Renderive/Core/execution/Render_Executor.cpp

#include "Render_Executor.h"
#include "../architecture/Render_Time.h"
#include <algorithm>
#include <array>
#include <atomic>
#include <memory>
#include <vector>
#include <taskflow/taskflow.hpp>
#include <thread>

namespace renderive {
struct Taskflow_Graph_Impl {
    explicit Taskflow_Graph_Impl(tf::Taskflow& taskflow) : taskflow(taskflow) {}
    tf::Taskflow& taskflow;
    std::vector<tf::Task> tasks;
};

Render_Task_Graph::Render_Task_Graph(void* impl) noexcept : impl(impl) {}

Render_Task_Graph_Node Render_Task_Graph::emplace(Task work) {
    auto* graph = static_cast<Taskflow_Graph_Impl*>(impl);
    if (!graph || !work)
        return {};
    auto work_ptr = std::make_shared<Task>(std::move(work));
    auto task = graph->taskflow.emplace([work_ptr]() mutable {
        if (work_ptr && *work_ptr)
            (*work_ptr)();
    });
    graph->tasks.push_back(task);
    return {static_cast<std::uint32_t>(graph->tasks.size() - 1)};
}

void Render_Task_Graph::precede(Render_Task_Graph_Node before, Render_Task_Graph_Node after) {
    auto* graph = static_cast<Taskflow_Graph_Impl*>(impl);
    if (!graph || !before || !after)
        return;
    if (before.index >= graph->tasks.size() || after.index >= graph->tasks.size())
        return;
    graph->tasks[before.index].precede(graph->tasks[after.index]);
}

namespace {

struct Render_Executor_Metrics {
    std::size_t worker_count{};
    std::atomic<std::size_t> active_workers{0};
    std::atomic<std::size_t> active_frame_jobs{0};
    std::atomic<std::size_t> queued_frame_jobs{0};
    std::atomic<std::uint64_t> submitted_tasks{0};
    std::atomic<std::uint64_t> started_tasks{0};
    std::atomic<std::uint64_t> completed_tasks{0};
    std::atomic<std::uint64_t> rejected_frame_jobs{0};
    std::atomic<std::uint64_t> dropped_frame_attempts{0};
    std::atomic<std::uint64_t> peak_concurrency{0};
    std::atomic<std::uint64_t> queue_wait_total_ns{0};
    std::atomic<std::uint64_t> queue_wait_max_ns{0};
    std::atomic<std::uint64_t> task_duration_total_ns{0};
    std::atomic<std::uint64_t> task_duration_max_ns{0};
    std::atomic<std::uint64_t> curve_tasks{0};
    std::atomic<std::uint64_t> waterfall_tasks{0};
    std::atomic<std::uint64_t> primitive_tasks{0};
    std::atomic<std::uint64_t> compose_tasks{0};
    std::atomic<std::uint64_t> performance_tasks{0};
};

void update_peak(std::atomic<std::uint64_t>& target, std::uint64_t value) {
    std::uint64_t current = target.load(std::memory_order_acquire);
    while (current < value && !target.compare_exchange_weak(current, value, std::memory_order_acq_rel, std::memory_order_acquire)) {}
}

void add_kind_count(Render_Executor_Metrics& metrics, Render_Task_Kind kind) {
    switch (kind) {
    case Render_Task_Kind::Frame:
    case Render_Task_Kind::Primitive:
        metrics.primitive_tasks.fetch_add(1, std::memory_order_relaxed);
        break;
    case Render_Task_Kind::Curve:
        metrics.curve_tasks.fetch_add(1, std::memory_order_relaxed);
        break;
    case Render_Task_Kind::Waterfall:
        metrics.waterfall_tasks.fetch_add(1, std::memory_order_relaxed);
        break;
    case Render_Task_Kind::Compose:
        metrics.compose_tasks.fetch_add(1, std::memory_order_relaxed);
        break;
    case Render_Task_Kind::Performance:
        metrics.performance_tasks.fetch_add(1, std::memory_order_relaxed);
        break;
    }
}

