alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. All rights reserved. |
| 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | */ |
| 17 | #include "Profiler.hpp" |
| 18 | |
| 19 | #include <cstring> |
| 20 | #include <string> |
| 21 | #include <sstream> |
| 22 | |
| 23 | namespace arm { |
| 24 | namespace app { |
| 25 | |
| 26 | template<class T> |
| 27 | static void writeStatLine(std::ostringstream& s, |
| 28 | const char* desc, |
| 29 | T total, |
| 30 | uint32_t samples, |
| 31 | T min, |
| 32 | T max) |
| 33 | { |
| 34 | s << "\t" << desc << total << " / " |
| 35 | << ((double)total / samples) << " / " |
| 36 | << min << " / " << max << std::endl; |
| 37 | } |
| 38 | |
| 39 | Profiler::Profiler(hal_platform* platform, const char* name = "Unknown") |
| 40 | : _m_name(name) |
| 41 | { |
| 42 | if (platform && platform->inited) { |
| 43 | this->_m_pPlatform = platform; |
| 44 | this->Reset(); |
| 45 | } else { |
| 46 | printf_err("Profiler %s initialised with invalid platform\n", |
| 47 | this->_m_name.c_str()); |
| 48 | } |
| 49 | } |
| 50 | |
| 51 | bool Profiler::StartProfiling(const char* name) |
| 52 | { |
| 53 | if (name) { |
| 54 | this->SetName(name); |
| 55 | } |
| 56 | if (this->_m_pPlatform && !this->_m_started) { |
| 57 | this->_m_pPlatform->timer->reset(); |
| 58 | this->_m_tstampSt = this->_m_pPlatform->timer->start_profiling(); |
| 59 | this->_m_started = true; |
| 60 | return true; |
| 61 | } |
| 62 | printf_err("Failed to start profiler %s\n", this->_m_name.c_str()); |
| 63 | return false; |
| 64 | } |
| 65 | |
| 66 | bool Profiler::StopProfiling() |
| 67 | { |
| 68 | if (this->_m_pPlatform && this->_m_started) { |
| 69 | this->_m_tstampEnd = this->_m_pPlatform->timer->stop_profiling(); |
| 70 | this->_m_started = false; |
| 71 | |
| 72 | this->_AddProfilingUnit(this->_m_tstampSt, this->_m_tstampEnd, this->_m_name); |
| 73 | |
| 74 | return true; |
| 75 | } |
| 76 | printf_err("Failed to stop profiler %s\n", this->_m_name.c_str()); |
| 77 | return false; |
| 78 | } |
| 79 | |
| 80 | bool Profiler::StopProfilingAndReset() |
| 81 | { |
| 82 | if (this->StopProfiling()) { |
| 83 | this->Reset(); |
| 84 | return true; |
| 85 | } |
| 86 | printf_err("Failed to stop profiler %s\n", this->_m_name.c_str()); |
| 87 | return false; |
| 88 | } |
| 89 | |
| 90 | void Profiler::Reset() |
| 91 | { |
| 92 | this->_m_started = false; |
| 93 | memset(&this->_m_tstampSt, 0, sizeof(this->_m_tstampSt)); |
| 94 | memset(&this->_m_tstampEnd, 0, sizeof(this->_m_tstampEnd)); |
| 95 | } |
| 96 | |
| 97 | std::string Profiler::GetResultsAndReset() |
| 98 | { |
| 99 | std::ostringstream strResults; |
| 100 | |
| 101 | for (const auto& item: this->_m_series) { |
| 102 | auto name = item.first; |
| 103 | ProfilingSeries series = item.second; |
| 104 | |
| 105 | uint32_t samplesNum = series.size(); |
| 106 | |
| 107 | uint64_t totalNpuCycles = 0; /* Total NPU cycles (idle + active). */ |
| 108 | uint64_t totalActiveNpuCycles = 0; /* Active NPU cycles. */ |
| 109 | uint64_t totalCpuCycles = 0; /* Total CPU cycles. */ |
| 110 | time_t totalTimeMs = 0; |
| 111 | |
| 112 | uint64_t minActiveNpuCycles = series[0].activeNpuCycles; |
| 113 | uint64_t minIdleNpuCycles = series[0].npuCycles - minActiveNpuCycles; |
| 114 | uint64_t minActiveCpuCycles = series[0].cpuCycles - minActiveNpuCycles; |
| 115 | time_t minTimeMs = series[0].time; |
| 116 | |
| 117 | uint64_t maxIdleNpuCycles = 0; |
| 118 | uint64_t maxActiveNpuCycles = 0; |
| 119 | uint64_t maxActiveCpuCycles = 0; |
| 120 | time_t maxTimeMs = 0; |
| 121 | |
| 122 | for(ProfilingUnit& unit: series){ |
| 123 | totalNpuCycles += unit.npuCycles; |
