diff options
author | Morph <39850852+Morph1984@users.noreply.github.com> | 2023-03-19 08:19:25 +0100 |
---|---|---|
committer | Morph <39850852+Morph1984@users.noreply.github.com> | 2023-03-22 00:17:32 +0100 |
commit | 306840a5808cae10bf5d91e4b6e8a91cd619386b (patch) | |
tree | 22d42022abbf8ab522ee046919920135d86be42c /src/common/bounded_threadsafe_queue.h | |
parent | Merge pull request #9970 from bunnei/string-util-view (diff) | |
download | yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.tar yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.tar.gz yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.tar.bz2 yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.tar.lz yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.tar.xz yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.tar.zst yuzu-306840a5808cae10bf5d91e4b6e8a91cd619386b.zip |
Diffstat (limited to '')
-rw-r--r-- | src/common/bounded_threadsafe_queue.h | 311 |
1 files changed, 199 insertions, 112 deletions
diff --git a/src/common/bounded_threadsafe_queue.h b/src/common/bounded_threadsafe_queue.h index 14e887c70..e03427539 100644 --- a/src/common/bounded_threadsafe_queue.h +++ b/src/common/bounded_threadsafe_queue.h @@ -1,159 +1,246 @@ -// SPDX-FileCopyrightText: Copyright (c) 2020 Erik Rigtorp <erik@rigtorp.se> -// SPDX-License-Identifier: MIT +// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project +// SPDX-License-Identifier: GPL-2.0-or-later #pragma once #include <atomic> -#include <bit> #include <condition_variable> -#include <memory> +#include <cstddef> #include <mutex> #include <new> -#include <type_traits> -#include <utility> #include "common/polyfill_thread.h" namespace Common { -#if defined(__cpp_lib_hardware_interference_size) -constexpr size_t hardware_interference_size = std::hardware_destructive_interference_size; -#else -constexpr size_t hardware_interference_size = 64; -#endif +namespace detail { +constexpr size_t DefaultCapacity = 0x1000; +} // namespace detail + +template <typename T, size_t Capacity = detail::DefaultCapacity> +class SPSCQueue { + static_assert((Capacity & (Capacity - 1)) == 0, "Capacity must be a power of two."); -template <typename T, size_t capacity = 0x400> -class MPSCQueue { public: - explicit MPSCQueue() : allocator{std::allocator<Slot<T>>()} { - // Allocate one extra slot to prevent false sharing on the last slot - slots = allocator.allocate(capacity + 1); - // Allocators are not required to honor alignment for over-aligned types - // (see http://eel.is/c++draft/allocator.requirements#10) so we verify - // alignment here - if (reinterpret_cast<uintptr_t>(slots) % alignof(Slot<T>) != 0) { - allocator.deallocate(slots, capacity + 1); - throw std::bad_alloc(); - } - for (size_t i = 0; i < capacity; ++i) { - std::construct_at(&slots[i]); - } - static_assert(std::has_single_bit(capacity), "capacity must be an integer power of 2"); - static_assert(alignof(Slot<T>) == hardware_interference_size, - "Slot must be aligned to cache line boundary to prevent false sharing"); - static_assert(sizeof(Slot<T>) % hardware_interference_size == 0, - "Slot size must be a multiple of cache line size to prevent " - "false sharing between adjacent slots"); - static_assert(sizeof(MPSCQueue) % hardware_interference_size == 0, - "Queue size must be a multiple of cache line size to " - "prevent false sharing between adjacent queues"); - } - - ~MPSCQueue() noexcept { - for (size_t i = 0; i < capacity; ++i) { - std::destroy_at(&slots[i]); + void Push(T&& t) { + const size_t write_index = m_write_index.load(); + + // Wait until we have free slots to write to. + while ((write_index - m_read_index.load()) == Capacity) { + std::this_thread::yield(); } - allocator.deallocate(slots, capacity + 1); + + // Determine the position to write to. + const size_t pos = write_index % Capacity; + + // Push into the queue. + m_data[pos] = std::move(t); + + // Increment the write index. + ++m_write_index; + + // Notify the consumer that we have pushed into the queue. + std::scoped_lock lock{cv_mutex}; + cv.notify_one(); } - // The queue must be both non-copyable and non-movable - MPSCQueue(const MPSCQueue&) = delete; - MPSCQueue& operator=(const MPSCQueue&) = delete; + template <typename... Args> + void Push(Args&&... args) { + const size_t write_index = m_write_index.load(); + + // Wait until we have free slots to write to. + while ((write_index - m_read_index.load()) == Capacity) { + std::this_thread::yield(); + } + + // Determine the position to write to. + const size_t pos = write_index % Capacity; + + // Emplace into the queue. + std::construct_at(std::addressof(m_data[pos]), std::forward<Args>(args)...); + + // Increment the write index. + ++m_write_index; + + // Notify the consumer that we have pushed into the queue. + std::scoped_lock lock{cv_mutex}; + cv.notify_one(); + } - MPSCQueue(MPSCQueue&&) = delete; - MPSCQueue& operator=(MPSCQueue&&) = delete; + bool TryPop(T& t) { + return Pop(t); + } - void Push(const T& v) noexcept { - static_assert(std::is_nothrow_copy_constructible_v<T>, - "T must be nothrow copy constructible"); - emplace(v); + void PopWait(T& t, std::stop_token stop_token) { + Wait(stop_token); + Pop(t); } - template <typename P, typename = std::enable_if_t<std::is_nothrow_constructible_v<T, P&&>>> - void Push(P&& v) noexcept { - emplace(std::forward<P>(v)); + T PopWait(std::stop_token stop_token) { + Wait(stop_token); + T t; + Pop(t); + return t; } - void Pop(T& v, std::stop_token stop) noexcept { - auto const tail = tail_.fetch_add(1); - auto& slot = slots[idx(tail)]; - if (!slot.turn.test()) { - std::unique_lock lock{cv_mutex}; - Common::CondvarWait(cv, lock, stop, [&slot] { return slot.turn.test(); }); + void Clear() { + while (!Empty()) { + Pop(); } - v = slot.move(); - slot.destroy(); - slot.turn.clear(); - slot.turn.notify_one(); + } + + bool Empty() const { + return m_read_index.load() == m_write_index.load(); + } + + size_t Size() const { + return m_write_index.load() - m_read_index.load(); } private: - template <typename U = T> - struct Slot { - ~Slot() noexcept { - if (turn.test()) { - destroy(); - } - } + void Pop() { + const size_t read_index = m_read_index.load(); - template <typename... Args> - void construct(Args&&... args) noexcept { - static_assert(std::is_nothrow_constructible_v<U, Args&&...>, - "T must be nothrow constructible with Args&&..."); - std::construct_at(reinterpret_cast<U*>(&storage), std::forward<Args>(args)...); + // Check if the queue is empty. + if (read_index == m_write_index.load()) { + return; } - void destroy() noexcept { - static_assert(std::is_nothrow_destructible_v<U>, "T must be nothrow destructible"); - std::destroy_at(reinterpret_cast<U*>(&storage)); - } + // Determine the position to read from. + const size_t pos = read_index % Capacity; + + // Pop the data off the queue, deleting it. + std::destroy_at(std::addressof(m_data[pos])); + + // Increment the read index. + ++m_read_index; + } - U&& move() noexcept { - return reinterpret_cast<U&&>(storage); + bool Pop(T& t) { + const size_t read_index = m_read_index.load(); + + // Check if the queue is empty. + if (read_index == m_write_index.load()) { + return false; } - // Align to avoid false sharing between adjacent slots - alignas(hardware_interference_size) std::atomic_flag turn{}; - struct aligned_store { - struct type { - alignas(U) unsigned char data[sizeof(U)]; - }; - }; - typename aligned_store::type storage; - }; + // Determine the position to read from. + const size_t pos = read_index % Capacity; - template <typename... Args> - void emplace(Args&&... args) noexcept { - static_assert(std::is_nothrow_constructible_v<T, Args&&...