} // namespace

class Render_Executor_Observer final : public tf::ObserverInterface {
public:
    explicit Render_Executor_Observer(std::shared_ptr<Render_Executor_Metrics> metrics)
        : metrics(std::move(metrics)) {}

    void set_up(std::size_t worker_count) override {
        if (metrics)
            metrics->worker_count = worker_count;
        worker_entry_ns.clear();
        worker_entry_ns.resize(worker_count);
    }

    void on_entry(tf::WorkerView worker, tf::TaskView) override {
        if (!metrics)
            return;
        std::size_t worker_id = worker.id();
        if (worker_id < worker_entry_ns.size())
            worker_entry_ns[worker_id] = steady_now_ns();
        std::size_t active = metrics->active_workers.fetch_add(1, std::memory_order_acq_rel) + 1;
        metrics->started_tasks.fetch_add(1, std::memory_order_relaxed);
        update_peak(metrics->peak_concurrency, static_cast<std::uint64_t>(active));
    }

    void on_exit(tf::WorkerView worker, tf::TaskView) override {
        if (!metrics)
            return;
        std::size_t worker_id = worker.id();
        if (worker_id < worker_entry_ns.size()) {
            std::uint64_t begin_ns = worker_entry_ns[worker_id];
            std::uint64_t end_ns = steady_now_ns();
            std::uint64_t duration_ns = end_ns > begin_ns ? end_ns - begin_ns : 0;
            metrics->task_duration_total_ns.fetch_add(duration_ns, std::memory_order_relaxed);
            update_peak(metrics->task_duration_max_ns, duration_ns);
        }
        metrics->completed_tasks.fetch_add(1, std::memory_order_relaxed);
        metrics->active_workers.fetch_sub(1, std::memory_order_acq_rel);
    }

private:
    std::shared_ptr<Render_Executor_Metrics> metrics;
    std::vector<std::uint64_t> worker_entry_ns;
};

class Render_Executor_Private {
    struct Task_Run_Record {
        std::uint64_t begin_ns{};
        std::uint64_t queue_wait_ns{};
    };

public:
    explicit Render_Executor_Private(std::size_t worker_count)
        : metrics(std::make_shared<Render_Executor_Metrics>()),
          executor(worker_count) {
        metrics->worker_count = worker_count;
        observer = executor.make_observer<Render_Executor_Observer>(metrics);
    }

    ~Render_Executor_Private() {
        shutdown();
    }

    bool try_submit(Render_Executor_Task task) {
        if (shutting_down.load(std::memory_order_acquire) || (!task.work && !task.build_graph)) {
            reject(task);
            return false;
        }

        std::uint64_t enqueue_ns = steady_now_ns();
        record_submission(task);
        submit_to_executor(std::move(task), enqueue_ns);
        return true;
    }

    void shutdown() {
        if (shutting_down.exchange(true, std::memory_order_acq_rel))
            return;
        executor.wait_for_all();
    }

    Render_Executor_Snapshot snapshot() const {
        std::uint64_t submitted = metrics->submitted_tasks.load(std::memory_order_acquire);
        std::uint64_t completed = metrics->completed_tasks.load(std::memory_order_acquire);
        std::uint64_t started = metrics->started_tasks.load(std::memory_order_acquire);
        std::uint64_t queue_total = metrics->queue_wait_total_ns.load(std::memory_order_acquire);
        std::uint64_t duration_total = metrics->task_duration_total_ns.load(std::memory_order_acquire);
        std::size_t active = metrics->active_workers.load(std::memory_order_acquire);
        std::size_t workers = metrics->worker_count;
        return {
            workers,
            active,
            workers > active ? workers - active : 0,
            metrics->active_frame_jobs.load(std::memory_order_acquire),
            metrics->queued_frame_jobs.load(std::memory_order_acquire),
            submitted,
            started,
            completed,
            metrics->rejected_frame_jobs.load(std::memory_order_acquire),
            metrics->dropped_frame_attempts.load(std::memory_order_acquire),
            metrics->peak_concurrency.load(std::memory_order_acquire),
            submitted ? queue_total / submitted : 0,
            metrics->queue_wait_max_ns.load(std::memory_order_acquire),
            completed ? duration_total / completed : 0,
            metrics->task_duration_max_ns.load(std::memory_order_acquire),
            metrics->curve_tasks.load(std::memory_order_acquire),
            metrics->waterfall_tasks.load(std::memory_order_acquire),
            metrics->primitive_tasks.load(std::memory_order_acquire),
            metrics->compose_tasks.load(std::memory_order_acquire),
            metrics->performance_tasks.load(std::memory_order_acquire)
        };
    }

    std::size_t worker_count() const {
        return metrics->worker_count;
    }

private:
    void record_submission(Render_Executor_Task& task) {
        metrics->submitted_tasks.fetch_add(1, std::memory_order_relaxed);
        add_kind_count(*metrics, task.kind);
        if (task.frame_job)
            metrics->queued_frame_jobs.fetch_add(1, std::memory_order_relaxed);
        if (task.frame_stat)
            task.frame_stat->executor_at_submit = snapshot();
    }