| 124 | totalActiveNpuCycles += unit.activeNpuCycles; |
| 125 | totalCpuCycles += unit.cpuCycles; |
| 126 | totalTimeMs += unit.time; |
| 127 | |
| 128 | maxActiveNpuCycles = std::max(maxActiveNpuCycles, |
| 129 | unit.activeNpuCycles); |
| 130 | maxIdleNpuCycles = std::max(maxIdleNpuCycles, |
| 131 | unit.npuCycles - maxActiveNpuCycles); |
| 132 | maxActiveCpuCycles = std::max(maxActiveCpuCycles, |
| 133 | unit.cpuCycles - maxActiveNpuCycles); |
| 134 | maxTimeMs = std::max(maxTimeMs, unit.time); |
| 135 | |
| 136 | minActiveNpuCycles = std::min(minActiveNpuCycles, |
| 137 | unit.activeNpuCycles); |
| 138 | minIdleNpuCycles = std::min(minIdleNpuCycles, |
| 139 | unit.npuCycles - minActiveNpuCycles); |
| 140 | minActiveCpuCycles = std::min(minActiveCpuCycles, |
| 141 | unit.cpuCycles - minActiveNpuCycles); |
| 142 | minTimeMs = std::min(minTimeMs, unit.time); |
| 143 | } |
| 144 | |
| 145 | strResults << "Profile for " << name << ": " << std::endl; |
| 146 | |
| 147 | if (samplesNum > 1) { |
| 148 | strResults << "\tSamples: " << samplesNum << std::endl; |
| 149 | strResults << "\t Total / Avg./ Min / Max" |
| 150 | << std::endl; |
| 151 | |
| 152 | writeStatLine<uint64_t>(strResults, "Active NPU cycles: ", |
| 153 | totalActiveNpuCycles, samplesNum, |
| 154 | minActiveNpuCycles, maxActiveNpuCycles); |
| 155 | |
| 156 | writeStatLine<uint64_t>(strResults, "Idle NPU cycles: ", |
| 157 | (totalNpuCycles - totalActiveNpuCycles), |
| 158 | samplesNum, minIdleNpuCycles, maxIdleNpuCycles); |
| 159 | |
| 160 | #if defined(CPU_PROFILE_ENABLED) |
| 161 | writeStatLine<uint64_t>(strResults, "Active CPU cycles (approx): ", |
| 162 | (totalCpuCycles - totalActiveNpuCycles), |
| 163 | samplesNum, minActiveCpuCycles, |
| 164 | maxActiveCpuCycles); |
| 165 | |
| 166 | writeStatLine<time_t>(strResults, "Time in ms: ", |
| 167 | totalTimeMs, samplesNum, minTimeMs, maxTimeMs); |
| 168 | #endif |
| 169 | } else { |
| 170 | strResults << "\tActive NPU cycles: " << totalActiveNpuCycles |
| 171 | << std::endl; |
| 172 | strResults << "\tIdle NPU cycles: " |
| 173 | << (totalNpuCycles - totalActiveNpuCycles) |
| 174 | << std::endl; |
| 175 | #if defined(CPU_PROFILE_ENABLED) |
| 176 | strResults << "\tActive CPU cycles: " |
| 177 | << (totalCpuCycles - totalActiveNpuCycles) |
| 178 | << " (approx)" << std::endl; |
| 179 | |
| 180 | strResults << "\tTime in ms: " << totalTimeMs << std::endl; |
| 181 | #endif |
| 182 | } |
| 183 | } |
| 184 | this->Reset(); |
| 185 | return strResults.str(); |
| 186 | } |
| 187 | |
| 188 | void Profiler::SetName(const char* str) |
| 189 | { |
| 190 | this->_m_name = std::string(str); |
| 191 | } |
| 192 | |
| 193 | void Profiler::_AddProfilingUnit(time_counter start, time_counter end, |
| 194 | const std::string& name) |
| 195 | { |
| 196 | platform_timer * timer = this->_m_pPlatform->timer; |
| 197 | |
| 198 | struct ProfilingUnit unit; |
| 199 | |
| 200 | if (timer->cap.npu_cycles && timer->get_npu_total_cycle_diff && |
| 201 | timer->get_npu_active_cycle_diff) |
| 202 | { |
| 203 | unit.npuCycles = timer->get_npu_total_cycle_diff(&start, &end); |
| 204 | unit.activeNpuCycles = timer->get_npu_active_cycle_diff(&start, &end); |
| 205 | } |
| 206 | |
| 207 | if (timer->cap.cpu_cycles && timer->get_cpu_cycle_diff) { |
| 208 | unit.cpuCycles = timer->get_cpu_cycle_diff(&start, &end); |
| 209 | } |
| 210 | |
| 211 | if (timer->cap.duration_ms && timer->get_duration_ms) { |
| 212 | unit.time = timer->get_duration_ms(&start, &end); |
| 213 | } |
| 214 | |
| 215 | this->_m_series[name].emplace_back(unit); |
| 216 | } |
| 217 | |
| 218 | } /* namespace app */ |
| 219 | } /* namespace arm */ |