>, - "T must be nothrow constructible with Args&&..."); - auto const head = head_.fetch_add(1); - auto& slot = slots[idx(head)]; - slot.turn.wait(true); - slot.construct(std::forward<Args>(args)...); - slot.turn.test_and_set(); - cv.notify_one(); + // Pop the data off the queue, moving it. + t = std::move(m_data[pos]); + + // Increment the read index. + ++m_read_index; + + return true; } - constexpr size_t idx(size_t i) const noexcept { - return i & mask; + void Wait(std::stop_token stop_token) { + std::unique_lock lock{cv_mutex}; + Common::CondvarWait(cv, lock, stop_token, [this] { return !Empty(); }); } - static constexpr size_t mask = capacity - 1; + alignas(128) std::atomic_size_t m_read_index{0}; + alignas(128) std::atomic_size_t m_write_index{0}; - // Align to avoid false sharing between head_ and tail_ - alignas(hardware_interference_size) std::atomic<size_t> head_{0}; - alignas(hardware_interference_size) std::atomic<size_t> tail_{0}; + std::array<T, Capacity> m_data; - std::mutex cv_mutex; std::condition_variable_any cv; + std::mutex cv_mutex; +}; + +template <typename T, size_t Capacity = detail::DefaultCapacity> +class MPSCQueue { +public: + void Push(T&& t) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::move(t)); + } + + template <typename... Args> + void Push(Args&&... args) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::forward<Args>(args)...); + } + + bool TryPop(T& t) { + return spsc_queue.TryPop(t); + } + + void PopWait(T& t, std::stop_token stop_token) { + spsc_queue.PopWait(t, stop_token); + } + + T PopWait(std::stop_token stop_token) { + return spsc_queue.PopWait(stop_token); + } + + void Clear() { + spsc_queue.Clear(); + } + + bool Empty() { + return spsc_queue.Empty(); + } + + size_t Size() { + return spsc_queue.Size(); + } + +private: + SPSCQueue<T, Capacity> spsc_queue; + std::mutex write_mutex; +}; + +template <typename T, size_t Capacity = detail::DefaultCapacity> +class MPMCQueue { +public: + void Push(T&& t) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::move(t)); + } - Slot<T>* slots; - [[no_unique_address]] std::allocator<Slot<T>> allocator; + template <typename... Args> + void Push(Args&&... args) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::forward<Args>(args)...); + } - static_assert(std::is_nothrow_copy_assignable_v<T> || std::is_nothrow_move_assignable_v<T>, - "T must be nothrow copy or move assignable"); + bool TryPop(T& t) { + std::scoped_lock lock{read_mutex}; + return spsc_queue.TryPop(t); + } + + void PopWait(T& t, std::stop_token stop_token) { + std::scoped_lock lock{read_mutex}; + spsc_queue.PopWait(t, stop_token); + } - static_assert(std::is_nothrow_destructible_v<T>, "T must be nothrow destructible"); + T PopWait(std::stop_token stop_token) { + std::scoped_lock lock{read_mutex}; + return spsc_queue.PopWait(stop_token); + } + + void Clear() { + std::scoped_lock lock{read_mutex}; + spsc_queue.Clear(); + } + + bool Empty() { + std::scoped_lock lock{read_mutex}; + return spsc_queue.Empty(); + } + + size_t Size() { + std::scoped_lock lock{read_mutex}; + return spsc_queue.Size(); + } + +private: + SPSCQueue<T, Capacity> spsc_queue; + std::mutex write_mutex; + std::mutex read_mutex; }; } // namespace Common |