path: root/src/armnn/Profiling.cpp

//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Profiling.hpp"
#include "JsonPrinter.hpp"

#if ARMNN_STREAMLINE_ENABLED
#include <streamline_annotate.h>
#endif

#include <algorithm>
#include <iomanip>
#include <iostream>
#include <fstream>
#include <map>
#include <stack>

#include <boost/algorithm/string.hpp>
#include <boost/core/ignore_unused.hpp>
namespace armnn
{

// Controls the amount of memory initially allocated to store profiling events.
// If chosen carefully, the profiling system will not make any additional allocations, thus minimizing its impact on
// measured times.
constexpr std::size_t g_ProfilingEventCountHint = 1024;

// Whether profiling reports should include the sequence of events together with their timings.
constexpr bool g_WriteProfilingEventSequence = true;

// Whether profiling reports should also report detailed information on events grouped by inference.
// This can spam the output stream, so use carefully (or adapt the code to just output information
// of interest).
constexpr bool g_AggregateProfilingEventsByInference = true;

// Whether a call to Profiler::AnalyzeEventsAndWriteResults() will be made when the Profiler is destroyed.
// It can be convenient for local tests.
constexpr bool g_WriteReportToStdOutOnProfilerDestruction = false;

Measurement FindMeasurement(const std::string& name, const Event* event)
{

    BOOST_ASSERT(event != nullptr);

    // Search though the measurements.
    for (const auto& measurement : event->GetMeasurements())
    {
        if (measurement.m_Name == name)
        {
            // Measurement found.
            return measurement;
        }
    }

    // Measurement not found.
    return Measurement{ "", 0.f, Measurement::Unit::TIME_MS };
}

std::vector<Measurement> FindKernelMeasurements(const Event* event)
{
    BOOST_ASSERT(event != nullptr);

    std::vector<Measurement> measurements;

    // Search through the measurements.
    for (const auto& measurement : event->GetMeasurements())
    {
        if (measurement.m_Name.rfind("OpenClKernelTimer", 0) == 0
            || measurement.m_Name.rfind("NeonKernelTimer", 0) == 0)
        {
            // Measurement found.
            measurements.push_back(measurement);
        }
    }

    return measurements;
}

std::map<std::string, Profiler::ProfilingEventStats> Profiler::CalculateProfilingEventStats() const
{
    std::map<std::string, ProfilingEventStats> nameToStatsMap;

    for (const auto& event : m_EventSequence)
    {
        Measurement measurement = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME, event.get());

        double durationMs = measurement.m_Value;
        auto it = nameToStatsMap.find(event->GetName());
        if (it != nameToStatsMap.end())
        {
            ProfilingEventStats& stats = it->second;
            stats.m_TotalMs += durationMs;
            stats.m_MinMs = std::min(stats.m_MinMs, durationMs);
            stats.m_MaxMs = std::max(stats.m_MaxMs, durationMs);
            ++stats.m_Count;
        }
        else
        {
            nameToStatsMap.emplace(event->GetName(), ProfilingEventStats{ durationMs, durationMs, durationMs, 1 });
        }
    }

    return nameToStatsMap;
}

const Event* GetEventPtr(const Event* ptr) { return ptr;}
const Event* GetEventPtr(const std::unique_ptr<Event>& ptr) {return ptr.get(); }

template<typename ItertType>
void Profiler::AnalyzeEventSequenceAndWriteResults(ItertType first, ItertType last, std::ostream& outStream) const
{
    // Outputs event sequence, if needed.
    if (g_WriteProfilingEventSequence)
    {
        // Makes sure timestamps are output with 6 decimals, and save old settings.
        std::streamsize oldPrecision = outStream.precision();
        outStream.precision(6);
        std::ios_base::fmtflags oldFlags = outStream.flags();
        outStream.setf(std::ios::fixed);
        // Outputs fields.
        outStream << "Event Sequence - Name | Duration (ms) | Start (ms) | Stop (ms) | Device" << std::endl;
        for (auto event = first; event != last; ++event)
        {
            const Event* eventPtr = GetEventPtr((*event));
            double startTimeMs = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME_START, eventPtr).m_Value;
            double stopTimeMs = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME_STOP, eventPtr).m_Value;

