jobqueue.h

#pragma once
 
#include "gaia/config/config.h"
 
#include <atomic>
 
#include "gaia/cnt/sarray.h"
#include "gaia/config/profiler.h"
#include "gaia/core/utility.h"
#include "gaia/mt/jobhandle.h"
 
// MSVC might warn about applying additional padding around alignas usage.
// This is perfectly fine but can cause builds with warning-as-error turned on to fail.
GAIA_MSVC_WARNING_PUSH()
GAIA_MSVC_WARNING_DISABLE(4324)
 
namespace gaia {
  namespace mt {
 
    template <const uint32_t N = 1 << 12>
    class JobQueue {
      static_assert(N >= 2);
      static_assert((N & (N - 1)) == 0, "Extent of JobQueue must be a power of 2");
      static constexpr uint32_t MASK = N - 1;
 
      static_assert(sizeof(std::atomic_uint32_t) == sizeof(JobHandle));
      cnt::sarray<std::atomic_uint32_t, N> m_buffer;
      GAIA_ALIGNAS(GAIA_CACHELINE_SIZE) std::atomic_uint32_t m_bottom;
      GAIA_ALIGNAS(GAIA_CACHELINE_SIZE) std::atomic_uint32_t m_top;
 
    public:
      JobQueue() {
        clear();
      }
 
      ~JobQueue() = default;
      JobQueue(const JobQueue&) = default;
      JobQueue& operator=(const JobQueue&) = default;
      JobQueue(JobQueue&&) noexcept = default;
      JobQueue& operator=(JobQueue&&) noexcept = default;
 
      void clear() {
        m_bottom.store(0);
        m_top.store(0);
        for (auto& val: m_buffer)
          val.store(((JobHandle)JobNull_t()).value());
      }
 
      bool empty() const {
        GAIA_PROF_SCOPE(JobQueue::empty);
 
        const uint32_t b = m_bottom.load(std::memory_order_relaxed);
        const uint32_t t = m_top.load(std::memory_order_relaxed);
        return int32_t(b - t) <= 0; // b<=t, but handles overflows, too
      }
 
      GAIA_NODISCARD bool try_push(JobHandle jobHandle) {
        GAIA_PROF_SCOPE(JobQueue::try_push);
 
        const uint32_t b = m_bottom.load(std::memory_order_relaxed);
        const uint32_t t = m_top.load(std::memory_order_acquire);
        const uint32_t used = b - t;
        if (used > MASK)
          return false;
 
        m_buffer[b & MASK].store(jobHandle.value(), std::memory_order_relaxed);
        // Make sure the handle is written before we update the bottom
        std::atomic_thread_fence(std::memory_order_release);
        m_bottom.store(b + 1, std::memory_order_relaxed);
 
        return true;
      }
 
      GAIA_NODISCARD uint32_t try_push(std::span<JobHandle> jobHandles) {
        GAIA_PROF_SCOPE(JobQueue::try_push);
 
        const uint32_t cnt = (uint32_t)jobHandles.size();
        uint32_t b = m_bottom.load(std::memory_order_relaxed);
        const uint32_t t = m_top.load(std::memory_order_acquire);
        const uint32_t used = b - t;
        const uint32_t free = (MASK + 1) - used;
        const uint32_t freeFinal = core::get_min(cnt, free);
 
        for (uint32_t i = 0; i < freeFinal; i++, b++)
          m_buffer[b & MASK].store(jobHandles[i].value(), std::memory_order_relaxed);
        // Make sure handles are written before we update the bottom
        std::atomic_thread_fence(std::memory_order_release);
        m_bottom.store(b, std::memory_order_relaxed);
 
        return freeFinal;
      }
 
      GAIA_NODISCARD bool try_pop(JobHandle& jobHandle) {
        GAIA_PROF_SCOPE(JobQueue::try_pop);
 
        uint32_t jobHandleValue = ((JobHandle)JobNull_t{}).value();
 
        const uint32_t b = m_bottom.load(std::memory_order_relaxed) - 1;
        m_bottom.store(b, std::memory_order_relaxed);
        std::atomic_thread_fence(std::memory_order_seq_cst);
        uint32_t t = m_top.load(std::memory_order_relaxed);
 
        if (int(t - b) <= 0) { // t <= b, but handles overflows, too
          // non-empty queue
          jobHandleValue = m_buffer[b & MASK].load(std::memory_order_relaxed);
 
          if (t == b) {
            // last element in the queue
            const bool ret =
                m_top.compare_exchange_strong(t, t + 1, std::memory_order_seq_cst, std::memory_order_relaxed);
            m_bottom.store(b + 1, std::memory_order_relaxed);
            jobHandle = JobHandle(jobHandleValue);
            GAIA_ASSERT(jobHandle != (JobHandle)JobNull_t{});
            return ret; // false = failed race, don't use jobHandle; true = found a result
          }
 
          jobHandle = JobHandle(jobHandleValue);
          GAIA_ASSERT(jobHandle != (JobHandle)JobNull_t{});
          return true;
        }
 
        // empty queue
        m_bottom.store(b + 1, std::memory_order_relaxed);
        return false; // false = empty, don't use jobHandle
      }
 
      GAIA_NODISCARD bool try_steal(JobHandle& jobHandle) {
        GAIA_PROF_SCOPE(JobQueue::try_steal);
 
        uint32_t t = m_top.load(std::memory_order_acquire);
        std::atomic_thread_fence(std::memory_order_seq_cst);
        const uint32_t b = m_bottom.load(std::memory_order_acquire);
 
        if (int(b - t) <= 0) { // t >= b, but handles overflows, too
          jobHandle = (JobHandle)JobNull_t{};
          return true; // true + JobNull = empty, don't use jobHandle
        }
 
        const uint32_t jobHandleValue = m_buffer[t & MASK].load(std::memory_order_relaxed);
 
