrtSynchronizeDevice()); |
| 201 | + DIPU_CALLACLRT(aclprofStop(config_)); |
| 202 | + DIPU_CALLACLRT(aclprofDestroyConfig(config_)) |
| 203 | + |
| 204 | + config_ = nullptr; |
| 205 | + enable_ = false; |
| 206 | +} |
| 207 | + |
| 208 | +// override pure virtual function, do nothing |
| 209 | +void AscendDeviceActivity::teardownContext() {} |
| 210 | + |
| 211 | +// override pure virtual function, do nothing |
| 212 | +void AscendDeviceActivity::setMaxBufferSize(int32_t size) {} |
| 213 | + |
| 214 | +// NOLINTNEXTLINE(cppcoreguidelines-interfaces-global-init) |
| 215 | +const static int32_t Ascend_device_activity_init = []() { |
| 216 | + if (!dipu::environ::detail::getEnvOrDefault<bool>("FORCE_USE_DIPU_PROFILER", |
| 217 | + false)) { |
| 218 | + libkineto::device_activity_singleton = &AscendDeviceActivity::instance(); |
| 219 | + return 1; |
| 220 | + } |
| 221 | + return 0; |
| 222 | +}(); |
| 223 | + |
| 224 | +} // namespace dipu |
0 commit comments