            // Find the WallClock measurement if there is one.
            double durationMs = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME, eventPtr).m_Value;
            outStream << std::setw(50) << eventPtr->GetName() << " "
                      << std::setw(20) << durationMs
                      << std::setw(20) << startTimeMs
                      << std::setw(20) << stopTimeMs
                      << std::setw(20) << GetComputeDeviceAsCString(eventPtr->GetComputeDevice())
                      << std::endl;
        }
        outStream << std::endl;
        // Restores previous precision settings.
        outStream.flags(oldFlags);
        outStream.precision(oldPrecision);
    }

    // Aggregates results per event name.
    std::map<std::string, ProfilingEventStats> nameToStatsMap = CalculateProfilingEventStats();

    // Outputs aggregated stats.
    outStream << "Event Stats - Name | Avg (ms) | Min (ms) | Max (ms) | Total (ms) | Count" << std::endl;
    for (const auto& pair : nameToStatsMap)
    {
        const std::string& eventLabel = pair.first;
        const ProfilingEventStats& eventStats = pair.second;
        const double avgMs = eventStats.m_TotalMs / double(eventStats.m_Count);

        outStream << "\t" << std::setw(50) << eventLabel << " " << std::setw(9) << avgMs << " "
            << std::setw(9) << eventStats.m_MinMs << " " << std::setw(9) << eventStats.m_MaxMs << " "
            << std::setw(9) << eventStats.m_TotalMs << " " << std::setw(9) << eventStats.m_Count << std::endl;
    }
    outStream << std::endl;
}

Profiler::Profiler()
    : m_ProfilingEnabled(false)
{
    m_EventSequence.reserve(g_ProfilingEventCountHint);

#if ARMNN_STREAMLINE_ENABLED
    // Initialises streamline annotations.
    ANNOTATE_SETUP;
#endif
}

Profiler::~Profiler()
{
    if (m_ProfilingEnabled)
    {
        if (g_WriteReportToStdOutOnProfilerDestruction)
        {
            Print(std::cout);
        }
    }

    // Un-register this profiler from the current thread.
    ProfilerManager::GetInstance().RegisterProfiler(nullptr);
}

bool Profiler::IsProfilingEnabled()
{
    return m_ProfilingEnabled;
}

void Profiler::EnableProfiling(bool enableProfiling)
{
    m_ProfilingEnabled = enableProfiling;
}

Event* Profiler::BeginEvent(Compute compute, const std::string& label, std::vector<InstrumentPtr>&& instruments)
{
    Event* parent = m_Parents.empty() ? nullptr : m_Parents.top();
    m_EventSequence.push_back(std::make_unique<Event>(label, this, parent, compute, std::move(instruments)));
    Event* event = m_EventSequence.back().get();
    event->Start();

#if ARMNN_STREAMLINE_ENABLED
    ANNOTATE_CHANNEL_COLOR(m_Parents.size(), GetEventColor(compute), label.c_str());
#endif

    m_Parents.push(event);
    return event;
}

void Profiler::EndEvent(Event* event)
{
    event->Stop();

    BOOST_ASSERT(!m_Parents.empty());
    BOOST_ASSERT(event == m_Parents.top());
    m_Parents.pop();

    Event* parent = m_Parents.empty() ? nullptr : m_Parents.top();
    boost::ignore_unused(parent);
    BOOST_ASSERT(event->GetParentEvent() == parent);

#if ARMNN_STREAMLINE_ENABLED
    ANNOTATE_CHANNEL_END(m_Parents.size());
#endif
}

int CalcLevel(const Event* eventPtr)
{
    int level=0;
    while (eventPtr != nullptr)
    {
        eventPtr = eventPtr->GetParentEvent();
        level++;
    }
    return level;
}

void Profiler::PopulateInferences(std::vector<const Event*>& outInferences, int& outBaseLevel) const
{
    outInferences.reserve(m_EventSequence.size());
    for (const auto& event : m_EventSequence)
    {
        const Event* eventPtrRaw = event.get();
        if (eventPtrRaw->GetName() == "EnqueueWorkload")
        {
            outBaseLevel = (outBaseLevel == -1) ? CalcLevel(eventPtrRaw) : outBaseLevel;
            outInferences.push_back(eventPtrRaw);
        }
    }
}

