framework/instruments/SchedulerTimer.cpp

/*
 * Copyright (c) 2017-2021 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include "SchedulerTimer.h"

#include "Instruments.h"
#include "WallClockTimer.h"
#include "arm_compute/core/CPP/ICPPKernel.h"
#include "arm_compute/graph/DataLayerVisitor.h"
#include "arm_compute/graph/INode.h"
#include "support/Cast.h"

namespace arm_compute
{
namespace test
{
namespace framework
{
template <bool output_timestamps>
std::string    SchedulerClock<output_timestamps>::id() const
{
    if(output_timestamps)
    {
        return "SchedulerTimestamps";
    }
    else
    {
        return "SchedulerTimer";
    }
}

template <bool    output_timestamps>
class Interceptor final : public IScheduler
{
public:
    /** Default constructor. */
    Interceptor(std::list<struct SchedulerClock<output_timestamps>::kernel_info> &kernels,
                std::map<std::string, SchedulerTimer::LayerData> &layers, IScheduler &real_scheduler,
                ScaleFactor scale_factor)
        : _kernels(kernels), _layer_data_map(layers), _real_scheduler(real_scheduler), _timer(scale_factor), _prefix()
    {
    }

    void set_num_threads(unsigned int num_threads) override
    {
        _real_scheduler.set_num_threads(num_threads);
    }

    void set_num_threads_with_affinity(unsigned int num_threads, BindFunc func) override
    {
        _real_scheduler.set_num_threads_with_affinity(num_threads, func);
    }

    unsigned int num_threads() const override
    {
        return _real_scheduler.num_threads();
    }

    void set_prefix(const std::string &prefix)
    {
        _prefix = prefix;
    }

    void schedule(ICPPKernel *kernel, const Hints &hints) override
    {
        _timer.start();
        _real_scheduler.schedule(kernel, hints);
        _timer.stop();

        typename SchedulerClock<output_timestamps>::kernel_info info;
        info.name         = kernel->name();
        info.prefix       = _prefix;
        info.measurements = _timer.measurements();
        _kernels.push_back(std::move(info));
    }

    void schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors) override
    {
        _timer.start();
        _real_scheduler.schedule_op(kernel, hints, window, tensors);
        _timer.stop();

        typename SchedulerClock<output_timestamps>::kernel_info info;
        info.name         = kernel->name();
        info.prefix       = _prefix;
        info.measurements = _timer.measurements();
        _kernels.push_back(std::move(info));
    }

    void run_tagged_workloads(std::vector<Workload> &workloads, const char *tag) override
    {
        _timer.start();
        _real_scheduler.run_tagged_workloads(workloads, tag);
        _timer.stop();

        typename SchedulerClock<output_timestamps>::kernel_info info;
        info.name         = tag != nullptr ? tag : "Unknown";
        info.prefix       = _prefix;
        info.measurements = _timer.measurements();
        _kernels.push_back(std::move(info));
    }

protected:
    void run_workloads(std::vector<Workload> &workloads) override
    {
        ARM_COMPUTE_UNUSED(workloads);
        ARM_COMPUTE_ERROR("Can't be reached");
    }

private:
    std::list<struct SchedulerClock<output_timestamps>::kernel_info> &_kernels;
    std::map<std::string, SchedulerTimer::LayerData> &_layer_data_map;
    IScheduler                  &_real_scheduler;
    WallClock<output_timestamps> _timer;
    std::string                  _prefix;
};

template <bool output_timestamps>
SchedulerClock<output_timestamps>::SchedulerClock(ScaleFactor scale_factor)
    : _kernels(),
      _layer_data_map(),
      _real_scheduler(nullptr),
      _real_scheduler_type(),
#ifdef ARM_COMPUTE_GRAPH_ENABLED
      _real_graph_function(nullptr),
#endif /* ARM_COMPUTE_GRAPH_ENABLED */
      _scale_factor(scale_factor),
      _interceptor(nullptr),
      _scheduler_users()
{
    if(instruments_info != nullptr)
    {
        _scheduler_users = instruments_info->_scheduler_users;
    }
}

template <bool output_timestamps>
void           SchedulerClock<output_timestamps>::test_start()
{
#ifdef ARM_COMPUTE_GRAPH_ENABLED
    // Start intercepting tasks:
    ARM_COMPUTE_ERROR_ON(_real_graph_function != nullptr);
    _real_graph_function  = graph::TaskExecutor::get().execute_function;
    auto task_interceptor = [this](graph::ExecutionTask & task)
    {
        Interceptor<output_timestamps> *scheduler = nullptr;
        if(dynamic_cast<Interceptor<output_timestamps> *>(this->_interceptor.get()) != nullptr)
        {
            scheduler = arm_compute::utils::cast::polymorphic_downcast<Interceptor<output_timestamps> *>(_interceptor.get());
            if(task.node != nullptr && !task.node->name().empty())
            {
                scheduler->set_prefix(task.node->name() + "/");

                if(_layer_data_map.find(task.node->name()) == _layer_data_map.end())
                {
                    arm_compute::graph::DataLayerVisitor dlv = {};
                    task.node->accept(dlv);
                    _layer_data_map[task.node->name()] = dlv.layer_data();
                }
            }
            else
            {
                scheduler->set_prefix("");
            }
        }

        this->_real_graph_function(task);

        if(scheduler != nullptr)
        {
            scheduler->set_prefix("");
        }
    };
#endif /* ARM_COMPUTE_GRAPH_ENABLED */

