QueueGenericSyclBase.hpp

Go to the documentation of this file.

/* Copyright 2024 Jan Stephan, Antonio Di Pilato, Luca Ferragina, Andrea Bocci, Aurora Perego
 * SPDX-License-Identifier: MPL-2.0
 */
 
#pragma once
 
#include "alpaka/acc/Tag.hpp"
#include "alpaka/dev/Traits.hpp"
#include "alpaka/event/Traits.hpp"
#include "alpaka/queue/Traits.hpp"
#include "alpaka/traits/Traits.hpp"
#include "alpaka/wait/Traits.hpp"
 
#include <algorithm>
#include <exception>
#include <memory>
#include <mutex>
#include <shared_mutex>
#include <type_traits>
#include <utility>
#include <vector>
 
#ifdef ALPAKA_ACC_SYCL_ENABLED
 
#    include <sycl/sycl.hpp>
 
namespace alpaka
{
    template<concepts::Tag TTag>
    class DevGenericSycl;
 
    template<concepts::Tag TTag>
    class EventGenericSycl;
 
    namespace detail
    {
        template<typename T, typename = void>
        inline constexpr auto is_sycl_task = false;
 
        template<typename T>
        inline constexpr auto is_sycl_task<T, std::void_t<decltype(T::is_sycl_task)>> = true;
 
        template<typename T, typename = void>
        inline constexpr auto is_sycl_kernel = false;
 
        template<typename T>
        inline constexpr auto is_sycl_kernel<T, std::void_t<decltype(T::is_sycl_kernel)>> = true;
 
        class QueueGenericSyclImpl
        {
        public:
            QueueGenericSyclImpl(sycl::context context, sycl::device device)
                : m_queue{
                    std::move(context), // This is important. In SYCL a device can belong to multiple contexts.
                    std::move(device),
                    {sycl::property::queue::enable_profiling{}, sycl::property::queue::in_order{}}}
            {
            }
 
            // This class will only exist as a pointer. We don't care about copy and move semantics.
            QueueGenericSyclImpl(QueueGenericSyclImpl const& other) = delete;
            auto operator=(QueueGenericSyclImpl const& rhs) -> QueueGenericSyclImpl& = delete;
 
            QueueGenericSyclImpl(QueueGenericSyclImpl&& other) noexcept = delete;
            auto operator=(QueueGenericSyclImpl&& rhs) noexcept -> QueueGenericSyclImpl& = delete;
 
            ~QueueGenericSyclImpl()
            {
                try
                {
                    m_queue.wait_and_throw();
                }
                catch(sycl::exception const& err)
                {
                    std::cerr << "Caught SYCL exception while destructing a SYCL queue: " << err.what() << " ("
                              << err.code() << ')' << std::endl;
                }
                catch(std::exception const& err)
                {
                    std::cerr << "The following runtime error(s) occured while destructing a SYCL queue:" << err.what()
                              << std::endl;
                }
            }
 
            // Don't call this without locking first!
            auto clean_dependencies() -> void
            {
                // Clean up completed events
                auto const start = std::begin(m_dependencies);
                auto const old_end = std::end(m_dependencies);
                auto const new_end = std::remove_if(
                    start,
                    old_end,
                    [](sycl::event ev) {
                        return ev.get_info<sycl::info::event::command_execution_status>()
                               == sycl::info::event_command_status::complete;
                    });
 
                m_dependencies.erase(new_end, old_end);
            }
 
            auto register_dependency(sycl::event event) -> void
            {
                std::lock_guard<std::shared_mutex> lock{m_mutex};
 
                clean_dependencies();
                m_dependencies.push_back(event);
            }
 
            auto empty() const -> bool
            {
                std::shared_lock<std::shared_mutex> lock{m_mutex};
                return m_last_event.get_info<sycl::info::event::command_execution_status>()
                       == sycl::info::event_command_status::complete;
            }
 
            auto wait() -> void
            {
                // SYCL queues are thread-safe.
                m_queue.wait_and_throw();
            }
 
            auto get_last_event() const -> sycl::event
            {
                std::shared_lock<std::shared_mutex> lock{m_mutex};
                return m_last_event;
            }
 
            // Perfect forwarding of the task is a backend-specific workaround for oneAPI SYCL internally converting
            // host tasks to std::function&&. The portable way is to pass the task by const reference, as done by
            // the other backends.
            template<bool TBlocking, typename TTask>
            auto enqueue(TTask&& task) -> void
            {
                {
                    using TaskType = std::decay_t<TTask>;
                    std::lock_guard<std::shared_mutex> lock{m_mutex};
 
                    clean_dependencies();
 
                    // Execute task
                    if constexpr(is_sycl_task<TaskType> && !is_sycl_kernel<TaskType>) // Copy / Fill
                    {
                        m_last_event = task(m_queue, m_dependencies); // Will call queue.{copy, fill} internally
                    }
                    else
                    {
                        m_last_event = m_queue.submit(
                            [this, captured_task = std::forward<TTask>(task)](sycl::handler& cgh) mutable
                            {
                                if(!m_dependencies.empty())
                                    cgh.depends_on(m_dependencies);
 