void Profiler::PopulateDescendants(std::map<const Event*, std::vector<const Event*>>& outDescendantsMap) const
{
    for (const auto& event : m_EventSequence)
    {
        const Event* eventPtrRaw = event.get();
        const Event* parent = eventPtrRaw->GetParentEvent();

        if (!parent)
        {
            continue;
        }

        auto it = outDescendantsMap.find(parent);
        if (it == outDescendantsMap.end())
        {
            outDescendantsMap.emplace(parent, std::vector<const Event*>({eventPtrRaw}));
        }
        else
        {
            it->second.push_back(eventPtrRaw);
        }
    }
}

void Profiler::Print(std::ostream& outStream) const
{
    // Makes sure timestamps are output with 6 decimals, and save old settings.
    std::streamsize oldPrecision = outStream.precision();
    outStream.precision(6);
    std::ios_base::fmtflags oldFlags = outStream.flags();
    outStream.setf(std::ios::fixed);
    JsonPrinter printer(outStream);

    // First find all the "inference" Events and print out duration measurements.
    int baseLevel = -1;
    std::vector<const Event*> inferences;
    PopulateInferences(inferences, baseLevel);

    // Second map out descendants hierarchy
    std::map<const Event*, std::vector<const Event*>> descendantsMap;
    PopulateDescendants(descendantsMap);

    JsonChildObject inferenceObject{"inference_measurements"};
    JsonChildObject layerObject{"layer_measurements"};
    std::vector<JsonChildObject> workloadObjects;
    std::map<unsigned int, std::vector<JsonChildObject>> workloadToKernelObjects;

    for (unsigned int inferenceIndex = 0; inferenceIndex < inferences.size(); ++inferenceIndex)
    {
        auto inference = inferences[inferenceIndex];
        Measurement measurement = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME, inference);
        inferenceObject.SetUnit(measurement.m_Unit);
        inferenceObject.AddMeasurement(measurement.m_Value);

        auto layerEventsIt = descendantsMap.find(inference);

        // Assuming 1 Execute per inference
        if (layerEventsIt != descendantsMap.end())
        {
            auto layerEvent = layerEventsIt->second[0];
            Measurement measurement = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME, layerEvent);
            layerObject.SetUnit(measurement.m_Unit);
            layerObject.AddMeasurement(measurement.m_Value);

            // Get Descendant Events for Execute
            auto workloadEventsIt = descendantsMap.find(layerEvent);
            for(unsigned int workloadIndex = 0; workloadIndex < workloadEventsIt->second.size(); ++workloadIndex)
            {
                auto workloadEvent = workloadEventsIt->second[workloadIndex];
                Measurement measurement = FindMeasurement(WallClockTimer::WALL_CLOCK_TIME, workloadEvent);
                std::vector<Measurement> kernelMeasurements = FindKernelMeasurements(workloadEvent);
                if (inferenceIndex == 0)
                {
                    // Only add second level once, in case of multiple inferences
                    JsonChildObject workloadObject{workloadEvent->GetName()};
                    workloadObject.SetUnit(measurement.m_Unit);
                    workloadObjects.push_back(workloadObject);
                }
                workloadObjects[workloadIndex].AddMeasurement(measurement.m_Value);

                for(unsigned int kernelIndex = 0; kernelIndex < kernelMeasurements.size(); ++kernelIndex)
                {
                    if (inferenceIndex == 0)
                    {
                        // Only add kernel measurement once, in case of multiple inferences
                        JsonChildObject kernelObject{kernelMeasurements[kernelIndex].m_Name};
                        kernelObject.SetUnit(kernelMeasurements[kernelIndex].m_Unit);
                        workloadToKernelObjects[workloadIndex].push_back(kernelObject);

                    }
                    workloadToKernelObjects[workloadIndex][kernelIndex].
                            AddMeasurement(kernelMeasurements[kernelIndex].m_Value);
                }
            }
        }
    }

    for (auto workloadToKernelPair : workloadToKernelObjects)
    {
        for (auto kernelObject : workloadToKernelPair.second)
        {
            workloadObjects[workloadToKernelPair.first].AddChild(kernelObject);
        }
    }

    for (auto workloadObject : workloadObjects)
    {
        layerObject.AddChild(workloadObject);
    }
    inferenceObject.AddChild(layerObject);

    printer.PrintHeader();
    printer.PrintArmNNHeader();