    void submit_to_executor(Render_Executor_Task task, std::uint64_t enqueue_ns) {
        auto task_ptr = std::make_shared<Render_Executor_Task>(std::move(task));
        auto completion = std::make_shared<Task>(std::move(task_ptr->completion));
        auto run_record = std::make_shared<Task_Run_Record>();
        auto topology = std::make_shared<tf::Taskflow>();
        auto entry = topology->emplace([this, enqueue_ns, task_ptr, run_record]() mutable {
            Render_Executor_Task& task = *task_ptr;
            std::uint64_t begin_ns = steady_now_ns();
            run_record->begin_ns = begin_ns;
            std::uint64_t queue_wait_ns = begin_ns > enqueue_ns ? begin_ns - enqueue_ns : 0;
            run_record->queue_wait_ns = queue_wait_ns;
            metrics->queue_wait_total_ns.fetch_add(queue_wait_ns, std::memory_order_relaxed);
            update_peak(metrics->queue_wait_max_ns, queue_wait_ns);
            if (task.frame_job) {
                metrics->queued_frame_jobs.fetch_sub(1, std::memory_order_relaxed);
                metrics->active_frame_jobs.fetch_add(1, std::memory_order_relaxed);
            }
        });
        auto exit = topology->emplace([this, task_ptr, run_record]() mutable {
            Render_Executor_Task& task = *task_ptr;
            std::uint64_t end_ns = steady_now_ns();
            std::uint64_t begin_ns = run_record->begin_ns;
            std::uint64_t queue_wait_ns = run_record->queue_wait_ns;
            std::uint64_t run_ns = end_ns > begin_ns ? end_ns - begin_ns : 0;
            if (task.frame_stat) {
                Frame_Worker_Stats& stat = *task.frame_stat;
                stat.task_count += std::max<std::uint32_t>(1, task.logical_task_count);
                stat.peak_parallelism = static_cast<std::uint32_t>(std::max<std::uint64_t>(stat.peak_parallelism, metrics->peak_concurrency.load(std::memory_order_acquire)));
                stat.queue_wait_total_ns += queue_wait_ns;
                stat.queue_wait_max_ns = std::max(stat.queue_wait_max_ns, queue_wait_ns);
                stat.worker_run_total_ns += run_ns;
                stat.parallel_stage_wall_ns += run_ns;
                stat.executor_at_finish = snapshot();
            }
            if (task.frame_job)
                metrics->active_frame_jobs.fetch_sub(1, std::memory_order_relaxed);
        });
        if (task_ptr->build_graph) {
            Taskflow_Graph_Impl graph_impl(*topology);
            graph_impl.tasks.push_back(entry);
            graph_impl.tasks.push_back(exit);
            Render_Task_Graph graph(&graph_impl);
            task_ptr->build_graph(graph, {0}, {1});
        }
        else {
            auto work = topology->emplace([task_ptr]() mutable {
                if (task_ptr->work)
                    task_ptr->work();
            });
            entry.precede(work);
            work.precede(exit);
        }
        executor.run(*topology, [this, topology, completion, task_ptr]() mutable {
            (void)this;
            if (completion && *completion)
                (*completion)();
        });
    }

    void reject(const Render_Executor_Task& task) {
        if (task.frame_job)
            metrics->rejected_frame_jobs.fetch_add(1, std::memory_order_relaxed);
        else
            metrics->dropped_frame_attempts.fetch_add(1, std::memory_order_relaxed);
        if (task.frame_stat)
            task.frame_stat->rejected_task_count++;
    }

    std::shared_ptr<Render_Executor_Metrics> metrics;
    tf::Executor executor;
    std::shared_ptr<tf::ObserverInterface> observer;
    std::atomic_bool shutting_down{false};
};

std::size_t resolve_render_worker_count(std::size_t configured_count) {
    if (configured_count)
        return configured_count;
    auto cpu_count = std::thread::hardware_concurrency();
    return cpu_count > 1 ? cpu_count - 1 : 1;
}

Render_Executor::Render_Executor(Render_Runtime_Config config)
    : d(std::make_unique<Render_Executor_Private>(resolve_render_worker_count(config.worker_count))) {}

Render_Executor::~Render_Executor() = default;

bool Render_Executor::try_submit(Render_Executor_Task task) {
    return d->try_submit(std::move(task));
}

void Render_Executor::shutdown() {
    d->shutdown();
}

Render_Executor_Snapshot Render_Executor::snapshot() const {
    return d->snapshot();
}

std::size_t Render_Executor::worker_count() const {
    return d->worker_count();
}

} // namespace renderive