                                if constexpr(is_sycl_kernel<TaskType>) // Kernel
                                    captured_task(cgh); // Will call cgh.parallel_for internally
                                else // Host
                                    cgh.host_task(std::move(captured_task));
                            });
                    }
 
                    m_dependencies.clear();
                }
 
                if constexpr(TBlocking)
                    wait();
            }
 
            [[nodiscard]] auto getNativeHandle() const noexcept
            {
                return m_queue;
            }
 
            std::vector<sycl::event> m_dependencies;
            sycl::event m_last_event;
            std::shared_mutex mutable m_mutex;
 
        private:
            sycl::queue m_queue;
        };
 
        template<concepts::Tag TTag, bool TBlocking>
        class QueueGenericSyclBase
            : public interface::Implements<ConceptCurrentThreadWaitFor, QueueGenericSyclBase<TTag, TBlocking>>
            , public interface::Implements<ConceptQueue, QueueGenericSyclBase<TTag, TBlocking>>
            , public interface::Implements<ConceptGetDev, QueueGenericSyclBase<TTag, TBlocking>>
        {
        public:
            QueueGenericSyclBase(DevGenericSycl<TTag> const& dev)
                : m_dev{dev}
                , m_spQueueImpl{std::make_shared<detail::QueueGenericSyclImpl>(
                      dev.getNativeHandle().second,
                      dev.getNativeHandle().first)}
            {
                m_dev.m_impl->register_queue(m_spQueueImpl);
            }
 
            friend auto operator==(QueueGenericSyclBase const& lhs, QueueGenericSyclBase const& rhs) -> bool
            {
                return (lhs.m_dev == rhs.m_dev) && (lhs.m_spQueueImpl == rhs.m_spQueueImpl);
            }
 
            friend auto operator!=(QueueGenericSyclBase const& lhs, QueueGenericSyclBase const& rhs) -> bool
            {
                return !(lhs == rhs);
            }
 
            [[nodiscard]] auto getNativeHandle() const noexcept
            {
                return m_spQueueImpl->getNativeHandle();
            }
 
            DevGenericSycl<TTag> m_dev;
            std::shared_ptr<detail::QueueGenericSyclImpl> m_spQueueImpl;
        };
    } // namespace detail
 
    namespace trait
    {
        //! The SYCL blocking queue device type trait specialization.
        template<concepts::Tag TTag, bool TBlocking>
        struct DevType<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>>
        {
            using type = DevGenericSycl<TTag>;
        };
 
        //! The SYCL blocking queue device get trait specialization.
        template<concepts::Tag TTag, bool TBlocking>
        struct GetDev<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>>
        {
            static auto getDev(alpaka::detail::QueueGenericSyclBase<TTag, TBlocking> const& queue)
            {
                ALPAKA_DEBUG_MINIMAL_LOG_SCOPE;
                return queue.m_dev;
            }
        };
 
        //! The SYCL blocking queue event type trait specialization.
        template<concepts::Tag TTag, bool TBlocking>
        struct EventType<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>>
        {
            using type = EventGenericSycl<TTag>;
        };
 
        //! The SYCL blocking queue enqueue trait specialization.
        template<concepts::Tag TTag, bool TBlocking, typename TTask>
        struct Enqueue<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>, TTask>
        {
            template<typename UTask>
            static auto enqueue(alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>& queue, UTask&& task) -> void
            {
                ALPAKA_DEBUG_MINIMAL_LOG_SCOPE;
                queue.m_spQueueImpl->template enqueue<TBlocking>(std::forward<UTask>(task));
            }
        };
 
        //! The SYCL blocking queue test trait specialization.
        template<concepts::Tag TTag, bool TBlocking>
        struct Empty<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>>
        {
            static auto empty(alpaka::detail::QueueGenericSyclBase<TTag, TBlocking> const& queue) -> bool
            {
                ALPAKA_DEBUG_MINIMAL_LOG_SCOPE;
                return queue.m_spQueueImpl->empty();
            }
        };
 
        //! The SYCL blocking queue thread wait trait specialization.
        //!
        //! Blocks execution of the calling thread until the queue has finished processing all previously requested
        //! tasks (kernels, data copies, ...)
        template<concepts::Tag TTag, bool TBlocking>
        struct CurrentThreadWaitFor<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>>
        {
            static auto currentThreadWaitFor(alpaka::detail::QueueGenericSyclBase<TTag, TBlocking> const& queue)
                -> void
            {
                ALPAKA_DEBUG_MINIMAL_LOG_SCOPE;
                queue.m_spQueueImpl->wait();
            }
        };
 
        //! The SYCL queue native handle trait specialization.
        template<concepts::Tag TTag, bool TBlocking>
        struct NativeHandle<alpaka::detail::QueueGenericSyclBase<TTag, TBlocking>>
        {
            [[nodiscard]] static auto getNativeHandle(
                alpaka::detail::QueueGenericSyclBase<TTag, TBlocking> const& queue)
            {
                return queue.getNativeHandle();
            }
        };
    } // namespace trait
} // namespace alpaka
#endif