    // print inference object, also prints child layer and kernel measurements
    printer.PrintJsonChildObject(inferenceObject);

    // end of ArmNN
    printer.PrintNewLine();
    printer.PrintFooter();

    // end of main JSON object
    printer.PrintNewLine();
    printer.PrintFooter();
    printer.PrintNewLine();

    // Restores previous precision settings.
    outStream.flags(oldFlags);
    outStream.precision(oldPrecision);
}

void Profiler::AnalyzeEventsAndWriteResults(std::ostream& outStream) const
{
    // Stack should be empty now.
    const bool saneMarkerSequence = m_Parents.empty();

    // Abort if the sequence of markers was found to have incorrect information:
    // The stats cannot be trusted.
    if (!saneMarkerSequence)
    {
        outStream << "Cannot write profiling stats. "
            "Unexpected errors were found when analyzing the sequence of logged events, which may lead to plainly "
            "wrong stats. The profiling system may contain implementation issues or could have been used in an "
            "unsafe manner." << std::endl;
        return;
    }

    // Analyzes the full sequence of events.
    AnalyzeEventSequenceAndWriteResults(m_EventSequence.cbegin(),
                                        m_EventSequence.cend(),
                                        outStream);

    // Aggregates events by tag if requested (spams the output stream if done for all tags).
    if (g_AggregateProfilingEventsByInference)
    {
        outStream << std::endl;
        outStream << "***" << std::endl;
        outStream << "*** Per Inference Stats" << std::endl;
        outStream << "***" << std::endl;
        outStream << std::endl;

        int baseLevel = -1;
        std::vector<const Event*> inferences;
        PopulateInferences(inferences, baseLevel);

        // Second map out descendants hierarchy
        std::map<const Event*, std::vector<const Event*>> descendantsMap;
        PopulateDescendants(descendantsMap);

        std::function<void (const Event*, std::vector<const Event*>&)>
            FindDescendantEvents = [&](const Event* eventPtr,
                std::vector<const Event*>& sequence)
            {
                sequence.push_back(eventPtr);

                if (CalcLevel(eventPtr) > baseLevel+2) //We only care about levels as deep as workload executions.
                {
                    return;
                }

                auto children = descendantsMap.find(eventPtr);
                if (children == descendantsMap.end())
                {
                    return;
                }

                for (const Event* child : children->second)
                {
                    return FindDescendantEvents(child, sequence);
                }
            };

        // Third, find events belonging to each inference
        int inferenceIdx = 0;
        for (auto inference : inferences)
        {
            std::vector<const Event*> sequence;

            //build sequence, depth first
            FindDescendantEvents(inference, sequence);

            outStream << "> Begin Inference: " << inferenceIdx << std::endl;
            outStream << std::endl;
            AnalyzeEventSequenceAndWriteResults(sequence.cbegin(),
                                                sequence.cend(),
                                                outStream);
            outStream << std::endl;
            outStream << "> End Inference: " << inferenceIdx << std::endl;

            inferenceIdx++;
        }
    }
}

std::uint32_t Profiler::GetEventColor(Compute compute) const
{
    switch(compute)
    {
        case Compute::CpuRef:
            // Cyan
            return 0xffff001b;
        case Compute::CpuAcc:
            // Green
            return 0x00ff001b;
        case Compute::GpuAcc:
            // Purple
            return 0xff007f1b;
        default:
            // Dark gray
            return 0x5555551b;
    }
}

// The thread_local pointer to the profiler instance.
thread_local Profiler* tl_Profiler = nullptr;

ProfilerManager& ProfilerManager::GetInstance()
{
    // Global reference to the single ProfileManager instance allowed.
    static ProfilerManager s_ProfilerManager;
    return s_ProfilerManager;
}

void ProfilerManager::RegisterProfiler(Profiler* profiler)
{
    tl_Profiler = profiler;
}

Profiler* ProfilerManager::GetProfiler()
{
    return tl_Profiler;
}

} // namespace armnn