        // We fail if concurrent pop()/steal() operation changed the current top
        const bool ret = m_top.compare_exchange_strong(t, t + 1, std::memory_order_seq_cst, std::memory_order_relaxed);
        jobHandle = JobHandle(jobHandleValue);
        GAIA_ASSERT(jobHandle != (JobHandle)JobNull_t{});
        return ret; // false = failed race, don't use jobHandle; true = found a result
      }
    };
 
    template <class T, const uint32_t N = 1 << 12>
    class MpmcQueue {
      static_assert(N >= 2);
      static_assert((N & (N - 1)) == 0, "Extent of MpmcQueue must be a power of 2");
      static constexpr uint32_t MASK = N - 1;
 
      struct Node {
        std::atomic_uint32_t sequence{};
        T item;
      };
      using view_policy = mem::data_view_policy_aos<Node>;
 
      static constexpr uint32_t extent = N;
      static constexpr uint32_t allocated_bytes = view_policy::get_min_byte_size(0, N);
 
      // MSVC might warn about applying additional padding to an instance of StackAllocator.
      // This is perfectly fine, but might make builds with warning-as-error turned on to fail.
      GAIA_MSVC_WARNING_PUSH()
      GAIA_MSVC_WARNING_DISABLE(4324)
 
      mem::raw_data_holder<Node, allocated_bytes> m_data;
      GAIA_ALIGNAS(GAIA_CACHELINE_SIZE) std::atomic_uint32_t m_pushPos;
      GAIA_ALIGNAS(GAIA_CACHELINE_SIZE) std::atomic_uint32_t m_popPos;
 
      GAIA_MSVC_WARNING_POP()
 
    public:
      MpmcQueue() {
        init();
      }
      ~MpmcQueue() {
        free();
      }
 
      MpmcQueue(MpmcQueue&&) = delete;
      MpmcQueue(const MpmcQueue&) = delete;
      MpmcQueue& operator=(MpmcQueue&&) = delete;
      MpmcQueue& operator=(const MpmcQueue&) = delete;
 
    private:
      GAIA_NODISCARD constexpr Node* data() noexcept {
        return GAIA_ACC((Node*)&m_data[0]);
      }
 
      GAIA_NODISCARD constexpr const Node* data() const noexcept {
        return GAIA_ACC((const Node*)&m_data[0]);
      }
 
      void init() {
        Node* pNodes = data();
 
        GAIA_FOR(extent) {
          Node* pNode = &pNodes[i];
          core::call_ctor(&pNode->sequence, i);
        }
 
        m_pushPos.store(0, std::memory_order_relaxed);
        m_popPos.store(0, std::memory_order_relaxed);
      }
 
      void free() {
        Node* pNodes = data();
 
        uint32_t enqPos = m_pushPos.load(std::memory_order_relaxed);
        uint32_t deqPos = m_popPos.load(std::memory_order_relaxed);
        for (uint32_t pos = deqPos; pos != enqPos; ++pos) {
          Node* pNode = &pNodes[pos & MASK];
          if (pNode->sequence.load(std::memory_order_relaxed) == pos + 1)
            pNode->item.~T();
        }
 
        GAIA_FOR(extent) {
          Node* pNode = &pNodes[i];
          core::call_dtor(&pNode->sequence);
        }
      }
 
    public:
      bool empty() const {
        GAIA_PROF_SCOPE(MpmcQueue::empty);
 
        const uint32_t pos = m_popPos.load(std::memory_order_relaxed);
        const auto* pNode = &data()[pos & MASK];
        const uint32_t seq = pNode->sequence.load(std::memory_order_acquire);
        return pos >= seq;
      }
 
      template <typename TT>
      bool try_push(TT&& item) {
        GAIA_PROF_SCOPE(MpmcQueue::try_push);
 
        Node* pNodes = data();
 
        Node* pNode = nullptr;
        uint32_t pos = m_pushPos.load(std::memory_order_relaxed);
        while (true) {
          pNode = &pNodes[pos & MASK];
          uint32_t seq = pNode->sequence.load(std::memory_order_acquire);
          int32_t diff = int32_t(seq) - int32_t(pos);
          if (diff == 0) {
            if (m_pushPos.compare_exchange_weak(pos, pos + 1, std::memory_order_relaxed))
              break;
          } else if (diff < 0) {
            // The queue is full, we can't push
            return false;
          } else {
            pos = m_pushPos.load(std::memory_order_relaxed);
          }
        }
 
        core::call_ctor(&pNode->item, GAIA_FWD(item));
        pNode->sequence.store(pos + 1, std::memory_order_release);
        return true;
      }
 
      bool try_pop(T& item) {
        GAIA_PROF_SCOPE(MpmcQueue::try_pop);
 
        Node* pNodes = data();
 
        Node* pNode = nullptr;
        uint32_t pos = m_popPos.load(std::memory_order_relaxed);
        while (true) {
          pNode = &pNodes[pos & MASK];
          uint32_t seq = pNode->sequence.load(std::memory_order_acquire);
          int32_t diff = int32_t(seq) - int32_t(pos + 1);
          if (diff == 0) {
            if (m_popPos.compare_exchange_weak(pos, pos + 1, std::memory_order_relaxed))
              break;
          } else if (diff < 0) {
            // The queue is empty, we can't pop
            return false;
          } else {
            pos = m_popPos.load(std::memory_order_relaxed);
          }
        }
 
        item = GAIA_MOV(pNode->item);
        core::call_dtor(&pNode->item);
        pNode->sequence.store(pos + MASK + 1, std::memory_order_release);
        return true;
      }
    };
  } // namespace mt
} // namespace gaia
 
GAIA_MSVC_WARNING_POP()