    ARM_COMPUTE_ERROR_ON(_real_scheduler != nullptr);
    _real_scheduler_type = Scheduler::get_type();
    //Note: We can't currently replace a custom scheduler
    if(_real_scheduler_type != Scheduler::Type::CUSTOM)
    {
        _real_scheduler = &Scheduler::get();
        _interceptor    = std::make_shared<Interceptor<output_timestamps>>(_kernels, _layer_data_map, *_real_scheduler, _scale_factor);
        Scheduler::set(std::static_pointer_cast<IScheduler>(_interceptor));
#ifdef ARM_COMPUTE_GRAPH_ENABLED
        graph::TaskExecutor::get().execute_function = task_interceptor;
#endif /* ARM_COMPUTE_GRAPH_ENABLED */

        // Create an interceptor for each scheduler
        // TODO(COMPID-2638) : Allow multiple schedulers, now it assumes the same scheduler is used.
        std::for_each(std::begin(_scheduler_users), std::end(_scheduler_users),
                      [&](ISchedulerUser * user)
        {
            if(user != nullptr && user->scheduler() != nullptr)
            {
                user->intercept_scheduler(std::make_unique<Interceptor<output_timestamps>>(_kernels, _layer_data_map, *user->scheduler(), _scale_factor));
            }
        });
    }
}

template <bool output_timestamps>
void           SchedulerClock<output_timestamps>::start()
{
    _kernels.clear();
}

template <bool output_timestamps>
void           SchedulerClock<output_timestamps>::test_stop()
{
    // Restore real scheduler
    Scheduler::set(_real_scheduler_type);
    _real_scheduler = nullptr;
    _interceptor    = nullptr;
#ifdef ARM_COMPUTE_GRAPH_ENABLED
    graph::TaskExecutor::get().execute_function = _real_graph_function;
    _real_graph_function                        = nullptr;
#endif /* ARM_COMPUTE_GRAPH_ENABLED */

    // Restore schedulers
    std::for_each(std::begin(_scheduler_users), std::end(_scheduler_users),
                  [&](ISchedulerUser * user)
    {
        if(user != nullptr)
        {
            user->restore_scheduler();
        }
    });
}

template <bool              output_timestamps>
Instrument::MeasurementsMap SchedulerClock<output_timestamps>::measurements() const
{
    MeasurementsMap measurements;
    unsigned int    kernel_number = 0;
    for(auto kernel : _kernels)
    {
        std::string name = kernel.prefix + kernel.name + " #" + support::cpp11::to_string(kernel_number++);
        if(output_timestamps)
        {
            ARM_COMPUTE_ERROR_ON(kernel.measurements.size() != 2);
            for(auto const &m : kernel.measurements)
            {
                if(m.first.find("[start]") != std::string::npos)
                {
                    measurements.emplace("[start]" + name, m.second);
                }
                else if(m.first.find("[end]") != std::string::npos)
                {
                    measurements.emplace("[end]" + name, m.second);
                }
                else
                {
                    ARM_COMPUTE_ERROR("Measurement not handled");
                }
            }
        }
        else
        {
            measurements.emplace(name, kernel.measurements.begin()->second);
        }
    }

    return measurements;
}

template <bool output_timestamps>
std::string    SchedulerClock<output_timestamps>::instrument_header() const
{
    std::string output{ "" };
    output += R"("layer_data" : {)";
    for(auto i_it = _layer_data_map.cbegin(), i_end = _layer_data_map.cend(); i_it != i_end; ++i_it)
    {
        output += "\"" + i_it->first + "\" : {";
        if(i_it->second.size() != 0)
        {
            // Print for each entry in layer
            for(auto entry_it = i_it->second.cbegin(), entry_end = i_it->second.cend(); entry_it != entry_end; ++entry_it)
            {
                output += "\"" + entry_it->first + "\" : \"" + entry_it->second + "\"";
                if(std::next(entry_it) != entry_end)
                {
                    output += ",";
                }
            }
        }
        output += "}";
        if(std::next(i_it) != i_end)
        {
            output += ",";
        }
    }
    output += "}";
    return output;
}

} // namespace framework
} // namespace test
} // namespace arm_compute

template class arm_compute::test::framework::SchedulerClock<true>;
template class arm_compute::test::framework::SchedulerClock<false>;